Suli Balazs Marton

Download as pdf or txt
Download as pdf or txt
You are on page 1of 55

Etvs Lornd University

Faculty of Science

Yield Curve Modeling


Thesis

Balzs Mrton Sli


Actuarial and Financial Mathematics MSc

Quantitative Finances Ma jor

Supervisors:

Dr. Andrs Zemplni


associate professor
Department of Probability Theory and Statistics

Dr. Daniel Niedermayer


Solvency Analytics

Budapest, 2014
.
Contents
1 Introduction 5
1.1 Fixed Income Securities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.2 Yield Curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.3 No-arbitrage Condition . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
1.4 Dierent Types of Curves . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

2 Interest Rate Modelling 12


2.1 One-factor Short Rate Models . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Connections . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.3 Two Popular Short Rate Models . . . . . . . . . . . . . . . . . . . . . . . . 16
2.4 Yield Curve Calibrating . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3 Statistical Yield Curve Models 20


3.1 Coupon Stripping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3.2 Interpolation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
3.3 Including Errors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.4 Parameterised Curves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.5 Polynomial Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.6 Spline Yield Curve Models . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.7 Smoothing Conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.8 Non-linear Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

4 Additional Features 35
4.1 Corporate Bond Valuation using Credit Spread . . . . . . . . . . . . . . . 35
4.2 Pricing with CDS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
4.3 Corporate Bond Spreads . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
4.4 Liquidity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

5 Applications 42
5.1 Data Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42
5.2 Modeling in Python . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.3 The results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
CONTENTS CONTENTS

6 Summary 51

7 Appendix 52
7.1 Appendix 1: An example of B-splines . . . . . . . . . . . . . . . . . . . . . 52

4
1 INTRODUCTION

1 Introduction
The main topic of this thesis is yield curve modeling. I have tried to collect the most
relevant information on that but still not to exceed the limits of an MSc thesis. The idea
of a thesis about yield curve modeling has come from the swiss Solvency Analytics group.
Reliable yield curve models can be very useful when calculating sensitivites and capital
charges of corporate bonds within the Solvency II framework. For the thesis to be useful
for Solvency Analytics, I have focused mostly on corporate bonds and I have chosen to
write it in English.
The rst few pages of the thesis is concentrated on concepts such as xed income
securities, risks aecting them, corporate bonds, YTM, zero-coupon yield curve, discount
curve, forward curve and no-arbitrage. After those the concepts of discount function and
instantenous forward rates are introduced. The next section of the thesis is about one
factor short rate models. After a general description of these types of interest rate models
two popular models are introduced: the Vasicek and Cox-Ingersoll-Ross models. In this
section, I have relied on the knowledge I have learned at the university lectures of Dr.
Gyrgy Michaletzky [1] and I used similar notations. The Statistical Yield Curve Mod-
els section presents some methods to model the yield curve based on observable market
prices and bond properties. It starts with a method called Coupon Stripping and after
that other types of yield curve models follow such as polynomial or spline-based models
and Nelson-Siegel type curves. I have relied on two books mostly: Handbook of Fixed
Income Securities [2] and Interest Rate Modelling [3]. The Additional Features section
presents some alternative but still popular ways to model the yield curve. They can be
very useful when the construction of statistical yield curve models are not possible. The
last section, Applied Methods summarizes the numerical implementations I have written
to be able to t some models to real data. I have focused on the polynomial and spline
estimation models here and presented some outputs of the apllications.

I would like to thank my supervisors Dr.Andrs Zemplni and Dr.Daniel Niedermayer


for the numerous advice and help they have provided me and made the following thesis
much better.

5
1.1 Fixed Income Securities 1 INTRODUCTION

1.1 Fixed Income Securities


Fixed income securities constitute a huge and important part of the nancial products
universe. There are a lot of dierent types of them, and as nancial markets became
more and more complex, the evolution of these products have started to speed up. The
basic xed income securities grant given cash ows on certain future dates to the security
holder. The payer of these cash ows is the issuer of the security. These are one of the most
simple type of nancial products, but even these are aected by several types of risks. It
is in the investor's interest to have a reliable mathematical model of the products, and a
good model always counts with the risks.
The most important type of risk aecting xed income securities is called interest-rate
risk. It is the risk arising from the constant change of the xed income securities market.
If interest rates increase, that means investors can expect a higher return on their new
investments in the market, and this lowers the value of older ones. The market value
of securities moves in the opposite direction of interest-rate movements, if rates rise the
market values fall while if rates fall the market values rise. Of course this type of risk does
not aect the investor who holds the security until maturity and doesn't plan to sell it
before that.
Another important risk aecting xed income securities is credit risk. The source of
this type is the issuer of the security and its ability to meet its duties. There is always
a possibility that the issuer can't or won't pay the amount when it must be payed, and
this possibility is dierent for dierent issuers. A bigger chance of default on the issuer's
side makes the security less valuable and this is usually compensated with higher yield,
to have someone buying them. Among others liquidity risk, currency risk and ination
risks also aect the price of a xed income security. To be able to precisely measure the
value of an investment, and to make solid investment choices, sophisticated mathematical
models are needed even for the above mentioned most simple types. In practice, there
are more complex features such as embedded options, seniority restrictions or convertible
bonds, which call for more complex models.
Corporate bonds refer to bonds issued by dierent corporations. The most important
dierence between governmental bonds and corporate bonds is the dierent credit risk
associated with the security. In most cases governmental bonds are considered to contain
very little credit risk, while some corporations have much bigger chance of default. This

6
1.2 Yield Curve 1 INTRODUCTION

extra risk beared by the investor is compensated with an increased return on the invest-
ment usually. Another distinction between governmental bonds and corporate bonds is
that corporate bonds usually contain additional features like convertibility or embedded
options, and are often less liquid than governmentals. These make the modelling of cor-
porate bonds harder mathematically, and less accurate models can be expected, than in
the case of governmentals. But as more and more complex nancial products arise, math-
ematical models become more and more sophisticated, trying to nd a solution for actual
nancial problems.

1.2 Yield Curve


How can the investor compare the return of several investments, and choose the one
that's the most approriate for his or her needs? There are dierent mathematical tools to
measure the return of an investment. In the case of coupon bonds, one of the most pop-
ular measure of return is Yield To Maturity. Let's consider a coupon bond with n years
time-to-maturity, xed $C coupon payments at the end of every year, and the redeem of
the principal which is $100 at the end of the n years. If the market price of this coupon
bond is PC now, the yield-to-maturity (YTM) is dened as:
n
X C 100
PC = i
+
i=1
(1 + Y T M ) (1 + Y T M )n

According to this equation, the YTM can be calculated every time the cash ow
pattern and the current market value of a security is known. Yield-to-maturity is an
easily computable, but not too sophisticated measure of a bond's yield. It assumes that
every cash ow is discounted with the same rate, which is not a realistic approach. In
practice cash ows further in time carry more risks, and usually should be discounted
with a larger discount rate than the closer ones. This is why the zero-coupon yield curve
is so important in the eld of modelling xed income securities.

7
1.3 No-arbitrage Condition 1 INTRODUCTION

For a zero-coupon bond the computation of the yield is more simple. Let PZ be the
current market price of the zero-coupon bond, and the only one cash ow it contains is
$100, t years from now. Then the eective yield (rE ) of this product comes from:

100
PZ =
(1 + rE )t

The compounded yield (rC ) comes from the equation:

PZ = 100erC t

The zero coupon yield curve refers to the illustration of zero-coupon yields of dierent
maturities in a given nancial market. This is one of the most important tools when
modelling xed income securities.

