ORF555 / FIN555
:
Fixed Income Models
Damir Filipovi´c
Department of
Operations Research and Financial Engineering
Princeton University
Fall 2002
2
Contents
1 Introduction 7
2 Interest Rates and Related Contracts 9
2.1 ZeroCoupon Bonds . . . . . . . . . . . . . . . . . . . . . . . . 9
2.2 Interest Rates . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.2.1 Market Example: LIBOR . . . . . . . . . . . . . . . . 12
2.2.2 Simple vs. Continuous Compounding . . . . . . . . . . 12
2.2.3 Forward vs. Future Rates . . . . . . . . . . . . . . . . 13
2.3 Bank Account and Short Rates . . . . . . . . . . . . . . . . . 14
2.4 Coupon Bonds, Swaps and Yields . . . . . . . . . . . . . . . . 15
2.4.1 Fixed Coupon Bonds . . . . . . . . . . . . . . . . . . . 16
2.4.2 Floating Rate Notes . . . . . . . . . . . . . . . . . . . 16
2.4.3 Interest Rate Swaps . . . . . . . . . . . . . . . . . . . 17
2.4.4 Yield and Duration . . . . . . . . . . . . . . . . . . . . 20
2.5 Market Conventions . . . . . . . . . . . . . . . . . . . . . . . . 22
2.5.1 Daycount Conventions . . . . . . . . . . . . . . . . . . 22
2.5.2 Coupon Bonds . . . . . . . . . . . . . . . . . . . . . . 23
2.5.3 Accrued Interest, Clean Price and Dirty Price . . . . . 24
2.5.4 YieldtoMaturity . . . . . . . . . . . . . . . . . . . . . 25
2.6 Caps and Floors . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.7 Swaptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3 Statistics of the Yield Curve 33
3.1 Principal Component Analysis (PCA) . . . . . . . . . . . . . . 33
3.2 PCA of the Yield Curve . . . . . . . . . . . . . . . . . . . . . 35
3.3 Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
3
4 CONTENTS
4 Estimating the Yield Curve 39
4.1 A Bootstrapping Example . . . . . . . . . . . . . . . . . . . . 39
4.2 General Case . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.2.1 Bond Markets . . . . . . . . . . . . . . . . . . . . . . . 45
4.2.2 Money Markets . . . . . . . . . . . . . . . . . . . . . . 46
4.2.3 Problems . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.2.4 Parametrized Curve Families . . . . . . . . . . . . . . . 49
5 Why Yield Curve Models? 65
6 NoArbitrage Pricing 67
6.1 SelfFinancing Portfolios . . . . . . . . . . . . . . . . . . . . . 67
6.2 Arbitrage and Martingale Measures . . . . . . . . . . . . . . . 69
6.3 Hedging and Pricing . . . . . . . . . . . . . . . . . . . . . . . 73
7 Short Rate Models 77
7.1 Generalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
7.2 Diﬀusion Short Rate Models . . . . . . . . . . . . . . . . . . . 79
7.2.1 Examples . . . . . . . . . . . . . . . . . . . . . . . . . 82
7.3 Inverting the Yield Curve . . . . . . . . . . . . . . . . . . . . 83
7.4 Aﬃne Term Structures . . . . . . . . . . . . . . . . . . . . . . 83
7.5 Some Standard Models . . . . . . . . . . . . . . . . . . . . . . 85
7.5.1 Vasicek Model . . . . . . . . . . . . . . . . . . . . . . . 85
7.5.2 Cox–Ingersoll–Ross Model . . . . . . . . . . . . . . . . 86
7.5.3 Dothan Model . . . . . . . . . . . . . . . . . . . . . . . 87
7.5.4 Ho–Lee Model . . . . . . . . . . . . . . . . . . . . . . . 88
7.5.5 Hull–White Model . . . . . . . . . . . . . . . . . . . . 89
7.6 Option Pricing in Aﬃne Models . . . . . . . . . . . . . . . . . 90
7.6.1 Example: Vasicek Model (a, b, β const, α = 0). . . . . 92
8 HJM Methodology 95
9 Forward Measures 97
9.1 TBond as Numeraire . . . . . . . . . . . . . . . . . . . . . . . 97
9.2 An Expectation Hypothesis . . . . . . . . . . . . . . . . . . . 99
9.3 Option Pricing in Gaussian HJM Models . . . . . . . . . . . . 101
CONTENTS 5
10 Forwards and Futures 105
10.1 Forward Contracts . . . . . . . . . . . . . . . . . . . . . . . . 105
10.2 Futures Contracts . . . . . . . . . . . . . . . . . . . . . . . . . 106
10.3 Interest Rate Futures . . . . . . . . . . . . . . . . . . . . . . . 108
10.4 Forward vs. Futures in a Gaussian Setup . . . . . . . . . . . . 109
11 MultiFactor Models 113
11.1 NoArbitrage Condition . . . . . . . . . . . . . . . . . . . . . 115
11.2 Aﬃne Term Structures . . . . . . . . . . . . . . . . . . . . . . 117
11.3 Polynomial Term Structures . . . . . . . . . . . . . . . . . . . 118
11.4 ExponentialPolynomial Families . . . . . . . . . . . . . . . . 122
11.4.1 Nelson–Siegel Family . . . . . . . . . . . . . . . . . . . 122
11.4.2 Svensson Family . . . . . . . . . . . . . . . . . . . . . 123
12 Market Models 127
12.1 Models of Forward LIBOR Rates . . . . . . . . . . . . . . . . 129
12.1.1 Discretetenor Case . . . . . . . . . . . . . . . . . . . . 130
12.1.2 Continuoustenor Case . . . . . . . . . . . . . . . . . . 140
13 Default Risk 145
13.1 Transition and Default Probabilities . . . . . . . . . . . . . . . 145
13.1.1 Historical Method . . . . . . . . . . . . . . . . . . . . . 146
13.1.2 Structural Approach . . . . . . . . . . . . . . . . . . . 148
13.2 Intensity Based Method . . . . . . . . . . . . . . . . . . . . . 150
13.2.1 Construction of Intensity Based Models . . . . . . . . . 156
13.2.2 Computation of Default Probabilities . . . . . . . . . . 157
13.2.3 Pricing Default Risk . . . . . . . . . . . . . . . . . . . 157
13.2.4 Measure Change . . . . . . . . . . . . . . . . . . . . . 160
6 CONTENTS
Chapter 1
Introduction
These notes have been written for a graduate course on ﬁxed income models
that I held in the fall term 2002–2003 at Princeton University.
The number of books on ﬁxed income models is growing, yet it is diﬃcult
to ﬁnd a convenient textbook for a onesemester course like this. There are
several reasons for this:
• Until recently, many textbooks on mathematical ﬁnance have treated
stochastic interest rates as an appendix to the elementary arbitrage
pricing theory, which usually requires constant (zero) interest rates.
• Interest rate theory is not standardized yet: there is no wellaccepted
“standard” general model such as the Black–Scholes model for equities.
• The very nature of ﬁxed income instruments causes diﬃculties, other
than for stock derivatives, in implementing and calibrating models.
These issues should therefore not been left out.
I will frequently refer to the following books:
B[3]: Bj¨ ork (98) [3]. A pedagogically well written introduction to mathe
matical ﬁnance. Chapters 15–20 are on interest rates.
BM[6]: Brigo–Mercurio (01) [6]. This is a book on interest rate modelling
written by two quantitative analysts in ﬁnancial institutions. Much
emphasis is on the practical implementation and calibration of selected
models.
7
8 CHAPTER 1. INTRODUCTION
JW[11]: James–Webber (00) [11]. An encyclopedic treatment of interest
rates and their related ﬁnancial derivatives.
J[13]: Jarrow (96) [13]. Introduction to ﬁxedincome securities and interest
rate options. Discrete time only.
MR[19]: Musiela–Rutkowski (97) [19]. A comprehensive book on ﬁnancial
mathematics with a large part (Part II) on interest rate modelling.
Much emphasis is on market pricing practice.
R[22]: Rebonato (98) [22]. Written by a practitionar. Much emphasis on
market practice for pricing and handling interest rate derivatives.
Z[27]: Zagst (02) [27]. A comprehensive textbook on mathematical ﬁnance,
interest rate modelling and risk management.
I did not intend to write an entire text but rather collect fragments of the
material that can be found in the above books and further references.
Chapter 2
Interest Rates and Related
Contracts
Literature: B[3](Chapter 15), BM[6](Chapter 1), and many more
2.1 ZeroCoupon Bonds
A dollar today is worth more than a dollar tomorrow. The time t value of
a dollar at time T ≥ t is expressed by the zerocoupon bond with maturity
T, P(t, T), for briefty also Tbond. This is a contract which guarantees the
holder one dollar to be paid at the maturity date T.
1 P(t,T)
t
 
T
→ future cashﬂows can be discounted, such as couponbearing bonds
C
1
P(t, t
1
) + +C
n−1
P(t, t
n−1
) + (1 +C
n
)P(t, T).
In theory we will assume that
• there exists a frictionless market for Tbonds for every T > 0.
• P(T, T) = 1 for all T.
• P(t, T) is continuously diﬀerentiable in T.
9
10 CHAPTER 2. INTEREST RATES AND RELATED CONTRACTS
In reality this assumptions are not always satisﬁed: zerocoupon bonds are
not traded for all maturities, and P(T, T) might be less than one if the issuer
of the Tbond defaults. Yet, this is a good starting point for doing the
mathematics. More realistic models will be introduced and discussed in the
sequel.
The third condition is purely technical and implies that the term structure
of zerocoupon bond prices T → P(t, T) is a smooth curve.
1 2 3 4 5 6 7 8 9 10
Years
0.2
0.4
0.6
0.8
1
US Treasury Bonds, March 2002
Note that t → P(t, T) is a stochastic process since bond prices P(t, T) are
not known with certainty before t.
1 2 3 4 5 6 7 8 9 10
t
0.2
0.4
0.6
0.8
1
PHt,10L
A reasonable assumption would also be that T → P(t, T) ≤ 1 is a de
creasing curve (which is equivalent to positivity of interest rates). However,
already classical interest rate models imply zerocoupon bond prices greater
than 1. Therefore we leave away this requirement.
2.2. INTEREST RATES 11
2.2 Interest Rates
The term structure of zerocoupon bond prices does not contain much visual
information (strictly speaking it does). A better measure is given by the
implied interest rates. There is a variety of them.
A prototypical forward rate agreement (FRA) is a contract involving three
time instants t < T < S: the current time t, the expiry time T > t, and the
maturity time S > T.
• At t: sell one Tbond and buy
P(t,T)
P(t,S)
Sbonds = zero net investment.
• At T: pay one dollar.
• At S: obtain
P(t,T)
P(t,S)
dollars.
The net eﬀect is a forward investment of one dollar at time T yielding
P(t,T)
P(t,S)
dollars at S with certainty.
We are led to the following deﬁnitions.
• The simple (simplycompounded) forward rate for [T, S] prevailing at t
is given by
1+(S−T)F(t; T, S) :=
P(t, T)
P(t, S)
⇔F(t; T, S) =
1
S −T
P(t, T)
P(t, S)
−1
.
• The simple spot rate for [t, T] is
F(t, T) := F(t; t, T) =
1
T −t
1
P(t, T)
−1
.
• The continuously compounded forward rate for [T, S] prevailing at t is
given by
e
R(t;T,S)(S−T)
:=
P(t, T)
P(t, S)
⇔R(t; T, S) = −
log P(t, S) −log P(t, T)
S −T
.
• The continuously compounded spot rate for [T, S] is
R(t, T) := R(t; t, T) = −
log P(t, T)
T −t
.
12 CHAPTER 2. INTEREST RATES AND RELATED CONTRACTS
• The instantaneous forward rate with maturity T prevailing at time t is
deﬁned by
f(t, T) := lim
S↓T
R(t; T, S) = −
∂ log P(t, T)
∂T
. (2.1)
The function T →f(t, T) is called the forward curve at time t.
• The instantaneous short rate at time t is deﬁned by
r(t) := f(t, t) = lim
T↓t
R(t, T).
Notice that (2.1) together with the requirement P(T, T) = 1 is equivalent to
P(t, T) = exp
−
T
t
f(t, u) du
.
2.2.1 Market Example: LIBOR
“Interbank rates” are rates at which deposits between banks are exchanged,
and at which swap transactions (see below) between banks occur. The most
important interbank rate usually considered as a reference for ﬁxed income
contracts is the LIBOR (London InterBank Oﬀered Rate)
1
for a series of
possible maturities, ranging from overnight to 12 months. These rates are
quoted on a simple compounding basis. For example, the threemonths for
ward LIBOR for the period [T, T + 1/4] at time t is given by
L(t, T) = F(t; T, T + 1/4).
2.2.2 Simple vs. Continuous Compounding
One dollar invested for one year at an interest rate of R per annum growths
to 1 + R. If the rate is compounded twice per year the terminal value is
(1 +R/2)
2
, etc. It is a mathematical fact that
1 +
R
m
m
→e
R
as m → ∞.
1
To be more precise: this is the rate at which highcredit ﬁnancial institutions can
borrow in the interbank market.
2.2. INTEREST RATES 13
Moreover,
e
R
= 1 +R +o(R) for R small.
Example: e
0.04
= 1.04081.
Since the exponential function has nicer analytic properties than power
functions, we often consider continuously compounded interest rates. This
makes the theory more tractable.
2.2.3 Forward vs. Future Rates
Can forward rates predict the future spot rates?
Consider a deterministic world. If markets are eﬃcient (i.e. no arbitrage
= no riskless, systematic proﬁt) we have necessarily
P(t, S) = P(t, T)P(T, S), ∀t ≤ T ≤ S. (2.2)
Proof. Suppose that P(t, S) > P(t, T)P(T, S) for some t ≤ T ≤ S. Then we
follow the strategy:
• At t: sell one Sbond, and buy P(T, S) Tbonds.
Net cost: −P(t, S) +P(t, T)P(T, S) < 0.
• At T: receive P(T, S) dollars and buy one Sbond.
• At S: pay one dollar, receive one dollar.
(Where do we use the assumption of a deterministic world?)
The net is a riskless gain of −P(t, S)+P(t, T)P(T, S) (1/P(t, S)). This
is a pure arbitrage opportunity, which contradicts the assumption.
If P(t, S) < P(t, T)P(T, S) the same proﬁt can be realized by changing
sign in the strategy.
Taking logarithm in (2.2) yields
S
T
f(t, u) du =
S
T
f(T, u) du, ∀t ≤ T ≤ S.
This is equivalent to
f(t, S) = f(T, S) = r(S), ∀t ≤ T ≤ S
14 CHAPTER 2. INTEREST RATES AND RELATED CONTRACTS
(as time goes by we walk along the forward curve: the forward curve is
shifted). In this case, the forward rate with maturity S prevailing at time
t ≤ S is exactly the future short rate at S.
The real world is not deterministic though. We will see that in general
the forward rate f(t, T) is the conditional expectation of the short rate r(T)
under a particular probability measure (forward measure), depending on T.
Hence the forward rate is a biased estimator for the future short rate.
Forecasts of future short rates by forward rates have little or no predictive
power.
2.3 Bank Account and Short Rates
The return of a one dollar investment today (t = 0) over the period [0, ∆t]
is given by
1
P(0, ∆t)
= exp
∆t
0
f(0, u) du
= 1 +r(0)∆t +o(∆t).
Instantaneous reinvestment in 2∆tbonds yields
1
P(0, ∆t)
1
P(∆t, 2∆t)
= (1 +r(0)∆t)(1 +r(∆t)∆t) +o(∆t)
at time 2∆t, etc. This strategy of “rolling over”
2
just maturing bonds leads
in the limit to the bank account (moneymarket account) B(t). Hence B(t)
is the asset which growths at time t instantaneously at short rate r(t)
B(t + ∆t) = B(t)(1 +r(t)∆t) +o(∆t).
For ∆t →0 this converges to
dB(t) = r(t)B(t)dt
and with B(0) = 1 we obtain
B(t) = exp
t
0
r(s) ds
.
2
This limiting process is made rigorous in [4].
2.4. COUPON BONDS, SWAPS AND YIELDS 15
B is a riskfree asset insofar as its future value at time t + ∆t is known (up
to order ∆t) at time t. In stochastic terms we speak of a predictable process.
For the same reason we speak of r(t) as the riskfree rate of return over the
inﬁnitesimal period [t, t +dt].
B is important for relating amounts of currencies available at diﬀerent
times: in order to have one dollar in the bank account at time T we need to
have
B(t)
B(T)
= exp
−
T
t
r(s) ds
dollars in the bank account at time t ≤ T. This discount factor is stochastic:
it is not known with certainty at time t. There is a close connection to the
deterministic (=known at time t) discount factor given by P(t, T). Indeed,
we will see that the latter is the conditional expectation of the former under
the risk neutral probability measure.
Proxies for the Short Rate
→ JW[11](Chapter 3.5)
The short rate r(t) is a key interest rate in all models and fundamental
to noarbitrage pricing. But it cannot be directly observed.
The overnight interest rate is not usually considered to be a good proxy
for the short rate, because the motives and needs driving overnight borrowers
are very diﬀerent from those of borrowers who want money for a month or
more.
The overnight fed funds rate is nevertheless comparatively stable and
perhaps a fair proxy, but empirical studies suggest that it has low correlation
with other spot rates.
The best available proxy is given by one or threemonth spot rates since
they are very liquid.
2.4 Coupon Bonds, Swaps and Yields
In most bond markets, there is only a relatively small number of zerocoupon
bonds traded. Most bonds include coupons.
16 CHAPTER 2. INTEREST RATES AND RELATED CONTRACTS
2.4.1 Fixed Coupon Bonds
A ﬁxed coupon bond is a contract speciﬁed by
• a number of future dates T
1
< < T
n
(the coupon dates)
(T
n
is the maturity of the bond),
• a sequence of (deterministic) coupons c
1
, . . . , c
n
,
• a nominal value N,
such that the owner receives c
i
at time T
i
, for i = 1, . . . , n, and N at terminal
time T
n
. The price p(t) at time t ≤ T
1
of this coupon bond is given by the
sum of discounted cashﬂows
p(t) =
n
¸
i=1
P(t, T
i
)c
i
+P(t, T
n
)N.
Typically, it holds that T
i+1
−T
i
≡ δ, and the coupons are given as a ﬁxed
percentage of the nominal value: c
i
≡ KδN, for some ﬁxed interest rate K.
The above formula reduces to
p(t) =
Kδ
n
¸
i=1
P(t, T
i
) +P(t, T
n
)
N.
2.4.2 Floating Rate Notes
There are versions of coupon bonds for which the value of the coupon is
not ﬁxed at the time the bond is issued, but rather reset for every coupon
period. Most often the resetting is determined by some market interest rate
(e.g. LIBOR).
A ﬂoating rate note is speciﬁed by
• a number of future dates T
0
< T
1
< < T
n
,
• a nominal value N.
The deterministic coupon payments for the ﬁxed coupon bond are now re
placed by
c
i
= (T
i
−T
i−1
)F(T
i−1
, T
i
)N,
2.4. COUPON BONDS, SWAPS AND YIELDS 17
where F(T
i−1
, T
i
) is the prevailing simple market interest rate, and we note
that F(T
i−1
, T
i
) is determined already at time T
i−1
(this is why here we have
T
0
in addition to the coupon dates T
1
, . . . , T
n
), but that the cashﬂow c
i
is
at time T
i
.
The value p(t) of this note at time t ≤ T
0
is obtained as follows. Without
loss of generality we set N = 1. By deﬁnition of F(T
i−1
, T
i
) we then have
c
i
=
1
P(T
i−1
, T
i
)
−1.
The time t value of −1 paid out at T
i
is −P(t, T
i
). The time t value of
1
P(T
i−1
,T
i
)
paid out at T
i
is P(t, T
i−1
):
• At t: buy a T
i−1
bond. Cost: P(t, T
i−1
).
• At T
i−1
: receive one dollar and buy 1/P(T
i−1
, T
i
) T
i
bonds. Zero net
investment.
• At T
i
: receive 1/P(T
i−1
, T
i
) dollars.
The time t value of c
i
therefore is
P(t, T
i−1
) −P(t, T
i
).
Summing up we obtain the (surprisingly easy) formula
p(t) = P(t, T
n
) +
n
¸
i=1
(P(t, T
i−1
) −P(t, T
i
)) = P(t, T
0
).
In particular, for t = T
0
: p(T
0
) = 1.
2.4.3 Interest Rate Swaps
An interest rate swap is a scheme where you exchange a payment stream
at a ﬁxed rate of interest for a payment stream at a ﬂoating rate (typically
LIBOR).
There are many versions of interest rate swaps. A payer interest rate
swap settled in arrears is speciﬁed by
• a number of future dates T
0
< T
1
< < T
n
with T
i
−T
i−1
≡ δ
(T
n
is the maturity of the swap),
18 CHAPTER 2. INTEREST RATES AND RELATED CONTRACTS
• a ﬁxed rate K,
• a nominal value N.
Of course, the equidistance hypothesis is only for convenience of notation
and can easily be relaxed. Cashﬂows take place only at the coupon dates
T
1
, . . . , T
n
. At T
i
, the holder of the contract
• pays ﬁxed KδN,
• and receives ﬂoating F(T
i−1
, T
i
)δN.
The net cashﬂow at T
i
is thus
(F(T
i−1
, T
i
) −K)δN,
and using the previous results we can compute the value at t ≤ T
0
of this
cashﬂow as
N(P(t, T
i−1
) −P(t, T
i
) −KδP(t, T
i
)). (2.3)
The total value Π
p
(t) of the swap at time t ≤ T
0
is thus
Π
p
(t) = N
P(t, T
0
) −P(t, T
n
) −Kδ
n
¸
i=1
P(t, T
i
)
.
A receiver interest rate swap settled in arrears is obtained by changing
the sign of the cashﬂows at times T
1
, . . . , T
n
. Its value at time t ≤ T
0
is thus
Π
r
(t) = −Π
p
(t).
The remaining question is how the “fair” ﬁxed rate K is determined. The
forward swap rate R
swap
(t) at time t ≤ T
0
is the ﬁxed rate K above which
gives Π
p
(t) = Π
r
(t) = 0. Hence
R
swap
(t) =
P(t, T
0
) −P(t, T
n
)
δ
¸
n
i=1
P(t, T
i
)
.
The following alternative representation of R
swap
(t) is sometimes useful.
Since P(t, T
i−1
) −P(t, T
i
) = F(t; T
i−1
, T
i
)δP(t, T
i
), we can rewrite (2.3) as
NδP(t, T
i
) (F(t; T
i−1
, T
i
) −K) .
2.4. COUPON BONDS, SWAPS AND YIELDS 19
Summing up yields
Π
p
(t) = Nδ
n
¸
i=1
P(t, T
i
) (F(t; T
i−1
, T
i
) −K) ,
and thus we can write the swap rate as weighted average of simple forward
rates
R
swap
(t) =
n
¸
i=1
w
i
(t)F(t; T
i−1
, T
i
),
with weights
w
i
(t) =
P(t, T
i
)
¸
n
j=1
P(t, T
j
)
.
These weights are random, but there seems to be empirical evidence that
the variability of w
i
(t) is small compared to that of F(t; T
i−1
, T
i
). This is
used for approximations of swaption (see below) price formulas in LIBOR
market models: the swap rate volatility is written as linear combination of
the forward LIBOR volatilities (“Rebonato’s formula” → BM[6], p.248).
Swaps were developed because diﬀerent companies could borrow at dif
ferent rates in diﬀerent markets.
Example
→ JW[11](p.11)
• Company A: is borrowing ﬁxed for ﬁve years at 5 1/2%, but could
borrow ﬂoating at LIBOR plus 1/2%.
• Company B: is borrowing ﬂoating at LIBOR plus 1%, but could borrow
ﬁxed for ﬁve years at 6 1/2%.
By agreeing to swap streams of cashﬂows both companies could be better
oﬀ, and a mediating institution would also make money.
• Company A pays LIBOR to the intermediary in exchange for ﬁxed at
5 3/16% (receiver swap).
• Company B pays the intermediary ﬁxed at 5 5/16% in exchange for
LIBOR (payer swap).
20 CHAPTER 2. INTEREST RATES AND RELATED CONTRACTS
Net:
• Company A is now paying LIBOR plus 5/16% instead of LIBOR plus
1/2%.
• Company B is paying ﬁxed at 6 5/16% instead of 6 1/2%.
• The intermediary receives ﬁxed at 1/8%.
5 5/16 % 5 3/16 %
LIBOR LIBOR LIBOR + 1%
5 1/2 %
Company A Intermediary Company B
Everyone seems to be better oﬀ. But there is implicit credit risk; this is
why Company B had higher borrowing rates in the ﬁrst place. This risk has
been partly taken up by the intermediary, in return for the money it makes
on the spread.
2.4.4 Yield and Duration
For a zerocoupon bond P(t, T) the zerocoupon yield is simply the continu
ously compounded spot rate R(t, T). That is,
P(t, T) = e
−R(t,T)(T−t)
.
Accordingly, the function T → R(t, T) is referred to as (zerocoupon) yield
curve.
The term “yield curve” is ambiguous. There is a variety of other ter
minologies, such as zerorate curve (Z[27]), zerocoupon curve (BM[6]). In
JW[11] the yield curve is is given by simple spot rates, and in BM[6] it is a
combination of simple spot rates (for maturities up to 1 year) and annually
compounded spot rates (for maturities greater than 1 year), etc.
2.4. COUPON BONDS, SWAPS AND YIELDS 21
1 2 3 4 5 6 7 8 9 10
Years
0.02
0.04
0.06
0.08
0.1
US Yield Curve, March 2002
Now let p(t) be the time t market value of a ﬁxed coupon bond with
coupon dates T
1
< < T
n
, coupon payments c
1
, . . . , c
n
and nominal value
N (see Section 2.4.1). For simplicity we suppose that c
n
already contains N,
that is,
p(t) =
n
¸
i=1
P(t, T
i
)c
i
, t ≤ T
1
.
Again we ask for the bond’s “internal rate of interest”; that is, the constant
(over the period [t, T
n
]) continuously compounded rate which generates the
market value of the coupon bond: the (continuously compounded) yieldto
maturity y(t) of this bond at time t ≤ T
1
is deﬁned as the unique solution
to
p(t) =
n
¸
i=1
c
i
e
−y(t)(T
i
−t)
.
Remark 2.4.1. → R[22](p.21). It is argued by Schaefer (1977) that the
yieldtomaturity is an inadequate statistics for the bond market:
• coupon payments occurring at the same point in time are discounted by
diﬀerent discount factors, but
• coupon payments at diﬀerent points in time from the same bond are
discounted by the same rate.
To simplify the notation we assume now that t = 0, and write p = p(0),
y = y(0), etc. The Macaulay duration of the coupon bond is deﬁned as
D
Mac
:=
¸
n
i=1
T
i
c
i
e
−yT
i
p
.
22 CHAPTER 2. INTEREST RATES AND RELATED CONTRACTS
The duration is thus a weighted average of the coupon dates T
1
, . . . , T
n
, and
it provides us in a certain sense with the “mean time to coupon payment”.
As such it is an important concept for interest rate risk management: it acts
as a measure of the ﬁrst order sensitivity of the bond price w.r.t. changes in
the yieldtomaturity (see Z[27](Chapter 6.1.3) for a thourough treatment).
This is shown by the obvious formula
dp
dy
=
d
dy
n
¸
i=1
c
i
e
−yT
i
= −D
Mac
p.
A ﬁrst order sensitivity measure of the bond price w.r.t. parallel shifts of
the entire zerocoupon yield curve T → R(0, T) is given by the duration of
the bond
D :=
¸
n
i=1
T
i
c
i
e
−y
i
T
i
p
=
n
¸
i=1
c
i
P(0, T
i
)
p
T
i
,
with y
i
:= R(0, T
i
). In fact, we have
d
ds
n
¸
i=1
c
i
e
−(y
i
+s)T
i
[
s=0
= −Dp.
Hence duration is essentially for bonds (w.r.t. parallel shift of the yield curve)
what delta is for stock options. The bond equivalent of the gamma is con
vexity:
C :=
d
2
ds
2
n
¸
i=1
c
i
e
−(y
i
+s)T
i
[
s=0
=
n
¸
i=1
c
i
e
−y
i
T
i
(T
i
)
2
.
2.5 Market Conventions
2.5.1 Daycount Conventions
Time is measured in years.
If t and T denote two dates expressed as day/month/year, it is not clear
what T −t should be. The market evaluates the year fraction between t and
T in diﬀerent ways.
