Virtual Barrels Quantitative Trading in The Oil Market

Download as pdf or txt
Download as pdf or txt
You are on page 1of 461

Springer Texts in Business and

Economics

Springer Texts in Business and Economics (STBE) delivers high-quality


instructional content for undergraduates and graduates in all areas of
Business/Management Science and Economics. The series is comprised
of self-contained books with a broad and comprehensive coverage that
are suitable for class as well as for individual self-study. All texts are
authored by established experts in their fields and offer a solid
methodological background, often accompanied by problems and
exercises.
Ilia Bouchouev

Virtual Barrels
Quantitative Trading in the Oil Market
Ilia Bouchouev
Pentathlon Investments, LLC, Westfield, NJ, USA

ISSN 2192-4333 e-ISSN 2192-4341


Springer Texts in Business and Economics
ISBN 978-3-031-36150-0 e-ISBN 978-3-031-36151-7
https://doi.org/10.1007/978-3-031-36151-7

© The Editor(s) (if applicable) and The Author(s), under exclusive


license to Springer Nature Switzerland AG 2023

This work is subject to copyright. All rights are solely and exclusively
licensed by the Publisher, whether the whole or part of the material is
concerned, specifically the rights of translation, reprinting, reuse of
illustrations, recitation, broadcasting, reproduction on microfilms or in
any other physical way, and transmission or information storage and
retrieval, electronic adaptation, computer software, or by similar or
dissimilar methodology now known or hereafter developed.

The use of general descriptive names, registered names, trademarks,


service marks, etc. in this publication does not imply, even in the
absence of a specific statement, that such names are exempt from the
relevant protective laws and regulations and therefore free for general
use.

The publisher, the authors, and the editors are safe to assume that the
advice and information in this book are believed to be true and accurate
at the date of publication. Neither the publisher nor the authors or the
editors give a warranty, expressed or implied, with respect to the
material contained herein or for any errors or omissions that may have
been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.
This Springer imprint is published by the registered company Springer
Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham,
Switzerland
To our children, Vlad and Valeria
Preface
The public perception of oil trading revolves around counting barrels,
predicting outcomes of OPEC meetings, estimating gasoline and jet fuel
consumption from high-frequency traffic measures, analyzing Middle
Eastern politics, assessing the impact of regulations and trade
sanctions, or hiring meteorologists to forecast the trajectory of a
hurricane that may impact offshore production and refining centers.
These are fascinating topics to discuss at the dinner table where
everyone has an opinion on some of numerous factors that impact the
price of oil. But how relevant is this information for professional oil
trading, given so much noise and almost no reliable facts that the
market does not already know? More importantly, what are the
channels that translate such fundamental information into the price
action in the futures market where the price of oil is determined, and
how do professional oil traders filter out the noise and turn the relevant
bits into consistently profitable trading strategies?
The book attempts to answer these questions following the author’s
own experiences in managing an energy derivatives trading business
for over twenty years. The language of the book is rather quantitative,
reflecting the demands of modern markets for more data, cutting-edge
technology, and advanced analytical modeling. The book tracks the
evolution of quantitative oil trading throughout its entire history.
Options play a special role throughout the book. With somewhat
elevated barriers to entry, the business of options and volatility trading
is where quants tend to shine. But, in contrast to many other financial
markets, the dynamics of oil options can hardly be separated from the
economics of the underlying physical system, the behavior of hedgers,
and the crucial role played by storage.
The need for storage is what led to the beginnings of oil trading in
the 1860s shortly after the process of oil drilling was discovered in
Western Pennsylvania. Oil quickly proved its worth, but the demand for
barrels at the well, constrained by complicated logistics, has not always
moved in unison with supply. To buy some time and buffer fluctuations
between supply and demand, oil had to be kept somewhere. The
cheapest storage facility at the time were the dumps. Not surprisingly,
the pioneers of oil trading, who were buying excess production at
discounts and storing it, were called “dump men.”
To move oil from the dumps to consumers, which was done largely
with the help of horses, oil was poured into empty whiskey barrels. The
whiskey itself helped to negotiate the price. The price of oil was
measured in dollars per barrel. Only a handshake agreement was
needed to secure the deal, as honor and trust were held in higher
regard than written contracts. The business of oil trading was built as
an exclusive private club, largely closed to outsiders.
Little has changed in the first hundred years of the oil business,
other than horses being replaced by pipelines and ships, and whiskey
by a much wider selection of fine spirits. Membership in the private oil
club was not open to newcomers until the 1990s when oil bosses
brought in a breed of young scientists, the quants. Hiring geeks for the
old-fashioned teams of oil traders was considered somewhat
progressive and even entertaining. The mandate for quants was clear.
The oil world was becoming more complex, and in a wider web of
interlinked prices, they must search for a low-hanging fruit, the
strategy called arbitrage.
An opportunity to make small but consistent profits from low-risk
arbitrage trades had a particular allure for quants. Not only did it
provide a more financially lucrative career alternative to teaching
mathematics at a university, but it also gave quants a chance to prove
that mathematics actually works in the real world. One challenge for
quants was the relatively small size of the strategy, which was dwarfed
by riskier directional wagers made by old-timers. In the oil market, the
size of the bet and the amount of risk taken determine the trading
hierarchy, and for a while, quants were largely viewed as second-class
citizens in the oil world. To earn their proper seat at the table, more
arbitrages had to be found. Fortunately, a new era of oil trading was
about to start where quants were destined to play a much larger role.
The world was globalizing and digitalizing. To meet fast-growing oil
demand, new sources of supply had to be found. The shale revolution in
the US led to unprecedented growth in the global petroleum
infrastructure. New energy assets were built with large risks held by
financial investors. In contrast to prior decades, when oil supply growth
was predominantly driven by sovereign-controlled entities, this time
the risks were held by private capital and managed in the market. The
risks were quite complex, requiring more sophisticated hedging
instruments and more advanced quantitative skills in managing them.
At the same time, oil was also rapidly financializing and moving
away from obscure voice negotiations to transparent electronic screens.
The private club of oil trading was finally open to the public. Buying the
new digital barrel of oil has proven to be particularly handy at times of
rising inflation and geopolitical conflicts, when many other financial
assets struggle. With new computer technologies coming in handy,
quants went on a journey to modernize the oil industry. The journey
was fully endorsed by the Wall Street marketing machine, which
smelled the potential and scalability of the new business–the business
of oil derivatives.
This book tells the inside story of this business. This story has been
molded into its current form by the energy trading course taught by the
author at the Courant Institute of Mathematical Sciences at New York
University, and I thank my students for their invaluable feedback. Many
of my former colleagues contributed to the ideas presented in this
book. I thank them all sincerely for being such a critical part of the
trading business that we have successfully run together for over two
decades. And I especially thank my family for their constant
encouragement and support of my project in very unique and special
ways.
Ilia Bouchouev
New York, USA
May 2023
Contents
1 Introduction
1.​1 The Ecosystem of Real Options
1.​2 The Structure of the Book
Reference
Part I Economic Foundations, Markets, and Participants
2 Oil, Money, and Yields
2.​1 Commodities and Money
2.​2 The Oil Own Rate of Interest
2.​3 Carry and Convenience Yield
2.​4 The Roll Yield
References
3 Fundamentals, Storage, and the Model of the Squeeze
3.​1 The Invisible Hand of Storage Boundaries
3.​2 The Canonical Theory of Storage
3.​3 A Stylized Model of the Squeeze
3.​4 Cushing:​Pipeline Crossroads of the World
3.​5 Negative Oil Prices
References
4 Financialization​and the Theory of Hedging Pressure
4.​1 The Theory of Normal Backwardation
4.​2 The Hedging Equilibrium
4.​3 The Genesis of Oil Financialization​
4.​4 Inflation Hedging and Risk Parity
4.​5 Inconvenience Yield, or the Theory of Normal Contango
References
Part II Quantitative Futures Strategies
5 Systematic Risk Premia Strategies
5.​1 The Evolution of Algos
5.​2 Myths and Realities of Oil Momentum
5.​3 Carry as a Transmitter of Fundamentals to Prices
5.​4 Value and Mean-Reversion
5.​5 The Reaction Function
References
6 Quantamentals
6.​1 Trading Curve and Convexity
6.​2 Time Spreads and Inventories
6.​3 WTI-Brent Accordion
6.​4 Cointegration and Energy Stat-Arb
6.​5 Disentangling Flows and Positioning
References
7 Macro Trading
7.​1 Dynamic Systems and Feedback Loops
7.​2 Oil, Dollar, and Commodity Terms of Trade Strategies
7.​3 Oil and Energy Equities
7.​4 Oil and Inflation
7.​5 Macro Fair-Value Model
Part III Volatility Trading
8 Options and Volatilities
8.​1 Options and “Théorie de la Spéculation”
8.​2 Local Volatility and Diffusions
8.​3 Delta Hedging and Option Replication
8.​4 Realized Volatility
8.​5 Implied Volatility and its Skew
References
9 The Hidden Power of Negative Gamma
9.​1 Options and Insurance
9.​2 The Most Powerful Option Greek
9.​3 The Smile of the Volatility Risk Premium
9.​4 The Art and Science of Delta Hedging
9.​5 The Behavior of Hedgers and Regime Changes
References
10 Volatility Smile Trading
10.​1 Producer Hedging and Volatility Market-Making
10.​2 Skew Delta and Two Types of Stickiness
10.​3 When Black Smirks, Bachelier Smiles
10.​4 Fat Tails and the Quadratic Normal Model
References
Part IV Over-the-Counter Options
11 Volatility Term Structure and Exotic Options
11.​1 Dark Pools and the Hacienda Hedge
11.​2 The Term Structure of Implied and Local Volatilities
11.​3 Volatility Discount from Price Averaging
11.​4 Early Expiry Options and Swaptions
11.​5 Multi-Factor Models and Other Exotics
References
12 Volatility Arbitrage and Model Calibration
12.​1 The Inverse Problem of Option Pricing
12.​2 Bootstrapping in Time
12.​3 Market-Implied Probability Distribution
12.​4 Local Volatility Smile
References
13 Spread Options and Virtual Storage
13.​1 The Synthetic Storage Strategy
13.​2 Triangular Correlation Arbitrage
13.​3 The Dichotomy of Spread Option Pricing
13.​4 Dealing with Unobservables
References
14 Epilogue
14.​1 The Roadmap for Energy Transition and Virtual
Commodities
Appendix A:​Diffusions and Probabilities
Appendix B:​Option Pricing under Normal and Lognormal
Distributions
Appendix C:​The Perturbation Method and the Quadratic Normal
Model
Appendix D:​Option Pricing with Time-Dependent Volatility
Appendix E:​Average Price Options
Appendix F:​The Inverse Diffusion Problem
Glossary
References
Index
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
I. Bouchouev, Virtual Barrels, Springer Texts in Business and Economics
https://doi.org/10.1007/978-3-031-36151-7_1

1. Introduction
Ilia Bouchouev1
(1) Pentathlon Investments, LLC, Westfield, NJ, USA

1.1 The Ecosystem of Real Options


Global oil trading is an example of a complex dynamic system, where
everything depends on everything else, and system components
communicate with each other via multiple feedback loops. Other
examples of such systems include the human body, the world economy,
climate change, or relationships within the family. Such systems are
capable of producing their own behavioral patterns which are often
governed by quantifiable rules. We will describe some of these rules
and refer to the resulting behavior of the system participants as mental
models.
In the core of the oil trading system lies the concept of optionality.
The business of producing oil is a call option on the price of oil. If the
price covers all costs of exploration and production, then the business
moves ahead. If it does not, then oil remains in the ground. The price of
oil is volatile, which makes oil production a valuable option to own, but
an expensive one to acquire, as large capital investment is needed at the
outset of the project. Investing in such a long-term option is a risky
business. By the time oil is found and the necessary infrastructure is
built, the price of oil could fall and make the entire project unprofitable.
It is possible, of course, to roll the dice and hope for the best in the style
of early wildcatters. Alternatively, one can do something about it and
secure a profitable forward price for the time when production is
expected to commence. This process of managing the forward price risk
is called hedging. The venue for such transactions is the derivatives
market, made up of futures, swaps, options, and highly customized
structured products.
When the term derivative is applied to oil futures, it becomes a
misnomer. A conventional textbook definition of a commodity
derivative suggests that the futures contract derives its value from the
spot price of a physical commodity. The oil market, however, is
somewhat unconventional. In contrast to many other raw commodities,
oil does not have a universal spot price with an instantaneous delivery.
There is no one large oil bazaar where all buyers and sellers show up
with barrels and gallon jars, negotiate the price with an instant delivery,
and take oil home. Oil is too bulky and a special infrastructure is
required for its transportation and storage. A buyer and a seller can
only agree on the price of oil to be delivered at a given time to a
specified location, typically to a storage tank connected to pipelines. In
other words, oil in commercial quantities can only trade forward. Even
though we use the term spot price throughout the book for brevity, it is
meant to represent the forward price with short delivery time.
A typical agreement for the delivery of physical barrels is highly
customized for the specific grade of oil, its delivery time, and location.
To compare simultaneous prices for many such agreements, one needs
to peg them to a common anchor. This anchor is usually the futures
contract with the nearest expiration, which is also referred to as the
prompt futures. The prices for physical transactions are then
negotiated as a basis to futures. Since the magnitude of the physical
basis typically represents only a relatively small adjustment to the
futures price, the primary price determination occurs in the futures
market. Unlike many other commodities, the oil futures contract can
hardly be called a derivative of the spot price. In fact, it would be much
more appropriate to say that the spot oil price is a financial derivative
of oil futures.
The need to transport oil to the desired location highlights an
important role that oil shippers play in the market. The most effective
way of transporting oil on the ground is via pipelines. Oil can also be
moved by rail cars and even by trucks but at a significantly higher cost.
To transport oil between continents, one must resort to special oil
tankers. The pipeline, the tanker, and other oil-transporting assets are
also examples of real options. They are options on the difference
between oil prices in two geographical locations, also known as spread
options. The asset owner will only commit to shipping oil if the price
differential between two locations covers the cost of transportation.
Spread options play an important role in the functioning of the entire
petroleum complex. They arise at virtually every step within the
petroleum value chain.
Once oil is shipped to the desired destination, it must be placed in
storage, since not all of it can be consumed instantaneously. The storage
asset shifts the allocation of the valuable resource in time, from times of
plenty to times of relative scarcity. If the futures price for later delivery
is higher, then the rational storage operator will postpone the sale of oil
and instead store it, provided that the forward price premium covers
the cost of storage. If the forward price premium over the spot price is
not sufficient to pay for storage costs, then the operator is better off
selling oil in the market than keeping it in storage. Storage is also a real
option. While the pipeline is an option on location, the storage asset is
an option on time.
Storage plays a prominent role in global oil trading, and it will
feature accordingly throughout the book. In fact, the largest single
participant in the futures market is not the producer, the consumer, or
even the speculator, but the hedger of oil inventories. We will see that
the behavior of the storage trader in the futures market translates the
fundamental information about supply and demand in the physical
market to futures prices, which in turn determines the price of the
physical barrel. The presence of storage makes physical and derivatives
oil markets operate as a mutually stimulating loop.
The final critical piece in the oil value chain is a refinery. Crude oil
has very limited use in its raw form. It must first be processed into a
usable product. A refinery that processes crude oil is also an option on
the spread. Here, the spread is between the basket of refined products
and the basket of crude oil inputs. The power of spread options in the
oil business is best exemplified by John D. Rockefeller, who built his
empire by monetizing real options to refine, transport, and store oil,
while leaving the more glamorous but much riskier option on the price
of oil to wildcatters and speculators.
The real optionality embedded in physical assets is the core of the
energy market. It transforms trading into a complex dynamic system
driven by multiple feedback loops. The system rests on its four primary
pillars of production, transportation, storage, and refining. Each of
these pillars represents a valuable but expensive option. The high cost
and associated risks imply that one cannot pre-build too many of such
real options in advance and keep them hanging around just in case. By
and large, the energy infrastructure has just enough production,
refining, pipelines, tankers, and storage tanks to keep it up and running,
but it has very little cushion and no margin for error. The oil ecosystem
operates near the edge in a highly optimized fashion with a just-in-time
and just-enough mentality. Such systems are characterized by an
inherent instability with periodic bursts of extreme volatility.
The economic function of financial derivatives is to provide asset
owners with tools to manage such volatility. Given the enormous cost of
the energy infrastructure, the demand for risk-transferring services
turned out to be so large that the market for oil derivatives stood out
not only among all commodities, but it also became one of the deepest
derivatives markets across all asset classes. In addition, its liquidity,
combined with the high volatility and relatively low correlation to other
financial assets, turned oil into a favorite playground of speculators.
With so many diverse factors impacting the price of oil, one can always
find something in the oil market to bet on, which can only stimulate the
marketing machine to make the market even larger.
At present, the world consumes approximately one hundred million
barrels of oil per day. While this may sound like a large number, it pales
in comparison to over two billion financial barrels that trade every day
via two benchmark oil futures contracts, West Texas Intermediate
(WTI) and Brent. Such an unprecedented growth of financially-traded
barrels is illustrated in Fig. 1.1.1 If one adds futures on other grades of
crude oil, futures on refined products, options, and over-the-counter
(OTC) derivatives, then the daily trading volume of petroleum
derivatives is at least fifty times larger than daily global oil
consumption. While such a comparison should only be interpreted with
great caution, it nevertheless shows the growing power of financial
markets.
Fig. 1.1 Daily trading volume of two benchmark oil futures contracts, WTI and Brent, is at
least twenty times larger than daily global oil consumption. Source: CME, ICE, EIA
The derivatives market has brought new entrants to the oil
marketplace. Besides producers, consumers, storage hedgers, and other
traditional market participants, the price of oil became highly sensitive
to the behavior and motivation of relative newcomers, including
inflation hedgers, systematic quantitative funds, and especially option
traders. The goal of this book is to demystify their behaviors that
collectively drive the elusive price of oil.

1.2 The Structure of the Book


The book consists of four main parts with each split into three chapters.
While all parts of the book are interconnected, each one can be read as
stand-alone based on individual interests and the level of the reader’s
mathematical background. The first part introduces market
participants and lays down the economic foundations of storage and
hedging pressure, upon which many trading mental models will
subsequently be built. These mental models and futures trading
strategies are presented in Part 2, which is the least technical and can
be easily read by anyone with a broad interest in financial markets. Part
3 targets option traders familiar with volatility trading and some basic
stochastic modeling. Part 4 is more suitable for professional volatility
arbitrageurs and anyone with a stronger interest in exotic oil options
and mathematical finance. To keep the main text accessible to a wider
audience, technical details are presented in appendices. Throughout the
book and unless noted otherwise, we use energy data from the U.S.
Energy Information Administration (EIA) and the Chicago Mercantile
Exchange (CME), and macroeconomic data from the Federal Reserve
Economic Data (FRED).
We begin Chap. 2 by looking at a hypothetical economy where oil
and the US dollar (USD) function as two alternative standards of money.
We relate the interest rate in the oil-denominated economy to the
interest rate on fiat money, following the first rigorously defined no-
arbitrage carry trade, formulated by Irving Fisher. This formula is now
better known as the Fisher inflation law, which establishes the close
linkage between oil and inflation. We revive a largely forgotten
Keynesian concept of commodity own rate of interest and illustrate it
with an example of an oil loan from the US Strategic Petroleum Reserve
(SPR). We then relate the oil own rate of interest to the convenience
yield in the physical market and to the roll yield that arises in trading
futures.
In the following two chapters we look at two alternative approaches
to modeling the behavior of futures prices. These approaches, known as
the theory of storage and the theory of hedging pressure, loosely
correspond to competing trading styles based, respectively, on
fundamentals and flows.
In Chap. 3, we use the conventional storage theory to illustrate the
dynamic feedback loop between prices and inventories and highlight
the challenge of its practical applications to the oil market. We then
borrow some concepts from the physics of extreme events and develop
a more practical alternative approach to the storage problem. We call it
a stylized model of the squeeze. For an example of such a squeeze, we
delve into the infamous episode of negative oil prices.
In Chap. 4, the focus shifts to flow imbalances. Particular attention is
paid to causes and consequences of the influx of financial investors to
the oil market, the phenomenon dubbed financialization. We track the
market transition from the early days of normal backwardation to the
subsequent regime of normal contango and combine them into a more
general economic framework that describes the hedging pressure
equilibrium.
In the second part of the book, we present mental models for
futures trading strategies deployed by professional speculators,
including quantitative hedge funds, sophisticated physical speculators,
and global macro traders. We devote one chapter to each of these
groups of traders and their marquee strategies. While appearing to be
very different in nature, many concepts supporting these strategies can
be seen as practical implications of economic theories presented in the
first part of the book. Our goal is to focus on modeling frameworks and
conditions under which various strategies tend to work, and not on
nuances of their technical implementation.
Chapter 5 is devoted to systematic quantitative funds, known as
commodity trading advisors (CTAs) or simply as algos. These traders
tend to look at markets through the lenses of risk premia, such as
momentum, carry, and value. We discuss how the source of these risk
premia in the oil market is ultimately linked to the theory of storage.
Important concepts of signal blending and the reaction function are
introduced.
In Chap. 6, we explore the considerably more challenging task of
using fundamental data for generating trading signals, and the class of
semi-systematic strategies, colloquially referred to as quantamentals.
We discuss the convergence strategy of WTI-Brent accordion and
extend the concept to construct a broader statistical arbitrage portfolio
of energy pairs. We use fractionation analysis to incorporate
fundamental data and provide some guidance and ideas for modeling
flows and positioning. Such quantamental strategies are more
effectively managed with a discretionary overlay. In the oil market, the
trader equipped with the machine is more powerful than either an
oilman or a machine alone.
Chapter 7 looks at oil as a sub-component of a broader macro
trading system. We illustrate oil linkages to currency, equity, and fixed
income markets with three cross-asset relative value strategies. We
conclude by constructing a simple and somewhat naïve fair-value
model for the price of oil. While this model is built on questionable
theoretical grounds, it is a good example of when something that
should not work in theory turns out to be helpful in practice.
The third part of the book deals with vanilla oil options and classic
volatility trading. The market for oil options is one of the most
developed options markets in the world. This should not come as a
surprise given that the entire oil market is driven by real options held
by owners of physical assets. However, from the quantitative
perspective the dynamics of oil volatility, which determines the price of
an option, remains largely undocumented. This book attempts to fill the
gap.
When it comes to trading, there is no free money. While trading
futures is easier to understand, popular linear strategies quickly
become crowded, as barriers to entry are relatively low. In contrast, the
barriers to entering the options market are set higher, and lucrative
trading opportunities there last longer. The highest barrier is the
quantitative skill of the oil trader, so more technical knowledge is
essential for success in volatility trading. The goal of this book is to help
aspiring option traders to develop a minimum competency level that
can then be applied and further improved in a professional trading
organization.
In Chap. 8, we summarize the main building blocks that make up the
business of volatility trading. We start by giving well-deserved credit to
Louis Bachelier, the father of modern quantitative finance, whose
Bachelier pricing formula is still being used by oil option traders. The
classical Black-Scholes-Merton (BSM) framework of option replication
by delta hedging is then derived in a more general setting of diffusion
processes. We highlight the importance of distinguishing between three
commonly used types of volatility: local volatility, realized volatility, and
implied volatility.
Chapter 9 looks at writing oil options from the perspective of the
insurance product, which is the essence of gamma trading and the
resulting volatility risk premium (VRP). We dissect the historical
performance of various delta-hedging strategies from multiple angles,
introduce the concept of the VRP smile, and identify regime changes
caused by changing behavior of large market participants.
In Chap. 10 we study the problem of the volatility smile and the
popular market-making strategy of vega trading. We demonstrate the
shortcoming of tracking oil volatility smile using popular heuristics of
sticky strike and sticky moneyness and replace these naïve methods with
a more flexible framework based on general diffusion processes. We
apply the technique of perturbation methods to develop a novel
quadratic normal (QN) model. This model corrects the Bachelier
formula for skewness and kurtosis. The three parameters of the model
are linked to the market prices of three primary option benchmarks: at-
the-money straddle, costless collar, and out-of-the-money strangle. The
QN model corrects the conventional pricing formula just enough to
capture important features of the oil market without introducing
excessive complexity.
In the final part of the main text, we describe core components of
managing large and often secretive over-the-counter (OTC) deals which
remain hidden from public view. We hope to unveil some of the mystery
behind these deals for the general public, while helping professional
option dealers with somewhat tricky challenges that arise in pricing
and calibration of models used for OTC energy derivatives. This part is
somewhat more advanced quantitatively, but once again, most technical
details are placed in appendices.
In Chap. 11, we start with the example of arguably the most
significant oil derivative trade, the large-scale annual put buying
program by the Government of Mexico. The complexity of OTC deals
highlights the importance of handling the volatility term structure and
the effect of volatility dampening by price averaging. A simplified and
more practical model for pricing and hedging average price options
(APO), which are very popular among end-users, is presented.
Swaptions and several other exotic options are also discussed.
Chapter 12 covers the more challenging problem of model
calibration. We present the bootstrapping method for calibrating
volatility time-dependency, including the volatility matrix for non-
homogeneous term structure of volatility. To calibrate the model across
strikes, we first back out market-implied probability distribution from
option prices. A more difficult problem is to reconstruct the market-
implied diffusion process, known as the inverse problem of option
pricing. Some readers may find it interesting that this important
problem is only partially solved, and in its general case, it presents a
rare example of an open mathematical problem. This problem arises
naturally in calibration of oil options.
Chapter 13 presents a relatively simple but quantitatively elegant
strategy based on the idea of virtual storage. In this strategy, the
physical storage asset is replicated with a financial derivative, a
calendar spread option (CSO), and the structural pricing dislocation
between the physical and financial market is then exploited. Other
spread option and correlation strategies are based on the concept of a
triangular arbitrage, which links prices of vanilla and spread options.
We highlight some challenges with an application of correlation-based
models and insist on connecting spread options models to
fundamentals of supply and demand.
The book concludes with a short epilogue related to the energy
transition. Its focus, however, is on the transition of virtual energy or
the market for energy derivatives and its impact on the broader market
for virtual commodities.

Reference
Bouchouev, I. (2023). Testimony to the House Subcommittee on Economic Growth, Energy Policy,
and Regulatory Affairs of the Committee on Oversight and Accountability, March 8. Reprinted in
Commodity Insights Digest (2023), Vol. 1.

Footnotes
1 From the author’s testimony to the U.S. House Subcommittee on Economic Growth, Energy
Policy, and Regulatory Affairs of the Committee on Oversight and Accountability. See Bouchouev
(2023).
Part I
Economic Foundations, Markets, and
Participants
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
I. Bouchouev, Virtual Barrels, Springer Texts in Business and Economics
https://doi.org/10.1007/978-3-031-36151-7_2

2. Oil, Money, and Yields


Ilia Bouchouev1
(1) Pentathlon Investments, LLC, Westfield, NJ, USA

If oil is so important to the economy, why cannot it be used as an


alternative form of money? Some oil transactions are settled directly
in barrels, and the US Government even lends oil with the interest
also paid in barrels.
The oil own rate of interest represents the return on lending oil
barrels. It can be calculated from the term structure of traded
futures. The own rate is analogous to the real rate in the famous
Fisher inflation law.
In fundamental trading, the oil own rate is better known as
convenience yield, net of storage costs. It represents the trade-off
between the intrinsic value of the commodity for its immediate
consumption and the cost of storing it.
Financial traders buy futures to synthesize the ownership of the
physical barrel but the investment in futures may lead to very
different results. The difference comes from the roll yield, which is a
financial equivalent of the oil own rate of interest.

2.1 Commodities and Money


The money-rate of interest is nothing more that the percentage
excess of a sum of money contracted for forward delivery. For
every kind of capital-asset there must be an analogue of the rate
of interest on money. Thus for every durable commodity we
have a rate of interest in terms of itself, ̶ a wheat-rate of interest,
a copper-rate of interest, a house-rate of interest, even a steel-
plant-rate of interest.
J. M. Keynes (1936)
Today, we live in the world of fiat money. For most of our ancestors,
however, money was represented by commodities. Gold, silver, copper,
salt, rice, maize, barley, cocoa beans, tobacco, and many other raw
materials have a long history of being used as money. From the very
beginnings of human civilization, people were choosing something
valuable to them to function as money. Unlike fiat money, which derives
its value solely from faith in the issuing government, commodities
possess value within themselves. The intrinsic value of a commodity is
its worth for consumption and for various productive processes. Since
oil has been the most valuable commodity for nearly a century, why not
use it as an alternative standard of money?
We can easily agree to settle all transactions directly in oil, borrow
and lend oil barrels, and negotiate interest on such loans, also to be
paid in oil. An oil rate of interest will depend on the value of oil today
relative to its value at the time when a debt must be settled. If
consumers are eager to receive oil barrels quickly, they may agree, for
example, to exchange 100 barrels today for 110 barrels to be returned
later. This would generate ten percent investment return to the lender
of oil barrels. The return comes from the intrinsic value of the
commodity, its utility, and the convenience of having it handy if it
becomes needed.
However, if immediate demand for oil is tepid, then the owner of oil
must find a place to keep it somewhere, while waiting for buyers to
come in. The process may involve renting a dedicated storage facility,
connecting it to a pipeline, buying insurance, and completing other
burdensome and costly logistical operations. Since all payments are
agreed to be made in barrels, the lender may need to give up some
barrels to remunerate a provider of storage services. For example, the
oil lender may agree with a storage manager to exchange 100 barrels
today for 95 barrels to be taken back later. In this case, the investment
return on lending oil barrels would be negative five percent. Any
decision making in such an oil-based economy represents the trade-off
between the intrinsic value of oil for its immediate consumption and
the cost of storing it.
Now consider a dual-currency economy, where besides oil, there
exists an alternative money standard, such as gold, the US dollar (USD),
or bitcoin. Let us assume that one can freely substitute between two
types of money at the prevailing exchange rate determined by the
market. We also assume that the market is efficient to preclude any
possibility of arbitrage, which means that it is deemed to be impossible
to extract riskless profits from borrowing in one money standard and
lending in another. In the absence of such arbitrage, interest rates on
two alternative types of money must be related. This relationship was
developed by Irving Fisher at the end of the nineteenth century.1 Today,
it is better known as the Fisher inflation law, and the transaction that
leads to it as the carry trade.
Fisher’s description of an equivalence between loans in two
commodity money standards is arguably the earliest rigorously
described example of the quantitative arbitrage trade in financial
markets. His work was motivated by the collapse of the bimetallic
monetary standard when gold and silver coins could both be used as
acceptable forms of money. To illustrate that the rates of interest on two
alternative commodity money standards must satisfy an equilibrium
no-arbitrage equation, Fisher used bushels of corn and ounces of gold.
His argument can be easily extended to a hypothetical bi-monetary
world where all transactions are settled either in barrels of oil or in
USD.
If one borrows D(t) dollars at time t and returns D(T) dollars to the
lender at time T, then the USD rate of interest, r, is conventionally
defined as

(2.1)

Likewise, if one chooses instead to take a loan in oil, borrow B(t)


barrels at time t, and return B(T) barrels at time T, then the oil rate of
interest, b, is similarly defined as

(2.2)

Let us assume that we also know the price of the second money
standard in terms of the first one. In other words, we know spot oil
prices S(t) and S(T) at times t and T, both expressed in USD per barrel
($/bbl)2:

From here, we express barrels as the ratio of dollars to oil prices, and
substitute into (2.2), which allows us to relate the oil rate of interest to
the dollar rate of interest, using the definition (2.1), as follows:

We then rearrange the terms and obtain that

(2.3)

where

represents the rate of change in prices measured by the alternative


money standard, which is oil, in terms of the original money standard,
which is USD. This formulation is Fisher’s original one-period version of
the inflation law. He defined b as the virtual rate of interest in
commodities. As we will see throughout the book, commodity prices
and inflation tend to go side by side.
With the development of commodity futures markets, the concept of
commodity rate of interest was revived by an Italian economist, Piero
Sraffa. Since the price of the futures contract F(t, T) for the delivery at
time T is observable from the market, one can use it to replace an
unknown future spot price by letting

In other words, one no longer needs to guess what commodity rate of


interest is. It can be calculated directly from the futures market. Sraffa
used such market-implied rates for different commodities to challenge
the proposal of Friedrich Hayek to fix the money rate of interest at
some unique natural commodity rate that can be found in a non-
monetary economy. Since the price of each commodity is driven by its
own supply and demand that lead to a different relationship between
spot and futures prices, Sraffa argued that “there might be at any one
moment as many ‘natural’ rates of interests as there are commodities …
and there would be no single equilibrium rate”3.
The idea of commodities having their own intrinsic rates was
further developed by John Maynard Keynes, who labeled the concept
the commodity own rate of interest. The own rate became the
centerpiece in one of Keynes’ most influential economic theories, on
how competition between commodity own rates and the interest rate
on money could lead to economic slumps. His application of this term,
however, became highly controversial, causing a considerable amount
of confusion among economists.4 Perhaps for these reasons, the
concept of a commodity own rate of interest has largely disappeared
from contemporary economic literature.

2.2 The Oil Own Rate of Interest


After being forgotten for decades, the concepts of commodity loans and
own rates of interest found important practical applications in the
modern oil market. Settling transactions directly in oil barrels turned
out to be very convenient for many market participants. Using oil in lieu
of the currency allows some resource rich but cash poor sovereign
producers to attract financial investors to develop oil properties and
pay them directly in oil. It also lets refineries secure profit margins via
netback agreements by exchanging their product output for the crude
oil input. Furthermore, storage owners, including governments, can
lend barrels to the market in exchange for more barrels to be returned
later. In other words, oil is already used as a money substitute by the
industry. Before we get to some specific case studies, let us set the stage
with some simple examples.
We first consider the scenario, when the demand for oil today at
time t is relatively high, and the spot price S(t) = $60 exceeds today’s
futures price for delivery at time T > t, which is F(t, T) = $50. In the
jargon of commodity markets, such a decreasing forward curve
corresponds to the market being in the state of backwardation. Let us
assume that the dollar rate of interest is r = 2.5%. Then one can use $60
today to either buy one barrel of oil, or, alternatively, buy
$61.50 = $60 ∗ (1 + 0.025) dollars of forward delivery. These $61.50 of
forward dollars can then be converted into oil barrels at
the forward oil price. Looking at this transaction from the perspective
of an oil lender means that one barrel of oil today is exchanged for 1.23
barrels to be returned later. Therefore, the oil own rate of interest in
this example is 23%.
Now consider the second scenario, where the forward price F(t,
T) = $70 exceeds the spot price S(t) = $60. The forward curve is now
increasing, which defines the market being in the state of contango. A
lower current price in comparison to the future price indicates that
current oil supplies are abundant, and the demand for oil is relatively
low. In this case, the same $61.50 of forward dollars can only buy
forward oil barrels. Therefore, the oil own rate of interest
is negative 12%. While lending barrels at a negative return may sound
counterintuitive, it may still be economical if the cost of storing oil is
significant.
While the concept of own rate of interest was originally introduced
in discrete-time setting using simple returns, the contemporary
analysis of financial markets usually operates in continuous-time
setting with continuously compounded returns.5 For consistency,
simple returns are usually replaced with logarithmic or log-returns,
which makes algebraic calculations simpler and more intuitive. Unlike
simple returns, which are not additive, the log-returns can be added
both in time and across multiple assets. Since in our book the impact of
interest rate plays only a minor role in oil futures strategies, for the
most part, we use log-returns in all analyses.
If we replace simple returns with log-returns, defined as

then repeating the previous steps and replacing unobservable future


spot price S(T) with the market futures price F(t, T), we obtain that
(2.4)

This is the continuous version of the Fisher’s inflation law. Here, b is


better known as the real rate, which is equal to the difference between
the nominal rate of interest r and the rate of inflation i.
The Eq. (2.4) is written for a single time period. If r and b are viewed
as annualized log-returns, then they must be multiplied by time to
maturity of the futures contract, so that

Therefore, the futures price can be expressed as the spot price that
grows exponentially at the rate equal to the difference between the
nominal rate of interest and the oil own rate of interest
(2.5)
Alternatively, one can back out the oil own rate of interest from the
futures curve

(2.6)

where the spot price is understood as the futures with the shortest
maturity.
Figure 2.1 illustrates how the own rate relates to backwardation
and contango of the futures curve with different maturities.
Fig. 2.1 If the oil own rate of interest exceeds (falls below) USD rate of interest then the
futures curve is in backwardation (contango)
If b > r and the intrinsic worth of the commodity is relatively high,
the futures curve is in backwardation. In contrast, if b < r and the
commodity own rate is relatively low, the futures curve is in contango.
If b = 0, then the futures curve is in a slight contango, monotonically
increasing at the rate r.
Having defined the basic terminology of the oil economy, we can
now illustrate it with an important case study. We consider an example
of oil loans from the US Strategic Petroleum Reserves (SPR), the world’s
largest supply of emergency oil stocks, held by the US Government. The
SPR was established in 1975 by the Energy Policy and Conservation Act
(EPCA) following the Oil Embargo of 1973–1974 by the Organization of
Arab Petroleum Exporting Countries (OAPEC). The mission of the SPR
is to store petroleum to diminish the impact of disruptions on
petroleum supplies and to carry out US obligations under the
International Energy Program. Oil stocks can be released from the SPR
either via competitive sales, or as oil exchanges. Oil exchanges are loans
where barrels are released to traders in exchange for a larger quantity
of barrels to be returned at some time in the future.6
Consider an example of the large SPR oil loan that was authorized in
November 2021. At that time, oil demand was quickly recovering from
the Covid-19 induced slowdown, while oil supplies from the OPEC+
group of large producers were released to the market at a slower pace.7
To reduce fundamental supply and demand imbalance, the US
Government offered oil from the SPR to traders in exchange for a larger
quantity of oil to be returned at a later time. The rates of return on
these oil loans were contractually set for five designated loan
maturities. The longest maturity loan stipulated 9.1% more barrels in
excess of the originally borrowed quantity to be returned to the SPR in
approximately thirty months. This allows the US Government not only
to replenish the reserve, but also to grow it at no cost.
The question is then why would anyone agree to borrow oil at such
a high rate, especially given that the interest rate on fiat money at that
time was nearly zero? The answer lies in the steeply backwardated
shape of the oil futures curve. To compare loan rates offered by the US
Government with the market-implied rates, one can apply formula (2.6)
to calculate oil own rates prevailing at the same time from WTI futures,
which is the primary US oil futures contract. It turns out that the thirty-
month oil own rate implied by the WTI futures curve at that time was
even higher, closer to 20%. Figure 2.2 illustrates.
Fig. 2.2 The rates on SPR oil loans (grey bars) offered by the US government in November
2021 are compared to WTI market implied rates (orange) computed using (2.6)
If oil markets were frictionless and conducted only using the WTI
grade of oil with all transactions executed at no cost, then the trader
could have theoretically made riskless profits by borrowing barrels
from the US Government at 9%, selling them in the spot market, buying
thirty-month futures at approximately 20% discount and taking the
delivery of physical barrels at that time. Obviously, the real market has
plenty of frictions, including high logistical costs associated with
physical oil delivery and large price uncertainty in the basis between
WTI and specific oil grades that must be returned to the SPR.
Nevertheless, for a physical oil trader, participation in such loan
transactions could be quite lucrative. This is our first example of the
trading strategy where the needs of one market participant, the US
Government, creates an arbitrage opportunity for a professional oil
trader.
Despite the fact that oil has many attributes of an alternative
standard of money, including even having its own rate of interest, we
still do not use oil as money. One reason is the complexity of storing oil,
but perhaps a more important one is its extreme volatility. To illustrate,
Fig. 2.3 shows the history of a one-year own rate of interest implied by
WTI futures.
Fig. 2.3 The history of a one-year oil own rate of interest as implied by WTI futures
The oil own rate of interest fluctuates wildly, as it is driven by the
constantly changing slope of the futures curve. For example, at the
onset of the Covid-19 pandemic when oil demand fell rapidly, storage
costs increased substantially, culminating in an unprecedented episode
of negative prices, which we will discuss in detail in the next chapter.
For a negative price, one cannot compute an own rate, but even with
this abnormal data point removed, the own rate during that time
reached its nadir of negative 100%. However, within two years it swung
to positive 30%, as the world recovered from the pandemic and the
convenience yield of owning physical barrels skyrocketed. With such a
high volatility in a lending rate, it would be extremely challenging to
ensure any stability in the oil-denominated economy.
While the oil own rate of interest may sound like an obscure
theoretical concept, it underpins many commercial transactions and
plays the overall crucial role in the oil market. We will see throughout
the book how large its impact is on different trading strategies. For
many market participants, however, this concept is better known under
different names. In fundamental trading, it is more often associated
with the concept of convenience yield.
2.3 Carry and Convenience Yield
The own rate of interest reflects the combined effect of the intrinsic
value of the commodity and the cost of storing it. It is calculated by
comparing the commodity price today to its price in the future. The
current price and the future price are connected by the carry trade. The
carry arbitrage trade has been well known to commodity merchants
from the early days of trading. It simply refers to merchants buying and
storing a physical commodity, while hoping to sell it later at a higher
price. If the forward price is also locked in, then the carry trade
effectively represents arbitrage in time.
Oil started to trade in the 1860s in Western Pennsylvania shortly
after the first oil well has been drilled there. The first oil traders were
carry arbitrageurs who developed some primitive forms of oil storage.
They were buying excess oil from local producers at discounts and
storing it in dumps. These pioneers of oil trading were called “dump
men”:

The first oil markets were, in my opinion, made by the dump


men on Oil Creek. Now, the dump men were the first speculators
in oil. A dump is a tank of any capacity from 10 barrels to 600.
The dump man developed in a refiner later on … The dump man
visited the small producer in all localities, bought his oil at so
much by contract, and if there was one or two feet in the tank
bottom he immediately bought in lump sum or so much per
barrel. He also bought the good oil, the merchantable oil, from
small producers, who could not hold their oil long enough to get
a full shipment of 50 or 100 barrels, whose wells had declined to
such a point that they could not hold for a month at a time, or
were obliged to sell from day to day, or week to week. The dump
man was their market.8

The storage carry strategy remains the backbone of modern oil


trading. It gives storage owners an option on time, an option to carry
barrels and sell them at a time when they become more valuable.
Professional storage traders avoid taking any directional exposure to
the price of oil, which is usually hedged by selling futures. The storage
business model is entirely driven by the economics of the carry trade. If
the spot price is discounted relative to futures, and the discount is
sufficiently large to cover the cost of storage, then, by and large, riskless
profits can be assured by buying a physical barrel for storage at the
spot price while simultaneously selling a higher priced futures contract.
Obviously, in competitive markets, the opportunity to make free
money is unlikely to last long. Carry traders will continue to bid up the
spot price while selling and pressuring down the futures price until the
arbitrage goes away. Therefore, to prevent the existence of a free lunch,
the futures price F(t, T) should not exceed the spot price S(t) by more
than the marginal cost of storage U and the cost of financing R:
(2.7)
The existence of this arbitrage boundary has a wide range of
consequences. Importantly, it implies that the futures prices cannot be
determined solely by market expectations of what the spot oil price will
likely be in the future. The arbitrage boundary holds regardless of such
expectations. Therefore, the role of expectations in the formation of
futures prices is rather limited. This partially explains why futures
prices are poor predictors of future spot prices, especially when the
market is in contango, and prices are dictated by storage economics.
Storage costs represent an important component in the economics
of the carry trade. Oil is bulky and relatively expensive. Keeping large
quantities of a pricey commodity in storage ties up a lot of capital. One
cannot logistically leave an oil storage facility empty, and some amount
must always be kept there irrespective of the storage economics. In
addition, the holder must spend some money on buying insurance and
maintaining proper security of the storage facility to protect the high
dollar value warehoused there. For these reasons, commercial market
participants rarely build oil storage in excess just to have it around in
case it ever becomes needed. Instead, they highly optimize the
operation of existing facilities. Like many other components of the
petroleum ecosystem, the storage business follows a just-in-time and
just-enough type of business model. In the end, if consumers run out of
oil, it will likely be the governments, and not commercial storage
operators, who will be on the hook to prevent an economic catastrophe.
The cheapest way to store oil is in underground salt caverns, which
is what the US Government does for the SPR. While salt cavern storage
works well as a long-term buffer for national security, it is less suitable
for commercial operations, where one needs to take oil in and out of
storage frequently in response to changing market prices. While state-
owned oil reserves function predominantly as the storage of last resort,
the role of managing short-term fluctuations in supply and demand is
left to private enterprises. They tend to store oil mostly in specialized
aboveground storage tanks that are connected to major production and
refining centers via pipelines.
The total storage cost typically includes the cost of transportation. If
storage tanks in nearby and easily accessible storage facilities fill up,
then oil must be transported to more remote storage locations,
preferably by pipelines. Additional shipping expenses increase the total
cost of transportation and storage, which raises the level of contango
needed for competitive storage owners to stay in business. As more and
more storage facilities accessible by pipelines become saturated, more
expensive transportation tiers, such as rail cars and trucks, are utilized
to ship oil. This further steepens the level of contango that is required
to offset higher costs. If all dedicated storage facilities reach their
maximum capacity, then the only other practical solution for storing oil
in commercial quantities is so-called floating storage, which is a
reference to oil being temporarily stored on large idle ships. Floating
storage is significantly more expensive, as it must incentivize ship
owners to divert tankers away from their primary business of oil
transportation.
The cost of storage and the carry trade set the floor on how far the
spot price can fall below futures. There is technically no ceiling on how
far the spot price can rise relative to futures. In financial markets, such
a ceiling is assured by the reverse carry trade. For example, if the price
of a common stock today is too high relative to its forward price, then
one can always borrow the stock to sell it short and receive it back at
the cheaper price by buying futures and taking delivery of the stock at
maturity. It is not even necessary to have the stock in possession, as the
existence of such riskless profits will force rational investors who own
the stock to either conduct such transactions themselves or lend the
stock to arbitrageurs. Since the stock is always owned by someone, the
incentive to make riskless profits will make reverse cash-and-carry
trading possible in the financial markets, and inequality (2.7) turns into
an identity.
In commodities, however, the reverse cash-and-carry strategy
cannot be guaranteed. The challenge comes from the consumption side
of commodities. If the entire inventory of a commodity has already
been used up and resupplies are running behind, then the commodity
may not be available for borrowing to enforce the reverse carry
arbitrage. Even if it is available, the owner may not be willing to part
with the valuable product if its consumption value is perceived to be
higher than the expected profit from lending it.
To quantify this additional consumption value of oil, we need to
bring in another balancing factor. For now, we simply define it as a
fudge factor by introducing an extra term Y that turns inequality (2.7)
into an equation
(2.8)
This balancing term was coined as convenience yield by Nicolas Kaldor
in 1939, which he interpreted as a negative component of the storage
cost.9 The relationship between futures and spot prices is then
determined by the relative magnitude of the total carrying costs and the
commodity convenience yield.
While the convenience yield in Eq. (2.8) is understood in dollar
terms, it is more common to write this equation in a continuous-time
setting, as follows10
(2.9)
Here, r is the short-term interest rate, u is the marginal rate of storage
costs, and y is the marginal convenience yield per unit of price. The
same no-arbitrage logic applies. At time t, one can secure the price of oil
to be delivered at time T either by buying the futures contract at the
price F(t, T), or by borrowing money to buy spot barrels at S(t) and
paying interest rate and storage costs while retaining convenience
benefits. The futures price must then be equal to the spot price, accrued
at the continuously compounded annualized rate of interest and
storage, net of the convenience yield.
The Eq. (2.9) resembles (2.5), which relates futures and spot prices
via nominal and oil own rate of interests. By comparing these two
equations, one can see that the oil own rate of interest is simply the
convenience yield, net of the marginal rate of storage costs
(2.10)
In other words, the own rate of a commodity is simply the difference
between its benefits for consumption and the cost to store it.
To grasp the meaning of the convenience yield better, it is useful to
compare it to the convenience of holding cash in the monetary
economy. As suggested by the Keynesian liquidity preference theory of
money, three primary motives for holding cash can be categorized as
transactional, precautionary, and speculative. First, cash is handy for
facilitating immediate transactions. Second, we keep some cash for
precautionary reasons, as a safety buffer against unforeseen
disruptions. Third, when we make large investments, such as
purchasing a house, we partially speculate on the value of money today
relative to its value in the future.
The convenience of holding oil inventories can largely be explained
by the same three motives. The primary consumer of crude oil is a
refinery. The transactional motive is driven by the rigidity of refining
operations. To accommodate short-term fluctuations in the end-user
demand, it is more efficient for a refinery to hold extra barrels in
storage instead of going through the hassle of frequently adjusting
output schedules. The precautionary motive, like in the case of money,
is simply a safety buffer against unexpected disruptions which could be
caused, for example, by hurricanes or pipelines outages. The
speculative motive for holding oil stocks though is more widespread
than it is for holding fiat money. Being in the very center of the
petroleum value chain, many refineries possess superior fundamental
knowledge and often attempt to leverage it by adjusting inventories
based on their views of supply and demand.
Besides their main processing function of turning crude oil into a
usable product, refineries also provide a valuable auxiliary service to
deliver the right quantities of each product to customers whenever they
need it. This is what a good retail convenience store would do and then
include the service charge for warehousing in the price of the product.
Pushing storage onto the shoulders of end-users would be impractical.
Without an investment in a specialized infrastructure, we can only store
as much gasoline as our vehicle’s fuel tank allows. While we may call it
a convenience yield for the refinery, from the end-user perspective,
such an environment causes more of an inconvenience, as low
inventories are associated with higher prices.
Neither convenience yield nor storage costs are directly observable
in the oil market. What one can infer from the futures curve is only
their combined effect, which is the own rate. Nevertheless, such a
theoretical decomposition is still useful for understanding that the own
rate is driven by the tug of war between the intrinsic value of the
commodity and the cost to store it. The own rate of interest is one of
the most important drivers of many trading strategies. Futures traders
though give it a different name, the roll yield.

2.4 The Roll Yield


So far, we have only looked at the market from the perspective of a
physical trader who has access to a storage facility. However, most
financial traders do not have this access. The best that they can do in
trying to approximate the behavior of spot oil prices is to buy the
nearest maturity futures contract and roll it to the next maturity
contract at some time prior to the futures expiration. Such an
approximation, however, is rather tricky, as holding and rolling futures
may result in a very different outcome from the strategy of holding
physical barrels.
Let us assume that one buys the futures contract at price F(t, T) and
holds it until expiration when the futures price converges to the spot
price S(T). The profit-and-loss (P&L) on this trade can then be
decomposed into two terms:
(2.11)
The first term represents the change in the spot price, which is of
course, is unknown at time t. The second term compares spot and
futures prices at time t, both of which are observable from the shape of
the futures curve at that time.
While most oil traders look at P&L in dollar terms, financial
investors often prefer to measure the strategy performance using
percentage returns which makes comparison easier across asset
classes. As explained above, we use log-returns to make analytics less
cumbersome. Similar to (2.11), the so-called excess return (ER) that
measures the strategy performance consists of two terms11:

(2.12)

The first term in this decomposition is defined as the spot return (SR). It
represents a hypothetical investment return on buying a physical barrel
without incurring any storage costs, which cannot be realized by
trading futures. The second term is the roll return (RR). The roll return
also cannot be generated by means of a futures trading strategy, as it is
calculated by comparing prices for contracts with different maturities.
Both the spot return and the roll return act merely as attribution terms
in the decomposition of the actual return on investment in futures. In
practice, the roll return is calculated as the difference between the
actual realized excess return and the hypothetical spot return.12
It is more common to look at spot and roll returns in annualized
terms, in which case all terms in the Eq. (2.12) must be divided by T − t.
We then define the roll yield as the annualized roll return on this
strategy as follows

(2.13)

In other words,
(2.14)
It follows from (2.5), (2.10), and (2.14) that the roll yield, the own rate,
and the convenience yield are related as
(2.15)
In other words, the roll yield can be thought of as the spread between
the oil own rate of interest and the USD rate of interest. Alternatively, it
can be understood as the convenience yield, net of storage and interest
costs. Thus, we have defined an important variable that characterizes
the shape of the futures curve and translated it into three different
languages used by different market participants.
To illustrate the importance of the roll yield, consider a simple
strategy where a financial investor attempts to synthesize the
continuous ownership of a physical barrel by holding futures with the
nearest maturity. Since futures are listed with monthly expirations, in
this simplified strategy an investor buys one-month expiry futures,
liquidates the contract when it expires, and buys the next maturity
futures at that time. Note that such a strategy is unlikely to be
implemented in practice and it is used only for illustration purposes. In
the real world, most investors must liquidate futures prior to expiration
with more realistic examples of a buy-and-roll strategy analyzed in
Chap. 4.
Let t = T0 and assume that the investor repeats the strategy for N
consecutive months by buying Ti, i = 1, 2, …N maturity futures and
successively liquidating them at the spot price at time Ti. The
cumulative excess return (CER) of this strategy is

In each period, we use (2.12) to decompose the excess return into the
spot return and the roll return. The cumulative spot return (CSR) is
then

as contributions of all intermediate spot returns cancel out.


The cumulative roll return (CRR) is the difference between CER and
CSR

Figure 2.4 highlights the significance of the roll return in futures


trading.
Fig. 2.4 An example of spot returns and roll returns (WTI, Jan-Dec 2006). While cumulative
spot return was close to zero, an investment in rolling futures lost 23% because of the
accumulation of negative toll returns
It shows the performance of the buy-and-hold futures strategy
during 2006. We chose this year for illustration because the spot price
at the beginning and at the end of the period was practically the same.
In other words, the spot returns oscillated throughout the year but
ended up at zero. At the same time, the monthly roll returns, while
being smaller in magnitude, were consistently negative. As a result, the
investment in futures suffered 23% loss during the period when the
spot price did not changed at all. The loss was entirely driven by the
accumulation of the negative roll yield as the futures curve throughout
the year was in contango. We will discuss this topic in a much greater
depth in Chap. 4 in the context of financialization and commodity index
investments.
It should be clear by now that whatever the name of this magic
variable it, the own rate, the convenience yield net of storage or the roll
yield, it plays a critical role in trading futures. We now move to analyze
the drivers of this key variable. We look at them from two alternative
perspectives. In the next chapter, we first take a fundamental view of oil
prices being driven by supply, demand, inventories, and the economics
of storage. In the following chapter, we take a different approach and
study it from the perspective of financial flows and the demand for
hedging services. We will see that fundamentals and flows in oil trading
are, in fact, deeply intertwined.

References
Bouchouev, I. (2022). The Strategic Petroleum Reserve strategies: Risk-free return or return-free
risk? The Oxford Institute for Energy Studies.

Boyle, P. C. (1899, September 6). Testimony at the Hearing of the US Industrial Commission on
Trusts and Industrial Combination.

Fisher, I. (1896). Appreciation and interest. Publications of the American Economic Association,
11(4), 331–442.

Geman, H. (2005). Commodities and commodity derivatives: Modeling and pricing for
agriculturals, metals and energy. Wiley.

Giddens, P. H. (1947). Pennsylvania petroleum 1750–1872: A documentary history. Pennsylvania


Historical and Museum Commission.

Hull, J. C. (2018). Options, futures, and other derivatives (10th ed.). Pearson.
[zbMATH]

Kaldor, N. (1939). Speculation and economic stability. The Review of Economic Studies, 7(1), 1–
27.
[Crossref]

Keynes, J. M. (1936). The general theory of employment, interest, and money. Macmillan.

Naldi, N. (2015). Sraffa and Keynes on the concept of commodity rates of interest. Contributions
to Political Economy, 34(1), 17–30.
[Crossref]

Smiley, A. W. (1907). A few scraps: Oily and otherwise. The Derrick Publishing Company.

Sraffa, P. (1932). Dr. Hayek on money and capital. The Economic Journal, 42 (165), 42–53.
[Crossref]

Whiteshot, C. A. (1905). The oil-well driller: A history of the world’s greatest enterprise, the oil
industry. Acme Publishing Company.

Footnotes
1 See Fisher (1896). For consistency with contemporary conventions adopted in the economic
literature, our definitions of certain terms differ from Fisher’s original formulation. For
example, we use commodity price appreciation in terms of money to characterize commodity
inflation, in contrast to Fisher’s choice of looking at inflation as money depreciation in terms of
commodities.

2 The term spot price is used only for pedagogical clarity. We will explain shortly that oil spot
prices with instantaneous delivery do not really exist, and the spot price should be understood
as the forward contract with the nearest delivery.

3 Sraffa (1932) only defined the concept by virtue of the following example of borrowing bales
of cotton: “The rate of interest … per hundred bales of cotton, is the number of bales that can be
purchased with the following sum of money: the interest on the money required to buy spot 100
bales, plus the excess (or minus the deficiency) of the spot over the forward prices of the 100
bales”.

4 See Keynes (1936), chapter 17. Keynes restated Sraffa’s definition algebraically, which
triggered a heated debate between the two economists in a series of letters. The debate was
caused by their disagreement on whether to use spot prices or forward prices in defining the
dollar cost of borrowing the commodity. While it may appear to be a small technical nuance, it
led to a more profound philosophical differences in the interpretation of the concept. For
details, see, e.g., Naldi (2015). In this book, we accept the definition as it was stated by Keynes,
which simply mimics the definition of the interest rate on fiat money.

5 The one-year continuously compounded interest rate is conventionally defined as the


following limit
where n is a number of times interest is compounded in a year.

6 The history of SPR sales and loans is summarized in Bouchouev (2022).

7 The Organization of Petroleum Exporting Countries (OPEC) was formed in 1960. The
membership of OPEC has varied over the years. As of 2022, OPEC included thirteen countries
with the largest producer being Saudi Arabia. A larger but more loosely structured
organization, known as OPEC+, which included several other major sovereign producers, was
formed in 2016.

8 From the testimony of Patrick C. Boyle, proprietor and publisher of The Oil City Derrick at the
Hearing of the US Industrial Commission on Trusts and Industrial Combinations, see Boyle
(1899). The testimony was subsequently reprinted in Whiteshot (1905). It describes the origins
of oil trading and highlights that dump men were the predecessors of the oil refiners. The dump
business is also described in Smiley (1907), who wrote that “the dumps in those days were
practically the exchange, and made the market price for oil each day”. Many other interesting
facts related to the early days of oil drilling in Pennsylvania, including the construction of first
wooden oil tanks, are given by Giddens (1947).

9 Kaldor (1939) writes that “in normal circumstances, stocks of all goods possess a yield,
measured in terms of themselves, and this yield which is a compensation to the holder of stocks,
must be deducted from carrying costs proper in calculating net carrying costs. The latter can,
therefore, be negative or positive.”

10 See, for example, Geman (2005) and Hull (2018).

11 The term excess return differs from the total return as the latter includes the interest on
collateral. Since in this book, for the most part, we ignore the interest on collateral for futures,
the two terms are often used interchangeably. The distinction will become clearer in Chap. 4 in
the context of fully collateralized commodity indices.

12 Unlike price changes and log-returns, simple percentage returns are not additive. This
makes a similar decomposition for simple returns more complex. For simple returns, the roll
return is formally defined as the difference between the excess return and the spot return.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
I. Bouchouev, Virtual Barrels, Springer Texts in Business and Economics
https://doi.org/10.1007/978-3-031-36151-7_3

3. Fundamentals, Storage, and the


Model of the Squeeze
Ilia Bouchouev1
(1) Pentathlon Investments, LLC, Westfield, NJ, USA

The oil market is an example of a complex dynamic system. Such


systems operate in feedback loops with their behavior driven by
important boundaries. In the oil system, one boundary is set by zero
inventories. Another boundary is the constraint on the available
storage capacity.
The conventional theory of storage analyzes the joint dynamics of the
oil system, including interactions between supply, demand,
inventories, and prices. However, its practical use is limited by the
low price elasticity of oil supply and demand and challenges in
measuring production and consumption in a timely manner.
In a more practical modeling alternative, prices are viewed as
financial derivatives of stochastic mean-reverting inventories that are
directly observable. The boundary conditions are defined by the
squeeze of futures traders when either inventories are low or storage
capacity is scarce.
The infamous episode of negative oil prices presents an important
case study of a financial squeeze. While storage was the initial
catalyst for this event, the enforced behavior of financial market
participants caused unprecedented chaos.

3.1 The Invisible Hand of Storage Boundaries


Storage plays a vital role for any consumable commodity. It allows
supplies to be shifted in time, from days when they are plentiful to days
when they are scarce. The problem of optimal resource allocation
through time in the presence of uncertainty goes back to the very
beginning of human civilization. It is a difficult problem to solve.
Imagine yourself being Robinson Crusoe on a remote island, where
every day the decision must be made on what portion of the existing
food supplies to consume today, when you do not know how much
more, if anything, you will be able to find tomorrow. The decision how
much to eat today impacts the availability of food tomorrow, which also
depends on how much you expect to find tomorrow, and tomorrow’s
decision will depend on expected supplies and availability of food the
day after tomorrow, and so on.
A solution to a problem of this nature depends on the boundary
conditions. For Robinson Crusoe, the boundary is harsh, as at some
point he may run out of food. The mere presence of this zero-inventory
boundary and the possibility of it being reached in the future impacts
the decision today. The less food remains in stock, the more strongly the
presence of the zero boundary is felt. Moreover, if the resource is bulky,
or subject to degradation, such as raw fish, then its storage is also
capped by the limited warehousing capacity. If the commodity cannot
be easily discarded, then the presence of the second boundary on the
available storage capacity becomes equally important.
To prevent either boundary from being breached, some changes
must happen to commodity supply or demand, which jointly regulate
the flow of inventories. In this problem, the critical level of the stock
affects the flow of the stock. Such a property is the hallmark feature of a
complex dynamic system. The property induces mutual causality
between components of the system that operates via multiple feedback
loops. Some loops are known as positive, or reinforcing, loops that let
the system grow. Other loops are negative, or balancing, loops that keep
the growth from spiraling out of control. The world economy, the
human brain, and climate evolution are examples of such dynamic
systems. The problem of commodity storage is another one. Its primary
components are supply, demand, inventories, and prices. The dynamic
theory of storage is the study of the behavior of such a system.
Early studies of commodity storage were focused on agricultural
markets, as oil prices remained largely regulated until the late 1970s.
The goal of these studies was to assist governments with market-based
ways to reduce commodity price volatility as an alternative to price
controls, quotas, or subsidies. The work on commodity storage and its
impact on food prices was pioneered by Holbrook Working during the
Great Depression.1 He looked at a storage operation as a commercial
enterprise whose profitability is determined in the marketplace by the
competition among professional storage operators for the business of
the carry trade. As a consequence, the price of storage and the quantity
of storage must be determined jointly in equilibrium by the intersection
of the supply and demand curves for the service of storage.
In this framework, the storage reaches its equilibrium state when
the marginal revenue earned from selling the product equals the
marginal cost of making it. Such equilibrium is effectively spelled out by
the Eq. (2.​8) of the carry trade, discussed in the previous chapter
(3.1)
The left-hand side of this equation, F − S, aptly labeled by Working the
price of storage, is the marginal revenue received by the storage owner.
The price of storage is determined by market prices. The right-hand
side represents the total cost. It consists of the physical cost of storage
operations U, plus the financing costs R, but it treats convenience
benefits Y as an offset, or as a negative cost. The price of storage,
therefore, can be either positive or negative depending on the relative
magnitude of storage costs, including the costs of financing, and
convenience yield.2
The core of Working’s thesis is the claim that both sides of the Eq.
(3.1) are driven primarily by the level of inventories, and that the
relationship between prices and inventories becomes highly nonlinear
when inventories are particularly low. Low inventories suggest high
convenience yield and correspond to a backwardated market when
S(t) > F(t, T). In the limiting case of inventories approaching zero, the
spot price can rise relative to futures without any bound. As inventories
accumulate, the contribution of the convenience yield diminishes, and
the price of storage is determined more by the cost of storage. The
storage cost in this framework is assumed to be mostly constant,
increasing only slightly for exceptionally high levels of inventories. The
approach is summarized graphically in Fig. 3.1, where for consistency
with conventions in the oil market, the sign of S(t) − F(t, T) is flipped
compared to Working’s original presentation.

Fig. 3.1 Working’s representation of the relationship between the price of storage and
inventories. The sign of the spot-futures spread S − F is flipped for consistency with the
convention in the oil market
Even though Working’s graph is sometimes referred to as the first
theory of storage, his argument did not follow any analytical theory or
model. It only conceptually represented the nonlinearity of convenience
yield at low inventories based on empirical observations in several
agricultural markets. The objective of a proper storage model is to
analytically describe the behavior of the system that generates such
nonlinear dynamics of prices versus inventories within the model itself.

3.2 The Canonical Theory of Storage


Storage is modeled differently in academic studies and by practitioners.
We start with a more theoretical outline of the problem, known as the
canonical theory of storage.3 The limitations of this theory will then
lead us to a more practical alternative. The idea behind the dynamic
modeling of storage is to consider zero inventory as a critical boundary
condition whose impact is propagated backward in time to the present.
If inventories fall to zero because of persistent excess of demand
relative to supply, then the spot price must continue to rise until the
point of demand destruction. The possibility of this zero-inventory
boundary being reached in the future impacts the price at all prior
times. This price premium resulting from the risk of running out of
stock induces the convenience yield. The closer the inventory level is to
the zero boundary, the higher and more nonlinear the convenience
yield becomes.
Let us specify the components of the oil inventory system. Let x(t)
and x(t − dt) represent the level of inventories at times t and t − dt,
where dt is an increment of time. Inventories change whenever supply
and demand are unbalanced. The flow of inventories is then equal to
the difference between today’s supply and demand D(t):

(3.2)
The supply is the sum of the local production and imports from
other regions. The total demand D(t) is made up of the local
consumption and exports. We use tilde to highlight that the supply
is uncertain, and, for simplicity, we assume the demand to be
deterministic. The problem remains conceptually similar if, instead, the
uncertainty is attributed to demand, or to both supply and demand. The
Eq. (3.2) is an accounting identity; it does not represent any model or
assumption.
The integration of supply and demand imbalances over a period (t,
T) results in the aggregate level of inventories accumulated over this
period

The cumulative level of inventories must satisfy two operational


constraints. The inventories cannot be negative, or more precisely, they
cannot fall below a minimum operating capacity level Xmin, as some
amount of inventories must be maintained at all times for operational
reasons. Similarly, the inventories cannot exceed the maximum
operating capacity Xmax, which is generally below the nominal shell
storage capacity X:

If demand exceeds supply, then cumulative inventories decrease. They


continue to decrease until the lower boundary is reached, a situation
sometimes called a stock-out. In practice, the boundary is unlikely to be
reached. As inventories approach some critically low level, the system
sends a signal in the form of a higher price to supply and demand,
indicating that the danger is near, and something must change to
prevent any further outflows. Either the supply must increase, or the
demand must fall.
The opposite occurs when supply exceeds demand. In this case, the
inventories can build up until they reach so-called tank-tops when the
storage is operationally full. The upper boundary of the maximum
storage capacity then sends a lower price signal to force either supply
to be restricted or demand to increase. If, for whatever reason, the
demand disappears altogether, then the supply must also fall to zero to
prevent a breach of the storage capacity boundary. The presence of a
hard boundary on storage capacity differentiates oil from many other
commodities, such as metals and agricultural products, for which
storage technology is abundant and relatively straightforward. We will
see throughout the book that this boundary induces distinctive price
behavior which makes it difficult to apply generic commodity models to
trading oil.
We can now add the spot price of oil, S, to the system. Its role is to
convey information about the boundaries to supply and demand. We
define the demand to be a decreasing function of price:

The price is then expressed in terms of demand as the inverse demand


function
The price elasticity of demand is defined as the ratio of the percentage
change in quantity demanded to the percentage change in price.4 For
oil, such elasticity is very low, which means that the inverse demand
function is extremely steep and nearly vertical. In other words, the oil
price must change by a large amount to induce even a small change in
oil consumption. In the hypothetical asymptotic case of a complete
stock-out, the price must rise towards infinity if, for a given supply,
consumption cannot be restricted.
The presence of storage helps to alleviate short-term deficits and
buys the system some time to adjust. Let us define the availability a(t)
as the sum of the supply and the inventory carried from the
previous period x(t − dt). The problem of resource allocation is the
problem of splitting availability between the current demand D(t) and
the inventory to be carried at the end of the period x(t), so that:

Likewise, the availability a(t + dt) for tomorrow is the sum of


tomorrow’s production and inventory carried from today:

The presence of the same variable x(t) in both equations for today and
for tomorrow highlights the intertemporal nature of the storage
problem. While this variable represents a carry-out portion for the
availability today, it also acts as a carry-in for the availability tomorrow.
The availabilities on both days are, therefore, linked as follows
(3.3)
In other words, tomorrow’s availability is equal to today’s availability
minus what is consumed today, plus whatever can be replenished
tomorrow, with the latter being uncertain. The theory of storage looks
at the price of oil as an implicit function of its availability, S(t) = S(a(t)).
Both price and availability depend on inventory decisions, and they
must be determined jointly in equilibrium.
We now differentiate between two states of inventories: one normal
state that corresponds to non-zero inventories, and a second stock-out
state when inventories are depleted. In the first state, when some oil is
available, the spot price must be linked to the forward price for the next
period F(a(t + dt)) via the economics of the carry arbitrage trade. In the
second case of no inventories, the price is entirely driven by immediate
supply and demand and is determined by the inverse demand function.
Therefore,

(3.4)

Here, C(x) represents the total storage cost, which can also depend on
the level of inventories. To simplify the problem, we assume that C(x) is
a known nonlinear function that rises without limit as x → Xmax.
Alternatively, one could have considered a third state where inventories
reach the maximum storage capacity, x = Xmax. In this simplified two-
state formulation of the problem, C(x) can be viewed as a mirror image
of the inverse demand function at low inventories. Note that in the
stock-out state, the availability is the same as the demand, and D(t) in
the argument of the inverse demand function is replaced by a(t).
In the traditional formulation of the theory of storage, the futures
price in (3.4) is replaced by the expected spot price at time t + dt, which
depends on the availability a(t + dt).5 The expectation is taken over all
uncertain realizations of future production , which impacts the
availability a(t + dt) via (3.3). This highlights the circular nature of the
problem caused by the feedback loop between prices and inventories.
The price today depends on the availability today via the demand
function. Likewise, the price tomorrow depends on the availability
tomorrow. The prices today and tomorrow are connected via the carry
trade. At the same time, the availability today and the availability
tomorrow are linked via the Eq. (3.3). This problem is intertemporal, as
a similar loop exists for tomorrow, the day after tomorrow, etc. To solve
such a problem, one must construct an iterative process which
converges to some equilibrium state for both price and availability. The
process also needs a boundary condition, which in this case is defined
by the possibility of a stock-out.
This is a complex problem. Think about Crusoe’s decision-making
process as a tree, where at each time period alternative decisions could
be made about food allocation between immediate consumption and its
storage for tomorrow. As time moves forward, the tree of possible
scenarios branches out. Each node represents total supplies that will be
held at that time, which is the sum of inventories carried from the prior
period and whatever else he can find on that day. The process continues
until some random path leads to the case of zero inventories. The stock-
out sets the terminal boundary condition for the process, and the
problem is solved iteratively backward in time to establish the
inventory management rule at each prior step that maximizes the usage
of all inventories.
In general, to solve this problem, techniques of stochastic dynamic
programming must be applied. Fortunately, for the purpose of this
book, a precise solution of this complex problem is not needed. Our
primary interest is twofold. First is to use this framework to illustrate
an important feedback loop that exists between prices and inventories.
Second is to explain its limitations when applied to oil markets by
simply looking at salient properties of a typical solution. We then use
some valuable insights resulting from these limitations to develop a
simpler and more practical alternative approach in the following
section.
A typical solution to this problem shows the equilibrium price as a
function of availability, as illustrated in Fig. 3.2. It is characterized by
the kink that corresponds to the level of zero inventories separating the
two states, like in the Eq. (3.4). This kink symbolizes the dual role of
commodities, as an investment asset and as a product for consumption.
The location of the kink that marks the danger zone of a stock-out is
what the storage problem is effectively solving for. The price in the
region to the left of the kink is solely determined by production and
consumption. It is computed via an inverse demand function, as there is
no storage buffer in this region. To the right of the kink, oil behaves as
an investment asset. In this state, the price is mostly determined by the
cost of storage, which increases nonlinearly near the boundary on the
maximum storage capacity. It should also be noted that conventional
models require prices to remain positive. However, for the oil market
this assumption, as we will see shortly, is too restrictive.
Fig. 3.2 A typical solution of the canonical storage problem has a characteristic kink that
separates the consumption region with zero inventories from the investment region driven by
the economics of the carry trade
Despite its popularity in academic studies, the standard approach to
the storage theory attracted only very limited interest among oil
traders. One practical challenge is the extremely low price elasticity of
oil demand. The solution to the problem at zero inventories effectively
degenerates into a straight vertical line. It means that if we truly run
out of oil stocks, and the supply of oil cannot meet the demand, then the
price must rise towards infinity. Therefore, the optimal solution is to
always store oil in sufficient quantities so that a stock-out is never
reached. However, this solution is impractical because of the high cost
of oil and the limitation on the storage capacity.
A similar problem occurs when inventories reach maximum storage
capacity. Like oil demand, the short-term oil supply is also inelastic
with respect to price, as many producers cannot shut down production
instantaneously without permanently damaging oil reservoirs. In this
case, if a significant portion of the oil demand disappears, like it did, for
example, during the early days of the Covid-19 pandemic, then the
excess oil must go to storage. Once all storage is filled, then the price
must theoretically fall without any boundary to force either production
or consumption to change. We will discuss shortly the infamous
episode of oil falling to negative forty dollars per barrel, which is
probably as close as one can get to negative infinity in financial
markets.
To summarize, given that the price elasticity of oil demand and oil
supply is extremely low, the optimal solution to the oil storage problem
collapses into something that resembles the graph in Fig. 3.3. It is
driven by two extremities. The left tail marks the case of zero
availability, when the price goes to infinity. The right tail indicates that
the storage capacity is full, and the price goes to minus infinity.
Anywhere in between, where inventories are deemed to be normal and
sufficiently far away from either boundary, the price becomes largely
insensitive to the availability of inventories.