1.3 No-arbitrage Condition


There is an essential condition when modelling xed incomes in a given market and
this is the no-arbitrage condition. Economic considerations suggest that there should be
no arbitrage opportunities in well behaving nancial markets. If any opportunity occurs
for an arbitrage, the arbitrageur's activity will move the maket prices in a direction that
extinguishes it. This phenomenon is very useful in mathematical modeling. It creates
a continuous and solid connection between the prices of nancial products in relating
markets. In the case of xed income securities, it is the no-arbitrage condition that ensures
that the knowledge of zero coupon yields itself denes the prices of coupon bonds in that
market.
In general it is possible to buy a coupon bond and sell its cash ows in tranches, like
dierent zero coupons. Because of that, the prices of this two types of bonds should be
in some kind of harmony. The price of a coupon bond should be the sum of the prices of
the zero-coupons of the cash ows it contains, or else an arbitrage opportunity arises.
Now, with the use of no-arbitrage conditions let's see how the zero-coupon yield curve
determines the prices of xed income securities. It is important to bear in mind that yield
curves are abstract theoretical concepts, usually adjusted to real market prices to model

8
1.4 Dierent Types of Curves 1 INTRODUCTION

reality accurately. Now, suppose that in a mathematical model of a nancial market, the
zero coupon yield curve is known. Let r(t) be the compounded yield of a zero coupon
bond maturing in t years for every t > 0. In this model, a coupon bond with known cash
ow properties can only have one specic price. With the use of the yield curve this price
can easily be determined. Assume that bond B has Ci cash ows at times ti . Then PB ,
the theoretical price of bond B is dened as:
X
PB = Ci er(ti )ti
i

This equation exhibits a strong relationship between the theoretical yield curve of a model,
and the prices of securities in the nancial market. If the market price of B diers from PB
an arbitrage opportunity arises in the model and that is inconsistent with the no-arbitrage
condition. It is important to mention that in practice, there is an error associated with
market prices and so there will always be a small dierence between the theoretical price
and the market price. The model is estimated usually by somehow minimizing these errors.
As shown, the knowledge of the yield curve is crucially important when pricing secu-
rities but it is an abstract concept. To have a model that describes reality well the yield
curve should be approximated with the use of market prices of securities in a market.
The starting set of securities of this approximation is very important. If one wants to
use the information provided by the yield curve, for example to anticipate the price of a
bond waiting to be issued, it is crucial that the curve is typical of that type of bond. The
estimation should start from similar type of products. Since the yield curve represents
a relationship between maturity and return, the types of risks (other than interest-rate
risk) should be considered when collecting the starting set of bonds.

1.4 Dierent Types of Curves


The three most popular types of curves used to represent the term structure of interest
rates are the discount curve, the zero-coupon yield curve and the instantenous forward
curve. They present the same core information in dierent ways and since that they can
be computed from each other. Let us denote the discount function with d(t). It refers to
the present value of 1 dollar (if dollar is the currency one wants to use) received in t years

9
1.4 Dierent Types of Curves 1 INTRODUCTION

(if a year is the time unit one wants to use). The yield of a zero-coupon bond maturing in
t years is written as r(t). The term structure of these are called the spot or zero-coupon
yield curve or shortly yield curve. The forward rates express the market's expectations of
interest rates in the future. On an investment beginning at time t and ending at time T
(where t < T ) the market expects a yield equal to f (t, T ).
The following relationships stand among these three rates. If the spot rates are known,
the discount rates can be calculated as:

d(t) = er(t)t

and conversely:

log d(t)
r(t) = . (1)
t

The forward rate can be obtained from the spot rates with the use of:

T r(T ) tr(t)
f (t, T ) = (2)
T t

The spot rate is the following forward rate trivially:

r(t) = f (0, t).

The conversion between the discount rates and the forward rates can be easily obtained
using the spot curve and the mentioned equations.

There is one more curve of importance when describing the term structure of interest
rates and that is the instantenous forward rate curve. The forward rates mentioned before
needed two parameters: the beginning of an investment t and the end T . If the spot yield
curve is dierentiable it is possible to dene the instantenous forward rate: f (t). The
instantenous forward rates represent the expected yield of an investment beginning at t
and ending in t + t when t 0. It describes the present market expectations of the

10
1.4 Dierent Types of Curves 1 INTRODUCTION

evolution of future short rates. Using equation (2), f (t) can be expressed as:

f (t) = lim f (t, t + t) =


t0
(t + t)r(t + t) tr(t)
= lim =
t0 t + t t
tr(t + t) + t(r(t + t) r(t))
= lim =
t0 t
r(t + t) r(t) r(t)
=r(t) + t lim = r(t) + t
t0 t t

In some yield curve models it is better to estimate the instantenous forward rate curve
and the spot and discount curves can be calculated from it.

11
2 INTEREST RATE MODELLING

2 Interest Rate Modelling


2.1 One-factor Short Rate Models
This section provides a few terms and concepts on the type of interest rate models
called one-factor short rate models. Short rate models are widely used mathematical
models to describe the stochastic evolution of dierent interest rates. They are called short
rate models because the basic concept of this model is the instantenous short rate r(t). It
refers to the (annualized and compounded) yield that can be earned on an innitesimally
short investment at time t. It is important to emphasize that altough the notation is
the same, this r(t) is not nearly the same as the one mentioned in the previous chapter.
Here, the t parameter refers to time in a model and not the time-to-maturity feature of a
product observed now.
In the ever-changing world of nancial markets the interest rate on actual short term
investment opportunities can vary quickly and unpredictably in time. The rates on longer
investments also change in time but they are considered much less volatile. In one-factor
short rate models the instantenous short rate's evolution is dened by a stochastic pro-
cess, usually an Ito-process under the risk-neutral measure:

dr(t) = a(t)dt + b(t)dW (t)

In this equation a(t) and b(t) are time-dependent coecients of the stochastic process,
and W (t) is a Wiener-process under the risk-neutral measure.
There is an other interest-rate process in this model and that is the forward rate. The
forward rate is the expected (annualized and compounded) instantenous interest rate of
time T , at time t. Of course T must be higher or equal than t. In the model the forward
rate is a stochastic process with two time parameters:

dt f (t, T ) = (t, T )dt + (t, T )dW (t) (3)

Here the (t, T ) and (t, T ) coecients are the functions of two time parameters.
The short rate model also incorporates two value processes of nancial products. One
of them is the value process of a zero coupon bond. For convenience the face value of

12
2.2 Connections 2 INTEREST RATE MODELLING

this zero coupon bond is set to be $1. Let's consider a $1 zero coupon bond maturing at
time T . The value of this bond at time t is denoted by P (t, T ). Since the values of bonds
depend on the level of interest rates and expectations of these, the evolution of P (t, T ) is
also described with a stochastic process in the model:

dt P (t, T ) = P (t, T )(m(t, T )dt + v(t, T )dW (t))

Here the m(t, T ) and v(t, T ) coecients are functions of two time dependent parameters.
It is important to mention that this mathematical model presumes that there is a bond
maturing at every time t.
Another participant in this model is the value of a bank deposit B(t). In opposite of
the bond where present value is derived from a future cash ow, the bank deposit's value
origins from the past. Bank deposits pay the timely short interest rate at every time t
and are doing it compoundedly. That's why the evolution of the value of a bank deposit
depends on the evolution of the short rate. The connection between the two is:

dB(t) = r(t)B(t)dt

The value of the deposit at time t can also be expressed as:


Rt
B(t) = B(0)e 0 r(s)ds
, (4)

where B(0) is the value of the deposit at t = 0.

2.2 Connections
There are certain connections between these processes. From Equation (4) it is appar-
ent that with the knowledge of B(0) the short rate process determines the value process
of the bank deposit. There is a connection between the value of the bond and the forward
rate process. The present value of a bond should be the discounted value of the future
cash ow. At time t the expectations of future interest rate levels are the f (t, s) values

13
2.2 Connections 2 INTEREST RATE MODELLING

where s is larger or equal than t. A fair valuation of the bond in this model should be:
RT
P (t, T ) = e t f (t,s)ds
(5)

This is the appropriate value here as the expected future value of a bank deposit of size
P (t, T ) at time t is exactly one dollar at time T :

RT RT
Bexp (T ) = B(t)e t f (t,s)ds
= P (t, T )e t f (t,s)ds
= e0 = 1

If it wouldn't be one dollar an arbitrage opportunity would arise.


