The Hidden Cost in Costless Put-Spread Collars:

Rebalance Timing Luck‡

January 24, 2023

Steven Braun Corey Hoffstein Roni Israelov David Nze Ndong

Prior research and empirical investment results demonstrate that strategy performance can be highly
sensitive to rebalance schedules, an effect called rebalance timing luck (“RTL”). In this paper we
extend the empirical analysis to option-based strategies. As a case study, we replicate a popular
strategy – the self-financing, three-month put-spread collar – with three implementations that vary
only in their rebalance schedule. We find that the annualized tracking error between any two
implementations is in excess of 400 basis points. We also decompose the empirically-derived
rebalance timing luck for this strategy into its linear and non-linear components. Finally, we provide
intuition for the driving causes of rebalance timing luck in option-based strategies.

KEYWORDS: Calendar-Based Rebalancing, Options, Portfolio Construction, Put-Spread Collars,

Rebalance Timing Luck, Systematic Investing

‡ Steven Braun is a Senior Quantitative Analyst and Chief Derivatives Risk Officer at Newfound Research LLC. Corey Hoffstein
is the Chief Investment Officer of Newfound Research LLC. Roni Israelov is the President and Chief Investment Officer of NDVR,
Inc. David Nze Ndong is a Research Associate at NDVR, Inc.

Newfound Research LLC (“Newfound”) is an investment advisor that may or may not apply similar investment techniques or
methods of analysis as those described herein when managing portfolios for its clients. The views expressed herein are those of the
authors and not necessarily those of Newfound. The views and information herein are not, and may not be relied on in any manner
as, investment, legal, tax, accounting or other advice provided by NDVR or as an offer to sell or a solicitation of an offer to buy
any security. Newfound does not provide legal or tax advice, and the information provided should not be considered legal or tax
advice. Consult an attorney, tax professional, or other advisor regarding your specific legal or tax situation.

NDVR, Inc. ("NDVR") is an investment advisor that may or may not apply similar investment techniques or methods of analysis
as those described herein when managing portfolios for its clients. The views expressed herein are those of the authors and not
necessarily those of NDVR. The views and information herein are not, and may not be relied on in any manner as, investment,
legal, tax, accounting or other advice provided by NDVR or as an offer to sell or a solicitation of an offer to buy any security.
NDVR does not provide legal or tax advice. Consult an attorney, tax professional, or other advisor regarding your specific legal or
tax situation.

Electronic copy available at: https://ssrn.com/abstract=4336419

Introduction

In the last decade, the interest in option-based investment strategies has grown dramatically
among both institutional and retail investors. To put this growing interest in perspective, the
Morningstar Options Trading category 1 has increased its assets under management from $7.8
billion in December 2012 to $54.2 billion as of December 2022. Allocators employ such strategies
both for the opportunity to introduce a diversifying risk premium as well as their ability to offer
defined risk controls.

The use of options introduces an important set of features: a defined expiration date and a
non-linear payoff profile. The defined expiration date makes the choice of when to reset exposure,
a necessary component of strategy design. This is not a unique feature of options: other
instruments, such as bonds, futures, and forwards, also have defined maturities. Their more direct
relationship with the underlying asset, however, means that the return difference between tenors
is often less profound. The combination of a defined expiration date and a non-linear payoff means
that tenor selection, as well as the choice of when to reset or roll exposure, can play an important
role in strategy returns. The question we aim to address in this paper is how important?

As an example, consider one such popular strategy: a three-month self-financing put-

spread collar on the S&P 500. Every three months, the strategy buys a three-month 5-20% put-
spread and sells a three-month out-of-the-money call such that the structure is costless. With index
options expiring every third Thursday of the month, there are three possible rebalance schedules
that can be easily implemented: December, March, June, and September; January, April, July, and
October; and February, May, August, and November. If each of these implementations constitutes
a unique strategy, what is the expected variation in their realized returns? How much does when
the rebalance occurs matter?

Inquiry into this question was first addressed by Blitz, Grient, and van Vliet (2010) in the
context of equity strategies. They demonstrate that the performance for an annually rebalanced

1
The Morningstar Options Trading category includes a wide range of option-based strategies, including: put writing,
options spreads, options-based hedged equity, and collar strategies.

Electronic copy available at: https://ssrn.com/abstract=4336419

fundamental index strategy (proposed by Arnott, Hsu, and Moore (2005)) is highly sensitive to the
month in which it is rebalanced. For example, a fundamental index rebalanced in March
outperformed one that rebalanced in September by over 1000 basis points in 2009. The authors
propose a solution for managing this undesirable performance ambiguity by blending multiple
indexes together in equal weight (known as “tranching” or “staggered rebalancing”).

Hoffstein, Sibears, and Faber (2019) prove, under certain assumptions, that equal-weight
tranching is the optimal strategy for minimizing exposure to the return variance that occurs from
different rebalance schedule choices. They go on to define “rebalance timing luck” as the
annualized tracking error between a specific rebalance choice implementation and the theoretically
optimal solution. Hoffstein, Faber, and Braun (2020) extend the empirical study of rebalance
timing luck to equity factor strategies, finding that prototypical implementations of popular long-
only factor portfolios can realize hundreds of basis points of annual tracking error between
rebalance date choices.

To illustrate the effects of rebalance timing luck within option-based strategies, we

implement the self-financing put-spread collar strategy described above. By varying the rebalance
dates, we generate three unique implementations, providing returns from which we empirically
estimate rebalance timing luck. To gain further intuition into the nature of rebalance timing luck
in option strategies, we decompose the empirical results into linear and non-linear components
using the methodology proposed by Israelov and Nielson (2015).2 We also provide an important
bias adjustment to empirical estimates for rebalance timing luck not accounted for in prior
literature.