The daycount convention decides upon the time measurement between
two dates t and T.
Here are three examples of daycount conventions:
2.5. MARKET CONVENTIONS 23
• Actual/365: a year has 365 days, and the daycount convention for
T −t is given by
actual number of days between t and T
365
.
• Actual/360: as above but the year counts 360 days.
• 30/360: months count 30 and years 360 days. Let t = (d
1
, m
1
, y
1
) and
T = (d
2
, m
2
, y
2
). The daycount convention for T −t is given by
min(d
2
, 30) + (30 −d
1
)
+
360
+
(m
2
−m
1
−1)
+
12
+y
2
−y
1
.
Example: The time between t=January 4, 2000 and T=July 4, 2002 is
given by
4 + (30 −4)
360
+
7 −1 −1
12
+ 2002 −2000 = 2.5.
When extracting information on interest rates from data, it is important
to realize for which daycount convention a speciﬁc interest rate is quoted.
→ BM[6](p.4), Z[27](Sect. 5.1)
2.5.2 Coupon Bonds
→ MR[19](Sect. 11.2), Z[27](Sect. 5.2), J[13](Chapter 2)
Coupon bonds issued in the American (European) markets typically have
semiannual (annual) coupon payments.
Debt securities issued by the U.S. Treasury are divided into three classes:
• Bills: zerocoupon bonds with time to maturity less than one year.
• Notes: coupon bonds (semiannual) with time to maturity between 2
and 10 years.
• Bonds: coupon bonds (semiannual) with time to maturity between 10
and 30 years
3
.
3
Recently, the issuance of 30 year treasury bonds has been stopped.
24 CHAPTER 2. INTEREST RATES AND RELATED CONTRACTS
In addition to bills, notes and bonds, Treasury securities called STRIPS
(separate trading of registered interest and principal of securities) have traded
since August 1985. These are the coupons or principal (=nominal) amounts
of Treasury bonds trading separately through the Federal Reserve’s book
entry system. They are synthetically created zerocoupon bonds of longer
maturities than a year. They were created in response to investor demands.
2.5.3 Accrued Interest, Clean Price and Dirty Price
Remember that we had for the price of a coupon bond with coupon dates
T
1
, . . . , T
n
and payments c
1
, . . . , c
n
the price formula
p(t) =
n
¸
i=1
c
i
P(t, T
i
), t ≤ T
1
.
For t ∈ (T
1
, T
2
] we have
p(t) =
n
¸
i=2
c
i
P(t, T
i
),
etc. Hence there are systematic discontinuities of the price trajectory at
t = T
1
, . . . , T
n
which is due to the coupon payments. This is why prices are
diﬀerently quoted at the exchange.
The accrued interest at time t ∈ (T
i−1
, T
i
] is deﬁned by
AI(i; t) := c
i
t −T
i−1
T
i
−T
i−1
(where now time diﬀerences are taken according to the daycount conven
tion). The quoted price, or clean price, of the coupon bond at time t is
p
clean
(t) := p(t) −AI(i; t), t ∈ (T
i−1
, T
i
].
That is, whenever we buy a coupon bond quoted at a clean price of p
clean
(t)
at time t ∈ (T
i−1
, T
i
], the cash price, or dirty price, we have to pay is
p(t) = p
clean
(t) +AI(i; t).
2.6. CAPS AND FLOORS 25
2.5.4 YieldtoMaturity
The quoted (annual) yieldtomaturity ˆ y(t) on a Treasury bond at time t = T
i
is deﬁned by the relationship
p
clean
(T
i
) =
n
¸
j=i+1
r
c
N/2
(1 + ˆ y(T
i
)/2)
j−i
+
N
(1 + ˆ y(T
i
)/2)
n−i
,
and at t ∈ [T
i
, T
i+1
)
p
clean
(t) =
n
¸
j=i+1
r
c
N/2
(1 + ˆ y(t)/2)
j−i−1+τ
+
N
(1 + ˆ y(t)/2)
n−i−1+τ
,
where r
c
is the (annualized) coupon rate, N the nominal amount and
τ =
T
i+1
−t
T
i+1
−T
i
is again given by the daycount convention, and we assume here that
T
i+1
−T
i
≡ 1/2 (semiannual coupons).
2.6 Caps and Floors
→ BM[6](Sect. 1.6), Z[27](Sect. 5.6.2)
Caps
A caplet with reset date T and settlement date T + δ pays the holder the
diﬀerence between a simple market rate F(T, T + δ) (e.g. LIBOR) and the
strike rate κ. Its cashﬂow at time T +δ is
δ(F(T, T +δ) −κ)
+
.
A cap is a strip of caplets. It thus consists of
• a number of future dates T
0
< T
1
< < T
n
with T
i
−T
i−1
≡ δ
(T
n
is the maturity of the cap),
• a cap rate κ.
26 CHAPTER 2. INTEREST RATES AND RELATED CONTRACTS
Cashﬂows take place at the dates T
1
, . . . , T
n
. At T
i
the holder of the cap
receives
δ(F(T
i−1
, T
i
) −κ)
+
. (2.4)
Let t ≤ T
0
. We write
Cpl(i; t), i = 1, . . . , n,
for the time t price of the ith caplet with reset date T
i−1
and settlement date
T
i
, and
Cp(t) =
n
¸
i=1
Cpl(i; t)
for the time t price of the cap.
A cap gives the holder a protection against rising interest rates. It guar
antees that the interest to be paid on a ﬂoating rate loan never exceeds the
predetermined cap rate κ.
It can be shown (→ exercise) that the cashﬂow (2.4) at time T
i
is the
equivalent to (1 + δκ) times the cashﬂow at date T
i−1
of a put option on a
T
i
bond with strike price 1/(1 +δκ) and maturity T
i−1
, that is,
(1 +δκ)
1
1 +δκ
−P(T
i−1
, T
i
)
+
.
This is an important fact because many interest rate models have explicit
formulae for bond option values, which means that caps can be priced very
easily in those models.
Floors
A ﬂoor is the converse to a cap. It protects against low rates. A ﬂoor is a
strip of ﬂoorlets, the cashﬂow of which is – with the same notation as above
– at time T
i
δ(κ −F(T
i−1
, T
i
))
+
.
Write Fll(i; t) for the price of the ith ﬂoorlet and
Fl(t) =
n
¸
i=1
Fll(i; t)
for the price of the ﬂoor.
2.6. CAPS AND FLOORS 27
Caps, Floors and Swaps
Caps and ﬂoors are strongly related to swaps. Indeed, one can show the
parity relation (→ exercise)
Cp(t) −Fl(t) = Π
p
(t),
where Π
p
(t) is the value at t of a payer swap with rate κ, nominal one and
the same tenor structure as the cap and ﬂoor.
Let t = 0. The cap/ﬂoor is said to be atthemoney (ATM) if
κ = R
swap
(0) =
P(0, T
0
) −P(0, T
n
)
δ
¸
n
i=1
P(0, T
i
)
,
the forward swap rate. The cap (ﬂoor) is inthemoney (ITM) if κ < R
swap
(0)
(κ > R
swap
(0)), and outofthemoney (OTM) if κ > R
swap
(0) (κ < R
swap
(0)).
Black’s Formula
It is market practice to price a cap/ﬂoor according to Black’s formula. Let
t ≤ T
0
. Black’s formula for the value of the ith caplet is
Cpl(i; t) = δP(t, T
i
) (F(t; T
i−1
, T
i
)Φ(d
1
(i; t)) −κΦ(d
2
(i; t))) ,
where
d
1,2
(i; t) :=
log
F(t;T
i−1
,T
i
)
κ
±
1
2
σ(t)
2
(T
i−1
−t)
σ(t)
√
T
i−1
−t
(Φ stands for the standard Gaussian cumulative distribution function), and
σ(t) is the cap volatility (it is the same for all caplets).
Correspondingly, Black’s formula for the value of the ith ﬂoorlet is
Fll(i; t) = δP(t, T
i
) (κΦ(−d
2
(i; t)) −F(t; T
i−1
, T
i
)Φ(−d
1
(i; t))) .
Cap/ﬂoor prices are quoted in the market in term of their implied volatil
ities. Typically, we have t = 0, and T
0
and δ = T
i
−T
i−1
being equal to three
months.
An example of a US dollar ATM market cap volatility curve is shown in
Table 2.1 and Figure 2.1 (→ JW[11](p.49)).
It is a challenge for any market realistic interest rate model to match the
given volatility curve.
28 CHAPTER 2. INTEREST RATES AND RELATED CONTRACTS
Table 2.1: US dollar ATM cap volatilities, 23 July 1999
Maturity ATM vols
(in years) (in %)
1 14.1
2 17.4
3 18.5
4 18.8
5 18.9
6 18.7
7 18.4
8 18.2
10 17.7
12 17.0
15 16.5
20 14.7
30 12.4
Figure 2.1: US dollar ATM cap volatilities, 23 July 1999
5 10 15 20 25 30
12%
14%
16%
18%
2.7. SWAPTIONS 29
2.7 Swaptions
A European payer (receiver) swaption with strike rate K is an option giving
the right to enter a payer (receiver) swap with ﬁxed rate K at a given future
date, the swaption maturity. Usually, the swaption maturity coincides with
the ﬁrst reset date of the underlying swap. The underlying swap lenght
T
n
−T
0
is called the tenor of the swaption.
Recall that the value of a payer swap with ﬁxed rate K at its ﬁrst reset
date, T
0
, is
Π
p
(T
0
, K) = N
n
¸
i=1
P(T
0
, T
i
)δ(F(T
0
; T
i−1
, T
i
) −K).
Hence the payoﬀ of the swaption with strike rate K at maturity T
0
is
N
n
¸
i=1
P(T
0
, T
i
)δ(F(T
0
; T
i−1
, T
i
) −K)
+
. (2.5)
Notice that, contrary to the cap case, this payoﬀ cannot be decomposed
into more elementary payoﬀs. This is a fundamental diﬀerence between
caps/ﬂoors and swaptions. Here the correlation between diﬀerent forward
rates will enter the valuation procedure.
Since Π
p
(T
0
, R
swap
(T
0
)) = 0, one can show (→ exercise) that the payoﬀ
(2.5) of the payer swaption at time T
0
can also be written as
Nδ(R
swap
(T
0
) −K)
+
n
¸
i=1
P(T
0
, T
i
),
and for the receiver swaption
Nδ(K −R
swap
(T
0
))
+
n
¸
i=1
P(T
0
, T
i
).
Accordingly, at time t ≤ T
0
, the payer (receiver) swaption with strike rate
K is said to be ATM, ITM, OTM, if
K = R
swap
(t), K < (>)R
swap
(t), K > (<)R
swap
(t),
respectively.
30 CHAPTER 2. INTEREST RATES AND RELATED CONTRACTS
Black’s Formula
Black’s formula for the price at time t ≤ T
0
of the payer (Swpt
p
(t)) and
receiver (Swpt
r
(t)) swaption is
Swpt
p
(t) = Nδ (R
swap
(t)Φ(d
1
(t)) −KΦ(d
2
(t)))
n
¸
i=1
P(t, T
i
),
Swpt
r
(t) = Nδ (KΦ(−d
2
(t)) −R
swap
(t)Φ(−d
1
(t)))
n
¸
i=1
P(t, T
i
),
with
d
1,2
(t) :=
log
Rswap(t)
K
±
1
2
σ(t)
2
(T
0
−t)
σ(t)
√
T
0
−t
,
and σ(t) is the prevailing Black’s swaption volatility.
Swaption prices are quoted in terms of implied volatilities in matrix form.
An x yswaption is the swaption with maturity in x years and whose un
derlying swap is y years long.
A typical example of implied swaption volatilities is shown in Table 2.2
and Figure 2.2 (→ BM[6](p.253)).
An interest model for swaptions valuation must ﬁt the given today’s
volatility surface.
2.7. SWAPTIONS 31
Table 2.2: Black’s implied volatilities (in %) of ATM swaptions on May 16,
2000. Maturities are 1,2,3,4,5,7,10 years, swaps lengths from 1 to 10 years.
1y 2y 3y 4y 5y 6y 7y 8y 9y 10y
1y 16.4 15.8 14.6 13.8 13.3 12.9 12.6 12.3 12.0 11.7
2y 17.7 15.6 14.1 13.1 12.7 12.4 12.2 11.9 11.7 11.4
3y 17.6 15.5 13.9 12.7 12.3 12.1 11.9 11.7 11.5 11.3
4y 16.9 14.6 12.9 11.9 11.6 11.4 11.3 11.1 11.0 10.8
5y 15.8 13.9 12.4 11.5 11.1 10.9 10.8 10.7 10.5 10.4
7y 14.5 12.9 11.6 10.8 10.4 10.3 10.1 9.9 9.8 9.6
10y 13.5 11.5 10.4 9.8 9.4 9.3 9.1 8.8 8.6 8.4
Figure 2.2: Black’s implied volatilities (in %) of ATM swaptions on May 16,
2000.
2
4
6
8
10
Maturity
2
4
6
8
10
Tenor
10
12
14
16
Vol
10
12
14
16
Vol
32 CHAPTER 2. INTEREST RATES AND RELATED CONTRACTS
Chapter 3
Some Statistics of the Yield
Curve
3.1 Principal Component Analysis (PCA)
→ JW[11](Chapter 16.2), [21]
• Let x(1), . . . , x(N) be a sample of a random n 1 vector x.
• Form the empirical n n covariance matrix
ˆ
Σ,
ˆ
Σ
ij
=
¸
N
k=1
(x
i
(k) −µ[x
i
])(x
j
(k) −µ[x
j
])
N −1
=
¸
N
k=1
x
i
(k)x
i
(k) −Nµ[x
i
]µ[x
j
]
N −1
,
where
µ[x
i
] :=
1
N
N
¸
k=1
x
i
(k) (mean of x
i
).
We assume that
ˆ
Σ is nondegenerate (otherwise we can express an x
i
as linear combination of the other x
j
s).
• There exists a unique orthogonal matrix A = (p
1
, . . . , p
n
) (that is,
A
−1
= A
T
and A
ij
= p
j;i
) consisting of orthonormal n1 Eigenvectors
p
i
of
ˆ
Σ such that
ˆ
Σ = ALA
T
,
33
34 CHAPTER 3. STATISTICS OF THE YIELD CURVE
where L = diag(λ
1
, . . . , λ
n
) with λ
1
≥ ≥ λ
n
> 0 (the Eigenvalues
of
ˆ
Σ).
• Deﬁne z := A
T
x. Then
Cov[z
i
, z
j
] =
n
¸
k,l=1
A
T
ik
Cov[x
k
, x
l
]A
T
jl
=
A
T
ˆ
ΣA
ij
= λ
i
δ
ij
.
Hence the z
i
s are uncorrelated.
• The principal components (PCs) are the n 1 vectors p
1
, . . . , p
n
:
x = Az = z
1
p
1
+ z
n
p
n
.
The importance of component p
i
is determined by the size of the cor
responding Eigenvalue, λ
i
, which indicates the amount of variance ex
plained by p
i
. The key statistics is the proportion
λ
i
¸
n
j=1
λ
j
,
the explained variance by p
i
.
• Normalization: let ˜ w := (L
1/2
)
−1
z, where L
1/2
:= diag(
√
λ
1
, . . . ,
√
λ
n
),
and w = ˜ w −µ[ ˜ w] (µ[ ˜ w]=mean of ˜ w). Then
µ[w] = 0, Cov[w
i
, w
j
] = Cov[ ˜ w
i
, ˜ w
j
] = δ
ij
,
and
x = µ[x] +AL
1/2
w = µ[x] +
n
¸
j=1
p
j
λ
j
w
j
.
In components
x
i
= µ[x
i
] +
n
¸
j=1
A
ij
λ
j
w
j
.
• Sometimes the following view is useful (→ R[22](Chapter 3)): set
σ
i
:= V ar[x
i
]
1/2
=
ˆ
Σ
ii
1/2
=
n
¸
j=1
A
2
ij
λ
j
1/2
v
i
:=
x
i
−µ[x
i
]
σ
i
=
¸
n
j=1
A
ij
λ
j
w
j
σ
i
, i = 1, . . . , n.
3.2. PCA OF THE YIELD CURVE 35
Then we have µ[v
i
] = 0, µ[v
2
i
] = 1 and
x
i
= µ[x
i
] +σ
i
v
i
.
It can be appropriate to assume a parametric functional form (→ re
duction of parameters) of the correlation structure of x,
Corr[x
i
, x
j
] = Cov[v
i
, v
j
] =
ˆ
Σ
ij
σ
i
σ
j
=
¸
n
k=1
A
ik
A
jk
λ
k
σ
i
σ
j
= ρ(π; i, j),
where π is some lowdimensional parameter (this is adapted to the
calibration of market models → BM[6](Chapter 6.9)).
3.2 PCA of the Yield Curve
Now let x = (x
1
, . . . , x
n
)
T
be the increments of the forward curve, say
x
i
= R(t + ∆t; t + ∆t +τ
i−1
, t + ∆t +τ
i
) −R(t; t +τ
i−1
, t +τ
i
),
for some maturity spectrum 0 = τ
0
< < τ
n
.
PCA typically leads to the following picture (→ R[22]p.61): UK market
in the years 19891992 (the original maturity spectrum has been divided into
eight distinct buckets, i.e. n = 8).
The ﬁrst three principal components are
p
1
=
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
0.329
0.354
0.365
0.367
0.364
0.361
0.358
0.352
, p
2
=
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
−0.722
−0.368
−0.121
0.044
0.161
0.291
0.316
0.343
, p
3
=
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
0.490
−0.204
−0.455
−0.461
−0.176
0.176
0.268
0.404
.
• The ﬁrst PC is roughly ﬂat (parallel shift → average rate),
• the second PC is upward sloping (tilt → slope),
• the third PC humpshaped (ﬂex → curvature).
36 CHAPTER 3. STATISTICS OF THE YIELD CURVE
Figure 3.1: First Three PCs.
2 3 4 5 6 7 8
0.8
0.6
0.4
0.2
0.2
0.4
0.6
Table 3.1: Explained Variance of the Principal Components (PCs).
PC Explained
Variance (%)
1 92.17
2 6.93
3 0.61
4 0.24
5 0.03
6–8 0.01
The ﬁrst three PCs explain more than 99 % of the variance of x (→Table 3.1).
PCA of the yield curve goes back to the seminal paper by Litterman
and Scheinkman (91) [17] (Prof. J. Scheinkman is at the Department of
Economics, Princeton University).
3.3 Correlation
→ R[22](p.58)
A typical example of correlation among forward rates is provided by
3.3. CORRELATION 37
Brown and Schaefer (1994). The data is from the US Treasury yield curve
1987–1994. The following matrix (→ Figure 3.2)
¸
¸
¸
¸
¸
¸
¸
1 0.87 0.74 0.69 0.64 0.6
1 0.96 0.93 0.9 0.85
1 0.99 0.95 0.92
1 0.97 0.93
1 0.95
1
shows the correlation for changes of forward rates of maturities
0, 0.5, 1, 1.5, 2, 3 years.
Figure 3.2: Correlation between the short rate and instantaneous forward
rates for the US Treasury curve 1987–1994
0.5 1 1.5 2 2.5 3
0.6
0.7
0.8
0.9
1
→ Decorrelation occurs quickly.
→ Exponentially decaying correlation structure is plausible.
38 CHAPTER 3. STATISTICS OF THE YIELD CURVE
Chapter 4
Estimating the Yield Curve
4.1 A Bootstrapping Example
→ JW[11](p.129–136)
This is a naive bootstrapping method of ﬁtting to a money market yield
curve. The idea is to build up the yield curve
from shorter maturities to longer maturities.
We take Yen data from 9 January, 1996 (→ JW[11](Section 5.4)). The
spot date t
0
is 11 January, 1996. The daycount convention is Actual/360,
δ(T, S) =
actual number of days between T and S
360
.
Table 4.1: Yen data, 9 January 1996.
LIBOR (%) Futures Swaps (%)
o/n 0.49 20 Mar 96 99.34 2y 1.14
1w 0.50 19 Jun 96 99.25 3y 1.60
1m 0.53 18 Sep 96 99.10 4y 2.04
2m 0.55 18 Dec 96 98.90 5y 2.43
3m 0.56 7y 3.01
10y 3.36
39
40 CHAPTER 4. ESTIMATING THE YIELD CURVE
• The ﬁrst column contains the LIBOR (=simple spot rates) F(t
0
, S
i
) for
maturities
¦S
1
, . . . , S
5
¦ = ¦12/1/96, 18/1/96, 13/2/96, 11/3/96, 11/4/96¦
hence for 1, 7, 33, 60 and 91 days to maturity, respectively. The zero
coupon bonds are
P(t
0
, S
i
) =
1
1 +F(t
0
, S
i
) δ(t
0
, S
i
)
.
• The futures are quoted as
futures price for settlement day T
i
= 100(1 −F
F
(t
0
; T
i
, T
i+1
)),
where F
F
(t
0
; T
i
, T
i+1
) is the futures rate for period [T
i
, T
i+1
] prevailing
at t
0
, and
¦T
1
, . . . , T
5
¦ = ¦20/3/96, 19/6/96, 18/9/96, 18/12/96, 19/3/97¦,
hence δ(T
i
, T
i+1
) ≡ 91/360.
We treat futures rates as if they were simple forward rates, that is, we
set
F(t
0
; T
i
, T
i+1
) = F
F
(t
0
; T
i
, T
i+1
).
To calculate zerocoupon bond from futures prices we need P(t
0
, T
1
).
We use geometric interpoliation
P(t
0
, T
1
) = P(t
0
, S
4
)
q
P(t
0
, S
5
)
1−q
,
which is equivalent to using linear interpolation of continuously com
pounded spot rates
R(t
0
, T
1
) = q R(t
0
, S
4
) + (1 −q) R(t
0
, S
5
),
where
q =
δ(T
1
, S
5
)
δ(S
4
, S
5
)
=
22
31
= 0.709677.
Then we use the relation
P(t
0
, T
i+1
) =
P(t
0
, T
i
)
1 +δ(T
i
, T
i+1
) F(t
0
; T
i
, T
i+1
)
to derive P(t
0
, T
2
), . . . , P(t
0
, T
5
).
4.1. A BOOTSTRAPPING EXAMPLE 41
• Yen swaps have semiannual cashﬂows at dates
¦U
1
, . . . , U
20
¦ =
11/7/96, 13/1/97,
11/7/97, 12/1/98,
13/7/98, 11/1/99,
12/7/99, 11/1/00,
11/7/00, 11/1/01,
11/7/01, 11/1/02,
11/7/02, 13/1/03,
11/7/03, 12/1/04,
12/7/04, 11/1, 05,
11/7/05, 11/1/06
.
For a swap with maturity U
n
the swap rate at t
0
is given by
R
swap
(t
0
, U
n
) =
1 −P(t
0
, U
n
)
¸
n
i=1
δ(U
i−1
, U
i
) P(t
0
, U
i
)
, (U
0
:= t
0
).
From the data we have R
swap
(t
0
, U
i
) for i = 4, 6, 8, 10, 14, 20.
We obtain P(t
0
, U
1
), P(t
0
, U
2
) (and hence R
swap
(t
0
, U
1
), R
swap
(t
0
, U
2
))
by linear interpolation of the continuously compounded spot rates
R(t
0
, U
1
) =
69
91
R(t
0
, T
2
) +
22
91
R(t
0
, T
3
)
R(t
0
, U
2
) =
65
91
R(t
0
, T
4
) +
26
91
R(t
0
, T
5
).
All remaining swap rates are obtained by linear interpolation. For
maturity U
3
this is
R
swap
(t
0
, U
3
) =
1
2
(R
swap
(t
0
, U
2
) +R
swap
(t
0
, U
4
)).
We have (→ exercise)
P(t
0
, U
n
) =
1 −R
swap
(t
0
, U
n
)
¸
n−1
i=1
δ(U
i−1
, U
i
) P(t
0
, U
i
)
1 +R
swap
(t
0
, U
n
)δ(U
n−1
, U
n
)
.
This gives P(t
0
, U
n
) for n = 3, . . . , 20.
42 CHAPTER 4. ESTIMATING THE YIELD CURVE
Figure 4.1: Zerocoupon bond curve
2 4 6 8 10
Time to maturity
0.2
0.4
0.6
0.8
1
In Figure 4.1 is the implied zerocoupon bond price curve
P(t
0
, t
i
), i = 0, . . . , 29
(we have 29 points and set P(t
0
, t
0
) = 1).
The spot and forward rate curves are in Figure 4.2. Spot and forward
rates are continuously compounded
R(t
0
, t
i
) = −
log P(t
0
, t
i
)
δ(t
0
, t
i
)
R(t
0
, t
i
, t
i+1
) = −
log P(t
0
, t
i+1
) −log P(t
0
, t
i
)
δ(t
i
, t
i+1
)
, i = 1, . . . , 29.
The forward curve, reﬂecting the derivative of T → −log P(t
0
, T), is very
unsmooth and sensitive to slight variations (errors) in prices.
Figure 4.3 shows the spot rate curves from LIBOR, futures and swaps. It
is evident that the three curves are not coincident to a common underlying
curve. Our naive method made no attempt to meld the three curves together.
→ The entire yield curve is constructed from relatively few instruments. The
method exactly reconstructs market prices (this is desirable for interest
rate option traders). But it produces an unstable, nonsmooth forward
curve.
4.1. A BOOTSTRAPPING EXAMPLE 43
Figure 4.2: Spot rates (lower curve), forward rates (upper curve)
2 4 6 8 10
Time to maturity
0.01
0.02
0.03
0.04
0.05
0.06
Figure 4.3: Comparison of money market curves
0.5 1 1.5 2
Time to maturity
0.005
0.006
0.007
0.008
0.009
0.01
0.011
0.012
→ Another method would be to estimate a smooth yield curve parametri
cally from the market rates (for fund managers, long term strategies).
The main diﬃculties with our method are:
• Futures rates are treated as forward rates. In reality futures rates are
greater than forward rates. The amount by which the futures rate is
above the forward rate is called the convexity adjustment, which is
44 CHAPTER 4. ESTIMATING THE YIELD CURVE
model dependend. An example is
forward rate = futures rate −
1
2
σ
2
τ
2
,
where τ is the time to maturity of the futures contract, and σ is the
volatility parameter.
• LIBOR rates beyond the “stup date” T
1
= 20/3/96 (that is, at S
5
=
11/4/96) are ignored once P(t
0
, T
1
) is found. In general, the segments
of LIBOR, futures and swap markets overlap.
• Swap rates are inappropriately interpolated. The linear interpolation
produces a “sawtooth” in the forward rate curve. However, in some
markets intermediate swaps are indeed priced as if their prices were
found by linear interpolation.
4.2 General Case
The general problem of ﬁnding today’s (t
0
) term structure of zerocoupon
bond prices (or the discount function)
x → D(x) := P(t
0
, t
0
+x)
can be formulated as
p = C d +,
where p is a vector of n market prices, C the related cashﬂow matrix, and
d = (D(x
1
), . . . , D(x
N
)) with cashﬂow dates t
0
< T
1
< < T
N
,
T
i
−t
0
= x
i
,
and a vector of pricing errors. Reasons for including errors are
• prices are never exactly simultaneous,
• roundoﬀ errors in the quotes (bidask spreads, etc),
• liquidity eﬀects,
• tax eﬀects (high coupons, low coupons),
• allows for smoothing.
4.2. GENERAL CASE 45
4.2.1 Bond Markets
Data:
• vector of quoted/market bond prices p = (p
1
, . . . , p
n
),
• dates of all cashﬂows t
0
< T
1
< < T
N
,
• bond i with cashﬂows (coupon and principal payments) c
i,j
at time T
j
(may be zero), forming the n N cashﬂow matrix
C = (c
i,j
) 1≤i≤n
1≤j≤N
.
Example (→JW[11], p.426): UK government bond (gilt) market, Septem
ber 4, 1996, selection of nine gilts. The coupon payments are semiannual.
The spot date is 4/9/96, and the daycount convention is actual/365.
Table 4.2: Market prices for UK gilts, 4/9/96.
coupon next maturity dirty price
(%) coupon date (p
i
)
bond 1 10 15/11/96 15/11/96 103.82
bond 2 9.75 19/01/97 19/01/98 106.04
bond 3 12.25 26/09/96 26/03/99 118.44
bond 4 9 03/03/97 03/03/00 106.28
bond 5 7 06/11/96 06/11/01 101.15
bond 6 9.75 27/02/97 27/08/02 111.06
bond 7 8.5 07/12/96 07/12/05 106.24
bond 8 7.75 08/03/97 08/09/06 98.49
bond 9 9 13/10/96 13/10/08 110.87
Hence n = 9 and N = 1 + 3 + 6 + 7 + 11 + 12 + 19 + 20 + 25 = 104,
T
1
= 26/09/96, T
2
= 13/10/96, T
3
= 06/11/97, . . . .