Expressing the forward rate from Equation (5):


f (t, T ) = ln P (t, T )
T

Using these strong connections between these two processes the knowledge of one of them
results that the other one can be expressed too. Assuming that the forward rate is known
in the form presented in Equation (3), so the coecient functions (t, T ) and (t, T )
are known. It can be shown that the coecients of the bond value's process m(t, T ) and
v(t, T ) can be expressed using (t, T ) and (t, T ) functions as follows:

1 2
m(t, T ) = f (t, t)
(t, T ) + (t, T )
2
and
v(t, T ) =
(t, T )

The (t, T ) and (t, T ) functions are integrals of (t, T ) and (t, T ):
Z T

(s, T ) = (s, u)du
s
and Z T

(s, T ) = (s, u)du
s

14
2.2 Connections 2 INTEREST RATE MODELLING

This denes the coecients of the bond's value process perfectly. The other way of ex-
pressing one from the other is also available. Assume that the stochastic process of the
bond value is known, so the coecient functions m(t, T ) and v(t, T ) are known. Using
this information and the connection between the two processes, the coecients (t, T )
and (t, T ) of the forward rate can be expressed:

(t, T ) = v(t, T )vT (t, T ) mT (t, T )

and
(t, T ) = vT (t, T ),

where:


mT (t, T ) = m(t, T )
T
and

vT (t, T ) = v(t, T ).
T

These equations dene a two-way connection between the bond value process and the
forward rate process.
Now let's consider the short rate process and its possibilities. It can be shown that
assuming the knowledge of the forward rate's evolution the short rate process can be ex-
pressed. The short rate process's a(t) and b(t) coecient functions take the following form:
Z t Z t
a(t) = (t, t) + T (s, t)ds + fT (0, t) + T (s, t)dW (s)
0 0
and
b(t) = (t, t)

In these expressions the T in the lower corner means the function's derivative by its sec-
ond variable similarly to mT (t, T ) and vt (t, T ) mentioned above.

15
2.3 Two Popular Short Rate Models 2 INTEREST RATE MODELLING

But what if one knows the short rate process and wants to express the forward rate from
that? It is provable that it's impossible to derive the forward rate or bond value processes
merely from the knowledge of the short rate process only, because an equivalent martingale
measure can't be obtained. The forward process contains an additional information and
that is the market's expectations of the future. It is needed to observe the market prices
of bonds and nd the appropriate market price of risk process, to be able to derive the
forward rate process and the bond value from the short rate.

2.3 Two Popular Short Rate Models


The following two models are among the most popular one factor short rate models.

Vasicek: The Vasicek-model uses a mean-reverting stochastic process with a constant


diusion coecient to model the short rate process's evolution:

dr(t) = (k r(t))dt + dW (t),

where t 0, k 0, > 0, > 0.


W (t) refers to a Wiener-process under the P probability measure. The P probability
measure is the one where the bank deposit is the numeraire process. This measure can
be obtained by observing the market prices of bonds and using the market price of risk
process with Girsanov's theorem. Under this measure the bond's value can be expressed as:
   R 
1 T
t r(u)du
(6)

P (t, T ) = B(t)EP F(t) = EP e F(t)
B(t)

CIR: The Cox-Ingersoll-Ross model constructs the short rate process from the sum
of squared, independent Ornstein-Uhlenbeck processes [13] . This way the result is a
non-negative stochastic process. Let us begin from the Ornstein-Uhlenbeck processes:
X1 , X2 , . . . , Xd for a positive integer d. Each of them can be described as:


dXj (t) = Xj (t)dt + dWj (t),
2 2

16
2.4 Yield Curve Calibrating 2 INTEREST RATE MODELLING

where the following stand: > 0, > 0 and Wj (t)-s are independent Wiener-processes.
Now the short rate process is:
d
X
r(t) = Xj2 (t)
j=1

The stochastic evolution of the short rate can be expressed as:


  p
dr(t) = r(t) dt + r(t)dW (t),

where

2
=d
4
and
d
X X (t)

dW (t) = pj dWj (t).
j=1 r(t)

This last equation relies on Levy's theorem.


In the Cox-Ingersoll-Ross model the dynamics of the zero coupon bond is just like in the
Vasicek-model:
   R 
1 tT r(u)du

P (t, T ) = B(t)EP F(t) = EP e F(t)
B(t)

2.4 Yield Curve Calibrating


Short rate models like these are very useful when one wants to analyze the evolution
of interest rates in time, but are they consistent with usual yield curve shapes observable
in the markets? Let's see the Vasicek model as an example.
According to Equation (6) the spot discount function can be evaluated at every future
points of time. The discount function d(t) is really the function of bond prices P (0, t)

17
2.4 Yield Curve Calibrating 2 INTEREST RATE MODELLING

evaluated now. Using Equation (6) it is the following in the Vasicek-model:


 R 
0t r(u)du

d(t) = EP e F(0)

It is easy to see that with the knowledge of the parameters of the model (k, and
) and r(0) the discount function can be constructed. Also the zero-coupon yield curve,
which is denoted by z(t) can be computed from the discount function easily. It takes the
following parametric form:

k t  2 1 t 2
z(t) = r(0)et e 1 e 1
2 2

This is a combination of et and e2t functions of time. With the three parameters of
r(t) it isn't always possible to match the shape of current yield curves. This is a short-
coming of the Vasicek-model and other parametric one-factor short rate models. If one
wants an interest rate model that can match the actual observable yield curves some im-
provements has to be made.

Extended Vasicek-model: An improved version of the original constant parame-


tered Vasicek model is the Extended Vasicek model. In order to be able to calibrate the
model to actual yield curves one of the constant parameters of the Vasicek-model is made
to be time-varying. The dynamics of the extended model is:

dr(t) = k(t) r(t) dt + dW (t),




where k(t) is an arbitrary deterministic function of time, and and are positive con-
stants. Deriving the actual zero coupon yield curve from this model results in:
t
2 1 t
Z
t
2
z2 (t) = r(0)e + k(u)e(tu) du + e 1
0 2 2

It is possible to nd a k(t) function that is appropriate in a sense, that the z2 (t) function

18
2.4 Yield Curve Calibrating 2 INTEREST RATE MODELLING

of the interest rate model will match the market zero coupon yield curve, derived from
observable prices of bonds.

A general method of calibrating interest rate models to market yield curves can be
shown through the CIR-model. Let z(t) be a process that is used for modelling short rate
dynamics in a CIR model:
p
dz(t) = z(t) dt + z(t)dW (t)


The short rate process now is a sum of z(t) and a deterministic function y(t):

r(t) = z(t) + y(t)

Here the value of a bond, P (t, T ) takes the following form:


RT
P (t, T ) = PCIR (t, T )e t y(u)du

The deterministic y(t) function can be found as follows:

y(t) = f (0, t) fCIR (0, t)

This approach is suitable to other kinds of short rate models as well. The types of yield
curve models that can be derived from interest rate models, like the ones mentioned above
called consistent. The need for model consistency depends on the goal of the modelling. In
some cases it is very important to be consistent with an interest rate model, while there
are other cases as well where other features are more desirable. The following section
introduces models where consistency isn't among the main requirements.

19
3 STATISTICAL YIELD CURVE MODELS

3 Statistical Yield Curve Models


There are several methods for modelling the yield curve, based on observable market
prices of securities. Some of them are more theoretical while others are often used in prac-
tice. Dierent situations appeal to dierent yield curve models. Each model has its own
advantages and disadvantages. They can dier for example in tractability, consistency
with stochastic interest rate models or in the type of rates they estimate. The discount
curve, the spot curve and the forward curve can also be the goal of a method. Once one
of this three curves is estimated, the others can be easily calculated from the one known.
In the followings, yield refers to the annualized and continuously compounded yield of a
security.

3.1 Coupon Stripping


Let us start with an easy model which is mostly theoretical. The goal is to estimate a
discount curve that is consistent with the market prices of securities of that kind. In most
of the situations one has to start from a set of coupon bond prices and properties and
derive the discount curve from that. A method called `Coupon Stripping' or `Bootstrap-
ping' [2] is an easy way to obtain the discount function's value on specic dates though
it requires very special and unrealistic starting set of information. This approach can be
used only if the following are given: a set of coupon bonds with cash ows like that on
every cash ow date there is exactly one maturing coupon bond.

Let's consider an example just to illustrate the method. Assume that we look at 10
bonds in the market with market prices Pi and annual coupon payments of size Ci and
the i-th bond matures i years from now:

i 1 2 3 4 5 6 7 8 9 10
Ci 4 7 3.5 4.25 2 7 6.5 3 4.5 3.25
Pi 101.74 107.01 98.5 99.68 87.93 108.89 101.6 76.71 83.36 72.18

20
3.1 Coupon Stripping 3 STATISTICAL YIELD CURVE MODELS

The goal is to estimate the discount function's value at bond maturity dates. The rst
step uses the rst period only. The discount function's value is:

P1 101.74
d(1) = = = 0.98
100 + C1 104

The next step uses the value resulted from the step before and the next bond's properties.