Fig. 3.3 The solution to the canonical storage problem degenerates at the boundaries when
the price elasticities of demand and supply approach zero

We now take this insight from the shortcomings of one model to


develop a much simpler and more practical alternative for the
relationship between prices and inventories that can actually be used
by oil traders. To do this, we borrow some wisdom from the physics of
extreme events.
3.3 A Stylized Model of the Squeeze
Boundaries play a critical role in the dynamics of any complex dynamic
system, such as the oil market. The conventional theory of storage
shows how the presence of two storage boundaries affects the behavior
of prices even away from the boundaries. One boundary marks the
scenario of running out of oil, in which case the spot price moves higher
until either supply or demand adjusts. The second boundary
corresponds to the scenario of running out of storage, which forces the
spot price to drop to prevent further inflows. These boundaries do not
need to be reachable. In fact, in most dynamic systems boundaries are
understood as unattainable limits, but their mere existence determines
the system behavior within the boundaries.
What makes the traditional theory of storage particularly complex is
the attempt to explicitly model the response of prices to changes in
supply and demand. In practice, the concept of an inverse demand
function that plays an important part in the theory has limited
applications. In the oil market, neither supply nor demand is easily
observable. Supply data is published with a significant lag and
frequently revised, sometimes even going back several years.
Measuring oil demand is even harder. The demand is typically
synthetically reconstructed from estimates of supply and measured
inventory data using the identity (3.2). The primary variable that
traders can track is inventory. For the storage model to be of any
practical use, it must be based on observable inputs. Instead of making
assumptions about supply and demand and calculating inventories, the
trader is better off modeling the behavior of inventories directly.
The presence of boundaries causes supply and demand to adjust in
a complicated way that we will not even attempt to model. Instead, we
model the uncertainty driven by these fundamental adjustments that
must cause the flow of inventories to reverse when the level of
inventories approaches the boundary. Such behavior can be
characterized by a mean-reverting process, where inventories are
pulled towards their long-term equilibrium level. We first illustrate
mean-reverting behavior in an idealized case where the inventory path
towards equilibrium is certain, or, in other words, deterministic. We
then add some noise or uncertainty to the process.
In the deterministic case, the change in the level of inventories dx
over a period dt can be described by the following equation
(3.5)
Here, represents the long-term equilibrium level of inventories, and
the parameter k defines the speed at which inventories revert to this
level. If current inventories are above their long-term level, i.e.,
, then the Eq. (3.5) implies that dx < 0, and, therefore,
inventories decrease. Likewise, if current inventories are below the
long-term level, i.e., , then dx > 0 and inventories rise.
If we assume some initial state of the inventories at time t = t0

then the solution to the ordinary differential Eq. (3.5) can be easily
found, as follows
(3.6)
One can easily verify it by direct substitution. As time increases, the
inventories x(t) are pulled from their current state x0 towards the
equilibrium level . The pull occurs at an exponential rate defined by
the parameter k. As t → ∞, inventories converge towards the
equilibrium, i.e., . The larger the parameter k, the faster the
speed of mean-reversion.
The solution (3.6) is illustrated in Fig. 3.4 for two scenarios. In the
first example of slow draws, the initial inventory level x0 is high and k is
relatively small, which could be the result of a prior fall in demand,
perhaps caused by an economic recession. As demand gradually
recovers, inventories decrease and slowly converge to the long-term
equilibrium. In the second scenario of rapid builds, initial inventories
are low but mean-reversion is fast, which could be the result of a short-
term supply disruption caused by a geopolitical event, but the supply is
assumed to be restored relatively quickly.
Fig. 3.4 Examples of slowly drawing and rapidly building inventories that converge to a long-
term equilibrium
The Eq. (3.5) describes an unrealistic dynamic where one can
predict future inventories with certainty. In the real world, the path
towards an inventory equilibrium state is likely to be noisy with some
random fluctuations along the way that can either delay the
convergence to the normal state or accelerate it. To capture this noise, a
second term is added to the Eq. (3.5):
(3.7)
The noise is assumed to come from an increment dz, which represents a
random variable, drawn from a normal distribution with mean of zero
and variance equal to dt. It is scaled by the volatility parameter σ, which
characterizes the magnitude of the uncertainty.
As time t moves forward, the market may experience large,
unexpected shocks to supply and demand that are described by large
random values of dz(t). During such episodes, the random term in (3.7)
dominates the change in inventories. However, over time the steady
gravitational pull of the mean-reverting term ensures that inventories
drift towards an equilibrium level. The further inventories deviate from
the normal level, the stronger the pull towards the long-term
equilibrium. The Eq. (3.7) is one of many examples of stochastic
processes that are commonly used to describe uncertainty in financial
markets. It is known as an Ornstein-Uhlenbeck process, which was
originally developed in physics to analyze Brownian motion with
friction. It has been widely adopted for modeling mean-reversion in
commodity prices.
When noise is introduced to the system, the forward dynamics is no
longer certain. It can only be understood in a probabilistic sense. Each
stochastic process is associated with a probability density function p(x,
t; xT, T) which describes transition probabilities of the random variable
between two states at times t and T. If the variable at time t < T is
known to be located at the point x, then the density function describes
the probability of this variable reaching various points xT at some
future time T. Alternatively, if the variable is known to be at some target
state xT at time T, then the same density function also represents the
probability of having reached this target state starting from different
levels x at an earlier time t < T. The probability density function
satisfies the Kolmogorov forward and backward equations, with the
former also known as the Fokker-Planck equation. These equations and
other basic facts related to transition probabilities and stochastic
processes are summarized in Appendix A.
The most well-known probability density function is a Gaussian
bell-shaped curve that represents the normal distribution, centered at
the point xT = x with variance σ2(T − t):

(3.8)

This normal probability density corresponds to the stochastic process


(3.7) without mean-reversion, i.e., for k = 0, which is known as
Arithmetic Brownian Motion (ABM).
Figure 3.5 illustrates the normal probability density functions (3.8)
for different T − t.
Fig. 3.5 Normal probability density for decreasing time variance
As T − t decreases, the normal probability density function becomes
thinner and taller, as the variance is reduced and the probability mass is
concentrated around the mean, where xT = x. In the limiting case when
t → T, the probability density takes a very peculiar form. It falls to zero
for all values of xT, except for a single point xT = x, where it goes to
infinity.
This limit is known as the Dirac delta function δ(xT − x), which is
defined as

(3.9)

The Dirac delta function describes the boundary of the system. It is


frequently used in physics to model extreme but largely unattainable
states. One popular example arises in modeling the dissipation of a
point heat impulse across a metal bar. As time passes, the heat diffuses
along the bar according to the differential equation of heat transfer, for
which the Dirac delta function serves as an initial condition. The Dirac
delta function possesses some interesting properties. Defined as the
limit of probability density functions which must integrate to one, it
also integrates to one. Also, when the Dirac delta function concentrated
at x is multiplied by an arbitrary function and integrated over all
possible values of xT, the integral returns the value of the function at the
point x. These properties are summarized in Appendix A.
We are now in a position to apply the technique of stochastic
calculus to describe the behavior of oil inventories and prices.
Stochastic processes are frequently used to model oil prices with so-
called reduced-form models. In such models, one assumes a certain
stochastic process for the spot price and for some additional state
variables, such as the convenience yield or interest rate. The theoretical
futures price is then derived as a function of parameters of stochastic
processes for state variables, which are calibrated to market prices of
traded futures. This approach has been the dominant modeling
framework for nearly three decades, starting from the 1980s, but its
relevance has waned over time.6 As the futures market developed and
became observable, the need to model futures using somewhat
nebulous state variables, such as convenience yields, largely
disappeared. Perhaps more importantly, reduced-form models ignore
the fundamental dynamics of the entire energy system, which made this
approach less popular among practitioners.
The canonical storage does take into consideration the dynamics of
the system but it is equally unpopular among practitioners, as it is
complex, unstable, and based on hard-to-measure variables, such as
supply and demand. In this chapter, we propose a more practical
approach to the problem by blending the two classical paradigms and
applying stochastic calculus instead directly to inventories, which is the
primary fundamental variable that traders track.
While the traditional theory of storage attempts to link inventory to
the outright price of oil, traders instead typically relate inventories to
the spread between the spot and the futures price. The spot-futures
spread isolates the exposure to fundamentals, as two legs of the spread
provide largely offsetting exposure to many exogenous non-
fundamental factors. It is also consistent with Working’s original
formulation of the storage problem. In practice, the spot oil price is
approximated with the nearest-delivery futures contract. The spread
between futures with two maturities then serves as a barometer for the
economics of the carry trade and the state of inventories.
In our practical formulation of the storage problem, we define the
futures time spread as a financial derivative of inventories x(t):
The inventory plays the role of an uncertain state variable whose
behavior is described by a stochastic process, such as, for example,
(3.7).
Ignoring, for simplicity, the effects of discounting, the price of a
financial derivative, as shown in Appendix A, must be equal to the
expected value of its terminal payoff s(xT, T) at the expiration time T:

(3.10)

We will see later in Chap. 8 that, in general, financial derivatives must


be priced as the cost of their dynamic replication by trading underlying
futures. In this case, the probabilities in formula (3.10) are understood
as so-called risk-neutral probabilities that correspond to a stochastic
process in which the drift term is zeroed out. However, when the
underlying instrument represents a non-tradable state variable, such as
oil inventories, the derivative of the state variable is simply equal to the
expected value of its payoff under regular real-world probabilities.7
To find the value of the futures spread, s(x, t), we need to specify the
boundary condition at time T = T1 when the first leg of the spread
expires, and the carry trade that connects two futures must be closed.
Inspired by insights from the conventional storage theory of the
previous section, we distinguish among three scenarios. In one normal
scenario, the futures spread is determined by the negative of the cost of
storage −C. Here, we do not need to impose any nonlinear cost of
storage function, as such nonlinearity will be generated endogenously
within the model, driven by the proximity to the storage boundary.
Two other scenarios correspond to two boundary cases when the
carry arbitrage cannot be completed, either because of zero inventories
or because of zero remaining storage capacity. In the case of zero
inventories, the holder of the short futures position has no choice but to
bid up the price of the nearby futures contract in the market. However,
in this scenario there will be no sellers since no one can deliver physical
barrels against the short futures position. The price can then keep
rising without any limit. Likewise, the holder of the expiring long
futures position who does not have access to storage capacity will try to
sell futures to avoid taking the delivery of physical barrels by pushing
the price down without any limits.
The spread behavior at these extremes that represent largely
unattainable asymptotic states can be described by a pair of Dirac delta
functions with opposite signs, concentrated, respectively, at Xmin and
Xmax:

(3.11)

One can think about infinite prices as the event of default by futures
traders. More practically, the market calls such an event a squeeze, and
we refer to this formulation of the storage problem as a stylized model
of the squeeze.
The solution to the storage problem now becomes simple and
intuitive. It is given by the derivative pricing formula (3.10) with the
boundary condition (3.11):

Note that the integration here is effectively performed over the range
xT ∈ (Xmin, Xmax), outside of which the probability density function is
equal to zero. Applying properties of the Dirac delta function (A.​4) and
(A.​5), we obtain that the price of the futures spread is given by
(3.12)
This representation explicitly relates the price of the futures spread to
market fundamentals, specifically to the probability distribution
function for stochastic inventories. It follows that the value of the
futures spread at time t is determined by the relative probabilities of
upside and downside squeezes, adjusted by the normal cost of storage.
One squeeze corresponds to the case of zero inventories, and the other
one to the case of zero available storage capacity. The proximity to each
boundary determines the value of the spread within the boundaries at
all times prior to expiration.
Let us now apply the generic formula (3.12) to a mean-reverting
stochastic process (3.7). It is shown in Appendix A that the probability
density for the mean-reverting process is also described by the normal
distribution

(3.13)

but it is applied to a modified state variable


(3.14)
and to the new time variable , which is defined by

(3.15)

The state variable y(t) resembles the function (3.6) that describes the
behavior of inventories without the added noise. The role of the new
time variable is to capture the impact of mean-reversion on the
process variance. It reduces the effective variance via an exponential
dampening factor that depends on the speed of mean-reversion k. For
example, if the mean-reversion is slow and k → 0, then and the
variance of the mean-reverting process converges to the variance of the
normal process. However, for k > 0 the new effective remaining time to
maturity is reduced, as . We should also note that at
time t = T, the new time , y(T) = xT, and the boundary condition
(3.11) remains intact in new variables.
The substitution of the probability density (3.13) into the
representation (3.12) results in the following pricing formula for the
futures spread:

(3.16)

where the variables y and are defined above by (3.14) and (3.15).
The solution (3.16) to our stylized model of the squeeze is
sufficiently flexible to cover the wide range of plausible spread
behaviors versus the level of inventories. Figure 3.6 illustrates that
faster mean-reversion makes the range of spread values more
contained. In this example, the inventories are expressed as a
percentage of total storage capacity, so that 0 ≤ x ≤ 1, and the long-term
equilibrium inventory level is set at . Since is slightly closer
to the upper boundary, this makes the probability of tank-tops higher
than the probability of stock-outs, resulting in wider contango than
backwardation near the boundaries. The opposite effect can be
generated by choosing to be closer to the lower boundary. One can
easily extend this framework and apply (3.12) to the more complex
dynamics of inventories characterized by different stochastic processes
and probability density functions.

Fig. 3.6 Examples of one-month futures spread functions (3.16) for fast (k = 4) and slow
(k = 1) mean-reversions with 0 < x < 1 and
We now move from theory to practice and illustrate the approach
for the world’s most important storage hub, where the futures market
meets the physical barrels. For nearly three decades, all roads in the oil
business have led to the rather nondescript US town of Cushing,
Oklahoma, which has often been called the oil capital of the world.

3.4 Cushing: Pipeline Crossroads of the World


While the theoretical thesis behind the theory of storage and the
nonlinear inverse relationship between inventories and prices is rather
intuitive, its validation in practice has been much more elusive. One
practical challenge is availability of high-quality inventory data. For the
storage model to be tested on real data, the storage facilities must be
located within the proximity of the corresponding pricing point.
Fortunately, the storage hub at Cushing, the delivery point for the
benchmark WTI contract, presents a unique opportunity to test the
theory.
Cushing is an important juncture in the oil trading system where the
connection between physical and financial markets is very explicit. If a
WTI futures trade is not closed before its expiration, then the holder of
the short position must deliver physical barrels to one of the designated
locations within Cushing tank farm, and the holder of the long futures
must take barrels from there. Taking or delivering barrels in Cushing
means reserving space on one of the dedicated pipelines. These
pipelines are not easily accessible by many financial traders. Further,
one cannot really deliver a truckload of oil and unload it there. If you do
not have a membership in this elite Cushing club, then you are out of
luck. In practice, if you do not close the expiring futures contract, then
your clearing house will close the trade without even telling you, then
ask one of the Cushing club members to handle physical barrels for a
healthy fee and send you a large bill to settle.
This should look sufficiently scary to disincentivize anyone without
a Cushing club membership from dabbling in WTI futures contracts
near their expiration. The vast majority of rational investors indeed
stay away from this game, especially when inventories are within reach
of the critical boundaries. However, from time to time, some traders,
driven either by greed or more likely by ignorance, decide to take their
chances. The brave ones can choose to play this game until the final
buzzer, often forgetting that the less time remains on the clock, the
more they are exposed to the squeeze. The most infamous example of
such a squeeze occurred on the day when oil prices went negative,
which we describe shortly.
Cushing is a landlocked location which is connected via pipelines to
production and consumption centers and to the Gulf Coast for exports
and imports. The supply of oil to Cushing is measured by flows on many
inbound pipelines from various production areas in the USA and
Canada. The demand is estimated by flows on outbound pipelines
delivering barrels to domestic refineries. Unlike many other oil trading
hubs located near ports, Cushing is landlocked and better ringfenced
from the impact of broader international factors. This relative isolation
from global trade factors makes local inventories at the Cushing storage
hub more suitable for forecasting local prices. Professional oil traders
use a variety of sources and sophisticated technologies, such as aerial
imagery and infrared cameras, to outsmart their peers in measuring
pipeline flows and the amount of stored inventories.
While the proper Cushing model requires higher-frequency private
data, the main characteristics of storage modeling can be illustrated
with some publicly available data. Such data is provided by the US
Energy Information Administration (EIA), the statistical agency of the
US Department of Energy (DOE). Every Wednesday, the agency
publishes a widely anticipated report with weekly estimates of various
fundamental variables, including inventories across various US regions.
While the state of inventories at Cushing is perceived to be the most
relevant to WTI prices, inventories in broader regions should also be
considered, as barrels can be shipped quickly between certain storage
locations. For example, the storage in the Midwest area, known as
PADD2 (Petroleum Administration for Defense District), is well
connected to Cushing and frequently used by analysts, along with
overall US inventories.
Figure 3.7 shows a time series of inventories in these three
locations. The data exhibits a common trend driven by the growth of US
shale production and the need to support the enhanced infrastructure.
The presence of the common trend in data significantly contaminates
the quality of statistical correlation analysis. It would be incorrect to
apply correlations, for example, to oil prices and raw inventories as
neither time series is stationary. However, the futures spread is
stationary as two legs of the spread negate a large portion of the
common trend. To remove such a common trend in raw inventories,
one must normalize them as well.
Fig. 3.7 Inventory levels for Cushing, PADD2, and US overall exhibit the common upward trend
driven by the growth of the shale industry
The methods of inventory scaling vary with overall objectives and
the availability of data. For the broader markets, where inventory data
is harder to collect, one often measures inventories relative to their
historical averages, which are often seasonally adjusted. For example,
OPEC uses a five-year inventory average as a benchmark for decision
making, which makes traders joke that OPEC can simply wait five years
to bring distorted inventories to the normal level. Other analysts
measure inventories in terms of days of demand, which has been
steadily growing over the years. For US markets, where more-granular
data are available, a more accurate metric can be constructed if
inventories are scaled with the overall storage capacity. The resulting
inventory variable, x(t)/Xmax, represents the inventory capacity
utilization.
In contrast to raw inventories, which are trending, the storage
capacity utilization for all three locations is stationary. Figure 3.8
clearly shows more oscillatory or mean-reverting behavior of
normalized inventories with no visible trends. This metric is
particularly important in modeling crucial nonlinear effects near
storage boundaries, which in this case are defined explicitly.
Fig. 3.8 Inventory levels for Cushing, PADD2, and US overall normalized by the storage
capacity are stationary
The historical correlation between the futures spreads and
normalized Cushing inventories over the entire sample since the
storage capacity data became public in 2011 is approximately negative
60%. While such a high correlation validates the basic thesis of the
storage theory that futures spread is inversely related to inventories,
the relationship is highly nonlinear, as shown in Fig. 3.9.
Fig. 3.9 The inverse relationship between WTI spreads and Cushing inventories is
approximated by the model of the squeeze
To illustrate the crux of the storage theory and the nonlinear
behavior around the boundaries, we overlay our stylized theoretical
solution (3.16) to the storage problem. This squeeze function captures
important nonlinear price signals when the market is running out of
inventory or out of storage capacity. The specific model parameters can
be fitted to sub-samples corresponding to prevailing market conditions.
We will leave this exercise to the reader, as econometric analysis of
historical data is not the main objective of this book.
So far, we have only illustrated the storage theory with a static
snapshot, where the inverse relationship between futures spreads and
inventories is measured contemporaneously. While it confirms the
thesis of the storage theory, it does not necessarily help to turn the
concept into trading profits, as we do not know yet whether inventories
have any predictive power for prices. The topic of actually trading
futures based on quantitative analysis of inventory data is more
nuanced. We will return to quantitative trading strategies based on
fundamental data in Chap. 6 when we discuss so-called quantamentals.
One can also notice a few outliers in Fig. 3.9, specifically when the
market fell into a steeper contango than was justified by prevailing
inventories. The cause of nearly all of these outliers is the forward-
looking nature of financial markets that trade based not on today’s level
of inventories, but rather based on expected forward inventories at the
time of the delivery of physical barrels. For the most part, the difference
between the two is minor, as inventories generally change slowly.
However, if something causes inventories to change quickly, then
forward expectations play the decisive role. If the development happens
to be near one of the two storage boundaries, then the consequences of
incorrect expectations could be dramatic. The best way to illustrate
such boundary behavior is to look more deeply at its most famous case
study, the unprecedented episode of negative oil prices.

3.5 Negative Oil Prices


On April 20, 2020, the oil market made history. The price of the prompt
WTI futures contract went negative. It did not just dip slightly below
zero; it collapsed to a mind-blowing negative forty dollars per barrel.
For the first time in the history of financial markets, an asset was priced
at the negative of its typical value. The historic price collapse was not
limited to futures. The prices for physical barrels also went negative, as
they were attached to the WTI benchmark via basis spreads which
were too slow to adjust to the fast pace of futures.
The speed at which the futures fell was staggering. It is unlikely to
be explained by any conventional economic models of rational behavior.
It took the prompt oil futures only thirty minutes to move from zero to
negative forty dollars per barrel, as illustrated in Fig. 3.10. At the same
time, the price of the second nearby contract barely changed. The front
spread, the benchmark for the price of storage, widened to negative
sixty dollars per barrel, far in excess of even the most expensive way of
storing oil; by moving it to the coast and loading it on ships for floating
storage. Why did it happen?
Fig. 3.10 On April 20, 2020, the price of the May 2020 futures contract traded at negative
$40/bbl and the spread between May and June contracts at negative $60/bbl
In the real world, the storage boundaries work their powers
somewhat differently from the theory. Prices are driven not by a magic
wand of storage but rather by the behavior of specific market
participants reacting to the information conveyed by storage. The
market was indeed oversupplied since the onset of the Covid-19
pandemic when oil demand fell virtually overnight. The supply was
slow to adjust, with the downward pressure on price further
exacerbated by disagreements and a short-term price war between
large sovereign oil producers. Inventories were increasing at an
alarming pace, but storage was nowhere close to full.8 Figure 3.11
illustrates.
Fig. 3.11 At the time when WTI fell below zero, Cushing storage was not yet full. However, the
market was extrapolating the recent inventory trend, which would have breached the
maximum storage capacity
Importantly, what matters for prices are not today’s inventories, but
rather, inventories at the time of the expected delivery, which in this
example corresponds to the following month. Forecasting future
inventories during a period of such uncertainty is a difficult task. The
most straightforward way to do it is simply to extrapolate the recently
observed trend in inventories. Since what goes into storage is the
excess of short-term supply over short-term demand, and since supply
and demand are slow to adjust, the inventory change tomorrow should
be broadly similar to the inventory change today. Therefore, in the short
term, the slope of the inventory trend should remain intact. Such naïve
extrapolation of the slope would indeed have breached the hard
boundary of the maximum storage capacity sometime during the
following month, when the delivery of physical barrels must take place.
This is where the important self-regulating property of the complex
dynamic system comes into play; the level of the stock approaching the
boundary must impact the flow of the stock to prevent the system from
overfilling. As discussed earlier, the dynamics of such a system is driven
by the interaction between the balancing and reinforcing loops. The
balancing loop here is enforced by storage. The presence of the hard
boundary forces either supply or demand to adjust and stop the flow of
the stock. The demand is always made up of the immediate
consumption by refineries and the demand for storage. While the
former was constrained by pandemic restrictions, the latter is a
function of the cost to transport oil elsewhere and to store it in more
remote locations. If the physical system was left to operate alone
without the disturbance caused by the financial markets, then the
futures spread would have likely traded around $5/bbl which was the
cost of moving oil out of Cushing at that time and storing it elsewhere.
The challenge is that financial futures markets operate via their own
feedback loop, which at a time of stress and panic, often acts in a self-
reinforcing manner. Let us briefly describe the mechanics of this panic
which caused some irrational behavior near the boundary, leading to
the squeeze. Recall that financial investors gain exposure to oil prices
by buying and rolling futures. It is tempting to think that the closer one
can hold futures to expiration, the more closely the future price should
mimic the price of the physical barrel. This argument is somewhat
dubious though, because the futures market is where the price for the
physical barrel is determined, and not the other way around.
Individual investors are rarely allowed by their clearing brokers to
hold futures near expiration due to extremely high volatility and the
risk of being squeezed. Professional traders, however, such as banks,
can hold futures for hedging purposes effectively until the final buzzer.
This also allows banks to offer clients OTC replicas of futures that can
be held closer to expiration. While most financial products avoid taking
this risk by rolling futures a few days or weeks prior to expiration, a
handful of dealers were still holding long futures on behalf of their OTC
investors on April 20, 2020, one day prior to the contract expiration.
One such dealer still holding a large quantity of long futures during
that time was the Bank of China. With oil prices already in a freefall
since the start of the Covid-19 pandemic, the bank’s investment
product named YuanYouBao, which was advertised as “oil cheaper than
water”, attracted enormous interest from domestic retail investors. The
product was marketed as a safe non-leveraged investment with the
entire notional value of oil prepaid upfront, in contrast to leveraged
futures contracts, which only require the initial margin to be posted.
The structure of the contract, of course, assumed that the notional
amount remains positive. From the investor’s perspective, it was then
reasonable to think that the price of OTC oil-linked product is bounded
by zero, like the price of a common stock.
When the bank sells an OTC replica of futures and buys actual
futures as a hedge, the bank must either liquidate or roll futures at the
futures settlement price on a given day. To make it easier for dealers,
the so-called Trading at Settlement (TAS) contract has been created by
futures exchanges. In this contract, buyers and sellers agree to
exchange futures at a price which will only be determined later, after
the trading session ends. If the dealer wants a particularly speedy
execution, then a TAS contract can be offered at a slight discount or
premium to the settlement price to incentivize counterparties to take
the other side. In a normal market, an extra one cent per barrel of
premium or discount is usually sufficient to attract enough
counterparties and ensure prompt execution of a TAS order.
However, given the tremendous uncertainty about the direction of
the Covid-19 pandemic, the behavior of the May 2020 WTI contract was
anything but normal. Extra demand from oil bargain hunters, further
stimulated by OTC dealers, led to a large quantity of long futures held
on behalf of individual investors. These futures had to be liquidated or
rolled by the dealer at the TAS price on April 20, 2020. The normal TAS
liquidity quickly dried up, forcing the dealers to offer progressively
larger discounts to sell futures relative to a still undetermined TAS
price. Even when the discount reached its maximum allowed $0.10/bbl,
the order to sell futures at the TAS price remained unfilled. This
imbalance telegraphed to the entire market that someone in the market
who must sell futures at the closing price is struggling to do it. It meant
that the only alternative for the holder of long futures was to sell them
instead at the regular market prices as close as possible to the end of
the trading session in the hope of matching the settlement price. With
dealers suddenly unable to trade TAS, panic kicked in and unfilled
futures were sold in a rapid-fire fashion into the vacuum within thirty
minutes before the market closed.
Shortly after the market closed at the historical price of negative
$37.63/bbl, the price bounced back to where it was trading a few hours
prior, as storage buyers rushed to capitalize on this unique opportunity.
The balancing loop of the physical storage came to the rescue and
contained the self-reinforcing loop of financial flows that went out of
control. Unfortunately, like in many other complex systems, the
balancing action came with a delay. The delayed reaction of storage
traders was largely driven by operational controls within professional
trading shops, such as the need to secure additional risk limits given
unprecedented market volatility. In the end, buying prompt futures at
nearly $60/bbl discount to the next-maturity futures was a real gift to
anyone with membership in the Cushing storage club. At such a price,
with a bit of creativity some could have found a way to store oil in their
backyards.
The financial loop of naïve investors had created and effectively
subsidized this unique opportunity for storage traders. Contractually,
the losses from sales at negative prices were supposed to be passed by
the dealer to investors. However, investors revolted, blaming the faulty
product design and pointing to an implicit zero price boundary in the
contract. After some government intervention, fines were imposed on
the dealer, who ended up paying most of the losses. Dealers learned a
painful lesson, that the market risk is not the only type of risk that must
be managed in the complex world of OTC oil derivatives.9
The nature of interactions between physical and financial markets
became more visible during this memorable episode of negative oil
prices. However, such interactions were not specific to that single day.
They happen every single day in the derivatives market for virtual
barrels. Similarly to how physical storage with its boundaries drives the
dynamics of one complex system, the intricate web of interactions
among participants in financial markets drives another. The latter is the
subject of our next chapter, the theory of hedging pressure.

References
Bouchouev, I. (2020, April 30). Negative oil prices put spotlight on investors, Risk.net.

Bouchouev, I. (2021). A stylized model of the oil squeeze, SSRN.

Brennan, M. J. (1958). The supply of storage. The American Economic Review, 48(1), 50–72.

Brennan, M. J. (1991). The price of convenience and the valuation of commodity contingent
claims. In D. Lund & B. Oksendal (Eds.), Stochastic models and option values. North Holland.
Brennan, M. J., & Schwartz, E. S. (1985). Evaluating natural resource investments. Journal of
Business, 58(2), 135–157.
[Crossref]

Carmona, R., & Ludkovski, M. (2004). Spot convenience yield models for the energy markets.
Contemporary Mathematics, 351, 65–79.
[MathSciNet][Crossref][zbMATH]

Casassus, J., & Collin-Dufresne, P. (2005). Stochastic convenience yield implied from commodity
futures and interest rates. The Journal of Finance, 60(5), 2283–2331.
[Crossref]

Clewlow, L., & Strickland, C. (2000). Energy derivatives: Pricing and risk management. Lacima
Publications.

Deaton, A., & Laroque, G. (1992). On the behavior of commodity prices. The Review of Economic
Studies, 59(1), 1–23.
[Crossref][zbMATH]

Dempster, M. A. H., Medova, E., & Tang, K. (2012). Determinants of oil futures prices and
convenience yields. Quantitative Finance, 12(12), 1795–1809.
[MathSciNet][Crossref][zbMATH]

Dvir, E., & Rogoff, K. (2009). Three epochs of oil, NBER Working Paper, 14927.

Eydeland, A., & Wolyniec, K. (2003). Energy and power risk management: New developments in
modeling, pricing, and hedging. Wiley.

Fernandez-Perez, A., Fuertes, A.-M., & Miffre, J. (2021, Summer). On the negative pricing of WTI
crude oil futures. Global Commodities Applied Research Digest, 6(1), 36–43.

Gibson, R., & Schwartz, E. S. (1990). Stochastic convenience yield and the pricing of oil
contingent claims. The Journal of Finance, 45(3), 959–976.
[Crossref]

Gustafson, R. L. (1958). Carryover levels for grains, U.S. Department of Agriculture, Technical
Bulletin, 1178.

Hamilton, J. D. (2009). Understanding crude oil prices. The Energy Journal, 30(2), 179–206.
[Crossref]

Interim Stuff Report. (2020). Trading in NYMEX WTI crude oil futures contract leading up to, on,
and around April 20, 2020, Commodity Futures Trading Commission, November 23.

Kilian, L. (2020). Understanding the estimation of oil demand and oil supply elasticities, Federal
Reserve Bank of Dallas Working Paper, 2027.

Ma, L. (2022). Negative WTI price: What really happened and what can we learn? The Journal of
Derivatives, 29(3), 9–29.
[MathSciNet][Crossref]
Miltersen, K. R. (2003). Commodity price modelling that matches current observables: A new
approach. Quantitative Finance, 3(1), 51–58.
[MathSciNet][Crossref][zbMATH]

Pirrong, C. (2012). Commodity price dynamics: A structural approach. Cambridge University


Press.

Routledge, B. R., Seppi, D. J., & Pratt, C. S. (2000). Equilibrium forward curves for commodities.
The Journal of Finance, 55(3), 1297–1338.
[Crossref]

Schwartz, E. S. (1997). The stochastic behavior of commodity prices: Implications for valuation
and hedging. The Journal of Finance, 52(3), 923–973.
[Crossref]

Williams, J. C., & Wright, B. D. (1991). Storage and commodity markets. Cambridge University
Press.
[Crossref]

Working, H. (1948). Theory of the inverse carrying charge in futures markets. Journal of Farm
Economics, 30(1), 1–28.
[Crossref]

Working, H. (1949). The theory of price of storage. The American Economic Review, 39(6), 1254–
1262.

Footnotes
1 Working’s original observations on commodity storage were published between 1929 and
1933 in a series of short articles issued by Wheat Studies of the Food Research Institute. The
more complete version of his argument is summarized in Working (1948) and Working (1949).
The approach was further extended by Brennan (1958).

2 For simplicity, we use the terms convenience benefits and convenience yield interchangeably
and refer to the previous chapter for their definitions.

3 The theory of storage was pioneered by Gustafson (1958) and subsequently extended in
multiple directions. The formulation that we present here is largely based on the ideas of
Deaton and Laroque (1992). This method has been applied to the oil market by Dvir and Rogoff
(2009). For other related methods of solving this problem, see Williams and Wright (1991),
Routledge et al. (2000), and Pirrong (2012).
4 More formally, the price elasticity of demand is defined as . While it is difficult to
measure this elasticity precisely, it has been steadily declining over time, as increasing overall
wealth made consumers less sensitive to energy prices. See, for example, Hamilton (2009) and
Kilian (2020).

5 The assumption that the futures price is equal to the expected spot price implies that the
futures price is fair and unbiased. It is equivalent to the so-called risk-neutral pricing that we
will formally introduce in later chapters. In the next chapter, we also consider the case when
futures price differs from the expected spot price, as futures may be distorted by imbalances in
the hedging market.

6 Brennan and Schwartz (1985) applied a one-factor lognormal model for the spot price with
deterministic convenience yield to derive futures prices. The model was further extended in
Brennan (1991). One-factor models, however, have quickly proven to be too restrictive, as they
allow futures across all maturities to move only in the same direction. A popular two-factor
model of Gibson and Schwartz (1990), which assumes a stochastic mean-reverting convenience
yield, generates a much richer dynamics for the futures curve and volatilities. Miltersen (2003)
allowed the equilibrium convenience yield to be time-dependent and showed how to make the
model consistent with the futures curve. Several three-factor models have been proposed, such
as Schwartz (1997), Casassus and Collin-Dufresne (2005), and Dempster et al. (2012). For
surveys of reduced-form models, we refer to Clewlow and Strickland (2000), Eydeland and
Wolyniec (2003), and Carmona and Ludkovski (2004).

7 In Bouchouev (2021), the author applied a slightly different methodology to this problem by
representing the futures spread as the solution to the Black-Scholes-Merton (BSM) partial
differential equation, where inventory is a state variable. Here, we use a somewhat simplified
approach, as BSM equation is formally introduced only later, in Chap. 8.

8 Some storage was already committed but not yet reflected in inventory data, nevertheless,
even including these additional volumes, the total inventory levels were substantially below the
maximum operating capacity.

9 For additional discussions of the episode of negative prices, see Interim Stuff Report (2020),
Bouchouev (2020), Fernandez-Perez et al. (2021), and Ma (2022).
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
I. Bouchouev, Virtual Barrels, Springer Texts in Business and Economics
https://doi.org/10.1007/978-3-031-36151-7_4

4. Financialization and the Theory of


Hedging Pressure
Ilia Bouchouev1
(1) Pentathlon Investments, LLC, Westfield, NJ, USA

The Keynesian theory of normal backwardation argues for structural


disequilibrium in the futures market caused by producer demand for
hedging. As a result, futures must trade at a discount to the expected
spot price to incentivize speculators to provide the service of risk
absorption.
The theory of hedging pressure is derived within a conventional
mean-variance framework of rational trading behavior. The
framework quantifies the hypothesis of normal backwardation and
relates it to the risk premium in the oil market.
The thesis of normal backwardation found strong empirical support
in the early days of oil trading. The existence of the structural oil risk
premium driven by the roll yield attracted financial investors and
spurred the financialization of the oil market.
Historically, jumps in oil prices preceded most of US economic
recessions. Oil also tends to appreciate during periods of high
inflation. To hedge inflation risks in highly leveraged risk parity
portfolios, large allocations are made to oil futures.
As investor demand for oil futures overpowered producers’ hedging
needs, the oil market shifted to a regime of normal contango. The
equilibrium hedging framework is generalized to include inflation
hedgers, which allows oil risk premium to be negative.
4.1 The Theory of Normal Backwardation
The question whether commodity futures markets are driven by
fundamentals or by financial flows is as old as the market itself. Any
attempt to answer this question unambiguously, however, is a fruitless
endeavor, as fundamentals and flows are closely intertwined. The
ideological underpinnings of fundamental analysts lie in the theory of
storage, which seeks an equilibrium between the price of the
commodity, its supply, demand, and inventories. Financial traders, on
the other hand, look for another equilibrium that arises directly in the
futures market. This financial equilibrium is driven by the aggregate
demand for hedging services among market participants and the
supply of such services by professional speculators. While the price for
a barrel of spot oil is driven by futures, an efficient futures market
cannot exist without a deep connection to the physical market. The
financial and physical markets are largely inseparable, with the crucial
link between them provided by carry traders.
The primary economic function of the futures market is to provide a
mechanism for transferring unwanted risks to those better equipped to
manage them. The price determined in the futures market should
reflect the supply and demand for hedging services, but it can also be
significantly impacted by speculative flows. In the oil market, it is
rather difficult to draw the line between hedging and speculation. Many
hedging programs incorporate subjective views of company executives,
at least when it comes to the timing of a trade and the choice of the
hedging instrument. Vice versa, some trades that may appear to be
speculative are often implemented within the broader mandate to
optimize a portfolio of physical or financial assets. The study of the
interaction between hedgers and speculators in the futures market is
known as the theory of hedging pressure.
Much like the theory of storage, the theory of hedging pressure has
developed its roots in agricultural markets. It starts with the works of
John Maynard Keynes in the 1920s.1 As an avid speculator in the
commodity markets, Keynes was clearly influenced by some painful
lessons from his own trading experience. This led him to recognize the
importance of psychological and behavioral factors in trading,
especially in the market for commodity futures, where hedgers and
speculators play a dominant role. These financial traders participate in
the market for reasons unrelated to prevailing fundamentals, but they
can still have a significant impact on price. Their behavior reflects the
willingness of some traders to pay for the services of risk absorption
and the remuneration of other traders for providing such services. If
supply and demand for the service of risk transfer are unbalanced, then
additional price incentives are needed to lure more speculators to the
market. The speculators are then compensated by trading futures
contracts at a discount or premium to their fair value, which is
determined by the supply of and demand for the physical commodity.
To transfer unwanted price risks, commodity producers and holders
of inventories sell futures, while consumers of commodities with
opposite price risks are expected to buy futures. In an ideal world,
consumers should be as eager as producers to get rid of the price risk,
in which case the two sides should be able to exchange futures in the
market at some fairly negotiated price. In practice, however, the
propensity to hedge between producers and consumers of commodities
tends to be asymmetric. The price risk typically has a more substantial
impact on the economics of a commodity producer, whose exposure is
less diversified than it is for a typical consumer.
Another important difference between producer and consumer
hedging interest lies in the timing of their business decisions. A
renowned economist and Nobel Prize winner, John Hicks, wrote that

… supplies in the near future are largely governed by decisions


taken in the past, so that if these planned supplies can be
covered by forward sales, risk is reduced. But … with planned
purchases … technical conditions give the entrepreneur a much
freer hand about the acquisition of inputs … than about the
completion of outputs. Thus, while there is likely to be some
desire to hedge planned purchases, it tends to be less insistent
than the desire to hedge planned sales. If forward markets
consisted entirely of hedgers, there would always be a tendency
for a relative weakness on the demand side; a smaller
proportion of planned purchases than of planned sales would be
covered by forward contracts.2
Given the structural imbalance between producer and consumer
demand for hedging, producers have little choice but to offer futures at
a discount to the fair value in order to attract professional speculators
to take the other side of their hedging needs.
Let us assume that the futures price is deemed to be fair if it is
determined by market collective expectations of the spot price at the
time of the futures delivery. In other words, if the market is fair, then
one cannot consistently make or lose money by trading futures. The
Keynes-Hicks hypothesis argues that the futures price should not be
fair, as there are more natural sellers of commodity futures than buyers.
Therefore, in the natural equilibrium state of the market, the futures
price must fall below the expected spot price by an amount equal to
normal backwardation,3 defined as:

In the modern jargon of financial markets, the Keynesian term of


normal backwardation is more commonly known as risk premium (RP).
The risk premium provides an incentive to speculators to participate in
an otherwise unbalanced market for hedging services. Anyone willing
to withstand short-term market fluctuations should be able to buy
discounted futures and get rewarded for offering the service of risk
absorption. The reward comes in the form of a positive expected return,
provided that the speculator has enough financial power to stay in the
game.
The concept of normal backwardation should be distinguished from
the conventional definition of market backwardation. The latter does
not involve expectations; it only refers to spot and futures prices
observed at the same time. To see how the two concepts are related, we
recall the decomposition of futures profitability (2.​11), and split the
risk premium in a similar way, as follows
(4.1)
The first term reflects the expected change in the spot price, which, of
course, is not known today. The second term, on the other hand, is
directly observable from the current shape of the futures curve. This
term is positive when the market is in backwardation, and it is negative
when the market is in contango. As explained in Chap. 2, it is the same
term that drives the oil own rate of interest, the convenience yield, net
of storage costs, and the roll yield.
Keynes argued that normal backwardation, which is equivalent to a
positive risk premium, exists regardless of whether the market is in the
state of backwardation or contango. His thesis is illustrated in Figs. 4.1
and 4.2. The blue line is the observable futures curve. The dashed
orange line represents the expected forward trajectory of the spot price
moving from left to right, i.e., from t to T. The difference between the
two lines is the measure of normal backwardation, or the expected
profit to the buyer of futures.

Fig. 4.1 Normal backwardation, or the risk premium, in a backwardated market


Fig. 4.2 Normal backwardation, or the risk premium, in a contango market
In both examples, the risk premium is positive, and the two terms in
the decomposition (4.1) have opposite signs. In the first scenario of
market backwardation, shown in Fig. 4.1, a long futures position is
expected to make money because the magnitude of the market
backwardation exceeds the expected depreciation of the spot price.
While the spot price is expected to fall, it falls only by a fraction of what
is priced in the futures curve. In the second scenario of market
contango, shown in Fig. 4.2, the rise in the spot price is expected to
exceed the amount of contango priced in the futures curve. Here, the
market contango only partially offsets a more substantial appreciation
of the spot price. In both scenarios, the futures price is suppressed by
producer demand for hedging services, which results in an expected
profit to the buyer.
The decomposition of the risk premium in (4.1) corresponds to two
potential sources of revenue for the commodity speculator. The first
one is highly uncertain as it comes from speculating on the volatile spot
price of a commodity. The second term does not require any
forecasting. It simply reflects the discount or premium, which is
measured by the current shape of the futures curve. By and large, the
two terms correspond to two types of speculators, eloquently labeled
by Keynes as prophets and risk-bearers. Being a prophet requires a
special talent for predicting future spot prices, something that Keynes
himself was highly skeptical about, as

… it presumes that the speculator is better informed on the


average than the producers and the consumers themselves,
which, speaking generally, is rather a dubious proposition.4
In contrast, the second type of speculator, who does not pretend to
possess any unique trading insights, is not involved in the game of price
forecasting, as

… he is not so much a prophet (though it may be a belief in his


own gifts of prophecy that tempts him into the business), as a
risk-bearer.5

The business model of the risk bearer is very different from the
business of price forecasting. Risk bearing resembles the business of a
casino dealer. The primary requirement for the risk bearer is an ability
to provide capital and withstand short-term losses resulting from
random fluctuations in spot prices. Since the spot price is likely to
mean-revert, as implied by the theory of storage, cumulative spot price
changes in the first term of (4.1) are expected to net out. In the
meantime, the contribution of the second term steadily accumulates as
long as the market remains backwardated. This is the reason why
casinos tend to win; they stay in the game and repeat the same bet with
a tiny statistical edge over and over. For them, wins and losses of
individual gamblers, or aspiring prophets, only represent noise that
eventually cancels out. By the law of large numbers, the casino’s
statistical edge turns into profits over time. For risk bearers in
commodity markets, this casino-like edge comes from the
backwardated shape of the futures curve.
We next quantify the Keynesian hypothesis of normal
backwardation and expected profits for risk absorbers with a simple
analytical framework based on the rational behavior of market
participants.
4.2 The Hedging Equilibrium
Let us consider a simplified model of the futures market which is made
up of three participants: a producer, a consumer, and a speculator. We
assume that they follow a conventional one-period mean-variance
decision-making framework, where participants are maximizing their
expected wealth, , while being penalized for risks measured by

the variance of wealth, . In other words, all participants


maximize their quadratic utility function of the form

(4.2)

where α represents the risk aversion coefficient.6


We denote wealth functions and risk aversion
coefficients αp, αc, αs with corresponding subscripts for producer,
consumer, and speculator, and use tilde to represent random variables
whose values are unknown at the beginning of the period.
Assume that the producer is endowed with Qp barrels of oil and,
thus, is exposed to potential losses if the spot oil price falls. The
producer also participates in the futures market and trades Np units of
futures that expire at the end of the period. Since the producer is long
physical barrels of oil, one should expect the producer to be net short
futures, so that Np < 0. We do not differentiate between the producer of
oil and the owner of oil inventory, as both have similar exposure to oil
prices. The main practical difference in their market participation is in
the duration of their typical hedges. Producers usually seek a longer-
term price hedge against the value of their expected production and
reserves, while inventory hedgers trade only a few months forward to
protect the value of oil that has already been taken out of the ground. In
our simplified one-period framework, their exposure is largely
identical.7
The producer’s wealth is determined by the dollar value of the oil
endowment, plus the profit or loss on Np futures traded at price F, less
some fixed operational cost Up:
The only uncertain variable here is the spot oil price . The expected
value of the wealth function is then given by

and the variance of wealth is proportional to the variance of the oil


price

The producer wealth maximization problem is to determine the


optimal position in the futures market Np that maximizes the quadratic
utility (4.2):

We apply the standard first-order optimality condition, differentiate


this expression with respect to Np, and equate the derivative to zero

Solving this equation for Np, we obtain that the producer optimal
position in the futures market is given by

(4.3)

As anticipated, the producer is likely to be short futures to hedge the


value of the oil endowment against the risk of falling prices. If the
producer is highly risk averse, so that αp → ∞, then the optimal hedging
decision is to cover the entire value of Qp. However, a less risk averse
producer may reduce the size of the short futures position if there is a
substantial positive risk premium in the market and the hedge is
expected to lose money. This adjustment depends on the market
volatility and the individual producer’s risk tolerance. The optimal
hedging decision, therefore, includes a speculative component.
We repeat the same argument for the consumer, but while doing so,
we also illustrate how to handle the basis risk in hedging. Since oil can
only be consumed in its processed form, the end-user is usually
exposed to the price of the refined consumer product , and not to the
price of crude oil . We assume that is correlated to the price of oil
with the correlation coefficient . The consumer,
whose risk aversion coefficient is αc, needs to purchase Qc units of the
refined product at price , which is used as an input into a
manufacturing process that generates fixed revenue Uc. The consumer
trades Nc futures contracts which are expected to be long positions, i.e.,
Nc > 0.
The consumer net wealth is then defined as

Here, we have two random variables, and . The variance of the


sum of two random variables is equal to the sum of their variances, plus
the covariance between them.
To maximize the expected utility of wealth, we substitute

and into (4.2)

To find the optimal futures hedge, we differentiate with respect to Nc


and equate the derivative to zero:

Solving for Nc, we obtain that


(4.4)

Here,

is the standard beta sensitivity of the asset with respect to the


benchmark index, commonly used in the equity market. The consumer
optimal hedge is based on the beta-weighted exposure of the refined
product to crude oil, further adjusted for a speculative view. The beta
hedge ratio increases with higher correlation between the refined
product and crude oil, while it decreases when crude oil volatility rises.
If necessary, a similar beta adjustment can be applied to the producer
hedging problem if the type of oil being produced is different from the
benchmark used for hedging. In practice, however, the correlation
among various grades of crude oil is much higher than the correlation
between crude oil and refined products, so, for simplicity, we let the
producer beta to the benchmark be equal to one.
Finally, the same logic is applied to the speculator with no
endowment, liability, or any other hedging mandate. The only motive of
the speculator is to make money by trading futures and to capture the
risk premium caused by the hedging imbalance. Repeating the previous
steps, we obtain that the optimal position of the speculator is simply
dictated by the market risk premium, adjusted for volatility and the
speculator risk aversion

(4.5)

We can now construct the financial equilibrium in the futures market.


The futures market is a zero-sum game, so if the producer, the
consumer, and the speculator are the only three participants in the
market, then for futures to clear, their net position must be zero

Adding up Eqs. (4.3), (4.4) and (4.5) for optimal futures positions held
by producer, consumer and speculator, we obtain that

This equation can be solved for the risk premium, as follows

(4.6)

where the coefficient α is given by

The Eq. (4.6) explicitly relates the Keynesian risk premium to the
aggregate hedging imbalance between producers and consumers. Given
that the producer has a much more concentrated price exposure than
the consumer, the producer propensity to hedge is likely to be higher,
and Qp ≫ Qc. This implies that in this simple market model, the
resulting oil risk premium in the Eq. (4.6) is indeed expected to be
positive, as is suggested by the theory of normal backwardation.
The size of the risk premium depends on the volatility of the market
and the risk aversion of all market participants. The risk tolerance of
speculators cannot change the sign of the risk premium, but the
speculators’ lower risk aversion can dilute the magnitude of the
imbalance. The risk in this hedging equilibrium model is shared among
all market participants. The parameter α can be thought of as the
market collective risk aversion coefficient. It is proportional to the
harmonic mean of the individual risk aversions. Remarkably, the idea of
using a harmonic mean in risk sharing goes back to Ancient Greece and
Aristotle, who considered the price of voluntarily exchanged goods to
be fair if it is calculated as the harmonic mean of the buyer’s bid and the
seller’s offer.8
To validate the existence of the structural risk premium in the oil
market, we now test the hypothesis of normal backwardation
empirically.

4.3 The Genesis of Oil Financialization


The search for a Keynesian free lunch for risk absorbers in commodity
futures got off to a rough start in agricultural markets. Numerous
attempts have been made to trace any long-term structural bias, but
little empirical evidence was found to support the hypothesis of normal
backwardation. In fact, the term structure of many agricultural futures
was more often in contango than it was in backwardation. It turns out
that the treasure hunt was going on in the wrong place. Only after the
introduction of oil futures in the 1980s did the thesis of normal
backwardation get its second wind.
The structure of petroleum markets appears to fit the Keynes-Hicks
consumer-producer hedging paradigm quite well. Investments in oil
exploration and production are extremely capital intensive and are
accompanied by large price uncertainty with a multi-year payback time.
These decisions are typically irreversible. Once an investment decision
is made, hedging the value of the expected output becomes the only
practical way to mitigate the impact of price fluctuations. Things work
differently for many oil consumers. The primary usage of petroleum
products is in transportation, where the demand mostly comes from
individuals buying gasoline and diesel to drive their cars. While
collectively, we all share a price risk of a similar magnitude to oil
producers, this risk is spread out across millions of us. The fuel cost
represents a relatively small portion of our routine expenses, and very
few of us would seriously contemplate the idea of hedging personal
gasoline consumption.9 The much higher concentration of the energy
price risk makes hedging more important for producers than for
consumers.
One notable exception of an active consumer hedger is the airline
industry. For an airline, the purchases of jet fuel represent the second
largest cost after the cost of labor. Since jet fuel, which is tied to the
price of oil, is much more volatile than salaries, higher oil prices could
easily dwarf the slim operational profit margins of the industry. While
to a certain degree an airline might be able to pass on some fuel costs to
travelers by raising ticket prices, such pass-throughs are often limited
by competitive pressures. Overall, however, airline hedging is
substantially smaller than producer hedging. The first part of the
Keynes-Hicks hypothesis, which argues for the structural imbalance
between producers’ and consumers’ propensity to hedge, clearly
applies to the oil market. The second part, which claims the existence of
a structural risk premium that rewards the buyers of futures, is, of
course, a more interesting question.
In contrast to agricultural markets dominated by contango, in the
early days of oil trading the futures curve exhibited a more persistent
backwardation. To test the Keynesian hypothesis of a free lunch for
futures buyers, we look at the passive long strategy that buys prompt
WTI futures contract and rolls it at the end of each calendar month to
the second nearby contract. Figure 4.3 presents the performance of this
strategy since the inception of the WTI futures contract until 2004. It
shows the picture as it was seen by prospective oil investors at that
time.
Fig. 4.3 Cumulative performance of fully funded long WTI futures during the regime of normal
backwardation
Driven by the preponderance of backwardation in the oil market,
the strategy of buying futures at a discount relative to spot prices was
indeed spectacularly profitable. During this period, the passive oil
investor would have realized over 10% annualized logarithmic return
on an unleveraged, or fully funded, investment in oil futures.10 Such
returns were comparable, if not superior, to long-term investment
returns in the stock market. Moreover, since in practice only a small
portion of the notional value of the futures contract is required to be
paid as the initial margin, the returns on the actual cash deployed in
highly leveraged oil futures contracts were five to ten times higher.11
The source of such impressive returns turned out to be even more
striking. Figure 4.3 also shows the decomposition of the excess return
into the spot return and the roll return based on the identity (2.​12). As
one can see, nearly entire profitability of the strategy came from the roll
return, as cumulatively WTI spot price did not change by much over
this period. While the spot price fluctuated, which is what one should
expect from the supply and demand induced price mean-reversion, the
positive roll return steadily accumulated as the oil market was
predominantly backwardated.
The superb performance of an investment in oil futures has finally
vindicated the Keynesian theory of normal backwardation. The profits
were driven not by a highly questionable ability to forecast the
direction of the spot price, but rather by providing liquidity and taking
the other side of the dominant producer hedging at advantaged prices.
The strategy of holding oil futures looked very compelling. It required
no unique skills, just enough money to withstand occasional losses. Its
appeal for financial investors has become so powerful that it marked
the beginning of a new era in the history of oil trading, the era of oil
financialization.
The opportunity for a free lunch promised by the theory of normal
backwardation became evident. The lunch was not entirely free though.
It was a compensation for committing large amount of capital for a long
period of time, which was needed to stay in this game of risk
absorption. To bring such capital to the market for oil futures, the
industry had to overcome significant challenges. A typical oil speculator
at that time neither had much capital, nor any desire to lock it up for
more than a few months ahead. A better candidate for the role of a risk
bearer was a longer-term investor with large financial coffers, like a
pension fund or an insurance company.
Bringing pension funds to trade oil was not straightforward since
many institutional investors were explicitly prohibited from trading
futures. Moreover, oil, along with many other commodity futures,
suffered from an image problem. For a long time, trading commodity
futures has been perceived by the public as a speculative, borderline
gambling enterprise, which made it difficult for large strictly regulated
asset managers to participate in this market. Traditional investments,
like stocks and bonds, serve a clear investment purpose; they provide
capital that allows businesses to operate and grow their enterprises. In
contrast, futures were viewed as highly leveraged speculative bets that
do not even require as much capital to be put upfront. Oil futures did
not appear to be suitable for deep-pocketed conservative financial
investors who were looking for steady long-term returns. The
breakthrough for turning commodity futures into a new asset class
came from the clever packaging of futures contracts into investable
indices by the banks.
Commodity indices were designed to function as a bridge that
connects the capital of long-term financial investors to the Keynesian
opportunity in highly speculative commodity futures markets. The idea
was for the banks to step in between and take care of futures trading
themselves. The resulting economics of the futures trade can then be
transferred to investors via an equity-like product, whose returns are
determined by profit and loss on futures. Since only a small portion of
the notional value of the futures contract is required as collateral, the
remaining cash can be invested in safe securities, such as US Treasury
bills. Not only does this cash provide an additional return in the form of
interest, but it also eliminates unwanted leverage, as the entire notional
value of commodity futures is effectively prepaid. Such fully
collateralized commodity indices look a lot closer to traditional
financial and capital investments like stocks and bonds.
To address concerns about the role of investor capital in commodity
investment, the Keynesian thesis has once again been recruited to help.
While in equity and bond markets capital provides explicit funding of
business operations, commodity index providers argued that in their
products it does so implicitly. The capital invested in commodity indices
can also be perceived as productive if it lets producers free up their
own capital. Otherwise, producers’ own capital would be tied up
unproductively and held as a buffer against price fluctuations.
Such an interpretation would allow producers to view oil hedges as
an integral part of their new capital structure along with equities and
bonds. The ability to hedge reduces volatility of producer earnings and
lowers borrowing costs, thus allowing more capital to be redeployed
towards expanding the primary business of producing oil. Producer
hedging and the credit market become intertwined. While hedging was
explicitly mandated by credit departments of lending banks, the deals
were also often executed by the derivatives desks of the same banks,
allowing them to capture profits on both sides. In contrast to many
tightly regulated financial markets, there were no restrictions on such
an intimate arrangement between oil and credit markets, which is one
of the reason why oil OTC market grew to become so large.
From the investor’s perspective, the capital allocated to commodity
indices is rewarded for risk bearing, much like the capital invested in
insurance products. The reward is expected to accrue over time from
the passive accumulation of normal backwardation, which does not
require any price forecasting. The opportunity to generate long-term
casino-like returns on capital via investments in commodity indices
without any need to develop unique commodity expertise quickly
became a home run with pension funds and other long-term financial
investors.
The commodity index business was designed for a diversified
basket of commodities but crude oil and other petroleum products
were its crown jewels. The Goldman Sachs Commodity Index (GSCI),
the largest and the most successful commodity index at that time,
epitomizes it well. In this index, seventy percent of its assets was
allocated to the energy sector. GSCI weights for individual commodities
are calculated based on the total value of commodity annual
production, where petroleum dominates by a wide margin. It leaves
little doubt that such a choice for the index construction was influenced
by the superb historical performance of energy futures, which
happened to be more often in backwardation than in contango. Unlike
many other commodities, oil had a high convenience yield that
dominated the cost of storage, resulting in a positive own rate of
interest. This intrinsic interest, or the roll yield, was what financial
investors were after.12
While energy futures were primary contributors to the index
returns, other commodities provided valuable diversification.
Individual commodities tend to have low correlations with each other
given their idiosyncratic risks, such as, for example, their sensitivity to
weather. Commodity prices have vastly different seasonal production
and consumption profiles. Even within the energy sector, gasoline has
higher upside risks during the summer driving season, while heating oil
and natural gas are more susceptible to price spikes during the winter.
An investment in a broader commodity index isolates the contribution
from the roll, and it dilutes the impact of random fluctuations in spot
prices. While backwardated commodities, such as oil, are expected to
drive the index return, other commodities help to dampen the volatility
of unpredictable spot returns. In addition, marketing the diversified
index for the commodities, as a new asset class, is much easier than
marketing oil futures, which would have been seen as too speculative.
One other substantial benefit that results from an investment in a
diversified commodity basket is the so-called rebalancing effect. Some
commodity indices target allocations in dollar terms that must be
periodically rebalanced as commodity prices change. Since commodity
spot prices tend to mean-revert over time, the rebalanced index would
buy more futures of relatively depreciated commodities and sell the
appreciated ones to keep their dollar notional contract value the same.
While GSCI does not rebalance frequently, with rebalancing impact
muted by its high concentration of energy futures, many other sector-
diversified commodity indices are able to capture additional returns
from this rebalancing effect by buying lows and selling highs.
The Keynesian normal backwardation revived by the energy market
and adopted by long-term passive investors was only the beginning of
oil financialization. The pace of financialization accelerated as another
even more powerful financial force, fueled by leverage, was about to
enter the oil market.
4.4 Inflation Hedging and Risk Parity
The introduction of commodity indices turned commodities into a new
asset class. The precise definition of what constitutes an asset class is
somewhat murky and highly debatable. In a very broad sense, one can
define an asset class as any type of investment that embeds a structural
risk premium which cannot be replicated by a portfolio of traditional
assets. An investment in commodity indices appeared to carry such a
risk premium in the form of normal backwardation, predominantly
driven by energy futures. Among all commodities, oil stood out not only
as the most valuable commodity in terms of its global production value
but also as the one that has the largest futures market. Since oil also
happened to have higher historical returns with low and often even
negative correlation to other financial asset classes, oil by itself was
often perceived to be a convenient proxy for commodities asset class.
Cross-asset correlations are crucial inputs into conventional asset
allocation frameworks, many of which are based on the Capital Asset
Pricing Model (CAPM). According to the CAPM, the expected returns on
all assets are proportional to the asset’s contribution to non-
diversifiable risk in the overall market portfolio that includes all assets.
The optimal asset allocations are then constructed based on assets’
expected returns, volatilities, and correlations, which are typically
estimated from historical data. Assets with lower correlation to the rest
of the portfolio warrant higher allocations. Based on the data available
in the early 2000s, the CAPM-optimal investment portfolio could have
justified up to 15–25% allocation of all assets to the petroleum-heavy
GSCI index.
Unfortunately, the concept of linear correlation that underpins
CAPM-based asset allocation methodologies is not well suited for
analyzing highly nonlinear systems such as the oil market. Any linear
measure of dependency is inevitably prone to instability. Correlations
computed using historical time series are extremely sensitive to
outliers and to the selection of the lookback period, which is often
cherry-picked by analysts to improve the optics of the desired outcome.
The reliance on correlation estimates in the presence of outliers does
not let one see the forest for the trees. Instead, one is better off focusing
on specific events that might have caused these outliers. As we have
already seen in the study of storage constraints in Chap. 3, the outliers
tend to reflect the presence of natural boundaries in the nonlinear
system. For the overall economy and financial markets, one such
boundary is an economic recession.13 It turns out that the behavior of
oil prices near this boundary is what gave oil its prominent role in
investment portfolios.
It has been long observed that most of US postwar economic
recessions were preceded by significant increases in oil prices which
became known as oil shocks.14 Prior to the 1970s, price increases were
fairly modest, as during those years oil prices were largely regulated.
The first large oil shock occurred in 1973 when the price of oil tripled
after an oil embargo was imposed by the Organization of Arab
Petroleum Exporting Countries (OAPEC). The second energy crisis
came in 1979 when oil price doubled in the aftermath of the Iranian
Revolution followed by the Iran-Iraq War. In both episodes oil was
blamed for severe economic recessions. The next oil spike of 1990 was
initially driven by the Persian Gulf War. While it happened shortly after
the US recession has already been announced, it also contributed to the
economic slowdown. Figure 4.4 illustrates.
Fig. 4.4 Oil price appreciation by at least 50% either preceded or coincided with most US
recessions since the 1970s
These episodes of large disruptions in oil supplies with severe
consequences secured a very special place for oil among key
macroeconomic variables. Furthermore, the pattern continued to hold
with oil prices appreciated again by more than 50% versus its prior
three-year average just before the recession of 2001 and then again in
the beginning of the Global Financial Crisis. One exception from this
statistical anomaly is a brief recessionary period at the onset of the
Covid-19 pandemic when oil prices also weakened.
Does it mean that rising oil prices cause economic recessions? This
question has been the subject of heated debates for nearly forty years.
Undoubtedly, the answer depends on whether higher oil prices were
driven by supply disruptions, unexpectedly strong demand, or other
factors, such as financial speculation.15 Our preference is to stay away
from this debate as the data is just too noisy and results are too
sensitive to the sample selection and to the choice of the econometric
methodology. Regardless of whether this pattern of oil being a
precursor of economic recessions is just a statistical fluke or created by
some invisible market force, its mere existence made oil very special in
the eyes of financial investors. An investment in oil provided
diversification to financial portfolios precisely when it was most
needed.
While recessions tend to be bad for the economy, periods of high
inflation are even more challenging for financial investors. The prices of
all financial assets are generally determined by the present value of
their future cash flows. If inflation unexpectedly rises, causing higher
interest rates, then discounting decreases the present value of all future
cash flows. For bonds, the negative impact is straightforward as future
cash flows are fixed. For equities, the value of real assets owned by
companies generally also rises with inflation, but it is offset by higher
costs of labor and capital. When inflation moves up quickly and
unexpectedly, input costs tend to increase at a faster pace than
revenues. In general, rising inflation hurts both stocks and bonds, and
oil futures happen to be one among very few investments that tends to
be profitable during periods of high inflation.
US inflation is typically measured by changes in the Consumer Price
Index (CPI), which represents the price for a basket of goods and
services paid by a typical consumer. Figure 4.5 shows how closely US
annual headline inflation tracks the annual change in the price of WTI.
The only visible dislocation between two variables occurred during the
recovery period after the Covid-19 pandemic when inflation lagged the
change in oil prices.

Fig. 4.5 US inflation measured by year-over-year changes in CPI is largely explained by annual
changes in the price of oil

This relationship is simply too good to be purely statistical. In fact, it


is much more mechanical and driven by the algebra of oil pass-through
into the consumer cost basket. The price of oil is the primary driver of
retail gasoline prices, which make up the majority of the motor fuel
component of the CPI. Even though this sub-component represents less
than 4% of the CPI basket, it explains the large portion of the CPI
monthly variance simply because other components of the CPI move
much more slowly. We will return to this topic in Chap. 7 at greater
length, where we develop a relative-value trading strategy between
energy futures and inflation swaps which is centered around the pass-
through of futures into the CPI.
Besides direct mechanical impact on the CPI via retail gasoline
prices, oil also affects inflation indirectly. Higher oil price raises the cost
of feedstocks for many manufacturing processes, often leading to
higher prices of many consumer goods. In addition, rising prices for
consumer goods may also force workers to demand higher wages. This
could lead to a potentially dangerous inflationary spiral, which is
understandably a major concern of policymakers.
Since the indirect impact of oil on future inflation is difficult to
observe quickly, it has been long suggested that consumers often use
highly visible and frequently changing gasoline prices to form their
expectations about future inflation. Observed trends in gasoline prices
are often extrapolated by consumers as trends in broader inflationary
pressures. Measuring consumer expectations via surveys, however, is a
rather daunting task, so it is not surprising that analysts and
policymakers switched to an alternative market-based measure of
inflation expectations.
To gain exposure to US inflation in the market, one buys inflation-
linked bonds, called Treasury Inflation-Protected Securities (TIPS),
which pay the real yield, and simultaneously sells nominal Treasury
bonds with the same maturity. The difference between the nominal
yield of the Treasury bond and the real yield paid by TIPS is known as
the inflation breakeven rate. Five-year and ten-year breakevens are two
primary inflation benchmarks in the marketplace. They are often
combined to form the so-called 5y5y forward breakeven rate, which
represents five-year inflation measured five years forward. One can
approximate the 5y5y forward inflation rate as twice the ten-year
inflation rate minus the five-year inflation rate. This metric became the
de facto standard used by policymakers in estimating market-based
measures of inflation expectations.
Figure 4.6 shows a rather surprising picture of how closely the 5y5y
breakeven rate follows the price of WTI futures. Even though,
theoretically, the oil price today should not have anything to do with
such long-dated inflation expectations, the 5y5y inflation breakeven
rate tracks the price of WTI quite well.
Fig. 4.6 5y5y inflation breakeven closely tracks the price of WTI
This co-movement is very bizarre from the statistical perspective. It
shows that the level of the commodity co-moves with the rate of change
in the consumer basket. The statistical relationship is usually properly
measured only when both time series are stationary, and for most
financial variables the price levels are not stationary, but price changes
or returns are.16 On the one hand, this relationship can be viewed
purely as a coincidence and discarded. On the other hand, as we will see
in Chap. 7, the connection between the two markets may indeed be
reinforced by cross-asset arbitrageurs. While its merit is still subject to
debate, its optics clearly contributes to the belief in the oil’s special role
as a valuable hedge against the risk of inflation.
Furthermore, it turns out that oil is also somewhat unique in its
ability to hedge against inflation surprises, or unexpected shifts in
inflation. Obviously, inflation surprises are hard to measure statistically.
Measuring them by comparing realized inflation against prior surveys
of forecasts by economists has proven to be highly unreliable. A more
objective measure of historical inflation surprises can be given by the
change in the rate of CPI. Using such a definition, one can then analyze
statistically the performance of different assets during such unexpected
changes in inflation. Contrary to common beliefs, oil performs much
better than gold, real estate, or virtually any other hedging alternative
in countering jumps in unexpected inflation, and this is what pension
funds that own a large portfolio of stocks and bonds care about.17
While the existence of normal backwardation and commodity
indices brought large institutional investors to the oil market, the next
wave of oil financialization was spurred by growing interest in hedging
the risk of inflation. A large role in this growth is attributed to the
popular investment strategy dubbed risk parity. The earliest versions of
the risk parity framework are usually credited to the so-called “All
Weather” strategy launched by Ray Dalio at a then relatively unknown
hedge fund called Bridgewater Associates. Subsequently, it became one
of the most profitable hedge funds in the world, and somewhat
surprisingly its risk parity strategy turned the fund into one of the
world’s largest oil traders.
The “All Weather” strategy was designed as an alternative to CAPM-
based asset allocation. It looks at returns on all assets through shifts in
two primary driving factors, growth and inflation. This rationale is
intuitive since all expected future cash flows are driven by the volume
of economic activity, which is growth, and how this activity is
discounted today relative to the future. The latter is largely based on
future inflation expectations. The standard portfolio of stocks and
bonds is only diversified with respect to the growth factor. If the
economy is strong, then equity markets should perform well, as stocks
represent claims on future company earnings, which are expected to
increase when the economy is booming. On the other hand, bonds tend
to do better when the economy is weaker, as bonds provide investors
with fixed cash flows whose present value increases when interest
rates drop. While a conventional portfolio of stocks and bonds provides
investors with some balance with respect to shifts in the economic
regime, it is not diversified with respect to shifts in inflation
expectations. To hedge this risk, commodities and, in particular, oil with
its optically impressive inflation-hedging properties, must be added to
the portfolio.
The crucial part of the risk parity methodology is in how assets are
weighted. Since equities are more volatile than bonds, the traditional
60–40 equity-bond portfolio translates into an approximately 90–10
allocation in risk terms, which provides little real diversification.
Instead, in the risk parity allocation framework, all assets are weighted
equally based on their risks. Lower-volatility assets, such as bonds,
must then be purchased in larger quantities to contribute the same
amount of risk as higher-volatility assets. This forces risk parity
managers to heavily invest in bonds and leveraged fixed income
derivatives. Since inflation is the major risk for bonds, commodities
must also be purchased in large quantities, which in practice can only
be done in the futures market. While banks brought commodities to
unlevered investors via fully funded indices, risk parity went a step
further and offered the coveted exposure to commodities on a highly
leveraged basis. Among all commodities, oil with its superior inflation-
hedging properties and the best liquidity across all commodity futures
received the largest slice in the risk parity allocation pie.
Bridgewater’s original idea was adopted by many other providers of
risk parity investment strategies. While these strategies vary in
implementation, the core premise of allocating larger notional
investments to lower-volatility bonds requires fund managers to also
buy large amounts of inflation sensitive commodity futures, among
which oil stands out. The arrival of risk parity funds not only restored
the Keynesian imbalance between sellers and buyers in the futures
market, but it also started to shift the disequilibrium in the opposite
direction. If positive expected returns from risk bearing alone were not
sufficient to induce buying of oil futures, the additional benefit of
hedging the risk of inflation became the final straw that broke the
camel’s back.

4.5 Inconvenience Yield, or the Theory of


Normal Contango
Keynesian risk bearers came to the oil futures market with additional
capital to bring it back to equilibrium. They did their job, in fact they
did even more than what was needed. They did not stop when the
hedging equilibrium was restored, as diversification benefits of oil
futures brought further benefits to their broader financial portfolios.
They continued to buy oil futures even as their own activity started to
push futures prices above the expected spot price. As a result, the oil
market shifted into a new structural regime of normal contango.
Extending the concept of normal backwardation, we define normal
contango when the expected spot price falls below the current futures.
This is not the same as market contango. The existence of normal
contango implies that investors are expected to lose money by buying
and rolling oil futures. These expected losses have indeed materialized.
The sign of the risk premium has flipped from positive to negative.18
Figure 4.7 extends the graph presented in Fig. 4.3. The strategy
performances over two periods are nearly mirror images of each other.
Starting from 2005 the strategy of buying and rolling oil futures began
to steadily lose money. By the end of 2018, it lost nearly all of its entire
gain accumulated during the prior period of normal backwardation, as
the normal state of the market was contango. Remarkably, the spot
price at the beginning and at the end of this entire period was basically
the same. In other words, the entire loss came again from the
accumulation of the roll yield which during that period was
consistently negative.

Fig. 4.7 Cumulative performance of long WTI futures during the regime of normal contango
The futures market is a zero-sum game. For every loser, there must
be a winner. Where did the negative risk premium go to? The answer
has already been given in the previous chapter. Selling futures at a
premium to financial investors and buying physical barrels is the
business of the professional storage trader. From being a provider of a
service to oil producers, investors themselves turned into hedgers of
financial risks. Investors pay for this service with the risk premium by
rolling futures in the contango market with profits accruing to the
storage owner. The investor looking to buy futures must incentivize
new marginal suppliers of short futures to come to the market, if
existing short futures offered by producer hedgers are not sufficient.
These incentives have also provided the impetus for building new
storage capacity, which we discuss later in the context of specific
trading strategies in Chaps. 6 and 13.
Since investors cannot easily buy and store physical barrels
themselves, they have no choice but to pay for the privilege to invest in
oil. These payments are not explicit. Every month when the futures
curve is contango, investors must sell prompt futures before their
expiration, and reset the long position at a higher price by buying the
next available futures contract. Even though no money is exchanged on
the day of the roll, losses are accumulated over time as futures tend to
roll down towards the lower spot price. The convenience yield of
owning a physical commodity, net of storage costs, has turned negative.
The intrinsic yield of oil became much more of an inconvenience. While
the negative roll yield acted as the synthetic cost of storage, it also
compensated storage owners for providing the service of essentially
storing the bulky commodity on behalf of investors.
With the addition of another major market participant, the manager
of the broader financial portfolio, the hedging equilibrium Eq. (4.6)
must be revisited. An investor, such as a risk parity find, now has an
inflation-hedging mandate that counters the producer’s need to sell
futures. Let us assume that such an investor is tasked to manage the
notional inflation exposure Qi in a financial portfolio that generates
income Ui while being exposed to inflation uncertainty . The investor
follows the same mean-variance framework with risk aversion
coefficient αi, and trades Ni oil futures to hedge the inflation risk. The
inflation hedger’s wealth is then given by19

Following the same steps as before, we find the futures position Ni that
maximizes the following expression for the investor wealth function

Like in the case of a consumer cross-hedging refined products exposure


with crude oil, the variance of the sum of two random variables
contains the covariance term. Here, represents the
correlation coefficient between inflation and oil.
As before, differentiating with respect to Ni and equating the partial
derivative to zero results in

The optimal number of futures held by the inflation manager is then


given by

(4.7)

where

represents inflation beta to oil. This is analogous to consumer hedging


of refined product exposure by trading crude oil futures, as in (4.4).
The futures now must clear among the producer, the consumer, the
inflation hedger, and the traditional speculator:
Adding up Eqs. (4.3), (4.4), (4.5) and (4.7) for optimal hedging positions
held by each participant and solving for the risk premium, we obtain
that

(4.8)

where the risk is shared among all futures traders, as

The theory of hedging pressure is now generalized to cover all the main
market participants. The resulting risk premium is a function of the net
imbalance between producers, consumers, and inflation hedgers.
Traditional speculators are incentivized to take either long or short
positions depending on the sign of the net hedging imbalance. Since
financial markets are significantly larger than oil markets and the
demand for inflation hedging has steadily increased over time, the
structural risk premium during the era of financialization has become
negative. In this case, systematically buying and rolling oil futures is a
losing value proposition.
All good things in financial markets eventually come to an end.
Storage operators and short speculators enjoyed a long run of
spectacular performance during the regime of normal contango.
Financial investors, in turn, have learned their lessons and became
more dynamic with their oil investment strategies. By and large, the oil
markets have largely matured, and the structural directional oil risk
premium has vanished. It became nimbler, responding to faster changes
in fundamentals of storage and hedging imbalances. To capture it,
investment strategies must become dynamic as well. These strategies
employed by different types of investors and speculators are presented
in the following three chapters of the book.

References
Acharya, V. V., Lochstoer, L. A., & Ramadorai, T. (2013). Limits to arbitrage and hedging:
Evidence from commodity markets. Journal of Financial Economics, 109(2), 441–465.
[Crossref]

Alexander, C. (2001). Market models. Wiley.

Ashton, M., & Greer, R. (2008). History of commodities as the original real return asset class. In
Inflation risk and products (pp. 85–109). Risk Books.

Backhouse, R. E. (2002). The ordinary business of life. Princeton University Press.

Baker, S. D. (2021). The financialization of storable commodities. Management Science, 67(1),


471–499.
[Crossref]

Bodie, Z., & Rosansky, V. I. (1980, May–June). Risk and return in commodity futures. Financial
Analysts Journal, 36(3), 27–39.

Bouchouev, I. (2012). Inconvenience yield, or the theory of normal contango. Quantitative


Finance, 12(12), 1773–1777.
[Crossref]

Bouchouev, I. (2020). From risk bearing to propheteering. Quantitative Finance, 20(6), 887–894.
[MathSciNet][Crossref]

Bü yü kşahin, B., & Robe, M. A. (2014). Speculators, commodities and cross-market linkages.
Journal of International Money and Finance, 42, 48–70.
[Crossref]

Cheng, I.-H., Kirilenko, A., & Xiong, W. (2015). Convective risk flows in commodity futures
markets. Review of Finance, 19(5), 1733–1781.
[Crossref]

Erb, C. B., & Harvey, C. R. (2006). The strategic and tactical value of commodity futures.
Financial Analysts Journal, 62(2), 69–97.
[Crossref]

Fama, E. F., & French, K. R. (1987). Commodity futures prices: Some evidence on forecast power,
premiums, and the theory of storage. Journal of Business, 60(1), 55–73.
[Crossref]

Fattouh, B., & Mahadeva, L. (2014). Causes and implications of shifts in financial participation in
commodity markets. The Journal of Futures Market, 34(8), 757–787.
[Crossref]

Gorton, G. B., Hayashi, F., & Rouwenhorst, K. G. (2012). The fundamentals of commodity futures
returns. Review of Finance, 17(1), 35–105.
[Crossref]

Gorton, G., & Rouwenhorst, K. G. (2006). Facts and fantasies about commodity futures. Financial
Analysts Journal, 62(2), 47–68.
[Crossref]
Greer, R. J. (1978, Summer). Conservative commodities: A key inflation hedge. Journal of
Portfolio Management, 4(4), 26–29.
[Crossref]

Hamilton, J. D. (1983). Oil and the macroeconomy since World War II. Journal of Political
Economy, 91(2), 228–248.
[Crossref]

Hamilton, J. D. (2003). What is an oil shock? Journal of Econometrics, 113(2), 363–398.


[MathSciNet][Crossref][zbMATH]

Hamilton, J. D., & Wu, J. C. (2014). Risk premia in crude oil futures prices. Journal of
International Money and Finance, 42, 9–37.
[Crossref]

Hicks, J. R. (1939). Value and capital: An inquiry into some fundamental principles of economic
theory. Oxford University Press.

Hirshleifer, D. (1988). Residual risk, trading costs, and commodity futures risk premia. The
Review of Financial Studies, 1(2), 173–193.
[MathSciNet][Crossref]

Keynes, J. M. (1923, March 29). Some aspects of commodity markets. The Manchester Guardian
Commercial, Reconstruction Supplement.

Keynes, J. M. (1930). A treatise on money (Vol. II). Macmillan.

Kilian, L. (2009). Not all oil price shocks are alike: Disentangling demand and supply shocks in
the crude oil market. American Economic Review, 99(3), 1053–1069.
[Crossref]

Kilian, L., & Murphy, D. P. (2014). The role of inventories and speculative trading in the global
market for crude oil. Journal of Applied Econometrics, 29(3), 454–478.
[MathSciNet][Crossref]

Kilian, L., & Vigfussion, R. J. (2017). The role of oil price shocks in causing U.S. recessions.
Journal of Money, Credit and Banking, 49 (8), 1747–1776.
[Crossref]

Neville, H., Draaisma, T., Funnell, B., Harvey, C. R., & Van Hemert, O. (2021). The best strategies
for inflationary times, SSRN.

Stoll, H. R. (1979). Commodity futures and spot price determination and hedging in capital
market equilibrium. The Journal of Financial and Quantitative Analysis, 14(4), 873–894.
[Crossref]

Tang, K., & Xiong, W. (2012). Index investment and the financialization of commodities.
Financial Analysts Journal, 68(6), 54–74.
[MathSciNet][Crossref]

Till, H., & Eagleeye, J. (Eds.). (2007). Intelligent commodity investing. Risk Books.
Footnotes
1 Among many studies of commodity markets in the 1920s, the newspaper article published
by Keynes (1923) is credited with a particularly significant contribution. In this article, the
author highlighted that the value of agricultural inventories after the harvest is so large relative
to financial resources of producers that they cannot themselves bear the risk of inventory value
to drop. Thus, the imbalance requires the service of professional speculators. Keynes was
careful to distinguish between agricultural and extraction commodities, such as oil, for which
the demand for temporary credit is lower due to the ratable nature of production.

2 Hicks (1939), chapter 10.

3 The term normal backwardation was introduced in Keynes (1930), volume II, chapter 29.

4 Keynes (1923).

5 Keynes (1923).

6 This approach follows the mean-variance hedging framework originally developed by Stoll
(1979) and Hirshleifer (1988).

7 The theory of hedging pressure has been combined with the canonical theory of storage by
Gorton et al. (2012), Acharya et al. (2013), and Baker (2021).

8 See Backhouse (2002).

9 Several attempts have been made to provide individual drivers with hedging instruments,
typically by selling them prepaid gasoline cards to be used at retail gas stations, but such
attempts never gained the economy of scale.

10 The term fully funded means that the entire notional value of the futures contract is held as
collateral. We use log-returns which are additive for the reasons of analytical tractability, as
explained in Chap. 2. While it is not recommended to average simple returns, the annualized
arithmetic average of monthly returns over this period would have been even higher at 15.6%.
These returns do not include any additional return on collateral.

11 A typical initial cash margin requirement for energy futures is 10–20% of the contract’s
notional value. Therefore, the returns on cash required to hold futures are 5–10 times higher
than returns on a fully funded position.

12 The idea of investments in diversified commodity indices has been covered extensively in
the literature. For earlier studies of this subject that predate the introduction of energy futures,
see Greer (1978), Bodie and Rosansky (1980), and Fama and French (1987). The addition of
petroleum futures markedly improved the performance and the overall attractiveness of
commodity indices for investors, as documented in the influential work of Gorton and
Rouwenhorst (2006), and Erb and Harvey (2006). See also Till and Eagleeye (2007), Ashton
and Greer (2008), Tang and Xiong (2012), Fattouh and Mahadeva (2014), Bü yü kşahin and Robe
(2014), Hamilton and Wu (2014), Cheng et al. (2015), and references therein.

13 In the USA, a recession is formally defined by the National Bureau of Economic Research
(NBER) as “a significant decline in economic activity spread across the market, lasting more
than a few months, normally visible in real gross domestic product (GDP), real income,
employment, industrial production, and wholesale-retail sales.” In practice, two consecutive
quarters of negative GDP growth is often taken as an indicator of a recession.

14 This observation was first made in Hamilton (1983).

15 Hamilton (2003) developed a nonlinear metric for net oil price increases and applied it to
explain oil price shocks mostly with supply disruptions driven by geopolitical events. In
contrast, Kilian (2009) attributed a much larger role in many oil shocks to growth in global
demand for commodities and to the so-called precautionary demand, which is associated with
speculative buying in anticipation of rising uncertainty and shifts in future expectations. The
latter approach was formalized via a structural vector autoregressive model (VAR) in Kilian and
Murphy (2014) that decomposes contribution of supply disruptions, business cycle-related
demand, and speculative oil-specific demand for inventories to the real price of oil. Another
VAR model was used by Kilian and Vigfussion (2017) to quantify the contribution of oil shocks
to past US recessions.

16 See, for example, Alexander (2001).


17 Many of the previous references on commodity indices also discuss superior investment
performance of commodity futures during periods of high inflation. Neville et al. (2021)
conducted the comprehensive empirical study of hedging against inflation surprises using not
only various financial assets but also some systematic trading strategies. Passive investment in
petroleum futures is shown to be by far the best hedge. Similar conclusions have been reached
in many publications by sell-side research analysts.

18 The thesis of normal contango was suggested in Bouchouev (2012).

19 This approach to the derivation of the hedging equilibrium was proposed in Bouchouev
(2020).
Part II
Quantitative Futures Strategies
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
I. Bouchouev, Virtual Barrels, Springer Texts in Business and Economics
https://doi.org/10.1007/978-3-031-36151-7_5

5. Systematic Risk Premia Strategies


Ilia Bouchouev1
(1) Pentathlon Investments, LLC, Westfield, NJ, USA

Systematic commodity trading attempts to identify and harvest


alternative risk premia caused by structural market imbalances and
behavioral biases of market participants. The primary risk premia
are momentum, carry, and value.
Oil momentum has its roots in the theory of storage. Oil supply and
demand are slow to adjust, inducing persistent trends in inventories.
Many conventional formulations of momentum strategies, however,
are sensitive to parameter selection and prone to overfitting.
Carry strategies also follow from the theory of storage but, unlike
momentum, the carry signal is forward-looking and largely model-
independent. It transmits fundamental information about supply and
demand to the futures market via the behavior of inventory hedgers.
Multiple risk premia can be combined to strengthen the predictive
power of the signal. While carry can be viewed as describing the
prevailing state of inventories, momentum applied to carry is
associated with expected changes in inventories.
The value risk premium is a mean-reverting strategy of buying low
and selling high. Even though the spot price of oil fluctuates cyclically,
betting on price mean-reversion in the futures market is not a viable
strategy due to the headwinds of negative carry.
An important component of systematic trading is dynamic position
sizing. The strength of the signal can be transformed into the optimal
risk exposure via the reaction function. This function is often
characterized by an inflection point beyond which the position is
reduced.