Through the results of our study, we extend the literature by validating that rebalance
timing luck is a meaningful source of performance variation in option strategies. Our results have
important implications for portfolio managers and allocators alike, as failing to inoculate these
effects can increase the ambiguity in returns between luck and skill and introduce an
uncompensated source of variance. This is a particularly timely topic given the popularization of
option-based strategies over the last decade.

2 The linear component can be described as the portfolio delta exposures, while the nonlinear component can be described as the
remaining higher-order option greeks exposures as well as implied volatility related exposures.

Electronic copy available at: https://ssrn.com/abstract=4336419

Data & Methodology

We begin by constructing a self-financing put-spread collar strategy. The strategy is

comprised of three components: (1) passive exposure to the S&P 500; (2) a put spread on the S&P
500 (long a 5% out-of-the-money put and short a 20% out-of-the-money put); and (3) a short out-
of-the-money call option on the S&P 500 (selected to have equal value to the put spread). The
options are all three months in tenor and the entire structure is rebalanced at each expiration.

S&P 500 Index options data, including mid-point prices and option greeks, is provided by
iVolatility. The S&P 500 Total Return and Price Indexes are from Reuters and risk-free rates of
return are the 30-day Treasury bill returns provided by the French Data Library.

We estimate the performance of this strategy from 2008 through 2022 using standard index
options that expire on the third Friday of each month. Given this setup, we can calculate three
different rebalance schedule implementations: (1) December, March, June, and September
(“DMJS”); (2) January, April, July, and October (“JAJO”); and (3) February, May, August, and
November (“FMAN”).

We then construct an equal-weight portfolio of DMJS, JAJO, and FMAN. This portfolio
reflects an estimate of strategy returns neutralizing the impact of rebalance schedule decisions, as
suggested by Hoffstein, Sibears, and Faber (2019). Measuring the annualized tracking error
between the idiosyncratic implementations (DMJS, JAJO, and FMAN) and the equal-weight
portfolio provides an estimate3 of rebalance timing luck.

Following the methodology of Israelov and Nielson (2015), we decompose our results into
three sources: passive equity exposure, equity timing exposure (“linear”), and delta neutral
exposure (“non-linear”). The passive equity exposure is the long-term average delta of the
strategy, the linear component is the stochastic delta of each strategy, and the non-linear
component captures sensitivity to both implied and realized volatilities as well as higher order
greeks.

3 Note that this estimate is a biased estimate and must be adjusted. See Appendix A for adjustment calculation.

Electronic copy available at: https://ssrn.com/abstract=4336419

 Specifically, we assume the three strategies share a common component, but also have
idiosyncratic components.

𝑟𝑡 = 𝑟̅𝑡 + 𝑟𝑡∗ (1)

The common component is obtained by taking the equally-weighted average of the returns
of the three strategies. Component-specific idiosyncratic returns are then obtained by subtracting
the component-specific common component from the actual return.4

∗ ∗ ∗
𝑟𝑡 = (𝑟̅𝑝𝑒,𝑡 + 𝑟𝑝𝑒,𝑡 ) + (𝑟̅𝑡𝑒,𝑡 + 𝑟𝑡𝑒,𝑡 ) + (𝑟̅𝑑𝑛,𝑡 + 𝑟𝑑𝑛,𝑡 ) (2)

Calculation of 𝑟𝑝𝑒,𝑡 and 𝑟𝑡𝑒,𝑡 (whether common or idiosyncratic) is achieved by utilizing

the return of the S&P 500 as well as the underlying delta of the underlying strategies. The passive
equity exposure is measured via the long-term average delta of the portfolio, whereas the equity
timing exposure is measured via the delta deviation from the long-term average at each point in
time.

Empirical Results & Discussion

To highlight the effects of dispersion caused by rebalance timing luck, Exhibit 1 depicts
the various equity curves of the self-financing put-spread collar strategies. Exhibit 2 provides
relevant summary statistics. Most striking is that JAJO outperforms FMAN over the sample period
by 100 basis points annualized while its annualized volatility is 80 basis points lower. Similarly,
JAJO has a max drawdown 900 basis points lower than FMAN. Arbitrarily selecting JAJO versus
FMAN would have led to a 40-percentage point difference in total return over the sample period.

EXHIBIT 1

4 In equation 2, 𝑟𝑝𝑒
is the return due to passive equity exposure, 𝑟𝑡𝑒 is the return due to timing exposure (linear exposure), and
𝑟𝑑𝑛 is the return due to delta-neutral exposure.

Electronic copy available at: https://ssrn.com/abstract=4336419

 EXHIBIT 2

Exhibit 3 isolates the cumulative difference in returns between DMJS, JAJO, and FMAN.
The average annualized volatility of these cumulative difference series is 360 basis points. Exhibit
4 plots the rolling one-year average return as well as the rolling maximum spread between
underlying strategies, highlighting that the magnitude of the spread is frequently larger than the
average annual return itself.

EXHIBIT 3

EXHIBIT 4

With empirical returns for DMJS, JAJO, and FMAN, we construct an equal-weight
portfolio of the three strategies. We then calculate the difference in daily returns between each
implementation and the equal-weight portfolio. The square-root of the average of these variances
provides us our estimate of rebalance timing luck.