P2 C2 d(1) 100.15
d(2) = = = 0.94
100 + C2 107

Following this method every new discount factor can be calculated with the properties of
the bond maturing at that date and the past discount factor values:
Pn1
Pn i=1 Cn d(i)
d(n) =
100 + Cn

The resulting discount function of this method is:

d(1) d(2) d(3) d(4) d(5) d(6) d(7) d(8) d(9) d(10)
0.98 0.94 0.87 0.84 0.79 0.73 0.64 0.58 0.52 0.48

It is useful to set d(0) as 1 because the present value of money available now is its
face value.

With a linear algebraic approach the coupon stripping method can be presented as fol-
lows. Let P be the vector of coupon bond prices maturing on dierent incremental dates:
ti i {1, . . . , n} and C be the cash-ow matrix such as C(i, j) denotes the i-th bond's
cashow on date tj . From now on C(i, j) is denoted by Ci (tj ) for convenience. C is a
square matrix due to the restrictions on the dataset. The discount factors on each ti date
are denoted by d(ti ) and D = (d(t1 ), . . . , d(tn )) is a vector comprised of them.

21
3.2 Interpolation 3 STATISTICAL YIELD CURVE MODELS

In an arbitrage-free solution the bonds' prices should be equal to the discounted present
values of their cash ows, so

Pi = d(t1 )C(i, t1 ) + + d(tn )C(i, tn )

should stand for every i {1, . . . , n}. Now by solving the CD = P linear system D is
obtained and so the discount values on the dened ti dates.

As it was mentioned before this approach is mostly theoretical and rarely used in
practice. One of the main problems is that a starting set of information as it is required
for the Coupon Stripping method is very unrealistic in practice. There is rarely available
a representative set of similar type of bonds maturing exactly one years (or any other
given period) from each other in the market.

3.2 Interpolation
To have a discount value for every t in a time interval, one can consider the discount
curve linear between the dates where it's values are known. With this a continuous but
not smooth (in most cases the rst derivative is not continuous) yield curve is estimated.
Another way to t a continuous curve onto the points known can be using polynomial
approximation techniques. In spite of the fact that a perfect t on the vertices can be
obtained by a polynomial function with the use of Lagrange approximation, it is not
suitable in practical use. Although the estimated curve is innitely dierentiable and
connects the points this approach doesn't result in realistic yield curves. This is because
between the initial points and usually on the short and long ends of the time interval very
unrealistic shapes can occur.
Figure 1 shows the discount factor values obtained by the coupon stripping method
example before. The dark blue marks represent the estimated discrete discount factor
values. The dark blue line is a linear interpolation while the light blue is a Lagrange
polynomial interpolation. It can be seen that the Lagrange interpolation results in some
extra curvatures at both ends of the sample.

22
3.2 Interpolation 3 STATISTICAL YIELD CURVE MODELS

Figure 1: Linear and Lagrange interpolation of the discount function.

While the dierence between the two interpolation method is not that big in the case
of the discount function, when computing the spot yield curve it becomes much more
signicant. The spot yield cuve is computed from the values of the discount curve using
the connection mentioned in Equation (1). Figure 2 shows the resulting curves. The dark
blue marks represent the yields computed from the original results of the coupon strip-
ping. The dark blue is the yield curve estimated from the linearry interpolated discount
function while the light blue is the one derived from the Lagrange interpolated discount
function.

23
3.2 Interpolation 3 STATISTICAL YIELD CURVE MODELS

Figure 2: Yield curves calculated from the interpolated discount functions.

The instantenous forward curve approximation derived from the spot yield curve is pre-
sented in Figure 3. This curve is computed by dividing the time interval between every
year into ten equal small intervals. Then using Equation (2) to calculate f (ti ) as:

f (ti ) = f (ti , ti+1 ),

for every i {0, 1, . . . , 99} where this represents the new time lattice.
Because of the absence of a continuous rst derivative in the case of linear interpola-
tion, this approximation of the instantenous forward curve becomes very ragged as shown
in Figure 3 . It is a good example of why a smooth curve is required. While the forward
curve computed from the Lagrange-interpolation is smooth, it has apparently too much
curvature which makes the results unrealistic.

24
3.3 Including Errors 3 STATISTICAL YIELD CURVE MODELS

Figure 3: Forward curves derived from the spot curves.

3.3 Including Errors


The Coupon Stripping method doesn't count with errors in the prices of securities.
In reality the bond market prices rarely equal the theoretical prices obtained (sum of
the discounted cash ows). There can be smaller or larger dierences arising from several
reasons like liquidity, the cash ow structure or tax eects. It is reasonable to include an
error term in the arbitrage-free equation:

Pi = d(ti )C(i, t1 ) + + d(tn )C(i, tn ) + ei ,

where ei refers to the error term, the dierence between the theoretical and quoted market
price of the i-th bond. This approach results in more realistic models of the yield curve
but requires more sophisticated estimation methods.

25
3.3 Including Errors 3 STATISTICAL YIELD CURVE MODELS

Suppose a set of coupon bond prices and properties are known, now without the re-
strictions stated before. Let P be the vector comprised of the market prices of the bonds,
C(i, j) the cashow matrix and D the vector of the discount values. In the Coupon Strip-
ping case the cashow matrix is a squared matrix due to the strict initial conditions.
Without those restrictions C has usually much more columns than rows and has a lot
of zero values. Let be the vector of the ei error terms as mentioned above. Now the
following holds:

P = CD +

One way to estimate D is by minimizing the error terms. Without any other restrictions
on D this can be done by Ordinary Least Squares (OLS) estimation. The task is to nd
D where

D = min T | = P CD

D

The solution of the OLS regression is

D = (C T C)1 C T P

It is important to mention that C T C is not always an invertible matrix, but with the
use of a pseudo-inverse this problem is solvable. Although this is an elegant approach in
theory, it is not a good choice in practice. Even with carefully chosen starting data very
unrealistic shapes can occur as a result if one wants to somehow produce a continuous
function out of the estimated values. This method doesn't place any restrictions on the
d(t) function and it is a source of some problems. One of them is that estimated values on
similar maturities often don't have similar values which results in a ragged curve. That is
not realistic at all.

26
3.4 Parameterised Curves 3 STATISTICAL YIELD CURVE MODELS

3.4 Parameterised Curves


Using OLS technique without any restrictions on the d(t) discount function is not a
good way of modelling the yield curve. To obtain better results it is benecial to look
for more specic curves only. A popular approach is to look for curves which are param-
eterised functions of time-to-maturity. For example if a model's goal is to estimate the
d(t) discount curve one should look for a d(t; a, b, ...) parameterised function of time-to-
maturity with parameters a, b, . . . etc. These are called parameterised yield curve models.
A big advantage of using parameterised yield curve models is that they result in real curve
and not just some set of individual estimated discount factors. An other advantage is that
much less parameters are needed to be estimated than in the simple OLS regression case
where there was a parameter for every cashow time.
Parameterised yield curve models can be linear or non-linear. Linear curves can be ex-
pressed as the linear combination of some basis elements and so they are easy to optimise
for a best t. Non-linear curves can not be expressed like that. In linear models it is much
easier to obtain a best t because only the linear coecients should be estimated and
OLS regression provides a good way of doing this. For a given set of i (t) basis functions
where i = 1, 2, . . . , K , the function looked for can always be expressed as:
K
(7)
X
d(t) = i i (t)
i=1

Now a vector of parameters (1 , . . . , K ) determines a d(t) function so one has to estimate


these parameters only to obtain a yield curve.
With the use of = (1 , . . . , K ) and as a matrix like that (i, j) = j (ti ) where
i = 1, . . . , n and j = 1, . . . , K , the vector D = (d(ti ), . . . , d(tn )) can be expressed as:

D =

According to the basic problem the following also stands:

P = CD +

27
3.5 Polynomial Estimation 3 STATISTICAL YIELD CURVE MODELS

Now with the use of the notation:

D
b = C, (8)

the linear problem can be expressed as:

P = D
b + (9)

The solution can be easily obtained with the use of OLS regression by minimalizing T .
It is:

= (D b 1 D
b T D) bT P (10)

Once the vector of the i coecients are estimated, the discount function, d(t) can be
obtained with the use of Equation (7) .