5.1 The Evolution of Algos


Traditional technical analysis can be viewed as the predecessor of
quantitative futures trading. The idea of identifying and betting on
recurrent patterns in commodity prices is as old as commodity trading
itself. Japanese rice traders were known to use candlestick charts in the
eighteenth century, and many other techniques developed by
commodity traders remained secret and largely undocumented. It is
impossible to give the entire credit for pioneering technical trading of
commodity futures to a single person or trade shop. The secret sauce
has been built by generations of traders learning as apprentices from
their masters and passing new tricks to their successors.
The quantitative trading industry grew side by side with the
development of the computing power required to handle large sets of
data. By the late 1950s, several trading units that resembled modern
quantitative hedge funds were already in existence. Perhaps, an
honorary mention should go to Commodities Corp., the fund set up in
1969 in an old farmhouse near Princeton University by Helmut
Weymar. His MIT dissertation on modeling cocoa prices was arguably
the first blueprint for the quantitative analysis of commodity markets.1
Among the founding partners of Commodities Corp. were other MIT
luminaries, including Novel Prize winner Paul Samuelson and Paul
Cootner, best known for his bestseller on random behavior of stock
prices. Much like Keynes, the legends of the academic world were ready
to put their money where their mouth was.2
The original vision of Commodities Corp. was to collect unique
fundamental data and use it to run sophisticated statistical prediction
models. To obtain the data, the fund even hired agents to monitor the
state of cocoa trees in Africa. While the theoretical rationale behind
fundamental models appeared to be sound, the actual trading
performance turned into a fiasco. The markets stubbornly defied
rationality and often moved further away from fundamentally
justifiable fair values. To mitigate mounting losses, the fund introduced
risk management filters to penalize traders for betting against the
prevailing trend. Surprisingly, the risk management system itself
generated consistent profits. It was then applied across a wider
universe of futures markets, which gave rise to what was possibly the
first diversified commodity trend-following system. Besides, its
systematic nature brought significant cost savings, as trading decisions
were entirely based on quantitative rules with no need to employ
agents to track the state of the crops.
Today, quantitative futures traders are usually called commodity
trading advisors (CTAs), even though this term is somewhat of a
misnomer. A CTA is a broad regulatory definition that applies to anyone
who provides investment advice on trading commodity futures. In
addition, the regulatory definition of the term commodity is so broad
that it even includes interest rate and currency futures as financial
commodities. Most money managers must be registered as CTAs,
regardless of how they trade, and vice versa, many CTAs are
discretionary traders. In practice, however, the term CTA became
synonymous with the business of managed futures, or any systematic
futures trading based on quantitative algorithms.
The business of a quantitative CTA is based on the science of
statistical inference, which identifies repeatable historical patterns and
attempts to extrapolate them into the future. A hypothesis is made
about the relationship among certain variables and a prediction is
formed that this relationship will continue to hold with a certain
probability. Trades are placed based on such predictions, generally with
no discretionary intervention. The goal is to eliminate human biases,
such as the avoidance of buying high or selling low, or the reluctance to
size up when it feels the least comfortable. Since the trading rules are
known, the performance of the strategy can be evaluated by
backtesting, or, in other words, by simulating the strategy’s historical
results. To validate the hypothesis with the arrival of new data, the
simulation process must always be tested out of sample. The statistical
edge in each prediction is rather small, so one needs to apply it either
many times repeatedly or diversify across many assets, much like in the
business of running a casino.
Commodity futures turned into fertile ground for CTAs for a number
of reasons. Futures trading provides high leverage as the initial cost is
only a small fraction of the notional value of the asset. Furthermore,
different commodities are impacted by their idiosyncratic factors that
make strategies less correlated and the overall portfolio better
diversified. Unlike equities, there are no restrictions on shorting
futures, and one does not need to pay the dealer to borrow the stock;
selling futures is as easy as buying. Finally, commodities have larger
probabilities of extreme moves. These moves can be captured with
systematic momentum and breakout strategies. It would be more
difficult to execute such a strategy discretionarily as buying high and
selling low goes against human contrarian biases.
In addition, commodity momentum strategies bring valuable
diversification to broader investment portfolios. This diversification
comes from the negative correlation during major market selloffs, when
many financial portfolios struggle but momentum strategies tend to do
well. Much like oil investments functioning as a hedge against inflation,
the momentum strategy performs well precisely when it is the most
needed. Some hedge funds even developed investable products to
capture a crisis alpha with an asymmetric momentum signal which is
only triggered on the downside. Such one-sided momentum strategies
are marketed as an alternative to buying insurance in the form of put
options. In addition, investors in a systematic strategy may feel that
they are in control since trading decisions can be anticipated, in
contrast to less predictable judgments of discretionary managers.
As normal backwardation largely disappeared from many
commodity markets and passive long-only indices stopped being
profitable, banks had a desperate need to find new investable
commodity products to retain large financial investors. The catalyst
came from equity markets, which were enthusiastically exploring the
novel concept of factor investing. The idea was to replace traditional
portfolio construction by asset class with allocations to common risk
factors that can explain price behaviors across different markets.
The concept of risk factors was quickly adopted and subsequently
enhanced by commodity futures traders. It was relatively
straightforward to replace struggling long commodity indices with
more dynamic commodity portfolios that combine both long and short
positions. The signals would be based on the common risk factors that
investors were familiar with from equity markets but applied across a
broader universe of commodities. The goal once again was to extract a
small risk premium from each market and reap diversification benefits
from the large number of mostly uncorrelated commodity sub-sectors.
The three most powerful commodity risk factors are momentum,
carry, and value, which can be combined in many different ways. A
diversified systematic commodity portfolio is typically constructed by
ranking its constituents on the cross-asset basis. A common method to
do this is to sort all commodities within the portfolio based on the
strength of various factors and assign them some sort of score.
Commodities with the highest scores are then bought, and the ones that
have the lowest scores are sold. Many alternative risk factors, such as
inventory, speculative positioning, volatility, skewness, and open
interest have also been considered, but as we explain in the next
chapter, these factors in the oil market are better used in a less
systematic manner.3
It turns out that despite the complexity of a diversified commodity
risk factor portfolio, the largest portion of the profits is typically
generated by energy futures. The role of many other commodity sectors
is mostly to provide diversification and to reduce the denominator of
the risk-adjusted returns. In the following sections we will explain that
the success of these strategies in the energy market can be explained by
the special role of storage and the related behavior of large market
participants.

5.2 Myths and Realities of Oil Momentum


The basic thesis of momentum or trend-following strategies is buying
what has already gone up and selling what has gone down. One can use
a popular analogy that the market, like a horse, is easier to ride in the
direction that it is already going. For many discretionary traders,
however, it is undoubtedly challenging to embrace the counterintuitive
concept of buying high and selling low. Well-known human behavioral
biases direct many of us to do exactly the opposite and attempt to buy
on dips and sell on rallies. We are reluctant to lock in losses and often
retain falling assets for too long in the hope of their eventual recovery,
and we tend to take profits prematurely to enjoy faster gratification.
Momentum trading takes the other side of such human biases. To
eliminate these emotions from decision making, the momentum
strategy is typically implemented in a rule-based systematic manner
that does not allow for any human intervention.
There is no shortage of behavioral theories trying to explain the
root cause of momentum observed across many financial markets.
Some argue that momentum is created by investors underreacting to
new information which is slow to diffuse. Others think that it is driven
by investors extrapolating past returns due to inertia. One can also add
overconfidence bias as a factor, as we are often reluctant to adjust our
subjective views quickly if new information does not support them. We
also tend to chase the desired investment more aggressively for the fear
of missing out when fresh news confirms our beliefs. More recently,
herding started to play a larger role in short-term momentum, ignited
by interactions and mutual reinforcement on social media.
While such behavioral explanations may have some merit for
diversified equity markets with broad participation of retail investors, a
more plausible explanation for momentum in the oil market lies in its
fundamentals. Oil supply and demand are price inelastic, and they are
very slow to adjust. If oil consumption exceeds oil production today,
then it is more likely that the status quo will prevail tomorrow, which
results in drawing inventories. Oil consumption tends to follow longer-
term business cycles, but relatively rigid production cannot adjust
quickly in response to unexpected demand shocks. As a consequence of
the theory of storage, such persistence in the dynamics of oil
inventories must then translate into persistence of prices.
We will use the concept of momentum synonymously with trend
following. In the context of a single market, the two terms are identical.
More generally, trend following should be identified as the time series
momentum that generates trading signals solely based on the asset’s
own price history. In contrast, cross-sectional momentum, which is
often used in the construction of broad commodity momentum
portfolios, ranks assets based on the relative strength of their
corresponding momentum signals. In such portfolios, one buys
commodities with the highest momentum rank and sells the ones with
the lowest rank, while maintaining overall relatively neutral exposure
to an asset class. If desired, one can also retain the directional bias of
the overall portfolio by buying all prior winners and selling all prior
losers within the asset class.
As a first-order approximation, one can interpret a portfolio of time
series momentum for individual commodities as the sum of the cross-
sectional portfolio and some common trend across all commodities.
The contribution of a common trend, however, is much smaller for
commodities than it is for equities. In normal market environments
individual commodities are not as highly correlated with each other.
These low correlations dilute the commodities’ common trend, except
for occasional episodes of high inflation and global demand shocks that
may have similar impact across many commodities. One example of
such a common trend is the negative demand shock at the onset of the
Covid-19 pandemic, followed by the positive inflationary shock driven
by the subsequent recovery. If there is no strong trend in the
commodities sector overall, then the performance of a cross-sectional
momentum portfolio will be similar to the combined performance of
trend-following strategies for its constituents. If, in addition, all
commodities trend together, then the portfolio of time series
momentum will outperform the sector-neutral portfolio of cross-asset
momentum.
An important, but often overlooked, feature of the long-short
momentum portfolio is the disproportionately large contribution of
energy futures that tend to exhibit stronger momentum properties. As
discussed previously, the complexity of oil and gas storage leads to
structurally different return distributions for energy commodities. In
contrast to many agricultural and other seasonal commodities that tend
to have a price distribution with a larger upside tail, oil follows an up
the stairs, down the elevator dynamics. Such a behavior is more typical
for a financial asset than for a typical consumption commodity. Large
periodic breakdowns of the oil market make a diversified momentum
strategy highly desirable to financial investors. As we will see later in
the discussion of oil options, the downside of oil momentum is also
exacerbated by delta hedgers of short put options.
There are many ways to define the time series momentum. The
most commonly used approach is to compare the latest price to some of
its recent history. The latter is often measured by the moving average of
prices Pt over the prior n days, defined as
We should note that for consistency with standard academic literature
on time series analysis, time variable in this part of the book is
indicated by a subscript, so that P(t) = Pt. Likewise, for the futures price
F(t, T) = Ft(T) = Ft, where the maturity T is sometimes omitted when it
is clear from the context.
We then define basic momentum Mt as the spread between the
latest price and its moving average

The momentum trading signal, πM(Ft), buys and sells futures Ft based
on the sign of Mt4

(5.1)

The strategy P&L at time t is calculated using the momentum signal at


time t − 1, i.e.,

While this clarification may appear to be obvious, we mention it


explicitly to prevent analysts from making a grave mistake of
calculating daily P&L at time t using the signal calculated at the end of
the same trading day.
Note the strategy specification (5.1) trades futures Ft even though
the momentum signal could be defined for a different time series Pt. In
other words, the signal generation does not have to apply to the same
instrument that one trades. For example, momentum signal can be
calculated using the time series for fixed nearby or for fixed maturity
month futures contract. In the following chapter, we also provide
examples of applying the momentum signal to non-price variables, such
as fundamentals and flows. Whenever the price series Pt is clear from
the context, we do not reference it explicitly and simply use MAt(n) and
Mt(n) to denote, respectively, n-day moving average and n-day
momentum. In this chapter, we only use a simple price momentum, but
even in the basic case, some implementation nuances of the momentum
strategy must be handled with caution.
Consider, for example, the impact of futures rolls. CTAs tend to keep
their positions in the most liquid contract, which is typically the nearby
futures. Prompt futures are usually rolled within one or two weeks
prior to their expiration, when the liquidity starts shifting to the second
nearby futures contract. Handling the rolls in the definition of the price
momentum based on moving averages could be tricky. If one uses a
continuous time series of prompt futures, then the impact of the rolls
would generate artificial jumps in the time series which could produce
false momentum signals. Instead, it is more appropriate to calculate
moving averages using the roll-adjusted time series, which is a
synthetic time series constructed from cumulative price changes for the
appropriate contract.
The proper adjustment for rolls in signal generation is particularly
important for seasonal commodities. Consider, for example, RBOB
futures.5 The prompt futures price jumps when RBOB transitions from
its winter specification, which ends with the March contract, to the
summer specification, which starts with the April contract. Winter and
summer RBOB represent two rather different molecules, and the latter
always trades at a significant premium to the former due to additional
blending expenses required for compliance with stricter environmental
regulations during summer. If momentum is defined for the continuous
time series of the prompt contract, then a seasonal jump will generate a
long signal on the day of the March–April roll and a short signal in the
fall when the contract specification reverts to the winter grade. Such
jumps will generate false signals that have nothing to do with the
trendiness of a given futures contract. If moving averages are calculated
using different contracts, then additional seasonality adjustments are
usually required.
An alternative way that bypasses the issue of rolls is to redefine the
momentum signal in terms of price changes or roll-adjusted returns
instead of moving averages of price levels. Momentum then means
buying futures when past return over a given period are net positive
and selling futures if past returns are net negative. We show next that
these two definitions are similar, as moving averages can be expressed
as time-weighted price changes.
To illustrate the equivalence between the two alternative
definitions, consider an example of a simple short-term momentum
based on a five-day moving average which does not cross the rollover
day:

This formula can be rewritten as the linear combination of four


previous price changes, where the weights decrease linearly for older
observations:

This formulation is intuitive, as the largest weight is attributed to the


latest and arguably most relevant observation. The weights on past
price changes gradually decrease, which makes the signal less sensitive
to prices that drop out of the sample.
The same transformation can be repeated for any lookback period
n:

where

defines daily price changes. The weights on price changes again


monotonically decrease with time, and the largest weight is assigned to
the latest price change. This formulation is also convenient because
price changes can easily be used for the specific contract regardless of
what nearby futures it represents at different times.
One can further modify the definition of momentum and specify it
directly in terms of a weighted average of prior price changes:
For a very specific choice of weights given by , this definition
is equivalent to the one based on the moving average of prices.
However, the weights ωi can also be chosen in many other ways. For
example, exponentially decreasing weights can be used to assign higher
relevance to the recent data. One can also use equal weights ,
in which case the definition of momentum reduces to the difference
between the latest price Pt and the roll-adjusted price n − 1 days ago
P t − n + 1.
Figure 5.1 illustrates the long-term performance for the benchmark
20-day momentum strategy Mt(20) for WTI futures, which corresponds
to approximately one-month lookback. In this example, we use a
prompt futures contract which is rolled on the last business day of the
calendar month. Such a rolling schedule simplifies the analysis of
diversified portfolios of commodity futures, as it avoids dealing with
expiration calendars for individual commodities. It also makes it easier
for interested readers to replicate the results as an exercise. As before,
for simplicity we use log-returns which are additive.

Fig. 5.1 Historical returns of a 20-day momentum strategy for WTI futures (1993–2022)
At first glance the average annual return of 10.4% generated by this
simple strategy over such a long time may look appealing. However, it is
less attractive on a risk-adjusted basis, with an information ratio of only
0.27.6 In four years, the strategy drawdowns substantially exceeded
25%, which often marks the maximum loss that investors are willing to
tolerate for any hedge fund investment strategy. While the momentum
signal clearly has some informational content, the strategy is unlikely to
be good enough to be traded on a stand-alone basis.
One of the main objectives of this book is not only to share ideas
that have a proven track record, but also to highlight their pitfalls.
While momentum is a useful trading concept, it is, unfortunately, one
that is often abused. In systematic trading, such abuse comes from data
mining and overfitting when the trading signal is allowed to have too
many degrees of freedom that are fitted to produce better-looking
backtests. The more parameters a trading strategy has, the more likely
it is to fall apart when some of these parameters are modified.
Momentum strategies provide a good case study.
The most common extension of basic momentum is to smooth the
signal sensitivity to the latest data point. Financial markets are full of
noise, and the momentum trader may want to avoid being whipsawed
by false signals and instead to wait for some confirmation of the trend
to be firmly established. Such smoothing is typically done by replacing
the latest price in the definition of momentum with another, shorter-
term moving average. The crossover momentum signal is generated by
the difference between moving averages

The trading signal πM of the momentum strategy then switches


between long and short futures Ft when the two moving averages cross
over

While an introduction of additional degrees of freedom provides


traders with more flexibility to search for optimal parameters to
improve the historical backtest, it adversely effects the stability of the
output. Here, one effectively searches for local minima in a
multidimensional space of parameters, and such local minima are
rarely unique.
To illustrate the nature of instability inherent in many momentum
strategies, we can again translate the crossover of moving averages to
an equivalent definition based on price changes. Following the previous
transformation, we represent both moving averages of prices in terms
of price changes:

and

Then using these representations, we can write the crossover


momentum as follows

where the two Pt terms in the last line cancel out.


We then split the second summation into the first m − 1 and
remaining n − m data points, and combine the former with the first
summation to obtain

The crossover momentum is still represented as a weighted sum of


price changes, but the weights ωi are no longer monotonically
decreasing. Since
the largest weight is assigned to the date which corresponds to the start
of the short lookback window.
Figure 5.2 illustrates these weights for the M(m, 20) momentum
strategy. The weights reach their corresponding peaks for price
changes that occurred exactly m − 1 days ago.

Fig. 5.2 The weights for the M(m, 20) momentum strategy peak for the price change that
occurred exactly m − 1 days ago

This simple illustration exposes how easy it is to overfit the


crossover momentum strategy. By varying m in the signal construction,
one can be fooled by seeking to improve the historical performance,
while essentially fitting to the noise. An optimizer will always pick the
best combination of parameters, which will likely artificially overweight
the contribution of some large favorable price move that happened to
occur precisely m − 1 days ago. The model becomes unstable and any
attempt to dynamically adjust the set of optimal parameters will result
in nothing but fitting to randomness.
Figure 5.3 illustrates how sensitive the results of such momentum
strategies can be to even a small perturbation of parameters. For
example, in 2016 the base momentum strategy would have lost 40%, a
remarkably similar strategy where the short-term momentum averages
prices over the past three days would have made 20%, and returns on
strategies with price averaging over two and four days were close to
zero.

Fig. 5.3 An example of instability of crossover momentum strategies with respect to short-
term moving average. In 2016, M(1,20) lost 40%, M(3,20) made 20%, while returns on M(2,20)
and M(4,20) strategies were close to zero
There is nothing structural about this choice of parameters and one
can easily find another period when the reverse was true. The
difference is typically driven by how well the parameter optimizer
captures a small number of large returns that tend to dominate the
overall strategy performance. On the bright side, the sensitivity to the
longer-term moving average, n, becomes significantly smaller as the
weights on the corresponding price changes monotonically decrease.
Our goal here is not to find the best combination of parameters for
the oil momentum strategy. We believe that this would be an ill-
intentioned task given the non-stationarity of the financial time series.
It is clear that the oil momentum does exist. It contains some useful
information that one can see even in the historical performance of the
basic momentum strategy. A few other formulations of momentum
strategies are also worth mentioning, not because they are better, but
more because they are often cited by researchers who claim to have
found some other magic versions of momentum.
One popular variation of momentum is a class of breakout strategies
where the asset is bought or sold only when the latest price moves
beyond its prior maximum or minimum over some given lookback
period. In contrast to the basic momentum, the breakout strategy is
likely to remain on the sidelines for substantial periods of time, which,
unfortunately, makes it even more prone to overfitting. Further
variations could include filters to keep the moving averages crossover
intact for a certain number of days to avoid being whipsawed before
taking a position. Alternatively, one can impose the filter to enter the
trade only when the spread between two moving averages exceeds a
certain threshold, which becomes an additional parameter as well.
Similar conditions can be imposed around the exit of the position.
There are a few other general techniques that are helpful for
momentum trading regardless of its exact specification. One can often
benefit from combining momentum signals across multiple trading
frequencies. For example, one can use the crossover of the latest price
and its weekly price average as the short-term momentum signal, the
weekly versus monthly averages as the medium-term signal, and the
monthly versus annual moving averages as the longer-term momentum
signal. Then the three signals can be combined into an aggregate
momentum score which determines the overall size of the trading
position. The weights for each of the three momentum frequencies can
also be customized.
Undoubtedly, the more parameters the strategy backtest is allowed
to have, the better the fitted results will inevitably be. However, if one
normalizes numerous permutations of momentum strategies for the
same degrees of freedom, then their long-term performances tested
out-of-sample become broadly similar as well. Any claim of the
discovery of some magic momentum specification that works much
better than any other momentum strategies should be treated with a
great amount of skepticism. If one is sufficiently excited about the
rationale of the oil momentum strategy, then the best advice would be
not to chase the best backtest but to keep the strategy simple and to
avoid overfitting.
One of the reasons that explains the popularity of the oil
momentum strategy is its negative correlation to directional financial
investments during synchronized risk-off events when any benefits of
portfolio diversification are particularly appreciated. For example, a
strategy of buying and rolling oil futures lost over 50% in four years
since 2000, but simultaneous investments in a momentum strategy
would have largely offset these losses. Figure 5.4 highlights these years
with blue dots. Since momentum strategies tend to do well when prices
make large moves in either direction but underperform when prices
are range bound, the graph of momentum returns versus price returns
is sometimes referred to as momentum smile.

Fig. 5.4 Momentum strategies would have mostly offset large losses from directional
investments in oil futures (WTI, 2000–2022)

One important topic for systematic trading that we have not yet
discussed is the sensitivity of the strategy to transaction costs.
Fortunately, for liquid oil futures, such as WTI and Brent, the existence
of the TAS mechanism described in Chap. 3 makes the task much easier
as it allows systematic traders to lock in the settlement price practically
without crossing the bid-ask spread. We should note that similarly set
up minute marker contracts also allow traders to lock in prices during
pricing windows for physical markets in the European and Asian time
zones. The European prices are often used by managers of global
systematic portfolios to capture the best simultaneous liquidity across
the three main time zones. One can also run faster intraday momentum
strategies based on higher-frequency data. Such strategies are outside
the scope of this book, as they depend more on the market
microstructure than on the behavior of participants.
To summarize, one can consider practically infinite permutations
for different momentum strategies for oil futures, but in the long run,
most exhibit similar predictive power once the somewhat artificial
contribution driven by additional degrees of freedom is removed. Oil
momentum does add some useful information content, but its strength
appears to be declining as markets become more liquid and mature.
Practitioners observed that momentum strategies generally work
better in less liquid markets, where the information diffuses more
slowly. However, the execution of such strategies in illiquid markets is
more challenging due to higher transaction costs.
The oil momentum can be used to enhance other strategies with
more solid economic foundations, but it should not be abused by data
mining. As we will see later, the momentum in liquid oil markets can
still add significant value when it is combined with other drivers that
have stronger economic foundation. Momentum is a great optimizer of
other trading signals and should be used as such. As a stand-alone
trading strategy, it is not sufficiently robust and the economic
foundation behind the concept also remains rather weak. As traders
often quip, momentum works until it does not.

5.3 Carry as a Transmitter of Fundamentals to


Prices
One pivotal factor that threads through virtually every aspect of oil
trading is the shape of the futures curve. It arises everywhere while
often hiding under different names in specific applications. Some
economists may recognize it as the own rate of interest, fundamental
traders know it as the convenience yield, and financial investors call it
the roll yield. Systematic traders refer to the same term by yet another
name, the commodity carry.
In the world of systematic trading, the term carry is often
understood in a broader sense, which allows it to be applied across
many markets of different nature. Simplistically, carry can be defined as
P&L if nothing changes. When applied to commodity futures, this
definition means P&L of the futures trading strategy under the
assumption that the spot price does not change, and the term structure
of futures with fixed time to expiration retains its shape. The P&L is
then determined by how futures roll up or down along the curve, as
time passes. In this static spot price scenario, the expected spot price is
the same as the current spot price. Therefore, the first term in the
decomposition (4.​1) is zero:

and the risk premium is determined by the carry, which is the value of
the spot-futures spread, as observed at time t.
The concept of systematic carry trading came from foreign exchange
markets. It hinges on the same principle of no-arbitrage between two
alternative money standards that Irving Fisher illustrated for the case
of the commodity money, as it was explained in Chap. 2. The argument
applies to fiat money of different countries as well. Each currency pays
its own domestic rate of interest. The forward value of a currency with
a relatively high interest rate must be lower relative to a currency with
a lower interest rate to make investors indifferent in which currency to
hold their money. According to an economic theory, the magnitude of
the forward discount, or an expected currency depreciation, must
negate the value that can be gained from interest rate differentials, the
hypothesis known as an uncovered interest rate parity. In practice,
however, this economic theory does not hold. In the long run, holding
riskier currencies with higher yields is often more profitable, as on
average, expected currency depreciation is not fully realized. Obviously,
in the short-term such a strategy can experience large losses, but over
time investors are generally able to extract positive risk premium from
the foreign exchange carry trade.
Using the broad definition of carry as the strategy return when
nothing changes, the idea of carry trading can be extended to almost
every asset class. In equities, carry is simply given by the dividend yield.
The future stock price is then discounted relative to the current stock
price by the amount of the expected dividend, net of the risk-free rate.
The carry trade in the equity market buys single-stock futures that have
larger discounts relative to the current stock price, which is equivalent
to buying high-dividend stocks. The dividend yield behaves like the
foreign interest rate in the currency market, or like the convenience
yield of holding oil inventories. In the fixed income market, carry is
simply the coupon received from holding the bond, or the roll down in
the interest rate swap curve. The fixed income carry strategy then buys
bonds with higher coupons and steeper interest rate curves, where
short-term rates are typically lower than long-term rates.7
The power of carry in the oil market has already been shown in the
previous chapter. When carry is positive and the oil market is in
backwardation, buying oil futures tends to be a profitable strategy. As
the futures curve moves into contango and the carry becomes negative,
the investor returns deteriorate. Since, in the futures market, selling is
as easy as buying, one can replace the long-only futures investment
with the dynamic strategy of buying and selling based on the direction
of carry. Like momentum, the same idea can be applied across all
commodities to achieve some diversification, but the strategy again
works better for energy commodities. Oil carry is also rooted in the
theory of storage. Unlike momentum, however, the transmission of the
carry theory to the futures market is much more explicit. An oil carry
plays a unique role in the market as it transforms the fundamentals of
supply and demand into pressure on prices via the behavior of
inventory hedgers.
The business model of the storage trader is to generate low-risk
steady returns by providing the service of storing oil. In such a business
model, the risk of volatile oil prices must be eliminated by hedging in
the futures market. As soon as a physical barrel is purchased for
storage, a financial barrel is sold in the futures market to reduce the
risk. This downward pressure on futures price occurs when the
contango is steep enough to cover the cost of storage. If the market flips
to backwardation or the contango becomes too narrow to economically
hold inventories, then the storage hedger is incentivized to pull oil out
of storage and buy back short futures hedges. This creates an upward
pressure on the futures price. The carry transmission channel is
illustrated in Fig. 5.5.

Fig. 5.5 An inventory hedger sells futures when contango covers the cost of storage and buys
them back otherwise
Such buying and selling of futures by inventory hedgers driven by
the shape of the futures curve is rather mechanical and somewhat
predictable. The cost of carry, which includes transportation, logistics,
and borrowing costs, and the convenience yield of holding inventories
depend on the economics of individual traders, which makes hedging
pressure vary with the steepness of the curve. In general, the steeper
the contango, the more widespread hedging becomes among inventory
managers, which increases the selling pressure in the futures market.
However, when carry becomes too steep, most of the hedging is already
finished, and the pressure on price starts waning. We address this
important transition point in more detail in the last section of this
chapter.
To define the systematic carry signal more formally, let
represent the spread between the two futures contracts with M and
N months to expiration. The basic carry strategy buys futures when
carry Ct is positive, and sells futures when it is negative, i.e., the trading
signal πc can be defined as
(5.2)
The definition of the carry signal again does not have to use the same
contract that is traded. Like in the case of the momentum strategy,
multiple carry signals computed from different parts of the futures
curve can be combined to generate a single aggregate carry signal.
However, in the case of crude oil multiple carry signals measured from
different parts of the futures curve tend to have the same sign and
become largely redundant. If the futures curve is monotonically
decreasing or increasing, then the sign of each monthly carry signal is
the same for all M and N.
Figure 5.6 shows the historical returns of the benchmark WTI carry
strategy that buys and sells prompt futures contracts based on the sign
of the annual carry Ct(1, 13). The twelve-month spread is commonly
used to define carry to eliminate seasonal effects, which are more
substantial for refined products and natural gas. As before, the futures
are rolled at the end of the month. Since both momentum and carry
strategies maintain either a long or a short position at all times, the two
strategies have the same volatility. The carry strategy, however,
generated a much higher annualized log-return of 20.3% over the
thirty-year period with impressive Sharpe ratio of 0.53. In addition, the
drawdowns of the carry strategy are significantly smaller than the
drawdowns of the momentum strategy.
Fig. 5.6 Historical returns of the C(1,13) carry strategy for WTI futures (1993–2022)
Carry, or the shape of the futures curve, is the ultimate mechanism
that transmits fundamental information into futures prices. When we
read media reports that the price of oil went up because of stronger
Asian demand, or it went down because of increased OPEC production,
these reports do not tell the full story. OPEC itself does not buy futures
and very few oil consumers do. However, any change in supply and
demand impacts the spot-futures spread via variations in convenience
yields and storage costs. The actual futures are bought and sold by
storage traders, who react to the economics determined by the carry.
We will see in the following chapter that an inventory hedger is indeed
the largest trader in the oil futures market. The systematic carry
strategy described in this section essentially trades ahead of
anticipated behavior of the largest market participant, who is acting in
response to the arrival of new fundamental information.
Compared with the momentum strategy, carry trading has one
tremendous advantage. The strategy is essentially model-free. Unlike
momentum, carry is a forward-looking signal. It does not depend on
history and does not require any parameter estimation. The trading
signal is directly observable from the current shape of the futures
curve. It measures the economic incentives of inventory hedgers and,
by and large, the strategy front-runs the expected flows of hedgers in
the futures market.
The idea of oil carry is probably one of the most powerful and
widely used indicators for directional oil trading. Many physical traders
simply refuse to take any positions that go against the direction of
carry. A similar idea of investing in oil only when the carry is positive
has found some interesting applications in broader financial
portfolios.8
The primary drawback of the carry strategy is its inability to react
fast enough to rapidly changing market conditions. Since the shape of
the futures curve does not flip frequently between contango and
backwardation, the carry signal often retains the same sign for a long
time. Even though it may capture early gains quickly when the shape of
the curve changes, it then often gives up a portion of the gains when the
market starts correcting. By holding and rolling the same directional
position until the time spread crosses zero, or some pre-defined
threshold, the carry strategy often misses early warning signs that
fundamentals are beginning to change. To capture an anticipated
change in fundamentals, traders often look at the change in carry
instead of its level. One can think about carry being a measure of the
current state of supply and demand, and the change in carry as a
measure of expected changes in future supply and demand.
A popular way to make the carry strategy more dynamic is to blend
it with the basic momentum signal. This can be accomplished by
applying momentum not to the price of oil, but instead directly to carry,
or to the shape of the futures curve. The carry-momentum trading
signal πCM is then defined as
(5.3)
The results from such simple signal blending, shown in Fig. 5.7, are
quite remarkable. The strategy, which we call the dynamic carry, or
carry-momentum, generated 24.7% annualized returns with the Sharpe
ratio of 0.64 over the thirty-year period. Out of hundreds of different
blends of systematic signals that the author has traded over many
years, the carry-momentum signal stood out in its long-term
robustness. It highlights how a technical indicator, such as momentum,
can improve a trading concept that has stronger links to the market
fundamentals.

Fig. 5.7 Historical returns of the 20-day carry-momentum CM(1,13) strategy for WTI futures
(1993–2022)

5.4 Value and Mean-Reversion


The success of momentum and carry strategies in the oil market may
surprise traders who are more accustomed to a bargain-hunting value
trading style. Since oil prices are cyclical and expected to mean-revert,
the natural temptation for many discretionary traders would be to do
the opposite of momentum, and instead attempt to buy low and sell
high. The convergence strategy of betting against price deviations from
some long-term equilibrium level is an example of the value risk
premium. Value is a contrarian strategy where one buys what is
deemed to be underpriced and sells what is overpriced relative to some
price level that is considered to be fair.
Establishing the fair value for oil is difficult. Fundamentally, it only
makes sense for contracts with long maturities where the fair price can
be approximated by the marginal cost of production. Short-term futures
can then be viewed as the spread to the long-term fair value. However,
the volatility of this spread driven by short-term fluctuations in supply
and demand is too high for any meaningful fundamental definition of
the short-term fair value. As it was demonstrated in Chap. 3, oil price
can rise or fall without any limits when inventories approach zero or
reach the maximum level of the storage capacity. In Chap. 7, we will
take a different approach and construct another example of the fair
value of oil by calculating it as a function of macroeconomic variables.
In practice, one often estimates the fair value statistically by
comparing the current price to its moving averages, like in the
definition of momentum. The purpose of moving averages is to
incorporate ongoing structural changes that constantly occur in the oil
market. In its simplest form, one can define the value trading signal πV
as the mirror image of the momentum indicator, such as

This strategy buys futures when the price falls below its moving
average and sells when it rises above. Obviously, since the momentum
strategy is generally profitable, systematically running the opposite
value strategy would be a loser.
To capture the value risk premium, traders usually run their
strategies on slower frequencies. While it is critical for momentum
traders to react fast and trade as soon as the price crosses some trigger,
value traders generally prefer to remain patient and avoid rushing into
a trade. They often let the market run in the same direction for some
distance and only trade when the deviation exceeds a certain threshold,
ε. Such a threshold-based value signal can be defined as:

(5.4)

In this variation of the value strategy, the trader may remain on the
sidelines for a while and bet only when the deck is rich, i.e., when the
price deviates sufficiently far away from its normal range.
In general, calling the tops and the bottoms for the price of oil is not
a prudent strategy in the futures market, as one must be very precise
on the timing of anticipated price reversals. If one could trade physical
oil and keep some oil in storage without incurring any cost, then selling
high and buying low would undoubtedly work well, as spot price does
eventually mean-revert. However, storage is not free, and it is rarely
accessible by financial traders, so that the closest proxy that one can
trade is the futures contract. Buying low and selling high in the futures
market, however, is very different from doing it in the spot market. In
the futures market, the value trader constantly fights against the
punitive cost of carry, which is determined by the slope and the
convexity of the futures curve.
If the market is oversupplied, then the price of prompt futures is
likely to be lower than the forward price. The slope of the futures curve
reflects the cost of storage, which is typically higher for shorter-
maturity futures, where the fundamental imbalance is the most acute.
Storage buys time either for consumption to increase or for production
to be curtailed. As time moves forward and fundamentals normalize,
the cost of storage decreases, and consequently the slope of the futures
curve flattens. Since the slope, which is the first derivative with respect
to futures maturity, is decreasing, then the futures curve in an
oversupplied market is expected to be concave. We call this a state of
concave contango.
Likewise, at times when inventories are low, the slope of the futures
curve is the steepest in the front-end of the curve, which is dominated
by the highest convenience yield of owning a physical barrel. As
inventories normalize over time, the convenience yield decreases,
which makes the forward slopes flatter. This is a state of convex
backwardation. The futures curves that are shown in Fig. 5.5 are in
their normal states of concave contango and convex backwardation.
The convexity of the futures curve creates a dilemma for the value
trader: whether to use shorter-term futures and lock in better value at
the expense of a steeper carry, or trade further out on the curve where
the carry is less punitive, but the value is also less attractive. Figure 5.8
illustrates the challenge of buying cheap futures when the market is in
concave contango.
Fig. 5.8 Buying cheaper short-term futures in concave contango is offset by steeper carry,
while buying long-term futures has less value
The case of selling futures in convex backwardation is handled
similarly, as the mirror image of concave contango. In either case, for
the value strategy to be profitable the magnitude of the expected price
mean-reversion must dominate the accumulation of carry over the
holding period. One should avoid using short-term futures for
capturing value unless the reversal is deemed to be imminent as
indicated by other fundamental factors. Otherwise, time is not on the
side of the value trader. The apparent cheapness of the futures contract
could quickly become overpowered by the adverse buildup of the
negative carry. These types of strategies where a systematic signal is
best combined with a discretionary overlay based on fundamental and
flow factors are discussed in the next chapter.
In theory, it is possible for both systematic momentum and value
strategies to generate positive returns if the two are traded on different
frequencies. Momentum tends to work better on the shorter time scale,
while the contrarian value signal works better in the longer run,
especially when an additional buffer is provided by a threshold ε. One
can also combine momentum, carry and value into signal-integrated
portfolios. This is typically done on a cross-sectional basis within
broader commodities portfolio.
Overall, momentum and carry signals tend to overpower value in
directional oil trading, but in the next chapter we will show that value
strategies are more robust in trading futures spreads and construct one
of such diversified energy value portfolios. For now, we continue to
focus on single-asset strategies and conclude this chapter by showing
another way of blending multiple signals that can help systematic
traders to determine the optimal size of their bet.

5.5 The Reaction Function


The simplest way to size positions within broad systematic portfolios is
to apply so-called volatility targeting. Similarly to the risk parity
strategy, every asset or strategy in the systematic portfolio is often
allocated the same amount of risk. The risk is measured as the product
of the notional size of the strategy and its volatility. Therefore, to keep
the dollar risk the same across all assets, the size of the position must
be inversely proportional to volatility. If the asset volatility increases,
then the risk capital is withdrawn and the position must be reduced.
Such a forced liquidation, however, is suboptimal as it does not take
into consideration the strength of the systematic signal.
More advanced systematic traders incorporate not only volatility
but also the magnitude of multiple signals and combine them to
determine an optimal position size. To illustrate this for a single
commodity, one can scale the position by blending various risk premia,
such as momentum and value. While value is unlikely to be an attractive
directional strategy on its own, it can be used to optimize momentum
and carry signals. The idea is for the value metric to act as a control
check to indicate that momentum and carry might have gone too far in
one direction and the status quo is unlikely to remain for much longer.
In Chap. 7 we will discuss the fact that this dynamic is typical for a
complex dynamic system. In such a system the presence of natural
boundaries eventually limits its growth, and the value signal in a
systematic trading strategy is akin to measuring the distance from the
system fundamental boundary. If momentum or carry signals become
too strong, then the market might have already moved too far outside of
its normal range, and the likelihood of a price reversal increases. In
such cases, it would be prudent to reduce the size of the position
inversely to the strength of the momentum or carry signals.
To quantify the blending of directional momentum and carry signals
with value, we use a so-called signal transformation or reaction
function. This is a function that maps the strength of the systematic
signal to the size of the bet. This function typically increases along with
the strength of the signal but only up to a certain point, beyond which
any further position increases are no longer justified. We call this an
inflection point. For example, for the momentum signal such an
inflection point may indicate that the trade is too crowded, and the
momentum has high probability to subside, as everyone who wanted to
be in this trade is already in this trade.
The reaction function can be specified parametrically. Figure 5.9
provides one example of such a function defined as
(5.5)
The dependent variable in Fig. 5.9 is some metric of the strength of the
momentum. In diversified systematic portfolios the metric is often
normalized. This makes it easier to compare signals across different
markets and to ensure more balanced risk deployment for assets with
different volatilities. Without such normalization the strategy
performance will be dominated by more volatile futures. The standard
way to normalize systematic signals is to divide raw signals by the
standard deviation of the asset returns. The resulting normalized
metric is known as the z-score, which represents the strength of the
trading signal in terms of its number of standard deviations. In the
reaction function above, parameters were chosen to set the maximum
position of ±1 when the normalized signal is equal to ±1.
Fig. 5.9 An example of a reaction function for the momentum strategy
Many other reaction functions are used by systematic traders, but
the general idea is to grow the position size up to a certain maximum
level typically defined by the trader’s risk limit, and then gradually
reduce it when the signal becomes too strong. This avoids taking
excessive risks when prices move too far away from their normal range
and the likelihood of a reversal increases. Some traders prefer to use
piecewise linear reaction functions that allow for the fixed-size
maximum position to be held for a wider range of the signal. The
reaction function can be easily parametrized to optimize its slopes and
widths, but one must be careful not to overfit parameters, especially if
they are fitted to a single time series. It is also easy to modify a reaction
function that changes its sign at the extremes when the momentum
strategy is replaced with a mean-reverting strategy. This would
effectively blend momentum and value into a single strategy.
Similar reaction functions can also be applied to carry and carry-
momentum signals and these strategies can be optimized by using a
reaction function with an inflection point. In the case of carry, the
existence of an inflection point is consistent with the behavior of
inventory hedgers. When the negative carry is too large, then most of
the storage is already full and there is nothing left for inventory traders
to hedge. Likewise, when the backwardation is extreme, most short
futures held by inventory hedgers are already covered. Therefore,
beyond a certain threshold the pressure on price from storage hedgers
starts waning and other corrective mechanisms may play a larger role,
causing the market to mean-revert. For example, extreme contango
could force producers to shut down drilling, and extreme
backwardation may lead to product substitution among consumers.
The reaction function attempts to quantify an important rule of thumb
in oil trading, where short-term momentum is better combined with
mean-reversion.
So far, we have only considered trading signals that are based on the
information contained in prices, and all strategies were designed to be
entirely rule-based and executed systematically without any human
intervention. As we have already seen, the power of good trading often
lies in blending signals of different natures, and also in letting humans
contribute to the decision making. In the next chapter, we look at how
some systematic concepts can be improved with non-price fundamental
and flow information, which require additional inputs from the
discretionary trader.

References
Bakshi, G., Gao, X., & Rossi, A. G. (2019). Understanding the sources of risk underlying the cross
section of commodity returns. Management Science, 65(2), 619–641.
[Crossref]

Boons, M., & Prado, M. P. (2019). Basis-momentum. The Journal of Finance, 74(1), 239–279.
[Crossref]

Daskalaki, C., Kostakis, A., & Skiadopoulos, G. (2014). Are there common factors in individual
commodity futures returns? Journal of Banking and Finance, 40, 346–363.
[Crossref]

Fernandez-Perez, A., Frijns, B., Fuertes, A.-M., & Miffre, J. (2018). The skewness of commodity
futures returns. The Journal of Banking and Finance, 86, 143–158.
[Crossref]

Koijen, R. S. J., Moskovitz, T. J., Pedersen, L. H., & Vrugt, E. B. (2018). Carry. Journal of Financial
Economics, 127, 197–225.
[Crossref]

Lux, H. (2003, February 1). What becomes a legend? Institutional Investor.


Miffre, J. (2016). Long-short commodity investing: A review of the literature. Journal of
Commodity Markets, 1(1), 3–13.
[Crossref]

Szymanowska, M., De Roon, F., Nijman, T., & Van Den Goorbergh, R. (2014). An anatomy of
commodity futures risk premia. The Journal of Finance, 69(1), 453–482.
[Crossref]

Till, H. (2022, Winter). Commodities, crude oil, and diversified portfolios. Global Commodities
Applied Research Digest, 7(2), 65–74.

Tully, S. (1981, February 9). Princeton’s rich commodity scholars. Fortune.

Weymar, F. H. (1965). The dynamics of the world cocoa market. Ph.D. Thesis, Massachusetts
Institute of Technology.

Footnotes
1 See Weymar (1965). In addition to empirical study of cocoa prices using fundamental data,
Weymar developed one of the first theoretical models of storage, which links prices and
inventories via coupled differential equations.

2 The fund’s trading roster included the names of Paul Tudor Jones, Louis Bacon, and Bruce
Kovner, who subsequently became trading legends in their own right, creating multi-billion
dollar hedge funds known to be among the largest commodity traders. Jones founded Tudor
Investment Corporation, Bacon is the founder of Moore Capital Management, and Kovner
established Caxton Associates. More details about the history of Commodity Corp. are provided
in Tully (1981) and Lux (2003).

3 While the literature on long-short commodity risk premia portfolios is broad, we highlight
the following work by Szymanowska et al. (2014), Daskalaki et al. (2014), Fernandez-Perez et
al. (2018), Bakshi et al. (2019), and Boons and Prado (2019). See also Miffre (2016) for a
comprehensive survey of empirical results and additional references on long-short commodity
portfolios.

4 For brevity, when defining the trading signal with the sign function we assume that sign(0) =
+1, which ensures that either a long or a short position is held on each trading day.

5 RBOB stands for Reformulated Blendstock for Oxygenate Blending. It represents a primary
blending component used to create the finished gasoline product. Even though RBOB does not
represent the finished product, it is widely used as the most liquid benchmark for gasoline
prices in the futures market.

6 In general, the term information ratio is used to measure the performance of the strategy
relative to a certain benchmark. It is analogous to the conventional Sharpe ratio if the risk-free
return is replaced with the return on the benchmark. Since trading futures does not require
much initial capital, the benchmark for most strategies is typically set at zero. Therefore, here
the information ratio is simply defined as the ratio of the strategy’s annualized return to its
annualized volatility, and, for convenience, we use the term information ratio interchangeably
with the Sharpe ratio.

7 The empirical performance of the broad cross-asset carry portfolio is presented in Koijen et
al. (2018).

8 Till (2022) showed that a dynamic strategy that replaces a conventional 60–40 equity-bond
portfolio with a 30–40–30 equity-bond-oil futures portfolio when oil carry is positive
significantly outperforms the static 60–40 portfolio.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
I. Bouchouev, Virtual Barrels, Springer Texts in Business and Economics
https://doi.org/10.1007/978-3-031-36151-7_6

6. Quantamentals
Ilia Bouchouev1
(1) Pentathlon Investments, LLC, Westfield, NJ, USA

Trading time spreads is an example of a quantamental trading style


where quantitative curve signals are combined with fundamental
information about inventories and financial flows.
A discretionary overlay is particularly important in identification of
fundamental regimes perceived to be favorable for the performance
of systematic signals. One important case study is a regime-
dependent WTI-Brent convergence strategy.
Many energy assets represent a real option on the output-input
spread. Asset owners hedge profit margins, which induces spread
mean-reversion. A portfolio of mean-reverting energy spreads is akin
to a statistical arbitrage strategy.
The analysis of flows and positioning is another important strategy
overlay. It is effective in identifying crowded trades that can be used
as a reversal indicator to optimize position sizing and risk
management.

6.1 Trading Curve and Convexity


The systematic risk premia presented in the previous chapter are based
on rather generic concepts that can be applied across many different
markets and traded within a broader portfolio of commodity futures.
While risk premia strategies have proven to work better for harder-to-
store energy commodities, adding other commodities to the portfolio
brings a much-needed diversification that reduces the overall risk of
the strategy. In such portfolios, distinct properties of individual
commodities are largely ignored. In this chapter, we focus instead on
important idiosyncratic features of oil price behavior and show how a
generic cookie-cutter strategy can be turned into a more elaborate oil-
specific bet.
Many oil traders do leverage the ideas of risk premia, but applying
these strategies to a single asset without the diversification benefits of a
broader portfolio would be too risky. To make the strategy viable on its
own, systematic price-based signals are often enhanced with additional
information about market fundamentals and flows. Fundamental and
flow data, however, are much noisier, more difficult to systematize, and
better left to a discretionary interpretation. In such trading, a human
equipped with a machine tends to do better than what either a machine
or human can achieve alone. An application of quantitative signals to
fundamental and flow data, or, more generally, any strategy based on
quantitative signals with a fundamental overlay, is described as
quantamental trading.
It may come as a surprise, but professional oil traders do not like to
speculate on the price of oil. The oil price is driven by too many factors
that are difficult to get a handle on. In the short run, the price does not
have any well-defined boundaries, and the cost of being wrong could be
catastrophic. To keep the risks contained, oil specialists instead focus
their attention on trading futures spreads. The spreads are constructed
to isolate the desired exposure by eliminating a portion of the risk that
the trader is not willing to bet on. While managers of broad portfolios
reduce risk by diversification, oil traders rely on trade structuring and
supplementary fundamental information for risk mitigation. Spreads
are significantly less volatile than outright prices, and to generate the
same amount of profits, spread strategies must be leveraged and traded
in larger volumes. A very significant portion of the trading volume in oil
futures represents components of various spread-trading strategies.
The single most important and the most widely traded spread in the
oil market is the spread between the first- and the second-nearest
maturity futures. This is the spread that determines the own rate of
interest, the convenience yield, the roll yield, and the commodity carry.
In the previous chapter, we have used the spread as an indicator for
directional price trading. Now, we consider the spread itself as a trading
instrument. Trading short-term time spreads is not a trivial task. It is
driven by a complex interplay between fundamentals of supply and
demand in the physical market and financial flows in the futures
market, as traders must roll their positions before the expiration of the
contract. The previously discussed episode of negative oil prices is
perhaps the most famous example of such an interplay.
Much as price traders use the slope of the curve for decision
making, curve traders often start their analysis by looking at the
convexity of the futures curve. The idea is to use the same concept of
carry but to apply it in a higher dimension. Convexity is viewed by
practitioners as the spread between two adjacent futures spreads. For
example, for the first three nearest maturity contracts, the convexity
can be defined by

In other words, convexity is simply the carry of the carry.


As discussed in the previous chapter, in the absence of any hedging
distortions, the natural shape of the futures curve is either concave
contango or convex backwardation. This follows directly from the
theory of storage. The role of storage is to smooth out short-term
variations in supply and demand and to dilute the magnitude of the
fundamental imbalance by shifting some of its burden forward. Storage
literally buys time. Therefore, nearby time spreads that bear the
strongest immediate effect of fundamental imbalances are likely to be
wider than forward time spreads, as the futures curve is expected to
gradually normalize.
Figure 6.1 illustrates monotonicity of time spreads for the futures
curve in the state of concave contango. Similarly, in the state of convex
backwardation, which would be the mirror image of Fig. 6.1, time
spreads are positive and monotonically decreasing.
Fig. 6.1 If the futures curve is in concave contango then time spreads are expected to roll
down
Since carry, or the slope of the futures curve, has proven to be useful
in predicting the direction of the price, one can speculate that carry of
the carry, which is the curve convexity, could be relevant for the
dynamics of the curve slope. In other words, if the first derivative of
futures with respect to maturity has some predictive ability for the
price level, then the second derivative may be useful in predicting the
direction of the slope. For example, if the futures curve is in a state of
concave contango, as pictured in Fig. 6.1, then the spread carry strategy
is to sell time spreads in anticipation of them rolling down the curve.
Likewise, if the curve is in a state of convex backwardation, then the
spread carry strategy is to buy time spreads, which are expected to roll
up.
Another way to think about the convexity-trading mental model is
to assume that the fair value of forward time spreads is determined by
the spot spread, which contains the most up-to-date information about
the current state of inventories. Since inventories are slow to adjust,
forward spreads are expected to roll towards the value of the spot
spread. Therefore, it might be beneficial to buy forward time spreads if
they trade below the prompt spread and to sell forward time spreads if
they are above the prompt spread. It should be noted that such a simple
formulation of the convexity strategy can only apply to commodities
that do not exhibit strong seasonality, such as crude oil.
Even though convex backwardation and concave contango can be
viewed as two fundamentally normal states of the futures curve, in
practice, the curve is distorted by imbalances between buyers and
sellers in the futures market. The largest imbalance in the spot time
spreads is caused by financial investors holding large quantities of the
nearest-maturity futures, which are held as a substitute for owning the
physical commodity. Every month, prompt futures positions must be
rolled ahead of their expiration, which is done by selling the time
spread. To the big disadvantage of commodity investors, exact roll
schedules for main commodity indices and exchange-traded funds
(ETFs) are publicly available and well known to the market. Such
transparency incentivizes professional traders to sell time spreads
ahead of anticipated rolls and buy them back when spread values are
expected to weaken from the negative rolling pressure.
When the market is oversupplied and is in a state of contango, then
the pressure from investor rolls tends to exacerbate the degree of the
curve concavity. In fundamentally weak markets, the combination of
such rolls and storage capacity constraints can lead to the situation
known as a super-contango, when the prompt futures contract
dislocates and falls substantially below the contract with the next
maturity.
In the case of backwardation, the rolling pressure on the prompt
spread may result in a humped futures curve, which corresponds to the
case of concave backwardation. The hedging pressure on spreads can
also be positive, which occurs when inventory hedgers and refineries
roll their short futures positions. This can lead to the scenario of convex
contango. However, such episodes occur rather infrequently because
commercial hedgers are not driven by any fixed roll schedule, which
gives them more flexibility on when to roll. In fact, many refineries and
storage hedgers view rolling their short futures as a way to take
advantage of the more rigid schedule that must be followed by long
futures holders, such as commodity index investors. Finally, the
scenario of super-backwardation arises during unexpected supply
disruptions when refineries have to pay up for spot barrels,
overwhelming the negative hedging pressure from investor rolls.
Various shapes of the futures curve are illustrated in Fig. 6.2. Note
that here we define convexity locally by using the shape of the front-
end of the futures curve, while describing backwardation and contango
by comparing the spot price to long-dated futures.

Fig. 6.2 The impact of roll hedging pressure on the shape of the futures curve

The structural hedging pressure caused by investor rolls makes the


returns of buying and selling short-term WTI time spreads highly
asymmetric. The odds of making money are generally higher by
systematically selling prompt WTI time spreads than buying them. To
illustrate, consider a simple strategy of selling the spot spread and
rolling it to the next spread five business days prior to expiration.1 Over
the thirty-year period starting from 1993, such a naïve strategy would
have generated a Sharpe ratio of 0.50. The strategy would have done
even better if one started it in 2005, which we previously earmarked
for the start of the regime of normal contango. Not surprisingly, the
strategy of selling WTI time spreads became the darling of Wall Street.
It was packaged into investable indices and sold to investors as an
alternative risk premium specific to WTI market.
To illustrate the likely source of such an incredibly strong bias for
the static strategy, we decompose the historical P&L of the strategy by
the shape of the futures curve. Figure 6.3 splits the cumulative rolling
strategy P&L into four constituents, determined by the slope of the
futures curve and its convexity.

Fig. 6.3 Decomposition of short time spread strategy by slope and convexity. The number in
parenthesis indicates the percentage of time that the futures curve was in a given regime

The decomposition of P&L with respect to a certain explanatory


variable is often referred to as fractionation analysis. For the short time
spread strategy the fractionation analysis reveals that strategy profits
were generated mostly when the curve was concave, either when it was
in its normal state of concave contango, or in a state of concave
backwardation where flow distortions lead to a humped curve. A
significant contribution to positive P&L during a somewhat abnormal
state of concave backwardation indicates the importance of financial
rolls, the phenomenon that accelerated since the beginning of
financialization. In contrast, when the curve was convex, then selling
time spreads was a losing trade, irrespective of contango or
backwardation, but the losses were relatively small. It is interesting to
see that it is the convexity of the curve rather than its slope is the main
driving force for time spreads.
These observations can be combined into a dynamic long-short
spread strategy based on the curve convexity. Let us define the prompt
time spreads as S1 − 2, t = F(t, T1) − F(t, T2). Then the trading signal can
be specified as

and the spread carry strategy is to sell and roll the prompt spread when
the curve is concave and to buy it when the curve is convex.
This long-short strategy further improves the cumulative strategy
performance, mostly from having long spread positions, prior to the
beginning of financialization. Since 2005 nearly all profits from the
spread carry strategy came from selling spreads when the futures curve
was concave. Is the strategy performance then entirely driven by
financialization, or some other fundamental factors, such as rapid
growth of US shale production, also played a role in the structural
rolldown of WTI time spreads? To assess the impact of fundamental
factors, we next look at another important example of fractionation
analysis and decompose P&L of the time spread strategy by the level of
inventories.

6.2 Time Spreads and Inventories


An analysis of time spreads behavior versus inventories is motivated by
the theory of storage, presented in Chap. 3. In that chapter, we have
only looked at contemporaneous correlations between time spreads and
inventories. Now we address the more difficult question of whether the
level of inventories have any predictive power for the direction of time
spreads.
We fractionate P&L of the same short WTI time spread strategy by
the level of Cushing inventories. As discussed in Chap. 3, one can also
use PADD2 and overall US inventories as alternatives. In practice,
traders tend to look at all three inventory metrics to get a broader
assessment of the fundamental state of the market. In this analysis, we
compare the levels of Cushing inventories, which are reported weekly,
to P&L of the strategy during the following week. As before, we use the
normalized storage capacity utilization as our preferred measure of
inventories to better assess proximity to the upper storage boundary.2
The results are shown in Fig. 6.4.

Fig. 6.4 P&L of short WTI time spread strategy (2011–2022) split into deciles based on
utilized storage capacity in Cushing
At first glance, these results may look surprising. There is no visible
correlation between forward P&L of the short time spread strategy and
the previously observed level of inventories. The conventional theory of
storage only explains contemporaneous correlation between time
spreads and the level of inventories. However, contrary to a common
perception, neither inventory level nor recent changes in inventories
show any strong predictive power for the direction of time spreads.
In fact, a deeper econometric analysis shows that the reverse is true,
and time spreads have some predictive power for the future change in
inventories.3 This happens because storage traders often lock in
contango for several months ahead when the steepness of the curve
exceeds the net cost of storage for a longer time period. Consider a
trader who buys a heavily contangoed spread between one-month and
six-month futures, then takes a physical delivery on the long futures
contract and keeps oil in storage for six months. These barrels increase
oil inventory next month and decrease inventory in six months, when
oil is pulled out of storage and delivered against a short futures
contract. Therefore, the storage decision, which is made based on the
slope of the futures curve today, affects the level of inventories six
months from now. Moreover, the level of inventories in six months
depends on all prior storage decisions made based on the prevailing
shape of the futures curve.
P&L fractionation by inventory level does highlight an important
feature of the spread strategy, the impact of inventory boundaries.
When inventories drop to a low level then selling time spreads becomes
particularly dangerous as the probability of an upside squeeze
increases. In such environments, the strategy generated a small loss,
shown in the leftmost bucket of Fig. 6.4. A much larger loss in the case
of high inventories is nearly exclusively explained by an isolated event
that occurred when oil price and time spreads recovered from super-
contango levels set at the onset of the Covid-19 pandemic. In this
episode, discussed in detail in Chap. 3 and illustrated by Fig. 3.​11, the
market was extrapolating a trend in inventories that would have
breached the maximum available storage capacity. Since the boundary
on storage capacity cannot be violated, supply and demand have no
choice but to adjust, and the inventory trend must stop and reverse.
This boundary phenomenon resembles the structure of the reaction
function, presented in the previous chapter. The trend in inventories
and time spreads can continue up to a certain inflection point, beyond
which the presence of the hard boundary induces the reversal.
One lesson that a quantamental trader can learn from the
fractionation analysis is that it might be better not to trade the time
spread strategy when Cushing inventories approach either boundary.
The proximity of boundaries clearly introduces additional uncertainty
and makes spread trading particularly risky. Unfortunately, the useful
insight obtained from fractionation analysis is sometimes abused by
systematic traders, who are often driven by an overarching desire to
improve the historical backtest. Imagine a conditional systematic
strategy which is short a time spread for structural reasons but the
strategy switches to a long spread position when inventories are less
than 20% or more than 80% of the storage capacity. Obviously, such a
rule-based strategy conditioned on inventories produces a substantial
improvement to the backtest, as the two negative buckets in Fig. 6.4 are
flipped from losses to gains. If the purpose of this book was to sell a
systematic strategy to an investor, we might have even presented an
optimized backtest for such a systematic strategy. Our goal, however, is
to describe how the sausage is made, which is what the fractionation
analysis reveals.
The fractionation analysis highlights the danger of data mining and
the sensitivity of any systematic strategy to additional parameters,
much like an optimized momentum crossover strategy. For a
quantamental trader, however, such analysis is a valuable tool for
decision making. It confirms the structurally short bias of the naïve
static strategy that results from the joint impact of financial rolls and
the growth of US shale production. When the two forces work in
tandem, the spreads steadily roll down. However, when inventories
decrease towards particularly low levels, then fundamental and flow
forces pull the spread in opposite directions. During such periods, the
strategy performance deteriorates. If a trading strategy makes money
when two primary driving factors point in the same direction, but does
not lose much when factors diverge, then it is definitely an attractive
strategy. In addition, the analysis emphasizes the important role of
storage boundaries.
In this section we only illustrated the curve convexity and inventory
analysis for WTI futures. While similar techniques can be used for other
energy futures, the strategy implementations are more nuanced, which
makes it difficult to apply these concepts across broader systematic
portfolios. Recall from Chap. 3 that the value of the time spread is
driven by the relative probabilities of downside and upside squeezes.
For example, in contrast to crude oil, time spreads for refined products
have an upside skew. In the case of many refined products,
fractionation of spread P&L by inventories often shows more
discernible impact of zero inventory boundary, when time spreads tend
to appreciate. The downside for time spreads on refined products is
more limited as refineries usually can cut product runs quickly in
response to rising inventories and declining profit margins. The upside,
on the other hand, does not have a well-defined boundary due to
potentially unlimited convenience yield if the market runs out of an
essential product during its peak consumption period, such as heating
oil during cold weather. The statistical analysis of refined products
inventories is, however, more challenging, due to their strong
seasonality, which significantly limits the quantity of relevant historical
data.
The case study of inventories is only one of many tools that can be
utilized by a quantamental trader. Additional insights can be gained
from a traditional fundamental analysis that involves so-called barrel
counting. Since the spreads are only contemporaneously correlated to
inventories, to forecast where spreads will be in the future, one needs
to know where inventories will be in the future. Such forecasting of
future inventories is based on counting physical barrels of oil that are
being moved by pipelines and ships and by extrapolating the data
forward using estimates for production and refinery inputs. While this
analysis is outside the scope of the book, a good quantamental trader
would always attempt to incorporate such information in a
discretionary manner to improve a rule-based trading strategy.
We next discuss another popular quantamental strategy that trades
the most important locational spread between two primary oil
benchmarks, WTI and Brent.

6.3 WTI-Brent Accordion


So far, we have been focusing primarily on the WTI futures contract for
several reasons. This contract has the deepest price history, which is
particularly useful for long-term analyses. WTI also serves as a primary
anchor for the large OTC hedging market dominated by independent
North American oil producers. And, perhaps most importantly, the
physically deliverable nature of the WTI contract allows futures traders
to convert their financial bets into physical barrels when the futures
contract expires. Such conversion, which can occur in Cushing,
Oklahoma, ties the pricing of the WTI contract to supply and demand
for inventories in the regional storage hub.
The second equally important financial oil contract is Brent, which
represents the basket of crude oils originating from the North Sea.4
Unlike WTI, Brent is a waterborne contract, which allows barrels to be
easily shipped anywhere in the world. Most of the world’s physical
barrels are priced off Brent futures. Given the global scope of this
market, storage-based trading strategies are more difficult for Brent, as
these strategies work better for isolated storage hubs. To trade such a
strategy globally, one would need to count barrels held in storage and
on ships around the world, which is rather difficult to do in real time.
While many WTI strategies are linked to storage, a significant portion of
Brent trading revolves around its location and the connectivity of this
benchmark to other pricing locations around the world.
WTI and Brent are linked to each other by the economics of the
transportation arbitrage. Chartering a ship or leasing pipeline space
gives traders a real option to buy oil at a cheaper price in one location
and sell it at a higher price at another location, provided that the spread
between the two prices covers the cost of shipping. To monetize the
real optionality, physical traders attempt to sell the locational spread
high and buy it back low, having an ability to ship physical barrels as a
backstop. Such spread-trading behavior is analogous to delta hedging of
a long financial spread option, the topic that we will discuss in Chap. 13.
The existence of the physical arbitrage does induce some degree of
mean-reversion in the spread behavior, which prevents the two
benchmarks from diverging too far from each other. This makes the
WTI-Brent spread a suitable candidate for a convergence strategy, or
using the terminology of the previous chapter, it is the value risk
premium but applied to the locational spread.
As we have seen in the previous chapter, value strategies do not
work well for the price of oil due to the persistent headwind of the
negative carry, but they work much better for fundamentally linked
energy spreads. In cross-product spread trading, carry is less of a factor
because two forward curves tend to have similar shapes, where carry
on one leg of the spread is partially mitigated by an opposite carry on
the second leg. With less punitive cumulative impact of carry, one has
more time to wait for convergence to occur after selling on rallies and
buying on dips for the spread. Such a spread convergence strategy
implemented purely in the futures market is known as a paper
arbitrage.
The thesis of the paper arbitrage strategy is to trade ahead of an
anticipated behavior of physical traders. Since the economics of
shipping is generally known, a financial trader can attempt to mimic a
physical arbitrage strategy in the futures market, but without the
delivery of physical barrels. The financial convergence strategy
generally works well as long as the physical arbitrage flows remain
uninterrupted. The risk of running a spread arbitrage strategy on paper
is, of course, an unexpected disruption in physical flows that can no
longer keep the spread contained. While the physical trader has a
hedge in place secured with physical barrels, the financial trader does
not.
Spread convergence strategies of this nature, which generate steady
income but with a low probability of a large loss, are difficult to trade
purely on a systematic basis. The most important discretionary overlay
in the spread convergence strategy is the decision of when not to trade.
Spread strategies are highly dependent on a particular fundamental
regime. For systematic traders, identifying and forecasting such
regimes without having access to the physical market is rather difficult,
as the entire petroleum infrastructure is constantly adjusting to new
developments in supply and new trends in demand.
To illustrate this for the spread between the two primary oil
benchmarks, WTI and Brent, we identify three drastically different
fundamental regimes, illustrated in Fig. 6.5.
Fig. 6.5 Three fundamental regimes for WTI-Brent: the regime of the USA being a large oil
importer, the regime of shale growth and logistical bottlenecks, and the regime of global
interconnectedness
The first regime covers several decades of US dependency on oil
imports when the country consumed substantially more oil than it
produced. To incentivize non-US producers to ship barrels to the USA
versus alternative destinations, WTI at that time was trading at a small
premium relative to Brent. By and large, this premium reflected a
relatively stable cost of shipping oil from the Middle East to the US Gulf
Coast. During this period, sufficient pipeline capacity was also available
to further distribute imported oil to US inland refineries for processing,
or to Cushing for storage. This low-volatility price regime for the WTI-
Brent spread started to change with the rapid growth of shale
production, which eventually made the USA energy self-sufficient.5
While US shale oil production grew rapidly, the infrastructure
needed to support such growth lagged. This marked the beginning of a
fundamentally new regime for WTI-Brent behavior, the regime of
logistical bottlenecks, characterized by bouts of extreme volatility. As
US domestic production exceeded domestic consumption, oil had to go
to storage, which quickly saturated storage capacity near production
centers. Moving oil to alternative storage locations around the country
was also constrained by an insufficient pipeline infrastructure that
could not keep up with the fast pace of production growth. Even when
oil found its way from inland production centers towards the coasts,
excess barrels could not be exported, as at that time crude oil exports
were prohibited by US regulations. Running the financial convergence
strategy during these years would have been a disaster. The spread was
no longer capped by any fundamental boundaries. WTI and Brent
traded as two disjoint prices with fundamental connections between US
and global oil benchmarks temporarily broken.
The ban on US crude oil exports was lifted in December 2015,
starting the third and the current regime of global interconnectedness.
The economic link between the two primary oil benchmarks has been
re-established. Moreover, exports and imports started to flow in both
directions driven by the relative prices of the two regional benchmarks.
In the absence of any regulatory restrictions, any abnormal deviations
in the spread between the two prices beyond the cost of shipping can
be corrected by physical arbitrageurs. As a result, the behavior of WTI-
Brent became more oscillatory and mean-reverting. Given its constant
compressions and expansions, the convergence strategy was dubbed by
traders the WTI-Brent accordion.
The actual economics of the physical arbitrage is rather intricate.
Most importantly, the cost of waterborne shipping is highly uncertain,
as it fluctuates along with volatile market prices of freight futures. As
part of an arbitrage trade, a physical trader must not only lock in the
spread between WTI and Brent, but also fix the freight price. Moreover,
the impact of regional bases must be carefully taken into consideration.
Such bases reflect transportation costs between specific inland
production, refining centers, and export-import hubs. For a physical
trader involved in the shipping of oil, the precise calculation and
hedging of all moving parts is required to secure riskless profits. A
financial trader usually takes a shortcut. Instead of calculating the true
fundamental arbitrage boundary, which depends on freight and
regional bases, the futures trader often approximates it with a
statistical boundary.
The simplest way to define a paper arbitrage strategy is to take
positions against large statistical deviations of the spread from its
historical norm. The signal can be specified using moving averages, as
was done in the previous chapter for momentum strategies. Specifically,
if St is the spread between two futures, then a financial trader could
take a contrarian or value position when the spread deviates by more
than a certain threshold ε from its n-day moving average MAt (St; n)6:

(6.1)

The purpose of the moving average is to approximate the equilibrium


level of the spread, which might be changing dynamically to reflect
ongoing changes in the physical infrastructure. The threshold ε is
designed to function as a buffer that covers uncertainty in freight and
other logistical costs. This specification means that the paper arbitrage
strategy is implicitly betting on some mean-reversion in freight rates.
The risk is that if the freight rate spikes, then the paper arbitrage
strategy will generate a trading signal even though the physical export-
import arbitrage remains closed due to the higher cost of freight. A
more advanced trader can make ε explicitly depend on market prices of
freight futures. The only other parameter in this strategy is the length
of the lookback period, n.
Figure 6.6 illustrates the moving average of the spread for the third-
nearby WTI and Brent futures along with a daily deviation from such a
moving average, which defines the trading signal. This simple WTI-
Brent convergence strategy effectively trades the residuals by taking
positions against this deviation.

Fig. 6.6 The deviation of WTI-Brent (m3) from its moving average

This strategy has performed quite well since the beginning of the
third regime of global interconnectedness, reaching a Sharpe ratio of
nearly 1.0 with a fairly stable combination of model parameters.7 As
before, our preference is to avoid showing backtests for quantamental
strategies as it only creates an urge to improve them by introducing
additional conditions and filters. These filters can be identified by
fractionation analysis with respect to various fundamental factors, an
example of which was presented in the previous section. More
importantly, these strategies thrive only during specific fundamental
regimes, the identification of which is largely discretionary. Since oil
regimes do change rather frequently, any backtest for a quantamental
strategy can quickly become obsolete.
Certain regime features, however, are likely to be more long-lasting
and even irreversible. For example, increasing global
interconnectedness of energy markets is unlikely to disappear once
appropriate infrastructure is built. This latest regime opened a new
trading opportunity not only for WTI-Brent convergence strategies, but
also for many other fundamentally linked petroleum spreads. It allowed
the construction of a much broader mean-reverting spread portfolio
that can diversify idiosyncratic risks associated with individual spreads.
In the equity markets, such portfolios fall into the category of statistical
arbitrage, or stat-arb, strategies. In the next section, we explain how to
construct its analogue in the energy market.

6.4 Cointegration and Energy Stat-Arb


The mathematical foundation behind spread convergence strategies
and the stat-arb portfolio is based on the concept of cointegration. As
described in many statistical textbooks, cointegration between two
time series is often compared to a person walking a dog on a leash. The
two can wander around, but ultimately, they cannot deviate too far from
each other and tend to move together. In more precise statistical
language, a cointegration property means that a certain linear
combination of time series is stationary. If the process is stationary,
then it has a finite variance that does not allow it to move too far from
its mean. Thus, it exhibits a mean-reverting behavior. Oil prices are not
stationary, but many petroleum spreads are.
Cointegration should be distinguished from the better-known
concept of correlation. Correlation is only defined for stationary
variables, so in most cases applying it to the price of oil is rather
meaningless. Instead, it is applied to price changes or to percentage
returns, as taking the difference between prices generally makes the
time series stationary. For example, measuring correlation between the
spot-futures spread and the normalized inventory storage capacity, as
defined in Chap. 3, is meaningful because both time series are
stationary.
When applying correlation analysis to price changes, one effectively
eliminates the entire memory of price levels. While this might be
acceptable for some financial markets, in commodities the memory
contains valuable information about the long-term equilibrium price
level driven by supply and demand. Therefore, cointegration is a
preferred starting point for analyzing commodity spreads, as the
method applies directly to price levels rather than price changes or
investment returns. Only when prices are cointegrated does measuring
the correlation between them become statistically meaningful.
The concept of cointegration extends to the multivariate case. For
example, it can be applied to the vector of petroleum futures across
various geographical regions. Similarly, the vector is said to be
cointegrated if there exists a linear combination of prices that is
stationary or mean-reverting. In other words, if multiple assets are
connected with each other via various fundamental linkages, they all
tend to fluctuate around some equilibrium state for the entire system
but eventually they gravitate towards such an equilibrium state. It is
not our objective to provide rigorous statistical details for various
cointegration tests. The topic is well covered in many statistics books,
and most statistical packages have cointegration tests built-in.8
While one can indeed construct a trading strategy based on
multivariate cointegration relationship among many petroleum futures
that fluctuate around some natural equilibrium for the entire system,
practitioners tend to simplify the problem and look instead at a
portfolio of energy spreads. Like the WTI-Brent accordion strategy,
many other petroleum spreads exhibit stronger mean-reversion only
during certain fundamental regimes. The prevalent regime of global
interconnectedness is particularly favorable for spread convergence
strategies which can be combined into a broader portfolio of energy
pairs. The growth of shale production created new fundamental
linkages, and the removal of the US ban on crude oil exports
reconnected US petroleum markets with the rest of the world. For
example, shale oil spurred the rapid growth of US natural gas liquids,
which became an important alternative feedstock for the petrochemical
complex, competing with other refined products around the world.
The idea behind trading a diversified energy stat-arb portfolio is
again based on the concept of monetization of the real optionality
embedded in the physical asset. Pipeline and oil tankers are locational
spread options. A refinery is an option on the spread between a basket
of refined products and another basket of crude oil inputs. A storage
tank is not only an option on time, but it also represents a valuable
substitution option to blend different grades of crude oil. The owners of
these real options are driven by economic incentives to lock in the
appreciated value of an asset when the spread that drives the profit
margin widens. Likewise, when the value of the spread falls, the asset
owner can monetize an optionality by reducing the productive capacity
of the asset and taking profits on financial hedges.
In contrast to popular equity stat-arb strategies, where the secret
sauce typically lies in technical details of the trading signal, the keys to
the successful implementation of the strategy in the energy market are
different. For the most part, regardless of whether one uses a simple
mean-reverting rule, such as (6.1), or more elaborate quantitative
signals, the performance of a strategy is driven by other factors. What
matters more for the energy stat-arb is the selection of what spreads to
trade and when to trade them. The spreads must be carefully picked
when the trading behavior of physical arbitrageurs is deemed to be
significant relative to the size of the market, and when there are no
impediments constraining hedging flows that keep the pair together.
As in many other quantamental energy strategies, the composition
of a stat-arb portfolio is likely to change dynamically, as the energy
infrastructure always evolves. However, even a snapshot example of
such a portfolio constructed at the time of writing this book might be a
useful starting point for its future iterations. Let us first describe the
composition of a sub-portfolio made up of different grades of crude oil,
which is illustrated in Fig. 6.7.
Fig. 6.7 Physical arbitrage relationships actively traded within a global crude oil portfolio
This diagram represents oil spreads that are typically traded by a
physical arbitrage desk. It only shows pairs that meet certain minimum
liquidity thresholds, measured by volumes and open interest, and the
convergence of which is not impeded by regulatory or other
restrictions. For example, one notable exclusion is a recently developed
oil contract listed on the Shanghai International Energy Exchange
(INE). Even though it has quickly grown to become the third most
actively traded oil futures after WTI and Brent, the contract has not yet
been fully integrated into global arbitrage trading, as local regulatory
restrictions may preclude its convergence to other oil benchmarks.
In North America, oil grades are traded as differentials to the WTI
futures contract, which is often associated with the Cushing storage
hub. Unfortunately, many traders do not realize that the reference to
the name WTI in the futures contract is a misnomer. As it is clear from
the name WTI, which stands for West Texas Intermediate, the contract
is expected to represent oil produced in West Texas. However, oil that is
delivered against the WTI contract could be very different from that
produced in West Texas. Any blend of various grades of oils, whose
chemical characteristics, such as density, viscosity, and sulfur content,
are within a certain range, can be delivered against the futures contract.
These blends are collectively referred to as domestic sweet (DSW) oil.
Not surprisingly, the business of blending has become such a valuable
real option for the owners of storage tanks in Cushing. If a futures buyer
decides to take physical delivery of a WTI futures contract, then the
exact origin of such a blended barrel delivered by the seller may not
even be known to the buyer.
The price of a genuine WTI barrel is better represented by the WTI
Midland contract, which explicitly references one of its main
production centers. The same oil barrel also trades via the WTI
Houston contract, whose price includes the transportation cost to US
Gulf Coast export terminals.9 The three WTI contracts at Cushing,
Midland, and Houston form a fully integrated triangle of cointegrated
prices with corresponding spreads fluctuating around their pipeline
tariffs. Such linkages only stabilized after a sufficient pipeline
infrastructure was built that connected shale production areas to
storage tanks and export terminals. In addition, WTI, which represents
a light sweet barrel of oil, is also linked to prices of heavy sour grades.10
One such grade, Western Canadian Select (WCS), is the benchmark for
oil produced in Canada that can be shipped to Cushing for blending to
produce a futures-deliverable grade. Alternatively, WCS can be shipped
directly to US Gulf Coast refineries, where it competes with Mars, which
is the benchmark for US heavy oil produced in the Gulf of Mexico.
Internationally, as discussed above, WTI is connected to Brent via
the transportation arbitrage. In Asia, the primary oil benchmark is
Dubai, which represents another grade of heavy oil. It trades at a
differential to Brent that reflects not only the cost of transportation, but
also a discount for the lower quality of heavy sour oil relative to light
sweet. Likewise, the WTI-Dubai spread captures both transportation
and quality differentials. At the same time, Mars-Dubai represents the
spread between two similar grades of oil and this spread reflects
mostly the cost of transportation. Two other European oil spreads to
Brent that play an important role in the global oil ecosystem include
dated Brent, which differs from Brent futures in timing for the physical
delivery of barrels, and Urals.
The construction of the petroleum stat-arb portfolio does not stop
with crude oil. One can picture the petroleum complex as a large
shopping mall, where all crude oils trade on the first floor, and refined
products for each brand of crude oil trade on the second floor. One can
walk on either floor and shop by product, or, alternatively, one can shop
for both crude oil and its affiliated refined product in the same branded
store. On the one hand, the price of each refined product is linked to the
local crude oil by the regional economics of refining. On the other hand,
all refined products, like all crude oils, are also connected with each
other via the transportation arbitrage. For example, US ultra-low-sulfur
diesel (ULSD) impacts the profitability of US refineries, and its
European equivalent, called gasoil, is crucial for the profitability of
European refineries. At the same time, ULSD and gasoil are also linked
via the shipping arbitrage, as they are comparable refined products.
Similar multifaceted relationships driven by substitution and
transportation options exist for many other refined products.
Figure 6.8 displays the map of the second floor of this petroleum
shopping mall, where for transparency, only three major regional crude
oil benchmarks are shown, among which Brent plays the central role.

Fig. 6.8 Physical arbitrage relationships actively traded within a global refined-products
portfolio

To navigate through this diagram, one can follow contracts by the


type of refined product. In the middle distillate complex, European
gasoil connects with ULSD and Singapore gasoil. In addition, each of
them is linked to corresponding jet fuel markets that trade as basis
spreads to regional distillate benchmarks. In the gasoline complex, the
Eurobob (EBOB) swap contract, which is a European counterpart of US
RBOB futures, and Singapore Mogas swaps, form the triplet of gasoline
products. A similar chain exists for the market for another refined
product, fuel oil. However, it is omitted from the graph, as the fuel oil
market has undergone significant structural changes in 2020 driven by
the introduction of a new sulfur specification, which makes the
historical data required for a cointegration analysis less relevant.
One can further expand the petroleum stat-arb portfolio by adding
natural gas liquids (NGLs) which include ethane, propane, and butane.
NGLs not only have strong fundamental linkages among themselves,
but they also compete as the feedstock for the petrochemical business
with naphtha, another European refined product linked to European
gasoline. Going beyond petroleum futures, one can add natural gas
markets, where in the US alone there are over thirty different pricing
locations that can be grouped together and traded for convergence
towards an equilibrium based on their pipeline connectivity. With the
expansion of liquefied natural gas shipping facilities, US natural gas is
expected to be more closely linked to natural gas prices in Europe and
Asia. Furthermore, natural gas competes against coal for local power
generation, allowing utilities to switch their inputs based on the
relative prices of competing fuels, which could bring coal markets along
with emission credits into the global energy stat-arb portfolio as well.
Trading energy pairs is an example of a complex dynamic system,
which will be revisited in more detail in the next chapter in the context
of broader financial markets. Such systems are characterized by
multiple nonlinear interdependencies. In the world of oil spreads, these
interdependencies are driven by the constantly evolving economic
incentives of arbitrageurs and asset owners that drive their decisions
where to ship barrels and what products to make out of them. For
example, if additional pipelines are built to move oil from US
production areas to the Gulf Coast for exports, then less oil will be sent
to Cushing, decreasing the safety buffer of oil inventories available for
delivery against WTI futures. A single change within the infrastructure
will impact many other spreads tied to WTI. The incentives for
producers and shippers change over time, as local production and
pipeline capacity do not grow synchronously. While production, once it
is established, tends to grow gradually, the takeaway pipeline capacity
moves in steps as new pipelines become operational. As a result, during
some periods production overwhelms the pipeline capacity, but during
other periods production lags, leaving some pipelines unfilled. One
should always carefully distinguish between such fundamentally
different regimes.
As it has been previously highlighted, because of its high cost, the
petroleum ecosystem tends to operate with a just-enough and just-in-
time mentality, leaving very little room for error. Such a highly
optimized complex system is susceptible to periodic extreme events
that could be caused by unexpected disruptions. Consider, for example,
Canadian WCS oil, which is linked to WTI via pipelines, but has little
buffer in terms of spare pipeline capacity. If the pipeline capacity is
sufficient to handle variations in oil flows, then a mean-reverting
strategy generally works. However, in the absence of an adequate
capacity buffer, a single event that disrupts the critical pipeline flow
could cause a downward jump in the WCS-WTI spread since no other
comparable transportation alternatives are readily available to move oil
out of Canada. The spread then must immediately widen to reflect the
economics of the next available pricing tier, transportation by rail,
which is much more expensive than moving oil by a pipeline. Such
jumps will inevitably lead to trading losses in the convergence strategy
for this pair.
The energy stat-arb portfolio is an example of a short volatility
strategy characterized by the negative skewness in its returns.
Strategies of this type tend to generate steady income, but at the
expense of taking a large tail risk. The art of managing the systematic
spread convergence portfolio is in knowing when not to trade
individual pairs. Such warnings are best seen in the fundamental data,
such as the level of inventories. For example, selling elevated refining
margin spreads ahead of a hurricane may be a much safer bet when
refined product inventories are high, but one would be better off
skipping the convergence signal when inventories are low. The
fractionation analysis by inventories presented earlier has proven to be
a particularly useful tool in making such decisions.
In the oil market, physical arbitrageurs and stat-arb traders
mimicking the behavior of asset owners can be viewed as liquidity
providers of last resort. Their incentives are typically to trade in the
direction opposite to systematic traders, who follow trend and carry
signals. To some degree, momentum and carry traders exacerbate
volatility and add fuel to the fire by buying when the price is already
high and selling when the price is low. Asset owners take the other side
of these flows and attempt to extinguish the fire. Despite being on
opposite sides of the trade, trend followers and cross-asset arbitrageurs
can both make money, as they trade different strategies with different
investment horizons. Systematic traders focus on shorter-term
directional price movements, while physical and paper arbitrageurs
trade longer-term relative value spread strategies which are often
structured to be market neutral.
Large negative tail risks coupled with constraints on human
intervention prevent traditional CTAs from fully embracing spread
convergence strategies, as quants often lack the fundamental insights
that are essential for forecasting the likelihood and potential impact of
short-term dislocations. However, with technological advances and the
proliferation of unique fundamental data, quantitative traders are
expanding their participation in petroleum spread trading. Quants can
now use up-to-date inventory data from satellites, track tankers
carrying oil in transit, and even estimate gasoline demand from
downloads of driving maps on mobile devices. While such fundamental
data is useful in determining when not to trade the strategy,
implementing decision rules in a systematic manner remains
challenging.