46 CHAPTER 4. ESTIMATING THE YIELD CURVE
No bonds have cashﬂows at the same date. The 9 104 cashﬂow matrix is
C =
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
0 0 0 105 0 0 0 0 0 0 . . .
0 0 0 0 0 4.875 0 0 0 0 . . .
6.125 0 0 0 0 0 0 0 0 6.125 . . .
0 0 0 0 0 0 0 4.5 0 0 . . .
0 0 3.5 0 0 0 0 0 0 0 . . .
0 0 0 0 0 0 4.875 0 0 0 . . .
0 0 0 0 4.25 0 0 0 0 0 . . .
0 0 0 0 0 0 0 0 3.875 0 . . .
0 4.5 0 0 0 0 0 0 0 0 . . .
4.2.2 Money Markets
Money market data can be put into the same price–cashﬂow form as above.
LIBOR (rate L, maturity T): p = 1 and c = 1 + (T −t
0
)L at T.
FRA (forward rate F for [T, S]): p = 0, c
1
= −1 at T
1
= T, c
2
= 1+(S−T)F
at T
2
= S.
Swap (receiver, swap rate K, tenor t
0
≤ T
0
< < T
n
, T
i
− T
i−1
≡ δ):
since
0 = −D(T
0
−t
0
) +δK
n
¸
j=1
D(T
j
−t
0
) + (1 +δK)D(T
n
−t
0
),
• if T
0
= t
0
: p = 1, c
1
= = c
n−1
= δK, c
n
= 1 +δK,
• if T
0
> t
0
: p = 0, c
0
= −1, c
1
= = c
n−1
= δK, c
n
= 1 +δK.
→ at t
0
: LIBOR and swaps have notional price 1, FRAs and forward swaps
have notional price 0.
Example (→ JW[11], p.428): US money market on October 6, 1997.
The daycount convention is Actual/360. The spot date t
0
is 8/10/97.
LIBOR is for o/n (1/365), 1m (33/360), and 3m (92/360).
4.2. GENERAL CASE 47
Futures are three month rates (δ = 91/360). We take them as forward
rates. That is, the quote of the futures contract with maturity date (settle
ment day) T is
100(1 −F(t
0
; T, T +δ)).
Swaps are annual (δ = 1). The ﬁrst payment date is 8/10/98.
Table 4.3: US money market, October 6, 1997.
Period Rate Maturity Date
LIBOR o/n 5.59375 9/10/97
1m 5.625 10/11/97
3m 5.71875 8/1/98
Futures Oct97 94.27 15/10/97
Nov97 94.26 19/11/97
Dec97 94.24 17/12/97
Mar98 94.23 18/3/98
Jun98 94.18 17/6/98
Sep98 94.12 16/9/98
Dec98 94 16/12/98
Swaps 2 6.01253
3 6.10823
4 6.16
5 6.22
7 6.32
10 6.42
15 6.56
20 6.56
30 6.56
Here n = 3 + 7 + 9 = 19, N = 3 + 14 + 30 = 47, T
1
= 9/10/97,
T
2
= 15/10/97 (ﬁrst future), T
3
= 10/11/97, . . . . The ﬁrst 14 columns of
48 CHAPTER 4. ESTIMATING THE YIELD CURVE
the 19 47 cashﬂow matrix C are
c
11
0 0 0 0 0 0 0 0 0 0 0 0 0
0 0 c
23
0 0 0 0 0 0 0 0 0 0 0
0 0 0 0 0 c
36
0 0 0 0 0 0 0 0
0 −1 0 0 0 0 c
47
0 0 0 0 0 0 0
0 0 0 −1 0 0 0 c
58
0 0 0 0 0 0
0 0 0 0 −1 0 0 0 c
69
0 0 0 0 0
0 0 0 0 0 0 0 0 −1 c
7,10
0 0 0 0
0 0 0 0 0 0 0 0 0 −1 c
8,11
0 0 0
0 0 0 0 0 0 0 0 0 0 −1 0 c
9,13
0
0 0 0 0 0 0 0 0 0 0 0 0 −1 c
10,14
0 0 0 0 0 0 0 0 0 0 0 c
11,12
0 0
0 0 0 0 0 0 0 0 0 0 0 c
12,12
0 0
0 0 0 0 0 0 0 0 0 0 0 c
13,12
0 0
0 0 0 0 0 0 0 0 0 0 0 c
14,12
0 0
0 0 0 0 0 0 0 0 0 0 0 c
15,12
0 0
0 0 0 0 0 0 0 0 0 0 0 c
16,12
0 0
0 0 0 0 0 0 0 0 0 0 0 c
17,12
0 0
0 0 0 0 0 0 0 0 0 0 0 c
18,12
0 0
0 0 0 0 0 0 0 0 0 0 0 c
19,12
0 0
with
c
11
= 1.00016, c
23
= 1.00516, c
36
= 1.01461,
c
47
= 1.01448, c
58
= 1.01451, c
69
= 1.01456, c
7,10
= 1.01459,
c
8,11
= 1.01471, c
9,13
= 1.01486, c
10,14
= 1.01517
c
11,12
= 0.060125, c
12,12
= 0.061082, c
13,12
= 0.0616,
c
14,12
= 0.0622, c
15,12
= 0.0632, c
16,12
= 0.0642,
c
17,12
= c
18,12
= c
19,12
= 0.0656.
4.2.3 Problems
Typically, we have n < N. Moreover, many entries of C are zero (diﬀerent
cashﬂow dates). This makes ordinary least square (OLS) regression
min
d∈R
N
¦
2
[ = p −C d¦ (⇒C
T
p = C
T
Cd
∗
)
unfeasible.
4.2. GENERAL CASE 49
One could chose the data set such that cashﬂows are at same points in
time (say four dates each year) and the cashﬂow matrix C is not entirely full
of zeros (Carleton–Cooper (1976)). Still regression only yields values D(x
i
)
at the payment dates t
0
+x
i
→ interpolation technics necessary.
But there is nothing to regularize the discount factors (discount factors of
similar maturity can be very diﬀerent). As a result this leads to a ragged
spot rate (yield) curve, and even worse for forward rates.
4.2.4 Parametrized Curve Families
Reduction of parameters and smooth yield curves can be achieved by using
parametrized families of smooth curves
D(x) = D(x; z) = exp
−
x
0
φ(u; z) du
, z ∈ Z,
with state space Z ⊂ R
m
.
For regularity reasons (see below) it is best to estimate the forward curve
R
+
÷ x → f(t
0
, t
0
+x) = φ(x) = φ(x; z).
This leads to a nonlinear optimization problem
min
z∈?
p −C d(z) ,
with
d
i
(z) = exp
−
x
i
0
φ(u; z) du
for some payment tenor 0 < x
1
< < x
N
.
Linear Families
Fix a set of basis functions ψ
1
, . . . , ψ
m
(preferably with compact support),
and let
φ(x; z) = z
1
ψ
1
(x) + + z
m
ψ
m
(x).
50 CHAPTER 4. ESTIMATING THE YIELD CURVE
Cubic Bsplines A cubic spline is a piecewise cubic polynomial that is
everywhere twice diﬀerentiable. It interpolates values at m + 1 knot points
ξ
0
< < ξ
m
. Its general form is
σ(x) =
3
¸
i=0
a
i
x
i
+
m−1
¸
j=1
b
j
(x −ξ
j
)
3
+
,
hence it has m+3 parameters ¦a
0
, . . . , a
4
, b
1
, . . . , b
m−1
¦ (a kth degree spline
has m+k parameters). The spline is uniquely characterized by speciﬁcation
of σ
t
or σ
tt
at ξ
0
and ξ
m
.
Introduce six extra knot points
ξ
−3
< ξ
−2
< ξ
−1
< ξ
0
< < ξ
m
< ξ
m+1
< ξ
m+2
< ξ
m+3
.
A basis for the cubic splines on [ξ
0
, ξ
m
] is given by the m + 3 Bsplines
ψ
k
(x) =
k+4
¸
j=k
k+4
¸
i=k,i,=j
1
ξ
i
−ξ
j
(x −ξ
j
)
3
+
, k = −3, . . . , m−1.
The Bspline ψ
k
is zero outside [ξ
k
, ξ
k+4
].
Figure 4.4: Bspline with knot points ¦0, 1, 6, 8, 11¦.
2 4 6 8 10 12
0.01
0.02
0.03
0.04
0.05
0.06
4.2. GENERAL CASE 51
Estimating the Discount Function Bsplines can also be used to esti
mate the discount function directly (Steeley (1991)),
D(x; z) = z
1
ψ
1
(x) + +z
m
ψ
m
(x).
With
d(z) =
¸
¸
D(x
1
; z)
.
.
.
D(x
N
; z)
=
¸
¸
ψ
1
(x
1
) ψ
m
(x
1
)
.
.
.
.
.
.
ψ
1
(x
N
) ψ
m
(x
N
)
¸
¸
z
1
.
.
.
z
m
=: Ψ z
this leads to the linear optimization problem
min
z∈R
m
p −CΨz.
If the n m matrix A := CΨ has full rank m, the unique unconstrained
solution is
z
∗
= (A
T
A)
−1
A
T
p.
A reasonable constraint would be
D(0; z) = ψ
1
(0)z
1
+ +ψ
m
(0)z
m
= 1.
Example We take the UK government bond market data from the last
section (Table 4.2). The maximum time to maturity, x
104
, is 12.11 [years].
Notice that the ﬁrst bond is a zerocoupon bond. Its exact yield is
y = −
365
72
log
103.822
105
= −
1
0.197
log 0.989 = 0.0572.
• As a basis we use the 8 (resp. ﬁrst 7) Bsplines with the 12 knot points
¦−20, −5, −2, 0, 1, 6, 8, 11, 15, 20, 25, 30¦
(see Figure 4.5).
The estimation with all 8 Bsplines leads to
min
z∈R
8
p −CΨz = p −CΨz
∗
 = 0.23
52 CHAPTER 4. ESTIMATING THE YIELD CURVE
Figure 4.5: Bsplines with knots ¦−20, −5, −2, 0, 1, 6, 8, 11, 15, 20, 25, 30¦.
5 10 15 20 25 30
0.01
0.02
0.03
0.04
0.05
0.06
0.07
with
z
∗
=
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
13.8641
11.4665
8.49629
7.69741
6.98066
6.23383
−4.9717
855.074
,
and the discount function, yield curve (cont. comp. spot rates), and for
ward curve (cont. comp. 3monthly forward rates) shown in Figure 4.7.
The estimation with only the ﬁrst 7 Bsplines leads to
min
z∈R
7
p −CΨz = p −CΨz
∗
 = 0.32
with
z
∗
=
¸
¸
¸
¸
¸
¸
¸
¸
¸
17.8019
11.3603
8.57992
7.56562
7.28853
5.38766
4.9919
,
4.2. GENERAL CASE 53
and the discount function, yield curve (cont. comp. spot rates), and
forward curve (cont. comp. 3month forward rates) shown in Figure 4.8.
• Next we use only 5 Bsplines with the 9 knot points
¦−10, −5, −2, 0, 4, 15, 20, 25, 30¦
(see Figure 4.6).
Figure 4.6: Five Bsplines with knot points ¦−10, −5, −2, 0, 4, 15, 20, 25, 30¦.
5 10 15 20 25 30
0.005
0.01
0.015
0.02
0.025
0.03
0.035
The estimation with this 5 Bsplines leads to
min
z∈R
5
p −CΨz = p −CΨz
∗
 = 0.39
with
z
∗
=
¸
¸
¸
¸
¸
15.652
19.4385
12.9886
7.40296
6.23152
,
and the discount function, yield curve (cont. comp. spot rates), and for
ward curve (cont. comp. 3monthly forward rates) shown in Figure 4.9.
54 CHAPTER 4. ESTIMATING THE YIELD CURVE
Figure 4.7: Discount function, yield and forward curves for estimation with
8 Bsplines. The dot is the exact yield of the ﬁrst bond.
2 4 6 8 10 12
0.2
0.4
0.6
0.8
1
2 4 6 8 10 12
0.02
0.04
0.06
0.08
0.1
0.12
0.14
2 4 6 8 10 12
0.05
0.1
0.15
0.2
0.25
4.2. GENERAL CASE 55
Figure 4.8: Discount function, yield and forward curves for estimation with
7 Bsplines. The dot is the exact yield of the ﬁrst bond.
2 4 6 8 10 12
0.2
0.4
0.6
0.8
1
2 4 6 8 10 12
0.02
0.04
0.06
0.08
0.1
0.12
0.14
2 4 6 8 10 12
0.05
0.1
0.15
0.2
0.25
56 CHAPTER 4. ESTIMATING THE YIELD CURVE
Figure 4.9: Discount function, yield and forward curves for estimation with
5 Bsplines. The dot is the exact yield of the ﬁrst bond.
2 4 6 8 10 12
0.2
0.4
0.6
0.8
1
2 4 6 8 10 12
0.02
0.04
0.06
0.08
0.1
0.12
0.14
2 4 6 8 10 12
0.05
0.1
0.15
0.2
0.25
4.2. GENERAL CASE 57
Discussion
• In general, splines can produce bad ﬁts.
• Estimating the discount function leads to unstable and nonsmooth
yield and forward curves. Problems mostly at short and long term
maturities.
• Splines are not useful for extrapolating to long term maturities.
• There is a tradeoﬀ between the quality (or regularity) and the correct
ness of the ﬁt. The curves in Figures 4.8 and 4.9 are more regular than
those in Figure 4.7, but their correctness criteria (0.32 and 0.39) are
worse than for the ﬁt with 8 Bsplines (0.23).
• The Bspline ﬁts are extremely sensitive to the number and location of
the knot points.
→ Need criterions asserting smooth yield and forward curves that do not
ﬂuctuate too much and ﬂatten towards the long end.
→ Direct estimation of the yield or forward curve.
→ Optimal selection of number and location of knot points for splines.
→ Smoothing splines.
Smoothing Splines The least squares criterion
min
z
p −C d(z)
2
has to be replaced/extended by criterions for the smoothness of the yield or
forward curve.
Example: Lorimier (95). In her PhD thesis 1995, Sabine Lorimier sug
gests a spline method where the number and location of the knots are deter
mined by the observed data itself.
For ease of notation we set t
0
= 0 (today). The data is given by N
observed zerocoupon bonds P(0, T
1
), . . . , P(0, T
N
) at 0 < T
1
< < T
N
≡
T, and consequently the N yields
Y
1
, . . . , Y
N
, P(0, T
i
) = exp(−T
i
Y
i
).
58 CHAPTER 4. ESTIMATING THE YIELD CURVE
Let f(u) denote the forward curve. The ﬁtting requirement now is for the
forward curve
T
i
0
f(u) du +
i
/
√
α = T
i
Y
i
, (4.1)
with an arbitrary constant α > 0. The aim is to minimize 
2
as well as the
smoothness criterion
T
0
(f
t
(u))
2
du. (4.2)
Introduce the Sobolev space
H = ¦g [ g
t
∈ L
2
[0, T]¦
with scalar product
'g, h`
H
= g(0)h(0) +
T
0
g
t
(u)h
t
(u) du,
and the nonlinear functional on H
F(f) :=
¸
T
0
(f
t
(u))
2
du +α
N
¸
i=1
Y
i
T
i
−
T
i
0
f(u) du
2
¸
.
The optimization problem then is
min
f∈H
F(f). (*)
The parameter α tunes the tradeoﬀ between smoothness and correctness of
the ﬁt.
Theorem 4.2.1. Problem (*) has a unique solution f, which is a second
order spline characterized by
f(u) = f(0) +
N
¸
k=1
a
k
h
k
(u) (4.3)
where h
k
∈ C
1
[0, T] is a second order polynomial on [0, T
k
] with
h
t
k
(u) = (T
k
−u)
+
, h
k
(0) = T
k
, k = 1, . . . , N, (4.4)
4.2. GENERAL CASE 59
and f(0) and a
k
solve the linear system of equations
N
¸
k=1
a
k
T
k
= 0, (4.5)
α
Y
k
T
k
−f(0)T
k
−
N
¸
l=1
a
l
'h
l
, h
k
`
H
= a
k
, k = 1, . . . , N. (4.6)
Proof. Integration by parts yields
T
k
0
g(u) du = T
k
g(T
k
) −
T
k
0
ug
t
(u) du
= T
k
g(0) +T
k
T
k
0
g
t
(u) du −
T
k
0
ug
t
(u) du
= T
k
g(0) +
T
0
(T
k
−u)
+
g
t
(u) du = 'h
k
, g`
H
,
for all g ∈ H. In particular,
T
k
0
h
l
du = 'h
l
, h
k
`
H
.
A (local) minimizer f of F satisﬁes
d
d
F(f +g)[
=0
= 0
or equivalently
T
0
f
t
g
t
du = α
N
¸
k=1
Y
k
T
k
−
T
k
0
f du
T
k
0
g du, ∀g ∈ H. (4.7)
In particular, for all g ∈ H with 'g, h
k
`
H
= 0 we obtain
'f −f(0), g`
H
=
T
0
f
t
(u)g
t
(u) du = 0.
Hence
f −f(0) ∈ span¦h
1
, . . . , h
N
¦
60 CHAPTER 4. ESTIMATING THE YIELD CURVE
what proves (4.3), (4.4) and (4.5) (set u = 0). Hence we have
T
0
f
t
(u)g
t
(u) du =
N
¸
k=1
a
k
−T
k
g(0) +
T
k
0
g(u) du
=
N
¸
k=1
a
k
T
k
0
g(u) du,
and (4.7) can be rewritten as
N
¸
k=1
a
k
−α
Y
k
T
k
−f(0)T
k
−
N
¸
l=1
a
l
'h
l
, h
k
`
H
T
k
0
g(u) du = 0
for all g ∈ H. This is true if and only if (4.6) holds.
Thus we have shown that (4.7) is equivalent to (4.3)–(4.6).
Next we show that (4.7) is a suﬃcient condition for f to be a global
minimizer of F. Let g ∈ H, then
F(g) =
T
0
((g
t
−f
t
) +f
t
)
2
du +α
N
¸
k=1
Y
k
T
k
−
T
k
0
g du
2
(4.7)
= F(f) +
T
0
(g
t
−f
t
)
2
du +α
N
¸
k=1
T
k
0
f du −
T
k
0
g du
2
≥ F(f),
where we used (4.7) with g −f ∈ H.
It remains to show that f exists and is unique; that is, that the linear sys
tem (4.5)–(4.6) has a unique solution (f(0), a
1
, . . . , a
N
). The corresponding
(N + 1) (N + 1) matrix is
A =
¸
¸
¸
¸
0 T
1
T
2
T
N
αT
1
α'h
1
, h
1
`
H
+ 1 α'h
1
, h
2
`
H
α'h
1
, h
N
`
H
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
αT
N
α'h
N
, h
1
`
H
α'h
N
, h
2
`
H
α'h
N
, h
N
`
H
+ 1
. (4.8)
Let λ = (λ
0
, . . . , λ
N
)
T
∈ R
N+1
such that Aλ = 0, that is,
N
¸
k=1
T
k
λ
k
= 0
αT
k
λ
0
+α
N
¸
l=1
'h
k
, h
l
`
H
λ
l
+λ
k
= 0, k = 1, . . . , N.
4.2. GENERAL CASE 61
Multiplying the latter equation with λ
k
and summing up yields
α
N
¸
k=1
λ
k
h
k
2
H
+
N
¸
k=1
λ
2
k
= 0.
Hence λ = 0, whence A is nonsingular.
The role of α is as follows:
• If α → 0 then by (4.3) and (4.6) we have f(u) ≡ f(0), a constant
function. That is, maximal regularity
T
0
(f
t
(u))
2
du = 0
but no ﬁtting of data, see (4.1).
• If α →∞ then (4.7) implies that
T
k
0
f(u) du = Y
k
T
k
, k = 1, . . . , N, (4.9)
a perfect ﬁt. That is, f minimizes (4.2) subject to the constraints (4.9).
To estimate the forward curve from N zerocoupon bonds—that is, yields
Y = (Y
1
, . . . , Y
N
)
T
—one has to solve the linear system
A
f(0)
a
=
0
Y
(see (4.8)).
Of course, if coupon bond prices are given, then the above method has
to be modiﬁed and becomes nonlinear. With p ∈ R
n
denoting the market
price vector and c
kl
the cashﬂows at dates T
l
, k = 1, . . . , n, l = 1, . . . , N, this
reads
min
f∈H
T
0
(f
t
)
2
du +α
n
¸
k=1
log p
k
−log
¸
N
¸
l=1
c
kl
exp
¸
−
T
l
0
f du
¸
2
.
If the coupon payments are small compared to the nominal (=1), then this
problem has a unique solution. This and much more is carried out in Lorim
ier’s thesis.
62 CHAPTER 4. ESTIMATING THE YIELD CURVE
ExponentialPolynomial Families
Exponentialpolynomial functions
p
1
(x)e
−α
1
x
+ +p
n
(x)e
−αnx
(p
i
=polynomial of degree n
i
)
form nonlinear families of functions. Popular examples are:
Nelson–Siegel (87) [20] There are 4 parameters z
1
, . . . , z
4
and
φ
NS
(x; z) = z
1
+ (z
2
+z
3
x)e
−z
4
x
.
Svensson (94) [26] (Prof. L. E. O. Svensson is at the Economics Depart
ment, Princeton University) This is an extension of Nelson–Siegel, in
cluding 6 parameters z
1
, . . . , z
6
,
φ
S
(x; z) = z
1
+ (z
2
+z
3
x)e
−z
4
x
+z
5
e
−z
6
x
.
Figure 4.10: Nelson–Siegel curves for z
1
= 7.69, z
2
= −4.13, z
4
= 0.5 and 7
diﬀerent values for z
3
= 1.76, 0.77, −0.22, −1.21, −2.2, −3.19, −4.18.
5 10 15 20
2
4
6
8
Table 4.4 is taken from a document of the Bank for International Settle
ments (BIS) 1999 [2].
4.2. GENERAL CASE 63
Table 4.4: Overview of estimation procedures by several central banks. BIS
1999 [2]. NS is for Nelson–Siegel, S for Svensson, wp for weighted prices.
Central Bank Method Minimized Error
Belgium S or NS wp
Canada S sp
Finland NS wp
France S or NS wp
Germany S yields
Italy NS wp
Japan smoothing prices
splines
Norway S yields
Spain S wp
Sweden S yields
UK S yields
USA smoothing bills: wp
splines bonds: prices
Criteria for Curve Families
• Flexibility (do the curves ﬁt a wide range of term structures?)
• Number of factors not too large (curse of dimensionality).
• Regularity (smooth yield or forward curves that ﬂatten out towards the
long end).
• Consistency: do the curve families go well with interest rate models?
→ this point will be exploited in the sequel.
64 CHAPTER 4. ESTIMATING THE YIELD CURVE
Chapter 5
Why Yield Curve Models?
→ R[22](Chapter 5)
Why modelling the entire term structure of interest rates? There is no
need when pricing a single European call option on a bond.
But: the payoﬀs even of “plainvanilla” ﬁxed income products such as caps,
ﬂoors, swaptions consist of a sequence of cashﬂows at T
1
, . . . , T
n
, where
n may be 20 (e.g. a 10y swap with semiannual payments) or more.
→ The valuation of such products requires the modelling of the entire covari
ance structure. Historical estimation of such large covariance matrices
is statistically not tractable anymore.
→ Need strong structure to be imposed on the comovements of ﬁnancial
quantities of interest.
→ Specify the dynamics of a small number of variables (e.g. PCA).
→ Correlation structure among observable quantities can now be obtained
analytically or numerically.
→ Simultaneous pricing of diﬀerent options and hedging instruments in a
consistent framework.
This is exactly what interest rate (curve) models oﬀer:
• reduction of ﬁtting degrees of freedom → makes problem manageable.
=⇒ It is practically and intellectually rewarding to consider noarbitrage
conditions in much broader generality.
65
66 CHAPTER 5. WHY YIELD CURVE MODELS?
Chapter 6
NoArbitrage Pricing
This chapter brieﬂy recalls the basics about pricing and hedging in a Brown
ian motion driven market. Reference is B[3], MR[19](Chapter 10), and many
more.
6.1 SelfFinancing Portfolios
The stochastic basis is a probability space (Ω, T, P), a ddimensional Brow
nian motion W = (W
1
, . . . , W
d
), and the ﬁltration (T
t
)
t≥0
generated by W.
We shall assume that T = T
∞
= ∨
t≥0
T
t
, and do not a priori ﬁx a ﬁnite
time horizon. This is not a restriction since always one can set a stochastic
process to be zero after a ﬁnite time T if this were the ultimate time horizon
(as in the Black–Scholes model).
The background for stochastic analysis can be found in many textbooks,
such as [14], [25], [23], etc. From time to time we recall some of the funda
mental results without proof.
Financial Market We consider a ﬁnancial market with n traded assets,
following strictly positive Itˆ o processes
dS
i
(t) = S
i
(t)µ
i
(t) dt +
d
¸
j=1
S
i
(t)σ
ij
(t) dW
j
(t), S
i
> 0, i = 1, . . . , n
and the riskfree asset
dS
0
(t) = r(t)S
0
(t) dt, S
0
(0) = 1
⇔S
0
(t) = e
t
0
r(s) ds
.
67
68 CHAPTER 6. NOARBITRAGE PRICING
The drift µ = (µ
1
, . . . , µ
n
), volatility σ = (σ
ij
), and short rates r are assumed
to form adapted processes which meet the required integrability conditions
such that all of the above (stochastic) integrals are welldeﬁned.
Remark 6.1.1. It is always understood that for a random variable “X ≥ 0”
means “X ≥ 0 a.s.” (that is, P[X ≥ 0] = 1), etc.
Theorem 6.1.2 (Stochastic Integrals). Let h = (h
1
, . . . , h
d
) be a mea
surable adapted process. If
t
0
h(s)
2
ds < ∞ for all t > 0
(the class of such processes is denoted by L) one can deﬁne the stochastic
integral
(h W)
t
≡
t
0
h(s) dW(s) ≡
d
¸
j=1
t
0
h
j
(s) dW
j
(s).
If moreover
E
¸
∞
0
h(s)
2
ds
< ∞
(the class of such processes is denoted by L
2
) then h W is a martingale and
the Itˆ o isometry holds
E
¸
t
0
h(s) dW(s)
2
¸
= E
¸
t
0
h(s)
2
ds
.
Selfﬁnancing Portfolios A portfolio, or trading strategy, is any adapted
process
φ = (φ
0
, . . . , φ
n
).
Its corresponding value process is
V (t) = V (t; φ) :=
n
¸
i=0
φ
i
(t)S
i
(t).
The portfolio φ is called selfﬁnancing (for S) if the stochastic integrals
t
0
φ
i
(u) dS
i
(u), i = 0, . . . , n
6.2. ARBITRAGE AND MARTINGALE MEASURES 69
are well deﬁned and
dV (t; φ) =
n
¸
i=0
φ
i
(t) dS
i
(t).
Numeraires All prices are interpreted as being given in terms of a nu
meraire, which typically is a local currency such as US dollars. But we may
and will express from time to time the prices in terms of other numeraires,
such as S
p
for some 0 ≤ p ≤ n. The discounted price process vector
Z(t) :=
S(t)
S
p
(t)
implies the discounted value process
˜
V (t; φ) :=
n
¸
i=0
φ
i
(t)Z
i
(t) =
V (t; φ)
S
p
(t)
.
Up to integrability, the selfﬁnancing property does not depend on the choice
of the numeraire.
Lemma 6.1.3. Suppose that a portfolio φ satisﬁes the integrability conditions
for S and Z. Then φ is selfﬁnancing for S if and only if it is selfﬁnancing
for Z, in particular
d
˜
V (t; φ) =
n
¸
i=0
φ
i
(t) dZ
i
(t) =
n
¸
i=0
i=p
φ
i
(t) dZ
i
(t). (6.1)
Since Z
p
is constant, the number of terms in (6.1) reduces to n.
Often (but not always) we chose S
0
as the numeraire.
6.2 Arbitrage and Martingale Measures
Contingent Claims Related to any option (such as a cap, ﬂoor, swaption,
etc) is an uncertain future payoﬀ, say at date T, hence an T
T
measurable
random variable X (a contingent (T)claim). Two main problems now are:
• What is a “fair” price for a contingent claim X?