3.5 Polynomial Estimation


One of the easiest ways to model the yield curve is to look for a polynomial func-
tion that ts the data most accurately but still has the characteristics of realistic yield
curves. High degree polynomials would t the data probably very well but they usually
result in unrealistic shapes. The most popular way is to nd a cubic polynomial that
ts the data well. Polynomial yield curve estimations are linear yield curve models since
every n-th degree polynomial can be expressed as the linear combination of the xj power
functions for j {0, . . . ., n}. Using the xj power functions as a basis, the discount curve is:
n
X
d(t) = j xj (t)
j=0

With those basis functions 's elements are: (i, j) = xj (ti ).


The solution can be obtained using Equation (8), Equation (9) and Equation (10). While
in some rare cases polynomial yield curve estimations lead to quick and good solutions

28
3.6 Spline Yield Curve Models 3 STATISTICAL YIELD CURVE MODELS

they are not the best choice most of the times. A more appropriate model is spline
approximation.

3.6 Spline Yield Curve Models


Using splines to model the yield curve can lead to good results. Spline yield curve
models are linear in the coecients so they are easy to t. Splines are piecewise polynomial
functions with some restriction on the derivatives. It is called knot points where the
dierent polynomial parts meet. If the spline's domain is a closed interval the ends are
sometimes called knot points too. This section considers the ends as knot points too. A
k -th degree spline is a function that is a k -th degree polynomial between the knot points
and k 1 times dierentiable everywhere. Because of that a spline of order m with n knot
points needs only n + m 1 parameters to be fully determined. The polynomial on the
rst interval is dened by m + 1 parameters and all the other parts need only one more
since their derivatives must be the same in the knot points. There are n 2 parts other
than the rst, so this gives m + 1 + n 2 = m + n 1. As mentioned spline models are
linear in the coecients so rst let's introduce the basis functions. A good choice is to use
basis splines or B-splines. It is possible to dene the basis splines recursively. If a dened
set of knot points is known: {P1 , . . . , Pn } , the denition of B-splines of order zero is the
following:
if Pi t < Pi+1
(
1
B0,i (t) =
0 else

In this denition the lower index i refers to the i-th basis spline available for denition
on the set of knot points in an incremental order. Starting from the basis spilnes of order
zero, it is possible to dene B-splines of higher order recursively:

t Pi Pi+m+1 t
Bm,i (t) = Bm1,i (t) + Bm1,i+1 (t),
Pi+m Pi Pi+m+1 Pi+1

where the lower index m refers to the order (also called degree) of the spline.

Bm,i (t) is a spline function of order m that is zero outside the [Pi , Pi+m ] interval. Be-
cause of these on a set of n knot points n m 1 basis splines of order m are available.

29
3.6 Spline Yield Curve Models 3 STATISTICAL YIELD CURVE MODELS

In Appendix 1, I give an example of B-splines on a set of knot points.

As mentioned every m-th degree spline with n knot points on a closed interval needs
n + m 1 parameters to be determined. So for example a cubic spline with knot points
{P1 , . . . , Pn } needs n + 2 parameters and so n + 2 basis functions to determine the spline.
There are n 4 B-splines that belong to the knot points so there are 6 more basis func-
tions required. To get those 6 more `out of interval' knot points should be dened such
as: P2 < P1 < P0 < P1 < < Pn < Pn+1 < Pn+2 < Pn+3 . This new set of knot
points determines n + 2 B-splines. Now the basis can be this n + 2 basis splines restricted
to the interval [P0 , Pn ]. The d(t) discount function can be expressed now as the linear
combination of the basis functions:
n1
X
d(t) = i B3,i (t)
i=2

Let's use the notations = (2 , 1 , . . . , n1 ) and B as a k (n + 3) sized matrix


like that B(i, j) = B3,j (ti ), where there is k cash ow times. If D denotes the vector
(d(t1 ), . . . , d(tk )) the following equation stands:

D = B

Now denoting D
b = CB and minimizing the term T using OLS techniques, the solution is:

= (D b 1 D
b T D) bT P

This is called the solution of the unconstrained problem. If d(0) = 1 is required it is


called the constrained problem. The constrained problem gives a better t on the short
end of the yield curve by xing the rst value to 1. This is a reasonable requirement in
accordance to practice.

30
3.7 Smoothing Conditions 3 STATISTICAL YIELD CURVE MODELS

By denoting the vector of the B-splines values at 0 as W , like that:

W = (B3,2 (0), B3,1 (0), . . . , B3,n1 (0))

and set w = 1, the constraint becomes the following additional requirement in the prob-
lem:

W = w

Using this condition the problem can be expressed as:

= min T = P D,

b W = w

And the solution of this problem is:

= (D b 1 W T (W (D
b T D) b 1 W T )1 (W w).
b T D)

The results of spline methods like these rely heavily on the selection of the knot points.
One suggestion made in the literature is to try to position the knot points so there are
similar number of observations between them. This produces better ts or more realistic
curves usually but sometimes there are exceptions when it is better to try an other starting
set of knot points.

3.7 Smoothing Conditions


The OLS method used for spline approximations minimized the following criterion
with changing the parameter:

c1 = (P CB T )2

31
3.7 Smoothing Conditions 3 STATISTICAL YIELD CURVE MODELS

This c1 term refers to the goodness-of-t of the curve only in consideration of the distance
between the tted curve and the observed prices. A smaller c1 value means the observed
values are closer to the tted curve. On one hand this a desirable property but on the
other hand it can carry some adverse eects on other important properties of the yield
curve such as smoothness. It is showed by Daz and Skinner that signicant liquidity and
tax induced errors occur when modelling corporate yield curves compared to Treasury
yield curve models [9]. The universe of corporate bonds are less heterogenous by nature
than governmentals. Sometimes it is better to have a yield curve with less curvature than
to try to overt the values. It can be reasonable to use some further criterion that controls
the smoothness of the estimated curve. These are called Smoothing Criterions and there
are several dierent ones. The total curvature of an estimated d(t) curve between P0 and
Pn can be measured as:
Z Pn  m1 d(t) 2
c2 = dt
P0 tm1

for an m-th order spline approximation.


An m-th order spline is called natural if the following is true:

m1 d(t) m1 d(t)
= =0
tm1 P0 tm1 Pn

The natural smoothing criterion is the following:

= min{c2 |c1 < S}


This criterion minimizes the curvature of a t while keeping c1 's value under S . A big-
ger S value places more importance on smoothness and less on the close t of the values
while a smaller S emphasizes the t more. The solution of the natural smoothing criterion
problem is a (2m 3)-th order natural spline with knots at the data points. That is why
it is called the natural smoothing criterion.

Fisher, Nychka and Zervos's studies show that better ts can occur if the forward rate
curve's curvature is restricted and not the discount function's [4]. The following is the

32
3.8 Non-linear Models 3 STATISTICAL YIELD CURVE MODELS

smoothing spline criterion:


Z T 2
 2 h(t|, )
c3 = 2.
dt + (P Cd(t|, ))
0 t2

where h is an invertible function of the yield curve, is the set of knot points: =
(P3 , P2 , . . . , Pn+3 ) and is called the smoothing parameter. Now it is possible to use
the curvature for example of the yield curve, the forward rate curve or the discount curve
as a basis for the approximation. carries the relative importance between the curvature
and the t in this approach. A large lays more weight on the rst part of the criterion.
An extension is to let vary by time. This method was proposed by Waggoner [11]. It
can be reasonable since yield curves usually have much more curvature in the short end
than in the long. Using the time-varying parameter, the criterion becomes:
Z T 2
 2 h(t|, )
c3 = (t) 2.
dt + (P Cd(t|, ))
0 t2