6.5 Disentangling Flows and Positioning


Traditional methods of analyzing and forecasting the price of oil are
based on fundamental arguments of supply and demand. As the oil
market matured, it attracted many new participants, including financial
investors, corporate risk managers, and systematic hedge funds, whose
trading motives have little to do with fundamentals. Even though their
trading strategies cannot always be rationalized by market
fundamentals, what these traders do in the market critically impacts
the price. In the end, the price rises when there are more buyers than
sellers, and it falls when there are more sellers than buyers, regardless
of their rationality or motivation. In much the same way that
fundamental analysts count the supply and demand for physical
barrels, professional traders also count the supply and demand for
financial barrels, which are typically measured by futures positions
held by different groups of market participants.
Fortunately, some information about behavioral patterns followed
by various types of traders can be extracted from the data on their
aggregate holdings reported by regulators. Such positioning reports
originated in the 1920s for certain agricultural commodities and
subsequently morphed into the so-called Commitments of Traders (CoT)
report. The reporting format evolved over time, but its basic idea is to
segregate hedgers and speculators, which are categorized, respectively,
as commercial (CM) and non-commercial (NC) market participants. All
traders carrying positions above a certain threshold are categorized.
Separate reports are published for futures, and for futures and options
combined where options are counted on a delta-adjusted basis. In
addition, the CoT report also includes the number of large traders,
which is sometimes helpful in the analysis of the position
concentration.
According to the CoT definition, the CM designation applies to
traders engaged in business activities hedged with the use of futures
and options. Therefore, any trading entity that owns a physical asset, or
a bank that hedges an OTC market-making book with exchange-listed
products, can be classified as a commercial market participant. This
definition was originally designed for agricultural commodities where
the primary CM is a farmer, and all other traders can be deemed to be
speculators. However, in the oil market the categorization is rather
confusing. The ownership of a storage asset puts many large physical
speculators under the CM designation. In addition, the existence of a
large OTC market and the need for market-makers to hedge risks from
bilateral derivatives deals puts many banks into the CM category as
well, despite the fact that the same banks are also large speculators.
In an attempt to provide additional granularity, in 2006 CoT started
to publish the so-called disaggregated report, with ICE adopting a
similar format five years later. In this report, which became the primary
source of positioning information, CM traders are separated into two
sub-categories that distinguish between end-users and market-makers.
One sub-category includes producers, merchants, processors, and users
(PMPU). The other sub-category captures swap dealers (SD) and applies
mostly to banks, even though several registered market-making entities
are also owned by oil majors.
Unfortunately, the reference to producers within the PMPU category
is a misnomer, as oil producers and many consumers prefer to trade
bilaterally with banks to avoid posting margins for exchange-listed
products. Therefore, to make some sense of positions held by corporate
hedgers, one instead should be looking at the SD category, as banks
effectively hold futures on behalf of end-users. Unlike many other
commodities, the PMPU category for the oil market is dominated by
large trading houses that own storage and other assets. Their primary
business model is based on a physical arbitrage, where the ownership
of a physical barrel is promptly hedged with futures. However, this
business model is supplemented by aggressive speculation as traders
try to take advantage of unique insights derived from the physical
market by trading highly leveraged futures and options strategies.
The dynamics of positions held by PMPUs and SDs are substantially
different for the WTI and Brent markets.11 Fig. 6.9 shows that the
PMPU category is structurally short Brent futures, as most of the
world’s physical barrels are priced referencing the waterborne Brent
contract, and nearly all oil in transit is diligently hedged by selling
futures. In fact, it is widely believed that Brent inventory hedgers are
the largest participants in the entire oil futures market.
Fig. 6.9 Net ICE Brent futures positions held by PMPU, SD, and NC traders. Source: ICE
In contrast to Brent, net WTI positions of PMPUs, shown in Fig. 6.10,
are more mixed. While some short WTI futures are held to hedge
barrels held in domestic storage facilities, such shorts are partially
offset by the long leg of WTI-Brent spread arbitrageurs. Physical
arbitrageurs buy WTI-Brent spread to hedge US exports or to lock in
the economics of pipelines that connect WTI production centers and
inland storage facilities to the US Gulf Coast export terminals, where
pricing is determined by a waterborne Brent contract.
Fig. 6.10 Net CME WTI futures positions held by PMPU, SD, and NC traders. Source:
Commodity Futures Trading Commission (CFTC)
Positions held by SDs also differ for WTI and Brent, reflecting
regional imbalances between producer and consumer hedgers. In
North America, the aggregate position held by dealers is heavily
dominated by futures shorts held against OTC producer hedging. If one
is interested in modeling the hedging behavior of genuine oil
producers, then holdings of WTI SDs are likely to be the best proxy. In
contrast to WTI, OTC Brent flows are more skewed to the consumer
side, as European and Asian airlines and some industrial consumers are
more active hedgers than their US counterparts.12 In addition, some
European utilities purchase natural gas via long-term supply contracts,
the pricing of which is formulaically linked to Brent. This exposure is
also hedged via swaps with OTC dealers who then buy Brent futures on
the exchanges.
The disaggregated CoT report also splits the NC category into two
sub-categories. One sub-category corresponds to positions held by
managed money (MM), which refers to all investors registered with US
regulators. The second sub-category, referred to as other reportables
(OTH), is essentially the residual. However, this additional granularity
in splitting NC category brings more confusion than any informational
value. Many analysts focus only on positions held by MMs and overlook
the important OTH sub-category, mistakenly assuming that this
residual is rather small. However, the OTH captures all financial
institutions that are exempt for various reasons from US registration
requirements, including some affiliates of non-US state-owned wealth
funds, and private investment vehicles of large family offices. For all
practical purposes, MMs and OTHs must be combined and analyzed as
overall NCs.
Unfortunately, NC positions are also a mixed bag. This category
combines the aggregate holdings of many different types of financial
investors, including pension funds that own futures as part of their
portfolio and more dynamically managed hedge funds that can trade
futures in either direction. The contribution of passive investments in
oil futures makes the NC category structurally long. On aggregate,
passive investors take the other size of producers and inventory
hedgers, as the sum of all futures positions must be zero. Net NC
positions in WTI are substantially longer than in Brent due to the
contribution of hedges against ETF flows, the largest of which are
linked to WTI. Finally, what makes the deciphering of the oil-
positioning puzzle particularly difficult is that only a portion of
financial flows is reported under the NC category. Since some investors
are using commodity indices intermediated by banks, these flows are
reported under the SD, but exact split of financial investments between
the two categories is difficult to estimate.
Given the complexity of oil-positioning data with multiple overlaps,
it would be very difficult to use this data in a systematic manner, as any
signal can hardly be separated from noise. For this reason, we again
avoid presenting any historical backtests, which, unfortunately, are
frequently seen in marketing presentations by sell-side analysts.
Instead, we follow the same path as before, and highlight a few useful
mental models that can be applied to positioning data. These concepts
are unlikely to be tradable as a stand-alone strategy based on flows, but
they are sufficiently robust with respect to the sample selection to be a
useful add-on to other trading strategies.
One popular idea for trading based on positioning is the follow the
flow or to follow the smart money strategy. In many financial markets,
smart money is generally associated with hedged funds, or the NC
category. In the oil market, the answer is more subtle. In fact, as we
have seen in Chap. 4, over long periods of time the buyers of oil futures
have lost money and the sellers have made money. Since the longs are
generally NCs and the shorts are CMs, one could argue that in contrast
to many other asset classes, the smart oil money is structural shorts
held by inventory hedgers and physical arbitrageurs. However, this
argument is also debatable, as over the long run a large portion of P&L
from futures comes from the roll yield and not from forecasting the
direction of prices. The answer on where the smart money is depends
on the investment horizon. Hedge funds generally have an edge in
short-term momentum-type trading and use liquidity provided by
commercial traders that use futures to monetize optionality embedded
in physical assets. In contrast, physical traders tend to do better in
capturing long-term value via mean-reverting strategies, where
liquidity is provided by long-term investors.13
To approximate hedge funds’ short-term behavior, one can consider
a trading strategy where the standard momentum indicator, introduced
in the previous chapter, is applied not to the price, but instead to hedge
fund positions. Buy and sell signals πNC can be defined by the
momentum in positioning as the crossover of moving averages

where moving averages are calculated for positions held by NCs

Our intent is only to provide the generic signal specification but avoid
fitting specific parameters. Given the high sensitivity of momentum
signals to the lookback period, to avoid the risk of overfitting, we would
rather steer away from optimization and let the parameters be
customized by the user.
In practice, any trading based on positioning indicators can only be
implemented with a lag. Each weekly report reflects positions held as of
Tuesday, but the report is released on Friday afternoon after the market
is already closed. If one backtests the strategy using daily closing prices,
then the earliest day when the trader can react to the latest positioning
information would be the following Monday. The existence of such a lag
is not specific to strategies based on positioning. Similar issues arise
with any non-price-based trading signals where fundamental and
economic data become publicly available only with a lag.
Another important concept in trading based on positioning
resembles the idea of an inflection point used in the reaction function
in the previous chapter. While it may pay to follow hedge funds up to a
certain point, if too many funds end up in the same bandwagon, then
their own buying or selling can bring the price to an unsustainable level
and the trend can quickly reverse. As this turning point is reached, it
might be better to enter into a contrarian trade. This strategy is known
as fading extreme positioning or fading the crowded trade. This concept
is popular among physical traders, as crowded trades are often
accompanied by large fundamentally unjustifiable price moves that
may open opportunities for physical arbitrages.
There are numerous ways of measuring the crowdedness of the
trade. One metric that has proven to be particularly useful is the so-
called sentiment index (SI). The index normalizes raw positioning data
for a certain category, for example the NC, by measuring it relative to its
previous minimum and maximum levels over a specified lookback
period, as follows

(6.2)

The sentiment index takes values between 0 and 1. Such normalization


removes the overall long bias of the NC category that arises from long
futures held by passive investors and focuses only on more dynamic
funds.
The trading strategy takes a contrarian position defined by

(6.3)

For example, one can let ε = 0.20 and take long and short positions,
respectively, in lower and upper quintiles of the sentiment index.
Like in previously discussed systematic strategies, multiple signals
can be blended. One can easily combine a follow the flow strategy with
fading extreme positioning using the idea similar to the reaction
function from the previous chapter. Here, one would follow hedge funds
up to a certain inflection point, beyond which the position is reduced
and then switched to a contrarian direction as the trade gradually
becomes crowded. Obviously, the inflection point is hard to pinpoint,
and here it becomes a strategy parameter that can be calibrated to
historical data.
Even more powerful signals can be constructed if positioning
indicators are further confirmed by other signals. For example, one can
strengthen the idea of fading the crowded momentum trade by taking a
contrarian signal only when some combination of positioning and the
price moves is extreme. One way to do this is to apply the algebraic
structure (6.2) of the sentiment index to construct a normalized price
(NP) indicator, as follows

Then the two scaled metrics for positioning and price can be combined,
using, for example, the Euclidean distance formula

and the same trading signal (6.3) can be applied where SIt is replaced
with dt.
Any information that can be deduced from the positioning analysis
is highly sought after by oil traders. It is analogous to playing a poker
game with an educated guess about the opponent’s cards. Complete
knowledge of the opponent’s hand is unlikely to be possible, but even a
quick glimpse could lead to a powerful edge. However, given the
incompleteness of the overall picture, which is contaminated by
overlapping categories in reported data, running a systematic trading
strategy purely based on positioning data is unlikely to be sustainable.
Like any other quantamental strategy, this concept works better when
it is combined with other discretionary inputs. In the next chapter, we
extend the idea of quantamental trading to macro factors that connect
oil to the broader world of financial markets.
References
Alexander, C. (2001). Market models. Wiley.

Bouchouev, I., & Zuo, L. (2020, Winter). Oil risk premia under changing regimes. Global
Commodities Applied Research Digest, 5(2), 49–59.

Ederington, L. H., Fernando, C. S., Holland, K. V., Lee, T. K., & Linn, S. C. (2021). The dynamics of
arbitrage. Journal of Financial and Quantitative Analysis, 56(4), 1350–1380.
[Crossref]

Fattouh, B. (2011). An anatomy of the crude oil pricing system. Oxford Institute for Energy
Studies, Working Paper, 40.

Imsirovic, A. (2021). Trading and price discovery for crude oils: Growth and development of
international oil markets. Palgrave Macmillan.
[Crossref]

Kang, W., Rouwenhorst, K. G., & Tang, K. (2020). A tale of two premiums: The role of hedgers
and speculators in commodity futures markets. The Journal of Finance, 75(1), 377–417.
[Crossref]

Footnotes
1 Many financial speculators are required to roll their position at least five days prior to the
expiration of the futures contract to avoid regulatory position limits and risks of physical
delivery. The static short spread strategy is relatively insensitive to the exact rolling schedule,
provided that a short position is entered prior to the scheduled index rolls.

2 The choice of this metric does shorten the lookback period as Cushing capacity data is only
available since 2011. However, the conclusions are substantially similar if the analysis is
repeated for longer lookbacks using private estimates for the storage capacity prior to 2011.

3 See Ederington et al. (2021).

4 Brent futures are based on the so-called BFOET basket, which includes Brent, Forties,
Oseberg, Ekofisk, and Troll, all produced in the North Sea. US Midland oil produced in the
Permian Basin was added to the basket in 2023. For a detailed description of the price setting
mechanism in the Brent market, we refer to Fattouh (2011). See also Imsirovic (2021).
5 The USA continues to import some oil to better match refinery needs that are optimized to
run on heavier crude oil. While the USA imports heavy oil, the light shale oil is exported to less
complex refineries located mostly in Asia and Latin America. With the growth of shale
production, US oil exports started to exceed its imports.

6 In practice, arbitrage traders often shift one leg of the spread by one month to account for the
shipping time. For example, to approximate the economics of oil exports from the USA to
Europe, traders are more likely to use the spread between Brent futures that expire at time
T + 1 and WTI futures that expire at time T.

7 The WTI-Brent convergence strategy and its sensitivity to model parameters are analyzed in
Bouchouev and Zuo (2020).

8 For a good introductory discussion of cointegration and its application to financial markets,
we refer to Alexander (2001).

9 There are, in fact, several competing US Gulf Coast WTI contracts listed by two major
exchanges, CME and ICE, but for simplicity, we do not differentiate between them.

10 Light and heavy oil correspond to the oil density and sweet and sour to its sulfur content.

11 For illustration, we use holdings of two main futures benchmarks, WTI traded on CME and
Brent traded on ICE. Adding options and other less liquid exchange-traded WTI and Brent
futures does not change the conclusions.

12 To simplify the exposition, we do not include positions held in refined products, such as
diesel, gasoil, or RBOB. The notional size of such positions is smaller than it is for crude oil
futures, even though it is comparable and sometimes even larger if measured relative to the size
of the market for refined products.

13 The question whether speculators or hedgers make money by trading commodity futures
has been studied extensively in the academic literature but with largely inconclusive results.
Given the complexity of differentiating between hedgers and speculators in the oil market and
the lack of available data, establishing such a causality using purely statistical tools is extremely
difficult. One interesting attempt to separate the impact of hedgers and speculators on prices by
the investment horizon was made by Kang et al. (2020) for a broader commodity portfolio.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
I. Bouchouev, Virtual Barrels, Springer Texts in Business and Economics
https://doi.org/10.1007/978-3-031-36151-7_7

7. Macro Trading
Ilia Bouchouev1
(1) Pentathlon Investments, LLC, Westfield, NJ, USA

The oil ecosystem is a sub-component of a larger complex dynamic


system of financial markets. Such systems are characterized by
feedback loops with two-way causality and a critical role of
boundaries that keep the system from falling apart.
The relationship between oil and the US dollar (USD) is an aggregate
of multiple feedback loops competing for dominance. While the joint
dynamics of oil and USD is too unstable for trading, the linkages
between oil and currencies of oil-exporting countries are more
robust.
The idea of cross-asset stat-arb trading can also be applied to the
equity market, where oil is traded as a spread to the basket of stocks
of independent oil producers. The stock selection can be optimized
using hedging profiles of individual companies.
The relative value strategy between gasoline futures and short-term
inflation swaps is another example of the cross-asset stat-arb. The
two markets operate at different speeds, creating trading
opportunities when the needs of respective market participants
diverge.
An attempt is made to estimate the fair value of oil as a function of
macroeconomic variables. While the model is based on shaky
theoretical grounds, it has still proven to be helpful to macro traders.
Its modifications include dynamic factor selection and methods of
machine learning.
7.1 Dynamic Systems and Feedback Loops
In this final chapter dedicated to trading linear instruments, such as
futures and swaps, we look at how oil trading connects to other
financial asset classes. So far, we have only considered strategies
confined to the petroleum complex, implicitly assuming that these
trades are immune to exogenous financial factors. However, the entire
petroleum ecosystem is a sub-component of a larger complex dynamic
system of financial markets, where everything impacts everything else,
the relationships between system components are reciprocal, and
information flows in both directions in the form of multiple feedback
loops.
Financial markets are filled with numerous loops between
interconnected components. Consider, for example, one of many causal
macroeconomic chain reactions that could result from an exogenous
rise in oil prices. As explained in Chap. 4, an increase in the price of oil
raises the headline inflation mechanically via oil pass-through into
retail gasoline prices. Rising inflation impacts monetary policy,
increasing the likelihood of interest rate hikes. Higher US interest rates
then induce capital inflows to the USA from countries that have lower
domestic interest rates, leading to a stronger USD. However, since the
price of oil is denominated in USD, dollar appreciation creates
downward pressure on the price of oil, which contains its initial rise.
This chain is an example of a negative, or balancing, feedback loop,
shown on the left side of Fig. 7.1. In a balancing loop, any change in a
given variable forces other system components to react in a way that
dampens the initial change. Such balancing loops are generally
associated with mean-reversion.
Fig. 7.1 Examples of negative (balancing) and positive (reinforcing) macroeconomic feedback
loops that involve oil
Consider now an example of a different causal loop. The rise in oil
prices can also be associated with stronger demand in emerging
markets (EM). Any indication of EM growth then stimulates capital
flows to move from stable assets, such as US Treasuries, towards riskier
investments, triggering a so-called risk-on move across broader
financial markets. Weakening USD can lead to higher US inflation via
more expensive imports of consumer goods. This brings additional
demand for financial oil products as a hedge against inflation, further
contributing to the rise in oil prices. This is a positive, or self-reinforcing
feedback loop. In contrast to the balancing loop, the positive feedback
loop amplifies the initial change in the given variable. Self-reinforcing
loops are associated with momentum.
The behavior of any complex dynamic system is characterized by
multiple positive and negative feedback loops, all acting concurrently. If
the positive loop dominates, then the system maintains its current
direction. However, it cannot run in the same direction forever. There
are certain hard boundaries that must eventually limit the move, much
like storage boundaries do in the case of trending oil inventories. The
presence of system boundaries forces a shift in dominance when the
balancing loop takes over from the reinforcing one. Obviously, the
crucial element in the analysis of such systems is identification of the
turning point when the dominance shifts. This point is what drives the
trader’s decision whether to buy or to sell. The reaction function with
an inflection point, introduced in Chap. 5, provides an example of how
traders can quantify such a shift in dominance.
Modeling the dynamics of the entire system is a gargantuan task.
Conventional statistical methods, such as linear regression with a pre-
defined direction of causality, are unlikely to be of much value here. A
more promising approach is to use the concept of cointegration,
introduced in the previous chapter. It searches for some stable
equilibrium among system components. While the system revolves
around the equilibrium state, it eventually converges to it. Deriving a
universal equilibrium for the entire dynamic system of financial
markets, however, is more interesting for economic theorists than for
traders. In practice, such an equilibrium is impossible to quantify given
the lack of timely data and the large amount of noise that surrounds it.
Practitioners instead tend to focus their attention only on specific
components of the system, where mutual feedback loops are more
visible and sufficiently ring-fenced from unrelated noise.
In this chapter, we briefly explore pairwise relationships between
oil and three major financial markets. First, we look at the foreign
exchange market and analyze the widely debated but very unstable
relationship between oil and USD. We eliminate some USD-specific
noise by focusing on the relative value strategy between commodity
futures and currencies of commodity-exporting countries. We then
extend the concept of cross-asset stat-arb to the equity market and
consider a strategy of trading oil futures as the spread to the basket of
stocks of independent oil and gas producers. Finally, we revisit an
important linkage between oil and inflation that has already come up
on several occasions in the book. This time, we outline a specific
arbitrage-like trading strategy between petroleum futures and short-
term inflation swaps. To conclude, we combine some of these
observations into a directional macro model that trades based on the
estimate of the fair value for the price of oil derived predominantly
from various macroeconomic variables.
7.2 Oil, Dollar, and Commodity Terms of Trade
Strategies
The relationship between oil and USD is too multifaceted to be molded
into any prescriptive format. It is one of the most extensively studied
economic relationships that, nevertheless, still leaves traders scratching
their heads when its dominant driver suddenly changes. One common
pitfall of such analyses is the application of conventional regression
methods that implicitly assume the direction of causality from one
asset to another, which is not consistent with the dynamics of a highly
nonlinear system. The oil-USD link in the complex system of financial
markets is the result of the confluence of many forces and feedback
loops acting with varying levels of strength at different times while
fighting for dominance. The dominant factor shifts quite frequently,
leading to drastic changes in correlations, which makes regression-
based estimates extremely unstable and sensitive to the sample
selection.
Like members of a family or parts of the universe, oil and USD
impact each other in many different ways. There is little doubt that in
the long run oil and USD are related. However, to turn this relationship
into a viable strategy on shorter-term trading frequencies, one must
find a way to eliminate less predictable factors and zoom in on the
behavior of more stable residuals. This is the same principle that
underlies previously studied energy stat-arb strategies. Now we take
another step and extend similar ideas across different asset classes,
starting with the elusive relationship between oil and USD.
The question of how the price of oil impacts USD has been actively
debated ever since the dollar was de-pegged from gold in 1971. Interest
in this topic has gone way beyond the basics of macroeconomics. It has
been a question of US national security that played a central role in
geopolitics during previous episodes of high oil prices. Following the
steep rise of oil prices and fuel shortages in the aftermath of the Oil
Embargo of 1973–1974 by the Organization of Arab Petroleum
Exporting Countries (OAPEC) that sent the USA into a recession,
Washington convinced Saudi Arabia to invest excess oil revenues into
US Treasuries in exchange for military aid. These investments were
subsequently dubbed petrodollar recycling. The deal was so clandestine
that its details only become known to the public several decades later.
Effectively, Saudi Arabia was financing a significant portion of American
spending, while at the same time higher oil prices were increasing
demand for USD from many other petroleum-importing countries. This
medium of exchange transmission channel contributed to a positive
correlation between oil and USD.
For years, the positive linkage from the medium of exchange
channel was largely offset by an inverse relationship between oil and
USD from the US terms of trade channel. Since the USA was the largest
importer of oil, rising oil prices negatively affected the US overall trade
deficit, which, in turn, put downward pressure on the US currency. Over
time, the contribution of both petrodollar recycling and the US terms of
trade channel became muted. The impact of petrodollar recycling on
USD even turned negative when sovereign wealth funds of oil-
producing countries started to diversify their petrodollars by selling
USD and buying other currencies. The US terms of trade linkage also
vanished nearly entirely, as the rapid growth of US shale production
made the country energy independent.
For completeness, however, we list both the medium of exchange
and the US terms of trade transmission channels as legacy factors in
Fig. 7.2, as oil analysts must be aware of their past contribution when
backtesting trading strategies.

Fig. 7.2 Multiple transmission channels between oil and USD

As demand for oil grew globally, particularly outside of the USA, the
dominant role in oil-USD dynamics shifted to the denomination, or the
numeraire effect, which is the consequence of oil being priced
predominantly in USD. The denomination effect leads to an inverse
relationship, as stronger USD implies that it takes fewer USD to buy a
barrel of oil.
This inverse relationship can be visualized from the supply and
demand perspective. A stronger dollar translates into higher domestic
prices in countries outside of the USA, which over the last decade were
the primary drivers of oil demand growth. Higher prices have a
negative impact on demand, and, therefore, create downward pressure
on the price of oil. In addition, USD appreciation also causes a negative
impact on price via increasing supply by oil-producing countries with
free floating currencies. Since a portion of production costs is
denominated in local currencies, stronger USD improves profit margins
and incentivizes oil exports. The supply-side channel is more
complicated though due to secondary factors, such as the value of USD-
denominated debt issued by foreign producers.
There are several macroeconomic factors that counter the
predominantly negative oil-USD relationship. One is driven by oil direct
pass-through into inflation and its impact on changes in monetary
policy. As we have already mentioned, rising inflation increases the
likelihood of interest rate hikes, which, in turn, positively impact
demand for short-term US Treasury bonds. Another positive
contribution to the oil-USD relationship comes from the US growth
effect. More favorable economic prospects in the USA relative to the
rest of the world could result in stronger USD given the higher
likelihood of monetary tightening and stronger oil prices, as the USA is
still the world’s largest petroleum consumer. Likewise, both USD and oil
can falter when the USA experiences an idiosyncratic slowdown. This
happened, for example, in the early days of the Covid-19 pandemic in
the USA when both oil demand and USD fell, and for a short period of
time the correlation between the two falling assets was positive.
Subsequently, as the pandemic escalated around the world, USD
regained its stability, while oil continued to fall because of the reduction
in global demand. As a result, the correlation returned to negative.
Numerous academic studies of macroeconomic links between oil
and USD have been conducted, but they only produced conflicting and
inconclusive results. This highlights the time-varying nature of the
relationship and the high sensitivity of econometric analyses to the
sample selection. Many attempts have also been made to establish an
aggregate direction of causality. However, given so many comingled
channels acting at once, disentangling causality in any robust way using
statistical methods is practically impossible. While slow-moving
macroeconomic forces remain more relevant for a long-run
relationship, in the short-run their combined impact is
indistinguishable from non-tradable noise. In fact, prior to the
financialization of oil markets that began in the 2000s, the dynamics of
short-run oil-USD correlation largely resembled a random walk around
zero mean. Financialization has brought drastic changes to this
relationship. Markets started to move much faster, and the dominance
of fundamental forces has been challenged and superseded by fast-
moving financial flows and cross-asset spillovers.
Perhaps the most important driver of the oil-USD relationship in the
modern financialized world of oil trading is the risk aversion or safe
haven channel, which is driven by risk-on, risk-off sentiment shifts.
When adverse economic news triggers a large selloff in the equity
market, many other risky assets, including oil, are also liquidated by
financial portfolio managers tasked to maintain fixed allocation across
all risky assets. During such a risk-off event, investors reallocate capital
towards safer securities, such as US Treasuries which, in turn, increases
demand for USD.
The financial safe haven channel has become the dominant driver of
the predominantly inverse oil-USD relationship. Since approximately
2004, the correlation between returns on oil and a broad USD index has
had a strong negative bias, as illustrated in Fig. 7.3. The correlation was
the strongest, i.e., the most negative, during the years of quantitative
easing following the global financial crisis. To stimulate investment
demand, large bond purchases were made by the US Federal Reserve
Bank, which led to low interest rates and downward pressure on USD,
pushing investors into riskier assets, including oil.
Fig. 7.3 Rolling one-year and sub-sample WTI-USD correlations during three different regimes
Occasionally, the risk aversion channel also contributes to a positive
relationship between oil and USD. This typically occurs either during
periods of high inflation caused by supply-side constraints, or when
geopolitical tensions jeopardize the stability of global oil production.
During such episodes, oil itself, along with USD, is perceived by the
market to be a safe haven instrument, and the dominant negative
relationship is interrupted by shorter spells of positive correlation. For
example, the inflation factor became more visible when the economy
was recovering from the Covid-19 pandemic and higher inflation was
driven by supply disruptions and rising commodity prices.
Finally, the conventional denomination effect also started to play a
larger role in financial oil markets. Many international investors have
oil allocation mandates set in their local currency, such as EUR. If USD
strengthens, i.e., the EURUSD exchange rate weakens, then the notional
value of oil holdings measured in EUR exceeds the investor’s target
allocation. Oil positions must then be reduced to bring the allocation
back to the target, creating downward pressure on the price of oil. The
numeraire effect is even more pronounced for highly leveraged risk
party funds and CTAs that manage allocations based on the asset
volatility. If oil and USD are negatively correlated, then oil volatility
measured in EUR is generally smaller than oil volatility measured in
USD. Consider, for example, oil trading at 90 USD per barrel and a
EURUSD exchange rate at 1.20, which translates to an oil price of 75
EUR per barrel. If the oil price rises 10% to 99 USD per barrel and EUR
vs USD strengthens 10% to 1.32, then the oil price measured in EUR
remains unchanged at 75 EUR per barrel despite its 10% change in
USD.
With so many overlapping driving factors of the oil-USD
relationship, trying to use one of the two variables to predict the other
rarely results in a viable trading strategy. One is better off trading them
as a spread that does not depend on the direction of causality. For the
spread to be a viable candidate for the convergence strategy, some
cointegration criteria must be satisfied. While such criteria usually fail
for the spread between oil and a broad USD index, they are often met
for a subset of the so-called commodity currencies of producing
countries, such as Canada or Norway. In contrast to petrodollar
recycling by producers whose currencies are pegged to USD, many
commodity exporters with free-floating currencies convert excess USD-
denominated oil revenues into domestic currencies, which makes them
highly sensitive to the price of oil. This transmission channel is
sometimes referred to as a wealth effect.
In fact, oil plays such an important role in the economics of these
countries that the price of oil is frequently used by Central Banks as an
input for modeling the fair value of their currencies. Similar models for
commodity currencies have also been developed and widely adopted by
foreign exchange traders. This opened an opportunity for a cross-asset
stat-arb-like strategy to trade certain commodities as the spread
relative to their currencies whenever the two markets dislocate. We
illustrate this with a popular convergence strategy between CADUSD, a
Canadian dollar expressed in USD, and WTI.
Since the CADUSD exchange rate and WTI represent values of two
assets both denominated in USD, trading one as the spread against the
other eliminates a large portion of the idiosyncratic USD-specific noise.
As a result, the co-movement between CADUSD and WTI is more
visible, as illustrated in Fig. 7.4. In other words, CADUSD and WTI are
generally considered to be cointegrated.
Fig. 7.4 Co-movement between CADUSD and WTI futures prices (m3)
There are effectively two primary factors that drive CADUSD. One
factor is the relative monetary policy in Canada versus the USA. This
factor is typically modeled by interest rate differentials, which steer
financial flows towards the direction of higher interest rates. The other
primary factor is the price of oil. When monetary policies in both
countries are on hold, then the contribution of the first factor weakens,
and the price of oil starts playing the dominant role. However, when
monetary policies are changing, the exchange rate becomes more
driven by interest rate differentials and the relationship with oil
weakens. One can clearly see this during uneven recovery after the
Covid-19 pandemic and rapidly rising inflation.
If the monetary policies in both countries are expected to remain
relatively stable, then the technical implementation of the convergence
strategy becomes very simple, as in the energy convergence strategies
discussed in the previous chapter. We define the cross-asset spread as

and then buy and sell the pair when the spread moves outside of a
certain band. The band can be specified, for example, by the spread
deviation from its moving average, as follows:
(7.1)

The most challenging technical part of this strategy is in defining β,


which is generally much more of an art than it is a science. Since the
idea of cross-asset stat-arb is based on the concept of cointegration,
which measures the long-term relationship between price levels, in
theory, one should be estimating betas using the price-level regression
with a fairly long lookback period.
In practice, cross-arb traders pay particular attention to the choice
of the lookback period. It must be chosen carefully, corresponding to
regimes that are less contaminated by changes in monetary policies. As
in the case of the energy stat-arb portfolio, the crucial decision here is
when not to trade the strategy. In a more sophisticated version of the
strategy, one can consider hedging against the uncertainty resulting
from changes in monetary policies. This can be done by supplementing
CADUSD versus WTI spread trade with additional positions in the
spread between Canadian and US interest rate futures.
Like energy stat-arb, the cross-asset stat-arb strategy is also better
traded within a broader portfolio. Besides the CADUSD and WTI pair, a
similar statistically robust relationship exists between NOKUSD, a
Norwegian krone expressed in USD, and Brent. Trading this pair
together with CADUSD and WTI as a mini portfolio can substantially
improve the overall performance, as it provides some degree of
geographical diversification both across currencies and in the oil
market. One can further diversify the strategy by adding EM currencies
of oil producers, but it should be done with care due to higher
unrelated risks in these countries. Furthermore, it could be beneficial to
replace benchmark WTI and Brent futures with oil grades that are
specific to the country’s exports, such as WCS (West Canadian Select)
for Canada. The addition of less liquid oil contracts, however, increases
transaction costs and requires more advanced execution capabilities.
The terms of trade strategy is not limited to energy futures. It can
also be applied to some metals and agricultural products. For example,
Chile is the largest producer of copper, with the Chilean peso, CLPUSD,
highly correlated to the price of copper. Figure 7.5 lists commodity
currency pairs commonly used for constructing more diversified
commodity terms of trade (CToT) portfolios. This pair selection is based
on various correlation and cointegration thresholds with an additional
adjustment for liquidity and geographical diversification. For some
countries that export many different commodities, the strategy can be
improved by using the basket of corresponding commodity futures that
mimic the basket of exports.

Fig. 7.5 An example of pair selection for a CToT portfolio


To conclude, we highlight the primary risk embedded in such
strategies. Since many commodities are produced in countries known
for their geopolitical instability, challenges arise when domestic
commodity production is disrupted. The strategy can then experience a
double whammy loss if the commodity price spikes due to the
reduction of supply from the producing country, while the local
currency simultaneously weakens, as financial investors reduce their
financial holdings in the country in response to increasing domestic
uncertainty.
We next extend the concept of cross-asset stat-arb to the equity
market, where oil futures can be traded as a spread to a portfolio of
energy stocks.

7.3 Oil and Energy Equities


The relationship between oil and broad equity indices is no less
complex than the relationship between oil and USD. Many of the
macroeconomic transmission channels listed in Fig. 7.2 stem from
overall financial conditions that impact both equities and oil markets.
Transmission linkages between broad equity markets and oil are also
time-varying and have low signal-to-noise ratios, which makes it
difficult to trade the relationship. The idea behind a cross-asset stat-arb
strategy is again to eliminate a portion of the macro noise and trade oil
futures relative to the subset of the equity market where connections
are more direct. This subset is the equity basket of independent oil and
gas producers, whose profitability is highly dependent on the price of
oil. With the exception of a few gas-focused companies, the majority of
the revenue in this sector is derived from oil production.
While professional oil-equity arbitrageurs are more likely to design
their own customized equity baskets, the strategy works quite well
even for standardized baskets of stocks, using, for example, ETFs. For
illustration, we use XOP, the most popular and the most liquid ETF of US
independent oil and gas producers. XOP has been in existence since
2006, the period that covers the growth in US oil production. Figure 7.6
shows that the correlation between WTI and XOP is naturally much
higher than the correlation between WTI and the broader US equity
market.
Fig. 7.6 Rolling one-year returns correlation of WTI futures (m3) to XOP and SPY equity ETFs
The XOP-WTI spread satisfies conventional cointegration criteria,
making this pair a suitable candidate for the cross-asset convergence
strategy. Figure 7.7 illustrates how closely the two assets tracked each
other over time.

Fig. 7.7 Co-movement between XOP and WTI futures prices (m3)

The actual strategy implementation is again kept simple. We apply


the basic trading rule (7.1) to the spread, defined by

To avoid the hassle with the estimation of hedging beta, practitioners


often construct the equity basket, which has similar volatility to WTI.
Since the volatilities of XOP and WTI happen to be roughly similar, in
this example one can trade the strategy with β = 1, which means that
the pair is notionally weighted.
We should also caution readers against backtesting the strategy
using publicly available closing prices for XOP and WTI. All stat-arb
trading strategies critically rely on simultaneity of price data for both
legs. While the WTI market closes at 14:30 Eastern Standard Time
(EST), US equity markets remain open for another ninety minutes.
Using both settlement prices for backtesting introduces a look-ahead
bias that could overestimate the historical performance of the strategy.
To properly analyze this strategy, one must use intra-day data.
Obviously, like in all other trades, the basic oil-equity strategy can
be enhanced. Perhaps the most unique approach which is specific to
this strategy is to replace the standardized XOP ETF with a customized
basket of energy stocks. One can also run the strategy for each
individual stock and then construct an optimal basket of stocks that
exhibit the strongest cointegration with oil. However, such cherry-
picking introduces additional degrees of freedom and must be used
carefully to avoid overfitting. It is also possible to construct a portfolio
of oil and natural gas futures weighted based on the share of oil and gas
revenues for companies in the chosen basket.
One important factor to consider is the impact of hedging by
producers, which could reduce the sensitivity of the company’s stock to
the price of oil. A particularly novel concept is to design a customized
equity basket with weights that explicitly incorporate producers’
individual hedging ratios. Alternatively, all producer stocks can be split
into basket of hedgers and non-hedgers, which theoretically must have
different oil betas but the market may not fully recognize such a
distinction. While this approach is quite elegant, its implementation is
very complex. The hedging data are generally available only from
regulatory producer filings, which are published quarterly with a lag.
However, this data can be combined with information about hedging
deals that are reported live by Swap Data Repositories (SDRs). Even
though real-time SDR reporting does not disclose names of trading
counterparties, one can attempt to deduce them by establishing
hedging patterns using quarterly filings. We will leave this idea for
future research that can probably qualify for a doctoral thesis in
finance.
Having analyzed cross-asset arbitrage between two fast-moving
markets, we next look at another interesting example of the cross-asset
lead-lag strategy, between fast-moving energy futures and the slower-
moving OTC market for inflation swaps.

7.4 Oil and Inflation


The relationship between oil and inflation that we have already
encountered on multiple occasions provides another example of a
bidirectional causality, a typical attribute of complex dynamic systems.
As discussed in Chap. 4, oil causes inflation directly via its pass-through
to retail gasoline prices and indirectly as the feedstock cost of other
goods. Inflation, on the other hand, impacts oil more via forward
expectations, which generates demand for hedging, in particular, from
risk parity funds. The strength of both transmission channels depends
on the time horizon.
One simple way to trade inflation relative to oil would be to
replicate the statistical convergence strategies presented above for
commodity currencies and for oil and gas producer equities. This
indeed can be done but a better trading opportunity exists, which is
more specific to the structure of the inflation market. This opportunity
is to trade short-term inflation swaps, whose linkages to energy futures
happen to be more algebraic than statistical. While this strategy is not
free money, it is much closer to being a model-free arbitrage than to a
statistical arbitrage.
The world’s most developed inflation market is in the USA, where
the primary trading instruments are inflation-linked bonds, or TIPS,
introduced in Chap. 4, and bilaterally negotiated inflation swaps. Both
TIPS and inflation swaps are settled based on the CPI, which was also
introduced in Chap. 4. The index represents the weighted average of
prices for the aggregate consumer basket of goods and services, with
the reference base set at 100. Inflation is quoted as the rate of change in
the CPI typically on a month-over-month or year-over-year basis. While
TIPS are only issued with maturities in January, April, and July, inflation
swaps exist for each monthly CPI print, which is also called a CPI fixing.
Such a finer granularity of swaps for monthly CPI fixings makes them a
better match to trade against energy futures, which are also listed with
monthly expirations.
As it was shown in Fig. 4.​5, the headline CPI is driven by volatile oil
prices via their pass-through to retail gasoline prices. While consumer
gasoline expenditure, which is reported under the motor fuel
component of the CPI, represents less than 4% of the total basket, it
explains approximately 60% percent of the CPI monthly variance.1 Fig.
7.8 illustrates this relationship over a long period of thirty years. In fact,
during periods of low inflation the explanatory power of retail gasoline
prices tends to be even larger, as other CPI components exhibit low
volatility.

Fig. 7.8 A large portion of CPI monthly variance is explained by retail gasoline prices (1993–
2022)
To describe the trading strategy, we use RBOB futures, which are
most closely related to retail gasoline prices, even though a similar
strategy can be constructed using more liquid WTI futures, or a basket
of petroleum futures that may also include diesel futures and Brent.
The trading opportunity exists largely because energy futures and
inflation markets operate at different speeds. While the electronic
market for futures moves very fast, inflation swaps largely trade OTC,
where transactions are still negotiated in the old-fashioned manner
bilaterally between clients and their bank dealers. If the inflation
market were efficient, then one would expect every move in energy
futures to be instantaneously repriced in CPI swaps. In practice,
however, this does not always happen, as the slow-moving OTC inflation
market often exhibits some inertia. This market is dominated by a
different set of market participants, predominantly by fixed income
portfolio managers whose decisions and investment mandates are set
based on longer-term structural views. They largely ignore short-term
fluctuations in energy prices as many of them are not even authorized
to trade RBOB futures. Energy traders, on the other hand, tend to be
nimbler and more flexible with a better eye on short-term arbitrage-
type profits.
When the inflation market lags the move in energy futures, the
trader can take a position against the dislocation between RBOB
futures and inflation swaps and trade the pair as a cross-asset spread.
The strategy bets on the residual which we can define as ex-gasoline
inflation. The dislocation between the two prices may allow the trade to
be entered at sufficiently advantageous levels, so that other
components of the CPI do not have to be estimated with high precision.
The crux of the strategy is in separation and careful hedging of the CPI
gasoline component. The ex-gasoline component is effectively treated
as residual noise. To implement the strategy, however, one needs to
understand the somewhat intricate plumbing of inflation derivatives.
One important feature of the inflation market that often causes a
considerable amount of confusion to outsiders is the so-called base
effect. Another peculiar quirk of inflation derivatives is the lag between
the valuation date and the period over which inflation is measured. We
will briefly explain these features but without getting too deeply into
details. For simplicity, we use a benchmark one-year inflation swap.
The fixed rate of the swap is quoted as an annual expected rate of
inflation. This rate is swapped for the floating rate of inflation i(T),
which is determined by the ratio of the reference CPI index I(T) at the
future time T > t to the reference index I(t0) at past time t0 < t, as
follows:

The reference index is calculated using monthly CPI prints but with a
lag of between two and three months. If the observation day t is the
first day of a calendar month, then the reference index uses a monthly
CPI print for the third preceding calendar month. In other words, in this
case t0 = t − 3/12 and the annual inflation rate will be determined by
the CPI print nine months forward, i.e., at time T = t + 9/12. If the
observation date t is any other day of the calendar month, then the
reference index is determined by a linear interpolation between CPI
prints two and three months prior.2
If another one-year swap is traded tomorrow at time t + dt, then it
will reference a new base index I(t0 + dt) which could be substantially
different from I(t0) if energy futures happened to experience a large
move at time t0 + dt. This makes it rather difficult to compare market
inflation swap rates on a rolling basis and visualize the price dynamics
in the form of a time series. As the base index constantly rolls in time,
the quoted swap level changes daily even if market expectations about
the future inflation index remain the same.
The skill of trading short-term inflation relative to energy futures
hinges on the trader’s ability to separate and hedge out the volatile
gasoline contribution from the residual ex-gasoline inflation
component of the index, which we denote by ix. Let R(T) represent
retail gasoline prices, and ω denote the weight of gasoline expenditures
in the CPI basket. For simplicity, we assume that ω is a known constant,
even though, more precisely, it can also vary with energy prices. Then
the quoted rate of inflation can be decomposed as

(7.2)

Let us define the pass-through p of energy futures into retail gasoline.


This is typically estimated by the slope of the linear regression of
percentage changes in retail gasoline and RBOB prices. An example for
monthly data is given by Fig. 7.9.
Fig. 7.9 Retail gasoline prices are highly correlated to RBOB futures (2005–2022) (RBOB
futures were introduced in October 2005)
The magnitude of the pass-through increases if one aggregates data
over several months or if one uses year-over-year calculations, which
eliminates short-term noise caused by seasonal differences. However,
the results of such regressions should be interpreted with caution given
the presence of autocorrelation effects.
Furthermore, the goodness of the fit can be improved by allowing
for a lag between futures and retail prices. This lag, however, is not
trivial. In the short run, retail prices are stickier than futures, which is
one of the reasons that this trading opportunity exists. Moreover, the
retail lag is also asymmetric. Retailers tend to raise prices at the pump
quickly in response to rising futures, but they reduce pump prices at a
much slower rate when futures fall, as they are trying to maximize
profit margins. Without getting too sidetracked with the econometrics
of pass-through calculations, which is not the goal of this book, we can
generally use the industry standard assumption of approximately 60%
pass-through for RBOB futures into retail gasoline prices.
We then replace retail gasoline inflation in (7.2) with RBOB using
the pass-through of traded futures as follows
(7.3)

where F(t0) represents prompt RBOB futures prices observable at time


t0.3
Alternatively, the market-implied rate of ex-gasoline inflation can be
expressed as follows

(7.4)

If one were able to trade futures directly on retail gasoline prices, then
(7.4) would have applied with p = 100%. In fact, there exists an OTC
market for retail gasoline swaps, but, unfortunately, it is extremely
illiquid. A similar market for retail diesel swaps is more developed, and
some energy traders use the market forward curve for retail diesel
prices as a guide to estimate the forward price of retail gasoline, which
is the most uncertain variable in the CPI-RBOB convergence strategy.
Let us consider a trading position in CPI swap i hedged with n
contracts of RBOB futures

What makes this strategy appealing is that the crucial hedging ratio can
be calculated algebraically rather than estimated statistically.
Let us assume that the futures price experiences an instantaneous
shock dF. Then from (7.3) we obtain that the value of the inflation swap
contract changes by

To illustrate this with a numerical example, let p = 0.60, ω = 0.04, and


F(t0) = 3.00 per gallon. Then a ten cent per gallon move in the RBOB
futures price changes the annual rate of inflation by
. Since one RBOB futures contract is for
42,000 gallons, to hedge a $100 million notional position in the
inflation swap, the trader needs approximately RBOB
futures contracts. In contrast to previous strategies, here the hedging
ratio is calculated fundamentally instead of using rolling regressions,
even though some statistical uncertainty is still hidden in the
assumption about the pass-through p.
Once the most volatile component of the headline inflation is
removed, the trading signal can be defined in a number of alternative
ways. Fixed income specialists tend to look at it more in inflation terms.
They calculate ix as in (7.4) and then compare it to some measure of
historical inflation ih. The latter is usually estimated from a range of
inflation realizations over prior years, adjusted for seasonality and
recent trends. The trading signal is to buy and sell inflation-RBOB
spread St when ix measured at time t, which we denote by ix, t, deviates
from its trend-adjusted historical seasonal norm, i.e.,

Since the idea of the strategy is to trade it only when the deviation
becomes particularly significant, one can avoid the need for high
precision in the estimation of ih. If the dislocation between inflation and
futures markets is very large, then ex-gasoline inflation ix,t can even fall
entirely outside of the range of previous historical observations. This
would suggest that the trading opportunity is so good that it would
have made money under all previous paths of realized inflation.
In contrast to inflation specialists, oil traders like to turn the
problem upside down and map all variables instead into energy terms
that they are more familiar with. The primary uncertain variable in this
strategy is the forward price of retail gasoline Rt(T), which at time t is
approximated using the shape of observable futures curves for RBOB
and, if necessary for longer maturities, crude oil futures. Therefore, by
and large, this strategy amounts to trading the spread between retail
gasoline prices implied by the inflation market and RBOB futures.
Instead of expressing market-implied ix,t via pass-through of
observable futures prices as in (7.4) and comparing it to ih, energy
traders assume that the inflation market already embeds some
consensus about historical average ex-gasoline inflation ih. In this case,
one can replace ix with ih in the Eq. (7.2), and instead back out the
market-implied future retail gasoline price from the headline
inflation rate it as follows

This market-implied retail gasoline price is then compared to


observable prices of RBOB futures. If inflation and futures are
misaligned, then the trader buys or sells retail-futures basis at a level
which falls outside of the historical seasonal norm for this basis. In
other words, using energy variables one can also consider an
alternative trading rule, specified as

where b(T) represents some historically normal retail-futures basis for


the month T. The oil trader can also recalculate for a range of
different ih to check the sensitivity of this trading rule to the
assumption about historical ex-gasoline inflation.
The dynamics of the retail-futures basis is the core component of
this trade. Besides the non-trivial lag mentioned earlier, the basis can
also be affected by regional anomalies in gasoline prices. The gasoline
component of the CPI represents the national average for finished
gasoline, but RBOB is only a particular blending component to be
delivered in the New York area against the futures contract. The spread,
therefore, could periodically dislocate during isolated regional events,
such as Gulf Coast hurricanes, or supply disruptions of gasoline-
blending components on the US East Coast. Such energy-specific
knowledge, which financial investors lack, creates an opportunity for
energy specialists to function as arbitrageurs in the short-term inflation
swap market. Given so many important nuances in the behavior of the
gasoline basis, this strategy is unlikely to be successful if it is managed
purely systematically. It presents another example of quantamental
trading where a human equipped with a machine is more powerful
than either a human or machine alone.
So far, we have focused our attention only on the most volatile
gasoline component of the CPI. One can also eliminate an even larger
portion of the CPI variance by hedging some other energy components
of the basket reported under energy commodities or energy services
which are also correlated to diesel, natural gas and electricity prices.
The algebra of the previous calculations remains the same with ω now
representing a larger weight of approximately 6%, which explains a
larger portion of the CPI monthly variance, but the pass-through must
be calculated with respect to the basket of energy futures.
Taking it one step further, one can extend the framework and
attempt to hedge out the second most volatile CPI category, food prices.
Let ω1 and ω2 represent relative importance weights for the energy and
food sub-components of inflation, and p1 and p2 represent pass-through
sensitivities of energy futures F1 and agricultural futures F2 to the
energy and food components of the CPI, respectively. Then one can
introduce the ex-energy-ex-food component of inflation, which is
analogous to the market-implied core inflation, icore, as follows

where F1(t0) and F1(t0) represent two baskets of energy and


agricultural prompt futures prices at time t0.
The market-implied core inflation is then calculated as follows

The practical challenge with this more advanced approach lies in the
estimation of the pass-through of agricultural futures into the food sub-
component of inflation. Given the larger diversity of products that make
up the food basket, such pass-through estimates for agricultural futures
are less reliable.
Having developed some tradable linkages between oil and other
main financial asset classes, we conclude this chapter by building a
simple model that aggregates the information from various macro
variables and attempts to use this information to estimate the fair value
for oil.

7.5 Macro Fair-Value Model


As we have highlighted throughout the book, professional oil traders
tend to focus more on capturing market inefficiencies and trading
spreads than on forecasting the direction of oil prices. Many successful
oil traders are effectively arbitrageurs. They tend to buy relatively
cheap assets and sell correlated but more expensive ones, trading them
as spreads and minimizing directional exposure to the price of oil. Such
relative value trades almost always have better risk-adjusted returns
than outright bets on oil prices.
However, for the vast majority of financial traders oil represents
only one of many assets in their portfolio. Macro traders are unlikely to
make a living by arbitraging energy prices, and instead they may be
content with a high-level opinion on whether the price of oil is cheap or
expensive. To help them navigate in the ocean of macroeconomic
variables that could affect the price of oil, we conclude this chapter with
one somewhat naïve model. It is based on rather shaky theoretical
grounds, so professional oil traders are unlikely to take it seriously.
However, as we stated previously, all models are wrong, but some are,
nevertheless, useful. This model attempts to estimate the fair value for
oil by applying some elementary statistical techniques to macro
variables.
The idea behind the model comes from the foreign exchange
market. Historically, currencies have been viewed as balancing valves
that adjust in response to various macroeconomic factors. Foreign
exchange analysts spend a considerable amount of time valuing
currencies using various techniques, ranging from purchasing power
parity and long-term equilibrium models, to faster market models,
which are based on short-term currency drivers and capital flows. For
example, to estimate the fair value of CADUSD, analysts often apply a
multidimensional regression, where explanatory variables are interest
rate differentials that impact cross-border capital flows, some measure
of the global risk appetite, such as the equity index or VIX, and the price
of oil. Traders then buy and sell the currency if the market price
deviates too far from the theoretical value generated by the fitted
model.
A similar approach can be applied to estimate the fair price for oil
, which is viewed as a linear combination of i = 1, . . , N macro factors
xi,t observed at time t. In other words, we let

(7.5)

For example, we can choose the three-factor model (N = 3) with factors


representing financial legs of the three previously discussed cross-asset
spread trading strategies. In this case, x1 = CADUSD exchange rate,
x2 = XOP equity ETF of oil producers, and x3 = CPI swap price. These are
representative factors from foreign exchange, equity, and fixed income
markets, which have close economic links to the price of oil.
In the cross-asset spread strategies, we were buying and selling oil
and taking the opposite position in these three financial markets. Here,
we consider a strategy which uses financial variables only as a signal
and not as trading instruments. For example, if oil is deemed to be
cheap relative to all three factors, or at least relative to some
combination of them, then it may indicate that oil is generally
undervalued. In this case, one may want just to invest in oil without
taking any offsetting position in other financial assets. Obviously, such a
directional position is much riskier, as oil along with other financial
factors could get even cheaper in the case of a macro-driven risk-off
event.
To determine the cheapness or expensiveness of oil relative to
financial factors, we use the simple price-level regression defined by
(7.5). As we discussed in the previous chapter, for the most part
running linear regressions on non-stationary prices is meaningless,
which is why we warned in advance about the shakiness of the
theoretical foundation of this approach. However, while being
undoubtably questionable, this approach is not necessarily wrong.
Applying regressions to price levels may be appropriate if variables are
cointegrated. Recall that cointegration means that while each variable
may not necessarily be stationary, there exists a linear combination of
them that makes the entire basket stationary. The Eq. (7.5) effectively
makes such an assumption.
In truth, for many other financial factors an Eq. (7.5) is an ambitious
assumption, but the method is still used by traders. We all know plenty
of examples where something that works in theory does not work in
practice. This model is somewhat the opposite. It should not work in
theory, but it has proven to be helpful in practice. In practice, traders
often use (7.5) by letting beta coefficients be time-varying and
estimating them by running rolling regressions. The lookback for such
regressions must be relatively long as the idea of cointegration is to find
some long-term equilibrium among the variables. At the same time,
running regressions on the rolling basis allows traders to incorporate
the latest data points. The output of the regression can be viewed as
representing the fair price of oil, which is then compared to the market
price Ft.
A simple trading strategy then buys oil futures Ft when they trade
below the estimated value by more than a certain threshold ε and sell
them when the market price exceeds the model price by this threshold:

This three-factor fair-value strategy based on USDCAD, XOP and CPI


swaps has been working well since approximately 2016, when the US
ban on oil exports was lifted and WTI became more exposed to global
macroeconomic forces. As with all other quantamental strategies, we
prefer to stay away from delving into backtests to avoid the natural
temptation to make them better with optimized parameters. When
working with daily settlement prices, one should also be careful not to
introduce the look-ahead bias in backtesting, as oil daily settlement
prices are published before prices of some financial factors, such as
equities. In these backtests, one must either wait until the following day
before taking a position, or better calculate signals using simultaneous
intra-day data for all factors.
The static factor-based strategy does have some merit, but it also
carries a significant tail risk. There are many factors not included in the
model specification that could cause large shifts in oil prices. OPEC
decisions, geopolitical events, or weather-driven supply disruptions can
have a large impact on price, but they cannot easily be modeled
systematically. The natural extension of this model would be to include
additional risk factors. Figure 7.10 lists a much wider set of potential
factors, which at various times have been important drivers of the price
of oil.

Fig. 7.10 Additional factors that can be included in the oil fair-value model
The factors are grouped into three categories that represent
tradable market prices of different assets, macroeconomic fundamental
factors, and oil-specific factors. Some non-tradable fundamental factors
may not be available on the daily basis, so a broader macro model can
only be implemented at a lower frequency, which does decrease its
performance characteristics. Another challenge comes from the fact
that very few of these factors are cointegrated with oil, making it
difficult to use price-level regressions of the form (7.5). In this case, one
can switch to using regressions on price changes which makes all
variables stationary.
One can take this idea a step further and make the factor-selection
process more dynamic in an attempt to identify the most relevant
factors during each period. This can be done by running individual
single-factor regressions on price differences and ranking factors based
on corresponding R2. One can then estimate the fair value of oil using a
multidimensional regression against several factors that have the
highest R2 in the previous period. However, we would caution against
the mechanical application of such a factor-selection approach given
the high degree of collinearity among them. At the end, a quantamental
trader still has the final say on which factors to use in any given period.
This problem of dynamic factor selection appears to be a good
candidate for using more sophisticated statistical techniques, such as
the methods of machine learning. While the author spent a
considerable amount of time in trying to apply these methods, so far,
the results produced by more advanced statistical models are only
marginally better as compared to a simple static three-factor model.
However, with recent advances in data science, this area of research
continues to evolve. If this book ever makes it to its second edition, the
chapter on application of machine learning methods will most certainly
be expanded. Until then, we can say that our analysis of relatively
simple linear trading strategies is now completed. We can move to a
more advanced, but also a more lucrative universe of nonlinear option
strategies.

Footnotes
1 As of December 2022, the relative importance of the motor fuel category was 3.275%, out of
which 3.172% was gasoline. Source: US Bureau of Labor Statistics (BLS).

2 The interpolation formula is , where tM is the


reference index for the first day of the calendar month in which t falls, tM + 1 is the reference
index for the first day of the calendar month immediately following t, and D is the number of
days in the month in which t falls.

3 Since inflation is measured on a calendar monthly basis, future prices must be interpolated
between two contracts with surrounding maturities. To calculate the base gasoline price, one
must also use linear interpolation similar to that described in the prior footnote for the
reference inflation index.
Part III
Volatility Trading
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
I. Bouchouev, Virtual Barrels, Springer Texts in Business and Economics
https://doi.org/10.1007/978-3-031-36151-7_8

8. Options and Volatilities


Ilia Bouchouev1
(1) Pentathlon Investments, LLC, Westfield, NJ, USA

– Options are nonlinear derivatives that inspired the development of


modern probability theory. An option price is driven by uncertainty,
which is measured by volatility. It is essential to distinguish between
local volatility, realized volatility, and implied volatility.
– Local volatility is a functional parameter of the diffusion process that
describes the behavior of futures prices. In contrast to one-
parameter normal and lognormal models, local volatility of a general
diffusion process is a function that depends on time and the futures
price.
– Options can be replicated by dynamically trading futures. The
original Black-Scholes-Merton framework extends to general
diffusions. The price of an option satisfies the equation of heat
transfer where conductivity of a non-homogeneous medium is
replaced with the local volatility function.
– Realized volatility is typically measured by the standard deviation of
asset returns for a particular realization of a diffusion process. It is a
noisy, backward-looking, and skewed statistic. In the oil market,
volatility of price changes is more meaningful than volatility of
percentage returns.
– Implied volatility is a forward-looking plug designed to tweak an
inaccurate pricing formula to make it match the market price. Its sole
purpose is to convert option prices into a more convenient
coordinate system. Volatility smile measures how far the model
deviates from the market.
8.1 Options and “Théorie de la Spéculation”
By some accounts, options embedded in commodity transactions
predate any other financial derivatives. Perhaps the most famous
example of an early commodity option trade is attributed to Thales of
Miletus, who around 550 BC, according to Aristotle, paid a little money
to secure the exclusive use of all olive presses in the towns of Chios and
Miletus. After an unexpectedly bumper harvest that the philosopher
claimed to have predicted using weather patterns, the option paid off
handsomely, as the demand for olive presses surged. The profit from
the option not only took the philosopher out of poverty, but it also
proved to sceptics that science may indeed be a worthwhile endeavor.
The payoff of an option contract is asymmetric, somewhat
analogous to a lottery ticket. The buyer pays a relatively small premium
and gets compensated if the price rises above a certain threshold in the
case of a call option, or if the price falls below a given threshold in the
case of a put option. This threshold, which is crucial for valuing an
option, is called the strike price.
More formally, a European call option C(F, t) pays the difference
between the futures price F(T) at time t = T and the contractually
specified strike price K if this difference is positive, and, otherwise, it
expires worthless:
(8.1)
Likewise, a European put option P(F, t) at time t = T pays the amount
by which the future falls below the strike price if this amount is
positive, and it pays zero, otherwise:
(8.2)
The payoffs of call and put options are nonlinear due to the
pronounced kink defined by the strike price. The option’s nonlinearity,
or its convexity, is the highest when the strike price is located near the
current futures price. An option with the strike K = F(t) is called at-the-
money (ATM). It is typically the most liquid option which serves as a
primary anchor for option prices with other strikes. Many options are
initially struck out-of-the-money (OTM) with K > F(t) for calls, and
K < F(t) for puts. OTM options are cheaper, as futures must cross some
distance before the strike price is reached for the option to start paying
off. If the current futures price exceeds the strike price, i.e., F(t) > K,
then the call is in-the-money (ITM), and likewise, the put is ITM when
F(t) < K.
Nonlinearity is what makes the mathematics more challenging for
options relative to futures. The value of a call option depends on the
probability that the futures price exceeds the strike price at the
expiration of the option. Since the magnitude of the option payoff varies
with the level of futures, one must know not only the cumulative
probability of reaching the strike price, but rather the collection of all
probabilities of futures reaching every possible price level. The sum of
such probability-weighted payoffs for all possible futures prices then
determines the value of the option. The tricky part, of course, is how to
get such probabilities.
The search for probabilities that determine the value of an option
contract created a foundation for the entire modern theory of
probability. Arguably, such a theory was born in 1900 when an
extraordinary “Théorie de la Spéculation” was presented by a French
mathematician, Louis Bachelier, as his doctoral thesis at the Sorbonne.1
For the first time in the history of science, the process of Brownian
motion, or the random motion of particles in a medium, was described
mathematically. Bachelier’s idea was to use Brownian motion to
characterize the random behavior of stocks and options in the financial
market. Remarkably, one of the most groundbreaking scientific
discoveries of all time was made by an options analyst. Only five years
later, Albert Einstein, unaware of Bachelier’s pioneering work, applied a
similar framework to describe the erratic motion of pollen particles
caused by water molecules.
The significance of Bachelier’s work cannot be overstated. Its reach
extends beyond the derivation of the first option pricing formula, which
is still frequently used today in the oil market. More importantly, while
trying to tackle the problem of option pricing, Bachelier introduced the
so-called law of radiation of probability, and derived that the normal
probability density, the famous bell-shaped Gaussian curve, satisfies the
Fourier equation of heat transfer. Thus, the first connection between
the concept of probability and the physics of diffusions that describes
uncertainty in many scientific disciplines was established in the
derivatives market. This connection subsequently gave rise to the
Fokker-Planck equation in physics, the Chapman-Kolmogorov equation
in probability and statistics, and the Black-Scholes-Merton (BSM)
equation in finance. The basics of the diffusion process and the
corresponding probabilities are summarized in Appendix A.
To price an option, Bachelier assumed that the speculator is not
expected to make or lose any money by trading the asset that underlies
the option contract. While individual expectations of buyers and sellers
of stocks or futures could differ, the market’s aggregate expectation, he
argued, must be zero, since there are as many willing buyers as willing
sellers. This principle of zero expectations for the speculator’s profit can
be interpreted as the earliest formulation of the efficient market
hypothesis. The fair price of an option is then established under this
assumption that the speculator’s own view about the direction of the
underlying asset price is irrelevant.
Bachelier paid particular attention to an ATM option which he
called a simple option. Interestingly, in his days, only simple options
were available for trading in commodity markets. The search for the
fair value of such an option led Bachelier to what he viewed as the most
important contribution of his entire study, summarized in the following
statement:

The value of a simple option must be proportional to the square


root of time.

The proportionality is characterized by what Bachelier defined as


the coefficient of instability, or what we know today as volatility. His
insight inspired the foundation of stochastic calculus, where the
variance, or the volatility squared, of the variable following the random
walk is proportional to an increment of time.
Bachelier was clearly ahead of his time. Back then traders did not
rule the world, and the science of financial markets was not perceived
to be a real science. Despite such a revolutionary invention, it took
Bachelier more than a decade to secure a permanent academic position.
His remarkable contribution remained largely forgotten until the
1960s, when the modern theory of stochastic calculus was developed
and applied to the analysis of financial markets.
The second big innovation in the theory of options came with
seminal papers by Black and Scholes and by Merton.2 They formally
derived the differential equation for the option price, by showing that
the option can be replicated by dynamic trading of the underlying asset.
In the next section, we apply their argument to a more generalized
setting of diffusion processes that have proven to be particularly useful
for describing the dynamics of oil futures. The derivation of the Black-
Scholes-Merton (BSM) equation confirmed that the expected value of
the underlying asset is indeed irrelevant for the valuation of
derivatives. Under certain additional assumptions, the risk carried by
an option contract can be offset by holding some variable quantity of
futures. As a result, options should be priced in the so-called risk-
neutral world, which is analogous to Bachelier’s principle of zero
speculator expectations. The only parameter that matters for option
pricing is the volatility of the asset, or Bachelier’s coefficient of
instability.
More broadly, volatility in financial markets is defined as the
measure of dispersion of uncertain prices around the mean. In other
words, it is the measure of risk. What makes the concept of volatility
often confusing to outsiders is the somewhat broad usage of the term
volatility. In contrast to the price of a financial asset, which is generally
observable, the volatility of the asset is not observable. Volatility can
only be understood and calculated in connection to a specific model or
assumption. Since the term volatility could mean different things in
different settings, the best way to remove any ambiguity is to reference
it with a clarifying adjective. In this chapter, we introduce three types of
volatility, the local volatility (LV) , the realized volatility (RV) , and the
implied volatility (IV) .

8.2 Local Volatility and Diffusions


Prices of financial assets, including oil futures, are stochastic, which
means that some part of their behavior is uncertain or, in other words,
random. The randomness is described by the probability distribution
for futures prices. For example, a random component of the price
change can be pulled from a normal distribution, whose probability
density is described by the well-known bell-shaped Gaussian curve.
The dynamics of the asset price could also carry a deterministic
component, for example, to characterize the growth trend or mean-
reversion towards a long-term equilibrium price level. Together,
deterministic and stochastic components form the stochastic
differential equation that describes the behavior of prices. We have
already encountered one example of a stochastic equation in Chap. 3
when we modeled oil inventories. To keep this part of the book self-
contained, we reintroduce the topic here in somewhat greater depth
and refer for more technical details to Appendix A.
Perhaps the simplest example of a stochastic process, which was
also used by Bachelier in his thesis, is arithmetic Brownian motion
(ABM). It assumes that future changes in the asset price are normally
distributed. The stochastic differential equation for ABM specifies
changes in futures prices dF over a small increment of time dt in the
following form:
(8.3)
The first term, denoted by μA, is deterministic. It represents the
drift, or the expected futures price change per unit of time. If it were the
only term in the equation, then futures prices would drift linearly in
time with a constant slope μA.
The second term is stochastic. It describes the uncertainty via
independent random increments dz, which is drawn from a normal
distribution with zero mean and whose variance is equal to an
increment of time dt. One can also write the random component dz as

where ε represents noise, or a random number taken from the normal


probability distribution with zero mean and variance equal to one. The
scaling by ensures that the variance of dz is equal to dt. Since
increments dz are independent and variances of normally distributed
variables are additive, the variance of uncertainty grows linearly with
time, and the standard deviation, which is the square root of the
variance, grows as the square root of time. It should also be noted that
for small increments of time the magnitude of is larger than dt,
and therefore, the second, random term in the stochastic equation
dominates the short-term deterministic component of dF.
The magnitude of the uncertainty in the ABM process is controlled
by a constant parameter σA which is often referred to as normal
volatility, or dollar volatility to emphasize its unit of measure. If dz
represents the source of risk, then normal volatility can be thought of
as representing the quantity of risk. To indicate the arithmetical nature
of the Brownian motion, we use corresponding subscripts for drift and
volatility coefficients. Under the ABM assumption, the futures price
drifts at a constant rate μA with uncertainty specified by a constant
volatility of price changes σA. Since, for the oil market, dF is measured
in dollars per barrel, the normal oil volatility σA is measured in dollars
per barrel per unit of time. The distribution of prices generated by ABM
process (8.3) is described by the normal probability density given by
(A.​12).
The assumption of normality to describe price changes is often
problematic for certain financial markets, such as equities, for which
many traditional derivatives models have been tailored. The ABM
process allows asset prices to move indefinitely in either direction and
does not prevent prices even from falling below zero, something that is
impossible for stocks and bonds. However, as we have seen previously,
negative prices cannot be ruled out for a storable commodity, such as
oil. Another challenge with an application of ABM to the equity market
comes from its inconvenient scaling, as a five-dollar price change is
much more impactful for a ten-dollar stock than it is for a hundred-
dollar stock. For equity investors, percentage returns are more intuitive
than absolute changes in stock prices.
To keep modeling standardized across assets and to bypass the
issue of negative stock prices, the financial industry applied the
assumption of normality instead to investment returns, which are
measured in percentages. Such a price dynamics is described by the
following stochastic differential equation, known as geometric
Brownian motion (GBM)

(8.4)
In the GBM case, asset returns are normally distributed, or
alternatively, the logarithm of prices is normally distributed.
Under these assumptions, the percentage return on a fully funded
futures position is expected to grow with a constant rate of return
μG and uncertainty characterized by a constant percentage volatility of
returns σG. For the most part, when a generic reference is made to the
term volatility, it is typically assumed to be σG. It is important to
highlight that unlike the price of the financial asset, the term volatility
can only be understood in reference to a specific model. Here, volatility
is understood as the standard deviation of percentage returns under
the assumption that prices are lognormally distributed. The
corresponding lognormal probability density function is given by (A.​
14).
The lognormal probability density is inherently asymmetric. It is
only defined for positive prices, and it has a much larger right-side tail
compared to the symmetric bell-shaped normal curve. The shapes of
two probability densities are illustrated in Fig. 8.1. While one can justify
using an asymmetric distribution in financial markets that tend to grow
over time, applying it to mean-reverting commodity markets brings
some adverse side-effects. As we explained in the chapters on storage
and hedging pressure, in the oil market, the frequency and the
magnitude of upward and downward price movements tend to be more
symmetric, if not even skewed to the downside.
Fig. 8.1 An example of normal and lognormal probability densities
By and large, the commodity industry took the path of least
resistance and followed the lognormal modeling paradigm prevalent in
the equity markets. Without any doubt, this assumption is extremely
convenient when comparing strategies across different commodities
using a standardized percentage-based volatility metric. No risk
manager would want to deal with volatility measured in dollars per
barrel for oil, in dollars per bushel for corn, or in dollars per ounce for
gold.
The convenience of such a one-size-fits-all measure comes at a
price. It forces traders to look at the oil market through an incorrect
frame of reference imposed by inflexible and highly standardized
operational setups of their trading systems. In contrast, dedicated oil
specialists customize their metrics, which allows them to capture
important salient features of the market and to take advantage of the
framing bias often suffered by generalists. This modeling duality adds
certain pedagogical challenges, as many market phenomena must be
presented in two alternative ways, first as they are seen by the general
crowd through standard but poor-quality lenses, and then, as perceived
by oil professionals equipped with more accurate viewing tools.
Both ABM and GBM assumptions are rather simplistic and chosen
primarily for their analytical tractability. It would be naïve to think that
oil price behavior, which is influenced by a multitude of driving factors,
can be fully characterized by only two parameters, the constant drift,
and the constant volatility. The real world of oil prices is much more
complex. Fortunately, as we will see in the following section, the drift
term in the stochastic equation plays no role in modeling options, and
our entire focus will be on volatility. However, volatility of oil futures is
anything but constant. It varies not only with time, but also with the
level of futures prices. For example, as futures move outside of some
normal price range, absolute volatility tends to rise. This indicates
much higher probabilities of extreme events that are nearly impossible
under either normal or lognormal distributions. In other words, the
distribution of oil prices has fat tails.
To allow ourselves more flexibility in modeling notorious tails of the
price distribution, we assume that the drift and volatility parameters of
the stochastic process are deterministic, but not yet specified, functions
of time and the futures price. Futures then follow a more general
stochastic process described by the following equation:
(8.5)
Such stochastic processes, which arise in many natural sciences and
applications, are called diffusions. In finance, the function σ(F, t) that
characterizes the volatility of the diffusion process is known as the local
volatility function. ABM and GBM processes represent only two special
cases of the more general class of diffusions. The local volatility of the
ABM process is constant σA. For the GBM process, the local volatility
σGF is proportional to the futures prices. All price changes for diffusions
are still driven by the same normally distributed single source of noise
dz.
The dependency of the volatility function on the futures price,
which itself is stochastic, allows diffusions to produce a much wider
universe of probability distributions for future price changes. One
simple way to generate a probability distribution by a diffusion process
is to use Monte Carlo simulations, where multiple sequences of random
variables dz(t) are drawn from the normal distribution. The
corresponding futures prices are then computed using the discretized
version of (8.5) and presented in the form of a histogram.
The science of diffusions largely hinges on one important result,
known as Itô’s lemma. It describes how a function of a stochastic
variable changes for a given change in the variable itself. In the
ordinary calculus, a small change in the value of the function G(F, t) is
approximated by its partial derivatives multiplied by a change in the
function’s arguments, the result known as Taylor’s formula. When one
of the variables F is stochastic, the analogous approximation is given by
the following Itô’s lemma:

(8.6)

This equation resembles the first-order approximation of a regular


function of two variables F and t, except for an additional term that
contains the second derivative with respect to F. In the regular calculus,
this term would have been multiplied by (dF)2 and discarded as a lower
second-order term, but in the stochastic calculus it can no longer be
omitted. The reason for keeping it in the approximation comes from the
property of the random noise dz. Since, by definition, dz is proportional
to the square root of time, the variance of dz is equal to dt. Therefore,
(dF)2 is also of the same order of magnitude as dt, which makes it
necessary to retain this term in (8.6). A rigorous proof of Itô’s lemma is
rather complicated and is not needed for our purposes.
Diffusion processes play a special role in modeling oil prices. They
provide a good balance between the complexity of the model and its
tractability. On the one hand, simple ABM and GBM models are not
sufficiently flexible to capture some important nuances of the oil price
dynamics. On the other hand, many sophisticated models inevitably
introduce too much complexity. As a result, their marginal value added
is dwarfed by detrimental side-effects of an excessive parametrization
that causes instability and often detracts from gaining valuable
intuition.3 The general diffusion framework outlined here appears to be
a sweet spot in modeling oil prices. It presents a relatively minor but
extremely powerful extension of simple ABM and GBM modeling
choices. Perhaps most importantly, the seminal argument of option
replication by delta hedging easily extends to this more flexible class of
diffusion models.