• How can one hedge against the ﬁnancial risk involved in trading con
tingent claims?
70 CHAPTER 6. NOARBITRAGE PRICING
Arbitrage An arbitrage portfolio is a selfﬁnancing portfolio φ with value
process satisfying
V (0) = 0 and V (T) ≥ 0 and P[V (T) > 0] > 0
for some T > 0. If no arbitrage portfolios exist for any T > 0 we say the
model is arbitragefree.
An example of arbitrage is the following.
Lemma 6.2.1. Suppose there exists a selfﬁnancing portfolio with value pro
cess
dU(t) = k(t)U(t) dt,
for some measurable adapted process k. If the market is arbitragefree then
necessarily
r = k, dt ⊗dPa.s.
Proof. Indeed, after discounting with S
0
we obtain
˜
U(t) :=
U(t)
S
0
(t)
= U(0) exp
t
0
(k(s) −r(s)) ds
.
Then (→ exercise)
ψ(t) := 1
k(t)>r(t)¦
yields a selfﬁnancing strategy with discounted value process
˜
V (t) =
t
0
ψ(s) d
˜
U(s) =
t
0
1
k(s)>r(s)¦
(k(s) −r(s))
˜
U(s)
ds ≥ 0.
Hence absence of arbitrage requires
0 = E[
˜
V (T)] =
A
1
k(t,ω)>r(t,ω)¦
(k(t, ω) −r(t, ω))
˜
U(t, ω)
. .. .
>0 on A
dt ⊗dP
where
^ := ¦(t, ω) [ k(t, ω) > r(t, ω)¦
is a measurable subset of [0, T] Ω. But this can only hold if ^ is a dt ⊗dP
nullset. Using the same arguments with changed signs proves the lemma.
6.2. ARBITRAGE AND MARTINGALE MEASURES 71
Martingale Measures We now investigate when a given model is arbi
tragefree. To simplify things in the sequel
• we ﬁx S
0
as a numeraire, and
•
˜
V will express the discounted value process V/S
0
.
But the following can be made valid for any choice of numeraire.
An equivalent probability measure Q ∼ P is called an equivalent (local)
martingale measure (E(L)MM) if the discounted price processes
Z
i
= S
i
/S
0
are Q(local) martingales.
Theorem 6.2.2 (Girsanov’s Change of Measure Theorem). Let Q ∼ P
be an equivalent probability measure. Then there exists γ ∈ L such that the
density process dQ/dP is the stochastic exponential c(γ W) of γ W
dQ
dP
[
Tt
= c
t
(γ W) := exp
t
0
γ(s) dW(s) −
1
2
t
0
γ(s)
2
ds
. (6.2)
Moreover, the process
˜
W(t) := W(t) −
t
0
γ(s) ds (6.3)
is a QBrownian motion.
Conversely, if γ ∈ L is such that c (γ W) is a uniformly integrable
martingale with c
∞
(γ W) > 0 — suﬃcient is the Novikov condition
E
¸
exp
1
2
∞
0
γ(s)
2
ds
< ∞ (6.4)
(see [23, Proposition (1.26), Chapter IV]) — then (6.2) deﬁnes an equivalent
probability measure Q ∼ P.
Market Price of Risk Let Q be an ELMM and γ (the stochastic logarithm
of the density process) and
˜
W given by (6.2) and (6.3). Integration by parts
yields the Zdynamics
dZ
i
(t) = Z
i
(t) (µ
i
(t) −r(t)) dt +Z
i
(t)σ
i
(t) dW(t)
= Z
i
(t) (µ
i
(t) −r(t) +σ
i
(t) γ(t)) dt +Z
i
(t)σ
i
(t) d
˜
W(t).
72 CHAPTER 6. NOARBITRAGE PRICING
Hence necessarily γ satisﬁes
µ
i
−r +σ
i
γ = 0 dt ⊗dQa.s. for all i = 1, . . . , n. (6.5)
If σ is nondegenerate (in particular d ≤ n and rank[σ] = d) then γ is
uniquely speciﬁed by
−γ = σ
−1
(µ −r1)
where 1 := (1, . . . , 1)
T
, and vice versa. This is why −γ is called the market
price of risk.
Conversely, if (6.5) has a solution γ ∈ L such that c(γ W) is a uniformly
integrable martingale (the Novikov condition (6.4) is suﬃcient) then (6.2)
deﬁnes an ELMM Q. If γ is unique then Q is the unique ELMM.
Notice that, by Itˆ o’s formula, Z
i
can be written as stochastic exponential
Z
i
= c(σ
i
˜
W).
Hence if σ
i
satisﬁes the Novikov condition (6.4) for all i = 1, . . . , n then the
ELMM Q is in fact an EMM.
Admissible Strategies In the presence of local martingales one has to be
alert to pitfalls. For example it is possible to construct a local martingale M
with M(0) = 0 and M(1) = 1. Even worse, M can be chosen to be of the
form
M(t) =
t
0
φ(s) dW(s)
(Dudley’s Representation Theorem), which looks like the (discounted) value
process of a selfﬁnancing strategy. This would certainly be a moneymaking
machine, say arbitrage. In the same way “suicide strategies” (e.g. M(0) = 1
and M(1) = 0) can be constructed. To rule out such examples we have to
impose additional constraints on the choice of strategies. There are several
ways to do so. Here are two typical examples:
A selfﬁnancing strategy φ is admissible if
1.
˜
V (t; φ) ≥ −a for some a ∈ R, OR
2.
˜
V (t; φ) is a true Qmartingale, for some ELMM Q.
Condition 1 is more universal (it does not depend on a particular Q) and
implies that V (t; φ) is a Qsupermartingale for every ELMMQ. Yet, “suicide
strategies” remain (however, they do not introduce arbitrage).
Both conditions 1 and 2, however, are sensitive with respect to the choice
of numeraire!
6.3. HEDGING AND PRICING 73
The Fundamental Theorem of Asset Pricing The existence of an
ELMM rules out arbitrage.
Lemma 6.2.3. Suppose there exists an ELMM Q. Then the model is arbi
tragefree, in the sense that there exists no admissible (either Condition 1 or
2) arbitrage strategy.
Proof. Indeed, let
˜
V be the discounted value process of an admissible strat
egy, with
˜
V (0) = 0 and
˜
V (T) ≥ 0. Since
˜
V is a Qsupermartingale in any
case (for some ELMM Q), we have
0 ≤ E
Q
[
˜
V (T)] ≤
˜
V (0) = 0,
whence
˜
V (T) = 0.
It is folklore (Delbaen and Schachermayer 1994, etc) that also the converse
holds true: if arbitrage is deﬁned in the right way (“No Free Lunch with
Vanishing Risk”), then its absence implies the existence of an ELMM Q.
This is called the Fundamental Theorem of Asset Pricing.
It has become a custom (and we will follow this tradition) to consider the
existence of an ELMM as synonym for the absence of arbitrage:
absence of arbitrage = existence of an ELMM;
→ the existence of an ELMM is now a standing assumption.
6.3 Hedging and Pricing
Attainable Claims A contingent claim X due at T is attainable if the
exists an admissible strategy φ which replicates/hedges X; that is,
V (T; φ) = X.
A simple example: suppose S
1
is the price process of the Tbond. Then
the contingent claim X = 1 due at T is attainable by an obvious buy and
hold strategy with value process V (t) = S
1
(t).
74 CHAPTER 6. NOARBITRAGE PRICING
Complete Markets The main problem is to determine which claims are
attainable. This is most conveniently carried out in terms of discounted
prices.
Suppose that σ is nondegenerate; that is
d ≤ n and rank[σ] = d, (6.6)
and that the unique market price of risk −γ given by (6.5) yields a uniformly
integrable martingale c(γ W) and hence a unique ELMM Q.
Lemma 6.3.1. Then the model is complete in the sense that any contingent
claim X with
X/S
0
(T) ∈ L
1
(T
T
; Q) (6.7)
is attainable.
Proof. Deﬁne the Qmartingale
Y (t) := E
Q
[X/S
0
(T) [ T
t
] , t ∈ [0, T].
Then
Y (t)D(t) = D(t)E
Q
[Y (T) [ T
t
]
Bayes
= E[Y (T)D(T) [ T
t
],
with the density process D(t) = dQ/dP[
Tt
= c
t
(γ W). Hence Y D is a P
martingale and by the representation theorem 6.3.3 we can ﬁnd ψ ∈ L such
that
Y (t)D(t) = Y (0) +
t
0
ψ(s) dW(s).
Applying Itˆ o’s formula yields
d
1
D
= −
1
D
γ dW +
1
D
γ
2
dt,
and
dY = d
(Y D)
1
D
= Y Dd
1
D
+
1
D
d(Y D) +d
Y D,
1
D
=
1
D
ψ −Y γ
dW −
1
D
ψ −Y γ
γ dt
=
1
D
ψ −Y γ
. .. .
=:
˜
ψ
d
˜
W.
6.3. HEDGING AND PRICING 75
Now deﬁne
φ
i
=
((σ
−1
)
T
˜
ψ)
i
Z
i
, (6.8)
then it follows that
n
¸
i=1
φ
i
dZ
i
=
n
¸
i=1
φ
i
Z
i
σ
i
d
˜
W = (σ
−1
)
T
˜
ψ σ d
˜
W =
˜
ψ σ
−1
σ d
˜
W =
˜
ψ d
˜
W = dY.
Hence φ yields an admissible strategy with discounted value process satisfying
˜
V (T; φ) = Y (T) = E
Q
[X/S
0
(T)] +
n
¸
i=1
T
0
φ
i
(s) dZ
i
(s) = X/S
0
(T). (6.9)
Hence nondegeneracy of σ (see (6.6) and (6.8)) implies uniqueness of Q
and completeness of the model. These conditions are in fact equivalent (see
for example MR[19](Chapter 10)).
Theorem 6.3.2 (Completeness). The following are equivalent:
1. the model is complete;
2. σ is nondegenerate, see (6.6);
3. there exists a unique ELMM Q.
Theorem 6.3.3 (Representation Theorem). Every Plocal martingale
M has a continuous version and there exists ψ ∈ L such that
M(t) = M(0) +
t
0
ψ(s) dW(s).
(This theorem requires the ﬁltration (T
t
) to be generated by W.)
Pricing In the above complete model the fair price prevailing at t ≤ T of
a Tclaim X which satisﬁes (6.7) is given by (6.9)
V (t, φ) = S
0
(t)
˜
V (t; φ) = S
0
(t)E
Q
[X/S
0
(T) [ T
t
] . (6.10)
We shall often encounter complete models. However, models can be gener
ically incomplete (as real markets are), and then the pricing becomes a dif
ﬁcult issue. The literature on incomplete markets is huge, and the topic
beyond the scope of this course.
76 CHAPTER 6. NOARBITRAGE PRICING
Stateprice Density It is a custom (e.g. for short rate models) to exoge
nously specify a particular ELMM Q (or equivalently, the market price of
risk) and then price a Tclaim X satisfying (6.7) according to (6.10)
price of X at t =: Y (t) = S
0
(t)E
Q
[X/S
0
(T) [ T
t
] .
This is a consistent pricing rule in the sense that the enlarged market
Y, S
0
, . . . , S
n
is still arbitragefree (why?).
Now deﬁne
π(t) :=
1
S
0
(t)
dQ
dP
[
Tt
.
By Bayes formula we then have
Y (t) = S
0
(t)E
Q
[X/S
0
(T) [ T
t
] = S
0
(t)
E
X
S
0
(T)
dQ
dP
[
T
T
[ T
t
dQ
dP
[
Tt
=
E[Xπ(T) [ T
t
]
π(t)
,
and, in particular, for the price at t = 0
Y (0) = E[Xπ(T)].
This is why π is called the stateprice density process.
The price of a Tbond for example is (if 1/S
0
(T) ∈ L
1
(Q), → exercise)
P(t, T) = E
¸
π(T)
π(t)
[ T
t
= E
Q
¸
S
0
(t)
S
0
(T)
[ T
t
.
Also one can check (→ exercise) that if Q is an EMM then
S
i
π are Pmartingales.
Chapter 7
Short Rate Models
→ B[3](Chapters 16–17), MR[19](Chapter 12), etc
7.1 Generalities
Short rate models are the classical interest rate models. As in the last sec
tion we ﬁx a stochastic basis (Ω, T, P), where P is considered as objective
probability measure. The ﬁltration (T
t
)
t≥0
is generated by a ddimensional
Brownian motion W.
We assume that
• the short rates follow an Itˆ o process
dr(t) = b(t) dt +σ(t) dW(t)
determining the savings account B(t) = exp
t
0
r(s) ds
,
• all zerocoupon bond prices (P(t, T))
t∈[0,T]
are adapted processes (with
P(T, T) = 1 as usual),
• noarbitrage: there exists an EMM Q, such that
P(t, T)
B(t)
, t ∈ [0, T],
is a Qmartingale for all T > 0.
77
78 CHAPTER 7. SHORT RATE MODELS
According to the last chapter, the existence of an ELMM for all Tbonds
excludes arbitrage among every ﬁnite selection of zerocoupon bonds, say
P(t, T
1
), . . . , P(t, T
n
). To be more general one would have to consider strate
gies involving a continuum of bonds. This can be done (see [4] or Mike
Tehranchi’s PhD thesis 2002) but is beyond the scope of this course.
For convenience we require Q to be an EMM (and not merely an ELMM)
because then we have
P(t, T) = E
Q
e
−
T
t
r(s) ds
[ T
t
(7.1)
(compare this to the last section). Let −γ denote the corresponding market
price of risk
c
t
(γ W) =
dQ
dP
[
Tt
and
˜
W = W −
γ dt the implied QBrownian motion.
Proposition 7.1.1. Under the above assumptions, the process r satisﬁes
under Q
dr(t) = (b(t) +σ(t) γ(t)) dt +σ(t) d
˜
W(t). (7.2)
Moreover, for any T > 0 there exists an adapted R
d
valued process σ
γ
(t, T),
t ∈ [0, T], such that
dP(t, T)
P(t, T)
= r(t) dt +σ
γ
(t, T) d
˜
W(t) (7.3)
and hence
P(t, T)
B(t)
= P(0, T)c
t
σ
γ
˜
W
.
Proof. Exercise (proceed as in the Completeness Lemma 6.3.1).
It follows from (7.3) that the Tbond price satisﬁes under the objective
probability measure P
dP(t, T)
P(t, T)
= (r(t) −γ(t) σ
γ
(t, T)) dt +σ
γ
dW(t).
This illustrates again the role of the market price of risk −γ as the excess of
instantaneous return over r(t) in units of volatility.
7.2. DIFFUSION SHORT RATE MODELS 79
In a general equilibrium framework, the market price of risk is given
endogenously (as it is carried out in the seminal paper by Cox, Ingersoll and
Ross (85) [7]). Since our arguments refer only to the absence of arbitrage
between primary securities (bonds) and derivatives, we are unable to identify
the market price of risk. In other words, we started by specifying the P
dynamics of the short rates, and hence the savings account B(t). However,
the savings account alone cannot be used to replicate bond payoﬀs: the
model is incomplete. According to the Completeness Theorem 6.3.2, this is
also reﬂected by the nonuniqueness of the EMM (the market price of risk).
A priori, Q can be any equivalent probability measure Q ∼ P.
A short rate model is not fully determined without the exogenous
speciﬁcation of the market price of risk.
It is custom (and we follow this tradition) to postulate the Qdynamics
(Q being the EMM) of r which implies the Qdynamics of all bond prices
by (7.1), see also (7.3). All contingent claims can be priced by taking Q
expectations of their discounted payoﬀs. The market price of risk (and hence
the objective measure P) can be inferred by statistical methods from histor
ical observations of price movements.
7.2 Diﬀusion Short Rate Models
We ﬁx a stochastic basis (Ω, T, (T
t
)
t≥0
, Q), where now Q is considered as
martingale measure. We let W denote a ddimensional (Q, T
t
)Brownian
motion.
Let Z ⊂ R be a closed interval, and b and σ continuous functions on
R
+
Z. We assume that for any ρ ∈ Z the stochastic diﬀerential equation
(SDE)
dr(t) = b(t, r(t)) dt +σ(t, r(t)) dW(t) (7.4)
admits a unique Zvalued solution r = r
ρ
with
r(t) = ρ +
t
0
b(u, r(u)) du +
t
0
σ(u, r(u)) dW(u)
and such that
exp
−
T
t
r(u) du
∈ L
1
(Q) (7.5)
80 CHAPTER 7. SHORT RATE MODELS
for all 0 ≤ t ≤ T. Notice that (7.5) is always satisﬁed if Z ⊂ R
+
.
Suﬃcient for the existence and uniqueness is Lipschitz continuity of b(t, r)
and σ(t, r) in r, uniformly in t. If d = 1 then H¨ older continuity of order 1/2
of σ in r, uniformly in t, is enough. A good reference for SDEs is the book
of Karatzas and Shreve [14] on Brownian motion and stochastic calculus.
Condition (7.5) allows us to deﬁne the Tbond prices
P(t, T) = E
Q
¸
exp
−
T
t
r(u) du
[ T
t
.
It turns out that P(t, T) can be written as a function of r(t), t and T. This is
a general property of certain functionals of Markov process, usually referred
to as Feynman–Kac formula. In the following we write
a(t, r) :=
σ(t, r)
2
2
for the diﬀusion term of r(t).
Lemma 7.2.1. Let T > 0 and Φ be a continuous function on Z, and assume
that F = F(t, r) ∈ C
1,2
([0, T]Z) is a solution to the boundary value problem
on [0, T] Z
∂
t
F(t, r) +b(t, r)∂
r
F(t, r) +a(t, r)∂
2
r
F(t, r) −rF(t, r) = 0
F(T, r) = Φ(r).
(7.6)
Then
M(t) = F(t, r(t))e
−
t
0
r(u) du
, t ∈ [0, T],
is a local martingale. If in addition either
1. ∂
r
F(t, r(t))e
−
t
0
r(u) du
σ(t, r(t)) ∈ L
2
[0, T], or
2. M is uniformly bounded,
then M is a true martingale, and
F(t, r(t)) = E
Q
¸
exp
−
T
t
r(u) du
Φ(r(T)) [ T
t
, t ≤ T. (7.7)
7.2. DIFFUSION SHORT RATE MODELS 81
Proof. We can apply Itˆ o’s formula to M and obtain
dM(t) =
∂
t
F(t, r(t)) +b(t, r(t))∂
r
F(t, r(t))
+a(t, r)∂
2
r
F(t, r(t)) −r(t)F(t, r(t))
e
−
t
0
r(u) du
dt
+∂
r
F(t, r(t))e
−
t
0
r(u) du
σ(t, r(t)) dW(t)
= ∂
r
F(t, r(t))e
−
t
0
r(u) du
σ(t, r(t)) dW(t).
Hence M is a local martingale.
It is now clear that either Condition 1 or 2 imply that M is a true mar
tingale. Since
M(T) = Φ(r(T))e
−
T
0
r(u) du
we get
F(t, r(t))e
−
t
0
r(u) du
= M(t) = E
Q
¸
exp
−
T
0
r(u) du
Φ(r(T)) [ T
t
.
Multiplying with e
t
0
r(u) du
yields the claim.
We call (7.6) the term structure equation for Φ. Its solution F gives the
price of the Tclaim Φ(r(T)). In particular, for Φ ≡ 1 we get the Tbond
price P(t, T) as a function of t, r(t) (and T)
P(t, T) = F(t, r(t); T).
Remark 7.2.2. Strictly speaking, we have only shown that if a smooth solu
tion F of (7.6) exists and satisﬁes some additional properties (Condition 1
or 2) then the time t price of the claim Φ(r(T)) (which is the right hand side
of (7.7)) equals F(t, r(t)). One can also show the converse that the expecta
tion on the right hand side of (7.7) conditional on r(t) = r can be written as
F(t, r) where F solves the term structure equation (7.6) but usually only in a
weak sense, which in particular means that F may not be in C
1,2
([0, T] Z).
This is general Markov theory and we will not prove this here.
In any case, we have found a pricing algorithm. Is it computationally
eﬃcient? Solving PDEs numerically in more than three dimensions causes
diﬃculties. PDEs in less than three space dimensions are numerically feasi
ble, and the dimension of Z is one. The nuisance is that we have to solve a
82 CHAPTER 7. SHORT RATE MODELS
PDE for every single zerocoupon bond price function F(, ; T), T > 0. From
that we might want to derive the yield or even forward curve. If we do not
impose further structural assumptions we may run into regularity problems.
Hence
short rate models that admit closed form solutions to the term
structure equation (7.6), at least for Φ ≡ 1, are favorable.
7.2.1 Examples
This is a (far from complete) list of the most popular short rate models. For
all examples we have d = 1. If not otherwise stated, the parameters are
realvalued.
1. Vasicek (1977): Z = R,
dr(t) = (b +βr(t)) dt +σ dW(t),
2. Cox–Ingersoll–Ross (CIR, 1985): Z = R
+
, b ≥ 0,
dr(t) = (b +βr(t)) dt +σ
r(t) dW(t),
3. Dothan (1978): Z = R
+
,
dr(t) = βr(t) dt +σr(t) dW(t),
4. Black–Derman–Toy (1990): Z = R
+
,
dr(t) = β(t)r(t) dt +σ(t)r(t) dW(t),
5. Black–Karasinski (1991): Z = R
+
, (t) = log r(t),
d(t) = (b(t) +β(t)(t)) dt +σ(t) dW(t),
6. Ho–Lee (1986): Z = R,
dr(t) = b(t) dt +σ dW(t),
7. Hull–White (extended Vasicek, 1990): Z = R,
dr(t) = (b(t) +β(t)r(t)) dt +σ(t) dW(t),
8. Hull–White (extended CIR, 1990): Z = R
+
, b(t) ≥ 0,
dr(t) = (b(t) +β(t)r(t)) dt +σ(t)
r(t) dW(t).
7.3. INVERTING THE YIELD CURVE 83
7.3 Inverting the Yield Curve
Once the short rate model is chosen, the initial term structure
T →P(0, T) = F(0, r(0); T)
and hence the initial yield and forward curve are fully speciﬁed by the term
structure equation (7.6).
Conversely, one may want to invert the term structure equation (7.6) to
match a given initial yield curve. Say we have chosen the Vasicek model.
Then the implied Tbond price is a function of the current short rate level
and the three model parameters b, β and σ
P(0, T) = F(0, r(0); T, b, β, σ).
But F(0, r(0); T, b, β, σ) is just a parametrized curve family with three degrees
of freedom. It turns out that it is often too restrictive and will provide a poor
ﬁt of the current data in terms of accuracy (least squares criterion).
Therefore the class of timeinhomogeneous short rate models (such as the
Hull–White extensions) was introduced. By letting the parameters depend on
time one gains inﬁnite degree of freedom and hence a perfect ﬁt of any given
curve. Usually, the functions b(t) etc are fully determined by the empirical
initial yield curve.
7.4 Aﬃne Term Structures
Short rate models that admit closed form expressions for the implied bond
prices F(t, r; T) are favorable.
The most tractable models are those where bond prices are of the form
F(t, r; T) = exp(−A(t, T) −B(t, T)r),
for some smooth functions A and B. Such models are said to provide an
aﬃne term structure (ATS). Notice that F(T, r; T) = 1 implies
A(T, T) = B(T, T) = 0.
The nice thing about ATS models is that they can be completely character
ized.
84 CHAPTER 7. SHORT RATE MODELS
Proposition 7.4.1. The short rate model (7.4) provides an ATS only if its
diﬀusion and drift terms are of the form
a(t, r) = a(t) +α(t)r and b(t, r) = b(t) +β(t)r, (7.8)
for some continuous functions a, α, b, β. The functions A and B in turn
satisfy the system
∂
t
A(t, T) = a(t)B
2
(t, T) −b(t)B(t, T), A(T, T) = 0, (7.9)
∂
t
B(t, T) = α(t)B
2
(t, T) −β(t)B(t, T) −1, B(T, T) = 0. (7.10)
Proof. We insert F(t, r; T) = exp(−A(t, T) −B(t, T)r) in the term structure
equation (7.6) and obtain
a(t, r)B
2
(t, T) −b(t, r)B(t, T) = ∂
t
A(t, T) + (∂
t
B(t, T) + 1)r. (7.11)
The functions B(t, ) and B
2
(t, ) are linearly independent since otherwise
B(t, ) ≡ B(t, t) = 0, which trivially would lead to be above results with
a(t) = α(t) ≡ 0. Hence we can ﬁnd T
1
> T
2
> t such that the matrix
B
2
(t, T
1
) −B(t, T
1
)
B
2
(t, T
2
) −B(t, T
2
)
is invertible. Hence we can solve (7.11) for a(t, r) and b(t, r), which yields
(7.8). Replace a(t, r) and b(t, r) by (7.8), so the left hand side of (7.11) reads
a(t)B
2
(t, T) −b(t)B(t, T) +
α(t)B
2
(t, T) −β(t)B(t, T)
r.
Terms containing r must match. This proves the claim.
The functions a, α, b, β in (7.8) can be further speciﬁed. They have to be
such that a(t, r) ≥ 0 and r(t) does not leave the state space Z. In fact, it can
be shown that every ATS model can be transformed via aﬃne transformation
into one of the two cases
1. Z = R: necessarily α(t) = 0 and a(t) ≥ 0, and b, β are arbitrary. This
is the (Hull–White extension of the) Vasicek model.
2. Z = R
+
: necessarily a(t) = 0, α(t) ≥ 0 and b(t) ≥ 0 (otherwise the
process would cross zero), and β is arbitrary. This is the (Hull–White
extension of the) CIR model.
Looking at the list in Section 7.2.1 we see that all short rate models except
the Dothan, Black–Derman–Toy and Black–Karasinski models have an ATS.
7.5. SOME STANDARD MODELS 85
7.5 Some Standard Models
We discuss some of the most common short rate models.
→ B[3](Section 17.4), BM[6](Chapter 3)
7.5.1 Vasicek Model
The solution to
dr = (b +βr) dt +σ dW
is explicitly given by (→ exercise)
r(t) = r(0)e
βt
+
b
β
e
βt
−1
+σe
βt
t
0
e
−βs
dW(s).
It follows that r(t) is a Gaussian process with mean
E[r(t)] = r(0)e
βt
+
b
β
e
βt
−1
and variance
V ar[r(t)] = σ
2
e
2βt
t
0
e
−2βs
ds =
σ
2
2β
e
2βt
−1
.
Hence
Q[r(t) < 0] > 0,
which is not satisfactory (although this probability is usually very small).
Vasicek assumed the market price of risk to be constant, so that also the
objective Pdynamics of r(t) is of the above form.
If β < 0 then r(t) is meanreverting with mean reversion level b/[β[, see
Figure 7.1, and r(t) converges to a Gaussian random variable with mean
b/[β[ and variance σ
2
/(2[β[), for t →∞.
Equations (7.9)–(7.10) become
∂
t
A(t, T) =
σ
2
2
B
2
(t, T) −bB(t, T), A(T, T) = 0,
∂
t
B(t, T) = −βB(t, T) −1, B(T, T) = 0.
The explicit solution is
B(t, T) =
1
β
e
β(T−t)
−1
86 CHAPTER 7. SHORT RATE MODELS
Figure 7.1: Vasicek short rate process for β = −0.86, b/[β[ = 0.09 (mean
reversion level), σ = 0.0148 and r(0) = 0.08.
50 100 150 200 250 300 350
Time in Months
0.07
0.08
0.09
0.1
0.11
0.12
Short Rates
and A is given as ordinary integral
A(t, T) = A(T, T) −
T
t
∂
s
A(s, T) ds
= −
σ
2
2
T
t
B
2
(s, T) ds +b
T
t
B(s, T) ds
=
σ
2
4e
β(T−t)
−e
2β(T−t)
−2β(T −t) −3
4β
3
+b
e
β(T−t)
−1 −β(T −t)
β
2
.
We recall that zerocoupon bond prices are given in closed form by
P(t, T) = exp (−A(t, T) −B(t, T)r(t)) .
It is possible to derive closed form expression also for bond options (see
Section 7.6).
7.5.2 Cox–Ingersoll–Ross Model
It is worth to mention that, for b ≥ 0,
dr(t) = (b +βr(t)) dt +σ
r(t) dW(t), r(0) ≥ 0,
7.5. SOME STANDARD MODELS 87
has a unique strong solution r ≥ 0, for every r(0) ≥ 0. This also holds when
the coeﬃcients depend continuously on t, as it is the case for the Hull–White
extension. Even more, if b ≥ σ
2
/2 then r > 0 whenever r(0) > 0.