3.8 Non-linear Models


There are some popular non-linear yield curve models. The most widely used is the
Nelson-Siegel model. This is a four parameter model that is used to estimate the forward
rate curve. It is very popular thanks to the relatively low parameter number and easy
computation. In spite of its small parameter number it can capture most of the yield
curve shapes. The disadvantage of the Nelson-Siegel model lies in precision. It is not a
good choice if very accurate estimations are required or when one is trying to model a
complex yield curve. The four parameters are: 0 , 1 , 2 , k. The forward rate curve esti-
mation is:

f (t) = 0 + (1 + 2 t)ekt

From this expression of the forward rate curve the spot curve can be written as:

33
3.8 Non-linear Models 3 STATISTICAL YIELD CURVE MODELS

 2  1 ekt 2 kt
r(t) = 0 + 1 + e .
k kt k

The short rate r(0) equals 0 + 1 and the long rate is limtinf r(t) = 0 . The other two
parameters 2 and k can adjust the height and location of a hump in the curve.
An extension of the original Nelson-Siegel model is the Svensson model. It is sometimes
called Nelson-Siegel-Svensson model since it is the original model with two more param-
eters:

f (t) = 0 + (1 + 2 t)ek1 t + 3 tek2 t

Wiseman [12] introduced a 2(n + 1) parameter model to estimate the forward curve. It
consists of n + 1 exponential decay terms with i coecients:
n
X
f (t) = i eki t
i=0

34
4 ADDITIONAL FEATURES

4 Additional Features
4.1 Corporate Bond Valuation using Credit Spread
There was a very important criterion of the statistical yield curve models of the pre-
vious section and that is the availability of data. All of the models are based on the
information a carefully gathered dataset of similar bonds contain. If for example one
wants to obtain a yield curve to use for pricing a bond, the starting set of information
should consist of other bonds with similar credit risk, liquidity, currency and country of
the bond that ought to be priced. That way the resulting yield curve provides a useful
tool for pricing the bond. If the information basis of the model is not chosen well it will
not provide an appropriate tool for valuation. But what if there aren't enough observable
information or simply there aren't enough time and resources to use a statistical yield
curve model? Situations can occur where only a quick valuation is needed and there isn't
enough time to carefully select the best datasets and precisely adjust statistical models.
And it can also happen that there is too little information available on a nancial product
or issuer company. There are ways to estimate a yield curve and to evaluate a security
without the need to construct statistical yield curve models.
The second most important type of risk aecting xed income securities is credit
risk (after interest-rate risk). Credit risk, more importantly the perceived credit risk of
investors, is an important factor that shapes the yield curve and aects the price of
a security aected by credit risk. When a company's perceived credit risk grows, that
means investors think that default is more probable than before. Investors can analyze
individually the credit risk of a company but the vast majority relies on the opinion
of big credit rating agencies. The biggest three are Standard and Poor's, Moody's and
Fitch Group. If a company becomes downgraded by one of these companies, it means its
creditworthiness has worsened and this implies that a larger yield margin is expected by
investors to buy the bonds' of that company. Skinner and Daz (2001) showed that it is
reasonable to pool bonds by broad rating agency categories when constructing statistical
yield curve models. Doing this doesn't aect the results signicantly in a bad way. They
also found that corporate yield curve models contain signicant liquidity and tax-induced
errors and thus they are less reliable than Treasury yield curve estimates. [9]
There are some rates in the markets considered (credit) risk-free. The U.S. Treasuries's

35
4.2 Pricing with CDS 4 ADDITIONAL FEATURES

yields as well as the LIBOR swap rates are usually the most popular when referring to risk
free rates. They contain nearly zero credit risk and that's why they are usually used as
benchmark rates. A not too precise approach is that corporate bonds dier from risk free
securities like U.S. Treasuries only in their perceived credit risk. Of course it is not true
since for example the U.S. Treasuries are usually much more liquid products than some
corporate bonds, but this approach can provide an easy and quick way with relatively
good results when estimating a yield curve or pricing products. If credit risk is the only
dierence in a model from the risk-free securities it is reasonable to use the risk-free yield
curve as a starting point and to transform it somehow to cope with the extra credit risk. A
big advantage of this approach is that precise statistical yield curves for risk-free interest
rates like U.S. Treasuries or LIBOR swap rates are usually available. Now one only has
to use some transformations to take in account the credit risk. One of the easiest ways is
to use a paralell shift of the original risk free curve of to get the yield curve of the credit
risky product. The dierence between the yield of a credit risky product and a risk-free
product of the same maturity is called credit spread.

4.2 Pricing with CDS


If there are CDSs' available for a company, this provide a way of pricing and mod-
elling the yield curve of bonds of that company. Credit Default Swaps are contracts that
provide insurance against the risk of default by a particular company [5]. The buyer of
the CDS pays money periodically to the seller and the seller pays money if the company
(or government) defaults. It is called credit event and the details are xed in the contract.
The CDS has a face or notional value. In the case of a credit event there are two popular
procedures. One is that if a credit event happens the CDS buyer gives the secured product
to the CDS seller while the seller pays the face value to the buyer. The other case is a
cash settlement payed to the buyer and the amount is usually determined by a calculation
agent or auction mechanism.

36
4.2 Pricing with CDS 4 ADDITIONAL FEATURES

The periodical payments (usually annual or semiannual) the buyer pays to the seller
is computed using the face value and the credit default swap spread (or CDS rate):

PCDS = F V CDSspread

Here PCDS refers to the annual cost of the CDS, F V denotes Face Value and CDSspread
is the credit default swap spread. The CDS spread of CDS rate is the factor that contains
the perceived credit risk of the company by the market. If market participants think a
company is becoming more credit risky the CDS rate rises while the opposite happens if
the company becomes more stable in the eyes of the public.
Credit Default Swaps have a maturity also. That is the time horizon when the contract
holds. Until maturity the buyer will pay the payments to the seller and the seller provide
security. The only case the contract terminates before maturity is a credit event. If that
occurs the buyer doesn't have to continue paying the periodical payments. CDS spreads
can be dierent on dierent time horizons and this way there is a time structure of
CDS spreads too. While it looks obvious that companies with lower credit ratings usually
have higher CDS spreads the relation between CDS maturity and the spread of the same
company is not that easy to see. Research on Eurobonds and domestic bonds from EU-
countries shows that the amount of the spread and maturity have a positive relationship
in the case of investment grade debt. [6]
Bond pricing with the use of CDS rates relies on no arbitrage considerations. If an
investor buys a bond and a CDS with maturity equal to the bond's maturity and face value
equal to the bond's face value this portfolio theoretically becomes a risk-free portfolio.
Then this portfolio should have a return equal to the return on a risk-free product with
the same maturity. Otherwise arbitrage opportunities would arise as the arbitrageur buys
the cheaper product and sells the more expensive contemporarily.

37
4.3 Corporate Bond Spreads 4 ADDITIONAL FEATURES

This no-arbitrage consideration provides an approach of constructing yield curves and


pricing credit risky bonds. The credit spread on bonds should be equal to the CDS spread
of the same bond and maturity. Using this, the yield curve of a credit risky product should
be a benchmark risk-free yield curve plus the term structure of CDS spreads of the prod-
uct. Figure 4 shows an arbitrary example of a risk-free yield curve, a credit spread time
structure and the yield curve aected by this credit spread.

Figure 4: An example of the risk-free yield curve, the credit spread and the yield curve
aected by credit risk as the sum of the rst two.

4.3 Corporate Bond Spreads


There are other ways to estimate the credit risk of a product too without the use
of CDS spreads. Credit Default Swaps are not always available in the market, so credit
spread should be estimated from something else. If a bond with similar maturity and
credit properties is observable in the market, one can derive the credit spread from that. An

38
4.4 Liquidity 4 ADDITIONAL FEATURES

example is a bond of the same company or a very similar company (same size, geopraphical
region, market sector, management stlye, leverage...etc.). Now the credit spread of this
bond can be derived from the market price, risk-free yield curve and properties of that
bond. The credit spread will be the value the risk-free yield curve should be shifted with
in order to reproduce the observable market price. It is a parallel shift of the risk-free
yield curve upwards. If CS denotes the credit spread, rf (t) is the risk-free interest rate,
Pc is the market price of the coupon bond that matures n years from now and pays $C
coupon annually and a principal of $100 at maturity, the following stands:
n
X
Pc = Ce(r(t)+CS)t + 100e(r(n)+CS)n
t=1

If the credit spread is obtained this way, it can be used to construct the yield curve
and price of the product where it is needed. The yield curve can be constructed two ways.
One of them is to simply use a paralell shift similar to the computed credit spread on the
risk-free curve. It is a quick and easy way to have a new yield curve that can be used to
price the product. The other way is to estimate a time structure of credit spreads and then
add this to the risk-free yield curve. The time structure can be obtained by evaluating
bond spreads on dierent maturities. Of course one should start the process from similar
type of bonds in credit properties.