Adjusting for estimation bias5, we calculate the rebalance timing luck of the strategy to be
292 basis points. This metric also provides guidance for the expected tracking error between any
two idiosyncratic implementations (derived in Appendix B). In the case of three samples, we scale
by √2, suggest a tracking error of 412 basis points.6

5See Appendix A for the derivation and proof of the bias adjustment.
6Our estimates are calculated using monthly returns. Using daily returns, we calculate rebalance timing luck to be 307 basis
points and the tracking error between idiosyncratic implementations to be 434 basis points.

Electronic copy available at: https://ssrn.com/abstract=4336419

 These results complicate both strategy analysis and manager selection, as the annual
performance between two otherwise identical managers can vary by hundreds of basis points due
only to the rebalance schedule they select. Where does this rebalance timing luck arise from?

In line with the methodology outlined by Israelov and Nielson (2015), we attempt to isolate
the impact of these different effects. Using the empirical returns and deltas of DMJS, JAJO,
FMAN, and the equal-weight portfolio, we implement Equation 2 to isolate passive equity (long-
term average delta), equity timing (stochastic delta), and delta neutral (non-delta greeks) effects.
The results are transcribed in Exhibit 5.

The idiosyncratic components capture the differences in DMJS, JAJO, and FMAN from
the common component and are the source of rebalance timing luck. We see that equity timing
creates 282 basis points of tracking error while delta neutral effects generate 147 basis points.
Note that if we assume these effects are independent, their cumulative tracking error would be 318
basis points 7, approximately 26 basis points higher than our empirical estimate of rebalance timing
luck, suggesting a slightly negative (albeit, not likely statistically significant) correlation between
the effects for this structure.

These results suggest that for our example self-financing put-spread collar strategy, equity
timing is responsible for nearly 80% of rebalance timing variance.8 This certainly need not be the
case for every option structure. For a continuously delta-hedged strategy, for example, the equity
timing component will have zero contribution to rebalance timing luck.

Developing Intuition

Consider that, in the case of our example strategies, options are held until expiration. This
means, under certain idealized assumptions, their payoff can be replicated by trading the S&P 500
in proportion to the structure’s cumulative delta. In fact, at any given point in time, the strategy
can be characterized by how much exposure it has to the S&P 500 and any deviations between the
three implementations will be driven by the differences in their exposure.

7 0.0318 = √0.02822 + 0.01472

8 Defining “rebalance timing variance” as rebalance timing luck squared.

Electronic copy available at: https://ssrn.com/abstract=4336419

 Exhibits 6 depicts how the delta profile of an individual strategy can change over time and
Exhibit 7 depicts how the delta profile across different strategies can vary on the same day. Exhibit
8 demonstrates that the underlying strategy deltas are clearly asynchronous.

EXHIBIT 6

EXHIBIT 7

EXHIBIT 8

While continuous delta replication is sufficient for matching the payoff profile of the
individual strategies at expiration and is a clear source of rebalance timing luck, it will not
adequately capture the daily mark-to-market of the option values themselves. For example,
consider a hypothetical case where the volatility surface is constant and price in the underlying is
static for three months. In such a case, all three strategies would end up holding options with
identical strikes but different expirations. If suddenly, the volatility surface repriced to a new level,
the option structures would respond differently due to their differing sensitivities to implied
volatilities (“vega”). Similarly, if price of the underlying suddenly jumped, the option structures
would also respond differently due to their differing deltas and gammas.

In fact, even without any shift in the underlying or volatility surface, the options would
reprice due to premium decay. Assuming a constant implied volatility surface, the self-financing
put-spread collar is short volatility (both realized and implied), and therefore earns theta over time.
The theta earned, however, is time-varying (due to decay in both gamma and vega) and therefore
will lead to return discrepancy among the implementations.

Electronic copy available at: https://ssrn.com/abstract=4336419

 “Which implied volatility,” however, is an important question with regards to vega as the
volatility surface can react very differently. Consider, for example, the extreme case of a strategy
that buys a one-year at-the-money straddle on the S&P 500 and does not roll until expiration. We
can easily imagine 12 different rebalance schedules: one for each month. Consider that the
schedule which rebalances in November will, once every four years, end up holding a less than
one-month straddle into the U.S. Presidential election. The event premium priced into the surface
will likely have substantially different effects upon that implementation than the implementation
which had just previously reset in October.

Broadly speaking, as a first order approximation, we would expect point-in-time

differences in delta and vega between rebalance schedule implementations to be the meaningful
driver of strategy dispersion. The greater the potential variance in a strategy’s delta and vega over
time, as well as the potential drift in the tenor of its options, the greater the potential magnitude of
rebalance timing luck.

Conclusion & Further Discussion

The decision of when to roll or reset an option structure is a critical element of an option-
based strategy’s design. Little research, however, has been performed on the potential impact this
arbitrary decision has upon strategy returns. Prior literature in the space of rebalance timing luck,
however, suggests that this decision is not without potentially meaningful consequences.

In this paper, we explore the impact of rebalance timing luck on one popular option-based
strategy: a three-month, self-financing put-spread collar. To empirically measure the impact, we
construct three possible rebalance schedule variations: (1) December, March, June, and September
(“DMJS”); (2) January, April, July, and October (“JAJO”); and (3) February, May, August, and
November (“FMAN”). Using these strategy returns, we then calculate rebalance timing luck over
the test period (2008 through 2022).