8.3 Delta Hedging and Option Replication


The breakthrough idea developed by Black and Scholes and
independently by Merton was based on an important insight that
follows from Itô’s lemma. Since a financial derivative, such as an option,
is a function of a random variable, Itô’s lemma provides the rule
relating a small change in the option price to a small change in the stock
price. Importantly, both the stock and the option, which is a function of
the stock, are driven by the same single source of uncertainty dz.
Therefore, one should be able to combine the option and the stock in
some smart way that eliminates this uncertainty, at least for a short
period of time. If the entire risk can indeed be eliminated, then in the
absence of an arbitrage, which rules out the existence of riskless profits,
the value of a combined portfolio that includes an option and some
quantity of the stock can only grow at the risk-free interest rate accrued
on the initial investment.
The BSM framework was initially designed for the equity market
and developed under the lognormal GBM assumption. Subsequently, it
was tailored by one of the authors to the futures market.4 The BSM
argument, however, remains intact for all diffusions of the form (8.5),
and we now replicate their framework in a more general setting.
Let C(F, t) denote the price of a call option struck at K that expires at
time T. The call depends on the stochastic futures price F and time t. To
simplify the notation, in this chapter we suppress the price dependency
on K and T, which are set contractually. To construct a mini portfolio of
an option and futures, we need to know how C(F, t) changes over a
small increment dt in response to the change in the futures price dF.
This change is described by Itô’s formula (8.6). Applying it to the
diffusion specification (8.5) results in the following stochastic equation
for the option price:

(8.7)
Note that both the futures price and the option price are driven by
the same source of uncertainty dz. This allows us to combine the option
and the future in a particular way that eliminates the source of
randomness.
Assume that we bought a call option and want to offset some risks
by selling a yet to be determined quantity of futures, denoted by the
Greek letter delta, Δ. We now have a portfolio Π that consists of a long
call option and a short position in Δ units of futures:

To calculate the change in the portfolio value dΠ over time


increment dt, we substitute dF and dC with their corresponding
dynamics given by (8.5) and (8.7) and obtain that

We can now choose the number of futures hedges, Δ, in a very


special way that makes the stochastic term dz disappear from the
equation. This can be accomplished by letting the hedging delta be:

Such a special choice of delta does more than the elimination of


randomness. It also removes the drift term μ(F, t) from the first,
deterministic part of the equation. This insight is the central part of the
options theory. It shows that the price of an option does not depend on
investor expectations, which are ultimately linked to individual risk
preferences. The expectation term μ(F, t) drops out from the equation
and plays no role in the option pricing. This argument developed by
Black, Scholes, and Merton could be viewed as a formal proof of
Bachelier’s original hypothesis that options must be evaluated under
the principle of zero expectations to the speculator. Since the drift term
is no longer present in the equation, the option price remains the same
even if the expected value of the future price changes is set to zero. This
approach became known as risk-neutral pricing of derivatives.
Without the random component dz, the resulting portfolio Π carries
no risk, at least, instantaneously. In the absence of an arbitrage, such a
portfolio can only accrue the risk-free interest rate r. If the portfolio
were growing at any other rate, then traders would have been able to
either borrow money and invest in the portfolio or short the portfolio
and invest the proceeds, making money without any risk, which is
deemed to be impossible. Therefore, over a small period dt, the change
in the value of this riskless portfolio should be equal to the interest
received on the initial investment:

Note that in the last term only the option premium C accrues
interest. The futures contract does not require any initial investment,
besides a small collateral, the impact of which for simplicity is ignored
here. Since it does not cost anything to enter into the futures contract,
there is no interest accrued on the futures position either.
Cancelling the dt term, we obtain the generalized BSM equation for
the option price written on futures, which follows a diffusion process
with local volatility σ(F, t):

(8.8)

This equation applies to any financial derivative on futures. The


specific nature of the derivative is defined by its boundary condition at
expiration T. For example, if the Eq. (8.8) is combined with the
boundary condition defined by the payoff (8.1), then it always admits a
unique solution that determines the price of the call option.
The Eq. (8.8) is analogous to the well-known equation of heat
transfer that describes the dissipation of heat impulse over time with
respect to the spatial variable that characterizes the medium. Here, the
spatial variable is given by the futures price, real time is replaced with
the time remaining to maturity of the option, and the initial condition is
specified by the option’s payoff at maturity. The behavior of option
prices versus futures with the impulse provided by the strike price is
analogous to the dissipation of temperature within the medium. The
local volatility plays the role of the thermal conductivity of a non-
homogeneous medium.
To solve the equation, in general, one must apply numerical
algorithms, such as finite difference methods. Simple analytic solutions
exist only for a few special but important cases. One special case, which
underpins Bachelier’s thesis, is the ABM process for which the local
volatility is constant

The solution to the Eq. (8.8) with the boundary condition (8.1) is
then given by the following Bachelier formula:
(8.9)
Here, standard notations are used for the normal Gaussian
probability density with zero mean and variance equal to one:

and the cumulative normal distribution function, which is

The quantity

represents time remaining to maturity, and

defines the normalized moneyness of the option, which scales the


distance between the futures price and the strike price by the total
volatility over the life of the option. The normalized moneyness mA can
be understood as the number of standard deviations that an option is
either ITM or OTM.
The Bachelier formula reduces to a particularly simple form for an
ATM option, for which F = K:
This formula reflects Bachelier’s original discovery that the value of
a simple option must be proportional to the square root of time.5
Without any doubt, the development of the entire modern probability
theory was inspired by this result.
In Appendix B, we provide more details on deriving the Bachelier
formula as the solution to the differential Eq. (8.8) by integrating the
option’s payoff with the normal probability density. Alternatively, one
can simply verify that it indeed solves (8.8) by calculating partial
derivatives of (8.9), which are also given in Appendix B, and
substituting them directly into the Eq. (8.8). These derivatives play an
important role in understanding the dynamics of option pricing. They
are known as Greeks, as they are typically labeled by letters of the
Greek alphabet.6 We have already introduced delta in the derivation of
the BSM equation, which is the first derivative of the option price with
respect to the futures price. The second partial derivative of the option
with respect to futures is gamma, which measures the degree of
convexity in the option’s payoff. Theta, the derivative with respect to
time, shows how quickly an option value decays as time passes. Vega is
the derivative with respect to the most important input into the pricing
model, the volatility of the stochastic process.7
The second important special solution to the pricing equation for
diffusions is given for the GBM case, where the local volatility is
assumed to be proportional to the futures price:

The solution to the Eq. (8.8) with the boundary condition (8.1) is
then given by the Black formula

(8.10)

The Black formula can also be derived by integrating the payoff with
the lognormal probability density or, alternatively, it can be verified by
the direct substitution of its partial derivatives into (8.8). The details
are provided in Appendix B.
The normalized log-moneyness term mG in (8.10) is defined as the
logarithmic ratio of the futures price to the strike price scaled by the
geometric volatility over the life of the option:
So far, we have only dealt with the valuation of a call option.
However, the Eq. (8.8) applies to any derivative of the futures price. The
fact that it is a call option was only specified by its boundary condition
at maturity. If the boundary condition (8.1) is replaced with (8.2), then
similar formulas can be obtained for a put option. However, an easier
way to derive pricing formulas for puts would be to utilize an important
no-arbitrage relationship that links prices of call and put options.
If we construct a portfolio which is long a call option and short a put
option with the same strike K, then the payoff of this portfolio at
expiration time T can be written as
(8.11)
In other words, holding this portfolio is identical to holding a long
futures position established at the price K. Since the portfolio of a long
call and a short put is itself a financial derivative of the futures, then the
solution to (8.8) with the boundary condition (8.11) is simply the
discounted futures payoff, reflecting the fact that option premia are
typically paid upfront at the initiation of the trade:
(8.12)
This formula, known as a put-call parity, allows one to determine
the price of a European put option given the price of a European call
and the price of futures. Alternatively, one can determine the price of a
call given the price of a put along with futures.
We have chosen to repeat the standard BSM replication algorithm
not because of its mathematical elegance, but rather because it
provides an explicit recipe for trading volatility. This book is not meant
to be a comprehensive reference of derivatives pricing models. Our goal
is to highlight certain features of these models that can be turned into
profitable strategies, specifically in the oil market. Not all models, of
course, present such unique trading opportunities. As such, to gain
more clarity in already complex topics, we make some simplifying
assumptions in areas where we do not see any particularly unique
trading opportunities.
For the most part, we ignore the impact of the interest rate,
assuming it to be zero. Unlike stocks and bonds, the interest rate plays
only a relatively minor role in the valuation of options on oil futures.
For European options that can only be exercised at expiration, the
impact of the interest rate amounts merely to the discounting factor
that appears as the multiplier in pricing formulas. The discounting
comes from the fact that an option premium is assumed to be paid
upfront, but the settlement of the option occurs later, at the expiration.
In the oil market, the premium is paid upfront only for WTI options,
while for exchange-traded Brent options, the premium is netted against
the final settlement. Therefore, Brent options are marked-to-market
like futures, in which case the discounting factor should be removed. It
is interesting that this was also the case for options considered by
Bachelier, which is why the interest rate was also omitted in his study.
Regular exchange-traded oil options are American options that can
be exercised at any time prior to the expiration. Thus, an American
option must be slightly more expensive than an equivalent European
option. The precise calculation of an early exercise premium is rather
complicated as the partial differential equation for the price of an
option turns into an inequality, where one needs to solve a complex free
boundary value problem or to apply other optimization techniques.8 To
simplify this, several analytical approximations have been developed,
and the one developed by Barone-Adesi and Whaley has proven to be
adequate for the oil market.9 These models show that it is optimal to
exercise an American option early only if the benefits from receiving a
non-discounted cash settlement sooner outweigh an additional value
from holding an option longer. The latter is driven by the remaining
volatility. Since the impact of oil volatility on the option price is much
larger than the contribution of the interest rate, it is rarely optimal to
exercise an American option early, except for deep ITM options at times
of high interest rates. Furthermore, under our simplified assumption of
zero interest rates, the early exercise premium is equal to zero.
While the interest rate does matter for oil trading, using it
generically in pricing formulas can cause more harm than it adds value.
The interest rate for all traders is nearly always tied to the cost of
funding specific to the trading company and to customized collateral
agreements with clearing brokers and counterparties. Some strategies
can indeed be enhanced by optimizing financing costs, but they are
ultimately inseparable from the management of the trader’s balance
sheet. For example, when a long-dated ITM option has a significant
intrinsic value that cannot be withdrawn, which is sometimes called the
trapped option value, an auxiliary agreement is often structured to use
it as the collateral for other trades. The true value of the early exercise
premium becomes heavily dependent on the details of such agreements
and remains highly tailored to the needs of individual traders. In other
words, this area is much closer to the field of structured finance than to
the topic of oil trading, and for this reason we do not consider it in the
book, mostly assuming zero interest rate, unless noted otherwise.
Finally, for some theoretical arguments we generally assume that
the maturity of the option coincides with the maturity of the underlying
futures. In practice, however, standard exchange-traded oil options
expire three business days prior to the expiration of the corresponding
futures. To simplify the exposition, in this chapter we ignore the
relatively minor impact of the three-day maturity mismatch between
options and futures and assume that both expire at time T. In Chap. 11,
we will cover a more general case of options whose expiration is
decoupled from the expiration of the underlying futures.
In the remainder of this chapter, we discuss two other types of
volatility, realized and implied, both of which can only be understood in
the context of particular simple modeling assumptions. Our initial goal
is to highlight potential challenges and pitfalls with blind application of
such simplified frameworks to the oil market. The actual recipes for
trading that feature models specifically tailored to opportunities in the
oil markets will be provided in the following five chapters of the book.

8.4 Realized Volatility


To price an option, we need to know the local volatility function of the
underlying diffusion process. The local volatility is not directly
observable, and it is rather difficult to estimate. Many traders start with
an easier route and attempt to get some sense of how this function
might look by measuring its traces from the historical time series of
futures prices. Unfortunately, this route is prone to dangerous pitfalls.
The purpose of this section is to caution traders against relying too
much on backward-looking volatility estimates in valuation of forward-
looking options. One should always remember that history is only one
particular realization of an unknown stochastic process. History does
not have information about the entire process. One can think of history
as a single random path among thousands of other paths generated by a
Monte Carlo simulation.
The volatility estimated statistically from the historical time series
is known as the realized volatility. It is usually defined as the standard
deviation of historical percentage returns over a particular lookback
period. Since historical returns could be calculated over different
frequencies and time horizons, the realized volatility is annualized.10
The realized volatility is typically measured on the rolling basis, where
the size of the data sample is fixed, but each day the new return is
added and the oldest one is removed.
The realized volatility is extremely sensitive to the choice of the
lookback period. The shorter the lookback period, the more volatile the
realized volatility is. The short-term realized volatility is highly variable
due to the large sampling error, which makes it difficult to use for any
investment decisions. The longer-term realized volatility is more stable,
but such a slow-moving estimate is rarely suitable for traders, whose
investment horizon tends to be much shorter.
Figure 8.2 illustrates different measures of the realized volatility for
the third nearby WTI futures calculated using one-month, three-month,
and one-year rolling windows.
Fig. 8.2 Realized volatilities of third nearby WTI futures computed for one-month, six-month,
and one-year lookback periods with highly visible “volatility ghost” in 2020
The calculation of the realized volatility is also very sensitive to
outliers, or particularly large price moves that have occurred in the
past. The realized volatility calculated using a rolling window could
drop nearly instantaneously when a single large historical return moves
out of sample. To illustrate, consider different calculations of the
realized volatility in 2020, shown in Fig. 8.2. The realized volatility for
all futures contracts spiked during several consecutive days in April
2020 after spot oil prices went negative, and futures across all
maturities moved down sharply. Subsequently, this event continued to
haunt six-month realized volatility for exactly six months, and one-year
realized volatility for exactly one year, leading to an abnormally high
volatility estimate until it suddenly dropped when the event fell out of
sample. Traders often refer to this phenomenon as volatility ghost.
This naïve method of estimating realized volatility does not
differentiate between an event that occurred a while ago at the
beginning of the sample and one that just happened yesterday. The
calculation of the standard deviation weighs all historical returns
equally, which creates a dangerous problem if such an estimate is used
for pricing an option. Some traders like to use the realized volatility
calculated over the lookback period that matches the investment
horizon, which for options is typically determined by time remaining to
maturity. For example, to value an option that expires N days from
today, they may use N-day realized volatility. This approach is very
problematic. If a large price move happened exactly N days ago, then
the realized volatility calculated tomorrow will be drastically different
from the one calculated today, as this large price move drops out of
sample. While the historical volatility calculation changes when the
sample shifts by a day, the market expectations of the future volatility
are unlikely to be very different from expectations the day prior.
The problem with equally weighted returns can be somewhat
mitigated by using instead an exponentially weighted average of
historical returns, where the largest weight is assigned to the latest
return and prior returns enter the calculation with exponentially
decreasing weights. In this case, the size of the lookback window
becomes less relevant as the contribution of past returns exponentially
approaches zero. The effect of gradually dissipating large prior returns
makes the volatility estimate more representative of the current market
conditions. The exponential parameter provides an additional degree of
freedom that controls the persistence of the short-term shock.
Many more sophisticated techniques have been proposed to
improve volatility estimation and forecasting using historical futures
data. However, these methods are not widely used by option traders in
practice. At the end, the realized volatility is only an estimation method
of the local volatility of the unknown stochastic process. For the most
part, historical volatility estimates tend to be poor predictors of future
volatility. An estimate of the realized volatility could mean different
things for different price distributions. If the nature of the stochastic
process that drives the underlying price distribution is not known, then
any interpretation of the realized volatility could be rather dubious.
Studying one realization of the unobservable stochastic process simply
does not get us too far in terms of volatility forecasting regardless of
how sophisticated the statistical model is.
Another challenge with the conventional application of a realized
volatility estimate to oil options comes from measuring uncertainty in
terms of percentage returns. Figure 8.3 shows that the volatility of oil
returns has strong inverse dependency on the price level.
Fig. 8.3 Realized 3-month volatility of third nearby futures (WTI, 2000–2022) spiked during
the periods of oil busts during which the price decreased by more than 50% from the previous
peak
The inverse relationship between prices and percentage-based
volatility measures is particularly vivid during the so-called oil busts.
Simplistically, we define oil busts as periods during which the price of
oil fell by more than 50% from its recent peak. There were three such
periods since the beginning of the century. First, it happened in the
aftermath of the global financial crisis, then at the time when OPEC
unexpectedly increased production to protect its market share from the
rapid growth of US shale, and finally when oil demand collapsed in the
beginning of the Covid-19 pandemic. The realized volatility jumped
during these periods of falling prices. This highlights an important
difference between oil and many other consumption commodities,
which tend to be more volatile when prices are high, and risks of supply
disruptions rise. In contrast, crude oil tends to be more volatile when
prices are lower, at least, if the volatility is measured in percentage
terms.
To see this effect even more clearly, Fig. 8.4 shows the relationship
between realized volatility and oil prices in the form of a scattergram.
The realized volatility as a function of price exhibits a strong negative
skew, something which is more typical for financial markets, such as
equities. As mentioned previously, these markets follow an up the stairs,
down the elevator dynamics which is characterized by frequent small
gains and less frequent but large losses. When it comes to oil trading,
measuring risk with volatility of percentage returns creates a strong
artificial bias which is caused by looking at the uncertainty through the
wrong lenses.

Fig. 8.4 Realized 3-month volatility of percentage returns versus third nearby futures (WTI,
2000–2022)
The problem with traditional percentage-based measures of
volatility starts with the definition of an investment return, which is
more ambiguous for trading futures. Since the only investment required
to enter into a futures contract is a relatively small initial margin posted
with the clearing broker, in theory, this posted margin should be used
as an initial investment in the definition of return. In practice, margin
requirements fluctuate along with prevailing risks and using the margin
for the standardized denominator of a return would be rather
cumbersome. Many financial analysts simply ignore this specificity of
futures trading, and for convenience, use the notional value of futures
in the definition of returns and subsequent calculation of volatility.
While such percentage-based metrics make sense for financial assets
that tend to grow over time, applying them to mean-reverting
commodity prices creates an artificial framing bias that could adversely
impact decision making. When oil prices are low, the division by a small
number blows up the magnitude of percentage returns, and even makes
calculation mathematically impossible if the asset price becomes
negative.
Since the profitability of highly leveraged futures trading is
measured by professional oil traders nearly exclusively in dollar terms,
corresponding risks are better measured accordingly. What oil traders
care about is the volatility of profits and losses driven by price changes,
not the volatility of somewhat artificially constructed investment
returns. If realized volatility is calculated instead as the standard
deviation of price changes instead of percentage returns, then this
results in a more accurate measure of uncertainty. We refer to such a
measure as the realized dollar volatility.
As shown in Fig. 8.5, the artificial negative skewness of the volatility
with respect to the price level disappears. The relationship between
volatility of price changes and futures prices becomes more intuitive.
The volatility tends to be relatively low when oil prices remain in some
normal range, but it increases sharply when prices move away from
such a range in either direction. In other words, the relationship
between the realized dollar volatility and prices is somewhat parabolic,
the insight that we will use later in developing an appropriate model for
pricing oil options.
Fig. 8.5 Realized 3-month volatility of price changes measured in dollars per barrel versus
third nearby futures (WTI, 2000–2022)
Figures 8.4 and 8.5 highlight the importance of choosing the correct
lenses through which to look at oil volatility. One measure is rather
poor, but it is often chosen for its operational convenience and for its
easy standardization across asset classes. Another one is a more
customized metric tailored to the specifics of the oil market.
Fortunately, for many practical purposes it is rather straightforward to
go back and forth between the two metrics. Since an investment return
is defined as the ratio of a price change to the initial futures price, the
volatility of returns can often be approximated by the ratio of the
volatility of price changes to the price level.
This distinction between alternative ways of measuring volatility
turns out to be vital for trading oil options. Switching the calculation of
the historical volatility from percentage returns to price changes does
not, of course, resolve any conceptual problems with the realized
volatility concept, but, at least, it removes an artificial skewness of risk
created by volatility of percentage returns. This skewness becomes
even more visible and problematic when statistical estimates of the
volatility of stochastic process are replaced with an alternative method
of estimating volatility by deducing it from the options market.
8.5 Implied Volatility and its Skew
Any model of financial markets is only a theoretical construct designed
to approximate the real behavior of prices. A typical financial model
transforms a certain set of inputs by means of some quantitative
machinery into a description of market prices. Regardless of how good
the machinery is, the output of the model can only be as good as the
quality of its inputs. The primary input to the models for option prices
is the volatility of the stochastic process that governs the dynamics of
the underlying futures contract. As discussed in the previous section,
estimating this input from history leaves a lot to be desired, as
historical realized volatility is all over the place. One can, of course,
build more complex statistical models to forecast volatility, but
regardless of the econometric technique, the resulting option price
based on the volatility estimate is likely to be different from the one
observed in the market.
When a choice is to be made between the model and the market,
practitioners tend to assign higher powers to the market. To reconcile
the model with the market, traders came up with a clever workaround
and turned the option pricing problem upside down. Instead of betting
on the model driven by a noisy input, they use the model in reverse and
back out the missing volatility input from the market price of an option.
In other words, the pricing formula is matched to the observable option
price and then inverted for volatility. The resulting measure of volatility
is known as the market-implied volatility.
The general pricing formulas (8.9) and (8.10) cannot be analytically
inverted for volatility, and implied volatility must be calculated
numerically using a root-finding algorithm. Since the option’s vega,
which is the derivative of an option price with respect to its volatility
input, is strictly positive, and an option price is a monotonically
increasing function of volatility, such numerical inversion is always
possible. Therefore, every option price is uniquely mapped to its
corresponding implied volatility, and one can view implied volatilities
as an alternative metric for option prices.
The market-implied volatility presents an alternative method of
estimating some properties of unobservable local volatility of the
stochastic process. In contrast to a backward-looking realized volatility
estimate, implied volatility is a forward-looking measure. It is simply a
plug to a particular option pricing formula that forces the model to
match the market. Only in a very special and somewhat unrealistic case,
when the futures market happens to behave exactly as prescribed by
the model, can implied and realized volatilities be meaningfully
compared. The implied volatility can then be interpreted as the market
expectation of the future realized volatility. But even in this case, the
two volatilities can still be very different if the option price contains an
additional risk premium caused by hedging imbalances. One can also
think of the implied volatility as the average of the expected local
volatility, adjusted by the volatility risk premium. We will discuss these
topics in more detail in subsequent chapters.
To distinguish the market-implied volatility from other types of
volatility, we denote it by a different letter. For an option with the strike
price K and expiration T, we use v(K, T) to represent the market
standard implied Black volatility (IBV), which is computed by inverting
the Black pricing formula (8.10). We will also use the less common, but
arguably more important measure of implied normal volatility (INV) or,
alternatively, implied dollar volatility, which we denote by vN(K, T). The
INV is backed out from the same market price of an option by inverting
the Bachelier formula (8.9). Regrettably, the market rarely explicitly
attaches Bachelier’s name to the normal volatility, and we reluctantly
follow the market lingo and adopt the term normal volatility. In
contrast to IBV, which is measured in percent, INV is expressed in
dollars per barrel.
The irony of the market standard IBV metric is that it debunks the
main assumption of the model that is responsible for its own creation.
For a given maturity, there are many options with different strikes, but
there is only one futures contract. The lognormal assumption of
constant proportional volatility is a property of the futures contract,
and it has nothing to do with options. If the Black model were correct,
then the inversion of any option price should result in the same
number, the same constant volatility of the futures that the model is
based on. In practice, implied Black volatilities computed for options
with different strikes and maturities are nearly always different.
A typical dependency of implied Black volatilities v(K) on the strike
price is shown in Fig. 8.6 for short-term, medium-term, and long-term
oil options. The graph is called the volatility smile. The reference to a
smile comes from its origins in the foreign exchange market, where the
plot is often more symmetric with similar curvature on both ends,
making it look like a smile. In the oil market, this graph is also known as
the volatility skew to highlight its predominantly negative slope, except
for the utmost right tail for shorter-term options, which makes the
graph look more like a smirk. The curvature of the smile on both ends
indicates that the options market expects a higher frequency and
magnitude of extreme events than can be generated by the lognormal
distribution. The market adjusts the pricing model by using higher
volatility input to capture such events. We will discuss alternative
normal volatility smiles and develop a more accurate model that
captures the fat tails of the price distribution in Chap. 10.

Fig. 8.6 Representative implied volatility skews for short-term, medium-term, and long-term
oil options
It is important to emphasize that the implied volatility smile is
nothing more than a collection of option prices presented in a more
convenient coordinate system. Simultaneous tracking of prices for all
options with different strikes and maturities in dollars per barrel would
be next to impossible, as option prices are constantly changing along
with futures based on their corresponding deltas. In contrast to many
statistical forecasting models, the main purpose of the Black model is to
translate market observable option prices to a different and more
convenient scale. Such a transformation does not require any statistical
methods.
Even though the implied volatility smile has a negative skew with
puts generally having higher implied volatilities than calls, this does not
mean that puts are overpriced relative to calls. It is simply another
artefact of looking at the market through incorrect lenses. The
skewness of the IBV curve versus strikes is a side-effect of the same
wrong unit of measure, previously illustrated by Fig. 8.4, where the
realized volatility was plotted versus futures. If the true oil price
distribution happened to be more symmetric, then observing a
phenomenon through skewed lognormal lenses would force us to
adjust our perception of option prices by raising percentage volatility
for lower strikes and lowering it for higher strikes. The lognormal
skewness is akin to prescribing lenses for astigmatism to someone who
does not need them, which will result in blurry vision. In Chap. 10, we
will construct more appropriate lenses to view the oil options market,
but for now, we proceed down the conventional route of expressing
option prices in terms of their IBVs.
Choosing the right metric is critical in trading as the frame of
reference has a strong impact on traders’ decision making. The
volatility smile, shown in Fig. 8.6 as a function of strike K, works well as
a static snapshot of option prices. However, this smile is more difficult
to track dynamically once futures move, as the range of actively traded
strikes also shifts. The most liquid option is typically ATM and the
strikes for many other options are often chosen by traders relative to
ATM, which makes it more practical to maintain the smile as a function
of option moneyness instead of fixed strikes. The simplest way to define
moneyness would be to use either the difference between the strike
price and futures, K − F, measured in dollars per barrel, or their ratio,
K/F, in percent. Such choices are intuitive and indeed both are often
utilized for short-term analysis and for options with the same
expiration. However, simple moneyness metrics become more
problematic when comparing implied volatilities across multiple time
horizons.
In both the Bachelier and Black formulas, the volatility is scaled
with the square root of time to maturity, . Therefore, when we back
out implied volatility from option prices, we always divide by . This
magnifies the resulting smile for shorter-term options and flattens it for
longer-term options, as clearly seen in Fig. 8.6. However, what matters
for option pricing is volatility-normalized moneyness, such as mA for
the normal distribution, and mG for the lognormal one. If we use such a
normalized moneyness as the primary unit of measure, then volatility
smiles will be more accurately compared across all maturities.
We could have stopped here and accepted a normalized moneyness
setup as the base case for tracking the smile, but option traders often
take it one step further and define moneyness instead directly in terms
of their hedging deltas. The deltas for Bachelier and Black models,
which are given in Appendix B, effectively transform volatility-
normalized moneyness to make it vary between minus one and plus
one. Keeping volatility as a function of delta has proven to be handy for
traders as the required hedging ratio comes directly as a biproduct of
the smile setup. For example, to hedge the sale of 100 units of 25-delta
calls, one would need to buy 25 futures. Black deltas are often used as
the market primary communication tool, even though traders typically
make further adjustments to the actual delta hedging, which we discuss
in more detail in Chap. 10.
To illustrate historically observed oil smiles, Fig. 8.7 shows the
average of IBV smiles for 10-, 20-, 30-, and 40-delta OTM puts and calls
and zero-delta straddle since 2000. Note that because of an asymmetry
embedded in the lognormal distribution, the Black delta for an ATM
straddle is counterintuitively not equal to zero. To preserve the uniform
spacing in the graph of the smile, we use a zero-delta straddle instead
of an ATM straddle.
Fig. 8.7 Average IBV versus delta-moneyness for different maturities (WTI, 2000–2022)
Average IBV smiles are shown for different days to maturity (DTM)
that correspond to one-, three-, six-, and twelve-month options. Implied
volatilities are always calculated using more liquid OTM options. Less
liquid ITM options trade infrequently, and their prices are usually
synthetically derived from corresponding OTM options and the put-call
parity relationship (8.12). We should also acknowledge that
maintaining volatilities as a function of delta has its own challenges, as
such a setup is somewhat circular. While volatility is commonly
measured versus delta, the delta itself depends on volatility. In practice,
implied volatilities are first calculated for fixed strikes along with
corresponding deltas, and then subsequently interpolated for the
desired grid of deltas.
It is also clear from Fig. 8.7 that for each moneyness, implied
volatility declines with increasing time to maturity. However, unlike the
contradictory presence of the skew for any given futures, the
decreasing term structure pattern of implied volatilities does not cause
any alarms, as it reflects volatilities of different futures which do not
have to be the same. In fact, the volatility should naturally decline for
longer maturities futures as short-term fundamental uncertainty fades
away while being smoothed out by the balancing role of storage. This
phenomenon is known as the Samuelson effect.11 The Samuelson effect,
which was originally observed for the realized volatility, applies to the
implied volatility as well. Under the normal market conditions, the term
structure of implied volatilities is expected to be a decreasing function
of time to maturity, as illustrated by the middle line in Fig. 8.8.

Fig. 8.8 While oil volatility generally declines with time to maturity, the short-term implied
volatility is more affected by realized volatility
If short-term realized volatility is particularly high, which, for
example, could be driven by falling oil prices, then the slope of the
implied volatility term structure steepens, as long-maturity futures
move less. This effect is exacerbated by a percentage-based volatility
metric, as weaker prices generally result in a contango market when
short-term futures prices fall below long-term futures. To compensate,
the percentage volatility must be raised for contracts with lower prices.
When volatility is measured instead in dollar terms, the volatility term
structure is also typically decreasing but with a flatter slope. During
periods of low realized volatility, the implied volatility term structure
may also exhibit a visible hump. The hump may reflect not only market
expectations of eventual increase in volatility, but also the risk premium
embedded in oil options, which we will study in the following chapter.
In Chap. 11, we will also see how critical the shape of the volatility term
structure is for pricing many exotic options.
To summarize, so far, we have only adapted some standard
methodologies for pricing options in the oil market. We chose to model
futures prices in the setting of general diffusions characterized by the
local volatility function. This framework has proven to be the sweet
spot in the oil market, providing a reasonable trade-off between model
accuracy and complexity. We introduced and discussed two essentially
defective, but, nevertheless, commonly used estimates of the local
volatility of the diffusion process. The realized volatility estimates it by
taking a look at the history, while the market-implied one attempts to
extract it from an unknowable future. Both volatilities can only be
understood in the context of simplified and largely unrealistic
assumptions about the distribution of futures prices.
The next step often taken by many amateur traders and occasionally
by some textbook writers is a comparison of one defective metric to
another, as an attempt to determine whether an option is cheap or
expensive. Our preference is to stay away from making such a naïve
comparison, as in the volatile oil markets, this simplified approach can
potentially bring more harm than value. While the difference between
implied and realized volatilities could indeed provide some information
about the richness of the option, the question of the fairness of the
option price cannot be properly answered without a careful
examination of the option’s gamma, the measure of its convexity.

References
Alexander, C. (2008). Market risk analysis, Vol. III: Pricing, hedging and trading financial
instruments. Wiley.

Bachelier, L. (1900). Théorie de la Spéculation, Annales scientifiques de l’Ê cole Normale


Supêrieure, Serie 3, 17, 21–86.

Barone-Adesi, G., & Whaley, R. E. (1987). Efficient analytic approximation of American option
values. Journal of Finance, 42(2), 301–320.
[Crossref]

Black, F. (1976). The pricing of commodity contracts. Journal of Financial Economics, 3(1/2),
167–179.
[Crossref]
Black, F., & Scholes, M. (1973). The pricing of options and corporate liabilities. The Journal of
Political Economy, 81(3), 637–654.
[MathSciNet][Crossref][zbMATH]

Hull, J. C. (2018). Options, futures, and other derivatives (10th ed.). Pearson.
[zbMATH]

Leoni, P. (2014). The Greeks and hedging explained. Palgrave Macmillan.


[Crossref]

Merton, R. C. (1973). Theory of rational option pricing. The Bell Journal of Economics and
Management Science, 4(1), 141–183.
[MathSciNet][Crossref][zbMATH]

Samuelson, P. A. (1965). Proof that properly anticipated prices fluctuate randomly. Industrial
Management Review, 6(2), 41–49.

Wilmott, P., Dewynne, J., & Howison, S. (1993). Option pricing: Mathematical models and
computation. Oxford Financial Press.
[zbMATH]

Footnotes
1 Bachelier (1900).

2 Black and Scholes (1973), and Merton (1973).

3 Some energy commodities, such as natural gas or power prices, are more susceptible to large
short-term spikes where diffusions are often combined with jump processes, but this modeling
paradigm is more complex as options can no longer be dynamically replicated with futures. We
discuss some limitations of diffusions in Chap. 12.

4 Black (1976).

5 For the type of options studied by Bachelier, the discounting was not necessary as the option
premium was netted against settlement at expiry. As a result, in his original derivation the
interest rate was ignored. We will explain shortly why a similar assumption can be made for
pricing many oil options. The discounting factor brings in an additional time dependency
resulting from the present value of money, rather than from the evolution of the variance.
6 Greeks are discussed in many standard derivatives textbooks, such as Alexander (2008) and
Hull (2018). For their practical interpretation, see also Leoni (2014).

7 Vega is not a letter of a Greek alphabet but it was adopted by option traders for its phonetic
similarity.

8 See, for example, Wilmott et al. (1993).

9 Barone-Adesi and Whaley (1987).

10 For example, if daily prices are used then realized volatility is the standard deviation of
returns multiplied by given approximately 250 trading days in a year, and for weekly
prices, it is multiplied by .

11 See Samuelson (1965).


© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
I. Bouchouev, Virtual Barrels, Springer Texts in Business and Economics
https://doi.org/10.1007/978-3-031-36151-7_9

9. The Hidden Power of Negative


Gamma
Ilia Bouchouev1
(1) Pentathlon Investments, LLC, Westfield, NJ, USA

– End-users and speculators often evaluate options actuarially, like an


insurance contract. Dealers, on the other hand, price options based
on the cost of their dynamic replication. The pricing dichotomy
results in alternative valuations and motivates trading among
different market participants.
– Option delta hedging is riskless only in an idealized setting. In
practice, the replication strategy is driven by the option’s gamma. Oil
gamma is dominated by producer demand for downside price
insurance, which contributes to a negative skewness of futures
returns.
– The volatility risk premium (VRP) can be extracted by selling and
dynamically delta hedging oil options. Its magnitude depends on
option moneyness and time to maturity. The strategy profitability is
characterized by the VRP smile and the VRP term structure.
– Option traders say that they do not need a model for pricing. The
price is given by a broker, but they need a model for hedging. The
option replication model is particularly sensitive to the hedging
frequency and to the volatility input used for delta calculation.
– VRP strategies evolve along with the changing needs of large market
participants. As the market matured, directional VRP became less
appealing, but new opportunities developed in trading relative VRP
strategies across moneyness and maturities.
9.1 Options and Insurance
Options can be viewed as insurance contracts. A relatively small
premium is paid upfront to ensure financial protection against some
adverse events in the future. A typical insurance contract pays off only
after a certain deductible is met, which is the responsibility of the
buyer. The insurer and the insured are effectively sharing the risks,
where the deductible sets the threshold for splitting their joint liability.
A smaller deductible means that larger risks are taken by the insurer
and a higher insurance premium is charged for taking on such risks. In
the jargon of commodity options, the equivalent of the deductible is the
moneyness of the option, determined by its strike price.
The seller of a commodity option, like the seller of an insurance
contract, is exposed to highly asymmetric risks. The seller receives a
steady income in the form of insurance premiums paid by the buyers.
To motivate the seller to take on undesired risks, the premium includes
some service surcharge on top of the fair price that reflects the risks. In
the long run, selling insurance is generally a profitable business, as the
willingness of buyers to pay up for the comfort of certainty allows
sellers to dictate the price that provides sufficient incentive to be in this
business. The insurance business thrives on diversification and the
underwriter’s ability to create sufficiently broad portfolios that can
withstand large periodic losses. The competition among sellers is often
constrained by the availability of capital and the limited appetite to
assume such highly asymmetric risks.
The market price of an insurance product can be viewed as the sum
of its fair price and the risk premium that compensates the seller for
being in a business with a negatively skewed payoff. The fair price is
determined by the amount that, in the long-run, is expected to offset
periodic payouts to the buyers. The entire insurance business model
hinges on its ability to accurately estimate the fair price by quantifying
the risks of adverse events. The fair value of insurance is typically
determined actuarially. A simple way to price an insurance contract
actuarially would be to average historical payoffs to buyers from
writing the same policy in the past.
Let us illustrate an application of this actuarial approach to the
valuation of oil options. We consider the so-called naked straddle
strategy that sells the benchmark three-month ATM straddle once a
month with 60 days to maturity and holds the trade until its expiration
without doing any hedging. The actuarial value of the straddle is then
equal to its average historical payoff. The payoff of an ATM straddle is
given by the distance between the initial strike level, determined by the
futures price on the day of the trade, and the futures price on the day
when the option expires. The risk premium is the P&L of this strategy. It
is the difference between the option premium collected on entry and its
actuarially determined fair value.
The frequency distribution of the P&L for the strategy of selling
unhedged ATM straddles is shown in Fig. 9.1. Its summary statistics are
also presented in the first column of Table 9.1.

Fig. 9.1 Frequency distribution of P&L ($/bbl) for the strategy of selling unhedged 3-month
ATM WTI straddles (2000–2022)

Table 9.1 The summary statistics of the P&L distribution for two strategies of selling
unhedged and daily delta-hedged 3-month ATM WTI straddles (2000–2022)

Actuarial (Unhedged) Volatility (Hedged)


Premium Collected 8.36 8.36
Fair Value 8.69 7.76
Actuarial (Unhedged) Volatility (Hedged)
Mean P&L (0.33) 0.60
Return on Premium –4.0% 7.1%
Standard Deviation 9.86 2.19
Skewness (2.09) 0.08
Excess Kurtosis 4.61 1.01
Minimum P&L (47.71) (8.81)
Maximum P&L 21.43 10.54
The profile of this P&L distribution can hardly excite anyone to be in
the business of selling naked oil straddles. On average, the seller would
have collected $8.36/bbl at the trade entry. Its actuarial value, or the
average historical payout, however, is higher, at $8.69/bbl, meaning that
on average the futures moved more over the life of the option than the
premium collected by the seller. Contrary to expectations of selling
insurance being a profitable business, selling naked oil straddles would
have lost, on average, $0.33/bbl. Moreover, the P&L distribution is
highly negatively skewed, with the maximum loss being much larger
than the maximum gain. This is puzzling, as nobody should be writing
insurance policies at a loss. In fact, the buyer, rather than the seller, of
such insurance would have generated a 4% three-month return on the
option premium invested by systematically investing in three-month
ATM oil straddles and holding them to expiration.1 Why then do oil
options appear to be so cheap?
The answer is that unlike traditional insurance, oil options are not
priced actuarially by their dealers. The delta-hedging technique
presented in the previous chapter allows volatility traders to
significantly alter the shape of the resulting P&L distribution. Figure 9.2
shows the P&L distribution from selling the same ATM straddle but
instead delta hedging it. The summary statistics are presented in the
second column of Table 9.1.
Fig. 9.2 Frequency distribution of P&L ($/bbl) for the strategy of selling 3-month ATM WTI
straddles (2000–2022) and delta hedging them daily
The difference between the two strategy profiles is striking. The
losing strategy of selling naked straddles turns into a winning one. It
generates an average profit of $0.60/bbl, as the fair value of a delta-
hedged straddle drops to $7.76/bbl. On average, the option seller
would have retained approximately 7% of three-month premiums
collected, which is a reasonable compensation for the risks of writing
an insurance-like product. Moreover, the overall P&L variance, the
negative skewness, and the excess kurtosis are significantly reduced.
Despite the contractual asymmetry in the option’s payoff, the resulting
P&L profile is much closer to a symmetric normal distribution.
We have a rather unique and a nearly magical situation here. Both
the buyer and the seller of a three-month ATM straddle would have
made good amounts of money by trading with each other, provided that
they were prudent enough to manage the same trade differently. The
buyer should have just left the option alone until its expiration, while
the seller should have hedged it daily with futures. But what if a third
trader comes to the market with another hedging strategy, for example,
by hedging the same option at a different frequency (every other day,
once a week, etc.). A fourth trader can come to the market and for some
proprietary reasons choose something more esoteric, such as hedging
the option every Thursday afternoon only if it rains. When it comes to
options, different hedging strategies lead to different fair values, and
the concept of the fair value is no longer uniquely defined. The
existence of such alternative valuations is the essence of volatility
trading. What appears to be fair to one trader may not look so fair to
another, and such disagreement is what motivates them to trade.
The P&L of the naked straddle strategy is easy to understand. It is
conceptually similar to the outcome of momentum and trend-following
strategies. If the futures price moves sufficiently far away from its initial
level determined by an ATM strike, then the buyer wins, and the seller
loses. On the other hand, if futures prices remain range-bound, then the
seller is more likely to win, and the buyer is more likely to lose. For the
delta-hedged strategy, understanding P&L is trickier, as it becomes
model dependent. Delta hedging only eliminates the futures price risk
instantaneously based on the slope, or the first derivative, of the option
pricing function with respect to the futures price. This is akin to
approximating a nonlinear function with a straight line. The accuracy of
such an approximation depends on the degree of the option’s convexity.
The convexity is measured by the second derivative of the option price
with respect to the futures price, which is known as the option’s
gamma.

9.2 The Most Powerful Option Greek


As elegant as the idea of dynamic option replication is, we should
accept that it is an idealization. It is derived under two major
simplifications, which are not feasible in the real market. The theory
assumes that trading can be done continuously without any friction,
and that the option’s risk can always be instantaneously offset by the
futures position, which is determined by some perfect foresight of the
option’s delta. In the real world of financial markets, nothing can be
done continuously, which results in some slippage error and
transaction cost. An even larger error comes from the uncertainty in
the calculation of delta. The only delta that truly eliminates the option
risk is the one computed using the future local volatility, something that
one can only guess ahead of time. In other words, the practical
implementation of the strategy cannot be riskless. We now move from
the theoretical lab of perfect delta hedging to the real world of running
such a strategy.
Consider a strategy of selling a put option to an oil producer and
delta hedging it until expiration. Its market price P(F, t; v), which can
also be quoted in terms of its implied Black volatility v, is determined
by supply and demand between buyers and sellers of this option. Let us
assume that the futures market follows the diffusion process (8.​5) with
the local volatility function σ(F, t). The market, of course, does not know
σ(F, t), as it is unobservable. In fact, if there are more buyers than
sellers, then the market might be willing to pay more than the option’s
fair price, which is determined by the solution to the diffusion Eq. (8.​8)
with the boundary condition (8.​2).
Since the risk of writing a put option comes from falling futures
prices, we can reduce this risk by also selling some quantity of futures
Δ < 0, so that the total portfolio is given by

We immediately face a challenge. In the traditional delta-hedging


framework that we are planning to use, delta depends on future
realized volatility, which we, like the rest of the market, do not know.
Since we need to know delta to hedge the option and delta depends on
volatility, if we are wrong on the future volatility then we will be wrong
on delta, and therefore, our hedging strategy is no longer riskless.
Let us see what happens to P&L if the actual volatility σ(F, t)
deviates from the constant implied volatility v. The option price P(F, t;
v) is a known Black formula that was made to fit the market price of the
option by tweaking the model parameter v. The formula is a function of
the stochastic variable F. Therefore, like in the previous chapter, the
small change dP in the value of this function over a time increment dt is
determined by Ito’s lemma. The P&L of the portfolio during this time
period is then given by

(9.1)

Now we need to make the choice of what delta to use to hedge the
option. Even though we believe that the implied Black volatility v may
not be accurate, we also notice that the only way for us to eliminate the
futures risk dF in (9.1) would be to hedge the option with deltas
calculated using the same implied Black volatility. Thus, we choose our
delta as

The fact that the only delta that instantaneously removes the
futures risk dF happens to be the one that is calculated using the wrong
volatility may sound a bit odd. This subtle but important point deserves
a short diversion. If somehow, we knew in advance what the future
actual volatility σ(F, t) would be and hedged the option continuously
using the corresponding deltas, then the option payoff could be
perfectly replicated. In such a purely hypothetical case, we would not
even care what the market thinks about daily volatility and what our
P&L is on a daily basis, because the terminal P&L would be known with
certainty. The terminal P&L is the difference between the sale price P(v)
and the cost of the option’s replication given by the diffusion (8.​8) with
the local volatility σ(F, t). However, even though hedging based on
perfect foresight ensures certainty of the terminal P&L, on a daily basis
P&L will fluctuate along with dF.2 This is because the market does not
know anything about our unique and exclusive foresight of the future
volatility, and it is not pricing the option correctly. We will illustrate this
argument shortly with specific strategy examples that use different
hedging deltas. For now, we take the easiest conventional route and
hedge the option using deltas computed with the implied volatility v.
With the contribution of dF instantaneously removed by the delta
hedge, we now have that

This equation is similar to the one obtained in the derivation of the


pricing equation in the previous chapter with the exception of one
important difference. Here, P(v) represents a particular theoretical
formula that does not know anything about the real market dynamics
characterized by σ(F, t). However, what we also know is that this
theoretical formula P(v) happens to satisfy the BSM equation with
constant geometric volatility v, which can be written as follows:

We then substitute this expression into the formula for dΠ, which
leads to the following P&L of the portfolio over period dt:

(9.2)

Here, ΓBL denotes Black gamma, or the second derivative of the


Black formula for the put option with respect to the futures price,
which is given by:

We should also note that call and put options with the same strike
have the same gamma. This can be easily seen by differentiating the
put-call parity relationship (8.​12) twice with respect to the futures
price. For volatility dealers, puts and calls with the same strike carry
essentially the same risk.
We are now in a good position to see why option sellers tremble
when they hear the term gamma. If the actual price moves on a given
day happens to be smaller than the one implied by the option market,
then the seller makes some money, but the gain is limited to the daily
option’s time decay, i.e., its theta. If the market moves exactly one
standard deviation, as measured by the option’s implied volatility,
which is often referred to as the daily breakeven point, then the daily
P&L is zero. However, if futures move N standard deviations, then the
option seller’s losses are proportional to N2 − 1 multiplied by the
option’s gamma. Things could quickly become troublesome for the
option seller if a large market move occurs at a time when the options
gamma is particularly large.
The option’s gamma is very dynamic, as it varies significantly with
the moneyness of the option and time remaining to maturity. To
illustrate, Fig. 9.3 shows Black gamma for one-, three-, and twelve-
month put options struck at $50/bbl.

Fig. 9.3 Black gamma for the $50/bbl put option with v = 0.30 and different times to
expiration
Algebraically, Black gamma is simply the lognormal probability
density function with the mean at the strike price. It is highest for an
option near ATM, and the peak gamma rises as time to maturity
shrinks. In the limiting case, the gamma of an ATM option right before
its expiration approaches infinity. Imagine now the scenario of a large
futures price move that unexpectedly crosses the strike price of the
option at expiration. The losses for the option seller determined by the
difference between realized and implied volatility are magnified by a
large gamma. This is where the big money in option trading is made or
lost, and why short-term volatility trading strategies are deservingly
named after this powerful Greek letter.
In addition to suffering losses from a large price move, an initially
delta-hedged option portfolio quickly loses its directional immunity,
and the residual delta must be promptly rebalanced. Delta hedging can
only keep the portfolio neutral with respect to futures instantaneously.
As soon as futures move, the desired risk neutrality disappears, as delta
itself changes according to its gamma profile. Gamma is the second
derivative of the option’s price, or alternatively, it is the first derivative
of the option’s delta. If the futures price falls and moves towards the
strike of a short OTM put option, then the put delta becomes more
negative, making the portfolio directionally longer futures. Additional
futures must be sold to bring net delta back to zero. Likewise, if the
futures price rises and moves away from the strike, then the put delta
becomes less negative, making the portfolio overall short. Some short
futures hedges which are no longer needed must then be bought to
bring net delta back to zero.
The short gamma trader must always sell futures on the way down
and buy them back on the way up to keep the portfolio delta neutral.
The quantity of futures needed for rebalancing is determined by the
combined impact of the option’s gamma and the magnitude of the
futures price move. Figure 9.4 shows the number of futures contracts
needed for rebalancing the delta of 100 contracts of short put option
struck at K = 50 with different times to maturity. The options have
gamma profiles, as in Fig. 9.3, but taken with the negative sign to
represent the short position. The initial futures price is F = 60 when the
hedged portfolios are delta neutral.
Fig. 9.4 Incremental futures required for rebalancing 100 short put options with K = 50,
v = 0.30 and different times to expiration. Initially, F = 60 and the portfolio is instantaneously
delta neutral
The gamma here is observed as the change in the slope of the delta
function. If the market moves down, then gamma increases as futures
move towards the strike. In this case, dealers have no choice but to sell
more futures to remain delta neutral. Larger gamma leads to more
aggressive selling of futures by volatility dealers as the futures price
drops. On the other hand, if futures move up, drifting further away from
the short strike, then the need for rebalancing becomes more muted.
Such an asymmetric gamma profile is very typical for the oil options
market, which is dominated by producer hedging demand for downside
protection. On the upside, the overall industry gamma profile is more
balanced, as some producers gain leverage by selling options to the
market, a topic that we discuss in more detail in the next chapter.
The rebalancing of short gamma portfolios is one of the most
powerful forces that drives the market for the underlying futures. This
process is completely mechanical. It has little to do with volatility
dealers’ own opinion about the direction of oil prices. When futures fall,
dealers who are short puts to producers must sell futures to remain
within tight volumetric risk limits on residual deltas, prudently
imposed by their risk managers. The larger the gamma, the more
futures they need to sell, pushing the price further down. This creates a
vicious cycle when prices fall in a downward spiral until strikes are
crossed and dealers’ short gamma exposure starts subsiding. The
gamma hedging profile can also be viewed as an alternative to the
reaction function previously introduced for the momentum strategy.
The insurance-like business of selling delta-hedged options, which
mandates selling futures on the lows and buying them back on the
highs with potentially unlimited losses, may appear to be a
questionable-value proposition for investors. This perception is driven
by cognitive biases that cause painful financial losses to be
overweighted relative to their frequency and magnitude. In the
meantime, every day when nothing major happens in the market, and
the futures price moves less than its daily breakeven, the option seller
retains a small portion of the option premium. The positive P&L on
uneventful days could easily add up to a substantial buffer which more
than offsets periodic large losses.
Is this strategy just an example of picking up pennies in front of the
steamroller, or can it be turned into a viable investment with its risk
kept under control? The answer is not so straightforward. It turns out
that one needs to be more selective on which options to sell, and which
ones are better to stay away from. We address this important topic in
the following sections with a comprehensive empirical study of the
volatility risk premium (VRP) strategy in the oil market.3

9.3 The Smile of the Volatility Risk Premium


To test the existence of the structural risk premium embedded in prices
of oil options, we look at the long-term performance of the VRP
strategy, starting from its earliest days when oil options gained
sufficient liquidity. In the base version of the VRP strategy, one sells a
particular option and hedges it daily using deltas computed with
conventional implied Black volatilities. The VRP profitability is
analyzed for options with different moneyness and times to maturity.
Then VRP trades are combined into a portfolio that can be analyzed as a
continuously managed systematic risk premium strategy. In the
following section, we present VRP implementations that are based on
alternative delta-hedging techniques.
When analyzing results of option-based strategies one must start
with a careful choice of appropriate performance metrics, which could
differ from conventional metrics used for studying stocks, bonds, and
futures. Consider first the investment return on an option trade. If an
investor buys an option, then the return is defined in the traditional
way, where the denominator of the return reflects the initial
investment, i.e., the option premium. This is the maximum amount that
the buyer can lose.
For selling options, the maximum loss is technically unlimited, and
additional capital must be set aside by the seller with a clearing broker
to ensure that all contractual obligations will always be met. However,
incorporating such highly variable and trader-specific funding
requirements would significantly distract us from our primary
objectives. To keep the presentation transparent and intuitive, we keep
the simple definition of the return, but address the risks of an option-
selling strategy that drive additional capital requirements separately.
We define the return on selling an option as the negative of the return
on buying the option. This also means that the return on the VRP trade
is the percentage of the premium retained after hedging.
Since selling options is akin to selling insurance, and option
moneyness corresponds to an insurance deductible, it is important to
analyze VRP performance for options with different moneyness. Using
moneyness as a function of an option’s delta, as defined in the previous
chapter, the performance of VRP strategies is measured for 0.10-, 0.20-,
0.30-, and 0.40-delta OTM puts and calls and for zero-delta straddle.
Option tenors are defined by the number of trading days remaining to
maturity (DTM) counting from the day when the option is sold. We use
increments of 20 DTM that approximately correspond to a monthly
option expiration schedule with roughly 20 trading days per month.
Figure 9.5 presents historical VRP returns, measured by the
percentage of the option premium retained with daily delta hedging
across different delta-moneyness for various investment horizons. In
the spirit of an implied volatility smile that describes option prices for
different moneyness, we label Fig. 9.5 the VRP smile. The VRP smile
characterizes the long-term historical profitability of selling delta-
hedged options for different moneyness, analogous to returns on
selling insurance with different deductibles.
Fig. 9.5 VRP smile, as the premium retained ratio versus moneyness with daily delta hedging
(WTI, 2000–2022)
For short-term one-month options, the graph does indeed look like
a smile. It confirms that a catastrophe-type oil insurance with the
highest deductible that only pays off in the case of an extreme price
move is the most overpriced one, at least, if the profit is measured as a
fraction of the option premium collected. In other words, the frequency
of large futures moves implied by the options market is much higher
than the frequency of their actual occurrences. Either market
participants overestimate the likelihood and the magnitude of such
events, or, more likely, they have no choice but to compensate writers of
deep OTM options for the higher risk taken per unit of premium
collected. The short-term VRP smile appears to be relatively symmetric,
confirming the higher risk premiums embedded in market prices for
both OTM calls and OTM puts.
For medium-term options, the VRP smile exhibits a higher degree of
asymmetry. The smile flattens for OTM puts but it remains steep for
OTM calls. The call risk premium is often driven by geopolitical
uncertainty and risks of supply disruptions. However, this risk is not
realized to the extent that it is priced in the options market. In contrast,
the risk premium embedded in put options, which is typically driven
more by macroeconomic weakness and occasional spillover from falling
equity markets, turns out to be more modest.
For longer-term options, VRP profitability erodes for both puts and
calls. For the most part, this is explained by the increasing denominator
in the definition of the return, as option values rise with time to
maturity. However, the profitability also declines even if the
performance is measured in absolute terms, i.e., in dollars per barrel. In
other words, it does not pay to sell options too early when the options’
time decay is too small to adequately compensate for the prolonged
risk exposure. Note also that selling deep OTM medium and long-term
put options results in trading losses, or, alternatively, buying such
options becomes profitable. This is another manifestation of the
negatively skewed realized price distribution. However, the conclusions
for longer-dated deep OTM options should be interpreted with a
greater degree of caution given the lower liquidity and the noisier
historical option data.
It is useful to contrast the shape of the VRP smile to the typical
shape of the implied volatility smile. The latter, as seen in Fig. 8.​7,
exhibits a distinctively negative put skew. At the first glance, such skew
may create an illusion that oil puts are overpriced, like they are, for
example, in the equity market. However, a positively skewed VRP smile
shows that oil puts are surprisingly much more reasonably priced than
oil calls if they are measured relative to their corresponding fair values.
Despite the fact that implied volatilities for most calls trade at discounts
to ATM volatility, call options are, nevertheless, still expensive. For calls
to be priced fairly, the implied call skew should be even more negative.
This is because the presence of an implied volatility skew is a side-
effect of looking at option prices through the lens of the lognormal
framework, which, as illustrated in the previous chapter, fails to capture
important properties of the oil price distribution. A more accurate
pricing model that corrects for such distortion will be introduced in the
next chapter.
Figure 9.6 presents results of the historical VRP performance versus
time remaining to maturity, which we define as the VRP term structure.
For both OTM calls and OTM puts, the term structure of VRP, measured
again by the premium retained ratio, resembles the term structure of
implied volatilities. Both curves generally decline with increasing DTM.
However, for zero-delta straddles the VRP term structure exhibits a
noticeable hump. This indicates that selling medium-term ATM options
is more profitable than selling either short-term ATM options that have
higher gamma or long-term options that maintain risk exposure for too
long. We will see shortly that this hump is magnified when the analysis
is extended to the portfolio of short oil options.

Fig. 9.6 VRP term structure, as the premium retained ratio versus maturity with daily hedging
for 0.20-delta puts (20P), zero-delta straddles (0S), and 0.20-delta calls (20C) (WTI, 2000–
2022)
While the historical returns of VRP strategies appear to be
attractive, returns only represent one side of the investment analysis.
The other side, which is arguably even more important for short option
strategies, is the amount of risk that must be taken to achieve such
returns. Here, we need to make an important distinction between the
risk characteristics of individual trades, which we define as positions,
and the risks of the portfolio made up of such positions. Since standard
oil options are only available with monthly maturities, this adds some
complexity to the analysis of risk-adjusted returns because of the
overlapping nature of multiple positions within each portfolio that
remain open at the same time. While returns generated by portfolios
with different maturities can be easily annualized and compared to
each other, the annualization of risks resulting from individual trades
with different maturities is prone to statistical anomalies, especially
when trades are highly correlated.4
To properly measure the relative performance of VRP strategies
across different time horizons, we construct continuously run
systematic strategies made up of corresponding VRP positions with the
same moneyness and the same DTM at the time of the option sale. Each
one-month portfolio (20 DTM) consists of only one position, as the new
trade is initiated within a few days of the expiration of the previous one.
However, it is not the same for portfolios with longer DTM. The two-
month (40 DTM) portfolio has two open trades with expirations
approximately twenty days apart, the three-month (60 DTM) portfolio
has three open trades, etc., and the twelve-month (240 DTM) portfolio
consists of twelve open trades. Only one option in a portfolio expires
each month. Therefore, all portfolios are exposed to the same futures
risk at the expiration of this option. However, since a new option is
added to the portfolio every month at the moneyness level prevalent at
the time, the strikes of options within the portfolio are spread out,
providing some valuable risk diversification benefits to the portfolio of
options.
For a proper apples-to-apples comparison of VRP investments
across time horizons, we construct continuous equity lines by
cumulating daily P&L from all open positions within the portfolio. We
can then calculate traditional investment characteristics for each
portfolio, such as the Sharpe ratio.5 Figures 9.7 and 9.8 display Sharpe
ratios for various VRP portfolios by moneyness and maturity.
Fig. 9.7 Sharpe ratios of VRP portfolios by moneyness (WTI, 2000–2022)

Fig. 9.8 Sharpe ratios of VRP portfolios by maturity for 0.20-delta puts (20P), zero-delta
straddles (0S), and 0.20-delta calls (20C) (WTI, 2000–2022)
The main difference between the portfolio VRP curves, shown in
Figs. 9.7 and 9.8, and position VRP curves, shown in Figs. 9.5 and 9.6, is
the materially improved performance of zero-delta straddles. ATM
options, being highly exposed to large gamma risks, benefit the most
from the strike diversification. The benefits of diversification for OTM
options are more muted. The reason again lies in the crucial role of the
option’s gamma, which becomes infinite if the futures cross the strike
price at the option’s expiration. Since strikes near ATM are more likely
to be crossed than strikes further OTM, the benefits from diversification
of the strike risk are more substantial for ATM options.
Similarly to other systematic risk premia strategies, oil VRP clearly
demonstrates the existence of an investment edge, but its actual
implementation is highly customized by professional volatility traders.
In the next section, we provide some examples of such practical
implementations. These implementations explicitly incorporate the
impact of transaction costs and optimize VRP by combining it with
short-term biases in the behavior of the underlying futures.

9.4 The Art and Science of Delta Hedging


Traders like to say that they do not need a model to price an option, as
the price is given to them by a broker, but they do need a model to
hedge an option. The hedging model is what determines the cost of the
dynamic replication, or the option’s value. The option fair value is not
uniquely determined, as different hedging models imply different
values. In this section, we test the robustness of VRP strategies with
respect to important parameters of the hedging model. As highlighted
earlier, the theory of dynamic option replication is developed for an
idealized world, where trading can be done continuously without any
friction, and delta can be calculated using the perfect foresight of the
future realized volatility. The real world is far from perfect, and the
trader constantly faces many practical challenges. The two most
commonly asked questions by volatility traders are how frequently to
hedge and what volatility to use to calculate the options’ delta.
For any other systematic strategy that trades relatively frequently,
the question of bid-ask and transaction costs comes to the fore. In many
academic studies, this important topic is often either skipped entirely
or sidestepped with simplistic assumptions, the impact of which is
buried within the strategy performance metrics. For traders, however,
the devil is in details. They only care about the existence of a theoretical
risk premium if it can be turned into real profits, net of all frictions
from market imperfections. A big question is whether transaction costs
associated with futures hedging chip away so much of the trading edge
that they could make the entire VRP strategy less appealing. We
address this important topic explicitly and illustrate how delta hedging
can be optimized to make the impact of execution costs more palatable.
One important instrument in the oil market that helps to reduce the
cost of hedging is the previously described Trading at Settlement (TAS)
contract. This contract allows traders to execute futures at a price to be
determined later during the daily settlement window. It is very liquid
for the nearest maturity futures, and for the most part, the trader can
execute any given number of futures effectively with zero bid-ask. The
TAS contract has proven to be quite useful for option traders, who often
choose to rebalance their portfolio deltas at settlement prices. One
caveat, however, is some uncertainty in the futures quantity required
for rebalancing, as the settlement delta is not known until the
settlement price is known. To tackle this issue, the option trader usually
adjusts the futures quantity for the TAS order throughout the day based
on updated estimates of the end-of-the-day portfolio delta. Any residual
exposure is then promptly cleaned up once the futures settlement is
published. This technique usually allows the trader to keep the average
delta-hedging slippage to a minimum, typically to less than $0.01/bbl.
Delta hedging of the short gamma strategy requires a trader to buy
futures after the price rises and sell futures after the price drops. Many
traders, reluctant to buy high and sell low, resist full delta rebalancing
and utilize instead strategies reduced-hedging strategies. One popular
choice is to hedge less frequently.6 This approach not only saves on
transaction costs, but also attempts to capitalize on short-term price
reversals that decrease the need to hedge. Obviously, reduced-hedging
strategies come with significantly higher risks if the price trend
continues. The question is whether such risks are worth taking.
Consider the hedging strategy where daily portfolio delta
rebalancing is replaced with re-hedging it only every N days. To keep
the presentation more concise, we only look at a representative case of
an ATM option with 60 DTM, but the conclusions for options with other
moneyness and maturities are broadly similar. Figure 9.9 shows the
performance of VRP strategies hedged every N = 1, 3, 5, and 10 days,
where N = 1 corresponds to the base case of daily hedging. Transaction
costs are assumed to be $0.01/bbl, which can be easily scaled to reflect
an individual trader’s assessment of the execution slippage.

Fig. 9.9 The performance of VRP strategies for 60 DTM ATM straddle for different hedging
frequencies (WTI, 2000–2022)
It is not surprising that with less frequent hedging, the risks
measured by the annualized standard deviation of the portfolio’s P&L
steadily increase. More interesting is the observation that annualized
profits also increase slightly if the portfolio does not rebalance the delta
for at least two days. This confirms that some additional gains can
indeed be captured from short-term price reversals by leaving the
portfolio unhedged for a few days. If instead of hedging every day, the
trader hedges only every two or three days, then higher profitability
adequately compensates for taking larger risks while simultaneously
saving on transaction costs. However, leaving the strategy unhedged for
more than a week makes it less attractive on a risk-adjusted basis.
While such a strategy allows traders to benefit from short-term price
reversals, it can suffer significant losses if the portfolio is left unhedged
for too long.
The second important driver of VRP profitability is the choice of the
hedging delta. Up until now, VRP strategies have been hedged with
deltas calculated using implied volatilities. This choice was motivated
by the desire to eliminate the stochastic term in (9.1). However, this
choice is somewhat inconsistent with the entire investment concept
behind VRP. We have already highlighted that the option’s payoff can be
perfectly replicated by trading futures only if the hedging delta is
calculated using the future actual volatility, which, of course, is not
known in advance. If the whole reason to trade VRP is driven by the
view that options are overpriced, which means that implied volatility is
too high relative to future realized volatility, then why would we be
hedging using a volatility that we ourselves do not even believe to be
correct?
To illustrate VRP sensitivity to the choice of the hedging delta, we
scale the hedging volatility by multiples of 0.5, 0.75, 1.25, and 1.5 of the
prevalent implied volatility. The multiple of 1.0 represents the base
case. The results of this experiment are summarized in Fig. 9.10.

Fig. 9.10 The performance of VRP strategy for 60 DTM ATM straddle versus scaled implied
volatility used for calculating hedging deltas (WTI, 2000–2022)
The most interesting takeaway from this analysis is P&L
improvement from hedging with low volatility and P&L decline for
hedging with high volatility. To interpret this result, recall from Fig. 9.1
and Table 9.1 how poorly the strategy of selling the naked straddle
performs. Selling unhedged straddles suffers from large losses when
futures periodically deviate too far from the initial ATM strike. This is
consistent with the overall trendiness of oil futures already established
in directional risk premia strategies. In the VRP strategy, when the
option crosses the strike and moves ITM, the sooner the seller hedges
changing delta, the better the protection against losses if the trend
persists. Using the lower volatility to calculate deltas does indeed
induce such a faster reaction. Lower input for the implied volatility
reduces the remaining variance in the price of the option and brings the
option payoff function closer to its terminal hockey stick-like payoff.
The deltas computed in this manner, therefore, increase for calls and
become more negative for puts when options are ITM. This forces the
dealer to hedge more aggressively in the direction of the trend. In
contrast to the reduced-hedging example, however, such aggressive
hedging increases transaction costs.
Even though one might easily jump to the conclusion that hedging
with a lower volatility is a better way to go, this approach has its own
challenges. If we hedge conventionally, using what we perceive to be an
incorrect implied volatility, then the second term in the Eq. (9.1) drops
out and the mark-to-market of the P&L becomes less volatile. In
contrast, if we hedge with any other volatility, even with a more
accurate one, then daily P&L will fluctuate along with dF, and the
standard deviation of P&L increases. This can also be seen from Fig.
9.10. Should we hedge then with a volatility that we believe to be
incorrect and benefit from relatively stable daily P&L, or is it better to
stick with our own opinion about future volatility and accept larger
daily swings?
While we may believe that we are right and the market is wrong,
our risk managers and controllers may have a different opinion. From
their independent perspective, the market is fair. Therefore, whatever
delta is implied by the market volatility should be the right one, and if
our creative delta-hedging strategy results in additional P&L from
futures, then it should be treated as a speculative position. Obviously, if
risk managers view an option price to be fair, then they should not be
letting the volatility trader sell an option to begin with. Some debate
with managers becomes inevitable, and to avoid it, many gamma
traders opt for a less controversial route and use conventional Black
deltas based on exchange-published futures settlement prices for the
base hedging case.
Despite the evidence of systematic biases highlighted by VRP
hedging alternatives presented in this section, actual optimization
decisions are usually made more tactically only when the volatility
trader has particularly strong views about the short-term expected
behavior of the market. To a certain degree, the problem of delta
hedging also has elements of quantamental trading, where a trading
algorithm is combined with human intervention. It would be a
disservice to attempt to provide readers with any more precise
systematic guidance on how frequently to rebalance a short VRP
portfolio, and what volatility to hedge it with, but these case studies
may steer hedgers towards better independent decisions.
The primary subjective decision in trading VRP, like in any other
systematic oil strategy, is the identification of regimes in which the
strategy is more likely to perform. For the oil VRP, such regimes are
mostly driven by the demand for hedging and the behavior of large
market participants.

9.5 The Behavior of Hedgers and Regime


Changes
Additional insights can be gained from looking at the performance of
VRP strategies over time. Many systematic risk premia strategies in
energy markets are known to be sensitive to regimes. The world of
energy constantly evolves, adjusting to new fundamental drivers, such
as the growth of shale oil, or flow factors resulting from market
financialization. These factors lead to changing behavior among
hedgers and speculators. Such behavioral changes impact the supply
and demand for risk management services and resulting risk premia.
The VRP strategy is not an exception. Figure 9.11 presents the
cumulative P&L of one-month VRP strategies for 0.25-delta OTM puts,
0.25-delta OTM calls, and zero-delta straddles. It clearly shows the
presence of the structural break that separates two distinct regimes.

Fig. 9.11 20 DTM VRP equity history for 0.25-delta puts (25P), zero-delta straddles (0S), and
0.25-delta calls (25C) (WTI, 2000–2022)
The idea of selling oil options as an overpriced insurance has
proven to work remarkably well for a while. Between 2000 and 2014,
VRP strategies generated impressive risk-adjusted returns with Sharpe
ratios exceeding 1.0 for many moneyness and maturity configurations.
After that, by and large, the risk premium contained in oil options
followed the path of Keynesian normal backwardation. It gradually
disappeared, like the structural futures risk premium did
approximately a decade earlier. The opportunity to make easy money in
any competitive market rarely lasts long. At some point, the strategy’s
consistent profitability motivates enough participants to invest in the
capability needed to manage the strategy, which helps to bring the
market towards an equilibrium.
The VRP strategy became a victim of its own success. As the
business of passive commodity investments struggled under the
pressure of contango and punitive rolling costs, capital shifted towards
more dynamic strategies designed to capture alternative risk premia. To
make it easier for investors, the VRP concept was packaged into
investable indices, and the cumbersome task of delta hedging was
effectively outsourced to index providers. Such investable volatility
indices provided large pools of capital held by institutional investors
with access to what used to be an obscure investment opportunity
previously dominated by oil quants. The barriers to entry were lifted,
and once again risk-bearing investors were compensated for providing
capital rather than for having any particularly unique trading skills. As
a consequence, the reward for offering relatively simple services of risk
absorption cratered. Another factor that contributed to the structural
break in VRP is a significant change in hedging strategies run by US
shale producers, a topic that we will discuss in more detail in the next
chapter.
The profitability of the VRP strategy will likely increase again, when
more buyers come to the market to buy oil insurance. This usually
happens shortly after the crisis that often forces insurance writers out
of business, resulting in higher premiums. At the time of writing this
book, there are indeed some indications that the Covid-19 pandemic
could well be another turning point, as the VRP performance in the
following two years improved substantially. This is driven by increasing
demand for hedging and the exodus of many option sellers after the
period of extreme volatility.
Even though the simple VRP strategy became less attractive to
financial investors, the aggregate risk premium in oil options has not
vanished. Instead, it spread out across various strikes and maturities,
but not in a uniform manner. Some options became consistently more
expensive than others. Like futures traders, many professional volatility
traders have also switched from outright directional bets to relative
value trades. In relative VRP strategies, one aims to sell options with
higher risk premia and hedge them by buying options with lower risk
premia.
A professional oil volatility portfolio is typically made up of many
relative value strategies based on the primary VRP building blocks
presented in this chapter. For example, the presence of the VRP smile
suggests that OTM calls, while being cheaper in dollar terms than ATM
calls, appear to contain a larger risk premium. One can then construct a
relative value strategy that takes advantage of the VRP call skew by
selling a larger quantity of OTM calls versus buying a smaller quantity
of ATM calls. Such a trade, known as a ratio call spread, is very common
in the marketplace. The weights between two legs are often chosen for
the trade to be either volatility-neutral or premium-neutral. However,
to successfully trade options across different moneyness, one needs to
better understand important quantitative linkages that connect prices,
which we discuss in the next chapter.
A relative value option portfolio typically includes options with
multiple expirations. For example, the decreasing term structure of VRP
may suggest hedging the sale of a short-term ATM option by buying a
longer-term ATM option that is less overpriced or even underpriced. If
the two options are traded in equal volumes, then the resulting
calendar spread trade is net short gamma, since gamma of the short leg
exceeds gamma of the longer-dated leg. However, the trade is net long
vega, since vega of the longer-dated leg exceeds vega of the short leg.
Buying deferred options can also be used as a valuable risk mitigant to
the base VRP strategy, as it provides significant mark-to-market offsets
when futures experience large moves. The weights between the two
legs can be adjusted to make the trade either gamma or vega neutral.
The proper construction of such strategies requires a more elaborate
modeling of the local volatility term structure, which we focus on in
Chap. 11.
The number of permutations that can be constructed from
individual options and distinct hedging strategies is nearly infinite, but
they are often based on the main features of VRP described in this
chapter. In addition, professional volatility traders rarely even hold
options until expiration. They might enter the trade when the
opportunity caused by end-user flows is particularly attractive and then
attempt to exit the trade once imbalances normalize. In the meantime,
they would delta hedge based on a particular model to preserve the
value of the option until the opportunity comes to exit the trade. This
brings an important additional dimension to volatility trading since the
trader is no longer betting on implied volatility versus gamma-scaled
realized volatility during the entire lifespan of the option. The trader
must also manage P&L fluctuations that come from changes in implied
volatility itself, which is often referred to as vega trading, the topic we
discuss next.
References
Bouchouev, I., & Johnson, B. (2022). The volatility risk premium in the oil market. Quantitative
Finance, 22(8), 1561–1578.
[MathSciNet][Crossref][zbMATH]

Derman, E., & Miller, M. B. (2016). The volatility smile. Wiley.


[Crossref]

Doran, J. S., & Ronn, E. I. (2006). The bias in Black-Scholes/Black implied volatility: An analysis
of equity and energy markets. Review of Derivatives Research, 8(3), 177–198.
[Crossref][zbMATH]

Doran, J. S., & Ronn, E. I. (2008). Computing the market price of volatility risk in the energy
commodity markets. Journal of Banking and Finance, 32(12), 2541–2552.
[Crossref]

Ellwanger, R. (2017). On the tail risk premium in the oil market, Bank of Canada Working Paper,
46.

Jacobs, K., & Li, B. (2023). Option returns, risk premiums, and demand pressure in energy
markets. Journal of Banking and Finance, 146, 1–26.
[Crossref]

Kang, S. B., & Pan, X. (2015). Commodity variance risk premia and expected futures returns:
Evidence from the crude oil market, SSRN.

Lo, A. W. (2002). The statistics of Sharpe ratios. Financial Analysts Journal, 58(4), 36–52.
[Crossref]

Prokopczuk, M., Symeonidis, L., & Simen, C. W. (2017). Variance risk in commodity markets.
Journal of Banking and Finance, 81, 136–149.
[Crossref]

Trolle, A. B., & Schwartz, E. S. (2010, Spring). Variance risk premia in energy commodities. The
Journal of Derivatives, 17(3), 15–32.
[Crossref]

Footnotes
1 Note that this analysis is presented on the trade level. Since the trades are entered once a
month and held for three months, multiple trades are open at the same time. The analysis of the
continuously managed strategy of selling options is presented later in this chapter.

2 This point is well covered in Derman and Miller (2016).


3 The existence of the volatility or variance risk premium in oil options has been documented
in several studies, see Doran and Ronn (2006, 2008), Trolle and Schwartz (2010), Kang and Pan
(2015), Prokopczuk et al. (2017), Ellwanger (2017), and Jacobs and Li (2023). The analysis of
VRP as the actual trading strategy, which we present here, is a much more difficult task due to
its extreme path-dependency and non-straightforward measurements of risks. The remainder
of this chapter is based on empirical results of Bouchouev and Johnson (2022). The author
would like to thank Brett Johnson for his highly valuable contribution to the data analysis.

4 The challenges that arise with annualization of risk-adjusted returns are well explained by Lo
(2002).

5 As before, we calculate the Sharpe ratio assuming zero risk-free interest rate. Since we do not
include the cost of capital and exchange margin in our calculations, we do not include any
interest that could be received on the initial margin either. Overall, the impact of financing costs
on the performance of VRP strategies is relatively minor. In addition, both the numerator and
the denominator of the Sharpe ratio here are computed in dollars per barrel, not in percent.

6 Another popular reduced-hedging strategy is to trade fewer futures than required by the
model. The details of this strategy can be found in Bouchouev and Johnson (2022).
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
I. Bouchouev, Virtual Barrels, Springer Texts in Business and Economics
https://doi.org/10.1007/978-3-031-36151-7_10

10. Volatility Smile Trading


Ilia Bouchouev1
(1) Pentathlon Investments, LLC, Westfield, NJ, USA

– Oil volatility traders like to quip that if you do not know who the fool
in the market is, then it is probably you. This does not literally mean
that someone is acting foolishly. Rather, it highlights the importance
of understanding the behavior of large market participants.
– Vega-trading strategies require modeling the evolution of the entire
volatility smile. Common smile heuristics, such as sticky moneyness
and sticky strike, perform poorly in the oil market. Instead, a more
realistic dynamics for the underlying futures contract must be
imposed.
– Empirical study of the relationship between futures prices and
implied volatilities guides model selection. The normal model
removes an artificial skewness embedded in the conventional
lognormal framework, but neither of the two models is capable of
capturing extreme events.
– A novel quadratic normal model is developed based on the
perturbation method for the general diffusion equation. The option
pricing formula incorporates skewness and fat tails and relates
model parameters to market prices for benchmark collars and
strangles.

10.1 Producer Hedging and Volatility Market-


Making
The strategy of harvesting an insurance-like risk premium by
systematically selling oil options turned out to be rather fragile. The
volatility risk premium that was clearly visible for nearly two decades
became scantier over the years, as the oil market transitioned to a
fundamentally different regime. The overall balance between buyers
and sellers of oil volatility has been largely restored. The supply of
volatility was increased through crafty financial products that allowed
investors to provide risk-bearing capital while outsourcing the
complexity of delta hedging to professional dealers. In addition, many
corporate hedgers, reluctant to pay up for overpriced options, switched
to more advanced risk management programs, many of which are
volatility neutral. The adoption of diverse and sophisticated hedging
strategies by end-users, however, opened new opportunities for
volatility market-makers in relative vega trading.
In the volatility market, traders say that if you do not know who the
fool in the market is, then it is probably you. This does not literally
mean that the counterparty of your trade is acting foolishly. Rather, it
highlights the importance of understanding the motivation of market
participants whose behavior may temporarily distort prices of certain
options and present attractive trading opportunities for the dealer.
Such understanding is particularly crucial in trading oil volatility.
The oil market is full of real options, many of which we have already
discussed. The volatility embedded in real options motivates asset
owners to utilize financial options in their hedging programs. The
largest hedger in the options market is the producer. Producers are
unlikely to be fools in making directional bets on oil prices. They are
arguably even more knowledgeable about market fundamentals than
dealers themselves. However, when it comes to volatility, end-users are
generally prepared to pay up for the service of tailoring risk
management products to their specific needs.
The risk management tools used by oil producers vary with the
ownership structure of the company, its overall business model, and the
level of indebtedness. The governments of oil-producing counties and
state-owned entities rarely get involved in the derivatives market. One
notable exception is a large-scale hedging program administered by the
Government of Mexico, which we will cover in more depth in the next
chapter. For the most part, other large sovereign producers manage
price risks with policy tools, such as taxation, subsidies, and rainy-day
funds. Large vertically integrated oil majors also rarely hedge in the
derivatives market. They tend to benefit more from related energy
businesses, such as refining, and often prefer to retain their exposure to
oil prices, which is the primary reason why some investors buy the
company’s stock. The most active participants in the oil hedging market
are independent oil producers, particularly the ones specializing in
shale oil.
The economics of shale drilling differs from conventional oil
production in several important ways. It is generally more expensive to
extract oil from shale than to produce it via traditional means in oil-rich
parts of the world. However, shale has a much shorter production cycle,
which makes its operation more like a mining or manufacturing
business. Any oil production represents a real call option on the price of
oil with the strike price determined by the cost of production. The real
option held by a shale producer can be thought of as being closer to
ATM because of its higher production cost, and ATM options have the
largest exposure to volatility.
Shale’s shorter production cycle is valuable for capturing short-
term volatility and monetizing the real option. It allows producers to be
nimbler in the market by quickly adjusting production levels in
response to market movements, while simultaneously locking in
favorable prices in the derivatives market when futures rise. If futures
subsequently fall, then producers can take profits on existing financial
hedges and reduce production. In contrast, real options held by low-
cost producers are effectively in-the-money. While these options have
larger intrinsic value, the producer’s decision making is less affected by
volatility. It is nearly always rational for any individual low-cost
producer to continue pumping oil, regardless of the price. The only
exception is the situation when large producers collectively agree to
reduce output with the specific goal to impact the price.
Unlike national oil companies and oil majors, independent shale
producers are notorious for their high leverage. Since shale production
has faster decline rates, producers must constantly drill just to
maintain the same output level. This strategy requires a persistent
inflow of capital, which is typically borrowed from the banks. The
lenders, unwilling to take the risk of a producer failing in the event of
lower oil prices, typically insist on hedging the price risk to support the
loan. This is not too different from the mandate to purchase house
insurance when the property is mortgaged to a bank. Shale lenders may
also hold a lien on the producer’s main property, oil reserves, as
collateral for the loan. However, bankers are not keen on ever taking
ownership of the physical oil. They would rather require producers to
protect themselves in the derivatives market against the adverse price
scenario, which can be enforced via lending covenants. Besides, the
same lending bank often acts as the counterparty on the derivatives
hedge as well, thus, making money on both transactions.
While the need for shale producers to hedge is apparent, their
financial resources to do so are rather limited. Producers can rarely
afford to buy traditional insurance and shell out upfront cash for put
options. With the strong competition for market share in the fast-
growing shale business, cash is usually preserved and used to maintain
and scale up production. Besides, producers often perceive an initial
investment in the oil property as a form of option premium that has
already been paid. Paying yet another premium to acquire a financial
option to hedge is a rather tough pill for many of them to swallow. What
the producer owns is a real call option; what is needed for a bank loan
is a financial put option. Thus, a more efficient solution for the producer
would be to swap one option for another and avoid additional cash
outflows. This is where the derivatives market with its flexibility comes
in handy.
The most popular producer hedging structure in the oil market is a
costless collar, sometimes also referred to as a fence. In this trade, a
producer buys a put option with a strike price that ensures some
minimum investment return on borrowed capital. The producer does
not pay any cash for the put option and instead gives the dealer a call
option of the same value. The call strike is iterated until the value of the
call matches the value of the put. In practice, the traded call strike is
shifted down to make the call slightly more valuable than the put to
incentivize the dealer to participate in the trade. The dealer’s
compensation comes not from retaining a portion of the option
premium, which is zero, but rather from receiving at zero cost the
collar, which has several cents of a positive value, at least according to
the dealer’s pricing model.
Since the iterated call strike in the costless collar trade is
customized specifically for a given deal, it is unlikely that exactly the
same collar will be quoted in the open market. In the jargon of option
traders, an end-user option trade is rarely a back-to-back transaction,
meaning that it cannot be hedged perfectly. To turn a theoretical profit
margin into real dollars, the dealer must reduce risks by trading more
liquid options and managing the residual exposure dynamically. With
many option trades added to the volatility market-making portfolio
throughout the day, the residual exposure is always managed on an
aggregate level. This can only be done if the pricing model used by the
dealer can handle options across all strikes in the portfolio consistently
with the market.
The costless collar is approximately volatility neutral. The vega from
the call largely offsets the opposite vega from the put, provided that
both options are struck with similar moneyness. Collar transactions
are, however, primary drivers of the skewness and the overall dynamics
of the volatility smile. If the end-user hedging program is executed
ratably, then the put strike is often kept the same, as it is often chosen
to meet lending requirements. However, call strikes are spread out
based on whatever strike makes the collar transaction costless for a
given futures price. This results in a rather convoluted dynamics for the
volatility smile. On the put side, the peak hedging demand is often tied
to the specific strike, while the supply of calls sold by a producer to
finance the put is more driven by the moneyness of the option. This
brings additional challenges to modeling the dynamics of the oil smile
using conventional techniques often employed in other financial
markets.
The costless collar is only one example that illustrates the leverage
that a producer can gain by selling optionality in the derivative market.
Some producers take the idea of monetizing their real option further by
selling a financial call option to the lender and collecting an option
premium without buying any downside price protection. Such deals can
hardly be classified as proper hedges. In fact, they are much closer to
hidden loans. The outright selling of options by producers is generally
frowned upon by lenders due to their large credit exposure. However,
as we will see in the next chapter, short calls are often embedded into
more complex exotic derivatives. These over-the-counter derivatives
are less transparent and sufficiently profitable for dealers to justify
taking on additional credit risks.
Another popular leveraged hedge in the oil market is a three-way
collar. In this trade, instead of buying a put option financed by selling a
call, the producer purchases a put spread. The put spread is essentially
a put option with capped payoff, where the purchased put is combined
with the sale of another put that has a lower strike. The put spread is
again financed by selling an OTM call that makes the entire structure of
three options costless. Since the put spread is generally cheaper than an
outright put option, less premium needs to be raised from the sale of
the call to make the net transaction cost zero. This lets the producer
push the call strike further OTM and retain larger potential upside from
higher prices. Alternatively, the producer can sell the call with the same
strike as in the regular collar, collect a larger premium and use it to buy
a more valuable put spread with a higher upper strike. The strikes are
again selected iteratively to make the three-way collar theoretically
costless, and then shifted slightly in favor of the dealer, as an incentive
to trade.
Terminal payoffs of a two-way collar and a three-way collar are
illustrated in Fig. 10.1. In this example futures trade at $60, and a
producer can consider buying $50 puts, which are financed by selling
$70 calls. If the distribution of oil prices happens to be perfectly
symmetric around its mean, which is determined by the futures price,
then such a collar trade should be theoretically costless. More generally,
the price for an equidistant collar represents the market measure of
asymmetry, or the skewness of the price distribution. If the distribution
were lognormal with a larger upside tail, then an equidistant collar
trade would have had positive premium to the call. Similarly, if the
market had a heavier left tail relative to the normal distribution, then
an equidistant collar would have traded at a premium to the put, which
is indeed more often the case in the oil market.
Fig. 10.1 The payoff profile for $50/$70 collar and $45/$55/$75 three-way collar with $60
futures price
In the case of a three-way collar, a producer can buy a $45/$55 put
spread with a maximum payout of $10 and retain larger unhedged
price upside by selling further OTM $75 calls. In return for raising the
effective floor to $55, the producer gives up any additional hedging
benefits if futures fall below $45. The producer often justifies the sale of
a lower strike put by the argument of already owning some real
optionality. If the strike of the short put is set at a sufficiently low level
that the shale production is no longer economical, then the producer
can monetize the physical option to close down production instead of
using the financial hedge. Such a rationale is highly debatable, and the
usage of three-way collars is widely considered to be a borderline
speculation by oil producers. It would be fair to accept that this
transaction lies somewhere in between a legitimate risk management
instrument and a speculative bet by a well-informed market
participant.
Another way to look at a three-way collar transaction is to
decompose it as a purchased put option financed by selling two other
options, an OTM call and an OTM put, collectively defined as a strangle.
Even though the transaction retains its costless structure in terms of an
aggregate option premium, it is unlikely to be neutral in terms of
volatility. The vega of the short strangle typically exceeds the vega of
the purchased put option, especially for longer-term deals, as two
options are sold for any one option bought. This means that in a three-
way collar trade the producer can hedge by selling volatility to the
market rather than by buying volatility as in the case of traditional
insurance. The proliferation of such leveraged hedging structures
increased the supply of volatility in the oil market by producers, which
led to a decreasing volatility risk premium, as documented in the
previous chapter.
Put options, two-, and three-way collars are the most common
hedging structures in the oil options market. Their market prices are
converted by option dealers into three primary volatility benchmarks:
an ATM straddle, an equidistant collar, and an OTM strangle. These
three benchmarks correspond to three moments of the underlying
price distribution. The price of an ATM straddle captures the market-
implied variance. An equidistant collar characterizes the skewness of
the distribution. An OTM strangle, which is embedded in the price of a
three-way collar, provides the market assessment of the tails of the
price distribution, or its kurtosis. At the end of this chapter, we develop
a new quadratic option pricing model whose three parameters are
mapped to the prices of three derivatives benchmarks.
The above-mentioned derivatives structures dominate the hedging
market, but the choice of strikes and maturities remains highly
dependent on the economics of an individual producer. The strike for
the purchased put usually depends on the cost of production, which
varies substantially across specific oil properties. The selection for
short options, on the other hand, is often driven by the hedger’s overall
business leverage and creditworthiness. The lender can also reduce the
leverage and credit risk by requiring the producer to pay some fixed
amount of cash premium, such as $1/bbl, for the collar transaction. In
these so-called premium collars, the transaction is no longer costless,
and strikes are iterated to match the target value for the entire package.
As a consequence, a typical oil option market-making portfolio has
open positions with numerous strikes resulting from various
transactions with end-users.
The additional selling of optionality by producers made the oil
volatility market more balanced. At the same time, the supply and
demand for individual strikes became more unbalanced and dynamic,
constantly changing as new hedging deals with different strikes come
to the market. Such an environment is a paradise for quantitatively
minded volatility traders. They could often charge one customer an
extra premium for an outright put option with a popular strike and
acquire potentially at a discount another option with a nearby strike
from a different client via a three-way collar or some other carefully
designed option package. Not only is the surcharge collected by the
dealer on both trades for providing the service, the dealer’s own market
risk is also simultaneously reduced.
To succeed in the competitive business of volatility market-making,
the trader must be ready to respond to any pricing request by end-
users and be able to act quickly in the market when an option structure
is perceived to be mispriced. What is mispriced, however, is model
dependent. If the trader can sell one option above its theoretical price
generated by a certain model and buy a similar option below its model
price, then it could be called mispricing only in the context of a specific
model. We would classify this as a model arbitrage. The question is, of
course, what model to use for tracking the constantly changing prices of
many different options, which are represented by the volatility smile.
However, before we proceed to a recommendation for such a model, we
first provide a cautionary warning about the application of
conventional approaches to this problem in the oil market.