The ATS equation (7.10) now becomes nonlinear
∂
t
B(t, T) =
σ
2
2
B
2
(t, T) −βB(t, T) −1, B(T, T) = 0.
This is called a Riccati equation. It is good news that the explicit solution is
known
B(t, T) =
2
e
γ(T−t)
−1
(γ −β) (e
γ(T−t)
−1) + 2γ
where γ :=
β
2
+ 2σ
2
. Integration yields
A(t, T) = −
2b
σ
2
log
2γe
(γ−β)(T−t)/2
(γ −β) (e
γ(T−t)
−1) + 2γ
.
Hence also in the CIR model we have closed form expressions for the bond
prices. Moreover, it can be shown that also bond option prices are explicit(!)
Together with the fact that it yields positive interest rates, this is mainly the
reason why the CIR model is so popular.
7.5.3 Dothan Model
Dothan (78) starts from a driftless geometric Brownian motion under the
objective probability measure P
dr(t) = σr(t) dW
P
(t).
The market price of risk is chosen to be constant, which yields
dr(t) = βr(t) dt +σr(t) dW(t)
as Qdynamics. This is easily integrated
r(t) = r(s) exp
β −σ
2
/2
(t −s) +σ(W(t) −W(s))
, s ≤ t.
Thus the T
s
conditional distribution of r(t) is lognormal with mean and
variance (→ exercise)
E[r(t) [ T
s
] = r(s)e
β(t−s)
V ar[r(t) [ T
s
] = r
2
(s)e
2β(t−s)
e
σ
2
(t−s)
−1
.
88 CHAPTER 7. SHORT RATE MODELS
The Dothan and all lognormal short rate models (Black–Derman–Toy and
Black–Karasinski) yield positive interest rates. But no closed form expres
sions for bond prices or options are available (with one exception: Dothan
admits an “semiexplicit” expression for the bond prices, see BM[6]).
A major drawback of lognormal models is the explosion of the bank ac
count. Let ∆t be small, then
E[B(∆t)] = E
¸
exp
∆t
0
r(s) ds
≈ E
¸
exp
r(0) +r(∆t)
2
∆t
.
We face an expectation of the type
E[exp(exp(Y ))]
where Y is Gaussian distributed. But such an expectation is inﬁnite. This
means that in arbitrarily small time the bank account growths to inﬁnity in
average. Similarly, one shows that the price of a Eurodollar future is inﬁnite
for all lognormal models.
The idea of lognormal rates is taken up later by Sandmann and Son
dermann (1997) and many others, which ﬁnally led to the so called market
models with lognormal LIBOR or swap rates.
7.5.4 Ho–Lee Model
For the Ho–Lee model
dr(t) = b(t) dt +σ dW(t)
the ATS equations (7.9)–(7.10) become
∂
t
A(t, T) =
σ
2
2
B
2
(t, T) −b(t)B(t, T), A(T, T) = 0,
∂
t
B(t, T) = −1, B(T, T) = 0.
Hence
B(t, T) = T −t,
A(t, T) = −
σ
2
6
(T −t)
3
+
T
t
b(s)(T −s) ds.
7.5. SOME STANDARD MODELS 89
The forward curve is thus
f(t, T) = ∂
T
A(t, T) +∂
T
B(t, T)r(t) = −
σ
2
2
(T −t)
2
+
T
t
b(s) ds +r(t).
Let f
∗
(0, T) be the observed (estimated) initial forward curve. Then
b(s) = ∂
s
f
∗
(0, s) +σ
2
s.
gives a perfect ﬁt of f
∗
(0, T). Plugging this back into the ATS yields
f(t, T) = f
∗
(0, T) −f
∗
(0, t) +σ
2
t(T −t) +r(t).
We can also integrate this expression to get
P(t, T) = e
−
T
t
f
∗
(0,s) ds+f
∗
(0,t)(T−t)−
σ
2
2
t(T−t)
2
−(T−t)r(t)
.
It is interesting to see that
r(t) = r(0) +
t
0
b(s) ds +σW(t) = f
∗
(0, t) +
σ
2
t
2
2
+σW(t).
That is, r(t) ﬂuctuates along the modiﬁed initial forward curve, and we have
f
∗
(0, t) = E[r(t)] −
σ
2
t
2
2
.
7.5.5 Hull–White Model
The Hull–White (1990) extensions of Vasicek and CIR can be ﬁtted to the
initial yield and volatility curve. However, this ﬂexibility has its price: the
model cannot be handled analytically in general. We therefore restrict ourself
to the following extension of the Vasicek model that was analyzed by Hull
and White 1994
dr(t) = (b(t) +βr(t)) dt +σ dW(t).
In this model we choose the constants β and σ to obtain a nice volatility
structure whereas b(t) is chosen in order to match the initial yield curve.
Equation (7.10) for B(t, T) is just as in the Vasicek model
∂
t
B(t, T) = −βB(t, T) −1, B(T, T) = 0
90 CHAPTER 7. SHORT RATE MODELS
with explicit solution
B(t, T) =
1
β
e
β(T−t)
−1
.
Equation (7.9) for A(t, T) now reads
A(t, T) = −
σ
2
2
T
t
B
2
(s, T) ds +
T
t
b(s)B(s, T) ds
We consider the initial forward curve (notice that ∂
T
B(s, T) = −∂
s
B(s, T))
f
∗
(0, T) = ∂
T
A(0, T) +∂
T
B(0, T)r(0)
=
σ
2
2
T
0
∂
s
B
2
(s, T) ds +
T
0
b(s)∂
T
B(s, T) +∂
T
B(0, T)r(0)
= −
σ
2
2β
2
e
βT
−1
2
. .. .
=:g(T)
+
T
0
b(s)e
β(T−s)
ds +e
βT
r(0)
. .. .
=:φ(T)
.
The function φ satisﬁes
∂
T
φ(T) = βφ(T) +b(T), φ(0) = r(0).
It follows that
b(T) = ∂
T
φ(T) −βφ(T)
= ∂
T
(f
∗
(0, T) +g(T)) −β(f
∗
(0, T) +g(T)).
Plugging in and performing performing some calculations eventually yields
f(t, T) = f
∗
(0, T) −e
β(T−t)
f
∗
(0, t) −
σ
2
2β
2
e
β(T−t)
−1
e
β(T−t)
−e
β(T+t)
+e
β(T−t)
r(t).
7.6 Option Pricing in Aﬃne Models
We show how to price bond options in the aﬃne framework. The discussion
is informal, we do not worry about integrability conditions. The procedure
has to be carried out rigorously from case to case.
7.6. OPTION PRICING IN AFFINE MODELS 91
Let r(t) be a diﬀusion short rate model with drift
b(t) +β(t)r,
diﬀusion term
a(t) +α(t)r
and ATS
P(t, T) = e
−A(t,T)−B(t,T)r(t)
.
Let λ ∈ C, and φ and ψ be given as solutions to
∂
t
φ(t, T, λ) = a(t)ψ
2
(t, T, λ) −b(t)ψ(t, T, λ)
φ(T, T, λ) = 0
∂
t
ψ(t, T, λ) = α(t)ψ
2
(t, T, λ) −β(t)ψ(t, T, λ) −1
ψ(T, T, λ) = λ.
This looks much like the ATS equations (7.9)–(7.10), and indeed, by plugging
the right hand side below in the term structure equation (7.6), one sees that
E
e
−
T
t
r(s) ds
e
−λr(T)
[ T
t
= e
−φ(t,T,λ)−ψ(t,T,λ)r(t)
.
In fact, we have
φ(t, T, 0) = A(t, T) and ψ(t, T, 0) = B(t, T).
Now let t = 0 (for simplicity only). Since discounted zerocoupon bond prices
are martingales we obtain for T ≤ S (→ exercise)
E
e
−
S
0
r(s) ds
e
−λr(T)
= E
e
−
T
0
r(s) ds
e
−A(T,S)−B(T,S)r(T)
e
−λr(T)
= e
−A(T,S)
E
e
−
T
0
r(s) ds
e
−(λ+B(T,S))r(T)
= e
−A(T,S)−φ(0,T,λ+B(T,S))−ψ(0,T,λ+B(T,S))r(0)
.
But
dQ
S
dQ
=
e
−
S
0
r(s) ds
P(0, S)
deﬁnes an equivalent probability measure Q
S
∼ Q on T
S
, the so called S
forward measure. Hence we have shown that the (extended) Laplace trans
form of r(T) with respect to Q
S
is
E
Q
S
e
−λr(T)
= e
A(0,S)−A(T,S)−φ(0,T,λ+B(T,S))+(B(0,S)−ψ(0,T,λ+B(T,S)))r(0)
.
92 CHAPTER 7. SHORT RATE MODELS
By Laplace (or Fourier) inversion, one gets the distribution of r(T) under
Q
S
. In some cases (e.g. Vasicek or CIR) this distribution is explicitly known
(e.g. Gaussian or chisquare). In general, this is done numerically.
We now consider a European call option on a Sbond with expiry date
T < S and strike price K. Its price today (t = 0) is
π = E
e
−
T
0
r(s) ds
e
−A(T,S)−B(T,S)r(T)
−K
+
.
The payoﬀ can be decomposed according to
e
−A(T,S)−B(T,S)r(T)
−K
+
= e
−A(T,S)−B(T,S)r(T)
1
r(T)≤r
∗
¦
−K1
r(T)≤r
∗
¦
where
r
∗
= r
∗
(T, S, K) := −
A(T, S) + log K
B(T, S)
.
Hence
π = E
e
−
S
0
r(s) ds
1
r(T)≤r
∗
¦
−KE
e
−
T
0
r(s) ds
1
r(T)≤r
∗
¦
= P(0, S)Q
S
[r(T) ≤ r
∗
] −KP(0, T)Q
T
[r(T) ≤ r
∗
].
The pricing of the option boils down to the computation of the probability
of the event ¦r(T) ≤ r
∗
¦ under the S and Tforward measure.
7.6.1 Example: Vasicek Model (a, b, β const, α = 0).
We obtain (→ exercise)
π = P(0, S)Φ
r
∗
−
1
(T, S, r(0))
2
(T)
−KP(0, T)Φ
r
∗
−
1
(T, T, r(0))
2
(T)
where
1
(T, S, r) :=
1
β
2
β
e
βT
(b +βr) −b
−a
2 −e
β(S−T)
−2e
βT
+e
β(S+T)
2
(T) :=
a
β
e
2βT
−1
and Φ(x) is the cumulative standard Gaussian distribution function.
A similar closed form expression is available for the price of a put option,
and hence an explicit price formula for caps. For β = −0.86, b/[β[ = 0.09
7.6. OPTION PRICING IN AFFINE MODELS 93
(mean reversion level), σ = 0.0148 and r(0) = 0.08, as in Figure 7.1, one gets
the ATM cap prices and Black volatilities shown in Table 7.1 and Figure 7.2
(→ exercise). In contrast to Figure 2.1, the Vasicek model cannot produce
humped volatility curves.
Table 7.1: Vasicek ATM cap prices and Black volatilities.
Maturity ATM prices ATM vols
1 0.00215686 0.129734
2 0.00567477 0.106348
3 0.00907115 0.0915455
4 0.0121906 0.0815358
5 0.01503 0.0743607
6 0.017613 0.0689651
7 0.0199647 0.0647515
8 0.0221081 0.0613624
10 0.025847 0.0562337
12 0.028963 0.0525296
15 0.0326962 0.0485755
20 0.0370565 0.0443967
30 0.0416089 0.0402203
Figure 7.2: Vasicek ATM cap Black volatilities.
5 10 15 20 25 30
0.04
0.06
0.08
0.1
0.12
0.14
94 CHAPTER 7. SHORT RATE MODELS
Chapter 8
Heath–Jarrow–Morton (HJM)
Methodology
→ original article by Heath, Jarrow and Morton (HJM, 1992) [9].
95
96 CHAPTER 8. HJM METHODOLOGY
Chapter 9
Forward Measures
We consider the HJM setup (Chapter 8) and directly focus on the (unique)
EMM Q ∼ P under which all discounted bond price processes
P(t, T)
B(t)
, t ∈ [0, T],
are strictly positive martingales.
9.1 TBond as Numeraire
Fix T > 0. Since
1
P(0, T)B(T)
> 0 and E
Q
¸
1
P(0, T)B(T)
= 1
we can deﬁne an equivalent probability measure Q
T
∼ Q on T
T
by
dQ
T
dQ
=
1
P(0, T)B(T)
.
For t ≤ T we have
dQ
T
dQ
[
Tt
= E
Q
¸
dQ
T
dQ
[ T
t
=
P(t, T)
P(0, T)B(t)
.
This probability measure has already been introduced in Section 7.6. It is
called the Tforward measure.
97
98 CHAPTER 9. FORWARD MEASURES
Lemma 9.1.1. For any S > 0,
P(t, S)
P(t, T)
, t ∈ [0, S ∧ T],
is a Q
T
martingale.
Proof. Let s ≤ t ≤ S ∧ T. Bayes’ rule gives
E
Q
T
¸
P(t, S)
P(t, T)
[ T
s
=
E
Q
P(t,T)
P(0,T)B(t)
P(t,S)
P(t,T)
[ T
s
P(s,T)
P(0,T)B(s)
=
P(s,S)
B(s)
P(s,T)
B(s)
=
P(s, S)
P(s, T)
.
We thus have an entire collection of EMMs now! Each Q
T
corresponds to
a diﬀerent numeraire, namely the Tbond. Since Q is related to the riskfree
asset, one usually calls Q the risk neutral measure.
Tforward measures give simpler pricing formulas. Indeed, let X be a
Tclaim such that
X
B(T)
∈ L
1
(Q, T
T
). (9.1)
Its fair price at time t ≤ T is then given by
π(t) = E
Q
e
−
T
t
r(s) ds
X [ T
t
.
To compute π(t) we have to know the joint distribution of exp
−
T
t
r(s) ds
and X, and integrate with respect to that distribution. Thus we have to
compute a double integral, which in most cases turns out to be rather hard
work. If B(T)/B(t) and X were independent under Q (which is not realistic!
it holds, for instance, if r is deterministic) we would have
π(t) = P(t, T)E
Q
[X [ T
t
] ,
a much nicer formula, since
• we only have to compute the single integral E
Q
[X [ T
t
];
9.2. AN EXPECTATION HYPOTHESIS 99
• the bond price P(t, T) can be observed at time t and does not have to
be computed.
The good news is that the above formula holds — not under Q though, but
under Q
T
:
Proposition 9.1.2. Let X be a Tclaim such that (9.1) holds. Then
E
Q
T [[X[] < ∞ (9.2)
and
π(t) = P(t, T)E
Q
T [X [ T
t
] . (9.3)
Proof. Bayes’s rule yields
E
Q
T [[X[] = E
Q
¸
[X[
P(0, T)B(T)
< ∞ (by (9.1)),
whence (9.2). And
π(t) = P(0, T)B(t)E
Q
¸
X
P(0, T)B(T)
[ T
t
= P(0, T)B(t)
P(t, T)
P(0, T)B(t)
E
Q
T [X [ T
t
]
= P(t, T)E
Q
T [X [ T
t
] ,
which proves (9.3).
9.2 An Expectation Hypothesis
Under the forward measure the expectation hypothesis holds. That is, the
expression of the forward rates f(t, T) as conditional expectation of the future
short rate r(T).
To see that, we write W for the driving QBrownian motion. The forward
rates then follow the dynamics
f(t, T) = f(0, T) +
t
0
σ(s, T)
T
s
σ(s, u) du
ds +
t
0
σ(s, T) dW(s).
(9.4)
100 CHAPTER 9. FORWARD MEASURES
The Qdynamics of the discounted bond price process is
P(t, T)
B(t)
= P(0, T) +
t
0
P(s, T)
B(s)
−
T
s
σ(s, u) du
dW(s). (9.5)
This equation has a unique solution
P(t, T)
B(t)
= P(0, T)c
t
−
T
σ(, u) du
W
.
We thus have
dQ
T
dQ
[
Tt
= c
t
−
T
σ(, u) du
W
. (9.6)
Girsanov’s theorem applies and
W
T
(t) = W(t) +
t
0
T
s
σ(s, u) du
ds, t ∈ [0, T],
is a Q
T
Brownian motion. Equation (9.4) now reads
f(t, T) = f(0, T) +
t
0
σ(s, T) dW
T
(s).
Hence, if
E
Q
T
¸
T
0
σ(s, T)
2
ds
< ∞
then
(f(t, T))
t∈[0,T]
is a Q
T
martingale.
Summarizing we have thus proved
Lemma 9.2.1. Under the above assumptions, the expectation hypothesis
holds under the forward measures
f(t, T) = E
Q
T [r(T) [ T
t
] .
9.3. OPTION PRICING IN GAUSSIAN HJM MODELS 101
9.3 Option Pricing in Gaussian HJM Models
We consider a European call option on an Sbond with expiry date T < S
and strike price K. Its price at time t = 0 (for simplicity only) is
π = E
Q
e
−
T
0
r(s) ds
(P(T, S) −K)
+
.
We proceed as in Section 7.6 and decompose
π = E
Q
B(T)
−1
P(T, S) 1(P(T, S) ≥ K)
−KE
Q
B(T)
−1
1(P(T, S) ≥ K)
= P(0, S)Q
S
[P(T, S) ≥ K] −KP(0, T)Q
T
[P(T, S) ≥ K] .
This option pricing formula holds in general.
We already know that
dP(t, T)
P(t, T)
= r(t) dt +v(t, T) dW(t)
and hence
P(t, T) = P(0, T) exp
¸
t
0
v(s, T) dW(s) +
t
0
r(s) −
1
2
v(s, T)
2
ds
where
v(t, T) := −
T
t
σ(t, u) du. (9.7)
We also know that
P(t,T)
P(t,S)
t∈[0,T]
is a Q
S
martingale and
P(t,S)
P(t,T)
t∈[0,T]
is a
Q
T
martingale. In fact (→ exercise)
P(t, T)
P(t, S)
=
P(0, T)
P(0, S)
exp
¸
t
0
σ
T,S
(s) dW(s) −
1
2
t
0
v(s, T)
2
−v(s, S)
2
ds
=
P(0, T)
P(0, S)
exp
¸
t
0
σ
T,S
(s) dW
S
(s) −
1
2
t
0
σ
T,S
(s)
2
ds
where
σ
T,S
(s) := v(s, T) −v(s, S) =
S
T
σ(s, u) du, (9.8)
102 CHAPTER 9. FORWARD MEASURES
and
P(t, S)
P(t, T)
=
P(0, S)
P(0, T)
exp
¸
−
t
0
σ
T,S
(s) dW(s) −
1
2
t
0
v(s, S)
2
−v(s, T)
2
ds
=
P(0, S)
P(0, T)
exp
¸
−
t
0
σ
T,S
(s) dW
T
(s) −
1
2
t
0
σ
T,S
(s)
2
ds
.
Now observe that
Q
S
[P(T, S) ≥ K] = Q
S
¸
P(T, T)
P(T, S)
≤
1
K
Q
T
[P(T, S) ≥ K] = Q
T
¸
P(T, S)
P(T, T)
≥ K
.
This suggests to look at those models for which σ
T,S
is deterministic, and
hence
P(T,T)
P(T,S)
and
P(T,S)
P(T,T)
are lognormally distributed under the respective
forward measures.
We thus assume that σ(t, T) = (σ
1
(t, T), . . . , σ
d
(t, T)) are determinis
tic functions of t and T, and hence forward rates f(t, T) are Gaussian dis
tributed.
We obtain the following closed form option price formula.
Proposition 9.3.1. Under the above Gaussian assumption, the option price
is
π = P(0, S)Φ[d
1
] −KP(0, T)Φ[d
2
],
where
d
1,2
=
log
P(0,S)
KP(0,T)
±
1
2
T
0
σ
T,S
(s)
2
ds
T
0
σ
T,S
(s)
2
ds
,
σ
T,S
(s) is given in (9.8) and Φ is the standard Gaussian CDF.
Proof. It is enough to observe that
log
P(T,T)
P(T,S)
−log
P(0,T)
P(0,S)
+
1
2
T
0
σ
T,S
(s)
2
ds
T
0
σ
T,S
(s)
2
ds
9.3. OPTION PRICING IN GAUSSIAN HJM MODELS 103
and
log
P(T,S)
P(T,T)
−log
P(0,S)
P(0,T)
+
1
2
T
0
σ
T,S
(s)
2
ds
T
0
σ
T,S
(s)
2
ds
are standard Gaussian distributed under Q
S
and Q
T
, respectively.
Of course, the Vasicek option price formula from Section 7.6.1 can now
be obtained as a corollary of Proposition 9.3.1 (→ exercise).
104 CHAPTER 9. FORWARD MEASURES
Chapter 10
Forwards and Futures
→ B[3](Chapter 20), or Hull (2002) [10]
We discuss two common types of term contracts: forwards, which are
mainly traded OTC, and futures, which are actively traded on many ex
changes.
The underlying is in both cases a Tclaim \, for some ﬁxed future date T.
This can be an exchange rate, an interest rate, a commodity such as copper,
any traded or nontraded asset, an index, etc.
10.1 Forward Contracts
A forward contract on \, contracted at t, with time of delivery T > t, and
with the forward price f(t; T, \) is deﬁned by the following payment scheme:
• at T, the holder of the contract (long position) pays f(t; T, \) and
receives \ from the underwriter (short position);
• at t, the forward price is chosen such that the present value of the
forward contract is zero, thus
E
Q
e
−
T
t
r(s) ds
(\ −f(t; T, \)) [ T
t
= 0.
This is equivalent to
f(t; T, \) =
1
P(t, T)
E
Q
e
−
T
t
r(s) ds
\ [ T
t
= E
Q
T [\ [ T
t
] .
105
106 CHAPTER 10. FORWARDS AND FUTURES
Examples The forward price at t of
1. a dollar delivered at T is 1;
2. an Sbond delivered at T ≤ S is
P(t,S)
P(t,T)
;
3. any traded asset S delivered at T is
S(t)
P(t,T)
.
The forward price f(s; T, \) has to be distinguished from the (spot) price
at time s of the forward contract entered at time t ≤ s, which is
E
Q
e
−
T
s
r(u) du
(\ −f(t; T, \)) [ T
s
= E
Q
e
−
T
s
r(u) du
\ [ T
s
−P(t, T)f(t; T, \).
10.2 Futures Contracts
A futures contract on \ with time of delivery T is deﬁned as follows:
• at every t ≤ T, there is a market quoted futures price F(t; T, \), which
makes the futures contract on \, if entered at t, equal to zero;
• at T, the holder of the contract (long position) pays F(T; T, \) and
receives \ from the underwriter (short position);
• during any time interval (s, t] the holder of the contract receives (or
pays, if negative) the amount F(t; T, \) − F(s; T, \) (this is called
marking to market).
So there is a continuous cashﬂow between the two parties of a futures con
tract. They are required to keep a certain amount of money as a safety
margin.
The volumes in which futures are traded are huge. One of the reasons
for this is that in many markets it is diﬃcult to trade (hedge) directly in the
underlying object. This might be an index which includes many diﬀerent
(illiquid) instruments, or a commodity such as copper, gas or electricity,
etc. Holding a (short position in a) futures does not force you to physically
deliver the underlying object (if you exit the contract before delivery date),
and selling short makes it possible to hedge against the underlying.
10.2. FUTURES CONTRACTS 107
Suppose \ ∈ L
1
(Q). Then the futures price process is given by the
Qmartingale
F(t; T, \) = E
Q
[\ [ T
t
] . (10.1)
Often, this is just how futures prices are deﬁned. We now give a heuristic
argument for (10.1) based on the above characterization of a futures contract.
First, our model economy is driven by Brownian motion and changes in
a continuous way. Hence there is no reason to believe that futures prices
evolve discontinuously, and we may assume that
F(t) = F(t; T, \) is a continuous semimartingale (or Itˆ o process).
Now suppose we enter the futures contract at time t < T. We face a con
tinuum of cashﬂows in the interval (t, T]. Indeed, let t = t
0
< < t
N
= t
be a partition of [t, T]. The present value of the corresponding cashﬂows
F(t
i
) −F(t
i−1
) at t
i
, i = 1, . . . , N, is given by E
Q
[Σ [ T
t
] where
Σ :=
N
¸
i=1
1
B(t
i
)
(F(t
i
) −F(t
i−1
)) .
But the futures contract has present value zero, hence
E
Q
[Σ [ T
t
] = 0.
This has to hold for any partition (t
i
). We can rewrite Σ as
N
¸
i=1
1
B(t
i−1
)
(F(t
i
) −F(t
i−1
)) +
N
¸
i=1
1
B(t
i
)
−
1
B(t
i−1
)
(F(t
i
) −F(t
i−1
)) .
If we let the partition become ﬁner and ﬁner this expression converges in
probability towards
T
t
1
B(s)
dF(s) +
T
t
d
1
B
, F
s
=
T
t
1
B(s)
dF(s),
since the quadratic variation of 1/B (ﬁnite variation) and F (continuous) is
zero. Under the appropriate integrability assumptions (uniform integrability)
we conclude that
E
Q
¸
T
t
1
B(s)
dF(s) [ T
t
= 0,
108 CHAPTER 10. FORWARDS AND FUTURES
and that
M(t) =
t
0
1
B(s)
dF(s) = E
Q
¸
T
0
1
B(s)
dF(s) [ T
t
, t ∈ [0, T],
is a Qmartingale. If, moreover
E
Q
¸
T
0
1
B(s)
2
d'M, M`
s
= E
Q
['F, F`
T
] < ∞
then
F(t) =
t
0
1
B(s)
dM(s), t ∈ [0, T],
is a Qmartingale, which implies (10.1).
10.3 Interest Rate Futures
→ Z[27](Section 5.4)
Interest rate futures contracts may be divided into futures on short term
instruments and futures on coupon bonds. We only consider an example
from the ﬁrst group.
Eurodollars are deposits of US dollars in institutions outside of the US.
LIBOR is the interbank rate of interest for Eurodollar loans. The Eurodollar
futures contract is tied to the LIBOR. It was introduced by the International
Money Market (IMM) of the Chicago Mercantile Exchange (CME) in 1981,
and is designed to protect its owner from ﬂuctuations in the 3months (=1/4
years) LIBOR. The maturity (delivery) months are March, June, September
and December.
Fix a maturity date T and let L(T) denote the 3months LIBOR for the
period [T, T + 1/4], prevailing at T. The market quote of the Eurodollar
futures contract on L(T) at time t ≤ T is
1 −L
F
(t, T) [100 per cent]
where L
F
(t, T) is the corresponding futures rate (compare with the example
in Section 4.2.2). As t tends to T, L
F
(t, T) tends to L(T). The futures price,
used for the marking to market, is deﬁned by
F(t; T, L(T)) = 1 −
1
4
L
F
(t, T) [Mio. dollars].
10.4. FORWARD VS. FUTURES IN A GAUSSIAN SETUP 109
Consequently, a change of 1 basis point (0.01%) in the futures rate L
F
(t, T)
leads to a cashﬂow of
10
6
10
−4
1
4
= 25 [dollars].
We also see that the ﬁnal price F(T; T, L(T)) = 1 −
1
4
L(T) = \ is not
P(T, T +1/4) = 1 −
1
4
L(T)P(T, T +1/4) as one might suppose. In fact, the
underlying \ is a synthetic value. At maturity there is no physical delivery.
Instead, settlement is made in cash.
On the other hand, since
1 −
1
4
L
F
(t, T) = F(t; T, L(T))
= E
Q
[F(T; T, L(T)) [ T
t
] = 1 −
1
4
E
Q
[L(T) [ T
t
] ,
we obtain an explicit formula for the futures rate
L
F
(t, T) = E
Q
[L(T) [ T
t
] .
10.4 Forward vs. Futures in a Gaussian Setup
Let S be the price process of a traded asset. Hence the Qdynamics of S is
of the form
dS(t)
S(t)
= r(t) dt +ρ(t) dW(t),
for some volatility process ρ. Fix a delivery date T. The forward and futures
prices of S for delivery at T are
f(t; T, S(T)) =
S(t)
P(t, T)
, F(t; T, S(T)) = E
Q
[S(T) [ T
t
].
Under Gaussian assumption we can establish the relationship between the
two prices.
Proposition 10.4.1. Suppose ρ(t) and v(t, T) are deterministic functions
in t, where
v(t, T) = −
T
t
σ(t, u) du
110 CHAPTER 10. FORWARDS AND FUTURES
is the volatility of the Tbond (see (9.7)). Then
F(t; T, S(T)) = f(t; T, S(T)) exp
T
t
(v(s, T) −ρ(s)) v(s, T) ds
for t ≤ T.