4.4 Liquidity
An other important factor that aects the prices and yields of nancial products is
liquidity. There are dierent denitions and measures of liquidity. Generally a highly
traded product is called liquid. Liquid products are easy to buy and easy to sell in short
period of time without the deterioration of their value. They can be converted to cash
quickly. Illiquid products cannot be sold quickly without some waste in their value. In
stock markets one way to measure liquidity is observing the bid-ask spread of a product.
A relatively big spread means a smaller incoming cashow when one wants to sell the
product quickly so it is considered illiquid while a small spread refers to a liquid product.
Liquidity carries an extra value for investors. Illiquid products usually have higher
yields than liquid ones for compensation. U.S. Treasuries are one of the most liquid -
nancial products available so they are usually used as a benchmark for valuation. Even

39
4.4 Liquidity 4 ADDITIONAL FEATURES

among Treasuries liquidity carries an extra value, this causes the slight dierence between
the yield to maturities of on-the-run and o-the-run treasuries. The most lately issued
treasuries are called on-the-run and the ones that are in the market for more time called
seasoned or o-the-run. New issues are usually traded more than old ones, that causes
the higher liquidity of them. Figure 5 shows the slight dierence between the yields of
U.S. Treasuries due to the liquidity of the product. On-the-run securities have a bit lower
yield to maturity than seasoned ones:

Figure 5: Yield to maturities' of on-the-run and o-the-run treasuries. [10]

When using this as the risk-free rate for calculating credit spreads, the results do not
only represent the risk of default of the issuer, it also incorporates the yield premium
arising from the much less liquidity on the secondary markets of corporate bonds. In a
research, Longsta [7] et al. (2005) shows that the default component explains only 50%
of the spread between the yields of Aaa/Aa rated bonds and treasuries. In the case of Baa

40
4.4 Liquidity 4 ADDITIONAL FEATURES

rated bonds it is 70%. The remaining part is mostly explained by the liquidity factor.
When one wants to estimate yield curves appropriate for evaluating a bond the men-
tioned ways starting from a risk-free yield curve can all lead to relatively good results.
However, Feldhtter suggests that the swap rate is the best as a benchmark risk-free rate
for corporate bond credit spread estimations. [8] The research also concludes that corpo-
rate bonds spreads are a better estimator of credit spread then CDS-spreads since the
latter can be disconnected from the credit component at times. One example according
to them is the Greek CDS and bond spread at 24 June, 2010. While CDS spread has
hit record high, the bond spreads didn't reacted that sensitively probably because of the
support from the European Central Bank bond buying program. Since it is possible to
buy CDSs without owning the bond of the subject company (this is called Naked CDS
contract ) there can be speculative motivations that drive the prices. Even if the bonds'
cash ows are protected some way and thus the values of them the credit event can hap-
pen and CDS owners can make prot from that. Figure 6 shows the Greek CDS and bond
spread around 24 June, 2010.

Figure 6: Historical data of the Greek CDS and bond spread values [8]

41
5 APPLICATIONS

5 Applications
5.1 Data Analysis
The data I have used is corporate bond data from Bloomberg. The plan was to focus
on CHF denominated bonds, altough the models are appropriate for any other currencies
of course. Swiss Bond Index was the starting point of collecting a representative sample of
bonds. The index consisted of 1276 bonds on the 30th September 2014. I used Microsoft
Excel and its Bloomberg extension to obtain the appropriate data.
From the known ISIN values I requested the following informations of the bonds:

NAME: name of the issuer company

PX_LAST: last quoted market price of the bond

COUPON: coupon rate

CPN_FREQ: frequency of the coupon payments

COUNTRY_FULL_NAME: country of the company

INDUSTRY_GROUP: industry group

RTG_MOODY: Moody's rating of the issuer company

The bold notations refer to the Bloomberg variables used in Excel. The data shows the
states of 14th, November 2014. It is observable in the spreadsheet 'ISINs from SBI' in
the attached Excel-le, titled data.xlsm. Since the bond index doesn't contain all bonds
of the companies included I needed some more work on that. First I separated the bonds
with the industry group 'Banks' as a focus of the analysis. From the original 1276 bonds
in the bond index 634 was in the 'Banks' industry group. I further narrowed the results
by centering the attention only on companies rated Aaa by Moody's. This way the set of
634 products decreased to 269. The next step was to collect all other bonds of the issuer
companies of these 269 products. I used the BOND_TO_EQY_TICKER function to
transform the ISIN to the Equity Ticker and after that the BOND_CHAIN function to
list all bonds available of the company the ticker denotes. It is in spreadsheet 'All bonds
of Aaa Banks'. After using a simple VBA-code to position the results I got 49998 bonds

42
5.1 Data Analysis 5 APPLICATIONS

in a column. Of course there were a lot of repetitions which I later corrected. For example
if there was a company with more than one bond in the Bond Index all of the company's
bonds appeared more than one times at this point. When nished removing duplicates
with Excel 7086 dierent bonds remained.
After that I requested the same data mentioned for the bonds: name, last price, ...etc.
In more than half of the 7068 bonds the result was not perfect because some or all of the
information wasn't available from Bloomberg. I cleared the data from the imperfect ones
and it resulted in 3029 dierent bonds with all the important information needed. The
only type where I left missing data is the Moody's rating category since that type is not
used in calculations directly. The spreadsheet 'All bonds of Aaa Banks 2' contains the
results. Figure 7 shows the rst few elements of the resulting dataset.

Figure 7: The rst few elements of the dataset after collecting and clearing it.

I have compared two types of linear yield curve models: polynomial approximation and
spline estimation. To be able to use the Ordinary Least Squares method I needed a cash-
ow matrix of the selected bonds. I have made a VBA code to construct a cashow matrix
from a selection of bonds. The code is in the attached data.xlsm le. The process starts
with collecting the maturities of the bonds into an array. Once it is done a new array
is made that contains all the dates when there is a cash ow from any one bonds of
the starting set. The date when the data is observed is included here even if there is no
payment. This date has to be an input in a cell and the code reads it from there. This

43
5.1 Data Analysis 5 APPLICATIONS

much bigger array is constructed with the use of the maturities and coupon frequency
data. I used a bubble sort algorithm to sort these dates in an incremental order and the
algorithm also unies repeating dates. The cash ow matrix's column will present the
cash ows of the products on these dates. The rows will be associated with the bonds.
I have set the matrix's elements to zero and after that lled with the appropriate cash
ows. In the case of every bond the algorithm started with the principal payment and
then moved backwards by the coupon payment frequencies to ll the selected cells with
the appropriate coupon payment amount. The selection of bonds I analyzed is Aaa rated
bonds of companies from Switzerland. Figure 8 shows an excerpt from the Cash-Flow
matrix result of the algorithm. The code also constructs three more arrays on dierent
spreadsheets. The vector of the last quoted prices, the incremental cash ow dates in a
column and the vector of the maturities. I found it easier to gather this four important
type of arrays from dierent spreadsheets when it came to Python implementation.

Figure 8: A part of the cash ow matrix constructed from Aaa rated bonds from Switzer-
land.

44
5.2 Modeling in Python 5 APPLICATIONS

5.2 Modeling in Python


After gathering and managing the data to obtain the appropriate cashow matrix and
arrays of prices, maturities and cashow dates I continued by constructing the models in
Python. This software was suggested by my Swiss co-supervisor, as they use it in their
daily work. I have tried two dierent methods mentioned in sections (3.5) and (3.6) to
t a yield curve to the data: polynomial and spline approximations. Both are linear yield
curve models and I have used the OLS method when minimizing the error terms.
I used the following packages under Python: numpy to handle the array datatype and
pandas to import the data from the excel le into Python. I also used the math and
matplotlib packages. After importing the data from the excel le I transformed the
arrays of dates to numbers that represent the true annualized time to maturites from
14th November. (Excel stores dates as numbers that represent days from 1st January,
1900.)
For the polynomial model I have made a function:
polynomial (cf, prices, dates, deg, con) , with ve arguments:

cf: The cashow matrix as a numpy array datatype.

prices: The market prices of the bonds as a numpy array datatype.

dates: The cashow dates in an ascending order as a numpy array datatype.

deg: The degree of the polynomial estimation.

con: A parameter to decide whether a constrained (con=1) or an unconstrained


(con=0) model is to be computed.