We find the empirical estimate of rebalance timing luck to be nearly 300 basis points,
implying an annual tracking error between any two implementations to be over 400 basis points.
For context, the annualized volatility of the three implementations is about 1100 basis points,
indicating that rebalance timing luck has nearly 30% the variability of the strategy itself. As these

Electronic copy available at: https://ssrn.com/abstract=4336419

are otherwise identically managed strategies, this deviation arises only from the choice in rebalance
schedule. To be clear, this choice does not necessarily affect strategy volatility, but it does cause
a meaningful dispersion in terminal wealth among the alternative choices.

These results call into question one’s ability to draw meaningful relative performance
conclusions between two option-based strategies or a strategy and its benchmark. Unless a
manager or benchmark explicitly controls for this effect, discerning between alpha and rebalance
timing luck may be difficult, even over a decade-long measurement period. Given that many
option-based strategy benchmarks are managed with a fixed rebalance schedule, allocators must
be aware of this effect.

To gain a better understanding of where rebalance timing luck is emerging from, we follow
the decomposition methodology of Israelov and Nielson (2015). We find that, in our example
study, approximately 80% of the variance emerges from idiosyncratic stochastic delta effects and
20% emerges from delta-neutral effects (e.g. vega or gamma).

With the growth of option-based strategies, the impact of rebalance timing luck can be seen
explicitly. Product sponsors now make available mutual funds and ETFs with identically managed
strategies but varying rebalance schedules. While many of these products are marketed for their
ability to create point-to-point “defined outcomes,” they are not always used this way. As investors
and allocators continue to adopt option-based strategies, they should do so fully aware of the
potential magnitude rebalance timing luck can play in their realized returns.

Electronic copy available at: https://ssrn.com/abstract=4336419

References

Arnott, R.D., Hsu, J., and Moore, P. (2005), “Fundamental Indexation”, Financial Analysts
Journal, Vol. 61, No. 2, 83-89.

Blitz, D., van der Grient, B., and van Vliet, P. (2010). “Fundamental Indexation: Rebalancing
Assumptions and Performance,” Journal of Index Investing, Vol. 1, No. 2, 82-88.

Hoffstein, C., Faber, N., Sibears, D. (2019). “Rebalance Timing Luck: The Difference Between
Hired and Fired,” Journal of Index Investing, Vol. 10, No. 1, 27-36.

Hoffstein, C., Faber, N., Braun, S. (2020). “Rebalance Timing Luck: The (Dumb) Luck of Smart
Beta.” Available at SSRN: https://ssrn.com/abstract=3673910 or
http://dx.doi.org/10.2139/ssrn.3673910

Israelov, R. and Nielsen, L. (2015). “Covered Calls Uncovered,” Financial Analysts Journal,
Vol. 71, No. 6, November/December 2015.

Electronic copy available at: https://ssrn.com/abstract=4336419

Exhibit 1: Compound Excess Returns
Compound excess returns for the three different rebalance schedule implementations of the self-financing
put-spread collar strategy.

250%

200%
Compounded Return

150%

100%

50%
2008 2012 2016 2020

DMJS JAJO FMAN

Source: Newfound Research, NDVR, Reuters and iVolatility. Past performance is not indicative of future returns.
Performance is backtested and hypothetical. Performance figures are gross of all fees, including, but not limited to,
manager fees, transaction costs, and taxes. Performance assumes the reinvestment of all distributions.

Electronic copy available at: https://ssrn.com/abstract=4336419

Exhibit 2: Summary Statistics for Three Tranches
Summary statistics for the three different rebalance schedule implementations of the self-financing put-
spread collar strategy.

DMJS JAJO FMAN Market

Mean (Excess) 5.7% 5.9% 4.9% 10.5%

Vol 10.6% 10.9% 11.7% 20.8%

Sharpe 0.54 0.54 0.42 0.51

Max DD 32% 25% 34% 52%

Beta 0.47 0.49 0.53 1.00

Alpha 0.8% 0.7% -0.6% 0.0%

Alpha t-stat 0.71 0.70 -0.56 0.00

Source: Newfound Research, NDVR, Reuters and iVolatility. Past performance is not indicative of future returns.
Performance is backtested and hypothetical. Performance figures are gross of all fees, including, but not limited to,
manager fees, transaction costs, and taxes. Performance assumes the reinvestment of all distributions.

Electronic copy available at: https://ssrn.com/abstract=4336419

Exhibit 3: Differences in Performances Across Tranches
Cumulative return difference between the three different rebalance schedule implementations of the self-
financing put-spread collar strategy.

40%

30%

20%
Compounded Return

10%

0%

-10%

-20%

-30%

-40%
2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022

DMJS-JAJO DMJS-FMAN JAJO-FMAN

JAJO-DMJS FMAN-DMJS FMAN-JAJO

Source: Newfound Research, NDVR, Reuters and iVolatility. Past performance is not indicative of future returns.
Performance is backtested and hypothetical. Performance figures are gross of all fees, including, but not limited to,
manager fees, transaction costs, and taxes. Performance assumes the reinvestment of all distributions.

Electronic copy available at: https://ssrn.com/abstract=4336419

Exhibit 4: One-Year Rolling Compounded Total Return
Rolling 1-year average return for the three different rebalance schedule implementations of the self-
financing put-spread collar strategy as well as the rolling one-year maximum return spread between the
three strategies.

40%

30%
One-Year Compounded Return

20%

10%

0%

-10%

-20%

-30%
2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022

Average Max Spread

Source: Newfound Research, NDVR, Reuters and iVolatility. Past performance is not indicative of future returns.
Performance is backtested and hypothetical. Performance figures are gross of all fees, including, but not limited to,
manager fees, transaction costs, and taxes. Performance assumes the reinvestment of all distributions.