Hence, if the instantaneous correlation of dS(t) and dP(t, T) is negative
d'S, P(, T)`
t
dt
= S(t)P(t, T)ρ(t) v(t, T) ≤ 0
then the futures price dominates the forward price.
Proof. Write µ(s) := v(s, T) −ρ(s). It is clear that
f(t; T, S(T)) =
S(0)
P(0, T)
exp
−
t
0
µ(s) dW(s) −
1
2
t
0
µ(s)
2
ds
exp
t
0
µ(s) v(s, T) ds
,
and hence
f(T; T, S(T)) = f(t; T, S(T)) exp
−
T
t
µ(s) dW(s) −
1
2
T
t
µ(s)
2
ds
exp
T
t
µ(s) v(s, T) ds
.
By assumption µ(s) is deterministic. Consequently,
E
Q
¸
exp
−
T
t
µ(s) dW(s) −
1
2
T
t
µ(s)
2
ds
[ T
t
= 1
and
F(t; T, S(T)) = E
Q
[f(T; T, S(T)) [ T
t
]
= f(t; T, S(T)) exp
T
t
µ(s) v(s, T) ds
,
as desired.
10.4. FORWARD VS. FUTURES IN A GAUSSIAN SETUP 111
Similarly, one can show (→ exercise)
Lemma 10.4.2. In a Gaussian HJM framework (σ(t, T) deterministic) we
have the following relations (convexity adjustments) between instantaneous
and simple futures and forward rates
f(t, T) = E
Q
[r(T) [ T
t
] −
T
t
σ(s, T)
T
s
σ(s, u) du
ds,
F(t; T, S) = E
Q
[F(T, S) [ T
t
]
−
P(t, T)
(S −T)P(t, S)
e
T
t
(
S
T
σ(s,v) dv
S
s
σ(s,u) du)ds
−1
for t ≤ T < S.
Hence, if
σ(s, v) σ(s, u) ≥ 0 for all s ≤ min(u, v)
then futures rates are always greater than the corresponding forward rates.
112 CHAPTER 10. FORWARDS AND FUTURES
Chapter 11
MultiFactor Models
We have seen that every timehomogeneous diﬀusion short rate model r(t)
induces forward rates of the form
f(t, T) = H(T −t, r(t)),
for some deterministic function H. This a onefactor model, since the driving
(Markovian) factor, r(t), is onedimensional. This is too restrictive from two
points of view:
• statistically: the evolution of the entire yield curve is explained by
a single variable. The inﬁnitesimal increments of all bond prices are
perfectly correlated
d'P(, T), P(, S)`
t
d'P(, T), P(, T)`
t
d'P(, S), P(, S)`
t
=
T
t
σ(t, u)du
S
t
σ(t, u)du
T
t
σ(t, u)du
S
t
σ(t, u)du
= 1.
• analytically: the family of attainable forward curves
H = ¦H(, r) [ r ∈ R¦
is only onedimensional.
To gain more ﬂexibility, we now allow for multiple factors. Fix m ≥ 1
and a closed set Z ⊂ R
m
(state space). A (m)factor model is an interest
rate model of the form
f(t, T) = H(T −t, Z(t))
113
114 CHAPTER 11. MULTIFACTOR MODELS
where H is a deterministic function and Z (state process) is a Zvalued
diﬀusion process,
dZ(t) = b(Z(t)) dt +ρ(Z(t)) dW(t)
Z(0) = z
0
.
Here W is a ddimensional Brownian motion deﬁned on a ﬁltered probability
space (Ω, T, (T
t
), Q), satisfying the usual conditions. We assume that
(A1) H ∈ C
1,2
(R
+
Z);
(A2) b : Z →R
m
and ρ : Z →R
md
are continuous functions;
(A3) the above SDE has a unique Zvalued solution Z = Z
z
0
, for every
z
0
∈ Z;
(A4) Q is the risk neutral local martingale measure for the induced bond
prices
P(t, T) = Π(T −t, Z
z
0
(t)),
for all z
0
∈ Z, where
Π(x, z) := exp
−
x
0
H(s, z) ds
.
Notice that the short rates are now given by r(t) = H(0, Z(t)). Hence the
assumption (A4) is equivalent to
(A4’)
Π(T −t, Z
z
0
(t))
e
t
0
H(0,Z
z
0(s)) ds
t∈[0,T]
is a Qlocal martingale, for all z
0
∈ Z.
Timeinhomogeneous models are included in the above setup. Simply set
Z
1
(t) = t (that is, b
1
≡ 1 and ρ
1j
≡ 0 for j = 1, . . . , d).
11.1. NOARBITRAGE CONDITION 115
11.1 NoArbitrage Condition
Since the function (x, z) → H(x, z) is in C
1,2
(R
+
Z) we can apply Itˆ o’s
formula and obtain
df(t, T) =
−∂
x
H(T −t, Z(t)) +
m
¸
i=1
b
i
(Z(t))∂
z
i
H(T −t, Z(t))
+
1
2
m
¸
i,j=1
a
ij
(Z(t))∂
z
i
∂
z
j
H(T −t, Z(t))
dt
+
m
¸
i=1
d
¸
j=1
∂
z
i
H(T −t, Z(t))ρ
ij
(Z(t)) dW
j
(t),
where
a(z) := ρ(z)ρ
T
(z). (11.1)
Hence the induced forward rate model is of the HJM type with
σ
j
(t, T) =
m
¸
i=1
∂
z
i
H(T −t, Z(t))ρ
ij
(Z(t)), j = 1, . . . , d.
The HJM drift condition now reads
−∂
x
H(T −t, Z(t)) +
m
¸
i=1
b
i
(Z(t))∂
z
i
H(T −t, Z(t))
+
1
2
m
¸
i,j=1
a
ij
(Z(t))∂
z
i
∂
z
j
H(T −t, Z(t))
=
d
¸
j=1
m
¸
k,l=1
ρ
kj
(Z(t))ρ
lj
(Z(t))∂
z
i
H(T −t, Z(t))
T
t
∂
z
i
H(u −t, Z(t)) du
=
m
¸
k,l=1
a
kl
(Z(t))∂
z
i
H(T −t, Z(t))
T
t
∂
z
i
H(u −t, Z(t)) du.
This has to hold a.s. for all t ≤ T and initial points z
0
= Z(0). Letting t →0
we thus get the following result.
116 CHAPTER 11. MULTIFACTOR MODELS
Proposition 11.1.1 (Consistency Condition). Under the above assump
tions (A1)–(A3), there is equivalence between (A4) and
∂
x
H(x, z) =
m
¸
i=1
b
i
(z)∂
z
i
H(x, z)
+
m
¸
i,j=1
a
ij
(z)
1
2
∂
z
i
∂
z
j
H(x, z) −∂
z
i
H(x, z)
x
0
∂
z
i
H(u, z) du
(11.2)
for all (x, z) ∈ R
+
Z, where a is deﬁned in (11.1).
Remark 11.1.2. Notice that, by symmetry, the last expression in (11.2) can
be written as
m
¸
i,j=1
a
ij
(z)∂
z
i
H(x, z)
x
0
∂
z
i
H(u, z) du
=
1
2
∂
x
m
¸
i,j=1
a
ij
(z)
x
0
∂
z
i
H(u, z) du
x
0
∂
z
j
H(u, z) du
.
There are two ways to approach equation (11.2). First, one takes b and
ρ (and hence a) as given and looks for a solution H for the PDE (11.2). Or,
one takes H as given (an estimation method for the yield curve) and tries
to ﬁnd b and a such that (11.2) is satisﬁed for all (x, z). This is an inverse
problem. It turns out that the latter approach is quite restrictive on possible
choices of b and a.
Proposition 11.1.3. Suppose that the functions
∂
z
i
H(, z) and
1
2
∂
z
i
∂
z
j
H(, z) −∂
z
i
H(, z)
0
∂
z
i
H(u, z) du,
for 1 ≤ i ≤ j ≤ m, are linearly independent for all z in some dense subset
T ⊂ Z. Then b and a are uniquely determined by H.
Proof. Set M = m+m(m + 1)/2, the number of unknown functions b
k
and
a
kl
= a
lk
. Let z ∈ T. Then there exists a sequence 0 ≤ x
1
< < x
M
such
that the M Mmatrix with kth row vector built by
∂
z
i
H(x
k
, z) and
1
2
∂
z
i
∂
z
j
H(x
k
, z) −∂
z
i
H(, z)
x
k
0
∂
z
i
H(u, z) du,
11.2. AFFINE TERM STRUCTURES 117
for 1 ≤ i ≤ j ≤ m, is invertible. Thus, b(z) and a(z) are uniquely determined
by (11.2). This holds for each z ∈ T. By continuity of b and a hence for all
z ∈ Z.
Remark 11.1.4. Suppose that the the parametrized curve family
H = ¦H(, z) [ z ∈ Z¦
is used for daily estimation of the forward curve in terms of the state vari
able z. Then the above proposition tells us that, under the stated assumption,
any Qdiﬀusion model Z for z is fully determined by H.
If T
t
= T
W
t
is the Brownian ﬁltration, then the diﬀusion coeﬃcient, a(z),
of Z is not aﬀected by any Girsanov transformation. Consequently, statistical
calibration is only possible for the drift of the model (or equivalently, for
the market price of risk), since the observations of z are made under the
objective measure P ∼ Q, where dQ/dP is left unspeciﬁed by our consistency
considerations.
11.2 Aﬃne Term Structures
We ﬁrst look at the simplest, namely the aﬃne case:
H(x, z) = g
0
(x) +g
1
(x)z
1
+ g
m
(x)z
m
.
Here the second order zderivatives vanish, and (11.2) reduces to
∂
x
g
0
(x) +
m
¸
i=1
z
i
∂
x
g
i
(x) =
m
¸
i=1
b
i
(z)g
i
(x) −
1
2
∂
x
m
¸
i,j=1
a
ij
(z)G
i
(x)G
j
(x)
,
(11.3)
where
G
i
(x) :=
x
0
g
i
(u) du.
Integrating (11.3) yields
g
0
(x)−g
0
(0)+
m
¸
i=1
z
i
(g
i
(x)−g
i
(0)) =
m
¸
i=1
b
i
(z)G
i
(x)−
1
2
m
¸
i,j=1
a
ij
(z)G
i
(x)G
j
(x).
(11.4)
118 CHAPTER 11. MULTIFACTOR MODELS
Now if
G
1
, . . . , G
m
, G
1
G
1
, G
1
G
2
, . . . , G
m
G
m
are linearly independent functions, we can invert and solve the linear equation
(11.4) for b and a. Since the left hand side is aﬃne is z, we obtain that also
b and a are aﬃne
b
i
(z) = b
i
+
m
¸
j=1
β
ij
z
j
a
ij
(z) = a
ij
+
m
¸
k=1
α
k;ij
z
k
,
for some constant vectors and matrices b, β, a and α
k
. Plugging this back
into (11.4) and matching constant terms and terms containing z
k
s we obtain
a system of Riccati equations
∂
x
G
0
(x) = g
0
(0) +
m
¸
i=1
b
i
G
i
(x) −
1
2
m
¸
i,j=1
a
ij
G
i
(x)G
j
(x) (11.5)
∂
x
G
k
(x) = g
k
(0) +
m
¸
i=1
β
ki
G
i
(x) −
1
2
m
¸
i,j=1
α
k;ij
G
i
(x)G
j
(x), (11.6)
with initial conditions G
0
(0) = = G
m
(0) = 0. This extends what we have
found in Section 7.4 for the onefactor case.
Notice that we have the freedom to choose g
0
(0), . . . , g
m
(0), which are
related to the short rates by
r(t) = f(t, t) = g
0
(0) +g
1
(0)Z
1
(t) + +g
m
(0)Z
m
(t).
A typical choice is g
1
(0) = 1 and all the other g
i
(0) = 0, whence Z
1
(t) is the
(nonMarkovian) short rate process.
11.3 Polynomial Term Structures
We extend the ATS setup and consider polynomial term structures (PTS)
H(x, z) =
n
¸
[i[=0
g
i
(x) (Z
t
)
i
, (11.7)
11.3. POLYNOMIAL TERM STRUCTURES 119
where we use the multiindex notation i = (i
1
, . . . , i
m
), [i[ = i
1
+ +i
m
and
z
i
= z
i
1
1
z
im
m
. Here n denotes the degree of the PTS; that is, there exists
an index i with [i[ = n and g
i
= 0.
Thus for n = 1 we are back to the ATS case.
For n = 2 we have a quadratic term structure (QTS), which has also been
studied in the literature.
Do we gain something by looking at n = 3 and higher degree PTS models?
The answer is no. In fact, we now shall show the amazing result that n > 2
is not consistent with (11.2).
For µ ∈ ¦1, . . . , n¦ and k ∈ ¦1, . . . , m¦ we write (µ)
k
for the multi
index with µ at the kth position and zeros elsewhere. Let i
1
, i
2
, . . . , i
N
be a
numbering of the set of multiindices
I = ¦i = (i
1
, . . . , i
m
) [ [i[ ≤ n¦, where N := [I[ =
n
¸
[i[=0
1.
As above, we denote the integral of g
i
by
G
i
(x) :=
x
0
g
i
(u) du.
Theorem 11.3.1 (Maximal Degree Problem I). Suppose that G
iµ
and
G
iµ
G
iν
are linearly independent functions, 1 ≤ µ ≤ ν ≤ N, and that ρ ≡ 0.
Then necessarily n ∈ ¦1, 2¦. Moreover, b(z) and a(z) are polynomials in
z with deg b(z) ≤ 1 in any case (QTS and ATS), and deg a(z) = 0 if n = 2
(QTS) and deg a(z) ≤ 1 if n = 1 (ATS).
Proof. Deﬁne the functions
B
i
(z) := b
k
(z)
∂z
i
∂z
k
+
1
2
m
¸
k,l=1
a
kl
(z)
∂
2
z
i
∂z
k
∂z
l
(11.8)
A
ij
(z) = A
ji
(z) :=
1
2
m
¸
k,l=1
a
kl
(z)
∂z
i
∂z
k
∂z
j
∂z
l
. (11.9)
Equation (11.2) can be rewritten
N
¸
µ=1
g
iµ
(x) −g
iµ
(0)
z
iµ
=
N
¸
µ=1
G
iµ
(x)B
iµ
(z) −
N
¸
µ,ν=1
G
iµ
(x)G
iν
(x)A
iµiν
(z).
(11.10)
120 CHAPTER 11. MULTIFACTOR MODELS
By assumption we can solve this linear equation for B and A, and thus
B
i
(z) and A
ij
(z) are polynomials in z of order less than or equal n. In
particular, we have
B
(1)
k
(z) = b
k
(z),
2A
(1)
k
(1)
l
(z) = a
kl
(z), k, l ∈ ¦1, . . . , m¦, (11.11)
hence b(z) and a(z) are polynomials in z with deg b(z), deg a(z) ≤ n. An
easy calculation shows that
2A
(n)
k
(n)
k
(z) = a
kk
(z)n
2
z
2n−2
k
, k ∈ ¦1, . . . , m¦. (11.12)
We may assume that a
kk
≡ 0, since ρ ≡ 0. But then the right hand side
of (11.12) cannot be a polynomial in z of order less than or equal n unless
n ≤ 2. This proves the ﬁrst part of the theorem.
If n = 1 there is nothing more to prove. Now let n = 2. Notice that by
deﬁnition
deg
µ
a
kl
(z) ≤ (deg
µ
a
kk
(z) + deg
µ
a
ll
(z))/2,
where deg
µ
denotes the degree of dependence on the single component z
µ
.
Equation (11.12) yields deg
k
a
kk
(z) = 0. Hence deg
l
a
kl
(z) ≤ 1. Consider
2A
(1)
k
+(1)
l
,(1)
k
+(1)
l
(z) = a
kk
(z)z
2
l
+ 2a
kl
(z)z
k
z
l
+a
ll
(z)z
2
k
, k, l ∈ ¦1, . . . , m¦.
From the preceding arguments it is now clear that also deg
l
a
kk
(z) = 0, and
hence deg a(z) = 0. We ﬁnally have
B
(1)
k
+(1)
l
(z) = b
k
(z)z
l
+b
l
(z)z
k
+a
kl
(z), k, l ∈ ¦1, . . . , m¦,
from which we conclude that deg b(z) ≤ 1.
We can relax the hypothesis on G in Theorem 11.3.1 if from now on we
make the following standing assumptions: Z ⊂ R
m
is a cone, and b and ρ
satisfy a linear growth condition
b(z) +ρ(z) ≤ C(1 +z), ∀z ∈ Z, (11.13)
for some constant C ∈ R
+
.
11.3. POLYNOMIAL TERM STRUCTURES 121
Theorem 11.3.2 (Maximal Degree Problem II). Suppose that
'a(z)v, v` ≥ k(z)v
2
, ∀v ∈ R
m
, (11.14)
for some function k : Z →R
+
with
liminf
z∈?,z→∞
k(z) > 0. (11.15)
Then necessarily n ∈ ¦1, 2¦.
Conditions (11.14) and (11.15) say that a(z) becomes uniformly elliptic
for z large enough.
Proof. We shall make use of the basic inequality
[z
i
[ ≤ z
[i[
, ∀z ∈ R
m
. (11.16)
This is immediate, since
[z
i
[
z
[i[
=
[z
1
[
z
i
1
[z
m
[
z
im
≤ 1, ∀z ∈ R
m
` ¦0¦.
Now deﬁne
Γ
k
(x, z) :=
N
¸
µ=1
G
iµ
(x)
∂z
iµ
∂z
k
(11.17)
Λ
kl
(x, z) = Λ
lk
(x, z) :=
N
¸
µ=1
G
iµ
(x)
∂
2
z
iµ
∂z
k
∂z
l
. (11.18)
Then (11.2) can be rewritten as (integration)
n
¸
[i[=0
(g
i
(x) −g
i
(0)) z
i
=
m
¸
k=1
b
k
(z)Γ
k
(x, z)
+
1
2
m
¸
k,l=1
a
kl
(z) (Λ
kl
(x, z) −Γ
k
(x, z)Γ
l
(x, z)) ,
(11.19)
Suppose now that n > 2. We have from (11.17)
Γ
k
(x, z) =
¸
[i[=n
G
i
(x)i
k
z
i−(1)
k
+ =: P
k
(x, z) + ,
122 CHAPTER 11. MULTIFACTOR MODELS
where P
k
(x, z) is a homogeneous polynomial in z of order n − 1, and
stands for lower order terms in z. By assumptions there exist x ∈ R
+
and
k ∈ ¦1, . . . , m¦ such that P
k
(x, ) = 0. Choose z
∗
∈ Z`¦0¦ with P
k
(x, z
∗
) = 0
and set z
α
:= αz
∗
, for α > 0. Then we have z
α
∈ Z and
Γ
k
(x, z
α
) = α
n−1
P
k
(x, z
∗
) + ,
where denotes lower order terms in α. Consequently,
lim
α→∞
Γ
k
(x, z
α
)
z
α

n−1
=
P
k
(x, z
∗
)
z
∗

n−1
= 0. (11.20)
Combining (11.14) and (11.15) with (11.20) we conclude that
L := liminf
α→∞
1
z
α

2n−2
'a(z
α
)Γ(x, z
α
), Γ(x, z
α
)`
≥ liminf
α→∞
k(z
α
)
Γ(x, z
α
)
2
z
α

2n−2
> 0. (11.21)
On the other hand, by (11.19),
L ≤
n
¸
[i[=0
[g
i
(x) −g
i
(0)[
[z
i
α
[
z
α

2n−2
+
b(z
α
)
z
α

Γ(x, z
α
)
z
α

2n−3
+
1
2
a(z
α
)
z
α

2
Λ(x, z
α
)
z
α

2n−4
,
for all α > 0. In view of (11.17), (11.18), (11.13) and (11.16), the right hand
side converges to zero for α → ∞. This contradicts (11.21), hence n ≤ 2.
11.4 ExponentialPolynomial Families
We consider the Nelson–Siegel and Svensson families. For a discussion of
general exponentialpolynomial families see [8].
11.4.1 Nelson–Siegel Family
Recall the form of the Nelson–Siegel curves
G
NS
(x, z) = z
1
+ (z
2
+z
3
x)e
−z
4
x
.
11.4. EXPONENTIALPOLYNOMIAL FAMILIES 123
Proposition 11.4.1. There is no nontrivial diﬀusion process Z that is con
sistent with the Nelson–Siegel family. In fact, the unique solution to (11.2)
is
a(z) = 0, b
1
(z) = b
4
(z) = 0, b
2
(z) = z
3
−z
2
z
4
, b
3
(z) = −z
3
z
4
.
The corresponding state process is
Z
1
(t) ≡ z
1
,
Z
2
(t) = (z
2
+z
3
t) e
−z
4
t
,
Z
3
(t) = z
3
e
−z
4
t
,
Z
4
(t) ≡ z
4
,
where Z(0) = (z
1
, . . . , z
4
) denotes the initial point.
Proof. Exercise.
11.4.2 Svensson Family
Here the forward curve is
G
S
(x, z) = z
1
+ (z
2
+z
3
x)e
−z
5
x
+z
4
xe
−z
6
x
.
Proposition 11.4.2. The only nontrivial HJM model that is consistent with
the Svensson family is the Hull–White extended Vasicek short rate model
dr(t) =
z
1
z
5
+z
3
e
−z
5
t
+z
4
z
−2z
5
t
−z
5
r(t)
dt +
√
z
4
z
5
e
−z
5
t
dW
∗
(t),
where (z
1
, . . . , z
5
) are given by the initial forward curve
f(0, x) = z
1
+ (z
2
+z
3
x)e
−z
5
x
+z
4
xe
−2z
5
x
and W
∗
is some Brownian motion. The form of the corresponding state
process Z is given in the proof below.
Proof. The consistency equation (11.2) becomes
q
1
(x) +q
2
(x)e
−z
5
x
+q
3
(x)e
−z
6
x
+q
4
(x)e
−2z
5
x
+q
5
(x)e
−(z
5
+z
6
)x
+q
6
(x)e
−2z
6
x
= 0, (11.22)
124 CHAPTER 11. MULTIFACTOR MODELS
for some polynomials q
1
, . . . , q
6
. Indeed, we assume for the moment that
z
5
= z
6
, z
5
+z
6
= 0 and z
i
= 0 for all i = 1, . . . , 6. (11.23)
Then the terms involved in (11.2) are
∂
x
G
S
(x, z) = (−z
2
z
5
+z
3
−z
3
z
5
x)e
−z
5
x
+ (z
4
−z
4
z
6
x)e
−z
6
x
,
∇
z
G
S
(x, z) =
¸
¸
¸
¸
¸
¸
¸
1
e
−z
5
x
xe
−z
5
x
xe
−z
6
x
(−z
2
x −z
3
x
2
)e
−z
5
x
−z
4
x
2
e
−z
6
x
,
∂
z
i
∂
z
j
G
S
(x, z) = 0 for 1 ≤ i, j ≤ 4,
∇
z
∂
z
5
G
S
(x, z) =
¸
¸
¸
¸
¸
¸
¸
0
−xe
−z
5
x
−x
2
e
−z
5
x
0
(z
2
x
2
+z
3
x
3
)e
−z
5
x
0
, ∇
z
∂
z
6
G
S
(x, z) =
¸
¸
¸
¸
¸
¸
¸
0
0
0
−x
2
e
−z
6
x
0
z
4
x
3
e
−z
6
x
,
x
0
∇
z
G
S
(u, z) du =
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
¸
x
−
1
z
5
e
−z
5
x
+
1
z
5
−
x
z
5
−
1
z
2
5
e
−z
5
x
+
1
z
2
5
−
x
z
6
−
1
z
2
6
e
−z
6
x
+
1
z
2
6
z
3
z
5
x
2
+
z
2
z
5
+
2z
3
z
2
5
x +
z
2
z
2
5
+
2z
3
z
3
5
e
−z
5
x
−
z
2
z
2
5
−
z
3
z
3
5
z
4
z
6
x
2
+
2z
4
z
2
6
x +
2z
4
z
3
6
e
−z
6
x
−
z
4
z
3
6
.
Straightforward calculations lead to
q
1
(x) = −a
11
(z)x + ,
q
2
(x) = a
55
(z)
z
2
3
z
5
x
4
+ ,
q
3
(x) = a
66
(z)
z
2
4
z
6
x
4
+ ,
deg q
4
, deg q
5
deg q
6
≤ 3,
11.4. EXPONENTIALPOLYNOMIAL FAMILIES 125
where stands for lower order terms in x. Because of (11.23) we conclude
that
a
11
(z) = a
55
(z) = a
66
(z) = 0.
But a is a positive semideﬁnite symmetric matrix. Hence
a
1j
(z) = a
j1
(z) = a
5j
(z) = a
j5
(z) = a
6j
(z) = a
j6
(z) = 0 ∀j = 1 . . . , 6.
Taking this into account, expression (11.22) simpliﬁes considerably. We are
left with
q
1
(x) = b
1
(z),
deg q
2
(x), deg q
3
≤ 1,
q
4
(x) = a
33
(z)
1
z
5
x
2
+ ,
q
5
(x) = a
34
(z)
1
z
5
+
1
z
6
x
2
+ ,
q
6
(x) = a
44
(z)
1
z
6
x
2
+ .
Because of (11.23) we know that the exponents −2z
5
, −(z
5
+ z
6
) and −2z
6
are mutually diﬀerent. Hence
b
1
(z) = a
3j
(z) = a
j3
(z) = a
4j
(z) = a
j4
(z) = 0 ∀j = 1, . . . 6.
Only a
22
(z) is left as strictly positive candidate among the components of
a(z). The remaining terms are
q
2
(x) = (b
3
(z) +z
3
z
5
)x +b
2
(z) −z
3
−
a
22
(z)
z
5
+z
2
z
5
,
q
3
(x) = (b
4
(z) +z
4
z
6
)x −z
4
,
q
4
(x) = a
22
(z)
1
z
5
,
while q
1
= q
5
= q
6
= 0.
If 2z
5
= z
6
then also a
22
(z) = 0. If 2z
5
= z
6
then the condition q
3
+q
4
=
q
2
= 0 leads to
a
22
(z) = z
4
z
5
,
b
2
(z) = z
3
+z
4
−2
5
z
2
,
b
3
(z) = −z
5
z
3
,
b
4
(z) = −2z
5
z
4
.
126 CHAPTER 11. MULTIFACTOR MODELS
We derived the above results under the assumption (11.23). But the set
of z where (11.23) holds is dense Z. By continuity of a(z) and b(z) in z, the
above results thus extend for all z ∈ Z. In particular, all Z
i
’s but Z
2
are
deterministic; Z
1
, Z
5
and Z
6
are even constant.
Thus, since
a(z) = 0 if 2z
5
= z
6
,
we only have a nontrivial process Z if
Z
6
(t) ≡ 2Z
5
(t) ≡ 2Z
5
(0).
In that case we have, writing shortly z
i
= Z
i
(0),
Z
1
(t) ≡ z
1
,
Z
3
(t) = z
3
e
−z
5
t
,
Z
4
(t) = z
4
z
−2z
5
t
and
dZ
2
(t) =
z
3
e
−z
5
t
+z
4
z
−2z
5
t
−z
5
Z
2
(t)
dt +
d
¸
j=1
ρ
2j
(t) dW
j
(t),
where ρ
2j
(t) (not necessarily deterministic) are such that
d
¸
j=1
ρ
2
2j
(t) = a
22
(Z(t)) = z
4
z
5
e
−2z
5
t
.
By L´evy’s characterization theorem we have that
W
∗
(t) :=
d
¸
j=1
t
0
ρ
2j
(s)
√
z
4
z
5
e
−z
5
s
dW
j
(s)
is a realvalued standard Brownian motion (→ exercise). Hence the corre
sponding short rate process
r(t) = G
S
(0, Z(t)) = z
1
+Z
2
(t)
satisﬁes
dr(t) =
z
1
z
5
+z
3
e
−z
5
t
+z
4
z
−2z
5
t
−z
5
r(t)
dt +
√
z
4
z
5
e
−z
5
t
dW
∗
(t).
Chapter 12
Market Models
Instantaneous forward rates are not always easy to estimate, as we have seen.