10.2 Skew Delta and Two Types of Stickiness


The business of vega trading is very different from the insurance-like
business of gamma trading described in the previous chapter. A gamma
trader speculates on what the realized volatility will likely be during
some time in the future, typically the entire lifespan of the option. In
contrast, a vega trader is an arbitrageur who cares more about relative
pricing of different options and their corresponding short-term price
dynamics. The profits in volatility market-making are expected to
accumulate from the slight edge in original customer transactions. The
goal of the trader is to buy enough time to rebalance strikes in the
portfolio. In the meantime, the portfolio must be neutralized in a way
that isolates and protects the mispricing premium which was
embedded in the original trade. This can only be accomplished if the
portfolio is hedged using a theoretical model that governs the joint
price dynamics of options across all strikes, and prices produced by the
model are consistent with market prices of options that are used for
hedging.
As discussed in Chap. 8, relative prices of options with different
strikes are maintained in the form of the volatility smile. The smile
converts the assortment of prices into more convenient standardized
units of measure mostly to facilitate price tracking. Option implied
volatilities are maintained as a function of either strike or moneyness.
The moneyness can also be defined in a variety of ways, such as the
strike’s distance to ATM, its ratio to ATM, the same ratio scaled with
time to maturity, or as an option’s delta. The smile itself does not say
much about the richness or the cheapness of the option. Its presence
only indicates that the futures price does not behave according to the
assumptions of the model. If the smile is steep, it simply means that the
chosen model for the dynamics of futures prices is a poor
approximation of the real price behavior. The smile is merely a
communication tool, a sort of universal language which all options
traders agreed to use to efficiently liaise with each other.
The smile presents only a static snapshot of prices taken at a given
instant. But what a volatility dealer needs is a movie of how the smile
evolves when futures move. The static smile does not tell the trader
what option prices will be tomorrow, or how to calculate option deltas
to keep the portfolio price neutral while looking for opportunities to
unload unwanted strikes. To compare option prices today and
tomorrow, the trader needs to understand what drives the dynamics of
the entire smile. Since the smile itself is an artefact of the questionable
assumption of lognormality, modeling the behavior of something that
does not correspond to reality may look a bit odd. However, this is what
many traders do for operational convenience. The traditional approach
to describing the behavior of the smile relies on ad hoc heuristics that
dynamically adjust model inputs in an attempt to artificially preserve
its validity. We first demonstrate the shortcomings of applying this
approach to the oil market before replacing it with a more accurate
pricing framework.
Consider the snapshot of option prices for a given maturity, which
are expressed via implied Black volatilities v(K, F)

We explicitly write the smile v(K, F) as a function of two variables,


the strike price K and the futures price F. Since the smile is simply a
plug-in to the Black formula designed to match market prices of options
for a given futures price, the shape of the smile is likely to change once
the futures price moves. This makes the life of a volatility dealer more
difficult, as a change in the smile also impacts the option’s delta.
Finding the delta that keep the overall portfolio of options neutral with
respect to futures is critical for the volatility trader. The impact of any
slippage in the calculation of delta can easily dominate the initial mark-
up in pricing the option, thus ruining the entire arbitrage opportunity.
When the smile moves, the option’s total delta, which is the partial
derivative of the option price with respect to futures, must be
calculated using the chain rule, as follows:

(10.1)

where VBL is Black vega. In addition to the standard Black delta, the
total delta in (10.1) acquired another component, which is equal to the
product of Black vega and the slope of the implied volatility function
with respect to futures. This second term in (10.1) is colloquially
referred to as a skew delta. It represents the change in the value of the
option resulting from the change in implied volatility, induced by the
change in futures.
Unfortunately, the smile today does not tell us much about its
possible shape tomorrow, so we do not really know the true hedging
delta. What we observe today is the slope of the smile with respect to K.
However, what we need for hedging is the slope of the smile with
respect to F, which arises in the second term of (10.1). To calculate the
latter, one must impose additional assumptions about the evolution of
the smile. Since the smile is predominantly maintained as a function of
option moneyness, the path of least resistance is to assume that implied
volatilities remain the same for a given moneyness.
Let us illustrate the dynamics of the implied volatility in the
simplified case when the smile is linearly decreasing with slope β > 0,
so that

Such a linear approximation is reasonable in the oil market for


strikes in the vicinity of ATM1. To make sure that implied volatility in
this example does not drop below zero given its negative slope, we
flatten the smile once volatility reaches 0.20, as our interest is
predominantly in the linearly decreasing portion of the graph. Figure
10.2 illustrates it.

Fig. 10.2 The volatility smile behavior under the sticky moneyness rule

We first consider the scenario of sticky moneyness for which the


shape of the smile stays intact for a given moneyness of the option.
Assume that at time t, the futures price F(t) = $50, and ATM volatility
v0 = 0.30. Assume that in the next period, futures move to F(t + 1) = $60,
and the entire smile shifts in parallel while retaining the same shape
versus moneyness. The graph in Fig. 10.2 simply slides to the right.
Algebraically, the dynamics of sticky moneyness with a linear
volatility skew can be written as follows:

where v0 is ATM volatility for the strike K = F. When futures move up,
then ATM volatility v0 always stays the same, but the implied volatility
for any fixed strike K rises. This is the consequence of a negatively
sloped volatility smile and the positivity of the second term in (10.1)
since

Likewise, when futures fall, then implied volatilities for all strikes
decrease making options somewhat less expensive. Such behavior is
counterintuitive and potentially indicative of a problem with the
assumption of sticky moneyness. Since oil futures tend to have larger
downside gaps and percentage volatility is generally inversely related
to the futures price, it would be strange for the optionality to become
less valuable following the futures move in the direction of greater
uncertainty.
The problem manifests itself more clearly when one considers its
impact on the option’s delta. Since for a negatively sloped smile, the
skew delta is positive, the total hedging delta for sticky moneyness
always exceeds a conventional Black delta:

This inequality holds for all options, calls, and puts.


Let us illustrate the hedging delta with a simple numerical example.
Consider a one-year ATM option struck at K = F = $50 with an implied
volatility v0 = 0.3, and β = 0.004, which corresponds to a typical slope of
the Black skew near ATM. Using formulas for Black delta and vega from
Appendix B, we calculate the total hedging delta for an ATM call option

and for an ATM put option


In this example, the assumption of sticky moneyness results in 28%
delta for an ATM straddle. Such a high delta is counterintuitive, as one
would have expected an ATM straddle with a symmetric payoff function
to have delta closer to zero. Part of the problem here comes, of course,
from Black delta itself, which is around 12% for a one-year ATM
straddle, as a consequence of the skewed lognormal assumption. Sticky
moneyness makes the hedging problem worse, as it magnifies the
artificially skewed positive delta bias. This suggests that the seller of
100 contracts of ATM straddle should buy 28 futures as a hedge. Such
an asymmetric hedge would be very bizarre in the oil market, especially
given its large downside price gaps.
Hedging a portfolio of options using the operationally convenient
setup of sticky moneyness often leads to catastrophic consequences.
Since the skew delta is positive for all options, managing a portfolio of
short options requires holding additional long futures. This means that
for any put option sold to a producer, sticky moneyness forces the
dealer to sell fewer futures than truly needed for delta hedging. A
portfolio which may appear to be delta neutral will behave with a long
futures bias. In a market that follows an up the stairs, down the elevator
dynamics, such a hedge could lead to large losses. Once sticky
moneyness traders realize their mistake, they often panic to unload the
excess of long futures. This often coincides with heavy pressure from
the negative gamma hedgers which periodically sends futures into a
downward spiral. Unfortunately, this is how many unsuccessful
volatility traders were forced to end their trading careers in the oil
market.
The conceptual problem with sticky moneyness is that implied
volatilities for all strikes are shifted in the wrong direction. Falling
implied volatility for each option in a falling futures market does not
feel right, as one should expect precisely the opposite to happen, at
least, if volatility is measured in percentage terms (recall Fig. 8.​4). Since
collectively dealers share the downside risk of a price collapse, they will
attempt to reduce their volatility exposure when futures fall closer to
put strikes sold to producers. The proximity to short strikes makes
dealers shorter volatility. To stay within allocated risk limits, dealers
have no choice but to buy back some short options, which will
inevitably push implied volatility higher. However, the naturally rising
percentage volatility in a falling futures market would be out of sync
with the dynamics imposed by the sticky moneyness assumption.
The problem with sticky moneyness does not go away for markets
with larger upside risks, such as markets for natural gas and refined
products. In these right-tailed markets, the implied volatility skew is
more likely to have a positive slope. If futures rise and the smile slides
in the sticky moneyness fashion with the same ATM volatility, then
given the positive slope of the smile, implied volatilities for each strike
decrease. This is again counterintuitive, as specific strike volatilities are
unlikely to fall when futures move in the direction of larger uncertainty,
which is now on the upside. In this setup, the skew delta is negative.
This forces the dealer to mistakenly under-hedge short OTM calls,
which is highly dangerous in positively skewed markets.
One simple way to mitigate the counterintuitive effects of sticky
moneyness is to assume an alternative heuristic of a sticky strike. Under
the sticky strike assumption, the volatility smile is assumed to retain its
shape across all given strikes, regardless of their moneyness. Since the
smile is still operationally maintained in terms of moneyness, to ensure
that implied volatilities stick to the same strikes, one must perpetually
adjust the shape of the smile by moneyness.
Figure 10.3 illustrates the sticky strike behavior. The smile graphed
versus specific strikes is static. The sticky strike essentially preserves
the Black formula with a one-off volatility adjustment for each strike. In
this case, ATM volatility does not shift in parallel when futures move,
but instead it slides up and down the skew.
Fig. 10.3 Volatility smile behavior under the sticky strike rule
To characterize the linear sticky strike rule, one only needs to
specify the slope β along with an initial futures price F0 for which the
corresponding volatility is v0. Its dynamics in the linear case is
described as

The second term in (10.1) then vanishes because the smile is static;
it depends only on fixed F0 but does not depend on future price F. The
total hedging delta in the sticky strike case is identical to Black delta for
calls:

and likewise, for puts

The sticky strike heuristic forces the trader to depart from the path
of least operational resistance determined by sticky moneyness.
Instead of doing nothing and letting the smile slide in parallel, the
sticky strike requires the trader to constantly repivot the smile to the
new ATM level and then tweak its shape by moneyness, as required by
operational standards.
At this point, one may wonder why not just keep the entire smile in
terms of fixed strikes? Theoretically, it is doable, especially for short
time periods when the range of popular strikes does not change. In the
long run, however, keeping the skew by strike is impractical. The
market always talks in terms of moneyness, and it uses ATM volatility
as a pivot point for the construction of the smile. Nearly all risk and
accounting systems require smiles to be saved by moneyness to
facilitate standardization across multiple assets. While the trader can
maintain the sticky strike smile for decision making, its system-
required shape by moneyness must be constantly bent. The reward for
this hassle comes from the elimination of undesirable and
counterintuitive skew delta corrections. The improvement, however, is
marginal. The delta of an ATM straddle is still positive, which is caused
by the assumption of lognormality. In the previous numerical example
even with no contribution from the skew delta, the ATM straddle still
has 12% delta. It is half as bad compared to the sticky moneyness case,
but it is still not a viable hedging solution as one should expect ATM
straddle delta to be closer to zero.
The obvious question arises now whether it is possible to come up
with other heuristics that shift artificially skewed Black deltas in the
right direction. Yes, it can be done. Nothing can stop the trader from
specifying any other arbitrary dynamics of the smile. The trader can
draw a certain glide path for an ATM volatility that does not simply
slide along the same shape, but instead increases or decreases with any
chosen slope. An equally arbitrary skew can then be attached to this
glide path of ATM volatility to complete the smile dynamics. However,
all such heuristics amount to more complex but still artificial tweaks to
a base pricing model that does not represent the real market behavior.
The problem can only be solved by treating its cause and not the
symptoms.
One challenge is the natural asymmetry embedded in the lognormal
distribution. While the assumption of lognormality might be reasonable
for some financial markets that tend to grow over time, it becomes
problematic for mean-reverting oil markets, where the underlying price
distribution is much more symmetric. In Chap. 8, we illustrated how the
desired symmetry can be restored if one switches from lognormal to
normal measures of the realized volatility. In the following section, we
take a similar route for the implied volatility and compare the behavior
of volatility smiles, as they are seen from two alternative angles.

10.3 When Black Smirks, Bachelier Smiles


The history of oil implied volatility, like many other topics related to oil
trading, is very colorful and multifaceted. Its complete story might even
warrant a separate publication to be filled up with fascinating
anecdotes of extreme volatility spikes caused by a panic of option
sellers running for cover in the aftermath of squeezes, wars, hurricanes,
and other unexpected market disruptions. Somewhat regretfully, we
leave these anecdotes for another time and only look at the history of
volatility through some observations from implied volatility data. Our
goal is to highlight some salient properties of oil volatility that can
guide us towards a better model, which will be developed in the
following section.
Certain key characteristics of the oil markets have already become
apparent from Chap. 8 where we looked at the historical behavior of
realized volatility versus the futures price. We now extend the same
view to implied volatility. We start with the benchmark ATM implied
Black volatility (IBV) and look at the history of implied volatility in the
form of a scatterplot versus the futures price, shown in Fig. 10.4.
Fig. 10.4 ATM IBV and futures prices for third nearby contract (WTI, 2000–2022)
Like realized volatility, IBV also exhibits a strong inverse
relationship with futures. To see more clearly the evolution of this
relationship, we split the historical sample into two sub-periods that
correspond to different regimes outlined in the previous chapter. In the
earlier period, IBV showed little dependency on the futures price,
providing some support to the assumption of lognormality. However, in
the latest regime of financialization and shale, the inverse relationship
became much stronger and highly nonlinear.
By now, a careful reader must have recognized that these
observations are also consequences of the skewed lognormal volatility
metric. The volatility of percentage returns rises algebraically when the
denominator of the ratio defining the return decreases. Recall our
analogy of looking at the market through lognormal lenses to glasses
for astigmatism worn by someone who does not need them. Imagine
now that over the years the overall vision of this patient somehow
normalized, but the patient is still wearing the same old glasses with
the wrong prescription. The perception that was blurry to begin with,
may not even let the person see anything at all.
Let us now switch to better lenses and look at the same historical
picture using implied normal volatility (INV). Thankfully, we do not
need to recalculate another set of implied volatilities, as the two
metrics can be easily transformed from one into the other. To relate IBV
and INV, we equate the Bachelier formula (8.​9) with normal volatility vN
and the Black formula (8.​10) with lognormal volatility v. By definition
of implied volatilities, the two formulas must produce the same
observable market price:

For an ATM option, when K = F, this pricing identity simplifies


considerably, as follows:

Since the normal probability density is the derivative of the


cumulative normal function, i.e., , we can apply the Taylor
rule to the right-hand side of this identity, and obtain that for short
maturities τ:

It follows that for ATM options with short maturities INV is


approximately equal to IBV multiplied by the futures price

This approximation is very accurate and commonly used by traders


to switch between the two alternative implied volatility metrics.
The shift from skewed lognormal lenses to normal ones provides a
different look at the historic relationship between prices and implied
volatilities. This alternative view is presented in Fig. 10.5, where INV is
shown versus the futures price.
Fig. 10.5 ATM INV and futures prices for third nearby contract (WTI, 2000–2022)
In the first period, ATM INV exhibited little convexity, generally
moving in the same direction as the futures price, which is consistent
with a more lognormal type of behavior. If IBV remains unchanged in
percentage terms and futures rise, then INV and the price of an ATM
option in dollars per barrel must also rise. In the second period, the
curvature on both sides of the graph is more visible. One can see
increasing volatility at lower prices, which is driven by financialization
and shale growth with increasing spillovers from equity markets and
periodic constraints on maximum storage capacity. The curvature of the
right tail can be attributed to larger impact of supply disruptions, given
a reduced buffer of spare production capacity and reluctance by
producers to commit to long-term investments in anticipation of energy
transition. As a result, the relationship between INV and futures
became somewhat parabolic.
While the analysis of ATM volatility sheds some light on the
underlying price distribution, a better diagnostic of the volatility
behavior can be obtained by looking at the entire volatility smile.
Conveniently, IBV and INV smiles can also be transformed from one to
the other by using the following simple approximation2:
(10.2)

It follows that for OTM options, the ratio of INV to IBV is


approximately equal to the ratio of dollar moneyness to the logarithmic
moneyness. This approximation has also proven to be remarkably
accurate. The average approximation error over the entire data set is
much smaller than 0.1% if measured in terms of Black volatilities.
Obviously, the approximation cannot be used for negative strike prices,
when the Black formula becomes meaningless and market participants
use INV as a primary volatility benchmark.
To complete the historical study, Fig. 10.6 shows the average Black
and Bachelier smiles calculated over the second period, which is more
reflective of the current market conditions. To make the two curves
visually comparable, the normal smile is scaled by the corresponding
futures price, so that the shapes of both volatility curves can be
displayed side by side in percent. Since the smile is shown here for
fixed maturity, for transparency, the moneyness is defined by a simple
ratio K/F.
Fig. 10.6 Historical average IBV and INV/F by moneyness for third nearby contract (WTI,
2015–2022)
It is important to remember that smiles are nothing more than
option prices displayed in different coordinate systems. Smiles attempt
to correct the flaws of simple models by tweaking their inputs to match
observable prices. Only a horizontal line for the smile indicates model
perfection. The magnitude of a smile’s deviation from a straight line is,
therefore, a measure of the model error, or the degree to which the
model misrepresents reality.
The Black smile in Fig. 10.6 is highly asymmetric, which is caused by
the natural skewness of the lognormal distribution. The presence of
such a steep skew confirms that the market disagrees with the
assumption of lognormality. This assumption underestimates the
probability of large downside price moves (down the elevator), while it
overestimates the likelihood of modest upside moves (up the stairs).
The market is trying to compensate for this shortcoming of the Black
model by assigning higher IBV parameters for lower strikes and lower
IBV for prices slightly above ATM.
In contrast, the INV smile is much more symmetric, and it appears
to be nearly parabolic. This symmetry is intuitive, as oil is, at least, as
likely to move down as it is to move up by the same dollar amount.
Note, however, that the vertex of the normal parabolic smile is also
typically located slightly above ATM. It often corresponds to the point of
maximum hedging pressure resulting from selling of slightly OTM calls
by oil producers to finance mandatory purchases of puts and put
spreads.
The benefits of using the more symmetrical INV are clearly seen in
managing the problem of deltas, which could not be rectified with
sticky moneyness and sticky strike band-aids applied to the naturally
skewed IBV metric. In the normal framework, the deltas for ATM calls
and puts are, respectively, positive and negative 50%, and the delta of
ATM straddle is zero.3 This is much more intuitive than the 12% ATM
straddle delta in the previous example of the lognormal sticky strike
model, or the 28% delta under the assumption of sticky moneyness.
The ATM normal delta simply reflects the probability of futures moving
up or down from the current level. In other words, the ATM delta also
represents the fair price for playing a coin-flipping game to receive one
dollar or nothing, depending on whether futures rise or fall. If a random
draw that determines the outcome of the game is taken from the
symmetric normal distribution, then nobody should be paying more
than fifty cents on a dollar to participate in such game.
Having addressed the issue of artificial lognormality-induced
skewness, we can now move to handling the more challenging problem
of capturing fat tails of the oil price distribution, or large price moves
that occur way more frequently than allowed by either the lognormal or
the normal assumption. The problem is evident from the wings of both
smiles in Fig. 10.6. Since neither assumption can generate sufficiently
large probabilities of extreme moves, these models require larger
volatility parameters to match prices for deep OTM options. This
creates the curvature on both ends of the smile. The upside tail is
driven by the risk of supply disruptions caused by unexpected
geopolitical or weather events. The downside tail reflects the negative
producer gamma and the risks of macroeconomic spillovers, which are
further amplified by periodic constraints on the available storage
capacity.
The simple one-parameter models of Bachelier and Black are too
rigid to be able to capture the important effects of skewness and fat
tails, which are typical for the oil price distribution. At a bare minimum,
at least two more parameters are needed to better characterize the
joint dynamics of prices and volatilities. Taking some inspiration from
the empirical evidence of parabolic normal smiles, we now develop a
novel extension of the option pricing framework based on a diffusion
process with quadratic local volatility.

10.4 Fat Tails and the Quadratic Normal Model


Modeling prices of financial instruments always involves a trade-off
between the desired accuracy of the model and its stability. At the one
extreme, there are simple one-parameter models, such as the ones
developed by Bachelier and Black. They are convenient to use, but, as
evidenced by the presence of volatility smiles, these models are not
flexible enough to describe the richer dynamics of oil prices. At the
other extreme, one can construct very complex models with numerous
parameters that can always be tweaked to fit the data, but such fits are
unlikely to last long. Once the market moves, overparametrized models
quickly fall apart, ruined by the instability of fitted parameters. Such
models are characterized by multiple minima in the space of
parameters, where various permutations of inputs compete to produce
the best fit. Besides, the more parameters the model has, the less
intuitive it is for traders. In this section, we propose a sweet spot and
modify a simple model just enough to capture the most important
features of the oil price dynamics, while retaining transparency and
intuition.
At this point, we should elaborate on the difference between what
we define as a proper model and ad hoc heuristics, such as sticky strike
and sticky moneyness. A heuristic is a certain pattern in the behavior of
outputs produced by some unknown pricing engine. These outputs are
option prices for different strikes, typically expressed in IBV terms, to
which a heuristic rule assigns a certain dynamic without much regard
to where such dynamics might have come from. Since the root cause
behind such dynamics is not even considered, heuristic rules often fail
in practical applications, for example, in the calculation of hedging
deltas. In contrast, a proper pricing model should not depend on any
option-specific characteristics, such as its strike price, as options across
all strikes are functions of the same futures contract. What needs to be
modeled is not the evolution of outputs generated by some undefined
pricing engine, but the pricing engine itself that describes the dynamics
of the underlying futures. Then one no longer needs to guess option
deltas and speculate on the dynamics of the smile which is implied by
the model.
In our preferred setting of generalized diffusions, the model is
described by the local volatility function, as in (8.​5). To price an option,
one needs to solve the differential Eq. (8.​8), where the volatility
coefficient is given by a general deterministic function. This equation
can be solved analytically only for a few special cases, including
constant and proportional volatilities, which we have already
considered, and several other volatility specifications that have not
proven to be particularly useful for the oil market. To solve the diffusion
equation with any other local volatility functions, one must resort to
numerical techniques, such as finite difference methods. However,
traders are always a bit leery of numerical black boxes built by quants.
Traders like to be in charge of their own destiny and often prefer to use
more intuitive analytical approximations rather than numerical
methods. The practical benefits of an analytical representation are also
evident in the calculation of Greeks, which can be computed by a
straightforward differentiation of the closed-form solution for the
option price.
A popular technique used in applied sciences for deriving analytical
approximations to solutions of complex nonlinear problems is based on
the method of perturbation or linearization. The idea is to represent a
solution to a complex problem as the sum of the solution to a simpler
problem and a first-order perturbation of the latter. Conceptually, this is
not too different from the idea of a Taylor series expansion, but it is
applied to the much more complex functional relationship between
option prices and local volatility governed by the diffusion equation.4
Given the substantial evidence of the superiority of the normal
assumption over the lognormal one, we use the former for the base
solution to more complex diffusion problems. The solution to a
simplified problem is then given by the Bachelier formula with constant
normal volatility. To correct it for its main flaw of underpricing extreme
events, a nonlinear perturbation is applied to constant normal volatility.
More precisely, we let the local volatility function to be described by a
relatively small perturbation of the constant normal volatility, specified
by the function ε(F)

The solution to (8.​8) is then constructed as the sum of the base


solution given by the Bachelier formula, CBC(F, t), and the skew
correction function U(F, t), which is designed to capture the impact of
non-normality:

Since the function C(F, t) must solve the diffusion equation with
local volatility σ(F), we substitute this decomposition into (8.​8), and
make use of the fact that CBC(F, t) also solves the same equation but
with constant coefficient σA. The main idea of the perturbation method
is to retain only the terms of order ε(F) and discard higher-order terms,
whose contribution is smaller for a relatively small ε(F).
It is shown in Appendix C that this substitution leads to the
following equation for U(F, t):

(10.3)

where, as before,

denotes the scaled normal moneyness. It follows that U(F, t) also solves
the BSM equation with constant coefficient σA, but its right-hand side
has an additional term which is proportional to the perturbation
function ε(F) and Bachelier gamma.
To solve (10.3), we need to supplement it with a boundary
condition. At the expiration time, C(F, T) is defined by the option’s
terminal payoff. The same payoff is also produced by the Bachelier
formula CBC(F, T) at expiration. Therefore, the terminal boundary
condition for the skew correction function is zero:

In the regular diffusion Eq. (8.​8), the impulse to an option price is


sent by its boundary condition at expiration. Here, instead the diffusion
is being driven by the new term in the right-hand side of the equation,
which impacts the dynamics throughout the life of the option.
The Eq. (10.3) is written for a yet to be specified function ε(F) which
can be chosen based on properties of the market and observed implied
volatility skews. Motivated by empirical observations of parabolic
normal smiles from the previous section, we specify the perturbation
function in the quadratic form:

Note that while constant parameter a may appear to be redundant


as it is added to another constant σA, in practice, it is convenient to
write it explicitly. Here, the perturbation is applied to constant volatility
σA, which is usually taken to be ATM volatility. It only plays the role of
the starting point to characterize the approximating normal
distribution. For any given σA, parameters a, b, c adjust base option
prices for variance, skewness and kurtosis of the underlying market
price distribution. Such a parametrization of the local volatility function
is known as a quadratic normal (QN) model.5
The Eq. (10.3) with a quadratic perturbation function is solved
analytically in Appendix C. The price of a call option in the presence of
skewness and fat tails is given by a simple formula

(10.4)

where

denotes Bachelier vega with the constant normal volatility σA, and

(10.5)

provides corrections to prices for OTM options.


The three-parameter parabolic model appears to represent a sweet
spot in modeling oil options that we were seeking. It is significantly
more flexible than single-parameter normal and lognormal models. At
the same time, the QN model preserves valuable intuition as its three
parameters are directly linked to moments of the probability
distribution of futures prices. Such intuition is very appealing to
traders, as these distributional characteristics are mapped to three
benchmark option trades in the market: ATM straddle, costless collar,
and OTM strangle.6 The closed-form nature of the formula (10.4 and
10.5) allows for explicit calculation of all Greeks that traders use for
hedging. The model generates a much richer set of deltas depending on
the interplay between the skewness and the curvature. It has proven to
be able to capture the actual dynamics of oil option prices very
accurately.
The method of perturbation is very powerful and can be extended
in several directions. If more accuracy is desired, then the perturbation
function ε(F) can be specified in the form of a higher-order polynomial
in which case the Eq. (10.3) can still be solved analytically. The
perturbation technique can also be applied to the Black formula instead
of Bachelier. The corresponding solution retains a structure similar to
(10.4) where σA is replaced with σGF, and normal vega VN is replaced
with the lognormal vega V. However, the exact formula is more
cumbersome as parabolas are naturally less compatible with geometric
volatilities. The formula simplifies only if the local volatility is assumed
to be quadratic with respect to lnF instead of F. Furthermore, in the
lognormal case, one would be correcting a base model which is worse
off to begin with, which makes the perturbation method less robust.
Given the inferiority of the quadratic lognormal model, we limit the
presentation in this book to a more practical quadratic normal model
that works particularly well in the oil market.
In this part of the book, we dealt with vanilla options that depend
only on a single futures contract. This allowed us to fix the time
remaining to maturity of the option and focus on the relationship
among option prices across different strikes. In the following three
chapters, we add more advanced features that commonly arise in OTC
oil options that depend on multiple futures contracts. Trading such
options brings an extra dimension. In addition to linkages between
strikes, no-arbitrage boundaries must also be established for option
contracts with different maturities. Furthermore, the idea of volatility
arbitrage presented in this chapter is not complete until we show how
to choose the parameters of the pricing model, the topic that will be
covered in the chapter on model calibration.

References
Andersen, L. (2011). Option pricing with quadratic volatility: A revisit. Finance and Stochastics,
15(2), 191–219.
[MathSciNet][Crossref][zbMATH]

Bouchouev, I. (1998). Derivatives valuation for general diffusion processes, Proceedings of the
Annual Conference of the International Association of Financial Engineers, New York, USA, pp.
91–104.

Bouchouev, I. (2000, August). Black-Scholes with a smile. Energy and Power Risk Management,
pp. 28–29.
Carr, P., Fisher, T., & Ruf, J. (2013). Why are quadratic normal volatility models analytically
tractable? SIAM Journal on Financial Mathematics, 4(1), 185–202.
[MathSciNet][Crossref][zbMATH]

Castagna, A., & Mercurio, F. (2007). The vanna-volga method for implied volatilities. Risk, 20(1),
106–111.

Derman, E., & Miller, M. B. (2016). The volatility smile. Wiley.


[Crossref]

Gatheral, J. (2006). The volatility surface: A Practitioner’s guide. Wiley.

Grunspan, C. (2011). A note on the equivalence between the normal and the lognormal implied
volatility: A model free approach, SSRN.

Ingersoll, J. E., Jr. (1997). Valuing foreign exchange rate derivatives with a bounded exchange
process. Review of Derivatives Research, 1, 159–181.
[Crossref][zbMATH]

Rebonato, R. (2004). Volatility and correlation: The perfect hedger and the fox. Wiley.
[Crossref]

Schachermayer, W., & Teichmann, J. (2008). How close are the option pricing formulas of
Bachelier and Black-Merton-Scholes? Mathematical Finance, 18(1), 155–170.
[MathSciNet][Crossref][zbMATH]

Zü hlsdorff, C. (2001). The pricing of derivatives on assets with quadratic volatility. Applied
Mathematics Finance, 8(4), 235–262.
[Crossref][zbMATH]

Footnotes
1 We only provide a simple illustration of this topic. For a more advanced exposition, we refer
to Rebonato (2004), Gatheral (2006), and Derman and Miller (2016).

2 For brevity, we omit the rigorous derivation of this approximation. It is again based on the
Taylor series expansion for both formulas, but with more cumbersome mathematical details
which are less essential for our goals. We refer to Schachermayer and Teichmann (2008) and
Grunspan (2011) for a more detailed comparison of the two formulas, and more advanced
approximation formulas.

3 Recall that we are using zero interest rate. In the general case, deltas must be discounted.
4 The method of linearization and the related parametrix method for the diffusion equation
were originally applied to option pricing in Bouchouev (1998, 2000).

5 Quadratic parametrizations of the local volatility function have been considered by several
authors, including Ingersoll (1997), Zü hlsdorff (2001), Andersen (2011), and Carr et al. (2013),
but resulting option pricing formulas are considerably more complicated.

6 The use of three parameters that correspond to the first three moments of the price
distributions is often used by market-makers to characterize the dynamics of the volatility
smile in financial markets. For example, Castagna and Mercurio (2007) developed a popular
vanna-volga model for options in foreign exchange markets.
Part IV
Over-the-Counter Options
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
I. Bouchouev, Virtual Barrels, Springer Texts in Business and Economics
https://doi.org/10.1007/978-3-031-36151-7_11

11. Volatility Term Structure and Exotic


Options
Ilia Bouchouev1
(1) Pentathlon Investments, LLC, Westfield, NJ, USA

– Exchange-traded contracts constitute only one side of the oil


derivatives market. The other one is hidden from public eyes in the
obscure world of over-the-counter (OTC) trading, where more
complex options trade bilaterally. Among thousands of OTC deals,
one particularly secretive sovereign hedging program stands out.
– Many OTC options depend on contracts with multiple maturities. The
pricing of such options is driven by the forward dynamics of the local
volatility. The quadratic mean of the local volatility over the lifespan
of an option determines the implied volatility quoted in the market.
– An average price option (APO) is the preferred hedging instrument
among end-users whose exposure to oil prices is ratable. While
precise APO valuation is not possible in the conventional lognormal
setting, a simple transformation inspired by the normal model
produces an accurate approximation.
– The expiration of certain OTC derivatives is decoupled from the
expiration of the underlying futures. One example of such a
derivative is a swaption, which represents a strip of early-expiring
European options. The pricing of swaptions and many other exotic
oil derivatives requires a multi-factor framework.

11.1 Dark Pools and the Hacienda Hedge


When the media talk about oil options, they almost certainly reference
options traded on organized exchanges. These are the only oil options
that individual investors can easily monitor and trade. Exchange-listed
options are relatively transparent. All transactions, prices, volumes, and
open interest are visible to the public. Professional volatility traders call
them vanilla options. These options are tied to corresponding monthly
futures contracts, expiring three business days prior to the futures
expiration. Vanilla options are listed for a wide range of strike prices
with $0.50/bbl increments. Such a fine granularity and transparency
makes them popular among speculators, who can pinpoint their wagers
and easily track their performance. However, these options represent
only one visible side of the oil options market; the other one, which
operates OTC, is covered with mysteries.
An OTC market is bilateral, where prices are privately negotiated
between professional dealers and end-users. The market for private oil
transactions developed long before any of the modern organized oil
exchanges.1 The introduction of oil exchanges did not replace the OTC
market, which is still very significant. One large incentive for corporate
hedgers to trade directly with banks is the need for credit. While vanilla
options are tightly margined by clearing houses, privately negotiated
collateral agreements that support OTC trades are more lenient. For
end-users, securing favorable margining terms to support their hedging
program is crucial. An oil producer can hardly afford to take the risk of
posting cash collateral against mark-to-market losses on forward
hedging. Since collateral must be posted with the exchange
immediately, while the production is still to come, any short-term price
spike can easily put the producer out of business. In contrast, banks
have proven to be very accommodating to producers, often waiving
cash margin requirements and instead accepting oil reserves as a form
of collateral against hedging deals. Likewise, in the case of consumer
hedging, there was even a precedent of an airline using an aircraft as
collateral on OTC oil hedging. Obviously, such credit accommodation by
dealers comes with a price. Banks usually act in a dual capacity as a
lender and as the dealer for derivatives hedges, which are sufficiently
profitable to compensate for carrying additional credit risk.
Another advantage of the OTC market is its versatility. In contrast to
vanilla options, bilateral oil derivatives are highly customized.
Standardized exchange-traded options are not suitable for many
producers and consumers whose exposure to oil prices is ratable.
Buying an option contract that settles based on the futures price on a
given day is unlikely to be a good hedge for the end-user. The futures
price on any single day can be easily impacted by a multitude of one-off
factors. Instead, corporate hedgers resort to the services of OTC
dealers, who can provide more tailored protection against unwanted
risks. In addition, the end-user can benefit from the credit line extended
by the dealer where no collateral is required against mark-to-market
hedging losses up to a certain limit.
Since end-users are usually exposed to oil prices daily, what they
need is a hedge against fluctuation in the average price of oil. End-users
can usually live with daily price volatility if the average oil price
remains within their budget. A better hedge for them would be an
option that settles based on the average futures price over a certain
period rather than on a particular day, when the price can be easily
distorted. Such OTC options are known as average price options (APOs),
or Asian options. A typical OTC oil option transaction involves a
quarterly or an annual strip of monthly settled APOs. Each day the
settlement price of the prompt futures contract contributes to the
calculation of the monthly average. If futures expire sometime during
the month, like in the case of a WTI contract, then the monthly
settlement is the weighted average of two futures contracts with
consecutive maturities. The corresponding weights are based on the
number of days that each futures represent the prompt contract.
Intuitively, price averaging must also reduce volatility and make
APOs cheaper in comparison to similar maturity vanilla options. This
additional benefit of volatility dampening has made APOs particularly
attractive for end-users. Conventional APOs are settled monthly, but a
hedger can achieve even larger volatility discount by trading so-called
term APOs, which settle based on the average price over a longer time
period, such as the entire year. A term APO, however, provides less
protection than a strip of monthly APOs. This is because the latter could
well have positive payoffs for certain months, even when the price
average over the entire term does not reach the strike price. The longer
the averaging period, the more significant the savings are from the APO
volatility discount.
The largest, and arguably the single most important derivative
transaction in the oil market is structured as a term APO. No book on oil
trading can be complete without describing this unique large-scale
sovereign hedging program executed annually by the Government of
Mexico. It is designed as an insurance policy to protect the country
budget, which is heavily dependent on revenues from oil exports. For
over a decade, this hedging program has rocked and rolled oil markets
with fortunes made by some traders and large losses suffered by
others.2 The market colloquially refers to it as the Hacienda hedge. The
program grew to be so important for the overall oil market that the
Government of Mexico stopped providing any information about
hedging transactions after 2020, unofficially treating it as a state secret.
The concept of hedging the country’s oil export revenues was first
introduced by Mexico in 1990 in response to the oil price spike at the
beginning of the Gulf War. The pilot program was largely successful and
ensured higher export prices for the following year. However, it took the
country over ten years to fully develop the hedging framework and
support it with in-house derivatives expertise. Convincing the
lawmakers to spend taxpayers’ money on buying oil put options has
proven to be challenging given the high political risks in making such
decisions. Finally, in 2001 a special structure was created in the form of
an oil stabilization fund with an annual budget allocation specifically
earmarked for paying the put option premium.
The program evolved over the years, but its core components
remained largely intact. The put option is settled as a term APO based
on the average futures price over a twelve-month period, usually
ending in November, to allow for one month to finalize the option
settlement before the end of the calendar year. The strike price, which
is typically set slightly below the forward price for the following year, is
linked to the price set in the country’s budget for the exports of a
Mexican oil grade, called Maya. The market price of Maya is determined
by a basket of more liquid futures with an additional adjustment by a
non-tradable factor set by Mexico itself.3 The presence of such an
unhedgeable risk complicates the task for the dealer, as proxy hedges
must be constructed to approximate the dynamics of Maya with liquid
futures. The price for this extra hassle and the compensation for
bearing an additional basis risk is embedded into the option premium.
Since the government’s primary objective is to protect the average
export price for the entire fiscal year, it does not necessarily need to pay
up for purchasing a more expensive protection against monthly price
fluctuations. It should be perfectly content with buying an option
settled based on the average price over the entire year. The long-term
performance of the Mexican hedging program is representative of a
typical insurance payout characterized by regular annual payments and
occasional large compensation. Figure 11.1 shows that large payouts
occurred in four years when oil prices collapsed. In aggregate, these
payouts approximately add up to the cumulative premium paid over the
years, confirming that the money was well spent. By using the
derivatives market, the government was able to reduce the volatility of
the country’s budget and accomplish this without paying any excess
risk premium. In other words, by and large, the program was executed
at a fair actuarial price.

Fig. 11.1 Profit and Loss of hedging program by the Finance Ministry of Mexico. Sources:
Auditoría Superior de la Federació n, Secretaría de Hacienda y Crédito Pú blico, Oxford Institute
for Energy Studies

The fact that Mexico was able to avoid paying any structural VRP
over the years by timing the execution well does not mean that option
dealers did not make any money. They sure did. The deal highlights the
importance of the difference between the actuarial value of an option
and its fair value determined by delta hedging, as discussed in Chap. 9.
For volatility market-makers, this hedging program turned into a high-
stakes skill competition for a slice of the approximately one-billion-
dollar premium paid by the government annually. The winners and
losers in this competition are often determined by how accurately they
can estimate the APO volatility discount for pricing the deal and
corresponding deltas for managing the risks.
This sovereign hedge by Mexico is only one among many other
examples of a customized OTC derivative, but it illustrates well the
complexity of the task faced by volatility traders. These complex deals
bring unparalleled profit opportunities to the dealers, but they are
inevitably accompanied by equally large risks of mis-hedging if the
price dynamics and resulting hedging ratios are not captured correctly.
The quantitative challenges here are substantial. In addition to tricky
handling of price averaging and basis risks, the trader must also
incorporate the volatility term structure, as the tenor of many OTC
derivatives spans periods covered by multiple futures contracts.
Up until now, we have dealt only with options written on a single
futures contract that allowed us to simplify the problem by fixing the
corresponding maturity. However, even a typical monthly APO volatility
for WTI depends on the volatilities of two different futures due to the
rolls that must occur sometime during the month. Likewise, the
volatility of an annual term APO is determined by volatilities of thirteen
vanilla options that correspond to prompt futures throughout the
calendar year. The volatility of a term APO depends on the slope of the
implied volatility term structure, which is particularly sensitive to
short-term fundamental uncertainty in the physical market. The
analysis of the volatility term structure is a prerequisite for valuing not
only APOs, but also many other OTC options which will be introduced
later in this chapter.

11.2 The Term Structure of Implied and Local


Volatilities
Modeling non-standard oil options can be a rather daunting exercise. As
in many other modeling choices that we have already encountered, one
should strive for some reasonable trade-off between the complexity of
the model and its robustness. This trade-off depends on the nature of
the specific trading opportunity. What works for one trade may not be
suitable for another. Traders must always pick their modeling battles
based on the dominant risk in the portfolio.
For example, the volatility smile trader focuses on capturing relative
value opportunities across options for various strikes, but typically for
the same maturity. For such a trader, modeling distributional properties
of a given futures contract is more important than linking up contracts
with different maturities. Since each vanilla option depends only on a
single futures contract, the smile for each maturity can be analyzed in
isolation. In contrast, an OTC trader is more concerned about volatility
over a period that spreads across futures with multiple expirations.
While capturing key distributional properties is still important for
some OTC trades, a more pressing task is to understand how the
volatility evolves with time. To get a better sense of the essential time-
varying properties of volatility, the term structure of volatility is usually
analyzed under simpler distributional assumptions.
In this chapter, we mostly focus on the conventional lognormal
framework and the GBM process. We let the geometric local volatility
σG(t, T) that drives this process depend on the current time t and on the
expiration of the futures contract T but we make it independent of the
futures price:
(11.1)
In parallel, we also consider the ABM process, defined by the time-
varying arithmetic local volatility:
(11.2)
The normal framework turns out to be particularly useful for
deriving closed-form option pricing formulas. These analytic formulas
are then used as convenient approximations in a more conventional
lognormal setting.
The diffusion differential Eq. (8.​8) applied to the GBM with time-
varying volatility is
(11.3)

Fortunately, one can easily eliminate volatility time-dependency in


(11.3) by switching to a new time variable scaled with the local
volatility. It is shown in Appendix D that the solution to (11.3) is given
by the regular Black formula with volatility v(T), which is calculated as
the quadratic mean of the local volatility σG(t, T) over the life of the
option:

(11.4)

This well-known extension of the BSM pricing framework, originally


developed by Merton,4 has some important implications. It shows that
time-dependent volatility does not bring any serious obstacles to the
valuation of options. The same Black pricing formula still applies for
options on futures. Its single constant volatility parameter is simply
replaced with the quadratic mean of the local volatility over the lifespan
of the option. Since we have chosen to ignore the impact of the volatility
smile for this analysis, v(T) can be taken to represent the term structure
of the market implied ATM volatility.
In general, the local volatility σ(t, T) represents a two-dimensional
function that describes the volatility behavior at time t of the futures
contract expiring at time T. In contrast, implied volatility v(T) is a one-
dimensional function of T. It only contains information about the
quadratic mean of the local volatility function over the entire period,
and it does not say anything about its path over time. Obviously, there
are many local volatility time paths that have the same quadratic mean,
i.e., the same implied volatility. This makes it impossible to create a
one-to-one mapping between local and implied volatilities without
additional assumptions.
To match the dimensionality of local volatility and implied volatility,
we impose an extra restriction on the local volatility. Specifically, we
assume that it only varies with time remaining to maturity τ = T − t.
This constraint turns a two-dimensional local volatility function into a
one-dimensional curve
This simplification is perfectly reasonable for non-seasonal
commodities, such as crude oil. However, it is not applicable to the
market for natural gas and refined products, where volatilities during
winter and summer months are markedly different. In the next chapter,
we will make some adjustments to incorporate the more complex case
of non-homogeneous volatility with respect to time to maturity. For
now, we also let the option expiration be the same as the futures
expiration T, an assumption that will also be relaxed shortly. Under
these assumptions, the mathematics simplifies considerably, allowing
us to identify some essential properties of time-dependent volatility
more clearly.
For simplicity, we let the current time to be equal to zero. Then the
implied volatility in (11.4) is written as the quadratic mean of the local
volatility, which is a function of time remaining to maturity:

(11.5)

To illustrate the important analytical properties implied by this


relationship with a specific example, we use the following frequently
used exponential parametrization of the local volatility function
(11.6)
Here, σ∞ is the asymptotic volatility of a hypothetical futures
contract with infinite maturity. This long-term volatility can be
associated with macroeconomic and structural factors that affect the
entire futures curve. The short-term volatility σ0 represents additional
volatility driven by fundamental factors of supply and demand. The
sum of the long-term and short-term volatilities represents the
volatility of a hypothetical futures contract with instantaneous oil
delivery. The closer the futures contract is to expiration, the larger the
contribution of the short-term volatility to the total volatility. The
volatility decay factor k measures how long it takes for the short-term
fundamental uncertainty to dissipate. Under normal market conditions,
the volatility function (11.6) is consistent with the Samuelson effect,
illustrated in Fig. 8.​8.
Implied volatility v(T) for each futures contract that corresponds to
the local volatility σG(τ) specification (11.6) is then easily calculated by
straightforward integration of (11.5), as follows:

Figure 11.2 illustrates the relationship between σG(τ) and v(T) for
options expiring up to two years forward.

Fig. 11.2 An example of implied and exponentially decreasing local volatility with σ∞ = 0.1,
σ0 = 0.2, k = 1

For each maturity T, the implied volatility represents the quadratic


mean of the local volatility over the lifespan of the option. For example,
for a two-year option, implied volatility v(2) = 0.194, which is
calculated by averaging the local volatility squared along its path from
σG(2) = 0.127 until the option expiration, when the local volatility rises
to σG(0) = 0.30. Likewise, v(1) = 0.231 is the quadratic average of the
local volatility along its path from σG(1) = 0.174 to the volatility at
expiration σG(0) = 0.30. Since the implied volatility represents a certain
average of the local volatility, it is smoother, and its slope is flatter than
the slope of the local volatility.
The reduced slope of the implied volatility term structure relative to
the local volatility is characteristic of a cumulative quantity, being an
attenuated version of the local building blocks that it is constructed
with. The same smoothing effect can be illustrated with a trivial
example from the interest rate market. The two-year interest rate is
effectively the average of two consecutive one-year forward interest
rates. If such one-year forward rates are 2% and 3%, respectively, then
their average, or the two-year rate, is 2.5%. If the second one-year
forward rate suddenly jumps from 3% to 4% while the first one
remains unchanged, then the two-year average rate only moves by half
as much, from 2.5% to 3%. Any move in a local quantity is smoothed by
averaging. In the case of implied volatility, such smoothing is done by
integrating the local variance.
The methodology for elimination of time-dependent volatility,
presented in Appendix D for the conventional lognormal assumption,
applies identically to time-dependent normal volatility. In the normal
case, the Bachelier formula still applies with its volatility input
calculated as the quadratic mean of the arithmetic local volatility

(11.7)

Having established this important linkage between implied and


local volatilities, we now proceed to the problem of calculating the
volatility discount that results from price averaging, which is crucial in
the valuation of APOs.

11.3 Volatility Discount from Price Averaging


In financial markets, APOs are considered to be exotic derivatives that
trade rather infrequently. In contrast, in the oil market an APO is the
primary derivative instrument used by corporate hedgers to manage
ratable exposure to prices. The impact of large APO deals conducted in
the OTC market is an important driver of volatility and skews. These
APOs are hedged using vanilla options with any arbitrage opportunities
between the two instruments quickly captured by market-makers.
However, these trading opportunities are not trivial as pricing and
hedging of APOs is model-dependent.
The challenge of handling APOs in the traditional lognormal
framework comes from the fact that the average of lognormal variables
is not lognormal. Conventional pricing approaches tend to focus on
approximation methods for the sum of lognormal variables and various
numerical techniques.5 They usually do not cover the term structure of
the local volatility, assuming it to be constant. In contrast, in the oil
market the challenge is rather different. Since the underlying price
distribution is much closer to being normal than lognormal, one can
bypass the analytical complexity by pricing them in the normal world,
where the average of normally distributed variables is also normal.
Instead, in the oil market it is much more essential to properly capture
the time-dependent nature of volatility, which also leads to some
important, but often overlooked model modifications.
What makes APOs unusual and somewhat cumbersome to price is
their path-dependent nature, as the value of an APO depends on the
history of prices during the averaging period. Let (Ta, T) denote the
period over which the futures prices are being averaged. Then once an
APO moves inside the averaging period, i.e., when t ≥ Ta then one needs
to track the history of prices that have already contributed to the
running average. To capture the running average of futures, we
introduce a new variable

For simplicity, we assume that the price averaging is applied only to


a single futures contract. The extension to a more general case where
the average is calculated over the rolling prompt contract conceptually
follows the same argument.
The payoff of an average price call option that expires at time T is
determined by the terminal value of the running average at time T:
(11.8)
When an APO moves inside the averaging period, i.e., t ≥ Ta, then the
problem becomes more complex as the option price depends not only
on the futures price F(t, T) but also on the running average of
accumulated prices A(t). At a first glance, one might be tempted to
discard this challenge since APOs usually trade before the pricing
period even starts and one may not even care about pricing it within
the averaging period. However, as we have seen in the previous section,
the price of an option depends on the volatility during the entire
lifespan of the option, which includes periods both before and after the
averaging starts. Thus, volatility reduction that occurs only during the
averaging period still affects the option price at all prior times.
Before the APO moves into the pricing period, i.e., when t < Ta, the
standard delta-hedging argument of Chap. 8 applies and the APO price
CAPO(F, t) satisfies the regular diffusion Eq. (8.​8):

(11.9)

However, to solve this equation for t < Ta one needs a boundary


condition at time t = Ta. This boundary condition can only be obtained
by first solving the problem for the subsequent period when Ta < t < T,
for which the boundary condition is given by the option’s terminal
payoff (11.8) at time t = T. Note that for brevity, we write local volatility
as σ(F, t) omitting the reference to T which is fixed in this problem.
The pricing problem within the averaging period is multi-
dimensional, and, thus, it is much more complicated. In Appendix E we
show that an APO price must generally solve a partial differential
equation with two spatial variables F and A. Fortunately, the specificity
of this problem allows the dimensionality to be reduced, resulting in
the following pricing equation for an APO

(11.10)

This equation is written with respect to an auxiliary spatial variable


which represents the time-weighted average of the accumulated
average A(t) and the futures price F(t, T) for the remainder of the
pricing period.
Note that at the beginning of the averaging period x(Ta) = F(Ta, T).
As the option moves through the pricing period, the running average
accumulates and weights in the calculation of x(t) gradually shift from
F(t, T) to A(t). At the end of the period, the entire average becomes
known and x(T) = A(T). Therefore, the boundary condition (11.8) for
the option’s payoff can also be written in term of x(T)

The second important feature of the Eq. (11.10) is the presence of


the multiplier that applies to the local volatility function. This
volatility adjustment is intuitive. As time to maturity of the option
shrinks and a larger portion of the average becomes known, the
remaining uncertainty decreases accordingly. At the time when the
option expires and the calculation of the entire average is completed,
the local volatility of an APO reduces to zero. Such volatility reduction
near the expiration of the option is very desirable for the volatility
dealer. The smoothing dilutes the high uncertainty associated with the
gamma risk at the expiration. In contrast to a vanilla option, an APO has
very little risk left when t → T, as by then, most of the average is already
determined.
While the Eq. (11.10) considerably simplifies the problem, it still
cannot be easily solved because the local volatility σ(F, t) in (11.10) for
the general diffusion process depends on F, while the equation itself is
written with respect to x. However, this equation can be solved for the
special case of the ABM process for which the volatility (11.2) does not
depend on futures. Applying the argument of the previous section, the
solution for the ABM process is given by the Bachelier formula, where
the normal APO volatility is calculated as the quadratic mean of its local
volatility, as in (11.7):

(11.11)
In (11.11), we omitted the subscript that corresponds to the normal
volatility as we will see shortly that the same formula can also be used
for lognormal volatility.
The local volatility of an APO is the same as local volatility of futures
outside of the pricing period, but it follows from (11.10) that within the
pricing period it is adjusted by the linear multiplier

(11.12)

Since the integration of the local variance must cover the entire
lifespan of the option, one needs to separate two periods when the
option is within and outside of the pricing period. If an APO is already
pricing out, i.e., Ta ≤ t ≤ T, then only a single integral is calculated from
current time t until maturity T. However, if the option is yet to start
pricing and t < Ta, then the integration period (t, T) must be split into
two sub-periods (t, Ta) and (Ta, T) as the local volatilities in the two
periods are different.
In other words, (11.11 and 11.12) can be combined as

(11.13)

What made this transformation possible for the ABM process is the
fact that the average of normally distributed variables is itself normal. If
F(t, T) is normal, then A(t) and x(t) are also normal, and the volatility
coefficient in (11.10) does not depend on any spatial variable. This
property does not hold for the GBM process or other diffusions, and for
the general volatility function a closed-form expression for the price of
an APO is not possible. However, the same volatility transformation
(11.13) can be used as a high-quality analytic approximation for other
diffusion processes.
By using this approximation for other diffusions, one effectively
assumes that the local volatility function behaves similarly with respect
to x and F, i.e., σ(x, t) ≈ σ(F, t), in which case (11.10) turns into a regular
pricing equation with respect to x. For example, under the lognormal
assumption it assumes that the average of lognormal variables is
approximately lognormal. From the practical standpoint, this is a
relatively benign assumption because the real oil price distribution, as
we have seen before, is closer to normal than to lognormal anyway, and
in the normal case the formula holds exactly. It means that one can still
use the Black formula to price APOs with the volatility given by (11.13),
which is now understood in percentage terms.
Let us illustrate this first for the constant local volatility

which can be understood either as normal or lognormal volatility. In


this case, the integration of (11.13) is straightforward. After some
simple algebra, we obtain the following formula that relates the
volatility of an APO to the implied volatility of a vanilla option:

(11.14)

The nature of the volatility reduction resulting from the price


averaging becomes very transparent. One particularly useful rule of
thumb can be obtained for pricing an APO precisely at the beginning of
the averaging period. When t = Ta, the formula (11.14) reduces to

It follows that when the averaging is about to start, the implied


volatility of an APO should be approximately 60% of the implied
volatility of a vanilla option with the same maturity. For the normal
Bachelier volatility, this rule of thumb is an exact analytical formula, but
for the lognormal Black model it represents an accurate approximation.
Figure 11.3 illustrates an APO volatility discount graphically for a
two-year option, where price averaging occurs over the second year
with constant Black volatility σ(t) = v = 0.30. Since the local volatility is
constant, the vanilla option must be priced with the same constant
volatility 0.30 regardless of time remaining to maturity. However, the
local volatility for an APO, marked by the dashed orange line, starts
decreasing when the option moves into the pricing period and time to
maturity falls under one year. Integrating the local APO variance over
two years, including the second year where it is linearly reduced, we
obtain that APO pricing volatility at the beginning of the averaging
period vAPO(1) is only 0.177. This is consistent with the 60% rule of
thumb derived above. Likewise, the APO volatility with two years to
expiration, vAPO(2) = 0.244 is also discounted relative to the Black
volatility, but the discount is smaller because outside of the averaging
period the local volatilities for the two options coincide.

Fig. 11.3 Volatility reduction for an APO with T − Ta = 1 and constant local volatility
The volatility discount for an APO is not too sensitive to the choice
of normality or lognormality, but it varies substantially with the
steepness of the volatility term structure in either framework. To
illustrate this, let us modify the example above. Instead of using
constant volatility, we calculate the APO volatility discount for an
exponentially decreasing local volatility specified in (11.6) with the
same set of parameters that we used above. The corresponding
volatility behavior is shown in Fig. 11.4.
Fig. 11.4 Volatility reduction for an APO with T − Ta = 1 and exponential decreasing local
volatility
The local volatilities for a vanilla option and for an APO are
represented by the corresponding dashed lines, which coincide prior to
the pricing period. The solid lines are implied volatilities which are the
quadratic means of the corresponding local volatilities for the same
time to maturity. As before, within the pricing period, a volatility
multiplier that linearly decreases with time to maturity applies to an
APO. For an exponential volatility function, however, volatility
attenuation is stronger as compared to the case of constant volatility, as
the multiplier dampens local volatility when it is at its highest near the
option expiration. For example, at the start of the averaging period APO
volatility vAPO(1) is approximately 50% of the corresponding vanilla
volatility in contrast to 60% for the constant volatility assumption.
Likewise, the vAPO(2) discount relative to the volatility of vanilla option
is also larger in comparison to the case of constant volatility. The
steeper the term structure of the local volatility curve, the larger the
discount for APOs relative to vanilla options across all maturities.
If the option dealer can capture even a fraction of the APO price
discrepancy driven by a more accurate modeling of the volatility term
structure, it will definitely lead to a stellar career in the oil market.
Since many less sophisticated market participants use off-the-shelf
software packages for APO pricing that are often constrained to
constant volatility specification, they could easily misprice an APO. The
APO volatility discount is sensitive to the slope of the volatility term
structure, which in turn, is linked to the speed of price mean-reversion.
The impact of the volatility term structure becomes even more
significant for another important type of OTC options that expire prior
to the expiration of the underlying futures.

11.4 Early Expiry Options and Swaptions


In this section, we describe another important feature often
encountered in OTC oil contracts, which brings additional flexibility to
an option buyer by decoupling the option expiration from the
expiration of the underlying asset. In contrast, expiration of a vanilla
option is always contractually tied to a futures contract that expires
during the same month. For example, one can only trade June options
written on June futures both expiring in May, or December options
written on December futures expiring in November, but an option on
December futures that expires in May is not available on exchanges. The
OTC market provides this flexibility, where an option can be chosen to
expire on any business day, regardless of the expiration schedule for the
underlying futures. Such an option in the oil market is known as an
early expiry option (EEO).
The choice of the option expiration date is often driven by the
timing of certain events that an end-user or a speculator is exposed to.
For example, an oil producer may be looking to hedge the long-term oil
price risk only during the negotiation process to acquire another oil
company. Such a producer or investor would benefit from buying some
protection on long-dated oil prices that impact the value of the target
company, but the hedge may only be needed for the duration of a few
months while negotiations are taken place. Likewise, a speculator may
buy a short-term option on long-term futures specifically to bet on the
outcome of a certain event, such as a possible new regulation that is
expected to affect forward prices more than spot prices.
Buying a short-term option on long-term futures is significantly
cheaper than buying a vanilla long-term option on the same long-term
futures because the former has a shorter lifespan. Furthermore, buying
a short-term option on long-term futures is also cheaper than buying a
vanilla short-term option on short-term futures. Here, the lifespans of
the two options are the same, but because the volatility of long-term
futures is lower than the volatility of short-term futures, an EEO is
worth less. For option pricing, volatility is always determined by
integrating the local variance of the underlying futures over the lifespan
of the option. For an EEO, this integration only covers a period when
the futures are still relatively far from their expiration, and, therefore,
they are less volatile during that period. As a result, the integration
chops off the most volatile segment of the local volatility, which makes
an EEO cheaper.
To quantify this effect, we show in Appendix D that the implied
Black volatility of a European option that expires at some time T0 < T is
given by the quadratic mean of the local volatility during the period (t,
T0), which is now shorter than the time remaining to the expiration of
the futures:

For clarity, we explicitly show the expirations of both the option and
the underlying futures as two arguments of the volatility function, and,
as before, we omit the subscript that differentiates between lognormal
and normal cases. For a vanilla option, where the option expiration is
tied to the expiration of futures, one can assume that T0 ≈ T and,
therefore,

Compared to an EEO, where we integrate over (t, T0), for a vanilla


option the same local variance is integrated over a longer period (t, T)
which includes a particularly volatile period near the expiration of the
futures. We illustrate the difference between the two volatilities using
the same example of an exponentially decreasing local volatility
function specified in (11.6) for a two-year option that expires in one
year.
The local and implied volatilities shown in Fig. 11.5 are the same as
in Fig. 11.2. The rightmost points of the graph correspond to the
current time, which is two years until the maturity of the vanilla option.
As before, the implied volatility of a vanilla two-year option is
v(2) = 0.194. The implied volatility of an EEO, however, is only vEEO(1,
2) = 0.148. It is calculated as the quadratic average of the local volatility
along its path going from right to left, starting from the current time,
when σG(2) = 0.127, and ending in one year when σG(1) = 0.174. The
orange line shows that the implied volatility of an EEO increases only
modestly as the option approaches its expiration. It does not exceed the
maximum local volatility σG(1) during the lifespan of an EEO. The
higher local volatility of the same futures that occurs later near the
expiration of the futures contract does not contribute to the volatility
and the price of an EEO.

Fig. 11.5 Volatility reduction for an EEO with T0 − t = 1 and exponential decreasing local
volatility
One can see some similarities between EEOs and APOs. For an APO,
the local volatility is dampened by a linear multiplier once the option
moves into the pricing period. For an EEO, the local volatility for t > T0
vanishes entirely, as the option no longer exists then. However, direct
comparison between the prices of these two types of OTC options is not
straightforward, as prices depend on the aggregate effect of the
different times to maturity and the slope of the volatility term
structure.
While EEOs do trade from time to time in the OTC market, they
primarily function as a building block for a more commonly traded OTC
derivative, known as a swaption, or an option on a swap. The standard
oil swap settles based on the calendar month average of prices for the
prompt futures contract. Like an APO, an oil swap typically trades in the
form of quarterly and annual strips with monthly settlements. This
means that the swap price can effectively be represented as the linear
combination of the M futures that make up the swap, which are also
referred to as swaplets:

(11.15)

The weights ωi on the futures are determined by the count of days


that correspond to each futures contract being a prompt contract and
by discounting factors. If we again ignore, for simplicity, the effect of
discounting and assume zero interest rate, then the futures weights
must add up to one6

A swaption is a European option that at time T0 can be exercised


into the strip of monthly swaps with expiries T1 < … < TM at the
predetermined strike K. Therefore, at the time of its expiration T0 that
occurs prior to T1, the call swaption payoff is given by

Unlike vanilla options or APOs that are settled financially, an oil


swaption gives its owner the right to enter into a swap at the
contractually specified fixed strike price. This right, of course, is
exercised only if the swaption is in-the-money at its expiration. The
swaption can be viewed as an EEO whose underlying contract is not a
single futures contract, but rather a basket of futures that comprises
the swap. To price a swaption, one still needs to integrate and average
the local variance of each swaplet over the lifespan of the swaption (t,
T0).
Since a swap is just a linear combination of futures, pricing
swaptions in the lognormal setting faces a similar challenge to pricing
an APO, namely the fact that the weighted average of lognormal
variables is not lognormal. The problem again simplifies if it is first
solved for normally distributed variables, whose variances are additive,
and then extended to the lognormal framework in the form of an
approximation. Let us assume that futures for all expirations Ti are
normally distributed and driven by the same single source of
uncertainty dz with corresponding local volatilities σA(t, Ti). In other
words, in a risk-neutral world, where the drift term is replaced with
zero, the dynamics of futures is

Then the swap, which is the linear combination of futures, also


follows the normal process driven by the same dz:

Therefore, the swaption in the normal setting can be priced using


the Bachelier formula. The implied normal volatility of the swaption
vSW is calculated as the quadratic average of swaplet volatilities during
the lifespan of the swaption (t, T0):

(11.16)

Here, we again omitted the subscript that corresponds to the


normal volatility, because the same formula can be used for the
lognormal volatility as well. As in the previous example of an APO, the
formula (11.16) is exact for normal variables, the average of which is
also normal, but it produces an accurate approximation for the
conventional lognormal volatility. In other words, if

then we assume that the swap, which is a linear combination of


lognormal variables, is approximately lognormal, and, therefore, it
follows the following stochastic dynamics:

The swaption can then be priced using the conventional Black


formula with its lognormal volatility given by (11.16).
For example, for the exponential local volatility of the form (11.6),
the corresponding Black swaption volatility is given by

which can be evaluated explicitly. Similarly to the case of an EEO, the


integration of the local swaption variance does not extend to more
volatile periods when futures that make up the swap approach their
corresponding expirations. Figure 11.6 illustrates this with an example
of a three-month option written on a strip of six monthly swaps,
sometimes referred to as a 3x6 swaption.
Fig. 11.6 Local volatilities for a 3x6 swaption. The shaded area represents the area over which
local variances must be integrated
The shaded area in Fig. 11.6 represents the swaption’s three-month
lifespan over which the integration of the local variances of the swap
constituents is calculated. The most volatile segments for each
component near the expiration of the corresponding futures fall outside
of this integration area. This results in a much lower implied volatility
for the swaption compared to the corresponding vanilla option with the
same expiration. We will return to this example in the next chapter on
model calibration. In the remainder of this chapter, we provide a brief
introduction to multi-factor models, which are more accurate for
pricing swaptions, many other complex OTC derivatives, and structured
deals that depend on multiple futures contracts.

11.5 Multi-Factor Models and Other Exotics


So far, we have only looked at so-called one-factor models, where all
futures are driven by the same single source of uncertainty dz. Such
models are generally sufficient for options written on a single futures
contract. However, for options that depend on multiple futures
contracts, such as swaptions, one-factor models are too restrictive. One-
factor models implicitly assume that futures across all maturities are
perfectly correlated. While futures with different expirations can move
by different amounts as determined by their corresponding local
volatilities, in one-factor models all futures can only move in the same
direction. The direction is always determined by the sign of their
common random component dz. This behavior is problematic for
swaptions and other options that depend on multiple futures. It is easy
to imagine a scenario where some futures that make up the swap move
in opposite directions in a way that keeps their linear combination or
the swap price unchanged. To allow futures to decorrelate, one must
introduce a second source of uncertainty.
Fortunately, the volatility algebra for many OTC options remains
essentially identical if we model the behavior of futures with a multi-
factor model. To keep the presentation simple, we only illustrate this
with the industry standard base case of a two-factor lognormal model
with exponentially decreasing volatility. The two-factor volatility
specification is analogous to (11.6). The only difference is that short-
term and long-term futures are now driven by two different but
correlated sources of uncertainty dz1 and dz2 with correlation
coefficient ρ:

(11.17)

The first factor is a hypothetical contract with an infinite maturity,


whose volatility σ∞ is assumed to be constant. The second factor
captures the additional volatility associated with short-term
fundamental uncertainty that dissipates exponentially with time. Since
the variance of the sum of two normal variables is equal to the sum of
the variances plus the covariance between them, the total lognormal
volatility can be written as

The pricing of a swaption in this two-factor lognormal framework is


a straightforward extension of the previous argument. Assuming again
that the swap itself is approximately lognormal, then
The usual Black formula applies, with the swaption volatility given
by the quadratic average of the corresponding local volatilities over the
lifespan of the swaption

Importantly, when ρ < 1, the introduction of the second factor


reduces the overall implied volatility of the swaption in comparison to
its volatility in a one-factor model. The lower the correlation between
the factors, the lower the value of the swaption, as the likelihood of
some swaplets moving in opposite directions increases, which dampens
the volatility of the swap. If the correlation ρ = 1, then the two-factor
model reduces to the one-factor model with the local volatility specified
in (11.6).
Occasionally, two-factor lognormal models are further extended into
N-factor models of the form:

These more complex specifications are more useful for modeling


natural gas and power prices. The purpose of the additional factors and
their volatility loadings σi(t, T) is to capture strong seasonal effects, as
natural gas volatility in the winter is substantially higher than it is in
the summer. In the power market, the usage of multiple exponential
decay factors ki also allows the model to separate the time scales of
short-term and long-term mean-reversions, which are typically driven
by different forces. Since the efficient storage of power is much more
difficult, at least at the time of writing this book, the power market is
characterized by large spikes with an extremely fast mean-reversion,
which requires a large parameter ki. At the same time, for long-term
natural gas and power contracts, price mean-reversion is driven by
structural factors that evolve much more slowly. Using two exponential
factors with different volatility decay parameters captures the
dynamics of such markets better.7
Multi-factor models are also used for pricing more complex exotic
options and structured derivatives in the oil market. In a structured
transaction, the dealer often attempts to embed some optionality and
present it in a less transparent, but optically attractive way for the end-
user. This is, of course, done to maximize the dealer’s expected profit
margins. For example, swaptions are often embedded in so-called
extendable swaps that are marketed to producers as a way to achieve
higher swap prices for hedges.
Consider a producer who wants to hedge an oil price for one year
forward but hesitates whether to hedge or not for the second year. If
the producer chooses to hedge for both years, then ignoring again the
discounting factor, the hedge price will be the average of two one-year
swap prices. Alternatively, the producer can commit to hedging only for
the first year and instead boosting the swap price by giving the dealer a
unilateral right rather than an obligation to extend the one-year swap
agreement for the second year. The producer, therefore, sells a call
swaption on the second-year swap that expires at the end of the first
year. The premium for selling a call swaption is not paid in cash, but
instead it is used to increase the swap price. In this case, the producer
improves the hedging price but loses the assurance that the second
year risk will be hedged at all. If the second-year price drops by the end
of the first year, the dealer will likely waive the option to extend the
deal, but at least, the producer receives the higher price in the first year.
If the oil price rallies by the end of the first year, then the producer
obtains higher price for both years, as compared to the average two-
year swap price at the time when the deal was struck.
The flexibility and the complexity of OTC options is practically
limitless. Many such options depend on the term structure of futures
and volatilities. Extendable options can be quoted with multiple nested
extension periods, where the decision for one period may depend on
the decision made in the previous period, which makes pricing
incredibly complex. Further variations could include expandable
options, where instead of owning the right to extend the deal in time,
the dealer acquires an option to unilaterally expand the deal or, in other
words, increase its original volume. For the most part, the hidden
objective of the dealer is to buy a relatively cheap optionality from an
end-user without making its fair value too apparent. The additional
value brought up by such optionality is often used to bridge the gap in
the negotiation of the hedging price.
An important component of pricing optionality embedded in
complex OTC deals is modeling the creditworthiness of a seller. The
higher the leverage in the transaction, the larger the risk of credit
default for the dealer. In fact, counterparties with lower credit and
limited free cash flows tend to have stronger motivation to sell
optionality to pay for the hedge. The value of the collateral often comes
into the valuation of such derivatives deals. To limit the credit risk,
dealers sometimes introduce caps on the maximum payoff of the deal.
One derivative structure with such a capped payout, called a target
redemption swap, was particularly popular during the global financial
crisis. The structuring possibilities in oil derivatives are endless, but
many of these complex deals heavily rely on the two key modeling
pillars described in this chapter, the impact on volatility from price
averaging and from an early expiration. Both effects are captured by
certain tricks applied to the local volatility function.
Without any doubt, one can build more sophisticated models to
price more complex deals. Analytic valuation for such deals is rarely
possible. The standard solution is to price them numerically using a
Monte Carlo simulation for the underlying stochastic process. The
question is what process to simulate. The results of such simulations
are highly sensitive to the chosen specification of the local volatility of
the stochastic process. Since all OTC options are ultimately hedged with
vanilla options, the simulated process for futures must also generate
prices for vanilla options that are consistent with observable market
prices. For the professional volatility trader, this process of reconciling
theoretical models with market prices is known as model calibration. It
is the core engine that drives OTC arbitrage, and we devote the next
chapter to this important topic.

References
Blas, J. (2017, April 4). Uncovering the secret history of Wall Street’s largest oil trade,
Bloomberg.
Bouchouev, I. (2000, July). Demystifying Asian options. Energy and Power Risk Management, pp.
26–27.

Geman, H. (2005). Commodities and commodity derivatives: Modeling and pricing for
agriculturals, metals and energy. Wiley.

Giddens, P. H. (1947). Pennsylvania petroleum 1750–1872: A documentary history. Pennsylvania


Historical and Museum Commission.

Kemna, A. G. Z., & Vorst, A. C. F. (1990). A pricing method for options based on average asset
values. Journal of Banking and Finance, 14(1), 113–129.
[Crossref]

Levy, E. (1992). Pricing European average rate currency options. Journal of International Money
and Finance, 11(5), 474–491.
[Crossref]

Lipton, A. (2001). Mathematical methods for foreign exchange: A financial engineer’s approach.
World Scientific.
[Crossref][zbMATH]

Merton, R. C. (1973). Theory of rational option pricing. The Bell Journal of Economics and
Management Science, 4(1), 141–183.
[MathSciNet][Crossref][zbMATH]

Swindle, G. (2014). Valuation and risk management in energy markets. Cambridge University
Press.
[Crossref]

Turnbull, S. M., & Wakeman, L. M. (1991). A quick algorithm for pricing European average price
options. Journal of Financial and Quantitative Analysis, 26(3), 377–389.
[Crossref]

Wilmott, P., Dewynne, J., & Howison, S. (1993). Option pricing: Mathematical models and
computation. Oxford Financial Press.
[zbMATH]

Footnotes
1 It should be noted that oil exchanges did exist in Pennsylvania and New York as early as in
the 1870s, see Giddens (1947). However, all of them were closed within a decade under
monopolistic pressure from the Standard Oil Company.

2 For a colorful description of the history of this program, see Blas (2017).
3 Prior to 2019, the Maya formula was calculated as 40% WTS (West Texas Sour), 40% HSFO
(High-Sulfur Fuel Oil), 10% LLS (Louisiana Light Sweet), 10% Dated Brent, plus a K-factor set
by Mexico. To reflect the growth in US shale and changes in the sulfur specification for marine
fuel, in 2019 the formula changed to 65% WTI Houston, 35% ICE Brent, plus a K-factor.

4 Merton (1973).

5 There is a large body of literature on pricing APOs. Kemna and Vorst (1990) formulated the
partial differential equation for the price of an APO and found a closed-form solution for an
option on the geometric average of prices. Such a solution is possible because the geometric
average of lognormal variables is also lognormal. Subsequently, many authors used the idea of
geometric averages to develop approximations of the probability density function for the
arithmetic average of lognormal variables and corresponding approximations for APO prices,
such as the formulas of Turnbull and Wakeman (1991) and Levy (1992). For other APO pricing
methods, see also Wilmott et al. (1993), Lipton (2001), and Geman (2005). The method
described in this book was originally suggested in Bouchouev (2000).

6 For non-zero interest rates ri that correspond to times Ti, the swap value S is determined by
equating the present value of the future cash flows to zero: Solving it

for S, we obtain that S is given by (11.15) with .