Electronic copy available at: https://ssrn.com/abstract=4336419

Exhibit 5: Average Summary Statistics Across Tranches
Table reporting the average summary statistics of return decomposition of DMJS, JAJO, and FMAN
portfolios. Returns are decomposed into common and idiosyncratic elements following the methodology
of Israelov and Nielson (2015). Passive Equity captures long-term average delta effects, Timing Equity
captures stochastic delta effects, and Delta Neutral captures non-delta greek effects. Calculations are not
adjusted for estimation bias.

AVERAGE Portfolio Common Idiosyncratic Common Idiosyncratic

Passive Timing Delta Passive Timing Delta
Total Total
Equity Equity Neutral Equity Equity Neutral

Mean (Excess) 5.5% 4.4% 0.7% 0.4% 0.0% 0.0% 0.0% 5.5% 0.0%

Vol 11.1% 8.7% 2.8% 2.4% 0.0% 2.3% 1.2% 10.8% 2.5%

Sharpe 0.50 0.51 0.25 0.17 0.17 0.00 0.00 0.51 0.00

Max DD 30.5% 24.5% 5.1% 5.2% 0.2% 10.3% 3.9% 29.2% 10.6%

Beta 0.50 0.42 0.05 0.03 0.00 0.00 0.00 0.50 0.00

Alpha 0.28% 0.00% 0.23% 0.05% 0.00% 0.00% 0.00% 0.28% 0.00%

Alpha t-Stat 0.28 0.80 0.33 0.08 - -0.01 0.01 0.34 -0.01

Source: Newfound Research, NDVR, Reuters and iVolatility. Past performance is not indicative of future returns.
Performance is backtested and hypothetical. Performance figures are gross of all fees, including, but not limited to,
manager fees, transaction costs, and taxes. Performance assumes the reinvestment of all distributions.

Electronic copy available at: https://ssrn.com/abstract=4336419

Exhibit 6: Strategy Delta for January-April-July-October Implementation
Delta profile for January-April-July-October implementation on different dates. The initial rebalance date
is January 20, 2017.

1.0

0.9

0.8

0.7

0.6
Delta

0.5

0.4

0.3

0.2

0.1

0.0
1200 1400 1600 1800 2000 2200 2400 2600 2800
S&P 500 Index Level

At launch After one month After two months Day before expiration

Source: Newfound Research, NDVR, Reuters, and iVolatility.

Electronic copy available at: https://ssrn.com/abstract=4336419

Exhibit 7: Strategy Deltas Across Implementations
Delta profiles for the three implementations on the same date (February 14, 2019).

1.0

0.9

0.8

0.7

0.6
Delta

0.5

0.4

0.3

0.2

0.1

0.0
1400 1600 1800 2000 2200 2400 2600 2800 3000 3200 3400
S&P 500 Index Level

FMAN DMJS JAJO

Source: Newfound Research, NDVR, Reuters, and iVolatility.

Electronic copy available at: https://ssrn.com/abstract=4336419

Exhibit 8: Strategy Net Delta Over Time
Net delta profiles of each strategy implementation plotted over time.

1.0

0.9

0.8

0.7

0.6
Delta

0.5

0.4

0.3

0.2

0.1

0.0
Dec-19 Mar-20 Jun-20 Sep-20 Dec-20 Mar-21 Jun-21 Sep-21 Dec-21 Mar-22 Jun-22 Sep-22

DMJS JAJO FMAN

Source: Newfound Research, NDVR, and iVolatility.

Electronic copy available at: https://ssrn.com/abstract=4336419

Appendix A: Derivation of the Rebalance Timing Luck Bias Adjustment

Assume we have N potential implementations of a strategy (different rebalance dates), from

which we select n. We can define the return of a specific implementation as:

𝑟𝑖 = 𝑟̅ + 𝑢𝑖

1
Where 𝑟̅ represents the pure, zero-timing-luck strategy return (𝑟̅ = 𝑁 ∑𝑁
𝑖=1 𝑟𝑖 ) and 𝑢𝑖 is the

idiosyncratic return of the ith implementation.

Unfortunately, 𝑟̅ is not observable. We can estimate it by averaging 𝑟𝑖 .

𝑛 𝑛
1 1
𝑟̂ = ∑ 𝑟𝑖 = 𝑟̅ + ∑ 𝑢𝑖
𝑛 𝑛
𝑖=1 𝑖=1

Let U be the covariance matrix of u’s. Below we prove that the sum of each row of U equals zero
(or, equivalently, the sum of each column equals zero as U is symmetric):

𝑁 𝑁

∑ 𝑐𝑜𝑣(𝑢𝑖 , 𝑢𝑗 ) = ∑ 𝑐𝑜𝑣(𝑟𝑖 − 𝑟̅ , 𝑟𝑗 − 𝑟̅ )
𝑗=1 𝑗=1

= ∑(𝑐𝑜𝑣(𝑟𝑖 , 𝑟𝑗 ) − 𝑐𝑜𝑣(𝑟𝑖 , 𝑟̅ ) − 𝑐𝑜𝑣(𝑟𝑗 , 𝑟̅ ) + 𝑐𝑜𝑣(𝑟̅ , 𝑟̅ ))