One may want to model other rates, such as LIBOR, directly. There has been
some eﬀort in the years after the publication of HJM [9] in 1992 to develop
arbitragefree models of other than instantaneous, continuously compounded
rates. The breakthrough came 1997 with the publications of Brace–Gatarek–
Musiela [5] (BGM), who succeeded to ﬁnd a HJM type model inducing log
normal LIBOR rates, and Jamshidian [12], who developed a framework for
arbitragefree LIBOR and swap rate models not based on HJM. The principal
idea of both approaches is to chose a diﬀerent numeraire than the riskfree
account (the latter does not even necessarily have to exist). Both approaches
lead to Black’s formula for either caps (LIBOR models) or swaptions (swap
rate models). Because of this they are usually referred to as “market models”.
To start with we consider the HJM setup, as in Chapter 9. Recall that,
for a ﬁxed δ (typically 1/4 = 3 months), the forward δperiod LIBOR for the
future date T prevailing at time t is the simple forward rate
L(t, T) = F(t; T, T +δ) =
1
δ
P(t, T)
P(t, T +δ)
−1
.
We have seen in Chapter 9 that P(t, T)/P(t, T + δ) is a martingale for the
(T +δ)forward measure Q
T+δ
. In particular (see (9.8))
d
P(t, T)
P(t, T +δ)
=
P(t, T)
P(t, T +δ)
σ
T,T+δ
(t) dW
T+δ
(t).
127
128 CHAPTER 12. MARKET MODELS
Hence
dL(t, T) =
1
δ
d
P(t, T)
P(t, T +δ)
=
1
δ
P(t, T)
P(t, T +δ)
σ
T,T+δ
(t) dW
T+δ
(t)
=
1
δ
(δL(t, T) + 1)σ
T,T+δ
(t) dW
T+δ
(t).
Now suppose there exists a deterministic R
d
valued function λ(t, T) such
that
σ
T,T+δ
(t) =
δL(t, T)
δL(t, T) + 1
λ(t, T). (12.1)
Plugging this in the above formula, we get
dL(t, T) = L(t, T)λ(t, T) dW
T+δ
(t),
which is equivalent to
L(t, T) = L(s, T) exp
t
s
λ(u, T) dW
T+δ
(u) −
1
2
t
s
λ(u, T)
2
du
,
for s ≤ t ≤ T. Hence the Q
T+δ
distribution of log L(T, T) conditional on T
t
is Gaussian with mean
log L(t, T) −
1
2
T
t
λ(s, T)
2
ds
and variance
T
t
λ(s, T)
2
ds.
The time t price of a caplet with reset date T, settlement date T + δ and
strike rate κ is thus
E
Q
e
−
T+δ
0
r(s) ds
δ(L(T, T) −κ)
+
[ T
t
= P(t, T +δ)E
Q
T+δ
δ(L(T, T) −κ)
+
[ T
t
= δP(t, T +δ) (L(t, T)Φ(d
1
(t, T)) −κΦ(d
2
(t, T))) ,
where
d
1,2
(t, T) :=
log
L(t,T)
κ
±
1
2
T
t
λ(s, T)
2
ds
T
t
λ(s, T)
2
ds
1
2
,
12.1. MODELS OF FORWARD LIBOR RATES 129
and Φ is the standard Gaussian CDF. This is just Black’s formula for the
caplet price with σ(t)
2
set equal to
1
T −t
T
t
λ(s, T)
2
ds,
as introduced in Section 2.6!
We have thus shown that any HJM model satisfying (12.1) yields Black’s
formula for caplet prices. But do such HJM models exist? The answer is yes,
but the construction and proof are not easy. The idea is to rewrite (12.1),
using the deﬁnition of σ
T,T+δ
(t), as (→ exercise)
T+δ
T
σ(t, u) du =
1 −e
−
T+δ
T
f(t,u) du
λ(t, T).
Diﬀerentiating in T gives
σ(t, T +δ)
= σ(t, T) + (f(t, T +δ) −f(t, T))e
−
T+δ
T
f(t,u) du
λ(t, T)
+
1 −e
−
T+δ
T
f(t,u) du
∂
T
λ(t, T).
This is a recurrence relation that can be solved by forward induction, once
σ(t, ) is determined on [0, δ) (typically, σ(t, T) = 0 for T ∈ [0, δ)). This
gives a complicated dependence of σ on the forward curve. Now it has to be
proved that the corresponding HJM equations for the forward rates have a
unique and wellbehaved solution. This all has been carried out by BGM [5],
see also [8, Section 5.6].
12.1 Models of Forward LIBOR Rates
→ MR[19](Chapter 14), Z[27](Section 4.7)
There is a more direct approach to LIBOR models without making ref
erence to continuously compounded forward and short rates. In a sense,
we place ourselves outside of the HJM framework (although HJM is often
implicitly adopted). Instead of the risk neutral martingale measure we will
work under forward measures; the numeraires accordingly being bond price
processes.
130 CHAPTER 12. MARKET MODELS
12.1.1 Discretetenor Case
We ﬁx a ﬁnite time horizon T
M
= Mδ, for some M ∈ N, and a probability
space
(Ω, T, (T
t
)
t∈[0,T
M
]
, Q
T
M
),
where T
t
= T
W
T
M
t
is the ﬁltration generated by a ddimensional Brownian
motion W
T
M
(t), t ∈ [0, T
M
]. The notation already suggests that Q
T
M
will
play the role of the T
M
forward measure. Write
T
m
:= mδ, m = 0, . . . , M.
We are going to construct a model for the forward LIBOR rates with matu
rities T
1
, . . . , T
M−1
. We take as given:
• for every m ≤ M − 1, an R
d
valued, bounded, deterministic function
λ(t, T
m
), t ∈ [0, T
m
], which represents the volatility of L(t, T
m
);
• an initial strictly positive and decreasing discrete term structure
P(0, T
m
), m = 0, . . . , M,
and hence strictly positive initial forward LIBOR rates
L(0, T
m
) =
1
δ
P(0, T
m
)
P(0, T
m+1
)
−1
, m = 0, . . . , M −1.
We proceed by backward induction and postulate ﬁrst that
dL(t, T
M−1
) = L(t, T
M−1
)λ(t, T
M−1
) dW
T
M
(t), t ∈ [0, T
M−1
],
L(0, T
M−1
) =
1
δ
P(0, T
M−1
)
P(0, T
M
)
−1
which is of course equivalent to
L(t, T
M−1
) =
1
δ
P(0, T
M−1
)
P(0, T
M
)
−1
c
t
λ(, T
M−1
) W
T
M
.
Now deﬁne the bounded (why?) R
d
valued process
σ
T
M−1
,T
M
(t) :=
δL(t, T
M−1
)
δL(t, T
M−1
) + 1
λ(t, T
M−1
), t ∈ [0, T
M−1
],
12.1. MODELS OF FORWARD LIBOR RATES 131
compare with (12.1).
This induces an equivalent probability measure Q
T
M−1
∼ Q
T
M
on T
T
M−1
via
dQ
T
M−1
dQ
T
M
= c
T
M−1
σ
T
M−1
,T
M
W
T
M
,
and by Girsanov’s theorem
W
T
M−1
(t) := W
T
M
(t) −
t
0
σ
T
M−1
,T
M
(s) ds, t ∈ [0, T
M−1
],
is a Q
T
M−1
Brownian motion.
Hence we can postulate
dL(t, T
M−2
) = L(t, T
M−2
)λ(t, T
M−2
) dW
T
M−1
(t), t ∈ [0, T
M−2
],
L(0, T
M−2
) =
1
δ
P(0, T
M−2
)
P(0, T
M−1
)
−1
,
that is,
L(t, T
M−2
) =
1
δ
P(0, T
M−2
)
P(0, T
M−1
)
−1
c
t
λ(, T
M−2
) W
T
M−1
,
and deﬁne the bounded R
d
valued process
σ
T
M−2
,T
M−1
(t) :=
δL(t, T
M−2
)
δL(t, T
M−2
) + 1
λ(t, T
M−2
), t ∈ [0, T
M−2
],
yielding an equivalent probability measure Q
T
M−2
∼ Q
T
M−1
on T
T
M−2
via
dQ
T
M−2
dQ
T
M−1
= c
T
M−2
σ
T
M−2
,T
M−1
W
T
M−1
,
and the Q
T
M−2
Brownian motion
W
T
M−2
(t) := W
T
M−1
(t) −
t
0
σ
T
M−2
,T
M−1
(s) ds, t ∈ [0, T
M−2
].
Repeating this procedure leads to a family of lognormal martingales
(L(t, T
m
))
t∈[0,Tm]
under their respective measures Q
Tm
.
132 CHAPTER 12. MARKET MODELS
Bond Prices
What about bond prices? For all m = 1, . . . , M, we then can deﬁne the
forward price process
P(t, T
m−1
)
P(t, T
m
)
:= δL(t, T
m−1
) + 1, t ∈ [0, T
m−1
].
Since
d
P(t, T
m−1
)
P(t, T
m
)
= δ dL(t, T
m−1
) = δL(t, T
m−1
)λ(t, T
m−1
) dW
Tm
(t)
=
P(t, T
m−1
)
P(t, T
m
)
σ
T
m−1
,Tm
(t) dW
Tm
(t)
we get that
P(t, T
m−1
)
P(t, T
m
)
=
P(0, T
m−1
)
P(0, T
m
)
c
t
σ
T
m−1
,Tm
W
Tm
, t ∈ [0, T
m−1
],
which is a Q
Tm
martingale.
From this we can derive, for 0 ≤ i < j ≤ m,
P(T
i
, T
j
) =
j
¸
m=i+1
P(T
i
, T
m
)
P(T
i
, T
m−1
)
=
j
¸
m=i+1
1
δL(T
i
, T
m−1
) + 1
. (12.2)
However, it is not possible to uniquely determine the continuous time dynam
ics of a bond price P(t, T
m
) in the discretetenor model of forward LIBOR
rates. The knowledge of forward LIBOR rates for all maturities T ∈ [0, T
M−1
]
is necessary.
LIBOR Dynamics under Diﬀerent Measures
We are interested in ﬁnding the dynamics of L(t, T
m
) under any of the forward
measures Q
T
k
.
12.1. MODELS OF FORWARD LIBOR RATES 133
Lemma 12.1.1. Let 0 ≤ m ≤ M − 1 and 0 ≤ k ≤ M. Then the dynamics
of L(t, T
m
) under Q
T
k
is given according to the three cases
k < m + 1 :
dL(t, T
m
)
L(t, T
m
)
= λ(t, T
m
)
m
¸
l=k
σ
T
l
,T
l+1
(t) dt +λ(t, T
m
) dW
T
k
(t);
k = m + 1 :
dL(t, T
m
)
L(t, T
m
)
= λ(t, T
m
) dW
T
m+1
(t);
k > m + 1 :
dL(t, T
m
)
L(t, T
m
)
= −λ(t, T
m
)
k−1
¸
l=m+1
σ
T
l
,T
l+1
(t)dt +λ(t, T
m
)dW
T
k
(t),
for t ∈ [0, T
k
∧ T
m
].
Proof. This follows from the equality
W
T
i
(t) = W
T
j
(t) −
j−1
¸
l=i
t
0
σ
T
l
,T
l+1
(s) ds, t ∈ [0, T
i
],
for all 1 ≤ i < j ≤ M.
Derivative Pricing
Here is a useful formula, which can be combined with (12.2).
Lemma 12.1.2. Let X ∈ L
1
(Q
Tm
) be a T
m
contingent claim, m ≤ M. Then
its price π(t) at t ≤ T
m
is given by
π(t) = P(t, T
m
)E
Q
Tm [X [ T
t
]
= P(t, T
n
)E
Q
Tn
¸
X
P(T
m
, T
n
)
[ T
t
,
for all m < n ≤ M (strictly speaking, this formula makes sense only for
t = T
j
, 0 ≤ j ≤ m, since we know P(t, T
n
) only for such t).
Proof. Notice that
dQ
T
k
dQ
T
k+1
[
Tt
= c
t
σ
T
k
,T
k+1
W
T
k+1
=
P(0, T
k+1
)
P(0, T
k
)
P(t, T
k
)
P(t, T
k+1
)
, t ∈ [0, T
k
].
134 CHAPTER 12. MARKET MODELS
Hence
dQ
Tm
dQ
Tn
[
Tt
=
n−1
¸
k=m
dQ
T
k
dQ
T
k+1
[
Tt
=
n−1
¸
k=m
P(0, T
k+1
)
P(0, T
k
)
P(t, T
k
)
P(t, T
k+1
)
=
P(0, T
n
)
P(0, T
m
)
P(t, T
m
)
P(t, T
n
)
.
Bayes’ rule now yields the assertion, since the ﬁrst equality was derived in
Proposition 9.1.2 (strictly speaking, we assumed there the existence of a sav
ings account. But even if there is no risk neutral but only forward measures,
the reasoning in Section 9.1 makes it clear that (9.3) is the arbitragefree
price of X).
Swaptions
Consider a payer swaption with nominal 1, strike rate K, maturity T
µ
and
underlying tenor T
µ
, T
µ+1
, . . . , T
ν
(T
µ
is the ﬁrst reset date and T
ν
the ma
turity of the underlying swap), for some positive integers µ < ν ≤ M. Its
payoﬀ at maturity is
δ
ν−1
¸
m=µ
P(T
µ
, T
m
)(L(T
µ
, T
m
) −K)
+
.
The swaption price at t = 0 (for simplicity) therefore
π(0) = δP(0, T
µ
)E
Q
Tµ
ν−1
¸
m=µ
P(T
µ
, T
m
)(L(T
µ
, T
m
) −K)
+
¸
¸
.
To compute π(0) we thus need to know the joint distribution of
L(T
µ
, T
µ
), L(T
µ
, T
µ+1
), . . . , L(T
µ
, T
ν−1
)
under the measure Q
Tµ
. This cannot be done analytically anymore, so one
has to resort to numerical procedures.
We sketch here the Monte Carlo method. Notice that by Lemma 12.1.1,
Itˆ o’s formula and the deﬁnition of σ
T
l
,T
l+1
(t) we have
d log L(t, T
m
) =
λ(t, T
m
)
m
¸
l=µ
δL(t, T
l
)
δL(t, T
l
) + 1
λ(t, T
l
) −
1
2
λ(t, T
m
)
2
dt
+λ(t, T
m
) dW
Tµ
(t),
12.1. MODELS OF FORWARD LIBOR RATES 135
for t ∈ [0, T
µ
] and m = µ, . . . , ν −1. Write α(t, T
m
) for the above drift term,
and let t
i
=
i
n
T
µ
, i = 0, . . . , n, n ∈ N large enough, be a partition of [0, T
µ
].
Then we can approximate
log L(t
i+1
, T
m
) = log L(t
i
, T
m
) +
t
i+1
t
i
α(s, T
m
) ds +
t
i+1
t
i
λ(s, T
m
) dW
Tµ
(s)
≈ log L(t
i
, T
m
) +α(t
i
, T
m
)
1
n
+ζ
m
(i),
where
ζ
m
(i) :=
t
i+1
t
i
λ(s, T
m
) dW
Tµ
(s),
such that ζ(i) = (ζ
µ
(i), . . . , ζ
ν−1
(i)), i = 0, . . . , n −1, are independent Gaus
sian (ν −µ)vectors with mean zero and covariance matrix
Cov[ζ
k
(i), ζ
l
(i)] =
t
i+1
t
i
λ(s, T
k
) λ(s, T
l
) ds,
which can easily be simulated.
Forward Swap Measure
We consider the above payer swap with reset dates T
µ
, . . . , T
ν−1
and cashﬂow
dates T
µ+1
, . . . , T
ν
(= maturity of the swap). The corresponding forward
swap rate at time t ≤ T
µ
is
R
swap
(t) =
P(t, T
µ
) −P(t, T
ν
)
δ
¸
ν
k=µ+1
P(t, T
k
)
=
1 −
P(t,Tν)
P(t,Tµ)
δ
¸
ν
k=µ+1
P(t,T
k
)
P(t,Tµ)
. (12.3)
Since for any 0 ≤ l < m ≤ M
P(t, T
l
)
P(t, T
m
)
=
P(t, T
l
)
P(t, T
l+1
)
P(t, T
m−1
)
P(t, T
m
)
=
m−1
¸
i=l
(1 +δL(t, T
i
)) ,
R
swap
(t) is given in terms of the above constructed LIBOR rates.
Deﬁne the positive Q
Tµ
martingale
D(t) :=
ν
¸
k=µ+1
P(t, T
k
)
P(t, T
µ
)
, t ∈ [0, T
µ
].
136 CHAPTER 12. MARKET MODELS
This induces an equivalent probability measure Q
swap
∼ Q
Tµ
, the forward
swap measure, on T
Tµ
by
dQ
swap
dQ
Tµ
=
D(T
µ
)
D(0)
.
Lemma 12.1.3. The forward swap rate process R
swap
(t), t ∈ [0, T
µ
], is a
Q
swap
martingale.
Proof. Let 0 ≤ m ≤ M and 0 ≤ s ≤ t ≤ T
m
∧ T
µ
. Then
E
Q
swap
¸
P(t, T
m
)
P(t, T
µ
)D(t)
[ T
s
=
1
D(s)
E
Q
Tµ
¸
P(t, T
m
)
P(t, T
µ
)D(t)
D(t) [ T
s
=
1
D(s)
P(s, T
m
)
P(s, T
µ
)
.
Now the lemma follows (set m = 0, µ) by (12.3).
The payoﬀ at maturity of the above swaption can be written as
δD(T
µ
) (R
swap
(T
0
) −K)
+
.
Hence the price is
π(0) = δP(0, T
µ
)E
Q
Tµ
D(T
µ
) (R
swap
(T
0
) −K)
+
= δP(0, T
µ
)D(0)E
Q
swap
(R
swap
(T
0
) −K)
+
= δ
ν
¸
k=µ+1
P(0, T
k
)E
Q
swap
(R
swap
(T
0
) −K)
+
.
Lemma 12.1.3 tells us that R
swap
is a positive Q
swap
martingale and hence of
the form
dR
swap
(t) = R
swap
(t)ρ
swap
(t) dW
swap
(t), t ∈ [0, T
µ
],
for some Q
swap
Brownian motion W
swap
and some swap volatility process
ρ
swap
. Hence, under the hypothesis
(H) ρ
swap
(t) is deterministic,
12.1. MODELS OF FORWARD LIBOR RATES 137
we would have that log R
swap
(T
µ
) is Gaussian distributed under Q
swap
with
mean
log R
swap
(0) −
1
2
Tµ
0
ρ
swap
(t)
2
dt
and variance
Tµ
0
ρ
swap
(t)
2
dt.
The swaption price would then be
π(0) = δ
ν
¸
k=µ+1
P(0, T
k
) (R
swap
(0)Φ(d
1
) −KΦ(d
2
)) ,
with
d
1,2
:=
log
Rswap(0)
K
±
1
2
Tµ
0
ρ
swap
(t)
2
dt
Tµ
0
ρ
swap
(t)
2
dt
1
2
.
This is Black’s formula with volatility σ
2
given by
1
T
µ
Tµ
0
ρ
swap
(t)
2
dt.
However, one can show that ρ
swap
cannot be deterministic in our log
normal LIBOR setup. So hypothesis (H) does not hold. For swaption pricing
it would be natural to model the forward swap rates directly and postulate
that they are lognormal under the forward swap measures. This approach
has been carried out by Jamshidian [12] and others. It could be shown, how
ever, that then the forward LIBOR rate volatility cannot be deterministic.
So either one gets Black’s formula for caps or for swaptions, but not simulta
neously for both. Put in other words, when we insist on lognormal forward
LIBOR rates then swaption prices have to be approximated. One possibility
is to use Monte Carlo methods. Another way (among many others) is now
sketched below.
We have seen in Section 2.4.3 that the forward swap rate can be written
as weighted sum of forward LIBOR rates
R
swap
(t) =
ν
¸
m=µ+1
w
m
(t)L(t, T
m−1
),
138 CHAPTER 12. MARKET MODELS
with weights
w
m
(t) =
P(t, T
m
)
D(t)P(t, T
µ
)
=
1
1+δL(t,Tµ)
1
1+δL(t,T
m−1
)
¸
ν
j=µ+1
1
1+δL(t,Tµ)
1
1+δL(t,T
j−1
)
.
According to empirical studies, the variability of the w
m
’s is small compared
to the variability of the forward LIBOR rates. We thus approximate w
m
(t)
by its deterministic initial value w
m
(0). So that
R
swap
(t) ≈
ν
¸
m=µ+1
w
m
(0)L(t, T
m−1
),
and hence, under the T
µ
forward measure Q
Tµ
dR
swap
(t) ≈ ( ) dt +
ν
¸
m=µ+1
w
m
(0)L(t, T
m−1
)λ(t, T
m−1
) dW
Tµ
, t ∈ [0, T
µ
].
We obtain for the forward swap volatility
ρ
swap
(t)
2
=
d 'log R
swap
, log R
swap
`
t
dt
≈
ν
¸
k,l=µ+1
w
k
(0)w
l
(0)L(t, T
k−1
)L(t, T
l−1
)λ(t, T
k−1
) λ(t, T
l−1
)
R
2
swap
(t)
.
In a further approximation we replace all random variables by their time 0
values, such that the quadratic variation of log R
swap
(t) becomes approxima
tively deterministic
ρ
swap
(t)
2
≈
ν
¸
k,l=µ+1
w
k
(0)w
l
(0)L(0, T
k−1
)L(0, T
l−1
)λ(t, T
k−1
) λ(t, T
l−1
)
R
2
swap
(0)
.
Denote the square root of the right hand side by ˜ ρ
swap
(t), and deﬁne the
Q
swap
Brownian motion (L´evy’s characterization theorem)
W
∗
(t) :=
t
0
d
¸
j=1
ρ
swap
j
(s)
ρ
swap
(s)
dW
swap
j
(s), t ∈ [0, T
µ
].
12.1. MODELS OF FORWARD LIBOR RATES 139
Then we have
dR
swap
(t) = R
swap
(t)ρ
swap
(t) dW
∗
(t)
≈ R
swap
(t)˜ ρ
swap
(t) dW
∗
(t).
Hence we can approximate the swaption price in our lognormal forward
LIBOR model by Black’s swaption price formula where σ
2
is to be replaced
by
1
T
µ
Tµ
0
ν
¸
k,l=µ+1
w
k
(0)w
l
(0)L(0, T
k−1
)L(0, T
l−1
)λ(t, T
k−1
) λ(t, T
l−1
)
R
2
swap
(0)
dt.
This is “Rebonato’s formula”, since it originally appears in his book R[22].
The goodness of this approximation has been tested numerically by several
authors, see BM[6](Chapter 8). They conclude that “the approximation is
satisfactory in general”.
Implied Savings Account
Given the LIBOR L(T
i
, T
i
) for period [T
i
, T
i+1
], for all i = 0, . . . , M − 1, we
can deﬁne the discretetime, implied savings account process
B
∗
(0) := 1,
B
∗
(T
m
) := (1 +δL(T
m−1
, T
m−1
))B
∗
(T
m−1
), m = 1, . . . , M,
that is,
B
∗
(T
n
) = B
∗
(T
m
)
n−1
¸
k=m
1
P(T
k
, T
k+1
)
, m < n ≤ M.
Hence B
∗
(T
m
) can be interpreted as the cash amount accumulated up to time
T
m
by rolling over a series of zerocoupon bonds with the shortest maturities
available.
By construction, B
∗
is a strictly increasing and predictable process with
respect to the discretetime ﬁltration (T
Tm
), that is,
B
∗
(T
m
) is T
T
m−1
measurable, for all m = 1, . . . , M.
Lemma 12.1.4. For all 0 ≤ m ≤ M we have
E
Q
T
M
[B
∗
(T
M
) [ T
Tm
] =
B
∗
(T
m
)
P(T
m
, T
M
)
.
140 CHAPTER 12. MARKET MODELS
Proof. Exercise.
Lemma 12.1.4 yields in particular
E
Q
T
M
[B
∗
(T
M
)P(0, T
M
)] = 1 and B
∗
(T
M
)P(0, T
M
) > 0,
so that we can deﬁne the equivalent probability measure Q
∗
∼ Q
T
M
on T
T
M
by
dQ
∗
dQ
T
M
= B
∗
(T
M
)P(0, T
M
).
Q
∗
can be interpreted as risk neutral martingale measure since
P(T
k
, T
l
) = E
Q
∗
¸
B
∗
(T
k
)
B
∗
(T
l
)
[ T
T
k
, 0 ≤ k ≤ l ≤ M. (12.4)
Indeed, in view of Lemma 12.1.4 we have for m ≤ M
dQ
∗
dQ
T
M
[
T
Tm
= B
∗
(T
m
)
P(0, T
M
)
P(T
m
, T
M
)
.
Hence (Bayes again)
E
Q
∗
¸
B
∗
(T
k
)
B
∗
(T
l
)
[ T
T
k
=
E
Q
T
M
B
∗
(T
k
)
B
∗
(T
l
)
B
∗
(T
l
)
P(T
l
,T
M
)
[ T
T
k
B
∗
(T
k
)
P(T
k
,T
M
)
= P(T
k
, T
l
),
which proves (12.4). Put in other words, (12.4) shows that for any 0 ≤ l ≤ M
the discretetime process
P(T
k
, T
l
)
B
∗
(T
k
)
k=0,...,l
is a Q
∗
martingale with respect to (T
T
k
).
12.1.2 Continuoustenor Case
We now specify the continuum of all forward LIBOR rates L(t, T), for T ∈
[0, T
M−1
]. Given the discretetenor skeleton constructed in the previous sec
tion, it is enough to ﬁll the gaps between the T
j
s. Each forward LIBOR rate
L(t, T) will follow a lognormal process under the forward measure for the
date T +δ.
The stochastic basis is the same as before. In addition, we now need a
continuum of initial dates:
12.1. MODELS OF FORWARD LIBOR RATES 141
• for every T ∈ [0, T
M−1
], an R
d
valued, bounded, deterministic function
λ(t, T), t ∈ [0, T], which represents the volatility of L(t, T);
• an initial strictly positive and decreasing term structure
P(0, T), T ∈ [0, T
M
],
and hence an initial strictly positive forward LIBOR curve
L(0, T) =
1
δ
P(0, T)
P(0, T +δ)
−1
, T ∈ [0, T
M−1
].
First, we construct a discretetenor model for L(t, T
m
), m = 0, . . . , M−1,
as in the previous section.
Second, we focus on the forward measures for dates T ∈ [T
M−1
, T
M
]. We
do not have to take into account forward LIBOR rates for these dates, since
they are not deﬁned there. However, we are given the values of the implied
savings account B
∗
(T
M−1
) and B
∗
(T
M
) and the probability measure Q
∗
. By
monotonicity there exists a unique deterministic increasing function
α : [T
M−1
, T
M
] → [0, 1]
with α(T
M−1
) = 0 and α(T
M
) = 1, such that
log B
∗
(T) := (1 −α(T)) log B
∗
(T
M−1
) +α(T) log B
∗
(T
M
)
satisﬁes
P(0, T) = E
Q
∗
¸
1
B
∗
(T)
, ∀T ∈ [T
M−1
, T
M
].
Let T ∈ [T
M−1
, T
M
]. Since (→ exercise) B
∗
(T) is T
T
measurable, strictly
positive and
E
Q
∗
¸
1
B
∗
(T)P(0, T)
= 1
we can deﬁne the Tforward measure Q
T
∼ Q
∗
on T
T
by
dQ
T
dQ
∗
=
1
B
∗
(T)P(0, T)
.
Then we have
dQ
T
dQ
T
M
=
dQ
T
dQ
∗
dQ
∗
dQ
T
M
=
B
∗
(T
M
)P(0, T
M
)
B
∗
(T)P(0, T)
.
142 CHAPTER 12. MARKET MODELS
By the representation theorem for Q
T
M
martingales there exists a unique
σ
T,T
M
∈ L such that (→ exercise)
dQ
T
dQ
T
M
[
Tt
= E
Q
T
M
¸
B
∗
(T
M
)P(0, T
M
)
B
∗
(T)P(0, T)
[ T
t
= exp
t
0
σ
T,T
M
(s) dW
T
M
(s) −
1
2
t
0
σ
T,T
M
(s)
2
ds
= c
t
σ
T,T
M
W
T
M
,
for t ∈ [0, T]. Girsanov’s theorem tells us that
W
T
(t) := W
T
M
(t) −
t
0
σ
T,T
M
(s) ds, t ∈ [0, T],
is a Q
T
Brownian motion.
Third, since T ∈ [T
M−1
, T
M
] was arbitrary, we can now deﬁne the forward
LIBOR process L(t, T) for any T ∈ [T
M−2
, T
M−1
] as
dL(t, T) = L(t, T)λ(t, T) dW
T+δ
(t),
L(0, T) =
1
δ
P(0, T)
P(0, T +δ)
−1
.