The return of the function is an array with the estimated discount function values on the
cashow dates. I used this to present the results in a diagram but the code can be easily
modied to return for example the estimated vector of the i parameters or the discount
function's values on dierent time to maturity values. The algorithm is based on the OLS
method mentioned in the 'Statistical Yield Curve Models' chapter of this thesis.
The function I wrote for spline tting is spline (matu, cf, prices, dates, deg,
con) . It has one more argument than the polynomial function and that is an array of
the maturities already imported from the excel le. This vector is used when computing

45
5.2 Modeling in Python 5 APPLICATIONS

the knot points. I have written a subfunction knotpoints (matu, deg) that computes
the vector of the knot points. There is three way the knot points can be calculated and
the user can choose which way he or she prefers. The rst is fully automated, the number
of knot points is the square root of the number of observations rounded down to the
nearest integer. The points are located so that there are approximately equal number of
observations between every point. This approach is suggested by J. Huston McCulloch
[14]. The second way is that the user inputs the number of knot points and then these are
located automatically similarly to the rst case. The third is fully manual. The location
of the knot points are typed in manually one by one. Either way has been chosen the
extension of the knot points is automatically made by the code using parameter deg .
After the knotpoints are generated the spline function estimates the spline yield curve
model with OLS technique and after that returns the arrray of the discount function's
values on cashow dates.
I have made two small functions: yieldcurve (discount, dates) and forwardcurve
(yieldcurve, dates) . These are short codes that generate the spot and instantenous
forward curves from the discount curve values or from the spot curve. The returns are the
corresponding curve's values on the cashow dates.

46
5.3 The results 5 APPLICATIONS

5.3 The results


I have used Aaa rated bonds of Swiss companies in the Bank industry to t the sta-
tistical yield curves. A total of 175 bonds were used. I have compared the results by the
outlook of the curves and the sum of squared errors of the ttings.

Polynomial Model: The discount function was modeled as a third and as a fth
degree polynomial function. The constrained versions (where a constraint ensured that
d(0) = 1) resulted in much better short ends especially for the yield and forward curves.
Of course this raised the error term a bit as constraints usually do. Figure 9 presents the
resulting discount functions. The yield curves computed from the discount functions are
presented by Figure 10.

Figure 9: Polynomial modeling of the discount function.

(a) 3rd degree polynomial. (b) 5th degree polynomial.

The yield curves computed from the discount functions are presented by Figure 10.

47
5.3 The results 5 APPLICATIONS

(a) Yield curve derived from the 3rd degree (b) Yield curve derived from the 5th degree

polynomial. polynomial.

Figure 10: Yield curves computed from the discount functions.

The forward curves calculated from the yield curves are presented by Figure 11.

(a) Forward curve of the 3rd degree dis- (b) Forward curve of the 5th degree dis-

count function. count function.

Figure 11: Forward curves calculated from the spot yield curves.

The sum of squared errors in the third degree case is 187.31. When tting the fth degree
polynomial the error decreased to 155.12. On these gures it is apparent how some extra
curvature in the discount function escalates when computing the yield and the forward

48
5.3 The results 5 APPLICATIONS

curves. Altough higher degree polynomials produce smaller errors, sometimes the results'
curvature may be considered too big to be realistic.

Spline Model: On the same data I examined two cubic spline models. Both of them
with the constraint d(0) = 1 because those produced much more realistic results. The
dierence was the number of the knot points. In the rst case a model with 4 knot points
was approximated while in the second case a model with six knot points. The knot points
were located automatically and the extensions were made automatically also with the use
of the knotpoints function. Figure 12 shows the estimated discount functions.

(a) Cubic spline with 4 knot points. (b) Cubic spline with 6 knot points.

Figure 12: Cubic spline estimations of the discount function.

The yield curves estimated from the discount functions above are presented in Figure
13.

The forward curves are presented in Figure 14.

49
5.3 The results 5 APPLICATIONS

(a) Yield curve computed from the spline (b) Yield curve computed from the spline

with 4 knot points. with 6 knot points.

Figure 13: Yield curves computed from the spline discount function estimations.

(a) Forward curve of the spline with 4 knot (b) Forward curve of the spline with 6 knot

points. points.

Figure 14: Forward curves estimated from the yield curves.

The sum of squared errors of the ttings are the following. In the case of the cubic spline
with 4 knot points the error was 170.56. The sum of squared errors in the case of 6 knot
points was 130.27. By comparing the outlook of the curves and the error values in this
four cases I conclude that the best t was the cubic spline with 4 knot points on this set
of bonds. Altough the one with 6 knot points resulted in smaller errors, the shapes of the
yield curve and forward curve are not realistic.

50
6 SUMMARY

6 Summary
I found that linear yield curve models are quite easy to implement and to work with
in applications. They are not slow even when working with much bigger observations
number and produced sensible results. The algorithm that slowed down the modeling is
the part that computes the cashow matrix from the list of the bonds in Excel. I plan to
improve that, for example by changing the bubble sort algorithm to a quicksort, and by
restructuring the algorithm. The excel le with the VBA function and the Python code
saved as an ipython notebook le can be downloaded from:

https://dl.dropboxusercontent.com/u/100401660/data.xlsm

and

https://dl.dropboxusercontent.com/u/100401660/yieldcurve.ipynb

I have plans to implement some nonlinear models in Python, such as Nelson-Siegel or


Svensson. And also to compare the dierent types of models with dierent statistical
methods.

51
7 APPENDIX

7 Appendix
7.1 Appendix 1: An example of B-splines
The following graphs show the B-splines belong to the set of knot points: {0, 2, 5, 8, 11, 14, 19}.

Figure 15: B-splines of order zero.

52
7.1 Appendix 1: An example of B-splines 7 APPENDIX

The basis splines of order one and two:

Figure 16: B-splines of order one.

Figure 17: B-splines of order two.

53
7.1 Appendix 1: An example of B-splines 7 APPENDIX

The basis spline of order three:

Figure 18: B-splines of order three.

54
REFERENCES REFERENCES

References
[1] Mitzhaletzky Gyrgy: Interest Rate Models (university lecture, 2014)

[2] Frank J. Fabozzi, Steven V. Mann: Handbook of Fixed Income Securities (McGraw-
Hill, 2005)

[3] Jessica James, Nick Webber: Interest Rate Modelling (John Wiley and Sons, 2000)

[4] Mark Fisher, Douglas Nychka, David Zervos: Fitting the term structure of interest
rates with smoothing splines, (FEDS 95-1, 1994)

[5] John C. Hull: Options, futures and other derivatives (6th edition, 2005)

[6] Stefan Trck, Matthias Laub, Svetlozar T. Rachev: The Term Structure of Credit
Spreads and Credit Default Swaps - an empirical investigation (2004)

[7] Francis A. Longsta, Sanjay Mithal, Eric Neis: Corporate Yield Spreads: Default Risk
of Liquidity? (2005)

[8] Peter Feldhtter: Where to look for credit risk? - The corporate bond market or the
CDS market? (Annual Financial Market Liquidity Conference presentation, 2014)

[9] Antonio Daz, Frank Skinner: Estimating Corporate Yield Curves (2001)

[10] Jerry Yi Xiao: Term Structure Estimations for U.S. Corporate Bond Yields

[11] Daniel F. Waggoner: Spline methods for extracting interest rate curves from coupon
bond prices, 1997

[12] Julian D. A. Wiseman: The exponential yield curve model, 1994

[13] Cox, J.C., J.E. Ingersoll, S.A. Ross: A Theory of the Term Structure of Interest Rates.
(Econometrica 53: 385407, 1985)

[14] J. Huston McCulloch: The Tax-adjusted Yield Curve (Journal of Finance, 1975)

55

You might also like