𝑗=1

𝑁 𝑁 𝑁 𝑁 𝑁
1 1 1
= ∑ (𝑐𝑜𝑣(𝑟𝑖 , 𝑟𝑗 ) − ∑ 𝑐𝑜𝑣(𝑟𝑖 , 𝑟𝑘 ) − ∑ 𝑐𝑜𝑣(𝑟𝑗 , 𝑟𝑘 ) + 2 ∑ ∑ 𝑐𝑜𝑣(𝑟𝑘 , 𝑟𝑚 ))
𝑁 𝑁 𝑁
𝑗=1 𝑘=1 𝑘=1 𝑘=1 𝑚=1

𝑁 𝑁 𝑁 𝑁 𝑁 𝑁
1 1
= ∑ 𝑐𝑜𝑣(𝑟𝑖 , 𝑟𝑗 ) − ∑ 𝑐𝑜𝑣(𝑟𝑖 , 𝑟𝑘 ) − ∑ ∑ 𝑐𝑜𝑣(𝑟𝑗 , 𝑟𝑘 ) + ∑ ∑ 𝑐𝑜𝑣(𝑟𝑘 , 𝑟𝑚 ) = 0
𝑁 𝑁
𝑗=1 𝑘=1 𝑗=1 𝑘=1 𝑘=1 𝑚=1

Therefore, the sum of all elements of the covariance matrix equals zero. Furthermore, because we
know the diagonal elements of U must be positive, for the sum of all the elements of the matrix to

Electronic copy available at: https://ssrn.com/abstract=4336419

equal zero, the sum of the diagonal elements must equal negative the sum of the off-diagonal
elements.

𝑁 𝑁 𝑁

∑ 𝑣𝑎𝑟(𝑢𝑖 ) = −2 ∑ ∑ 𝑐𝑜𝑣(𝑢𝑖 , 𝑢𝑗 )
𝑖=1 𝑖=1 𝑗=𝑖+1

Let 𝑆𝑛 define a subset of 𝑢𝑖 ’s of length 𝑛 from the larger set of 𝑁 (where 𝑁 > 1) 𝑢𝑖 ’s. Assume
the ansatz relationship:

𝑁−𝑛
𝐸 [𝐸 [∑ ∑ 𝑐𝑜𝑣(𝑢𝑖 , 𝑢𝑗 ) |𝑆𝑛 ]] = ( ) 𝐸 [𝐸 [ ∑ 𝑣𝑎𝑟(𝑢𝑖 ) | 𝑆𝑛 ]]
𝑁−1
𝑖∈𝑆𝑛 𝑗∈𝑆𝑛 𝑢𝑖 𝑖∈𝑆𝑛

Define the left-hand side as 𝐿𝐻𝑆𝑛 and the right-hand side as 𝑅𝐻𝑆𝑛 .

We will prove by induction9 that the equality holds for all 𝑛 where 1 ≤ 𝑛 ≤ 𝑁.

Base case: The equality is trivially true for both 𝑛 = 1 and 𝑛 = 𝑁 (as we proved the sum of the
covariance matrix equals zero above).

Induction step: We assume 𝐿𝐻𝑆𝑛 = 𝑅𝐻𝑆𝑛 . Adding 𝑢𝑘 , we find:

𝐿𝐻𝑆𝑛 + 2 ∑ 𝑐𝑜𝑣(𝑢𝑖 , 𝑢𝑘 ) − 𝑣𝑎𝑟(𝑢𝑘 ) = 𝑅𝐻𝑆𝑛 + 𝑣𝑎𝑟(𝑢𝑘 )

𝑖=1

Taking the expectation of both sides with respect to 𝑢𝑘 , we get:

9 The authors would like to thank @quantian1 on Twitter for their contribution to this proof.

Electronic copy available at: https://ssrn.com/abstract=4336419

 𝑁 𝑁 𝑁 𝑁
2 1 1
𝐿𝐻𝑆𝑛 + ∑ ∑ 𝑐𝑜𝑣(𝑢𝑖 , 𝑢𝑘 ) − ∑ 𝑣𝑎𝑟(𝑢𝑘 ) = 𝑅𝐻𝑆𝑛 + ∑ 𝑣𝑎𝑟(𝑢𝑘 )
𝑁 𝑁 𝑁
𝑘=1 𝑖=1 𝑘=1 𝑘=1

𝑁 𝑁 𝑁
1 1
𝐿𝐻𝑆𝑛 + 2 ( ∑ ∑ 𝑐𝑜𝑣(𝑢𝑖 , 𝑢𝑘 )) = 𝑅𝐻𝑆𝑛 + 2 ( ∑ 𝑣𝑎𝑟(𝑢𝑘 ))
𝑁 𝑁
𝑘=1 𝑖=1 𝑘=1

𝐿𝐻𝑆𝑛 + 2𝐿𝐻𝑆1 = 𝑅𝐻𝑆𝑛 + 2𝑅𝐻𝑆1

Since 𝐿𝐻𝑆1 = 𝑅𝐻𝑆1, we find that 𝐿𝐻𝑆𝑛 =𝑅𝐻𝑆𝑛 , which is true because we assumed it. By the
principle of induction, we have proven that the ansatz relationship is true for all 𝑛 where 1 ≤
𝑛 ≤ 𝑁.