7 For more details on the time-scale separation for pricing natural gas and power options, see
Swindle (2014).
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
I. Bouchouev, Virtual Barrels, Springer Texts in Business and Economics
https://doi.org/10.1007/978-3-031-36151-7_12

12. Volatility Arbitrage and Model


Calibration
Ilia Bouchouev1
(1) Pentathlon Investments, LLC, Westfield, NJ, USA

– The problem of option pricing is turned upside down. Instead of


making assumptions about the stochastic behavior of oil futures, this
behavior is reconstructed from market prices of oil options. This
inverse problem of model calibration is a blueprint for the volatility
arbitrage.
– The local volatility term structure for normal and lognormal
processes can be recovered from implied volatilities via a
bootstrapping algorithm. However, due to the incompleteness of the
oil option market with respect to maturities, this reconstruction is
uniquely defined only under additional assumptions on the local
volatility.
– In contrast, for a given maturity, the oil option market is relatively
complete with prices available for a wide range of strikes, allowing
recovery of the entire market-implied probability distribution.
Distortions in the shape of this distribution indicate potential
arbitrage opportunities.
– A more difficult problem is to reconstruct a diffusion process that
generates a given market-implied probability distribution. Simple
approximations can be used to extract the local volatility of the
diffusion directly from the implied volatility smile. The relationship
between the two functions is particularly intuitive for the quadratic
normal model.
12.1 The Inverse Problem of Option Pricing
A typical process of modeling system behavior starts with the
description of a phenomenon with certain characteristic variables. The
relationship between these variables is then established in the form of
mathematical equations. The equations are characterized by a set of
parameters that are assumed to be known and estimated from
empirical studies. If the equation describes an evolutionary process,
then for the solution to be uniquely defined, it must also be
supplemented with a boundary condition that defines the initial state
of the system. Such a boundary value problem is an example of a direct
problem. If the solution to the direct problem is unique and stable with
respect to small changes in the boundary condition or with respect to
parameters of the process, then the problem is well-posed. Most direct
problems, including the problem of solving the diffusion equation with
a given initial condition, are well-posed.
We have previously highlighted that the problem of option pricing is
mathematically analogous to the problem of heat dissipation within a
medium after an application of an initial point impulse. The diffusion
process is described by the partial differential equation of heat transfer.
The properties of the medium are characterized by its thermal
conductivity. If the medium is non-homogeneous then its conductivity
is modeled as a function of the spatial variable. The solution to the heat
equation for a given conductivity function describes the temperature
distribution across the medium over time.
The problem of heat dissipation is an example of a direct problem.
Such a problem can be written in the form y = f(x), where an input x is
transformed by a given model f into an output y. If small changes in x
produce small changes in y, then the problem is well-posed and stable.
The diffusion problem is indeed stable with respect to the initial
impulse and also with respect to properties of the medium.
Now consider the problem of heat transfer in a composite non-
homogeneous medium, whose properties, such as its thermal
conductivity, are not well understood. In fact, we would like to deduce
these properties by applying multiple heat impulses and measuring
how the medium responds to them. Mathematically, the process is still
described by the same equation, but the conductivity of the medium is
now an undetermined function. However, the solution itself, which is
the temperature distribution, can be measured for various boundary
conditions. The problem is turned upside down: by measuring the
output of the model, what can we say about the property of the
material, or its conductivity?
In such a formulation of the relationship defined by y = f(x), we now
observe both the input x and its measurable output y. However, the
properties of the engine itself f that is responsible for the
transformation are not fully specified. The question is whether it is
possible to observe enough input-output pairs (x, y), perhaps for
different initial impulses, to recover some information about an
unobservable engine f? This would be an example of an inverse problem.
Inverse problems constitute an important branch of applied
mathematics that has contributed to numerous scientific discoveries. In
many real-life problems, we cannot easily see what is inside the closed
surface, such as deep inside the earth or inside the human body. We can
only reconstruct some unobservable properties by sending various
signals and measuring the response of the system on the surface that
we can easily access. Inverse problems arise everywhere, from medical
tomography and the geophysics of oil exploration to military submarine
tracking and image recognition by machine learning algorithms. The
problem of volatility reconstruction is an example of such an inverse
problem in financial markets.
The analogy between the inverse conductivity problem in heat
transfer and the inverse problem of volatility reconstruction in financial
markets is striking. The spatial variable x that describes the physical
medium represents the range of possible futures prices in the market.
The conductivity of the medium is an unobservable local volatility of
the diffusion process for futures prices. The real-time variable is
replaced with time remaining to maturity, which turns the option’s
terminal payoff at expiration into an initial condition. The strike price
of a given option plays the role of an initial impulse sent by the option
to the market medium of futures. The price of an option is then the
reaction to this impulse, which reflects the temperature of the
underlying futures market. The response is easily measured, as the
market price of an option is provided by a broker. If the market gives us
enough option quotes for different strikes and times to expiration, can
we then reconstruct the entire engine, the diffusion process with its
local volatility that describes the underlying futures market?
Financial analysts are more accustomed to statistical models of the
form y = f(x), where an empirically estimated input x is plugged into an
econometric model f to produce a forecast for a variable y. In contrast,
volatility traders are not forecasters, they are arbitrageurs seeking to
profit from relative mispricing opportunities. They do not care as much
about the historical behavior of prices; for them the purpose of the
model is reversed. Volatility traders do not pretend to know what is
inside the body of the market, but they can observe how the body
reacts to various impulses by measuring its responses via option prices.
Both inputs x and outputs y in a stylistic representation y = f(x) are
observable, as they represent market prices of futures and options. The
goal of the volatility arbitrageur is to reconstruct some unknown
properties of the machine f that turns an observable x into an
observable y for many (x, y) pairs. We call this process of backing out
unknown properties of the futures market from market prices of traded
options the inverse problem of option pricing. More colloquially, it is
known as the problem of volatility calibration.
Why does the volatility trader need this calibration? The volatility
trader makes a living by arbitraging prices of options and hedging out
residual risks. Since potentially mispriced options are hedged with the
liquid ones, the model must reproduce the prices of liquid options
correctly. The residual risks are then reduced by delta hedging, and all
options within a portfolio must be delta hedged consistently using the
same calibrated model for the underlying futures prices. Recall that an
option can be dynamically replicated only if its delta is computed using
the correct diffusion process for futures. Even though nobody knows
with certainty what the correct process is, the one implied by the
market produces deltas which are likely to be reflective of the actual
changes in option prices, at least over short time periods.
In addition, lucrative arbitrage opportunities exist in trading exotic
options and hedging them with vanilla options. As we have seen in the
previous chapter, the world of OTC oil options is quite diverse and
complex. The pricing of these options is not transparent with closed-
form solutions rarely possible for many OTC derivatives products.
These deals are usually priced with Monte Carlo simulation, but the
trader needs to know what dynamics for futures prices to simulate. The
simulated process for futures must also produce prices for vanilla
options, which are consistent with market prices, as vanilla options are
used for hedging an exotic option. Once the stochastic process is
calibrated, many exotic options can be priced at once, using the same
Monte Carlo simulation.
Unfortunately, inverse problems, such as the problem of volatility
calibration, are generally much harder to solve. In contrast to direct
problems, many inverse problems tend to be unstable and ill-posed. The
ill-posedness of the problem means that a small change in the input
could lead to a large change in the solution. In our case the solution to
the inverse problem is the property of the medium that we are trying to
reconstruct. If the signal happens to be noisy, and many option prices
are, then it can only produce a blurry reconstructed image of the
market medium.
To solve an inverse problem, one usually needs to smooth, or
regularize, the problem. Regularization comes in different shapes and
forms, but the main objective of this technique is to find an optimal
trade-off between the accuracy of the reconstruction and stability with
respect to noisy data. By and large, finding an optimal balance between
accuracy and stability is the core premise behind many machine
learning algorithms. On the one hand, there are very flexible models
that can be tweaked to match almost any data set, but such models
suffer from extreme instability. On the other hand, one can build robust
and relatively rigid models which are only capable of producing a poor
resolution of the desired picture.
The same challenge applies to the inverse problem of volatility
calibration. If the model is chosen to be too complex, then its calibration
is inevitably more difficult, as many model parameters become less
stable with respect to small changes in input data. Finding a sweet spot
between model accuracy and stability is an art of option modeling. At
the one extreme would be the calibration of the one-parameter Black
model. The calculation of implied volatility is robust, but the model
description of the real world is rather poor, as the lognormal
assumption cannot capture the much richer dynamics of the oil market.
At the opposite extreme is the calibration of a multifactor model with
volatility smiles. Such a model has too much flexibility, and the solution
to the calibration problem is unlikely to be stable, as many
combinations of model parameters can produce similar results. An
optimally calibrated local volatility can then change drastically in
response to even small changes in observable option prices. This would
certainly create chaos in delta hedging of derivatives portfolios.
As much as mathematicians dream of developing a grand unified
model for pricing everything, such an approach does not work in
financial markets. The markets are driven by the behavior of humans
and not by the natural laws of any physical process. The oil market is
too complex and constantly evolving to be described by a single model.
The model and its calibration must always be tailored to the specific
nature of the trading strategy. For example, a calibration technique that
works for a smile arbitrageur is unlikely to be as useful to an OTC term
structure trader, and vice versa. To handle this complex task of volatility
calibration, we parse the problem into pieces and develop several
foundational building blocks upon which the proper calibration engine
can be built. The assembly part of this engine is left to the reader, as it
must be tailored to a particular trading opportunity.

12.2 Bootstrapping in Time


The first building block of the calibration process deals with the
relatively simple problem of reconstructing time-dependent local
volatility from the term structure of implied volatilities under simple
lognormal and normal distributional assumptions. This part of the
calibration process is based on the relationship between the two
functions developed in the previous chapter. We now invert this
relationship.
Let the dynamics of futures prices be described by the GBM process
with lognormal time-dependent volatility of the form (11.​1). Following
the assumption made in the previous chapter, we first consider the case
of local volatility that depends only on time remaining to maturity

To simplify the notation, we drop the subscript that corresponds to


the geometric volatility since the methodology presented in this section
is identical for geometric and arithmetic volatilities. The goal is to
recover the term structure of the local volatility σ(τ) from the term
structure of the market-implied volatility v(T). As before, if we ignore
the presence of the smile, then v(T) can be taken to represent the term
structure of the implied ATM volatilities.
For convenience, we restate the formula (11.​5), which represents
the implied volatility as the quadratic mean of the local volatility over
the lifespan of the option:

(12.1)

The calibration problem reduces to an inversion of (12.1) for the


local volatility. This can be done by differentiating (12.1) with respect
to the upper limit of integration, resulting in:

(12.2)

Since the local variance on the left-hand side of (12.2) cannot be


negative, it sets an important restriction on the shape of the implied
volatility function. While the term structure of implied volatilities is
expected to decrease with time remaining to maturity as short-term
fundamental uncertainty reduces with time, there is a limit on how fast
it can decrease.
Any violation of the restriction (12.2) could indicate possible
mispricing of options with certain maturities. However, such mispricing
turns into an arbitrage only if the assumption of the model holds,
specifically, the assumption that volatility depends solely on time
remaining to maturity. To distinguish this opportunity from
unconditionally riskless profits, we refer to any violation of this
boundary as a model arbitrage.
Let us illustrate the restriction on implied volatilities, stated by
(12.2) for the continuous-time case, with a more practical example of
discrete times to expiration. In the oil market, vanilla options are
available only with monthly expirations and the partial derivative with
respect to time in (12.2) must be discretized. In the discrete case, we
continue to use the same letters v(i) and σ(i) to denote implied and
local volatilities but i = 1, . . N now refers to the i-th nearby futures
contract and the volatility during the corresponding month.
The integral in (12.1) that represents implied volatility for N-th
nearby futures contract is then replaced with the sum of local monthly
variances over N prior months

As above, volatility is assumed to be time-homogeneous depending


only on time remaining to maturity.
One can also express the same implied variance v2(N) recursively, as
the time-weighted average of the implied variance for the previous
period v2(N − 1) and the local variance during the latest period σ2(N). It
is useful to write this recursive relationship explicitly, as follows:

In the final formula, the implied variance v2(N − 1) for the periods
N − 1 is combined with the local variance σ2(N) in the last period to
produce the variance v2(N) for N periods.
These recursive relationships can be inverted for the local variance
σ2(N), which is expressed as the weighted difference between two
implied variances for consecutive periods:
The last formula is the discrete version of (12.2). This simple
recursive algorithm to recover the local volatility term structure from
the implied volatility curve is often referred to as volatility
bootstrapping.
The local volatility can only be calculated if the implied volatility
does not decline too fast with increasing time to maturity. To illustrate
non-negativity restriction (12.2) with a numerical example, let us
assume that the implied volatilities observed in the market for the first
three monthly futures contracts are

Then corresponding local volatilities are computed as follows:

So far, the bootstrapping algorithm seems to be working fine, but let


us consider the scenario when the implied volatility for the fourth
month v(4) = 0.19. Then the local volatility during the fourth month
σ(4) cannot be computed because its variance becomes negative, which
is mathematically impossible. In this example, the restriction imposed
by (12.2) is violated. The implied volatility v(4) is too low relative to
v(3), and its term structure declines with a steeper slope than the
maximum allowed by the model. This violation suggests that something
is wrong either with the model or with the market.
If the trader believes in the validity of the model, then mispricing
could indicate an attractive trading opportunity. Since the implied
volatility curve is perceived to be unreasonably steep, the trader could
buy a relatively underpriced option on the fourth nearby futures
contract and sell a relatively expensive option on the third nearby
contract. If the volatility varies only with time remaining to maturity, as
specified by the model, then these two options share the same variance
during the first three months. However, the longer-dated option must
have an additional non-negative value that arises from the price
variance during the fourth month. The violation of the restriction (12.2)
implies that by selling a short-term option and buying a longer-term
option, the trader acquires the local volatility σ(4) during the fourth
month for free or even gets paid to own it.
Such a strategy would indeed be an arbitrage, but only if volatility is
time-homogeneous, changing only with time remaining to maturity. If it
does not, and if the slope of the implied volatility curve is impacted by
uncertainty specific to a particular futures contract, then the trader can
easily end up losing money by trying to capture this pseudo-arbitrage
opportunity. For example, the drop in implied volatility for the fourth
contract could represent a justifiable reduction of uncertainty after the
passage of a certain important event, such as an OPEC meeting, which
occurred during the previous period. In this case, there is no arbitrage,
as both implied and realized volatility will likely decrease during the
fourth period. To capture such a scenario, one needs a more flexible
modeling framework that incorporates a non-homogenous time-
dependency of the local volatility. In fact, the market-implied volatility
curve does not even have to monotonically decrease. It can easily take a
humped shape with the peak corresponding to the timing of the event
after which the market expects the uncertainty to decline.
To incorporate such events into the calibration process, we extend
the bootstrapping algorithm to a more general specification of the local
volatility σ(t, T) that now depends on both the real time t and the
contract expiration T. In the discretized setting, such a model is often
maintained in the form of a local volatility matrix, an example of which
is shown in Fig. 12.1.
Fig. 12.1 An example of a local volatility matrix
The second row above the matrix shows the implied Black
volatilities v(N) for the N-th maturity contract, which are observable in
the market. The implied volatility curve in this example is chosen to be
humped. The hump could reflect either some uncertainty associated
with a particular event or an abnormal risk premium in the specific
option contract caused by hedging imbalances.
A non-homogeneous local volatility in the discrete setting is
represented by an upper-diagonal matrix σ(i, N), where each cell
represents volatility measured i months forward for the contract, which
is the N-th nearby as of today, with i ≤ N. Today is defined by i = 1. Using
this notation, σ(i, i) is volatility of a rolling prompt contract measured i
months forward, σ(i, i + 1) is volatility of the second-nearby contract i
months forward, σ(i, i + 2) is volatility of the third-nearby contract i
months forward, etc. In other words, each row defines local volatility of
various maturity futures at a given forward time. Each column, on the
other hand, describes the evolution of local volatilities of the same
contract during its lifespan.
The matrix σ(i, N) of local volatilities must be constructed in such a
way that it is consistent with the implied volatilities v(N) observed in
the market. Recall that the implied volatility of each option is the
quadratic mean of the local volatilities of the same contract, which are
represented by the corresponding column in the matrix. Obviously,
there are infinitely many ways to fill in a two-dimensional local
volatility matrix and still ensure that the quadratic average of each
column matches the implied volatility specified in the second row of the
same column. We have some additional degrees of freedom at our
discretion. We can use them to impose a supplementary structure on
our matrix that acts as a form of regularization for this inverse problem.
One way to do this is to restrict the shape of the local volatility to a
certain functional form.
To illustrate, we use our standard example of an exponentially
decreasing volatility function, but introduce an extra degree of freedom
by letting the short-term volatility depend on time t

If T = t, then the futures contract represents the spot price with an


instantaneous delivery. Thus, σ0(t) can be understood as the excess of
total spot volatility over constant long-term volatility driven by short-
term fundamental uncertainty that now depends on time:

In the discrete setting of monthly local volatilities σ(i, N) this


assumption is equivalent to

Allowing spot volatility to be time-dependent is particularly helpful


for modeling seasonal commodities, such as natural gas and heating oil,
for which volatility during winter months is substantially higher. The
other two model parameters, the long-term asymptotic volatility σ∞
and the decay factor k, are kept fixed. As in the previous examples, we
assume that σ∞ = 0.10 and k = 1. This specification aligns the degrees of
freedom between the inputs and the outputs of the model and allows us
to construct a one-to-one mapping between implied and local
volatilities.
To extend the bootstrapping algorithm to a more general volatility
specification and to fill in a non-homogeneous time-dependent local
volatility matrix recursively in a unique way, we proceed as follows. The
local volatility for the first discrete period is the same as the implied
volatility for the first contract:
The first row of the matrix, which represents local volatilities for
various futures during the first month σ(1, N), is calculated using the
exponential volatility specification, as follows:

The second row starts with σ(2, 2), which is the volatility of the
contract that is second nearby today, but its volatility is measured
during the following month when this contract becomes the prompt.
Since the total two-month implied variance for the contract is the sum
of its variances during the first month and the second month,

We can then apply the bootstrapping technique and back out the
local volatility σ(2, 2) from the implied volatility v(2) and the local
volatility during the first month σ(1, 2) calculated in the previous step:

The knowledge of σ(2, 2) gives us the pivot spot volatility for the
second row, and we can fill in the rest of the second row for all N > 2,
using the specified exponential function, as follows

Moving on to the third row, the implied variance for the third
contract is the average of the local variances over three periods

Since we already know the local variance for the first two rows and
we also know the implied variance that contains the cumulative
information from all three periods, the local volatility during the third
month can be bootstrapped, as follows:

After obtaining the spot volatility during the third month σ(3, 3), we
again calculate σ(3, N) for all N > 3 using the exponential specification
for the local volatility. The process continues until the entire local
volatility matrix is filled in a unique way.
Having such a local volatility matrix in place is important for pricing
swaptions and other exotic options described in the previous chapter.
To illustrate its application to swaptions, we assume that the volatility
matrix in Fig. 12.1 is constructed for calendar months, rather than for
futures contracts, as one can easily convert between the two. To price a
swaption, the trader must calculate the quadratic mean of local swaplet
volatilities over the life of the swaption. For example, for a three-month
swaption into a six-month swap used in the previous chapter, the
volatility of the swaption spans the highlighted area in Fig. 12.1, which
is made up of eighteen discrete local volatility blocks. The quadratic
average of swaplet volatilities with maturities from four to nine months
over next three months produces the swaption volatility. In this
example, the swaption volatility is only 0.267, which is lower than any
of the implied volatilities for swaplets. It would be a major mistake by
the trader to price a swaption by calculating the quadratic mean of
implied volatilities for swap components. What goes into averaging are
the local variances of swaplets, not the implied ones.
In addition, if some swaption prices are observable in the broker
market, then this flexible format of the local volatility matrix also
allows its parameters to be further calibrated to the market prices of
these swaptions. For example, we still have parameters σ∞ and k that
remain entirely under our control. We can choose them to approximate
market prices of swaptions, EEOs, or some other exotic derivatives that
may be quoted in the market. This is typically done in an ad hoc manner
by trying out different combinations of σ∞ and k, while monitoring how
the entire local volatility matrix responds to different combination of
inputs. If more flexibility is desired, the decay parameter k(t) can also
be made time-dependent.1
A similar volatility matrix can also be built for the two-factor model
specification (11.​17). Its construction remains largely identical, except
for an additional degree of freedom introduced by the factor correlation
ρ. Using it, however, often causes more harm than it adds value as many
combinations of parameters may end up producing similar results. For
example, one can achieve a drop in the swaption volatility either by
lowering the factor correlation, or alternatively, by steepening the local
volatility function. The market, however, does not give us enough
observable information to distinguish between these two alternatives.
The problem becomes over-parametrized, and unless some model
parameters are fixed, it becomes prone to instability.
The inverse calibration problem in the time direction is relatively
simple analytically. However, this problem is inherently incomplete, as
there are too many degrees of freedom relative to the number of
available observations. The oil option market does not contain enough
information to allow us to deduce the rich dynamics of the volatility
time-dependency, so the term structure modeling is often
supplemented by econometric analysis of futures prices. In contrast, an
inverse problem in the strike dimension where the local volatility is
allowed to be space-dependent faces the opposite challenge. It is
relatively complete in terms of the availability of granular data for
option prices with different strikes, but it is more challenging from the
mathematical point of view. We turn to this problem next.

12.3 Market-Implied Probability Distribution


So far, we have only analyzed a rather simple inverse problem of
reverse engineering the local variance of a lognormal process from its
cumulative variance. What made the inversion possible is a special
property of normally distributed random variables whose variances are
additive. We only had to express the cumulative variance as the sum of
its constituent local variances, and then bootstrap the local variance
recursively by taking differences between cumulative variances for
options with nearby expirations. Such a bootstrapping requires only
ATM volatilities that are uniquely mapped to an implied volatility
parameter characterizing the width of the price distribution, which is
assumed to be either normal or lognormal. In general, the nature of the
price distribution is unknown. It turns out that by taking options prices
for all strikes, one can reconstruct the distribution itself. Such a
distribution is called a market-implied probability distribution.
Recall from (A.​10) that the price of a call option can be expressed in
terms of the probability density function and the option’s terminal
payoff as follows:
(12.3)

Here, we again ignore the impact of interest rate, assuming it to be


zero. The function p(F, t; FT, T) represents the probability of the futures
price reaching the level FT at time T, given that its price at time t is F.
This function is also the same as the fundamental solution to the
pricing Eq. (8.​8). As explained in Appendix A, the fundamental solution
solves the Eq. (8.​8) with the special boundary condition given by the
Dirac delta function.
Importantly, what we refer to as probability is not the real-world
probability, but rather it is the so-called risk-neutral probability. This is
because the real-world stochastic process (8.​5) has a non-zero drift
μ(F, t) and its corresponding probability density function satisfies (A.​8).
The pricing Eq. (8.​8), however, does not have any drift. Even though we
still refer to p(F, t; FT, T) as a probability, it should be distinguished from
the statistical probability of the historical price moves, which can only
be measured in the real-world. The Eq. (12.3) relates an option price to
the market-implied risk-neutral probabilities generated by an
imaginary stochastic process in which the real-world drift term is
replaced with zero. Much as the realized volatility is distinct from the
implied volatility, the historical distribution of prices is not the same as
the market-implied risk-neutral distribution.
To calculate the market-implied probability distribution, we need to
invert formula (12.3) for the function p(F, t; FT, T). Since the integration
in (12.3) is performed only over the range of FT for which the call
option payoff is non-zero, we can rewrite it as

We differentiate this formula with respect to K and obtain that

The first term comes from differentiation with respect to the lower
limit of integration. It is equal to the value of the integrand evaluated at
the point K, which is zero, and, therefore, this term vanishes.
We then take the second derivative with respect to K, differentiating
it again with respect to the lower limit of integration, which results in

(12.4)

This is a rather remarkable result. The given set of option prices can
be transformed into the set of market-implied risk-neutral probabilities
simply by differentiating the option prices twice with respect to the
strike price. This relationship was once again first discovered by
Bachelier in his seminal doctoral thesis.2
The transformation (12.4) is possible only because of the very
special terminal payoff of the call option, which is given by max(0,
FT − K). The first derivative of the payoff with respect to K taken with
the negative sign is given by the so-called digital, or the Heaviside,
function. This function is equal to either one or zero, depending on
whether FT is above or below K:

(12.5)

Furthermore, the second derivative of the payoff is equal to the


Dirac delta function centered at K

(12.6)

In other words, the second derivative of the terminal option payoff


is a point impulse function which represents the terminal state of the
probability density function p(F, T; K, T). As the impulse diffuses
backward in time for t < T, the probability density function p(F, t; K, T)
is described by the second derivative of the option price, as in (12.4).
For further background on the Dirac delta function, probabilities, and
related differential equations, we refer to Appendix A.
Let us illustrate the concept of market-implied probabilities in a
more practical discrete case. Since in the real market we only have
option prices for a finite set of strikes, the second derivative must be
discretized. In the previous section, when we applied the bootstrapping
technique to time-dependent volatility, we were able to extract some
information about the local quantity from the difference between two
closely related global quantities. We will do something similar here and
extract the local probability at each point by taking the difference
between option prices with nearby strikes, which contains some
information about the likelihood of futures being between two strike
prices.
Let us construct a call spread trade, where we purchase a call option
C(K) struck at K and sell a slightly further OTM call option C(K + dK)
struck at K + dK for some small increment dK. Both options have the
same expiration T, and, for brevity, we omit explicit references to other
option parameters. If we now take 1/dK units of these call spreads and
let dK be infinitesimally small, then this trading position represents the
first derivative of the option price with respect to the strike price taken
with the negative sign

The payoff of such call spreads are shown in Fig. 12.2 for decreasing
strike increments.
Fig. 12.2 The payoff of the call spread with strikes K and K + dK approaches the payoff of the
digital option when dK → 0
The maximum payoff of one call spread is dK. Therefore, the
maximum payoff of holding 1/dK units of such infinitesimally tight call
spreads is equal to one dollar. In the limit when dK → 0, the call spread
converges to a digital option D(FT − K), which is defined by (12.5). The
digital option pays one dollar if the futures price is above the strike K at
expiration and zero if the futures price is below the strike:

Digital options do trade in the OTC oil market. Volatility traders


typically manage the risk of selling a digital option by buying the
tightest possible call spreads to approximate the discontinuous payoff.
The first derivative of the option price is itself a popular financial
instrument. Many traders appreciate the simplicity of the digital payoff,
which is similar to a lottery ticket that pays a fixed amount in the event
that the futures price exceeds the strike price at expiration. The digital
payoff is also identical to the option’s delta, and prices of the digital
option under normal and lognormal assumptions are simply given by
the formulas for deltas provided in Appendix B.
We now consider the so-called butterfly spread, where we purchase
an equal quantity of calls struck at K − dK and K + dK and sell twice as
many calls struck at K. The butterfly can also be understood as the
spread of two call spreads, where one buys the C(K − dK) − C(K) call
spread and sells the C(K) − C(K + dK) call spread. If we scale the call
butterfly position by holding 1/(dK)2 units of this trade and make dK
again infinitesimally small, then this portfolio is equal to the second
derivative of the option price with respect to the strike K:

The terminal payoff of the butterfly call spread is shown in Fig. 12.3
for decreasing strike increments.
Fig. 12.3 The terminal payoff of 1/(dK)2 units of call butterflies with strikes K − dK, K, and
K + dK approaches the Dirac delta function when dK → 0
The butterfly trade is certain to generate a non-negative payoff
which is defined by the triangular region with a width of 2dK and peak
height dK. Therefore, if we hold 1/(dK)2 units of such butterflies, then
the area of the triangle is always equal to one, regardless of the size of
dK. If we let dK → 0, then the second derivative of the option price with
respect to K turns into the Dirac delta function, centered at K:

Note that since the area of the triangle in the butterfly payoff is
equal to one for any dK, the normalization property (A.​4) of the Dirac
delta function that integrates to one is also satisfied. The butterfly
payoff is the discrete analogue of (12.6).
Obviously, the second derivative of the option price does not
explicitly trade in the market as it would have an impossible infinite
payoff, but its approximations in the form of tight butterflies with small
dK are quite common even on the exchanges, where the smallest strike
increment is only fifty cents per barrel. The maximum payoff of the
infinitesimally tight butterfly occurs when FT = K, and traders use
butterflies to explicitly bet on the probability of the price reaching the
specific level at expiration. Since at expiration the payoff of the butterfly
approximates the Dirac delta function centered at K, the market prices
of various butterflies with various strikes make up the probability
density function for the futures price to be at the level FT = K at
expiration.
Market-implied probability distributions are very helpful for
volatility traders in spotting arbitrage opportunities in option prices
across different strikes. The probability density function is generally
expected to be smooth. Any visible kink in the density function
indicates that prices for some options with nearby strikes might be
inconsistent. It is rather difficult to identify such an inconsistency by
looking at option prices or at their implied volatilities, which are
cumulative quantities, because averaging tends to smooth the effects of
any mispricing. In contrast, for local quantities, such as probability
densities, there is nowhere to hide, and any distortion in its shape
exposes a pricing anomaly. Looking at the implied probability
distribution is like looking at option prices under the microscope.
The recovery of the market-implied distribution from option prices
is undoubtably useful in identifying potential trading opportunities.
However, to capture such opportunities, additional information is
needed. The market-implied distribution provides us only with a
snapshot of the price distribution taken at a given time, but it tells us
little about the evolution of prices between today and the option’s
maturity. Since the opportunity in trading options is captured by delta
hedging, which must be done throughout the life of the option, to
calculate the deltas, the trader needs to model the evolution of the
entire stochastic process that produces this distribution. This
information is contained in the local volatility function, which is also
needed for pricing and hedging OTC options.

12.4 Local Volatility Smile


The conditional price distribution contains much less information that
the stochastic process itself. In fact, many different stochastic processes
could produce the same distribution at a given time horizon. Consider,
for example, the trivial case of the normal distribution. Any time-
dependent local volatility with the same quadratic mean produces the
same implied volatility, and, therefore, the same terminal distribution.
Moreover, a mean-reverting process, such as the one used for modeling
oil inventories in Chap. 3, also has a normal probability density (A.​17)
which can alternatively be generated by a different ABM process
without mean-reversion. The problem of reconstructing the entire
diffusion process from a snapshot of the terminal distribution is ill-
posed as it admits multiple solutions. To create a one-to-one mapping
between the price distribution and the underlying stochastic process,
one must restrict the degrees of freedoms in the model specification for
the process.
Let us first consider a hypothetical case and assume that at present
time t when the futures price is F, all option prices C(F, t; K, T) are
available for the continuum of strikes K and maturities T. Then these
option prices are related to the local volatility function of the diffusion
process via the following equation:

(12.7)

This equation is known as the Dupire equation.3 In the Appendix F


we show how it can be easily derived by integrating the Fokker-Planck
equation for the probability density function.
While such an assumption about option prices may be unrealistic,
the Dupire equation plays an important role in solving inverse
calibration problems, much like the BSM equation does in solving direct
problems of option pricing. The structure of the Dupire equation
resembles the BSM Eq. (8.​8) which is written with respect to futures F
and current time t. However, in (12.7), the futures price F is replaced
with the strike price K, and real time t is replaced with option
expiration T. In addition, the direction of time is reversed. While the
BSM equation moves backward in real time for t < T starting from the
terminal boundary condition when t = T, the Dupire equation moves
forward with respect to maturities T > t with an initial boundary
condition at T = t.
The BSM and Dupire equations are complementary to each other.
They are somewhat analogous to the Kolmogorov backward and
forward (Fokker-Planck) Eqs. (A.​8) and (A.​6) for the probability density
function. However, the BSM equation holds for any derivative, but what
makes the existence of the dual Dupire equation possible is the unique
nature of the call option payoff, whose second derivative with respect to
the strike price happens to represent the probability density function,
as shown by (12.4). If the option had any other payoff profile, then a
complementary dual equation for option prices would not have been
possible.
The Dupire equation provides a stylish theoretical path towards
solving the inverse calibration problem. The local volatility function,
which generates theoretical option prices that match market prices
with all strikes and maturities, can be backed out from option prices as
follows:

(12.8)

Note that while the actual local volatility function depends on F and
t, when it is reconstructed from option prices, its functional arguments
are replaced with K and T.
To understand this representation better, let us revert to a discrete
approximation of partial derivatives of option prices with respect to
strikes and maturities. We know from the previous section that the
denominator of (12.8) is the probability density function, which can be
approximated by market prices of call butterflies. Similarly, the
numerator of (12.8) can be approximated by an infinitesimally tight call
spread with respect to T.
Consider a calendar call spread which consists of long and short call
options with the same strike K and corresponding expiration times
T + dT and T. If we hold 1/(dT) units of this calendar call spread and let
dT → 0, then such a trading position becomes the partial derivative of
the option price with respect to T:

Therefore, the local volatility at the point (K, T) can be


approximated by the square root of the ratio of 2/(dT) units of calendar
call spreads to 1/(dK)2 units of call butterflies. This relationship shows
how the local volatility function ties together option prices with nearby
strikes and maturities. In this discretized example, the local volatility at
each point is determined by a portfolio of four options, two from the
calendar spread and two more from the wings of the call butterfly. Since
these options also impact neighboring local volatility points, all option
prices across strikes and expirations are deeply intertwined via this not
so obvious relationship between their partial derivatives.
Unfortunately, direct application of the Dupire equation to the oil
market faces some limitations. While the assumption about availability
of options with the continuum of strikes is reasonable given that vanilla
options are listed with granular strike increments, a similar assumption
does not hold for options with different maturities. For a given futures
contract, vanilla oil options exist only for one maturity date in the same
month. It would be difficult to calculate the T-derivative in the Dupire
formula as options on the same contract with other maturities do not
exist. One would need to know prices for all EEOs on the same futures,
which are clearly not available. The only way to apply Dupire’s equation
to volatility calibration in the oil market is to estimate the T-derivative
from the market price of a calendar call spread with the same strike but
on different maturity futures. However, such an approach would
implicitly assume that all futures are identically distributed, which is
very precarious in the oil market where different maturity futures
reference different physical barrels.
In practice, the volatility calibration problem in the oil market must
be simplified. Instead of trying to recover two-dimensional local
volatility function σ(F, t) from option prices with all strikes K and all
maturities T, one usually fixes the maturity and focuses on recovering
time-independent local volatility σ(F) for a given futures contract using
option prices with various strikes. Unfortunately, for such a problem
the Dupire equation cannot be used as the partial derivative with
respect to T is not known.
It turns out that such an inverse problem has very unique and
important scientific applications. It is analogous to the problem of
reconstructing the conductivity of a non-homogenous medium from
temperature observations at a given time. It also represents the
problem of reconstructing a time-independent diffusion process from a
snapshot of the probability density function. Surprisingly, such a
problem has been solved only under additional assumptions, but in its
general case, at least to the best of the author’s knowledge, it presents a
rare example of an open mathematical problem. We formulate this
problem more precisely in Appendix F. It is fascinating how the
problem of option pricing that gave rise to an entire branch of modern
mathematics continues to push science to new frontiers. If any readers
choose to solve this inverse problem instead of taking advantage of
volatility arbitrage opportunities in the oil derivatives market, then the
author would consider the mission of this book to be largely
accomplished.
Despite its theoretical shortcomings, practitioners have no choice
but to come up with a way to back out the local volatility numerically.
Many computational techniques have been developed for recovering
σ(F) from option prices with fixed maturity, but nearly all of them suffer
from numerical instability. This instability is not a flaw of the
algorithms, but rather it is a salient property of this inverse problem.4
One way to simplify and tame an instable inverse problem is to
impose a certain parametric structure on the local volatility which acts
as its automatic stabilizer. Instead of trying to bootstrap the entire
function σ(F), one can assume that σ(F; θi) is defined by parameters θi
and then minimize the distance between theoretical C(σ(K; θi)) prices
that depend on θi and observable market prices C∗(K):

For example, in the QN volatility model of Chap. 10, parameters


characterize variance, skewness, and kurtosis of the underlying price
distribution. These parameters can be fitted to market price of options
with many different strikes. Alternatively, one can choose to use only
three liquid options benchmarks for ATM straddle, a costless collar and
a strangle and match them to three parameters of the QN model. In this
case, the least squares minimization is replaced with three nonlinear
equations with respect to three model parameters. In fact, many
traders prefer not to fit the model to all strikes, arguing that the three-
parameter model provides a proper balance by imposing enough
structure on the model while leaving the residuals from this parametric
fit as an indication of mispricing.
While parameter fitting approach is widely used by quants,
professional market-makers always look for some shortcuts that may
help them to gain some intuition into the cumbersome process of
volatility calibration. What the trader sees in the market is an implied
volatility smile v(K); what is needed for calibration is the local volatility
function of the diffusion process σ(F). Recall that for lognormal and
normal processes, implied and local volatilities are related via (12.1).
Since for these processes implied volatility is represented as some
average of instantaneous local volatility in the time direction, the
natural question is whether a similar relationship between implied and
local volatilities may also exist in the price direction?
Unfortunately, there are no simple analytic formulas that relate local
and cumulative variances for more complex price distributions.
Conceptually, the implied volatility still represents some average of the
local volatility, but this average is calculated along certain stochastic
paths taken by the futures price. Imagine pricing an option using a
Monte Carlo simulation, where many different paths for the futures
price are generated by the diffusion process with the calibrated local
volatility. The option value is then determined as the average payoff
across all simulated paths. Each path starts at the current futures price
F, but only certain paths that cross the strike price K contribute to the
option value because the payoff from the remaining ones is zero.
We can, therefore, internalize the implied volatility as an average of
the local volatility but compute the average only along such paths from
F to K that contribute non-zero values to the option price.5 However, in
the special case of the QN model we can derive a more elegant solution
for the relationship between implied and local volatilities. If we equate
the formula (10.​4) to option prices expressed in terms of implied
normal volatilities vN(K) and apply the Taylor formula, we obtain that

Therefore,

(12.9)
Furthermore, vQN(K, F) represents the average of the local volatility
between F and K, which is verified by direct integration:

Figure 12.4 illustrates the relationship between the quadratic local


volatility with respect to futures and the corresponding implied normal
and Black volatilities with respect to strike prices. Implied normal
volatility is also approximately quadratic. However, its curvature is only
one-third of the local volatility curvature, as implied volatility smooths
the perturbation of the local quantity via averaging.

Fig. 12.4 Implied volatilities generated by the local volatility σ(F) = 20 + 0.02 (F − 60)2

Another useful example is given by a linear local volatility function


for which the curvature c = 0. In this case, the implied volatility is also
approximately linear, but its slope is only one-half of the slope of the
local volatility. Again, the implied volatility is flatter than the
corresponding local volatility as the cumulative function smooths the
average of its local constituents.
Figure 12.4 also shows how option prices produced by the QN
model would be seen through the incorrect lenses of the lognormal
model. Equivalent Black volatilities that correspond to QN volatilities
can be calculated either numerically or by using the approximation
formula (10.​2). The QN model produces Black volatility skew which is
representative for the crude oil option market. The bottom of the skew,
which corresponds to slightly OTM calls, indicates the impact of
producer pressure that results from selling calls via two- and three-way
collars. The skew exhibits positive curvature on both ends, which is
consistent with fat-tailed distributions and the larger volatility risk
premium for OTM options.
The reconstruction of the local volatility and the corresponding
stochastic process for future prices forms the core of the pricing
framework that must be built by professional volatility traders. With
calibrated volatility, all hedging ratios are consistently calculated within
the same framework, and traders no longer need to guess whether the
smile is fixed or floating. The smile behavior is entirely determined by
the diffusion process with the calibrated local volatility function. In
addition, the local volatility allows option traders to price many exotic
derivatives of high complexity all at once. Once the local volatility is
reconstructed, complex options can be priced by using a Monte Carlo
simulation. Thousands of forward price paths are simulated by the
diffusion process with the reconstructed local volatility, and the value
of the complex derivative is computed as the derivative’s average
payout for such simulated paths.
The diffusion framework is, by far, our preferred choice for
modeling oil options. However, we should also recognize that no model
is perfect, and we briefly mention its shortcomings and alternatives
often proposed in the literature. One drawback of the diffusion
assumption is the difficulty of calibration to short-term maturity
options. The implied volatility smile is often very steep because of the
shortage of natural sellers for deep OTM options. The local volatility,
therefore, must be even steeper. To match observed market smiles, the
local volatility is then forced to rise with unrealistic slopes. This comes
from the continuous nature of the diffusion process, which does not
allow for jumps. Allowing local volatility to rise sharply for short-term
maturity options can also be viewed as an attempt to incorporate such
jumps. It should also be understood that since the short-term implied
smile contains a larger volatility risk premium, the corresponding local
volatility may not accurately represent the actual dynamics of the
futures.
Another popular alternative to the diffusion framework is allowing
volatility to be stochastic. Stochastic volatility processes can also be
calibrated to match the market observed smile. However, the forward
smile dynamics implied by such models is less intuitive than the one
generated by local volatility models. Many popular stochastic volatility
models produce inadequate deltas which resemble deltas from the
sticky moneyness heuristics. Even though for many other financial
markets, such as equities, stochastic volatility models are widely used,
as far as the oil market is concerned, diffusion seems to represent a
better choice.6
Perhaps most importantly, the diffusion pricing framework
guarantees the absence of arbitrage, as the risks of the option and the
futures are driven by the same source of uncertainty, which can be
hedged. This is not the case in jump and stochastic volatility models,
which introduce other non-tradable sources of risks that require more
assumptions and more unobservable parameters. Diffusion models are
complete in the sense of their ability to eliminate risk by hedging, while
more complex jump and stochastic volatility models are not, and for
that reason we chose not to cover them in this book. For the volatility
arbitrageur, the model completeness that allows for consistent pricing
and hedging is far more important than complex multi-factor
frameworks with unobservable parameters.
The inverse problem of volatility calibration is a non-trivial task. Its
implementation is often the primary assignment for quants on option
trading desks. Calibration techniques are typically customized to
individual markets and even to individual large deals, such as the
hedging program of the Government of Mexico. Going into more
specifics here would significantly distract us from the overall purpose
of this book. All traders, however, need to be aware of the importance of
model calibration, and reconstruction of the volatility engine from
observable prices of liquid options. The volatility calibration creates an
ultimate roadmap for identifying and capturing arbitrage opportunities
in the options market.
References
Alexander, C. (2008). Market risk analysis, Vol. III: Pricing, hedging and trading financial
instruments. Wiley.

Avellaneda, M., Friedman, C., Holmes, R., & Samperi, D. (1997). Calibrating volatility surfaces via
relative entropy minimization. Applied Mathematical Finance, 4(1), 37–64.
[Crossref][zbMATH]

Bachelier, L. (1900). Théorie de la Spéculation, Annales scientifiques de l’Ê cole Normale


Supêrieure, Serie 3 17, 21–86.

Bouchouev, I., & Isakov, V. (1997). The inverse problem of option pricing. Inverse Problems,
13(5), L11–L17.
[MathSciNet][Crossref][zbMATH]

Bouchouev, I., & Isakov, V. (1999). Uniqueness, stability and numerical methods for the inverse
problem that arises in financial markets. Inverse Problems, 15(3), R95–R116.
[MathSciNet][Crossref][zbMATH]

Bouchouev, I., Isakov, V., & Valdivia, N. (2002). Recovery of volatility coefficient by linearization.
Quantitative Finance, 2(4), 257–263.
[MathSciNet][Crossref][zbMATH]

Breeden, D. T., & Litzenberger, R. H. (1978). Prices of state contingent claims implicit in option
prices. Journal of Business, 51(4), 621–651.
[Crossref]

Carr, P., & Madan, D. (2001). Determining volatility surfaces and option values from an implied
volatility smile. In M. Avellaneda (Ed.), Quantitative analysis in financial markets (Vol. II, pp.
163–191). World Scientific.
[Crossref]

Chiarella, C., Craddock, M., & El-Hassan, N. (2003). An implementation of Bouchouev’s method
for short time calibration of option pricing models. Computational Economics, 22, 113–138.
[Crossref][zbMATH]

Dempster, M. A. H., & Richards, D. G. (2000). Pricing American options fitting the smile.
Mathematical Finance, 10(2), 157–177.
[MathSciNet][Crossref][zbMATH]

Derman, E., & Miller, M. B. (2016). The volatility smile. Wiley.


[Crossref]

Dupire, B. (1994). Pricing with a smile. Risk, 7(1), 18–20.

Gatheral, J. (2006). The volatility surface: A Practitioner’s guide. Wiley.

Javaheri, A. (2005). Inside volatility arbitrage: The secrets of skewness. Wiley.


Lipton, A., & Sepp, A. (2011, October). Filling the gaps. Risk, 24(10), 78–83.

Pilipović, D. (2007). Energy risk: Valuing and managing energy derivatives. McGraw-Hill.

Rebonato, R. (2004). Volatility and correlation: The perfect hedger and the fox. Wiley.
[Crossref]

Footnotes
1 Another example of a discrete volatility matrix is constructed in Pilipović (2007), where
instead of imposing the exponential structure on local volatilities, an additional constraint is set
by equating long-term local volatilities to historical realized volatilities. Our preference is to
avoid explicitly tying market-implied volatility matrix to realized volatilities due to hedging
imbalances and the presence of the volatility risk premium documented in Chap. 9.

2 See Bachelier (1900). Bachelier’s brief statement of this formula is often overlooked in the
literature, where the formula is typically attributed to a more comprehensive work on this topic
by Breeden and Litzenberger (1978).

3 The equation was presented at several conferences in 1993 and subsequently published in
Dupire (1994).

4 This inverse problem has been extensively studied in academic literature. See, among many
others, Avellaneda et al. (1997), Bouchouev and Isakov (1997, 1999), Dempster and Richards
(2000), Carr and Madan (2001), Bouchouev et al. (2002), Chiarella et al. (2003), Alexander
(2008), Lipton and Sepp (2011), and references therein.

5 For a more detailed discussion of this topic, we refer to Gatheral (2006), and Derman and
Miller (2016).

6 For further reading on jump-diffusions and stochastic volatility models, we refer to Rebonato
(2004), Javaheri (2005), Gatheral (2006), and Derman and Miller (2016).
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
I. Bouchouev, Virtual Barrels, Springer Texts in Business and Economics
https://doi.org/10.1007/978-3-031-36151-7_13

13. Spread Options and Virtual Storage


Ilia Bouchouev1
(1) Pentathlon Investments, LLC, Westfield, NJ, USA

– The payoff of a storage asset can be replicated with calendar spread


options (CSOs). A synthetic storage strategy buys long-dated CSOs at
discounted “wholesale” prices from asset owners, monetizes value
by delta hedging, and resells options at premium “retail” prices to
financial investors.
– The volatility of the spread cannot be lower than the difference
between the volatilities of the spread components. The relationship
between vanilla and spread options is bound by the triangular
arbitrage. Violation of the arbitrage boundary can generate riskless
profits that do not depend on any models or assumptions.
– While many energy spreads trade as independent assets, certain
pairs can also be constructed as spreads between two correlated
prices. This dichotomy leads to two alternative approaches to pricing
and hedging of spread options.
– Correlation-based spread option models are unstable at high levels of
correlation, which is typical for petroleum spreads. The volatility of
energy spreads is better estimated using analogue fundamental
regimes that are indicative of potential disruption risks.

13.1 The Synthetic Storage Strategy


Our last chapter on trading oil options can be viewed as its grand finale.
It is a remarkably simple strategy, but it encompasses many concepts
presented throughout the book, starting with the critical role of oil
storage. This trading opportunity has played a special role in the
author’s own trading career, which started with the launch of this
strategy in the 1990s when quantitative oil trading was yet in its
infancy.1 Importantly, this storage strategy may find new applications
during the energy transition, which has shown the desperate need for
new alternative forms of storage, including batteries.
Storage is arguably the most valuable asset for a professional oil
trading shop. It gives the owner a powerful option to shift a limited
resource from times of plenty to times of relative scarcity. This option is
particularly valuable if supply and demand balances are expected to
change. Storage pays off only when the market is in contango, and when
the trader can cover the cost of storage and financing by buying
physical barrels at a sufficiently large discount relative to the futures
price. When the market is backwardated, the storage option is
effectively out-of-the-money, and the option premium is largely
wasted.2 In other words, a storage asset represents a real put option on
the shape of the futures curve with the strike price determined by the
cost of storage.3
In the early days of the oil derivatives market, quantitative traders
were not part of the elite group of physical traders who had access to
physical storage facilities. To substitute, quants created virtual storage
and synthesized the storage asset on paper. This was done by means of
a financial put option written on the futures time spread. Today this
option, known as a calendar spread option (CSO), represents the most
liquid exotic oil option, with its trading volumes rapidly growing not
only OTC but also on the exchanges.
More formally, a call option on the spread between the two futures
F1 and F2 is defined by the following payoff at the expiration time T
(13.1)
where K is the strike price.
Similarly, the terminal payoff of a put option on the spread is given
by
(13.2)
In other words, the put option on F1 − F2 with the strike K is the
same as the call option on F2 − F1 with the strike −K. Here, we keep
generic notations F1 and F2 for the two futures that may represent
either contracts on the same commodity with different expirations or
futures on two different commodities. In the former case, which is a
CSO, F1 = F(t, T1) and F2 = F(t, T2), with T1 < T2. Standard CSOs
contractually expire on the so-called penultimate day, which is one
business day before the expiration of the futures, in this case the
futures with the shorter maturity T1.
Two types of CSOs trade both on exchanges and OTC. One type is
physically exercised into futures spreads at the expiration, similarly to
vanilla options. The other one is settled financially, based on the
settlement price of the spread on the penultimate day. When a
physically settled CSO is in-the-money at the expiration and is 100%
hedged with the futures spread, then the hedge automatically offsets
futures received from exercising the CSO. In contrast, a fully-hedged
financially settled in-the-money CSO requires the trader to liquidate
futures in the market as close as possible to the settlement price for the
spread, which is typically done using the previously discussed TAS
contract. The liquidation of these futures, which are held as hedges
against expiring CSOs, can cause additional spread volatility on the
penultimate day. For example, such a liquidation played an important
role in the episode of negative oil prices discussed in Chap. 3.
With the use of CSOs, financial traders can closely replicate cash
flows of physical storage without getting their hands dirty in the
somewhat cumbersome operation of a storage facility. The option
premium is equivalent to the rental fee which is paid by physical
traders to lease the storage facility. In the physical market, the storage
lease is typically set for the fixed term, such as one year, with payments
made monthly. A strip of monthly put options on the spread can then be
used to mimic the revenues of the storage asset.
In theory, the value of any option must increase if the option has
more time before its expiration. Surprisingly, this is not always the case
in the market for physical oil storage. Storage facilities are often owned
and operated by independent companies. The owner then leases
storage to professional traders, which effectively means selling asset
optionality in exchange for the monthly rent. The longer the rental
commitment, the more attractive the deal is to the risk-averse operator,
who places higher value on the certainty of securing fixed payments for
a longer time period. For a physical trader, however, it is rather difficult
to commit to a long-term storage lease due to the fluid nature of the
trading business with its relatively short-term vision and frequent
changes in personnel and investors’ risk appetite. As a result, the term
structure of rental fees in the storage market is often relatively flat. It
may even decrease slightly for long-term deals, which are often
discounted by operators to secure a longer-term income stream. Such a
pricing structure is similar to the real estate market and apartment
leases.
This is where an opportunity arises for quantitative traders who can
construct a more liquid version of synthetic storage in the derivatives
market. Ignoring the impact of interest rates, the value of a financial
option must increase with time remaining to maturity. If it did not, then
one would sell a shorter-term option, buy a longer-dated option, and
enjoy holding the latter with zero theta since the option only increases
in value even as its time to expiration shrinks. This is in contrast to the
much flatter term structure of rental fees in the physical market. It
gives a derivatives trader an opportunity to pay a higher premium to
the storage owner relative to the owner’s alternative of renting out the
asset to a physical trader.
From the operator’s perspective, selling an option to a
quantitatively minded trader instead of leasing it in the physical market
does require a bit more trading savviness. The physical barrels now
must be moved in and out of storage efficiently as storage economics
change. The owner must be able to monetize a real option to offset the
liability that arises from selling a financial option. The opportunity to
sell the storage option in the financial market at a higher price usually
provides sufficient incentives to the storage operator to develop such
basic trading capabilities. In addition, collecting an option premium
upfront from selling long-dated CSOs allows the storage operator to use
this cash in lieu of a bank loan and deploy it towards the construction of
additional storage facilities. The US storage capacity has been steadily
growing over the years to accommodate the corresponding growth of
shale production. A portion of this storage capacity has been implicitly
financed by the option premium that operators collected from selling
CSOs in the derivatives market.
As we have already seen in Chap. 9, oil options with longer
maturities tend to be somewhat underpriced from the volatility
perspective. For vanilla options, the excess supply of long-dated
volatility comes from leveraged producer hedgers that attempt to
monetize the real asset optionality in the derivatives market. For long-
dated spread options, such a discount to the fair value is provided by
storage operators, and its magnitude happens to be even more
significant. Taking the other side of this imbalance by buying a
discounted long-dated CSOs and delta hedging them to protect the
option value has become an attractive investment strategy. We call this
strategy a synthetic storage warehouse.
Figure 13.1 helps to visualize the rationale behind this trading
strategy. It shows a typical term structure of rental premia for physical
storage relative to market prices of equivalent CSO puts. It also shows
the realized fair value for the CSO, which is calculated here using the
realized volatility during the lifespan of such options. The fair value is
what a CSO owner might expect to generate by delta hedging a long
option position during the lifespan of the option. This example is
stylized for pedagogical convenience, where, for simplicity, we assume
that all options are ATM. The nature of the trading opportunity remains
the same for OTM options, where the fixed strike represents the cost of
physical storage.
Fig. 13.1 Storage rents and CSO market prices with their estimated fair value
One can see that rental fees vary only slightly with the duration of
the rental commitment. The term structure of fees even declines for
long-dated contracts. Like in the real estate market, the highest
premium is paid for a short-term lease. Long-dated leases may be
cheaper as the operator is willing to provide incentives by discounting
the price in return for stable longer-term cash flows. Short-term
physical deals are also favored by trading houses, looking to capture
additional value from blending different grades of oil. However, it
would be difficult for the blender to justify long-term commitments, as
imbalances among various oil grades that make the business of
blending profitable often disappear quickly.
The market price for a long-dated financial CSO is also typically
pressured down by discounted forward physical storage rates. This
makes the CSO an attractive buy versus its fair value, which is the
expected cost of its dynamic replication by trading futures spreads. In
some sense, the storage operator and the CSO buyer split the expected
profit. The operator gains by selling a financial option at a price higher
than the one offered in the physical market. The volatility trader, in
turn, extracts profits by buying and delta hedging the discounted
optionality in the derivatives market.
Figure 13.2 shows term structures of implied and local volatility
that correspond to the CSO prices in Fig. 13.1. For simplicity, the spread
volatility is assumed to be time-homogeneous, depending only on time
remaining to maturity. Implied volatility is computed using the
Bachelier formula, which is the market standard model for pricing
CSOs, and the units of implied normal volatility are dollars per barrel.
In contrast to the Black model, the normal assumption for spread
options allows spreads to be positive and negative as the market flips
between backwardation and contango. We will discuss pros and cons of
different pricing models for spread options later in this chapter.

Fig. 13.2 Implied and corresponding normal local volatility for a CSO computed using the
Bachelier formula

The local volatility curve for time spreads is typically very steep. As
a result, the spread volatility rolls up the curve so fast that the gain
from vega negates a large portion of the option theta decay. The
steepness of the curve is driven by a combination of the much higher
fundamental risks embedded in short-term options and discounted
prices offered by storage operators for longer-term deals. Occasionally,
the discount for forward CSOs becomes so large that it even violates the
no-arbitrage restriction (12.​2) on how fast implied spread volatility
term structure can decline.
The long-dated CSO that the derivatives trader acquires does not
stay cheap forever. In contrast to a physical storage facility, the lifespan
of the virtual storage is shorter. When the CSO expires, its time value
must decay to zero. It means that at some point prior to its expiration,
the CSO is likely to become expensive relative to its fair value that the
buyer expects to generate by delta hedging. Therefore, one challenge
for a synthetic storage strategy is to liquidate an option before the
escalation of theta turns into a real burden. But why would anyone in
the market at that time buy such an overpriced and rapidly decaying
financial asset?
Recall from Chap. 9 our study of options as insurance contracts that
carry a volatility risk premium paid by hedgers. The same argument
applies to CSOs. Here, the primary demand for hedging comes not from
end-users, but rather from financial investors. We have already seen
how punitive the negative roll yield is for long-only investors holding
futures in contango markets. By paying this roll yield, financial
investors effectively outsource the function of storage to carry traders,
but the magnitude of the cost to roll is subject to the vagaries of the
market. Buying a CSO provides financial investors with an interesting
alternative. Instead of taking the risk of an excessive roll cost in the
event of a super-contango, investors can hedge such a cost in the
financial market. They can pay the premium to buy CSO puts and
eliminate potentially the largest drag on the performance of their long
futures investment. Such an incremental demand for short-term CSO
puts from hedgers of the roll yield turns out to be substantial, especially
during times when the market is in a state of contango.
In addition, professional physical trading houses also like buying
short-term CSOs to place highly leveraged speculative bets on time
spreads. These options are used as effective wagers on the squeeze
probability that the market may either run out of oil or run out of
capacity to store it. CSOs are also routinely used for risk management
by trading desks to protect large speculative positions taken in the
futures market. Physical speculators trade time spreads based on
fundamental models that rarely look beyond just a few months ahead.
Their demand for spread volatility tends to be higher when futures
approach expiration. This hedging imbalance allows CSO holders to exit
the ownership of virtual storage just in time before its theta decay
starts to accelerate.
To summarize, the CSO strategy described here is akin to a three-
step virtual storage warehouse. First, one acquires a long-dated strip of
monthly CSOs from storage operators at a discounted “wholesale” price.
The time value of financial options is then protected by delta hedging.
Given the steepness of the volatility term structure and only a minimal
theta to cover, the bar for delta hedging gains is set rather low. Finally,
the trader parses a long-dated strip into individual month CSOs and
resells them to financial investors looking to hedge against the negative
roll yield, or to short-term physical speculators. An embedded volatility
risk premium allows these short-term sales to be made at higher
“retail” prices.
This strategy may sound like a remarkably simple way to make
money. One challenge is, of course, in knowing how much to pay for
such a long-dated optionality at the initiation of the trade. Another one
is the calculation of delta, which heavily depends on the pricing model
to be covered later in this chapter. But before we get into details of
spread option pricing, an important arbitrage boundary that plays the
central role in many spread option strategies must be introduced. This
boundary is somewhat easier to visualize first in the context of cross-
product and locational spread options.

13.2 Triangular Correlation Arbitrage


The idea behind the virtual storage strategy can be applied to cross-
asset spread options, many of which also represent financial replicas of
real options. We have already highlighted on several occasions that an
oil refinery represents a call option on the spread between a refined
product and crude oil, called the crack spread. Similarly, a pipeline asset
or an oil tanker is an option on the spread between prices of the same
commodity but in two geographical locations. A natural question would
be whether the virtual asset strategy that buys discounted financial
optionality from an asset owner can also be applied to a refinery or a
pipeline whose owner may wish to monetize the physical optionality.
This turns out to be more difficult, as in contrast to a storage facility,
refineries and pipelines are simply too expensive to be financed by
option premiums in the derivatives market.
The trading opportunity in crack options and locational spread
options, such as options on the WTI-Brent spread, comes with a slightly
different twist. While it is plausible for the volatility trader to be able to
buy such a spread option at a slight discount to the fair value and
effectively own a tiny refinery or a pipeline on paper, the owner of a
large asset is unlikely to be sufficiently motivated to offer any
substantial discounts to relatively small buyers. As a result, the implied
volatility curve for a cross-product spread option is usually much
flatter, and the volatility roll up does not cover as much of the theta
decay. Thus, a cross-product spread option represents a much faster
decaying financial asset. However, such a spread option is a vital
component of a more advanced strategy, known as triangular
correlation arbitrage.
In the triangular correlation strategy one attempts to isolate
correlation exposure between two assets while neutralizing overall
exposure to volatility by constructing a mini-portfolio of three options.
This is accomplished by combining an option on the spread with the
spread between two vanilla options on each leg of the spread. The
spread between two options is also known as a synthetic spread option,
where one sells a relatively expensive option on one commodity and
hedges it by buying a cheaper option on a correlated commodity. For
example, the trader can sell a call option on diesel at a premium relative
to a similar call option on crude oil. The higher price of a diesel call may
reflect not only the higher realized dollar volatility of diesel, but also
some imbalances in the hedging market. Diesel calls are often relatively
expensive as they are used by airlines to substitute for jet fuel hedging,
while crude oil calls are well supplied by oil producers via two- and
three-way collars. The term synthetic spread option is somewhat
misleading as it represents the spread of options, and not the option on
the spread.
In the synthetic spread option trade of selling diesel calls versus
buying crude oil calls, the dealer collects net premium in return for
taking diesel idiosyncratic risks that cannot be mitigated by crude oil
options. If diesel and crude oil futures happen to move up and down in
parallel by a similar amount in dollars per barrel, then gamma risks in
both markets largely offset each other. In this case, the dealer can retain
most of the premium collected upfront by delta hedging both options.
However, if the diesel happens to be more volatile than crude oil,
perhaps driven by an unforeseen interruption of refining processes,
then the dealer may end up suffering significant losses. Such refinery
outages could be caused by hurricanes or unexpectedly cold weather
which explains more pronounced seasonal patterns in the volatility of
refined products.
One way for the seller of a synthetic spread option to protect the
relative value trade against unexpected moves in the spread is to buy an
option on the spread itself, perhaps from a refinery that naturally owns
it. Then the strategy involves selling a relatively expensive call option
on diesel to an airline and buying two relatively cheap call options, one
option on crude oil from a producer, and another option on the spread
from a refinery. If the trader somehow manages to buy two cheaper
options at the price of the expensive one, then it will set an important
triangular arbitrage boundary. While this arbitrage boundary is
unlikely to be breached in practice as it is difficult to buy two options
for the price of one, it still plays a pivotal role in trading correlation and
spread options.
To construct the triangular boundary more precisely, we express the
variance of the spread between two random variables X and Y in terms
of the variances of each leg, net of the covariance between them:
(13.3)
One can also rewrite this formula as follows

It is clear from this decomposition that the volatility of the spread is


made up of two components, the spread between two volatilities, and
the contribution resulting from non-perfect correlation. The volatility
of the spread increases either when one commodity becomes more
volatile than the other, or when the correlation between the two assets
declines.
Since the correlation coefficient ρ ≤ 1, it follows that the volatility of
the spread cannot be lower than the spread of volatilities
(13.4)
Only in the very special case of two perfectly correlated assets, this
inequality becomes an identity.
Figure 13.3 illustrates this relationship between volatilities and
correlation geometrically. It is presented in the form of a triangle whose
sides are drawn in proportion to volatilities σ(X), σ(Y), and σ(X − Y) of
three options. The Eq. (13.3) is effectively the law of cosines if we think
of the correlation coefficient ρ as the cosine of the angle formed by σ(X)
and σ(Y).

Fig. 13.3 Triangular relationship between volatilities and correlation

The size of the angle α affects the length of the opposite side, which
represents the volatility of the spread σ(X − Y). If the two assets are
perfectly correlated, then the triangle collapses into a straight line as
the correlation, or the cosine of a zero angle, is equal to one. In this case
of perfect correlation, the volatility of the spread is determined entirely
by the spread between the two volatilities. If we keep the volatilities of
each leg unchanged, but instead widen the angle, which means
reducing correlation, then the length of the opposite side, which
corresponds to the volatility of the spread, increases. The lower the
correlation, the more volatile the spread, and, therefore, the more
expensive the option on the spread is.
This triangular relationship sets up an important boundary not only
for volatilities, but for option prices as well. It follows, for example,
from the Bachelier formula that prices of ATM options with the same
maturity are directly proportional to their dollar volatilities, and,
therefore, the same inequality must also hold for ATM call prices:
(13.5)
In other words, an option on the spread CSP cannot be worth less
than the spread of two options with the same moneyness and maturity.
If market prices for three options violate this inequality, then it is an
indication of a potential arbitrage opportunity.
Let us illustrate this with a simple numerical example. Let
X(t) = 60 and Y(t) = 55 represent diesel and crude oil prices in dollars
per barrel ($/bbl), as observed today at time t. Assume that their
implied ATM Black volatilities for options with one-year maturity are,
respectively, 30% and 29%. Then call options prices, obtained from the
standard Black formula, are $7.18/bbl for diesel and $6.36/bbl for
crude oil. By selling the diesel option and buying the crude oil option,
the dealer collects $0.82/bbl in exchange for taking unforeseen diesel-
specific risks. It is quite possible that the net hedging cost could exceed
the premium collected if diesel prices happen to be highly volatile. To
mitigate this risk, a prudent dealer would attempt to buy the same
maturity ATM option on the crack spread from a refinery.
If the dealer does buy the spread option, then the overall portfolio
consists of three options, one short option and two long:

Given the relationship (13.5), such a portfolio has a non-negative


payoff for any values of X and Y. In other words, despite being short one
option, the overall portfolio π itself can be viewed as owning an option,
as it cannot lose money under any scenario. If somehow the dealer can
convince a refinery to sell an ATM spread option at, say, $0.70/bbl, then
the dealer could even get paid $0.12/bbl to own such an option π! Since
the volatility of the crack spread during normal market conditions is
often low, selling a spread option may still look appealing to a refiner, as
it brings some cash and turns the naturally long asset exposure into a
covered call. At the same time, the dealer not only gets paid to own an
option π, but can even reap additional benefits in some particularly
favorable price scenarios. Under no circumstances does the dealer
stand to suffer any loss at expiration even if all three options are left
unhedged.
To illustrate, consider several price scenarios at the expiration of
these options.

Scenario 1 Both diesel and oil markets rally, and diesel outperforms
crude oil, so that X(T) = 70, Y(T) = 60. Then a ten-dollar loss from the
short diesel call is precisely offset by the sum of a five-dollar gain from
owning the oil call, and another five-dollar gain from owning an ATM
crack call struck at $5/bbl:

Scenario 2 Both diesel and oil markets rally, and the oil price
outperforms diesel, so that X(T) = 70, Y(T) = 67. Here, the gain from
owning the oil option exceeds the ten-dollar liability on the diesel call,
while the crack call expires worthless:

Scenario 3 Diesel rallies, but oil falls as a result, perhaps, of an


unexpected weather event that temporarily shuts the refinery, but
simultaneously reduces demand for crude oil. Let X(T) = 70, Y(T) = 50.
In this case, the oil call expires out-of-the-money, but the loss on the
diesel call is more than offset by the gain on the ATM crack call which is
$15/bbl in-the-money:

Scenario 4 The diesel price remains unchanged as inventories are


abundant, but oil rallies due to an unexpected OPEC production cut, so
that X(T) = 60, Y(T) = 60. Then the only option among the three that has
a non-zero value at expiration is the oil call:

Only in the first scenario, which is, admittedly, more likely than the
other three, the dealer’s payoff at the options expiration is precisely
equal to zero. This scenario corresponds to the case of equal dollar
variance and perfect correlation between diesel and crude oil. In the
other three scenarios, which are less likely, the dealer receives
additional benefits from options payoffs. In the second scenario, the
positive payoff comes from a favorable move in relative realized
volatilities when crude oil moves more than diesel. In the last two
scenarios, the profit is driven by decorrelation of the two commodities.
This portfolio provides an example of an arbitrage boundary which
is independent of any models or assumptions. If one can execute a
triangular trade in the market at net zero cost, it would genuinely
represent a free option. The only thing to worry about would be the
creditworthiness of the option sellers. The dealer can only profit from
an OTC trade if neither the refiner, nor the producer, default on their
short option obligations. A popular joke, or an interview question for a
trading job, is “how much to pay for an OTC option to a counterparty
that has an option not to pay you back? The answer is, of course, zero.
Some energy volatility dealers, however, learned painful lessons by
buying “cheap” options from asset owners with poor credit, who ended
up defaulting when their derivatives obligations became particularly
large. In one case, the dealer even ended up owning an entire power
plant which was placed as collateral against the financial derivative sold
by an asset owner. Unfortunately for the dealer, the value of the plant
covered only a fraction of the payoff due on the derivatives trade.
A perfect triangular arbitrage is unlikely to be found in competitive
markets. However, the pricing boundary created by options on two
correlated assets and an option on the spread forms the backbone of
spread option and correlation trading. We now discuss how this
relationship between the volatility of the spread and the correlation is
captured in two alternative modeling paradigms.

13.3 The Dichotomy of Spread Option Pricing


Modeling spread options using conventional methodologies developed
for financial markets could quickly turn into a quagmire. Petroleum
spread options are structurally different from spread options that trade
in interest rate and equity markets. A financial spread is typically a
derived variable, computed as the difference between the prices of two
assets. In contrast, in many energy markets, the spread itself is a
primary traded variable that connects a less liquid asset to an industry
benchmark. The value of an illiquid asset is then synthetically
constructed by adding the traded value of the spread to the price of the
liquid leg. For example, nearly all US oil grades trade primarily as the
basis to WTI with their prices calculated as the sum of the WTI price
and the basis spread.
Certain petroleum spreads exhibit a split personality. This happens
when some market participants view the spread as an independently
traded asset, while others see it as the difference between two prices.
Many refined products, such as diesel or gasoline, trade in such a dual
manner. Fundamental traders tend to analyze refined products through
the lens of the crack spread, as their trading strategies generally avoid
taking directional exposure to petroleum prices. In contrast, financial
and systematic traders, who often rely on covariance matrices in
managing diversified commodity portfolios, prefer to trade each leg of
the spread separately. The way different participants think about the
spread dynamics often leads to framing biases which are essential to
understand when selecting the model for trading spread options.
Such a trading dichotomy suggests looking at the price of the spread
option from two different angles. One approach is to model an option
written on an independently traded spread contract that follows a
certain stochastic process. An alternative is to specify the stochastic
dynamics of each leg of the spread and assume some correlation
structure between the two processes. Fortunately, over time the two
methodologies have somewhat converged to more practical hybrid
solutions, where the volatility of the spread in the former method and
the correlation input in the latter are connected to each other.
The first approach of handling the spread as an independent
variable is a straightforward extension of methods presented in
previous chapters. Let S be the spread between two futures contracts F1
and F2
that follows a diffusion process of the form:

where σ(S, t) represents the local volatility of the spread.


To apply our previously developed techniques, we only need to
replace the futures price F with the spread S and acknowledge that the
local volatility function refers to the volatility of the spread. Since many
petroleum spreads can be either positive or negative, the local spread
volatility must be measured in absolute or dollar terms, and not in
percent. As it was shown in Chap. 8, the drift term μ(S, t) does not play a
direct role in the option valuation. This may sound somewhat
counterintuitive, as one would expect relatively fast spread mean-
reversion to have an impact on option prices. The mean-reversion does
indeed play a role but indirectly, as the speed of mean-reversion affects
the steepness of the volatility term structure.4
The most straightforward approach for pricing oil spread options
assumes that the spread follows an ABM process with constant dollar
volatility, for which the option price is given by the Bachelier formula. It
is remarkable that a formula originally derived over a century ago to
price options on French government bonds retains its relevance a
hundred years later in the modern oil market. If spread option prices
are available in the market, then they can be easily expressed in terms
of implied normal spread volatilities and visualized in the form of the
volatility smile for the spread. Such a smile usually has a very
pronounced curvature on both sides, reflecting higher normal volatility
for out-of-the-money puts and calls.
To adjust the Bachelier model for non-normality and fat tails, which
are typical for the behavior of energy spreads, one can use the QN
model developed in Chap. 10. The QN model works particularly well for
options on the spread between assets that are highly correlated during
fundamentally balanced markets, but prone to periodic sharp
dislocations. Such dislocations occur when economic linkages between
two assets are disrupted by unforeseen events, such as pipeline
outages, weather, or geopolitical conflicts. When two assets suddenly
diverge, the volatility of the spread spikes, and the value of the spread
option explodes.
The parabolic nature of the QN model is very intuitive for many
petroleum spreads. The vertex of the parabola that marks the lowest
local volatility is often associated with the equilibrium spread level,
corresponding to the economics of the physical arbitrage that connects
two markets. For example, for CSOs the vertex of the parabola is often
located near the spread level that reflects the cost of storage. Likewise,
for WTI-Brent options, it is related to the cost of freight. The further
away the spread deviates from its normal trading range in either
direction, the higher the uncertainty and the resulting dollar volatility.
The quadratic local volatility with its increasing curvature on both ends
captures such dynamics quite well. Spread options tend to trade under
a sticky local volatility smile for specific strikes, which retains its shape
regardless of the option moneyness. The ability to capture the tails of
the spread behavior is the major advantage of modeling the spread as
an independent asset.
Occasionally, one also encounters non-standard spread options,
such as spread APOs, early expiry spread options, and even spread
swaptions. For example, APOs on the spread between gasoil and Brent
are particularly popular among European and Asian refineries. Some
spread APOs are listed on organized exchanges even though their
primary liquidity is still concentrated OTC. All methods developed in
Chap. 11 for exotic OTC options apply directly to non-standard spread
options, provided that the spread itself is modeled as a stand-alone
asset. One can also calibrate the local spread volatility and bootstrap it
from market prices of spread options. When the spread is treated as the
primary asset, the problem of pricing spread options, as
mathematicians often say, reduces to the previous case that has already
been solved. This is, by far, the most popular way of handling spread
options in petroleum markets.
There are, however, certain spread options for which modeling the
spread as a single asset has limitations. Consider a refinery that buys a
put option on a spread which approximates its profit margin with a
basket of refined product outputs and crude oil inputs. The weights of
the basket components are highly specific to individual refineries. For
example, one refinery may be producing a 50–50 basket of diesel and
gasoline, while another one may have a 60–40 yield split between two
of its main products. If one chooses to model such baskets as stand-
alone assets, such as basket one, basket two, basket three, etc. with
their own independent basket volatilities, then any direct connection
between volatilities of closely related baskets will be lost. To manage
the portfolio of such refinery margin options, it is better to specify the
dynamics for individual components of the spread along with their
correlation.
Perhaps a more pressing need to model individual components of
the spread arises in other energy markets, such as natural gas and
power. For example, a power plant can be modeled as a call option on
the spread between the price of electricity that trades in megawatt-
hours (MWh) and natural gas that trades in million British thermal
units (MMBtu), where the latter is multiplied by the heat rate that
measures the efficiency of the plant.5 Even though this approach of
modeling two assets independently is less frequently used for
petroleum spread options, it does provide an interesting alternative
perspective on valuation.
To keep the exposition simple, we only illustrate the correlation-
based framework for two assets as extending it to a multi-asset spread
basket is relatively straightforward. We assume that F1 and F2 follow
two diffusion processes

with random components dz1 and dz2 that have correlation ρ. As


before, the drift terms are eliminated by delta hedging and the spread
option value depends on local volatilities of two futures and the factor
correlation which, for simplicity, is assumed to be constant.
Consider first the simplest case where the local dollar volatilities of
the two legs are constant, σ1(F1, t) = σ1,A and σ2(F2, t) = σ2,A. Then both
futures are normally distributed, and, therefore, the spread between
them is also normally distributed with the arithmetic spread volatility
σS,A, given by
The problem, therefore, again reduces to the previous one, where
the spread is an asset by itself, and the option price is given by the
Bachelier formula with volatility σS,A. The normal volatility of the
spread is expressed in terms of the normal volatilities of the two legs
and the correlation between the two assets.
We next take a short-cut approach similar to the one that was used
for pricing APOs and swaptions. We obtain the exact formula for a
normal process and then use it to construct an approximation for
option prices under other diffusions. For example, consider two
conventional lognormal processes, where the volatilities for each leg
are proportional to the corresponding futures prices:

The spread between two lognormal variables is not lognormal. In


fact, it is closer to being normal than lognormal. In this case, one can
still use the Bachelier formula as an approximation with normal spread
volatility given by

(13.6)

Assuming that both futures prices and their implied volatilities are
observable in the market, this transformation allows the trader to
calculate the volatility of the spread given some assumption about the
correlation. Alternatively, if the price for the spread option is observed
in the market, then the trader can back out the implied correlation as
follows:

(13.7)

where, for simplicity, we kept the same letter to denote implied and
local geometric volatility.
When the spread is modeled as an independent asset, its delta and
other Greeks are reported with respect to the spread itself. In other
words, the delta with respect to the first leg of the spread is the
negative of the delta with respect to the second leg. This makes perfect
sense for closely related spreads between assets with similar
volatilities, as hedging is typically done by transacting directly in the
spread contract. For example, an option on the WTI-Brent spread will
always be hedged by trading in the spread itself. However, when one
asset is significantly more volatile than the other, the trader will most
likely be looking to hedge the risk of the more volatile asset differently
by delta hedging it outright rather than as a spread to something that is
moving a lot less. The problem becomes more evident when the two
legs of the spread trade in different units. This is more common for
cross-product spreads that arise in the natural gas and power markets
where the two legs of the spread are more disjoint.
An alternative valuation method attempts to mold spread options
into the standard lognormal BSM pricing framework. Its main idea is
based on the observation that the ratio of lognormal variables is also
lognormal. Furthermore, the variable

can be viewed as approximately lognormal if the strike price K ≪ F2.


More formally, the terminal payoff of the call option on the spread
(13.1) can be rewritten in terms of the variable x as follows:

Both the spread option price CSP and the underlying variable x are
scaled by the same quantity F2 + K. They are effectively expressed in
terms of the new scaled numeraire. The payoff of the scaled spread
option price is then the same as the payoff of the regular call option
written on the variable x with the strike price equal to one.
Unfortunately, the spread option admits an analytic solution in a
two-dimensional lognormal setting only in a very special case when the
strike price K = 0 and the variable x becomes lognormal. Such a solution
is known as the Margrabe formula, which was initially referred to as an
option to exchange one asset for another. For K ≠ 0, there exist several
convenient approximations, among which the following Kirk
approximation is the most widely used6:
(13.8)
where

Here, τ = T − t and the effective geometric volatility σG,x is given by

The formula (13.8) resembles the Black formula on the underlying


asset x for the option price scaled by F2 + K with the strike price equal
to one. If we let K = 0, then it turns into the Margrabe formula, which
represents an exact solution for the two-dimensional lognormal
distribution. Many other more sophisticated approaches to the
correlation-based pricing of spread options have been proposed, but
the accuracy of the approximation (13.8) is usually sufficient for all
practical applications in petroleum markets.7
Despite the popularity of the two-dimensional lognormal
framework in academic studies, it turns out to be of a limited use for
practitioners in the oil market. Using this framework, it is more difficult
to incorporate the tails of the distribution, which are critical for pricing
energy spread options. In theory, one can extend the idea of the
volatility smile for the spread and use (13.7) to construct the implied
correlation smile for spread options with different moneyness. In fact,
such a correlation smile would appear more like a frown, as lower
correlation is needed to generate higher prices for OTM spread options.
However, the process of constructing implied correlations for
different moneyness is somewhat ambiguous. For a given strike K of the
spread option, there are infinitely many pairs of strikes for individual
assets K1 and K2 that match the spread strike defined by K = K1 − K2.
Using different pairs of implied volatilities, σ(K1) and σ(K2), inevitably
produces different implied correlations. Therefore, the concept of
correlation frown is not uniquely defined. Additional constraints must
be imposed on the selection of K1 and K2, which further complicates the
calibration process.
Having two methodologies for pricing spread options in place, we
finish this chapter by highlighting some important practical challenges
with empirical estimation of inputs to these models and provide some
guidance on approaches favored by professional traders.