We would like to estimate the variance of 𝑢𝑖 , which is not observable. Define:

𝑢̂𝑖 = 𝑟𝑖 − 𝑟̂

We can rewrite this as:

𝑛 𝑛 𝑛
1 1 𝑛−1 1
𝑢̂𝑖 = 𝑟𝑖 − 𝑟̂ = (𝑟̅ + 𝑢𝑖 ) − (𝑟̅ + ∑ 𝑢𝑖 ) = 𝑢𝑖 − ∑ 𝑢𝑖 = 𝑢𝑖 − ∑ 𝑢𝑗
𝑛 𝑛 𝑛 𝑛
𝑖=1 𝑖=1 𝑗=1;𝑖≠𝑗

We define:

1
𝑣𝑎𝑟(𝑢) ≡ ∑ 𝑣𝑎𝑟(𝑢𝑖 )
𝑛
𝑖∈𝑆

𝑛
1 1 1
𝑣𝑎𝑟(𝑢̂) ≡ ∑ 𝑣𝑎𝑟(𝑢̂𝑖 ) = ∑ 𝑣𝑎𝑟 (𝑢𝑖 − ∑ 𝑢𝑖 )
𝑛 𝑛 𝑛
𝑖∈𝑆 𝑖∈𝑆 𝑖=1

We solve for 𝐸[𝐸[𝑣𝑎𝑟(𝑢̂)|𝑆𝑛 ]]:

Electronic copy available at: https://ssrn.com/abstract=4336419

 1 1 1
= 𝐸 𝐸 [ ∑ (𝑣𝑎𝑟(𝑢𝑖 ) + 𝑣𝑎𝑟 ( ∑ 𝑢𝑗 ) + 2𝑐𝑜𝑣 (𝑢𝑖 , − ∑ 𝑢𝑗 )) |𝑆𝑛 ]
𝑛 𝑛 𝑛
𝑖∈𝑆𝑛 𝑗∈𝑆𝑛 𝑗∈𝑆𝑛
[ ]

1 1 2
= 𝐸 [𝐸 [ ∑ (𝑣𝑎𝑟(𝑢𝑖 ) + 2 ( ∑ ∑ 𝑐𝑜𝑣(𝑢𝑗 , 𝑢𝑘 )) − ∑ 𝑐𝑜𝑣(𝑢𝑖 , 𝑢𝑗 )) |𝑆𝑛 ]]
𝑛 𝑛 𝑛
𝑖∈𝑆𝑛 𝑗∈𝑆𝑛 𝑘∈𝑆𝑛 𝑗∈𝑆𝑛

1 1 2
= 𝐸 [𝐸 [ (∑ 𝑣𝑎𝑟(𝑢𝑖 ) + 2 ∑ ( ∑ ∑ 𝑐𝑜𝑣(𝑢𝑗 , 𝑢𝑘 )) − ∑ ∑ 𝑐𝑜𝑣(𝑢𝑖 , 𝑢𝑗 )) |𝑆𝑛 ]]
𝑛 𝑛 𝑛
𝑖∈𝑆𝑛 𝑖∈𝑆𝑛 𝑗∈𝑆𝑛 𝑘∈𝑆𝑛 𝑖∈𝑆𝑛 𝑗∈𝑆𝑛

1 1 2
= 𝐸 [𝐸 [ (∑ 𝑣𝑎𝑟(𝑢𝑖 ) + ∑ ∑ 𝑐𝑜𝑣(𝑢𝑗 , 𝑢𝑘 ) − ∑ ∑ 𝑐𝑜𝑣(𝑢𝑖 , 𝑢𝑗 )) |𝑆𝑛 ]]
𝑛 𝑛 𝑛
𝑖∈𝑆𝑛 𝑗∈𝑆𝑛 𝑘∈𝑆𝑛 𝑖∈𝑆𝑛 𝑗∈𝑆𝑛

1 1
= 𝐸 [𝐸 [ (∑ 𝑣𝑎𝑟(𝑢𝑖 ) − ∑ ∑ 𝑐𝑜𝑣(𝑢𝑗 , 𝑢𝑘 ))] |𝑆𝑛 ]
𝑛 𝑛
𝑖∈𝑆𝑛 𝑗∈𝑆𝑛 𝑘∈𝑆𝑛

1 1
= 𝐸 [𝐸 [ ∑ 𝑣𝑎𝑟(𝑢𝑖 ) − 2 ∑ ∑ 𝑐𝑜𝑣(𝑢𝑗 , 𝑢𝑘 ) |𝑆𝑛 ]]
𝑛 𝑛
𝑖∈𝑆𝑛 𝑗∈𝑆𝑛 𝑘∈𝑆𝑛

1 1
= 𝐸 [𝐸 [ ∑ 𝑣𝑎𝑟(𝑢𝑖 ) |𝑆𝑛 ] − 𝐸 [ 2 ∑ ∑ 𝑐𝑜𝑣(𝑢𝑗 , 𝑢𝑘 ) |𝑆𝑛 ]]
𝑛 𝑛
𝑖∈𝑆𝑛 𝑗∈𝑆𝑛 𝑘∈𝑆𝑛

Electronic copy available at: https://ssrn.com/abstract=4336419

 1 1
= 𝐸 [𝐸 [∑ 𝑣𝑎𝑟(𝑢𝑖 ) |𝑆𝑛 ]] − 2 𝐸 [𝐸 [ ∑ ∑ 𝑐𝑜𝑣(𝑢𝑗 , 𝑢𝑘 ) |𝑆𝑛 ]]
𝑛 𝑛
𝑖∈𝑆𝑛 𝑗∈𝑆𝑛 𝑘∈𝑆𝑛

Substituting, we find that:

1 1 𝑁−𝑛
= 𝐸 [𝐸 [∑ 𝑣𝑎𝑟(𝑢𝑖 ) |𝑆𝑛 ]] − 2 ( ) 𝐸 [𝐸 [∑ 𝑣𝑎𝑟(𝑢𝑖 ) |𝑆𝑛 ]]
𝑛 𝑛 𝑁−1
𝑖∈𝑆𝑛 𝑖∈𝑆𝑛

𝑁 (𝑛 − 1) 1
𝐸[𝐸[𝑣𝑎𝑟(𝑢̂)|𝑆]] = 𝐸 [𝐸 [ ∑ 𝑣𝑎𝑟(𝑢𝑖 ) | 𝑆𝑛 ]]
(𝑁 − 1) 𝑛 𝑛
𝑢𝑖 𝑖∈𝑆𝑛

𝑁 (𝑛 − 1)
𝐸[𝐸[𝑣𝑎𝑟(𝑢̂)|𝑆𝑛 ]] = 𝐸[𝐸[𝑣𝑎𝑟(𝑢)| 𝑆𝑛 ]]
(𝑁 − 1) 𝑛

Taking the limit as 𝑁 → ∞,

(𝑛 − 1)
𝐸[𝐸[𝑣𝑎𝑟(𝑢̂)|𝑆𝑛 ]] = 𝐸[𝐸[𝑣𝑎𝑟(𝑢)| 𝑆𝑛 ]]
𝑛

𝑛
Therefor𝑒 𝑛−1
is our bias adjustment.

Electronic copy available at: https://ssrn.com/abstract=4336419

Appendix B: Derivation of Expected Tracking Error Between Idiosyncratic
Implementations

Assume we have n potential implementations of a strategy (different rebalance dates). We can

define the return of a specific implementation as:

𝑟𝑖 = 𝑟̅ + 𝑢𝑖

Where 𝑟̅ represents the pure, zero-timing-luck strategy return and 𝑢𝑖 is the idiosyncratic return of
the ith implementation.

We would like to estimate the average variance of 𝑢𝑖 .

1 𝑛 1 𝑛 1 𝑛
𝑣𝑎𝑟[𝑢] = ∑ 𝑣𝑎𝑟[𝑢𝑖 ] = ∑ 𝑣𝑎𝑟[𝑟𝑖 − 𝑟̅ ] = ∑ [𝑣𝑎𝑟[𝑟𝑖 ] + 𝑣𝑎𝑟[𝑟̅ ] − 2𝑐𝑜𝑣[𝑟𝑖 ,̅𝑟]]
𝑛 𝑖=1 𝑛 𝑖=1 𝑛 𝑖=1

1 𝑛 1 𝑛 𝑛 2 𝑛
= ∑ [𝑣𝑎𝑟[𝑟𝑖 ] + 2 ∑ ∑ 𝑐𝑜𝑣[𝑟𝑗 , 𝑟𝑘 ] − ∑ 𝑐𝑜𝑣[𝑟𝑖 , 𝑟𝑗 ]]
𝑛 𝑖=1 𝑛 𝑗=1 𝑘=1 𝑛 𝑗=1

1 𝑛 1 𝑛 𝑛
= [∑ 𝑣𝑎𝑟[𝑟𝑖 ] − ∑ ∑ 𝑐𝑜𝑣[𝑟𝑖 , 𝑟𝑗 ]]
𝑛 𝑖=1 𝑛 𝑖=1 𝑗=1

We would also like to estimate the average variance of 𝑟𝑖 − 𝑟𝑗 where 𝑖 ≠ 𝑗:

2 𝑛 𝑛
∑ ∑ 𝑣𝑎𝑟[𝑟𝑖 − 𝑟𝑗 ]
𝑛(𝑛 − 1) 𝑖=1 𝑗=𝑖+1

Because we know that 𝑣𝑎𝑟[𝑟𝑖 − 𝑟𝑖 ] = 0, we can rewrite:

1 𝑛 𝑛
= ∑ ∑ 𝑣𝑎𝑟[𝑟𝑖 − 𝑟𝑗 ]
𝑛(𝑛 − 1) 𝑖=1 𝑗=1

1 𝑛 𝑛
= ∑ ∑ [𝑣𝑎𝑟[𝑟𝑖 ] + 𝑣𝑎𝑟[𝑟𝑗 ] − 2𝑐𝑜𝑣[𝑟𝑖 , 𝑟𝑗 ]]
𝑛(𝑛 − 1) 𝑖=1 𝑗=1

1 𝑛 𝑛 𝑛 𝑛
= ∑ [∑ 𝑣𝑎𝑟[𝑟𝑖 ] + ∑ 𝑣𝑎𝑟[𝑟𝑗 ] − 2 ∑ 𝑐𝑜𝑣[𝑟𝑖 , 𝑟𝑗 ]]
𝑛(𝑛 − 1) 𝑖=1 𝑗=1 𝑗=1 𝑗=1

Electronic copy available at: https://ssrn.com/abstract=4336419

 2 𝑛 𝑛 𝑛
= (𝑛 ∑ 𝑣𝑎𝑟[𝑟𝑖 ] − ∑ ∑ 𝑐𝑜𝑣[𝑟𝑖 , 𝑟𝑗 ])
𝑛(𝑛 − 1) 𝑗=1 𝑖=1 𝑗=1

2 𝑛 1 𝑛 𝑛
= (∑ 𝑣𝑎𝑟[𝑟𝑖 ] − ∑ ∑ 𝑐𝑜𝑣[𝑟𝑖 , 𝑟𝑗 ])
(𝑛 − 1) 𝑗=1 𝑛 𝑖=1 𝑗=1

2𝑛
Demonstrating that the average variance of 𝑟𝑖 − 𝑟𝑗 is equal to 𝑛−1 times the average variance of

𝑢𝑖 .

Electronic copy available at: https://ssrn.com/abstract=4336419