13.4 Dealing with Unobservables


While correlation is a commonly used metric for describing co-
movements between financial assets, applying it to pricing oil spread
options is quite dangerous. Many petroleum futures are highly
correlated as they are linked to each other by the economics of storage,
transportation, and processing. Unfortunately, spread options are
extremely sensitive to the correlation input, especially when
correlation is particularly high.
Let us illustrate the challenge with a simplified example of two
futures contracts that have the same price F1 = F2 = F and the same
percentage volatilities σ1, G = σ2, G = σ. In this case, the volatility of the
spread between the two futures is related to the correlation between
them through (13.6):

The price of an ATM spread option is then given by the Bachelier


formula (8.​9), where, as usual, we ignore the impact of interest rates:

Figure 13.4 shows how the value of an ATM spread option changes
for a wide range of correlations.
Fig. 13.4 The price sensitivity of a one-year ATM spread option to correlation for F1 = F2 = 60,
and σ1, G = σ2, G = 0.30
Higher correlation means lower spread option value. It is easy to
see that the relationship becomes highly nonlinear for extremely high
levels of correlation. In fact, since

the sensitivity of the spread volatility to correlation becomes infinite


when correlation approaches one.
Now let us look at the magnitude of the correlation coefficient for
key petroleum spreads that underlie traded spread options.
Correlations are especially high for time spreads between futures on
the same commodity but with different maturities. The average
correlation between the first two nearby WTI futures is approximately
99%. Moreover, the correlation between longer-term futures on the
same commodity is basically indistinguishable from one. Imagine now
being asked to come up with a bid for a long-term storage asset using
the CSO pricing framework, which requires empirically calculated
correlation input. Using, for example, 0.996 correlation instead of 0.999
would double the price of a benchmark one-month CSO, or,
equivalently, double the value of the physical storage asset. One could
easily be relieved from duties as a desk quant for providing such an
unacceptable degree of precision. It is unlikely that anyone would be
comfortable purchasing an asset based on such an unreliable valuation
model.
Figure 13.5 shows other examples of realized correlations for key
petroleum spreads. While all correlations generally remain high, one
can also observe periodic bouts of rapid decorrelation. These periods of
decorrelation are driven by unexpected short-term disruptions in
economic linkages between components of the spread.8

Fig. 13.5 Realized correlations for many petroleum spreads are extremely high and unstable
(6-month rolling correlations)

Such correlation behavior is typical for petroleum markets, which is


also reflected in a parabolic function for the spread volatility versus the
spread itself. When the oil market is fundamentally balanced, most
futures tend to move together, with spreads exhibiting low volatility.
However, when disruptions occur unexpectedly, futures decorrelate and
the volatility of spreads rapidly increases. These observations lead us to
the following rule of thumb: for spread options on highly correlated
assets one should always model the volatility of the spread itself
instead of relying on unstable correlation-based models for its two
components.
In the case of a CSO, correlation-based pricing models face yet
another challenge. Not only is the correlation input hard to estimate,
the volatility of the second leg of the time spread is also not observable.
Recall from the discussion of EEOs in Chap. 11 that what matters for
pricing an option is the local volatility during the lifespan of the option,
which is not the same as the implied volatility of a vanilla option.
Consider the primary benchmark CSO on the prompt one-month
futures spread. The volatility of the front month futures can be
considered to be observable.9 However, for pricing the CSO using a
correlation-based model, one also needs to know the local volatility of
the second month futures measured only during the life of the CSO,
which expires a month earlier. The local volatility of the second leg is
not observable, and it can be drastically lower than the implied Black
volatility of the second-month futures.
To illustrate, we use as an example a one-month ATM CSO on the
spread between the first two futures. Let the implied Black volatilities
for these futures be v1 = 0.35 and v2 = 0.32, and the correlation between
the two contracts be empirically estimated at 0.995. If we assume again
that local volatility is time homogeneous, depending only on the time
remaining to maturity, then following the bootstrapping methodology
of Chap. 12, the local volatility of the second leg during the first month
when the CSO exists is then given by

Using the Bachelier formula, the corresponding price of an ATM CSO


call is $0.49/bbl. If the trader mistakenly uses implied volatility v2 for
the second leg of the spread instead of its local volatility σ2, then the
CSO price drops by more than one-third to $0.31/bbl. By selling an
option at such a lower price, the trader may be making a grave error.
This mistake by the seller of the spread option could even violate the
boundary set by the triangular arbitrage. An arbitrage would occur if
the CSO and another EEO on the second leg of the spread can be
purchased at a total price that does not exceed the price of a vanilla
option on the first leg. Surprisingly, such an arbitrage trade has indeed
been executed in the oil market on several occasions, rewarding the
dealer with a free option.
To summarize, the extreme sensitivity of a spread option price to
the correlation input and the need for an additional estimation of the
unobservable local volatility of the second leg of the spread makes
correlation-based methods practically unusable for pricing many
petroleum spread options. As a consequence, the Bachelier model
became a de facto industry standard for pricing CSOs and other spread
options on highly correlated futures.
The initially obscure market for oil CSOs has gained transparency
over the years. At the time of writing this book, market prices for one-
month ATM CSOs are regularly quoted by option brokers. One can,
therefore, easily calculate CSO implied normal volatility. However, this
information is insufficient for backing out implied correlation, as the
local volatility of the second leg is still unobservable. There are many
combinations of implied correlations and the second local volatility that
result in the same volatility of the spread. This is similar to the non-
uniqueness problem in model calibration that we have encountered in
the previous chapter for the local volatility matrix driven by a multi-
factor model. It is also impossible to construct an implied correlation
frown for CSOs in a unique way, because it requires the knowledge of
the volatility smile for all EEOs.
If the market for CSOs is deemed to be inefficient because of
hedging imbalances, then the ability to calculate implied CSO volatility
does not necessarily answer the question whether an option is cheap or
expensive. Some traders attempt to answer this challenge by using
estimates of the past realized volatility to approximate the future
implied volatility for the spread. However, for spread options this
approach must be taken with the greater degree of caution. The local
spread volatility, such as the one shown in Fig. 13.2, is very steep.
Therefore, the primary contributor to the total variance of the spread
and to the value of a CSO is volatility that is expected to occur in the
future near the contract expiration, and not the prior historical
volatility of the same spread, which is generally much lower.
For example, if we are pricing a CSO on a January–February spread
with, say, six months to maturity, then the recently observed volatility
of this spread tells us almost nothing about what volatility to expect in
December when the CSO is about to expire. Instead, more meaningful
information can be obtained by looking at the realized volatility of
previous spreads over the same six-month lifespan prior to their
corresponding expirations. Each previous expiration produces a single
data point for the realized volatility with given time to expiration. One
can then average monthly data to come up with some generic historical
term structure of spread volatility that can be compared to the implied
volatility of the spread.
This simple approach can only be used as a starting point for
modeling volatility term structure for time spreads. The proper pricing
of spread options must always be tied to fundamentals, as spread
volatility typically depends on the state of local inventories in nearby
storage facilities. Many large participants in the CSO market are
professional physical traders who use sophisticated technologies to
gain access to up-to-date inventory information and pipeline flows.
They tend to price spread options based on anticipated trajectory in
forward inventories, specifically based on the likelihood of inventories
reaching their boundaries, when the spread volatility is expected to be
particularly high. This is akin to modeling the spread dynamics based
on the probability of a squeeze, as it was presented in Chap. 3.
One way to link historical volatility estimates to current market
conditions is to utilize the concept of analogue periods. Instead of using
a simple average of volatility observations over a selected lookback
period, the trader can average only observations that are perceived to
be relevant to the prevalent fundamental regime. Obviously,
establishing what is relevant and what is not is much easier said than
done. Some degree of relevance can be quantified using inventory data
as a measure of fundamentally similar periods. For example, one could
choose analogues by identifying historical periods when inventories
were sufficiently close to the current level and average realized
volatility only during such periods. One can even assign different
weights to each historical volatility data point based on the distance
between inventories in the analogue period and their present level.
Numerous other schemes can be developed to weigh historical
volatility realizations that can incorporate seasonality or the slope of
inventory trends that indicate whether inventories are building or
drawing. We should always remember though that for the most part,
the value of a spread option is driven by the probability of a squeeze, or
the likelihood of an unexpected event. This probability is extremely
difficult to extract solely from historical price data. Identifying and
comparing the behavior during fundamental analogue periods is
probably the best that one can do. Even though history rarely repeats
itself, in the case of spread volatilities and correlations it often rhymes
with fundamental environments that led to prior squeezes. In the
spread options market, quantitative volatility models and fundamental
models of supply and demand are deeply intertwined.
These comments provide only a rough guide to estimation of
volatility from historical and fundamental data. Volatility forecasting
can also involve many other sources of information, such as speculative
positioning, open interest, and various measures of liquidity. Any
attempt to provide a more rigorous prescription for calculating the fair
value of a spread option would have been a disservice to the reader, as
volatilities and correlations are impacted by too many constantly
evolving factors. The problem of spread option pricing is representative
of many other problems that arise in quantitative oil trading. In fact,
none of the strategies discussed in this book should be read
prescriptively. If one ever claims to have a precise step-by-side guide to
making money in the oil market, then my advice would be to stay away
from such a guide, as promised money is likely to be an illusion.
Throughout the entire book, we emphasized that quantitative oil
trading is a blend of art and science, a blend of a human and a machine.
The objective of the book is to provide readers with some science that
has been helping the author to paint the picture of the oil market
throughout his twenty-five career as a trader. The actual trading
strategy is always an art. In this book, I’m happy to share some of mine,
both the pieces that worked and the ones that failed, and I’m looking
one day to learn about yours.

References
Carmona, R., & Durrleman, V. (2003). Pricing and hedging spread options. SIAM Review, 45(4),
627–685.
[MathSciNet][Crossref][zbMATH]
Considine, J., Galkin, P., & Aldayel, A. (2022). Inventories and the term structure of oil prices: A
complex relationship. Resources Policy, 77, 1–18.
[Crossref]

Dempster, M. A. H., & Hong, S. S. G. (2002). Spread option valuation and the fast Fourier
transform. In H. Geman, D. Madan, S. R. Pliska, & T. Vorst (Eds.), Mathematical finance, Bachelier
congress (Vol. 1, pp. 203–220). Springer.
[Crossref]

Eydeland, A., & Wolyniec, K. (2003). Energy and power risk management: New developments in
modeling, pricing, and hedging. Wiley.

Johnson, O. (2022). 40 classic crude oil trades: Real-life examples of innovative trading.
Routledge.

Kirk, E. (1995). Correlation in the energy markets. In Managing energy price risk (pp. 71–78).
Risk Publications.

Margrabe, W. (1978). The value of an option to exchange one asset for another. The Journal of
Finance, 33(1), 177–186.
[Crossref]

Shimko, D. C. (1994). Options on futures spreads: Hedging, speculation, and valuation. The
Journal of Futures Markets, 14(2), 183–213.
[Crossref]

Swindle, G. (2014). Valuation and risk management in energy markets. Cambridge University
Press.
[Crossref]

Venkatramanan, A., & Alexander, C. (2011). Closed form approximations for spread options.
Applied Mathematical Finance, 18(5), 447–472.
[MathSciNet][Crossref][zbMATH]

Footnotes
1 The story of how the strategy was originated and developed is described in Johnson (2022).

2 In this chapter, we ignore blending optionality of the storage asset, which allows the owner
to mix different grades of crude oil and to sell the blend at a premium by customizing it to the
rigorous specifications of a refinery.

3 Likewise, the cost of freight determines the strike price of the locational spread option on the
difference between oil prices in two different regions. Considine et al. (2022) relates the value
of such an option to inventories.
4 Formally, this connection follows from reduced-form mean-reverting models, the reference
to which are provided in Chap. 3. Less formally and more intuitively, one can see that faster
mean-reversion precludes the spread from deviating too far from its mean, which generally
results in volatility that declines with time to maturity.

5 For more background on spread options that arise in natural gas and power markets, see, for
example, Eydeland and Wolyniec (2003) and Swindle (2014).

6 Margrabe (1978) derived two-dimensional partial differential equation for the price of a
spread option and solved it for K = 0. Kirk (1995) proposed the extension for K ≠ 0 but did not
publish its derivation. This formula can be formally derived by applying the method of
perturbation introduced in Appendix C. The technical details, however, are rather cumbersome,
and given the limited usage of correlation-based pricing formulas in the oil market, we chose
not to present them.

7 Similar to (A.​10), the solution to the partial differential equation for the price of a spread
option can be represented as a double integral of the option payoff multiplied by the risk
neutral joint probability density for F1 and F2. Advanced methods for pricing spread options in
two-dimensional lognormal setting tend to focus on developing efficient integration techniques
and more advanced analytic approximations. See, for example, Shimko (1994), Dempster and
Hong (2002), Carmona and Durrleman (2003), and Venkatramanan and Alexander (2011). For
petroleum markets, a simple approximation (13.8) is generally sufficient.

8 The day when oil prices went negative is removed as neither percentage returns nor
standard correlation metrics can even be computed.

9 Technically, the volatility of the first leg is also not observable as a vanilla option expires
three days prior to futures but the corresponding CSO expires one day before futures. Traders
often make an additional ad hoc adjustment to the implied volatility of a vanilla option to
incorporate this effect.
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2023
I. Bouchouev, Virtual Barrels, Springer Texts in Business and Economics
https://doi.org/10.1007/978-3-031-36151-7_14
14. Epilogue
Ilia Bouchouev1
(1) Pentathlon Investments, LLC, Westfield, NJ, USA

14.1 The Roadmap for Energy Transition and


Virtual Commodities
As I finish writing this book, the world is actively engaged in the debate
about the energy transition. Yet very little attention is paid to the
transition of virtual energy in the market for energy derivatives, which,
by and large, sets the price for the energy that we all consume. I will
conclude the book by outlining the role that virtual barrels could play
in the energy transition and its impact on broader markets for virtual
commodities.
Regardless of the source of energy, the market will continue to
function as a complex dynamic system with multiple feedback loops
and bidirectional causality between its components. The entire energy-
trading ecosystem will still revolve around real options embedded in
physical assets, even if the nature of these assets and their ownership
takes a different form. These real options are the foundation for a self-
balancing mechanism that allows the entire system to adjust and
prevent prices from breaking through the system boundaries.
The nucleus of the energy ecosystem is an option on time, which is
provided by storage technology. As we have highlighted before, storage
buys time by shifting limited supplies from times of plenty to times of
relative scarcity. Without this technology, an energy market is unlikely
to exist. For the pioneers of oil trading of the 1860s the technology used
for storage was literally a dump. In the new world of the energy
transition, storage will take the form of a battery.
Batteries are expensive to build, and they are likely to be financed
and leased to professional traders who are in a better position to
monetize an embedded optionality. To pay for the lease, battery traders
will generate revenues by buying and storing power when the price is
low and discharging it when the price is high. In the jargon of virtual
barrels this is the strategy of delta hedging the real option. Thus, many
core concepts presented in the book, including the theory of storage,
the model of the squeeze, and the virtual storage warehouse strategy
with time spread options will find new applications in the business of
battery storage with only minimal modifications. The main change is
the speed of the market. While the owner of oil storage can afford to
trade rather infrequently, the value of the battery option is driven by
intraday volatility, and one must participate in the market continuously
throughout the day. Trading will inevitably become more data intensive,
creating an opportunity that the new generation of quants would most
certainly appreciate.
In the energy market, an option on time often goes side by side with
an option on location that allows traders to deliver energy from one
point to another. New locational spread options are already being
created in response to the needs of the physical energy transition. The
biggest beneficiary so far is the virtual market for natural gas, which is
on its way to become increasingly global, interconnected, and more
financialized. An important role in this transition is played by liquefied
natural gas (LNG) assets whose owners monetize the transportation
option by linking US natural gas with consumers in Europe and Asia.
Furthermore, the switching option held by power generators would
likely strengthen the link between the prices of power, natural gas, coal,
and emission credits. It will further expand the scope of quantitative
relative value trading, providing additional diversification to energy
stat-arb portfolios.
The growth potential for the energy derivatives market remains
highly significant. While this market is already larger than the market
for physical commodities, it is still tiny relative to the size of other
financial markets. The energy market represents no more than 0.1% of
the size of the global equity market, even though the energy share in
the world economy is undoubtedly more significant. If every financial
investor decides to allocate, for example, 5% of total assets to energy-
related strategies, then hypothetically, the market for energy
derivatives can grow nearly 50-fold. The larger the energy market, the
higher the need for a common benchmark. Even today, WTI and Brent
futures behave much more as financial benchmarks, setting the tone for
pricing physical barrels.
Does this mean that energy prices will disconnect more often from
the underlying physical market? This might be an incorrect question to
pose. From the technical pricing perspective, the two markets are
unlikely to diverge too far, as the price in the physical market is largely
determined by futures. A better question to ask is whether the price is
reflective of the prevalent state of physical supply and demand. While
the answer is debatable, fundamentals alone will not be able to explain
the price. The benchmark futures price will always reflect the combined
effect of supply and demand for physical and financial barrels, and
more often than not, the latter will dominate the former.
As the energy transition picks up its pace, the strategies of large
sovereign producers will change as well. The risk of ending up with
stranded oil assets is already forcing OPEC to change its long-term
goals. Its main priority has shifted from providing some market
stability to maximizing the value of its resources before oil demand
starts to wane. OPEC will attempt to exert some control with the goal of
setting price floors by collectively cutting production when the price is
perceived to be low. This strategy is inherently more volatile, requiring
dynamic decision making with higher frequency of OPEC meetings and
less predictable outcomes. The perception of such a price floor may
change the composition of market participants and their trading styles.
It may trim investor interest in using downside momentum strategies,
which may lead to an excess of buyers over sellers in the futures
market. This financial imbalance can be further exacerbated by
declining interest in hedging by independent producers, many of which
remain under investor pressure to divest fossil assets.
The inability to tame energy prices driven by financial markets will
undoubtedly frustrate policymakers of consumer nations, and one
should expect more frequent market interventions by their
governments. There will be more attempts to use strategic reserves,
price caps, subsidies, sanctions, tariffs, and other policy tools
specifically targeting energy prices. However, the track record of states
meddling with energy prices is very poor. Replacing a part of a dynamic
ecosystem with an artificial graft is unlikely to make the body of the
market any stronger. The system would lose its natural immunity and
become more susceptible to a complete collapse in the event of
unexpected exogenous shocks. Unfortunately, plenty of such episodes
have already been recorded. The way to keep the ecosystem under
control is to let it evolve organically under the power of market forces
and whenever possible adjust its permissible boundaries. For example,
excessive speculation can be better managed with tighter position
limits and more transparency around large positions, which will
dissuade traders from taking excessive risks.
Any transitionary period will inevitably bring more uncertainty and
volatility, as changing composition of market participants can distort
the supply and demand both for physical energy and for hedging
services. Since the entire derivatives ecosystem operates on a highly
leveraged basis, increasing volatility raises the cost of trading via more
stringent collateral requirements. This could force certain traders out
of the market, which, in turn, would adversely impact the market
liquidity, and lower liquidity leads to even higher volatility. Thus,
volatility begets more volatility.
The need to manage volatility brings us to the final piece in the
virtual energy transition: the growth and increasing sophistication of
the options market. The presence of real options, either developed by
new asset owners or implicitly set by regulators, will leave more
footprints in the market. Such real options will certainly be replicated
by quantitative traders in the derivatives market. The primary
motivation of these traders is to capture an arbitrage between the
actuarial value of optionality as it is perceived by end-users and
policymakers, and the volatility-based value determined by the cost of
dynamic replication.
The dynamics of energy options could start to resemble equity
options, where the demand for hedging services exceeds the supply,
and buyers are willing to pay up for insurance-like protection. In a less-
fossilized energy world, the supply of volatility provided by producer
hedgers could decline. As a result, the structural volatility risk premium
will likely return to the energy market. The other pillars of quantitative
option modeling, such as negative gamma, the volatility smile, its term
structure, and the parabolic nature of fat tails, could also play a larger
role across more diverse markets for energy options.
Similar quantitative concepts apply to metals and other commodity
markets which are widely expected to grow with the energy transition.
Industrial metals will play a critical role in the electrification of the
global economy. The markets for copper and aluminum are already
highly financialized, being predominantly driven by fluctuations in
global demand. Like oil, metals are also valued as an inflation hedge.
The markets for precious metals, such as gold and silver, continue to
compete with oil for the status of being a safe haven during selloffs
across broader financial markets. Finally, the development of corn-
based ethanol and biodiesel markets will further tie the agricultural
complex to energy. It will provide important diversification and a
unique seasonal component to broader portfolios of virtual
commodities.
To summarize, whether one likes it or not, commodity prices will be
heavily influenced by financial traders. The markets for raw
commodities have been studied for more than a century. The supply
and demand for physical commodities were much better at explaining
price behavior in the past. The present and, most importantly, the
future will be more impacted by the supply and demand for virtual
commodities traded in the derivatives market. This market requires a
different set of skills. While media and policymakers continue to
explain market drivers with a simplified narrative for a broader
audience, the computers that increasingly drive commodity prices are
unlikely to be able to hear them. They will continue to do what we have
attempted to describe in this book and a lot more that we cannot
pretend to know.
Further quantification and digitalization of commodity markets is
on its way. One cannot stop the market evolving from primitive towards
more complex ways of doing things. The transition towards more
quantitative commodity trading is no different from the technological
innovations that we are seeing in many other areas of our lives. The
virtual energy transition will inevitably be driven by big data and more
advanced quantitative trading methods. Virtual barrels describe only
the beginning of this quantitative transition; much more is yet to come
in the world of virtual commodities.
Appendix A: Diffusions and Probabilities
We assume that an asset price x follows the diffusion process of the
form
(A.1)
defined by drift μ(x, t) and the local volatility function σ(x, t). The
increment dz represents a normally distributed random variable with
zero mean and variance equal to dt.
A function G(x, t) that depends on a stochastic variable x also follows
the diffusion process which is specified by Itô’s lemma:

(A.2)

Itô’s lemma is the stochastic analogue of Taylor’s formula used in


the ordinary calculus, but it retains the additional second-order term.
This is because in the stochastic calculus (dx)2 is of the same order of
magnitude as dt. Itô’s lemma shows that the variable x and the function
G(x, t) are driven by the same source of uncertainty dz.
The Dirac delta function, or impulse function, centered at x0, is
informally defined as the abstract function that is equal to zero
everywhere, except for a single point x0, where its value is infinite:

(A.3)

More formally, the Dirac delta function can be understood as the


limit of normal probability density with the mean x0 and the standard
deviation σ → 0:

Since the integral of the probability density function over all


possible values of x is equal to one, the integral of the Dirac delta
function is also equal to one
(A.4)

When the Dirac delta function centered at x0 is multiplied by a


function f(x) and integrated over all possible values of x, then the
integral isolates the value of the function at the point x0:

(A.5)

Since (A.1) has a stochastic component, the behavior of variable x


can only be understood in a probabilistic sense. Let p(x, t; xT, T) denote
the probability density function generated by the process (A.1). It
describes the transition probability of the asset price reaching various
levels of xT at a future time T > t given that its price is equal to x at time
t. To find this probability density, one must solve the following Fokker-
Planck equation, which is also known as the Kolmogorov forward
equation:

(A.6)

The equation is written with respect to xT and it propagates forward


in time for T > t. The initial condition is given by the Dirac delta
function centered at a given asset price x at time T = t:
(A.7)
This initial condition means that the asset price at time t is certain,
and the entire probability mass is concentrated in a single point when
xT = x.
The same probability density function also describes the probability
of where the asset x was at an earlier time t, if its price at a later time
T > t is known to be xT. It satisfies the following Kolmogorov backward
equation, defined for t < T:

(A.8)
This equation is with respect to x and it moves backward in time for
t < T. The terminal boundary condition at t = T is given by the Dirac
delta function centered at x = xT:
(A.9)
In applied mathematics, Eqs. (A.6) and (A.8) are known,
respectively, as the forward and backward parabolic differential
equations that describe the diffusion of heat in a medium, and the
function p(x, t, xT, T) is called the fundamental solution to these
equations.
In this book, we are mainly interested in the backward Eq. (A.8),
which applies to the pricing of a financial derivative at time t if its
payoff is specified at a later time T > t. The function p(x, t, xT, T)
represents only one particular solution to the Eq. (A.8) with a very
special boundary condition given by the Dirac delta function. Any other
solution G(x, t) to (A.8) with a different boundary condition specified at
time T by

can be expressed in terms of its fundamental solution, or the


probability density function, as follows:

(A.10)

In other words, G(x, t) can be understood as the expected value of


the function g(x) when the random behavior of x is described by the
probability density function p(x, t, xT, T).
In Chap. 8, we show that if x represents the futures contract, then
the price of any financial derivative of x solves (A.8) with μ(x, t) = 0. The
derivatives price can then be expressed in the form of (A.10) with p(x, t,
xT, T) representing the so-called risk-neutral probability density that
corresponds to a modified diffusion process (A.1) for which the drift
term is eliminated.
We list three examples of probability density functions used
throughout the book. The first one corresponds to Arithmetic Brownian
Motion (ABM), which is defined by (A.1) with constant drift and
volatility coefficients:
(A.11)
The solution to (A.8) that corresponds to ABM is given by the
normal probability density:

(A.12)

The second important diffusion process is Geometric Brownian


Motion (GBM), where the drift and local volatility of the diffusion
process are assumed to be proportional to the random variable:
(A.13)
To find the corresponding probability density function, we define a
new variable

It follows from Itô’s lemma (A.2) that

This means that the logarithm of the variable x is normally


distributed with mean and variance . The corresponding
probability function that solves equations (A.6) and (A.8) is given by

(A.14)

Since the logarithm of the random variable is normally distributed,


(A.14) is called the lognormal probability density. It is defined only for
positive x and widely used for modeling financial assets, such as
equities, whose prices cannot be negative.
The third important diffusion is the mean-reverting (MR) or
Ornstein-Uhlenbeck process, which modifies ABM by allowing the drift
term to gravitate towards its long-term mean with speed of mean-
reversion k > 0:
(A.15)
Such a mean-reverting process is more suitable for modeling
commodity prices where mean-reversion is induced by the cyclicality of
supply and demand.
To derive the probability density for MR process (A.15), consider
the following shifted variable

which describes how x is pulled towards its equilibrium level , where


speed of mean-reversion is characterized by k.
Applying Itô’s lemma (A.2) to y(x, t), we obtain that

In other words, the shifted variable y(t) follows the ABM process,
where for brevity, we no longer explicitly show its dependence on x.
The drift of this process is zero, but its variance is reduced by the
time-dependent exponential factor. This factor can be eliminated by
switching to the new integrated-time variable defined by

(A.16)

Following the steps described in Appendix D for pricing options


with time-dependent volatility, one can reduce the Eq. (A.8) for MR
process (A.15) to the one with constant volatility in variables

Note that when t = T, then y(T) = xT and , which preserves the


boundary condition (A.9) in the new variables y and .
The solution to this equation, which is the probability density for
MR process (A.15), is then also given by the normal probability density

(A.17)

with shifted drift and reduced variance, where is defined by (A.16).

Appendix B: Option Pricing under Normal and


Lognormal Distributions
In Chap. 8, it is shown that the price for a financial derivative that
follows a stochastic process (A.1) satisfies the Kolmogorov backward
Eq. (A.8) with zero drift μ(F, t) = 0. Therefore, its solution is given by the
representation (A.10), where g(x) is the option payoff.
For ABM process (A.11) with constant volatility, the value of the call
option is, therefore, obtained by the following integral of the normal
probability density (A.12) with μA = 0 and the option payoff:

It can be shown that the integration results in the following


Bachelier pricing formula with constant normal volatility σA and the
time remaining to maturity τ = T − t is given by the Bachelier formula:
(B.1)
Here

represents the normalized option moneyness,

the time remaining to maturity, and N is the cumulative normal


distribution with zero mean and variance equal to one, whose density
is given by
(B.2)

The normal delta is the first derivative of the Bachelier formula with
respect to the futures price:

(B.3)

The normal gamma is the second derivatives of the Bachelier


formula with respect to the futures price:

(B.4)

The normal vega is the partial derivative of the Bachelier formula


with respect to the volatility σA:

(B.5)

The normal theta is the partial derivative of the Bachelier formula


with respect to time:

(B.6)

The derivatives pricing Eq. (8.​8) is satisfied as

Similarly, for GBM process (A.13) the pricing formula is given by the
integral of the payoff function with lognormal probability density with
μG = 0

The evaluation of this integral results in the following Black formula


with constant percentage volatility σG:
(B.7)

where

represents the normalized log-moneyness.1


The corresponding Black delta, gamma, vega, and theta are:

(B.8)

(B.9)

(B.10)

(B.11)

The following identity

was used in the calculation of lognormal partial derivatives. The pricing


equation

is satisfied. Partial derivatives for put options can be obtained from the
Greeks for the call option by differentiating the put-call parity
relationship (8.​12).
Appendix C: The Perturbation Method and the
Quadratic Normal Model
Let us assume that the local volatility function is defined by a small
perturbation ε(F) of the constant volatility σA
(C.1)
Then the partial differential Eq. (8.​8) for the option price is

(C.2)

We seek its solution C(F, t) as the sum of the Bachelier formula


CBC(F, t) that corresponds to constant normal volatility σA and the skew
correction function U(F, t):
(C.3)
To derive the equation for U(F, t), we substitute (C.3) into (C.2) and
regroup the terms as follows, where, for brevity, we suppress the
dependency of function ε on futures price F

The sum of the two terms grouped in the first parenthesis is zero
because CBC solves (C.2) with constant dollar volatility σA and ε = 0. The
terms grouped in the last parenthesis are of a higher order with respect
to ε. The method of perturbation or linearization assumes that for small
perturbation ε, these terms can be omitted.
For small ε, we are left with only the terms grouped in the middle
parenthesis, which leads to the following equation for U(F, t):

To simplify this equation, we let τ = T − t, and replace the second


derivative of the Bachelier formula, which is simply its gamma, with its
explicit formula (B.4). The equation for the skew correction then
becomes
(C.4)

We look for a solution to (C.4) of the following form

(C.5)

where p(F, τ) is a yet to be defined function. We evaluate partial


derivatives of U(F, τ) using the chain rule, suppressing arguments of
functions p and n:

and substitute them into (C.4). After cancelling the common multiplier
n and multiplying each term by , we arrive at the following
equation for the function p(F, τ)

(C.6)

We can now choose to specify the perturbation function to be a


quadratic of the form
(C.7)
In this case, the Eq. (C.6) can be solved by letting p(F, τ) be a
polynomial function:

where wi can be found by the method of undetermined coefficients.


Specifically, we equate the coefficients in the left-hand side and the
right-hand side of the Eq. (C.6) for polynomial terms of the same order,
which leads to the following system of linear equations:

This system is then inverted for coefficients wi as follows:

After some simple algebra, the polynomial function p(F, τ) simplifies


to

and representation (C.5) leads to the solution for the skew correction
function

(C.8)

Also, since both C(F, T) and CBC(F, T) at expiration are equal to the
option payoff, max(0, F − K), the skew correction function at time t = T
must be equal to zero. This boundary condition for the skew correction
function at τ = 0
is clearly satisfied by the formula (C.8).
If necessary, this method can be easily extended to a higher-order
polynomial volatility perturbation function, in which case p(F, τ) must
also be taken to be a polynomial of the same order with coefficients
determined by matching corresponding terms of the Eq. (C.6).

Appendix D: Option Pricing with Time-


Dependent Volatility
We consider the GBM process for the T-matuirty futures price F(t, T)
with time-dependent lognormal volatility σG(t, T) as in (11.​1). The
option price is then the solution of the following partial differential
equation:

(D.1)

In general, the option expires at time T0 ≤ T. The payoff of the call


option is given by

We introduce a new time variable , defined as

(D.2)

where v is constant.
Let represent the call option price using this new time
variable

We calculate the derivative with respect to time using the chain rule
The substitution of this partial derivative into Eq. (D.1) with time
variable volatility σG(t, T) replaces it with an identical equation but with
constant volatility v

Note that when t = T0 then so that the boundary condition


remains intact

The price of the call option is simply given by the Black formula
with respect to time and with constant volatility v:

It follows from (D.2) that the total variance in the Black formula in
the new variables is equal to the integrated local variance in the
original variables

Therefore, the option price C(F, t) is given by the same Black


formula with implied volatility v given by the quadratic mean of the
local volatility

The calculations above are identical for time-dependent normal


volatility, in which case the option price is given by the Bachelier
formula with volatility vN equal to the mean-square of the local
volatility
Appendix E: Average Price Options
Let the futures price follow the general diffusion process (A.1):
(E.1)
We consider the problem of pricing an APO within the averaging
period, i.e., Ta ≤ t ≤ T. The rolling average is defined as

By differentiating A(t) with respect to t, we obtain the stochastic


differential equation for the rolling average as

The option price CAPO(F, A, t) depends on two spatial variables, F


and A. The latter, however, does not have the stochastic term that
contains dz. The pricing equation for an APO follows from the two-
dimensional Itô’s lemma and the BSM hedging argument from Chap. 82:

(E.2)

Such two-dimensional partial differential equations are difficult to


solve analytically. However, the specific form of (E.2) allows for a
reduction of dimensionality using the auxiliary variable x

We make the change of variables

and evaluate partial derivatives using the chain rule, as follows:


Then the two-dimensional partial differential Eq. (E.2) for an APO
reduces to

The local volatility in this equation is adjusted by a linear multiplier


that reflects gradually decreasing volatility as the average accumulates
throughout the averaging period. Outside of the averaging period for
t > Ta, no adjustment is needed and the standard differential Eq. (8.​8)
with respect to t and F applies with the local volatility function σ(F, t, T).

Appendix F: The Inverse Diffusion Problem


To derive the Dupire equation, we use the fact that the risk-neutral
probability density p(F, t; K, T) satisfies the Fokker-Planck Eq. (A.6)
with zero drift with respect to variables K and T:

Using representation (12.​4), we replace the probability density


function with the second derivative of the call option payoff:

We then change the order of differentiation in the first term, and


integrate both terms twice with respect to K, which leads to the
following equation for the option price with respect to strikes K and
maturities T 3:

(F.1)
We next consider a more practical case, where the local volatility
only depends on the spatial variable and option prices are available for
the continuum of strikes but only for a fixed maturity T, which is the
case for the oil market.
Let τ = T − t represent the time remaining to maturity. Then the
Dupire Eq. (F.1) with time-independent volatility is

(F.2)

In applied mathematics, this is the equation of heat transfer in a


non-homogeneous medium, described by an unknown space-
dependent thermal conductivity σ(K).
The equation is supplemented with an initial condition, which is
defined by the payoff of a call option at maturity τ = 0:
(F.3)
The futures price F here plays the role of an exogenous parameter.
Furthermore, we assume that market prices of options with all
strikes C∗(K) are observed today at time t = 0, or when τ = T:
(F.4)
We can now formulate the following inverse problem:
The Inverse Problem with Final Over-Determination4: If the solution
to the boundary-value problem (F.2 and F.3) is measured at a given time
by (F.4), then is it possible to uniquely reconstruct the unknown
coefficient σ(K)?
This problem arises in many applications. For example, when
applied to the equation of heat transfer in a medium, it raises the
question whether it is possible to uniquely reconstruct unknown
conductivity properties of the medium from the measurement of the
temperature distribution across all points at a given time.
This problem can also be restated using the jargon of probabilities.
Let us consider the Eq. (F.2) but with a different initial condition, where
the call option payoff is replaced with the Dirac delta function
Then the solution represents the probability density function for a
diffusion process with unknown time-independent diffusion coefficient.
This means that the inverse problem can be restated in terms of
probability densities as follows:
The Inverse Diffusion Problem: Let p(F, 0; K, τ) represent the
diffusion probability of a particle being at the point K at time τ given its
starting point K = F at τ = 0. Is it possible to uniquely reconstruct the
time-independent diffusion coefficient σ(K) from a single observation
of the probability density at time T?5
The glossary summarizes key terms used by professional oil
derivatives traders. It reflects the jargon and conventions of the oil
market, as some terms may be used in different contexts in other
markets and in academic studies.

Glossary
Actuarial Valuation A method of pricing options based on their average
historical payoffs.
Arbitrage A strategy that involves buying and selling similar contracts
to capture discrepancies in their prices.
Average Price Option (APO) An option that settles based on the average
price of underlying futures contracts over a given period.
Backwardation A situation where the futures curves declines with
increasing time to maturity.
Barrel Counting Fundamental analysis of supply and demand.
Basis The spread between the price of a physical or a less liquid
financial contract and the benchmark futures price.
Bid-Ask Spread The spread between the price to buy (bid) and the price
to sell (ask).
Black Volatility Implied volatility computed using the Black (1976)
option pricing formula.
Butterfly Spread A three-legged option trade that buys options with
strikes K1 and K2 and sells twice as many options with the strike price
set half-way between K1 and K2, all with the same expiration.
Calendar Spread Option (CSO) An option on the spread between two
futures on the same commodity with different maturities.
Calibration A process of fitting volatility models to market prices of
traded options.
Carry The spread between futures on the same commodity with
different maturities, which is used to measure the expected price roll
up (down) the curve.
Cash Price The price of an over-the-counter contract with physical
delivery.
Cash-Settled Option An option whose payoff at expiration is calculated
financially as the difference between the futures settlement price and
the strike price.
Commodity Currencies Currencies of commodity-exporting countries.
Commitments of Traders (CoT) Report A weekly report on futures and
option positions held by different categories of market participants.
Commodity Trading Advisors (CTAs) In general, anyone who advises on
commodity trading, but in practice, the term is often used to describe
hedge funds that trade based on quantitative algorithms.
Contango A situation when the futures curve increases with time
remaining to maturity.
Convenience Yield Consumption benefits accrued to the owner of a
physical commodity, but not to the owner of a futures contract.
Crack Spread The spread between the price of a refined product and
the price of crude oil.
Cushing A reference to the storage hub and inventories at the delivery
location for WTI futures contract.
Digital (Binary) Option An option with a discontinuous payoff that pays
a fixed amount or zero depending on the futures price.
Denomination Effect An impact on oil price, resulting from oil being
denominated in US dollars (USD).
Excess Return (ER) A return on buying and rolling futures.
Fading the Crowded Trade A contrarian position taken when hedge
funds’ futures holdings are perceived to be excessive.
Follow the Flow Strategy A strategy of mimicking futures position held
by hedge funds.
Fractionation Analysis Decomposition of profit-and-loss (P&L) of a
systematic strategy by the value of a certain explanatory variable.
Gamma Hedging The process of option dealers rebalancing their
directional price exposure caused by shifts in the futures price.
Grades Various types of crude oil that trade as a differential to main
futures benchmarks, such as WTI and Brent.
GSCI Rolls The process of rolling futures between the fifth and the ninth
business days of each month by major commodity indices, such as
Goldman Sachs Commodity Index (GSCI).
Hacienda Hedge A reference to a large-scale hedging program by the
Government of Mexico.
Hedging Pressure An imbalance between buyers and sellers of a
particular futures or options contract.
Implied Volatility An input into an option pricing model that makes the
model price of an option to match its market price.
Local Volatility A volatility function that depends on futures prices and
time and characterizes uncertainty of the random variable in a
stochastic diffusion process.
Market-Implied Probability Distribution A probability distribution
reconstructed from market prices of options with various strikes and
the fixed expiration.
Mean-Reversion A tendency of prices to revert to a long-term
equilibrium.
Naked Option An option position (typically a sale) which is not delta
hedged.
Normal Backwardation (Contango) A situation where the futures price
is below (above) the expected spot price. The term is equivalent to a
risk premium.
Normal (Dollar) Volatility Implied volatility computed using the
Bachelier option pricing formula.
Quantamentals A trading style that combines rule-based decision-
making with discretionary overlay.
Penultimate Expiration An expiration of derivatives contracts one day
prior to the expiration of the futures.
Physically-Settled Option An option, which when it is in-the-money, is
exercised into futures at expiration.
Positioning A reference to positions held by hedge funds and other
large market participants.
Producer Collar (Fence) A derivatives structure where a producer buys
a put option and finances it by selling a call option of the same value.
Producer Three-Way Collar A derivatives structure where a producer
buys a put spread and finances it by selling a call option of the same
value.
Prompt (Nearby) Futures A futures contract with the shortest time to
expiration.
Reaction Function A function that maps the strength of the systematic
signal to the position size.
Realized Volatility A statistical measure of volatility calculated as the
annualized standard deviation of percentage changes in prices over a
given period.
Regime Change A structural shift in dominant market-driving factors.
Risk Parity An asset allocation framework that weighs assets inversely
to their volatility and invests in commodity futures as an inflation
hedge.
Roll Return (RR) The difference between the excess return (ER) and the
spot return (SR).
Signal Blending A combination of multiple systematic trading signals.
Skew Delta A correction to the hedging delta that results from the
change in option’s vega caused by the move in futures.
Spot Return (SR) A return on a hypothetical investment in a spot
futures contract with no convenience benefits or storage costs.
Spread Option An option on the difference between the prices of two
futures.
Squeeze A large price move caused by futures traders exiting positions
near the expiration of the futures contract.
Sticky Moneyness A heuristic rule that assumes that implied volatilities
for all options remain unchanged for a given moneyness.
Sticky Strike A heuristic rule that assumes that implied volatilities for
all options remain unchanged for a given strike.
Stock-Out A situation of zero inventories in storage.
Swap Data Repository (SDR) A registered entity that collects and
disseminates information about over-the-counter transactions on a
nearly real-time basis.
Synthetic Spread Option The spread between an option on a refined
product and an option on crude oil with the same expiration.
Synthetic Storage A portfolio of calendar spread options (CSOs)
designed to mimic cash flows of physical storage.
Swaption An option to enter into a swap.
Tank-Tops A situation where inventories reach the maximum storage
capacity.
Trading-at-Settlement (TAS) A futures contract that allows parties to
trade during a trading day at a settlement price which will only be
determined later after the market closes.
Underlying Futures A futures contract on which the price of an option
contract depends.
Vanilla Options Exchange-traded options, excluding spread options and
APOs.
Volatility Risk Premium An investment return of the systematic strategy
of selling and delta hedging short term options.
Volatility Skew (Smile) A graph of implied volatilities for options versus
their strikes or moneyness for the same underlying futures contract.
Volatility Targeting A strategy that adjusts a notional position size
required to bring expected portfolio volatility to a given target.
Volatility Term Structure A graph of implied volatilities versus time to
maturity.

References
Acharya, V. V., Lochstoer, L. A., & Ramadorai, T. (2013). Limits to arbitrage and hedging:
Evidence from commodity markets. Journal of Financial Economics, 109(2), 441–465.

Alexander, C. (2001). Market models. Wiley.

Alexander, C. (2008). Market risk analysis, Vol. III: Pricing, hedging and trading financial
instruments. Wiley.

Andersen, L. (2011). Option pricing with quadratic volatility: A revisit. Finance and Stochastics,
15(2), 191–219.
[MathSciNet][zbMATH]

Ashton, M., & Greer, R. (2008). History of commodities as the original real return asset class. In
Inflation risk and products (pp. 85–109). Risk Books.

Avellaneda, M., Friedman, C., Holmes, R., & Samperi, D. (1997). Calibrating volatility surfaces via
relative entropy minimization. Applied Mathematical Finance, 4(1), 37–64.
[zbMATH]

Bachelier, L. (1900). Théorie de la Spéculation. Annales scientifiques de l’Êcole Normale


Supêrieure, Serie 3, 17, 21–86.
[MathSciNet][zbMATH]
Backhouse, R. E. (2002). The ordinary business of life. Princeton University Press.

Baker, S. D. (2021). The financialization of storable commodities. Management Science, 67(1),


471–499.

Bakshi, G., Gao, X., & Rossi, A. G. (2019). Understanding the sources of risk underlying the cross
section of commodity returns. Management Science, 65(2), 619–641.

Barone-Adesi, G., & Whaley, R. E. (1987). Efficient analytic approximation of American option
values. Journal of Finance, 42(2), 301–320.

Black, F. (1976). The pricing of commodity contracts. Journal of Financial Economics, 3(1/2),
167–179.

Black, F., & Scholes, M. (1973). The pricing of options and corporate liabilities. The Journal of
Political Economy, 81(3), 637–654.
[MathSciNet][zbMATH]

Blas, J. (2017, April 4). Uncovering the secret history of Wall Street’s largest oil trade,
Bloomberg.

Bodie, Z., & Rosansky, V. I. (1980, May–June). Risk and return in commodity futures. Financial
Analysts Journal, 36(3), 27–39.

Boons, M., & Prado, M. P. (2019). Basis-momentum. The Journal of Finance, 74(1), 239–279.

Bouchouev, I. (1998). Derivatives valuation for general diffusion processes, Proceedings of the
Annual Conference of the International Association of Financial Engineers, New York, USA, pp.
91–104.

Bouchouev, I. (2000a, July). Demystifying Asian options. Energy and Power Risk Management,
pp. 26–27.

Bouchouev, I. (2000b, August). Black-Scholes with a smile. Energy and Power Risk Management,
pp. 28–29.

Bouchouev, I. (2012). Inconvenience yield, or the theory of normal contango. Quantitative


Finance, 12(12), 1773–1777.

Bouchouev, I. (2020a, April 30). Negative oil prices put spotlight on investors, Risk.net.

Bouchouev, I. (2020b). From risk bearing to propheteering. Quantitative Finance, 20(6), 887–
894.
[MathSciNet]

Bouchouev, I. (2021). A stylized model of the oil squeeze, SSRN.

Bouchouev, I. (2022). The Strategic Petroleum Reserve strategies: Risk-free return or return-free
risk? The Oxford Institute for Energy Studies.
Bouchouev, I. (2023). Testimony to the US House Subcommittee on Economic Growth, Energy
Policy, and Regulatory Affairs of the Committee on Oversight and Accountability, March 8.
Reprinted in Commodity Insights Digest (2023), Vol. 1.

Bouchouev, I., & Isakov, V. (1997). The inverse problem of option pricing. Inverse Problems,
13(5), L11–L17.
[MathSciNet][zbMATH]

Bouchouev, I., & Isakov, V. (1999). Uniqueness, stability and numerical methods for the inverse
problem that arises in financial markets. Inverse Problems, 15(3), R95–R116.
[MathSciNet][zbMATH]

Bouchouev, I., Isakov, V., & Valdivia, N. (2002). Recovery of volatility coefficient by linearization.
Quantitative Finance, 2(4), 257–263.
[MathSciNet][zbMATH]

Bouchouev, I., & Johnson, B. (2022). The volatility risk premium in the oil market. Quantitative
Finance, 22(8), 1561–1578.
[MathSciNet][zbMATH]

Bouchouev, I., & Zuo, L. (2020, Winter). Oil risk premia under changing regimes. Global
Commodities Applied Research Digest, 5(2), 49–59.

Boyle, P. C. (1899, September 6). Testimony at the Hearing of the US Industrial Commission on
Trusts and Industrial Combinations.

Breeden, D. T., & Litzenberger, R. H. (1978). Prices of state contingent claims implicit in option
prices. Journal of Business, 51(4), 621–651.

Brennan, M. J. (1958). The supply of storage. The American Economic Review, 48(1), 50–72.

Brennan, M. J. (1991). The price of convenience and the valuation of commodity contingent
claims. In D. Lund & B. Oksendal (Eds.), Stochastic models and option values. North Holland.

Brennan, M. J., & Schwartz, E. S. (1985). Evaluating natural resource investments. Journal of
Business, 58(2), 135–157.

Bü yü kşahin, B., & Robe, M. A. (2014). Speculators, commodities and cross-market linkages.
Journal of International Money and Finance, 42, 48–70.

Carmona, R., & Durrleman, V. (2003). Pricing and hedging spread options. SIAM Review, 45(4),
627–685.
[MathSciNet][zbMATH]

Carmona, R., & Ludkovski, M. (2004). Spot convenience yield models for the energy markets.
Contemporary Mathematics, 351, 65–79.
[MathSciNet][zbMATH]

Carr, P., Fisher, T., & Ruf, J. (2013). Why are quadratic normal volatility models analytically
tractable? SIAM Journal on Financial Mathematics, 4(1), 185–202.
[MathSciNet][zbMATH]
Carr, P., & Madan, D. (2001). Determining volatility surfaces and option values from an implied
volatility smile. In M. Avellaneda (Ed.), Quantitative analysis in financial markets (Vol. II, pp.
163–191). World Scientific.

Casassus, J., & Collin-Dufresne, P. (2005). Stochastic convenience yield implied from commodity
futures and interest rates. The Journal of Finance, 60(5), 2283–2331.

Castagna, A., & Mercurio, F. (2007). The vanna-volga method for implied volatilities. Risk, 20(1),
106–111.

Cheng, I.-H., Kirilenko, A., & Xiong, W. (2015). Convective risk flows in commodity futures
markets. Review of Finance, 19(5), 1733–1781.

Chiarella, C., Craddock, M., & El-Hassan, N. (2003). An implementation of Bouchouev’s method
for short time calibration of option pricing models. Computational Economics, 22, 113–138.
[zbMATH]

Clewlow, L., & Strickland, C. (2000). Energy derivatives: Pricing and risk management. Lacima
Publications.

Considine, J., Galkin, P., & Aldayel, A. (2022). Inventories and the term structure of oil prices: A
complex relationship. Resources Policy, 77, 1–18.

Daskalaki, C., Kostakis, A., & Skiadopoulos, G. (2014). Are there common factors in individual
commodity futures returns? Journal of Banking and Finance, 40, 346–363.

Deaton, A., & Laroque, G. (1992). On the behavior of commodity prices. The Review of Economic
Studies, 59(1), 1–23.
[zbMATH]

Dempster, M. A. H., & Hong, S. S. G. (2002). Spread option valuation and the fast Fourier
transform. In H. Geman, D. Madan, S. R. Pliska, & T. Vorst (Eds.), Mathematical finance, Bachelier
congress (Vol. 1, pp. 203–220). Springer.

Dempster, M. A. H., Medova, E., & Tang, K. (2012). Determinants of oil futures prices and
convenience yields. Quantitative Finance, 12(12), 1795–1809.
[MathSciNet][zbMATH]

Dempster, M. A. H., & Richards, D. G. (2000). Pricing American options fitting the smile.
Mathematical Finance, 10(2), 157–177.
[MathSciNet][zbMATH]

Derman, E., & Miller, M. B. (2016). The volatility smile. Wiley.

Doran, J. S., & Ronn, E. I. (2006). The bias in Black-Scholes/Black implied volatility: An analysis
of equity and energy markets. Review of Derivatives Research, 8(3), 177–198.
[zbMATH]

Doran, J. S., & Ronn, E. I. (2008). Computing the market price of volatility risk in the energy
commodity markets. Journal of Banking and Finance, 32(12), 2541–2552.
Dupire, B. (1994). Pricing with a smile. Risk, 7(1), 18–20.

Dvir, E., & Rogoff, K. (2009). Three epochs of oil, NBER Working Paper, 14927.

Ederington, L. H., Fernando, C. S., Holland, K. V., Lee, T. K., & Linn, S. C. (2021). The dynamics of
arbitrage. Journal of Financial and Quantitative Analysis, 56(4), 1350–1380.

Ellwanger, R. (2017). On the tail risk premium in the oil market, Bank of Canada Working Paper,
46.

Erb, C. B., & Harvey, C. R. (2006). The strategic and tactical value of commodity futures.
Financial Analysts Journal, 62(2), 69–97.

Eydeland, A., & Wolyniec, K. (2003). Energy and power risk management: New developments in
modeling, pricing, and hedging. Wiley.

Fama, E. F., & French, K. R. (1987). Commodity futures prices: Some evidence on forecast power,
premiums, and the theory of storage. Journal of Business, 60(1), 55–73.

Fattouh, B. (2011). An anatomy of the crude oil pricing system. Oxford Institute for Energy
Studies, Working Paper, 40.

Fattouh, B., & Mahadeva, L. (2014). Causes and implications of shifts in financial participation in
commodity markets. The Journal of Futures Market, 34(8), 757–787.

Fernandez-Perez, A., Frijns, B., Fuertes, A.-M., & Miffre, J. (2018). The skewness of commodity
futures returns. The Journal of Banking and Finance, 86, 143–158.

Fernandez-Perez, A., Fuertes, A.-M., & Miffre, J. (2021, Summer). On the negative pricing of WTI
crude oil futures. Global Commodities Applied Research Digest, 6(1), 36–43.

Fisher, I. (1896). Appreciation and interest. Publications of the American Economic Association,
11(4), 331–442.

Gatheral, J. (2006). The volatility surface: A Practitioner’s guide. Wiley.

Geman, H. (2005). Commodities and commodity derivatives: Modeling and pricing for
agriculturals, metals and energy. Wiley.

Gibson, R., & Schwartz, E. S. (1990). Stochastic convenience yield and the pricing of oil
contingent claims. The Journal of Finance, 45(3), 959–976.

Giddens, P. H. (1947). Pennsylvania petroleum 1750–1872: A documentary history. Pennsylvania


Historical and Museum Commission.

Gorton, G. B., Hayashi, F., & Rouwenhorst, K. G. (2012). The fundamentals of commodity futures
returns. Review of Finance, 17(1), 35–105.

Gorton, G., & Rouwenhorst, K. G. (2006). Facts and fantasies about commodity futures. Financial
Analysts Journal, 62(2), 47–68.
Greer, R. J. (1978, Summer). Conservative commodities: A key inflation hedge. Journal of
Portfolio Management, 4(4), 26–29.

Grunspan, C. (2011). A note on the equivalence between the normal and the lognormal implied
volatility: A model free approach, SSRN.

Gustafson, R. L. (1958). Carryover levels for grains, U.S. Department of Agriculture, Technical
Bulletin, 1178.

Hamilton, J. D. (1983). Oil and the macroeconomy since World War II. Journal of Political
Economy, 91(2), 228–248.

Hamilton, J. D. (2003). What is an oil shock? Journal of Econometrics, 113(2), 363–398.


[MathSciNet][zbMATH]

Hamilton, J. D. (2009). Understanding crude oil prices. The Energy Journal, 30(2), 179–206.

Hamilton, J. D., & Wu, J. C. (2014). Risk premia in crude oil futures prices. Journal of
International Money and Finance, 42, 9–37.

Hicks, J. R. (1939). Value and capital: An inquiry into some fundamental principles of economic
theory. Oxford University Press.

Hirshleifer, D. (1988). Residual risk, trading costs, and commodity futures risk premia. The
Review of Financial Studies, 1(2), 173–193.
[MathSciNet]

Hull, J. C. (2018). Options, futures, and other derivatives (10th ed.). Pearson.
[zbMATH]

Imsirovic, A. (2021). Trading and price discovery for crude oils: Growth and development of
international oil markets. Palgrave Macmillan.

Ingersoll, J. E., Jr. (1997). Valuing foreign exchange rate derivatives with a bounded exchange
process. Review of Derivatives Research, 1, 159–181.
[zbMATH]

Interim Stuff Report. (2020). Trading in NYMEX WTI crude oil futures contract leading up to, on,
and around April 20, 2020, Commodity Futures Trading Commission, November 23.

Jacobs, K., & Li, B. (2023). Option returns, risk premiums, and demand pressure in energy
markets. Journal of Banking and Finance, 146, 1–26.

Javaheri, A. (2005). Inside volatility arbitrage: The secrets of skewness. Wiley.

Johnson, O. (2022). 40 classic crude oil trades: Real-life examples of innovative trading.
Routledge.

Kaldor, N. (1939). Speculation and economic stability. The Review of Economic Studies, 7(1), 1–
27.
Kang, S. B., & Pan, X. (2015). Commodity variance risk premia and expected futures returns:
Evidence from the crude oil market, SSRN.

Kang, W., Rouwenhorst, K. G., & Tang, K. (2020). A tale of two premiums: The role of hedgers
and speculators in commodity futures markets. The Journal of Finance, 75(1), 377–417.

Kemna, A. G. Z., & Vorst, A. C. F. (1990). A pricing method for options based on average asset
values. Journal of Banking and Finance, 14(1), 113–129.

Keynes, J. M. (1923). Some aspects of commodity markets. The Manchester Guardian


Commercial, Reconstruction Supplement, March 29.

Keynes, J. M. (1930). A treatise on money (Vol. II). Macmillan.

Keynes, J. M. (1936). The general theory of employment, interest, and money. Macmillan.

Kilian, L. (2009). Not all oil price shocks are alike: Disentangling demand and supply shocks in
the crude oil market. American Economic Review, 99(3), 1053–1069.

Kilian, L. (2020). Understanding the estimation of oil demand and oil supply elasticities, Federal
Reserve Bank of Dallas Working Paper, 2027.

Kilian, L., & Murphy, D. P. (2014). The role of inventories and speculative trading in the global
market for crude oil. Journal of Applied Econometrics, 29(3), 454–478.
[MathSciNet]

Kilian, L., & Vigfussion, R. J. (2017). The role of oil price shocks in causing U.S. recessions.
Journal of Money, Credit and Banking, 49 (8), 1747–1776.

Kirk, E. (1995). Correlation in the energy markets. In Managing energy price risk (pp. 71–78).
Risk Publications.

Koijen, R. S. J., Moskovitz, T. J., Pedersen, L. H., & Vrugt, E. B. (2018). Carry. Journal of Financial
Economics, 127, 197–225.

Leoni, P. (2014). The Greeks and hedging explained. Palgrave Macmillan.

Levy, E. (1992). Pricing European average rate currency options. Journal of International Money
and Finance, 11(5), 474–491.

Lipton, A. (2001). Mathematical methods for foreign exchange: A financial engineer’s approach.
World Scientific.
[zbMATH]

Lipton, A., & Sepp, A. (2011, October). Filling the gaps. Risk, 24(10), 78–83.

Lo, A. W. (2002). The statistics of Sharpe ratios. Financial Analysts Journal, 58(4), 36–52.

Lux, H. (2003, February 1). What becomes a legend? Institutional Investor.

Ma, L. (2022). Negative WTI price: What really happened and what can we learn? The Journal of
Derivatives, 29(3), 9–29.
[MathSciNet]

Margrabe, W. (1978). The value of an option to exchange one asset for another. The Journal of
Finance, 33(1), 177–186.

Merton, R. C. (1973). Theory of rational option pricing. The Bell Journal of Economics and
Management Science, 4(1), 141–183.
[MathSciNet][zbMATH]

Miffre, J. (2016). Long-short commodity investing: A review of the literature. Journal of


Commodity Markets, 1(1), 3–13.

Miltersen, K. R. (2003). Commodity price modelling that matches current observables: A new
approach. Quantitative Finance, 3(1), 51–58.
[MathSciNet][zbMATH]

Naldi, N. (2015). Sraffa and Keynes on the concept of commodity rates of interest. Contributions
to Political Economy, 34(1), 17–30.

Neville, H., Draaisma, T., Funnell, B., Harvey, C. R., & Van Hemert, O. (2021). The best strategies
for inflationary times, SSRN.

Pilipović, D. (2007). Energy risk: Valuing and managing energy derivatives. McGraw-Hill.

Pirrong, C. (2012). Commodity price dynamics: A structural approach. Cambridge University


Press.

Prokopczuk, M., Symeonidis, L., & Simen, C. W. (2017). Variance risk in commodity markets.
Journal of Banking and Finance, 81, 136–149.

Rebonato, R. (2004). Volatility and correlation: The perfect hedger and the fox. Wiley.

Routledge, B. R., Seppi, D. J., & Pratt, C. S. (2000). Equilibrium forward curves for commodities.
The Journal of Finance, 55(3), 1297–1338.

Samuelson, P. A. (1965). Proof that properly anticipated prices fluctuate randomly. Industrial
Management Review, 6(2), 41–49.

Schachermayer, W., & Teichmann, J. (2008). How close are the option pricing formulas of
Bachelier and Black-Merton-Scholes? Mathematical Finance, 18(1), 155–170.
[MathSciNet][zbMATH]

Schwartz, E. S. (1997). The stochastic behavior of commodity prices: Implications for valuation
and hedging. The Journal of Finance, 52(3), 923–973.

Shimko, D. C. (1994). Options on futures spreads: Hedging, speculation, and valuation. The
Journal of Futures Markets, 14(2), 183–213.

Smiley, A. W. (1907). A few scraps: Oily and otherwise. The Derrick Publishing Company.

Sraffa, P. (1932). Dr. Hayek on money and capital. The Economic Journal, 42 (165), 42–53.
Stoll, H. R. (1979). Commodity futures and spot price determination and hedging in capital
market equilibrium. The Journal of Financial and Quantitative Analysis, 14(4), 873–894.

Swindle, G. (2014). Valuation and risk management in energy markets. Cambridge University
Press.

Szymanowska, M., De Roon, F., Nijman, T., & Van Den Goorbergh, R. (2014). An anatomy of
commodity futures risk premia. The Journal of Finance, 69(1), 453–482.

Tang, K., & Xiong, W. (2012). Index investment and the financialization of commodities.
Financial Analysts Journal, 68(6), 54–74.
[MathSciNet]

Till, H. (2022, Winter). Commodities, crude oil, and diversified portfolios. Global Commodities
Applied Research Digest, 7(2), 65–74.

Till, H., & Eagleeye, J. (Eds.). (2007). Intelligent commodity investing. Risk Books.

Trolle, A. B., & Schwartz, E. S. (2010, Spring). Variance risk premia in energy commodities. The
Journal of Derivatives, 17(3), 15–32.

Tully, S. (1981, February 9). Princeton’s rich commodity scholars, Fortune.

Turnbull, S. M., & Wakeman, L. M. (1991). A quick algorithm for pricing European average price
options. Journal of Financial and Quantitative Analysis, 26(3), 377–389.

Venkatramanan, A., & Alexander, C. (2011). Closed form approximations for spread options.
Applied Mathematical Finance, 18(5), 447–472.
[MathSciNet][zbMATH]

Weymar, F. H. (1965). The dynamics of the world cocoa market. Ph.D. Thesis, Massachusetts
Institute of Technology.

Whiteshot, C. A. (1905). The oil-well driller: A history of the world’s greatest enterprise, the oil
industry. Acme Publishing Company.

Williams, J. C., & Wright, B. D. (1991). Storage and commodity markets. Cambridge University
Press.

Wilmott, P., Dewynne, J., & Howison, S. (1993). Option pricing: Mathematical models and
computation. Oxford Financial Press.
[zbMATH]

Working, H. (1948). Theory of the inverse carrying charge in futures markets. Journal of Farm
Economics, 30(1), 1–28.

Working, H. (1949). The theory of price of storage. The American Economic Review, 39(6), 1254–
1262.

Zü hlsdorff, C. (2001). The pricing of derivatives on assets with quadratic volatility. Applied
Mathematics Finance, 8(4), 235–262.
[zbMATH]

Index
A
Actuarial valuation 188, 331
Agricultural markets 30, 56, 64, 65, 129, 155, 308
Airline hedging 65, 234
All Weather strategy 74
Aluminum 308
American options 174
Arbitrage
model 214, 261
paper 118, 120, 121
statistical 5, 109, 122–124, 139, 148, 306
triangular correlation 7, 281, 287–291
volatility 230, 257
Aristotle 64, 161
Arithmetic Brownian motion (ABM) 40, 165, 171, 238, 293, 311
Asian options 234
At-the-money (ATM) options 162, 163, 172, 183, 188–190, 194, 199,
200, 203–205, 207, 214–225, 289–291, 298, 300
Autocorrelation 151
Availability 34–37
Average price options (APOs) 7, 233–237, 241–247, 325–326, 331
B
Bachelier formula 6, 171, 222, 227, 295, 315
Bachelier, L. 6, 162
Backtesting 85
Backwardation 15–17, 57, 58, 98, 103, 110–113, 331
Bank of China 52
Barone-Adesi, G. 174
Barrel counting 117, 331
Basis risk 62
Basis spread 2, 126, 292, 331
Batteries 282, 305
Beta hedge ratio 62, 77, 144, 146
Biodiesel 308
Black, F. 164, 169
Black formula 173, 222, 316
Black-Scholes-Merton (BSM) model 6, 161, 163, 169, 171, 193, 273
Black volatility 181, 192, 245, 331
Bootstrapping 7, 257, 261–268
Breakout strategy 85
Brent 3, 117, 118, 123–125, 130, 132, 145, 174, 236, 287, 293, 295,
306
Bridgewater Associates 74
Brownian motion 40, 162
Bureau of Labor Statistics (BLS) 149
Butane 127
Butterfly spread 271, 272, 331
C
Calendar spread options (CSOs) 7, 281–286, 293, 299–301, 331
Calibration 7, 257, 259–279
Call options 162
Call spreads 270, 271, 274
Canadian dollar (CAD) 143–145, 156, 157
Capital Asset Pricing Model (CAPM) 69, 74
Carry 5, 83, 86, 97–101, 103–106, 110, 111, 114, 118, 128, 331
Carry-momentum 101, 102, 106
Carry trade 5, 12, 19–21, 30, 35, 36, 42, 43
Causality 132, 137, 139, 141, 143, 148, 305
Chapman-Kolmogorov equation 163
Chicago Mercantile Exchange (CME) 4, 125, 130, 131
Chilean peso (CLP) 145
Coal 127, 306
Coefficient of instability 163
Cointegration 122, 123, 139, 143–146, 148, 156, 157
Collateral 24, 65, 67, 170, 174, 211, 234, 255, 291, 307
Commitments of Traders (CoT) report 129, 331
Commodities Corp. 84
Commodity currencies 143, 145, 148, 331
Commodity Futures Trading Commission (CFTC) 131
Commodity indices 67, 68
Commodity own rate of interest 5, 14, 22, 23, 26
Commodity terms of trade (CToT) 139–146
Commodity trading advisors (CTAs) 5, 84, 332
Concave backwardation 112, 113
Concave contango 103, 104, 110–113
Consumer hedging 57, 62–64, 77, 130, 234
Consumer Price Index (CPI) 71, 73, 149–152, 154, 157
Contango 15–17, 21, 26, 58, 75, 76, 98, 103, 107, 110–116, 282, 286,
332
Convenience yield 5, 11, 19, 22, 23, 25, 31, 32, 41, 76, 97, 98, 100, 103,
110, 117, 332
Convex backwardation 103, 110–112
Convex contango 112
Convexity 103, 110, 162, 172, 186, 191
Cootner, P. 84
Copper 11, 145, 308
Corn 13, 167, 308
Correlation
frown 297, 301
implied 295, 297, 301
realized 299
Costless collars 211, 212, 229, 276
Crack options 287, 290–292
Crack spreads 287, 292, 332
Credit risks 212, 214, 234, 255, 291
Crisis alpha 85
Cross-sectional momentum 87
Cushing 46, 114–117, 119, 124, 125, 127, 332
D
Dalio, R. 74
Dated Brent 125, 236
Delta 169, 172, 183, 184, 189–196, 202–206, 208, 216, 218–220, 225,
226, 229, 259, 271, 279, 295, 315–317
Delta hedging 6, 169, 183, 187, 189–196, 202–206, 218, 259, 273, 281,
284–288, 294, 296, 305
Denomination effect 140, 143, 332
Derivatives market 1, 3, 4
Diesel 131, 149, 152, 154, 288, 290–292, 294
Diffusion process 6, 7, 161–171, 227, 228, 257–259, 276–278, 309, 328
Digital options 270, 271, 332
Dirac delta function 41, 43, 44, 268–272, 309, 310, 328
Direct problems 257
Disaggregated reports 129
Domestic sweet oil (DSW) 125
Dubai oil 125
Dump men 20
Dupire equation 273–275, 327
Dynamic systems 1, 3, 29, 30, 38, 105, 137–139, 148, 305
E
Early exercise premium 174
Early expiry options (EEOs) 233, 247–252, 300
Einstein, A. 163
Emerging markets 138
Emission credits 127, 306
Energy Information Administration (EIA) 4, 47
Energy Policy and Conservation Act (EPCA) 16
Energy transition 282, 305–308
Ethane 127
Ethanol 308
Euclidean distance 134
Euro (EUR) 143
Eurobob 126
European options 174
Excess return (ER) 24, 25, 65, 66, 332
Exchange-traded options 174, 233
Expandable options 255
Extendable options 254
F
Fading extreme positioning 133
Fading the crowded trade 133, 332
Fat tails 167, 182, 209, 225, 226, 228, 293, 307
Federal Reserve Economic Data (FRED) 4
Feedback loops 1, 3, 5, 29, 30, 35, 52, 137, 138, 305
Fence 211, 333
Financialization 5, 55, 64–68, 74, 78, 113, 114, 141, 206, 221, 223
Finite difference methods 171, 227
Fisher, I. 5, 12, 97
Fisher inflation law 5, 11, 12, 16
Floating storage 21, 50
Fokker-Planck equation 40, 163, 273, 310, 327
Follow the flow strategy 132, 332
Fourier equation 163
Fractionation analysis 5, 113, 114, 116, 122, 128, 332
Freight 120, 121, 282, 293
Fuel oil 126
Fundamental solution 268, 310
Futures margin requirements 66
G
Gamma 6, 172, 187, 191–196, 199, 200, 207, 215, 218, 225, 228, 307,
316, 317, 320, 332
Gasoil 126, 131, 294
Gasoline 64, 68, 71, 89, 126–128, 137, 148–154, 292, 294
Geometric Brownian motion (GBM) 166, 311
Gold 11–13, 73, 140, 167, 308
Goldman Sachs Commodity Index (GSCI) 67, 332
H
Hacienda hedge 233–237, 332
Harmonic mean 64
Hayek, F. 14
Heat equation 161, 171, 258, 310, 328
Heat rate 294
Heaviside function 269
Hedging
airline 65, 234
consumer 57, 62–64, 77, 130, 234
inflation 76
Mexico sovereign 7, 210, 235–237
pressure 4, 5, 55–64, 76–78, 112, 332
producer 57, 60–67, 209–214, 234
Hicks, J.R. 57
High-Sulfur Fuel Oil (HSFO) 236
I
Ill-posed problems 260
Implied Black volatility (IBV) 181, 221–225
Implied correlation 295, 297, 301
Implied normal volatility (INV) 181, 222–224
Implied volatility 6, 161, 164, 180–186, 215–226, 239–241, 261–268,
276, 277, 332
Impulse function 269, 309
Incomplete markets 257, 267
Inconvenience yield 76
Inflation 11–14, 16, 55, 69–78, 137–139, 148–154, 308
Inflation base effect 150
Inflation breakeven rate 72
Inflation hedging 76
Inflation pass-through 71, 138, 141, 148–154
Inflation swaps 139, 148–154
Inflection point 105, 106, 116, 133, 134, 139
Information ratio 91
Intercontinental Exchange (ICE) 4, 125, 130
In-the-money (ITM) options 162
Inventories 29–53, 83, 87, 98, 100, 102, 103, 114–117
Inverse demand function 33–36, 38
Inverse diffusion problem 327–328
Inverse problem of option pricing 7, 257–260
Inverse problems 258–260
Itô’s lemma 168, 309
J
Jet fuel 64, 126, 288
K
Kaldor, N. 22
Keynes, J.M. 11, 14, 56–60
K-factor 236
Kirk approximation 296
Kolmogorov equations 40, 274, 310
Kurtosis 6, 190, 214, 276
L
Law of cosines 289
Law of radiation of probability 163
Linearization methods 227, 319
Liquefied natural gas (LNG) 127, 145, 306
Liquidity preference theory of money 23
Local volatility 6, 161, 164, 167, 171, 172, 175, 177, 181, 185, 226–
230, 237–255, 261, 263–267, 273–279, 285, 292–294, 300, 301, 309,
319, 326, 327, 332
Lognormal distribution 166, 167
Louisiana Light Sweet (LLS) 236
M
Macro fair-value model 155–158
Managed money (MM) 131
Margrabe formula 296, 297
Market-implied diffusion 7
Market-implied probability distribution 7, 257, 268–273, 332
Mars oil 125
Maya oil 235
Mean-reversion 39, 40, 44, 45, 83, 101–104, 107, 109, 118, 120–123,
127, 132, 138, 247, 254, 293, 312, 332
Medium of exchange channel 140
Mental models 1, 4, 5
Merton, R.C. 164, 169, 238
Metals markets 145, 307
Model arbitrage 214, 261
Model of the squeeze 5, 38–46, 305
Momentum 5, 83–97, 132, 133, 138
Momentum smile 95
Monte Carlo simulation 168, 175, 255, 259, 276, 278
Multi-factor models 252–255
N
Naphtha 127
National Bureau of Economic Research (NBER) 69
Natural gas 68, 99, 127, 131, 148, 168, 219, 239, 254, 265, 294, 296,
306
Natural gas liquids (NGLs) 123, 127
Negative oil prices 5, 49–53
Normal backwardation 5, 55–60, 63–68, 74, 75, 85, 206, 332
Normal contango 5, 55, 75–78, 113, 332
Normal distribution 40, 164–167, 171, 225, 311
Normal volatility 165
Norwegian krone (NOK) 145
Numeraire effect 140
O
Oil
busts 177
grades 124, 332
heavy 125
light 125
loans 16–18
own rate of interest 15–19
sour 125
swap 249, 250
sweet 125
Open interest 86, 124, 233, 302
Options
American 174
Asian 234
at-the-money (ATM) 162, 163, 172, 183, 188–190, 194, 199, 200,
203–205, 207, 214–225, 289–291, 298, 300
average price (APOs) 7, 233–237, 241–247, 325–326, 331
calendar spread (CSOs) 7, 281–286, 293, 299–301, 331
call 162
crack 287, 290–292
digital 270, 271, 332
early expiry (EEOs) 233, 247–252, 300
European 174
exchange-traded 174, 233
expandable 255
extendable 254
Greeks 172
as insurance 187–191, 197, 198
in-the-money (ITM) 162
moneyness 172, 183, 315, 316
out-of-the-money (OTM) 162
over-the-counter (OTC) 7, 233, 234, 237, 247, 249, 252, 254, 255,
291
put 162
real 2–3, 6, 109, 118, 123, 125, 210–213, 282, 287, 305, 307
simple 163, 172
spread 2, 118, 281–302, 333
synthetic spread 288, 333
vanilla 233, 334
Organization of Arab Petroleum Exporting Countries (OAPEC) 16, 70,
140
Organization of the Petroleum Exporting Countries (OPEC) 17, 47, 100,
157, 177, 264, 291, 306
Ornstein-Uhlenbeck process 40, 312
Other reportables (OTH) 131
Out-of-the-money (OTM) options 162
Over-the-counter (OTC) options 7, 233, 234, 237, 247, 249, 252, 254,
255, 291
P
Paper arbitrage 118, 120, 121
Parametrix method 227
Penultimate expiration 282, 333
Perturbation method 6, 209, 227–229, 319–322
Petrodollar recycling 140
Petroleum Administration for Defense Districts (PADDs) 47, 114
Pipelines 2, 21, 23, 46, 117, 119, 123, 125, 127, 287, 293, 302
Positioning 128–134
Position limits 307
Power markets 254, 294, 296, 305
Premium collars 214
Premium retained ratio 197, 199
Price elasticities of demand and supply 29, 34, 36, 37
Price of storage 30, 31
Principle of zero expectations 163
Probability density 40, 309
Producer hedging 57, 60–67, 209–214, 234
Producers, merchants, processors, and users (PMPUs) 129
Prompt futures 2, 333
Propane 127
Put-call parity 173
Put option 162
Q
Quadratic mean 233, 238–241
Quadratic normal (QN) model 6, 228, 229, 276, 277, 293, 319–322
Quadratic utility function 60
Quantamentals 5, 6, 109, 333
R
Ratio call spread 207
Reaction function 5, 83, 105–107, 116, 133, 134, 139, 196, 333
Realized correlation 299
Realized volatility 6, 161, 164, 175–181, 333
Real options 2–3, 6, 109, 118, 123, 125, 210–213, 282, 287, 305, 307
Real rate 11, 16
Rebalancing effect 68
Recessions 55, 69–71, 140
Reduced-form models 41, 293
Refined products 3, 62, 77, 99, 125–128, 219, 239, 292, 294
Refineries 3, 23, 46, 51, 112, 119, 123, 287, 288, 290, 294
Reformulated gasoline blendstock for oxygenate blending (RBOB) 89,
126, 131, 149–154
Regularization 260
Relative vega trading 210
Reverse carry trade 21
Risk aversion channel 142
Risk aversion coefficient 60, 62, 64, 76
Risk-neutral pricing 35, 164, 170, 250
Risk-neutral probabilities 43, 268, 269, 311, 327
Risk parity 55, 69–76, 105, 148, 333
Risk premium 55, 57, 59, 63, 64, 75, 77, 78, 86
Rockefeller, J.D. 3
Roll return 24–26, 66, 75, 333
Roll yield 5, 11, 23, 25, 26, 55, 58, 68, 75, 76, 97, 110, 132, 286
S
Safe haven channel 142
Samuelson effect 185, 240
Samuelson, P. 84
Scholes, M. 164, 169
Seasonality 88, 89, 99, 111, 117, 151, 153, 239, 254, 265
Sentiment index 133
Shale oil 47, 114, 116, 119, 120, 123, 125, 140, 177, 206, 210, 211, 221,
223, 236, 284
Shanghai International Energy Exchange (INE) 124
Signal blending 5, 101, 333
Signal transformation function 105
Silver 11, 12, 308
Simple options 163, 172
Skew correction function 227, 228, 319, 321
Skew delta 215–221, 333
Skewness 6, 86, 128, 179, 180, 183, 187, 190, 209, 212–214, 225, 226,
228, 229, 276
Spot prices 1
Spot return 24–26, 66, 68, 333
Spread options 2, 118, 281–302, 333
Squeeze 5, 29, 38–46, 48, 52, 116, 117, 286, 302, 333
Sraffa, P. 14
Stationarity 47, 48, 73, 122, 156, 158
Statistical arbitrage 5, 109, 122–124, 139, 148, 306
Sticky moneyness 6, 209, 217–220, 333
Sticky strike 6, 209, 219, 220, 333
Stochastic dynamic programming 35
Stochastic volatility 279
Stock-out 33–36, 45, 333
Storage
boundaries 29, 32–38, 43, 50, 115–117
capacity 33, 35–37, 51, 76, 115–116
capacity utilization 48, 115
costs 2, 21, 22, 31, 35, 76, 282
floating 21, 50
optimization 29
price of 30, 31
synthetic 281, 283, 286, 333
tanks 21, 123, 125
theory of 5, 29–37, 83, 87, 305
virtual 7, 281, 282, 286, 287, 305
Straddles 184, 188–190, 214, 218, 229, 276
Strangles 213–214, 229, 276
Super-backwardation 112
Super-contango 112
Swap Data Repositories (SDRs) 148, 333
Swap dealers (SDs) 129
Swaplets 249–251, 253, 267
Swaption 7, 233, 249–254, 266, 267, 333
Synthetic spread options 288, 333
Synthetic storage 281, 283, 286, 333
T
Tank-tops 33, 45, 334
Target redemption swap 255
Taylor formula 168, 309
Thales of Miletus 161
Theory of hedging pressure 4, 5, 55–64, 76–78, 112, 332
Theory of storage 5, 29–37, 83, 87, 305
Theta 172, 193, 283, 286, 287, 316, 317
Three-way collars 212–214, 333
Time spreads 109–117
Trading at settlement (TAS) contract 52, 95, 202, 283, 334
Transaction costs 95, 96, 145, 191, 200–204
Trapped option value 174
Treasury Inflation-Protected Securities (TIPS) 72, 148
Trend following 87
Triangular correlation arbitrage 7, 281, 287–291
Two-factor models 42, 253, 267
U
Ultra-low-sulfur diesel (ULSD) 126
Uncovered interest rate parity 97
US Department of Energy (DOE) 47
US dollar (USD) 5, 12, 13, 137–146, 156, 157
US Strategic Petroleum Reserve (SPR) 5, 16–18, 21
US terms of trade 140
V
Value risk premium 83, 101, 102
Vanilla options 233, 334
Vector autoregressive models (VAR) 71
Vega 6, 172, 181, 208–218, 229, 316, 317
Virtual rate of interest in commodities 14
Virtual storage 7, 281, 282, 286, 287, 305
VIX index 156
Volatility
arbitrage 230, 257
Black 181, 192, 245, 331
exponential 239, 246–247, 251, 254, 266
ghost 176
heuristics 226
implied 6, 161, 164, 180–186, 215–226, 239–241, 261–268, 276,
277, 332
implied Black (IBV) 181, 221–225
implied normal (INV) 181, 222–224
local 6, 161, 164, 167, 171, 172, 175, 177, 181, 185, 226–230, 237–
255, 261, 263–267, 273–279, 285, 292–294, 300, 301, 309, 319, 326,
327, 332
matrix 264
normal 165
realized 6, 161, 164, 175–181, 333
risk premium (VRP) 6, 187, 196–208, 286, 307, 334
risk premium (VRP) smile 6, 187, 198
risk premium (VRP) term structure 187, 199
skew 182, 334
smile 6, 161, 182, 183, 209, 214–221, 226, 276–278, 334
stochastic 279
targeting 105, 334
term structure 7, 185, 237–241, 261, 334
W
Wealth effect 143
Well-posed problems 258
Western Canadian Select (WCS) 125, 127, 145
West Texas Intermediate (WTI) 3, 18, 19, 26, 46, 47, 49–52, 112–115,
117–121, 124, 125, 127, 130, 132
West Texas Sour (WTS) 236
Weymar, H. 84
Whaley, R.E. 174
Working, H. 30
WTI-Brent accordion 5, 120
WTI Houston 125, 236
WTI Midland 125
X
XOP ETF 146, 147, 157
Y
YuanYouBao 52

Footnotes
1 The integration details are presented in many standard derivatives textbooks, such as
Wilmott et al. (1993) and Hull (2018).

2 See, for example, Kemna and Vorst (1990) and Wilmott et al. (1993) for a more detailed
derivation.
3 Some technical regularity conditions must be imposed to make sure that boundary terms at
K → 0 and K → ∞ vanish, which for simplicity, we omit here.

4 This problem was formulated in Bouchouev and Isakov (1997, 1999) and solved analytically
only under additional assumptions. For numerical solutions, see also references in the footnote
in Chap. 12.

5 The drift of the diffusion process must be fixed, and here it is assumed to be zero. It is well
known that two diffusions with different drifts and time-independent volatilities can generate
the same probability density. For example, the mean-reverting process and arithmetic
Brownian motion represent different diffusion processes, but they can produce the same
normal distribution, as shown in Appendix A.

You might also like