This in turn deﬁnes the positive and bounded process
σ
T,T+δ
(t) :=
δL(t, T)
δL(t, T) + 1
λ(t, T), t ∈ [0, T],
for any T ∈ [T
M−2
, T
M−1
]. The forward measures for T ∈ [T
M−2
, T
M−1
] are
now given by
dQ
T
dQ
T+δ
= c
T
σ
T,T+δ
W
T+δ
.
Hence we have (→ exercise)
dQ
T
dQ
T
M
[
Tt
=
dQ
T
dQ
T+δ
[
Tt
dQ
T+δ
dQ
T
M
[
Tt
= c
t
σ
T,T+δ
W
T+δ
c
t
σ
T+δ,T
M
W
T
M
= c
t
σ
T,T
M
W
T
M
, t ∈ [0, T],
12.1. MODELS OF FORWARD LIBOR RATES 143
for any T ∈ [T
M−2
, T
M−1
], where
σ
T,T
M
:= σ
T,T+δ
+σ
T+δ,T
M
.
Proceeding by backward induction yields the forward measure Q
T
and
Q
T
Brownian motion W
T
for all T ∈ [0, T
M
], and forward LIBOR rates
L(t, T) for all T ∈ [0, T
M−1
].
This way, we obtain the zerocoupon bond prices for all maturities. In
deed, for any 0 ≤ T ≤ S ≤ T
M
, it is reasonable to deﬁne (why?) the forward
price process
P(t, S)
P(t, T)
:=
P(0, S)
P(0, T)
dQ
S
dQ
T
[
Tt
=
P(0, S)
P(0, T)
dQ
S
dQ
T
M
[
Tt
dQ
T
M
dQ
T
[
Tt
=
P(0, S)
P(0, T)
c
t
−σ
T,S
W
T
, t ∈ [0, T],
where (→ exercise)
σ
T,S
:= σ
T,T
M
−σ
S,T
M
.
In particular, for t = T we get
P(T, S) =
P(0, S)
P(0, T)
c
T
−σ
T,S
W
T
.
Notice that now P(T, S) may be greater than 1, unless S −T = mδ for some
integer m. Hence even though all δperiod forward LIBOR rates L(t, T) are
positive, there may be negative interest rates for other than δ periods.
144 CHAPTER 12. MARKET MODELS
Chapter 13
Default Risk
→ [24, Chapter 2], [1], etc.
So far bond price processes P(t, T) had the property that P(T, T) = 1.
That is, the payoﬀ was certain, there was no risk of default of the issuer. This
may be the case for treasury bonds. Corporate bonds however may bear a
substantial risk of default. Investors should be adequately compensated by
a risk premium, which is reﬂected by a higher yield on the bond.
For the modelling of credit risk we have to consider the following risk
elements:
• Default probabilities: probability that the debtor will default on its
obligations to repay its debt.
• Recovery rates: proportion of value delivered after default has occurred.
• Transition probabilities: between credit ratings (credit migration).
Usually one has to model objective (for the rating) and riskneutral (for the
pricing) probabilities.
13.1 Transition and Default Probabilities
There are three main approaches to the modelling of transition and default
probabilities:
• Historical method: rating agencies determine default and transition
probabilities by counting defaults that actually occurred in the past
for diﬀerent rating classes.
145
146 CHAPTER 13. DEFAULT RISK
• Structural approach: models the value of a ﬁrm’s assets. Default is
when this value hits a certain lower bound. Goes back to Merton
(1974) [18].
• Intensity based method: default is speciﬁed exogenously by a stopping
time with given intensity process.
We brieﬂy discuss the ﬁrst two approaches in this section. The intensity
based method is treated in more detail in Section 13.2 below.
13.1.1 Historical Method
Rating agencies provide timely, objective information and credit analysis of
obligors. Usually they operate without government mandate and are inde
pendent of any investment banking ﬁrm or similar organization. Among
the biggest US agencies are Moody’s Investors Service and Standard&Poor’s
(S&P).
After issuance and assignment of the initial obligor’s rating, the rating
agency regularly checks and adjusts the rating. If there is a tendency observ
able that may aﬀect the rating, the obligor is set on the Rating Review List
(Moody’s) or the Credit Watch List (S&P). The number of Moody’s rated
obligors has increased from 912 in 1960 to 3841 in 1997.
The formal deﬁnition of default and transition rates is the following.
Deﬁnition 13.1.1. 1. The historical oneyear default rate, based on the
time frame [Y
0
, Y
1
], for an Rrated issuer is
d
R
:=
¸
Y
1
y=Y
0
M
R
(y)
¸
Y
1
y=Y
0
N
R
(y)
,
where N
R
(y) is the number of issuers with rating R at beginning of year
y, and M
R
(y) is the number of issuers with rating R at beginning of
year y which defaulted in that year.
2. The historical oneyear transition rate from rating R to R
t
, based on
the time frame [Y
0
, Y
1
], is
tr
R,R
:=
¸
Y
1
y=Y
0
M
R,R
(y)
¸
Y
1
y=Y
0
N
R
(y)
,
13.1. TRANSITION AND DEFAULT PROBABILITIES 147
Table 13.1: Rating symbols.
S&P Moody’s Interpretation
Investmentgrade ratings
AAA Aaa Highest quality, extremely strong
AA+ Aa1
AA Aa2 High quality
AA Aa3
A+ A1
A A2 Strong payment capacity
A A3
BBB+ Baa1
BBB Baa2 Adequate payment capacity
BBB Baa3
Speculativegrade ratings
BB+ Ba1 Likely to fulﬁll obligations
BB Ba2 ongoing uncertainty
BB Ba3
B+ B1
B B2 High risk obligations
B B3
CCC+ Caa1
CCC Caa2 Current vulnerability to default
CCC Caa3
CC
C Ca In bankruptcy or default
D or other marked shortcoming
where N
R
(y) is as above, and M
R,R
(y) is the number of issuers with
rating R at beginning of year y and R
t
at the end of that year.
Transition rates are gathered in a transition matrix as shown in Ta
ble 13.2.
The historical method has several shortcomings:
• It neglects the default rate volatility. Transition and default probabili
ties are dynamic and vary over time, depending on economic conditions.
148 CHAPTER 13. DEFAULT RISK
Table 13.2: S&P’s oneyear transition and default rates, based on the
time frame [1980,2000] (Standard&Poor’s, Ratings Performance 2000,
see http://ﬁnancialcounsel.com/Articles/Investment/ARTINV0000069
2000Ratings.pdf).
Rating at end of year (R
t
)
Initial AAA AA A BBB BB B CCC D
rating (R)
AAA 93.66 5.83 0.40 0.09 0.03 0.00 0.00 0.00
AA 0.66 91.72 6.94 0.49 0.06 0.09 0.02 0.01
A 0.07 2.25 91.76 5.18 0.49 0.20 0.01 0.04
BBB 0.03 0.26 4.83 89.24 4.44 0.81 0.16 0.24
BB 0.03 0.06 0.44 6.66 83.23 7.46 1.05 1.08
B 0.00 0.10 0.32 0.46 5.72 83.62 3.84 5.94
CCC 0.15 0.00 0.29 0.88 1.91 10.28 61.23 25.26
• It neglects crosscountry diﬀerences and business cycle eﬀects.
• Rating agencies react too slow to change ratings. There is a systematic
overestimation of tr
R,R
and d
R
, and hence underestimation of tr
R,R
for
some R = R
t
.
13.1.2 Structural Approach
Merton [18] proposed a simple capital structure of a ﬁrm consisting of equity
and one type of zero coupon debt with promised terminal constant payoﬀ
X > 0 at maturity T. The obligor (=the ﬁrm) defaults by T if the total
market value of its assets V (T) at T is less than its liabilities X. Thus the
probability of default by time T conditional on the information available at
t ≤ T is
p
d
(t, T) = P[V (T) < X [ T
t
] ,
with respect to some stochastic basis (Ω, T, (T
t
)
t∈[0,T]
, P). The dynamics of
V (t) is modelled as geometric Brownian motion
dV (t)
V (t)
= µdt +σ dW(t), t ∈ [0, T],
13.1. TRANSITION AND DEFAULT PROBABILITIES 149
that is
V (T) = V (t) exp
σ(W(T) −W(t)) +
µ −
1
2
σ
2
(T −t)
, t ∈ [0, T].
Then we have
p
d
(t, T) = Φ
¸
log
X
V (t)
−
µ −
1
2
σ
2
(T −t)
σ
√
T −t
, t ∈ [0, T].
If the ﬁrm value process V (t) is continuous, as in the Merton approach,
the instantaneous probability of default (∂
+
T
p
d
(t, T)[
T=t
) is zero. To include
“unexpected” defaults one has to consider ﬁrm value processes with jumps.
Zhou (1997) models V (t) as jumpdiﬀusion process
V (T) = V (t)
¸
N(T)
¸
j=N(t)+1
e
Z
j
e
µ−
σ
2
2
(T−t)+σ(W(T)−W(t))
,
where N(t) is a Poisson process with intensity λ and Z
1
, Z
2
, . . . is a sequence
of i.i.d. Gaussian ^(m, ρ
2
) distributed random variables. It is assumed that
W, N and Z
j
are mutually independent. A dynamic description of V is
V (t) = V (0) +
t
0
V (s) (µds +σ dW(s)) +
N(t)
¸
j=1
V (τ
j
−)
e
Z
j
−1
,
where τ
1
, τ
2
, . . . are the jump times of N.
It is clear that the distribution of log V (T) conditional on T
t
and N(T) −
N(t) = n is Gaussian with mean
log V (t) +mn +
µ −
σ
2
2
(T −t)
and variance
nρ
2
+σ
2
(T −t).
150 CHAPTER 13. DEFAULT RISK
Hence the conditional default probability
p
d
(t, T) = P[log V (T) < log X [ T
t
]
=
∞
¸
n=0
P[log V (T) < log X [ T
t
, N(T) −N(t) = n] P[N(T) −N(t) = n]
=
∞
¸
n=0
Φ
¸
log
X
V (t)
−mn −
µ −
σ
2
2
(T −t)
nρ
2
+σ
2
(T −t)
e
−λ(T−t)
(λ(T −t))
n
n!
First passage time models make this approach more realistic by admitting
default at any time T
d
∈ [0, T], and not just at maturity T. That means,
bankruptcy occurs if the ﬁrm value V (t) hits a speciﬁed stochastic boundary
X(t), such that
T
d
= inf¦t [ V (t) ≤ X(t)¦.
In this case the conditional default probability is
p
d
(t, T) = P[T
d
≤ T [ T
t
] , t ∈ [0, T],
which can be determined by Monte Carlo simulation.
13.2 Intensity Based Method
Default is often a complicated event. The precise conditions under which
it must occur (such as hitting a barrier) are easily misspeciﬁed. The above
structural approach has the additional deﬁciency that it is usually diﬃcult
to determine and trace a ﬁrm’s value process.
In this section we focus directly on describing the evolution of the default
probabilities p
d
(t, T) without deﬁning the exact default event. Formally, we
ﬁx a probability space (Ω, T, P). The ﬂow of the complete market information
is represented by a ﬁltration (T
t
) satisfying the usual conditions. The default
time T
d
is assumed to be an (T
t
)stopping time, hence the rightcontinuous
default process
H(t) := 1
T
d
≤t¦
is (T
t
)adapted. The T
t
conditional default probability is now
p
d
(t, T) = E[H(T) [ T
t
] , t ∈ [0, T].
13.2. INTENSITY BASED METHOD 151
Obviously, H is a uniformly integrable submartingale. By the Doob–Meyer
decomposition ([14, Theorem 1.4.10]) there exists a unique (T
t
)predictable
1
increasing process A(t) such that
M(t) := H(t) −A(t)
is a (uniformly integrable) martingale (notice that A(t) = A(t ∧T
d
)). Hence
p
d
(t, T) = 1
T
d
≤t¦
+E[A(T) −A(t) [ T
t
] .
This formula is the best we can hope for in general.
We proceed in several steps towards an explicit expression for p
d
(t, T) by
making more and more restrictive assumptions (D1)–(D4).
(D1) There exists a strict subﬁltration ((
t
) ⊂ (T
t
) (partial market infor
mation) and a ((
t
)adapted process Λ such that
A(t) = Λ(t ∧ T
d
) and T
t
= (
t
∨ H
t
,
where H
t
:= σ(H(s) [ s ≤ t) and (
t
∨ H
t
stands for the smallest
σalgebra containing (
t
and H
t
.
A market participant with access to the partial market information (
t
cannot
observe whether default has occurred by time t (T
d
≤ t) or not (T
d
> t). In
other words, T
d
is not a stopping time for ((
t
). This nicely reﬂects the
aforementioned diﬃculties to determine the exact default event in practice.
Intuitively speaking, events in T
t
are (
t
observable given that T
d
> t.
The formal statement is as follows.
Lemma 13.2.1. Let t ∈ R
+
. For every A ∈ T
t
there exists B ∈ (
t
such that
A ∩ ¦T
d
> t¦ = B ∩ ¦T
d
> t¦. (13.1)
Proof. Let
T
∗
t
:= ¦A ∈ T
t
[ ∃B ∈ (
t
with property (13.1)¦ .
1
The (F
t
)predictable σalgebra on R
+
× Ω is generated by all leftcontinuous (F
t
)
adapted processes; or equivalently, by the sets {0} ×B where B ∈ F
0
and (s, t] ×B where
s < t and B ∈ F
s
.
152 CHAPTER 13. DEFAULT RISK
Clearly (
t
⊂ T
∗
t
. Simply take B = A. Moreover H
t
⊂ T
∗
t
. Indeed, for every
A ∈ H
t
the intersection A∩ ¦T
d
> t¦ is either ∅ or ¦T
d
> t¦, so we can take
for B either ∅ or Ω.
Since T
∗
t
is a σalgebra (→ exercise) and T
t
is deﬁned to be the smallest
σalgebra containing (
t
and H
t
, we conclude that T
t
⊂ T
∗
t
. This proves the
lemma.
(D2) The default probability by t as seen by a (
t
informed observer satisﬁes
0 < P[T
d
≤ t [ (
t
] < 1.
Hence we can deﬁne the positive ((
t
)adapted hazard process Γ by
e
−Γ(t)
:= P[T
d
> t [ (
t
] .
Notice that X(t) := P[T
d
> t [ (
t
] is a ((
t
)supermartingale and E[X(t)] is
rightcontinuous in t (→ exercise). Hence X(t), and thus Γ(t), admits a
rightcontinuous modiﬁcation, see e.g. [14, Theorem I.3.13]. We show below
(Lemma 13.2.4) the rather surprising fact that if Γ is regular enough then it
coincides with Λ on [0, T
d
].
A consequence of the next lemma is that for any T
t
measurable random
variable Y there exists an (
t
measurable random variable
˜
Y such that Y =
˜
Y
on ¦T
d
> t¦.
Lemma 13.2.2. Let t ∈ R
+
and Y a random variable. Then
E
1
T
d
>t¦
Y [ T
t
= 1
T
d
>t¦
e
Γ(t)
E
1
T
d
>t¦
Y [ (
t
. (13.2)
Proof. Let A ∈ T
t
. By Lemma 13.2.1 there exists a B ∈ (
t
with (13.1), and
so 1
A
1
T
d
>t¦
= 1
B
1
T
d
>t¦
. Hence, by the very deﬁnition of the (
t
conditional
expectation,
A
1
T
d
>t¦
Y P[T
d
> t [ (
t
] dP =
B
1
T
d
>t¦
Y P[T
d
> t [ (
t
] dP
=
B
E
1
T
d
>t¦
Y [ (
t
P[T
d
> t [ (
t
] dP
=
B
1
T
d
>t¦
E
1
T
d
>t¦
Y [ (
t
dP
=
A
1
T
d
>t¦
E
1
T
d
>t¦
Y [ (
t
dP.
13.2. INTENSITY BASED METHOD 153
This implies
E
1
T
d
>t¦
Y P[T
d
> t [ (
t
] [ T
t
= 1
T
d
>t¦
E
1
T
d
>t¦
Y [ (
t
,
which proves the lemma.
As a consequence of the preceding lemmas we may now formulate the
following results, which contain an expression for the aforementioned default
probabilities.
Lemma 13.2.3. For any t ≤ T we have
P[T
d
> T [ T
t
] = 1
T
d
>t¦
E
e
Γ(t)−Γ(T)
[ (
t
, (13.3)
P[t < T
d
≤ T [ T
t
] = 1
T
d
>t¦
E
1 −e
Γ(t)−Γ(T)
[ (
t
. (13.4)
Moreover, the processes
L(t) := 1
T
d
>t¦
e
Γ(t)
= (1 −H(t))e
Γ(t)
is an (T
t
)martingale.
Proof. Let t ≤ T. Then 1
T
d
>T¦
= 1
T
d
>t¦
1
T
d
>T¦
. Using this and (13.2) we
derive
P[T
d
> T [ T
t
] = E
1
T
d
>t¦
1
T
d
>T¦
[ T
t
= 1
T
d
>t¦
e
Γ(t)
E
1
T
d
>T¦
[ (
t
= 1
T
d
>t¦
e
Γ(t)
E
E
1
T
d
>T¦
[ (
T
[ (
t
= 1
T
d
>t¦
e
Γ(t)
E
e
−Γ(T)
[ (
t
,
which proves (13.3). Equation (13.4) follows since
1
t<T
d
≤T¦
= 1
T
d
>t¦
−1
T
d
>T¦
.
For the second statement it is enough to consider
E[L(T) [ T
t
] = E
1
T
d
>t¦
1
T
d
>T¦
e
Γ(T)
[ T
t
= 1
T
d
>t¦
e
Γ(t)
E
1
T
d
>T¦
e
Γ(T)
[ (
t
= L(t),
since by deﬁnition of Γ
E
1
T
d
>T¦
e
Γ(T)
[ (
t
= E
E
1
T
d
>T¦
[ (
T
e
Γ(T)
[ (
t
= 1.
154 CHAPTER 13. DEFAULT RISK
(D3) There exists a positive, measurable, ((
t
)adapted process λ such that
Γ(t) =
t
0
λ(s) ds.
Taking (formally) the righthand Tderivative at T = t in (13.4) gives λ(t).
Hence we refer to λ(t) as default intensity.
Here is the announced result for Γ.
Lemma 13.2.4. The process
N(t) := H(t) −
t
0
λ(s)1
T
d
>s¦
ds
is an (T
t
)martingale. Hence, by the uniqueness of the predictable Doob–
Meyer decomposition, we have
Λ(t ∧ T
d
) =
t
0
λ(s)1
T
d
>s¦
ds = Γ(t ∧ T
d
).
Proof. Let t ≤ T. In view of (13.3) we have
E[N(T) [ T
t
] = 1 −E
1
T
d
>T¦
[ T
t
−
t
0
λ(s)1
T
d
>s¦
ds
−
T
t
E
λ(s)1
T
d
>s¦
[ T
t
ds
= 1 −1
T
d
>t¦
E
e
−
T
t
λ(u) du
[ (
t
−
t
0
λ(s)1
T
d
>s¦
ds
−
T
t
1
T
d
>t¦
e
t
0
λ(u) du
E
λ(s)1
T
d
>s¦
[ (
t
ds
. .. .
=:I
.
We have further
I =
T
t
1
T
d
>t¦
e
t
0
λ(u) du
E
λ(s)E
1
T
d
>s¦
[ (
s
[ (
t
ds
= 1
T
d
>t¦
E
¸
T
t
λ(s)e
−
s
t
λ(u) du
ds [ (
t
= 1
T
d
>t¦
E
1 −e
−
T
t
λ(u) du
[ (
t
,
13.2. INTENSITY BASED METHOD 155
hence
E[N(T) [ T
t
] = 1 −1
T
d
>t¦
−
t
0
λ(s)1
T
d
>s¦
ds = N(t).
The next and last assumption leads the way to implement a default risk
model.
(D4) P[T
d
> t [ (
∞
] = P[T
d
> t [ (
t
] ∀t ∈ R
+
.
It can be shown that (D4) is equivalent to the hypothesis
(H) Every square integrable ((
t
)martingale is an (T
t
)martingale.
For more details we refer to [1, Chapter 6]. For the next lemma we only
assume (D1), (D2) and (D4).
Lemma 13.2.5. Suppose Γ is continuous. Then φ := Γ(T
d
) is an exponential
random variable with parameter 1 and independent of (
∞
. Moreover,
T
d
= inf ¦t [ Γ(t) ≥ φ¦ .
Proof. By assumption,
P[T
d
> t [ (
∞
] = e
−Γ(t)
.
Hence Γ(t) is nondecreasing and continuous. We can deﬁne its right inverse
C(s) := inf¦t [ Γ(t) > s¦.
Then Γ(t) > s ⇔t > C(s) and Γ(C(s)) = s, so
P[Γ(T
d
) > s [ (
∞
] = P[T
d
> C(s) [ (
∞
] = e
−Γ(C(s))
= e
−s
.
This proves that φ = Γ(T
d
) is an exponential random variable with parameter
1 and independent of (
∞
. Moreover,
T
d
= inf¦t [ Γ(t) ≥ Γ(T
d
)¦ = inf¦t [ Γ(t) ≥ φ¦.
156 CHAPTER 13. DEFAULT RISK
13.2.1 Construction of Intensity Based Models
The construction of a model that satisﬁes (D1)–(D4) is straightforward.
We start with a ﬁltration ((
t
) satisfying the usual conditions and
(
∞
= σ((
t
[ t ∈ R
+
) ⊂ T.
Let λ(t) be a positive, measurable, ((
t
)adapted process with the property
t
0
λ(s) ds < ∞ a.s. for all t ∈ R
+
.
We then ﬁx an exponential random variable φ with parameter 1 and inde
pendent of (
∞
, and deﬁne the random time
T
d
:= inf
t [
t
0
λ(s) ds ≥ φ
with values in (0, ∞]. Consequently, we have for t ≤ T
P[T
d
> T [ (
t
] = P
¸
φ >
T
0
λ(u) du [ (
t
= E
¸
P
¸
φ >
T
0
λ(u) du [ (
T
[ (
t
= E
e
−
T
0
λ(u) du
[ (
t
,
by the independence of φ and (
T
(this is a basic lemma for conditional
expectations). And it is an easy exercise to show that
0 < P[T
d
> t [ (
t
] = e
−
t
0
λ(u) du
< 1 and φ =
T
d
0
λ(u) du.
We ﬁnally deﬁne T
t
:= (
t
∨ H
t
, where H
t
= σ(H(s) [ s ≤ t). Conditions
(D1)–(D3) are obviously satisﬁed for
Λ(t) = Γ(t) :=
t
0
λ(s) ds.
As for (D4) we notice that
P[T
d
> t [ (
∞
] = P
¸
φ >
t
0
λ(u) du [ (
∞
= e
−
t
0
λ(u) du
= P[T
d
> t [ (
t
] .
13.2. INTENSITY BASED METHOD 157
13.2.2 Computation of Default Probabilities
When it comes to computations of the default probabilities (13.3) we need a
tractable model for the intensity process λ. But the righthand side of (13.3)
looks just like what we had for the riskneutral valuation of zerocoupon
bonds in terms of a given short rate process (Chapter 7). Notice that λ ≥ 0
is essential. An obvious and popular choice for λ is thus a square root (or
aﬃne) process. So let W be a ((
t
)Brownian motion, b ≥ 0, β ∈ R and σ > 0
some constants, and let
dλ(t) = (b +βλ(t)) dt +σ
λ(t) dW(t), λ(0) ≥ 0. (13.5)
The proof of the following lemma is left as an exercise.
Lemma 13.2.6. For the intensity process (13.5) the conditional default prob
ability is
p
d
(t, T) = P[T
d
≤ T [ T
t
] =
1 −e
−A(T−t)−B(T−t)λ(t)
, if T
d
> t
0, else,
where
A(u) := −
2b
σ
2
log
2γe
(γ−β)u/2
(γ −β) (e
γu
−1) + 2γ
,
B(u) :=
2 (e
γu
−1)
(γ −β) (e
γu
−1) + 2γ
,
γ :=
β
2
+ 2σ
2
.
13.2.3 Pricing Default Risk
The stochastic setup is as above. In addition, we suppose now that we are
given a riskneutral probability measure Q ∼ P and a measurable, ((
t
)
adapted short rate process r(t). Moreover, we assume that there exists a
positive, measurable, ((
t
)adapted process λ
Q
such that
Γ
Q
(t) :=
t
0
λ
Q
(s) ds < ∞ a.s. for all t ∈ R
+
,
and (D1)–(D3) are satisﬁed for Q, Λ
Q
:= Γ
Q
and Γ
Q
replacing P, Λ and
Γ, respectively (unfortunately, these conditions are not preserved under an
158 CHAPTER 13. DEFAULT RISK
equivalent change of measure in general). So that Lemmas 13.2.1–13.2.4
apply.
We will determine the price C(t, T) of a corporate zerocoupon bond with
maturity T, which may default. As for the recovery we ﬁx a constant recovery
rate δ ∈ (0, 1) and distinguish three cases:
• Zero recovery: the cashﬂow at T is 1
T
d
>T¦
.
• Partial recovery at maturity: the cashﬂow at T is 1
T
d
>T¦
+δ1
T
d
≤T¦
.
• Partial recovery at default: the cashﬂow is
1 at T if T
d
> T,
δ at T
d
if T
d
≤ T.
ZeroRecovery
The arbitrage price of C(t, T) is
C(t, T) = E
Q
e
−
T
t
r(s) ds
1
T
d
>T¦
[ T
t
.
In view of Lemma 13.2.2 this is
C(t, T) = 1
T
d
>t¦
e
t
0
λ
Q
(s) ds
E
Q
e
−
T
t
r(s) ds
1
T
d
>T¦
[ (
t
= 1
T
d
>t¦
e
t
0
λ
Q
(s) ds
E
Q
e
−
T
t
r(s) ds
E
Q
1
T
d
>T¦
[ (
T
[ (
t
= 1
T
d
>t¦
E
Q
e
−
T
t
(r(s)+λ
Q
(s))ds
[ (
t
.
(13.6)
Notice that this is a very nice formula. Pricing a corporate bond boils down
to the pricing of a nondefaultable zerocoupon bond with the short rate
process replaced by
r(s) + λ
Q
(s) ≥ r(s).
A tractable (hence aﬃne) model is easily found. For the short rates we chose
CIR: let W be a (Q, (
t
)Brownian motion, b ≥ 0, β ∈ R, σ > 0 constant
parameters and
dr(t) = (b +βr(t)) dt +σ
r(t) dW(t), r(0) ≥ 0. (13.7)
For the intensity process we chose the aﬃne combination
λ
Q
(t) = c
0
+c
1
r(t), (13.8)
for two constants c
0
, c
1
≥ 0.
13.2. INTENSITY BASED METHOD 159
Lemma 13.2.7. For the above aﬃne model we have
C(t, T) = 1
T
d
>t¦
e
−A(T−t)−B(T−t)r(t)
,
where
A(u) := c
0
u −
2b(1 +c
1
)
σ
2
log
2γe
(γ−β)u/2
(γ −β) (e
γu
−1) + 2γ
,
B(u) :=
2 (e
γu
−1)
(γ −β) (e
γu
−1) + 2γ
(1 +c
1
),
γ :=
β
2
+ 2(1 +c
1
)σ
2
.
Proof. Exercise.
A special case is c
1
= 0 (constant intensity). Here we have
C(t, T) = 1
T
d
>t¦
e
−c
0
(T−t)
P(t, T),
where P(t, T) is the CIR price of a defaultfree zerocoupon bond.
Partial Recovery at Maturity
This is an easy modiﬁcation of the preceding case since
1
T
d
>T¦
+δ1
T
d
≤T¦
= (1 −δ) 1
T
d
>T¦
+δ.
In view of (13.6) hence
C(t, T) = (1 −δ) 1
T
d
>t¦
E
Q
e
−
T
t
(r(s)+λ
Q
(s))ds
[ (
t
+δP(t, T),
where P(t, T) stands for the price of the defaultfree zerocoupon bond.
Partial Recovery at Default
A straightforward modiﬁcation of the proofs of Lemmas 13.2.1 and 13.2.2
shows that
E
Q
1
T
d
>t¦
Y [ (
∞
∨ H
t
= 1
T
d
>t¦
e
t
0
λ
Q
(s) ds
E
Q
1
T
d
>t¦
Y [ (
∞
160 CHAPTER 13. DEFAULT RISK
for every random variable Y . Combining this with Section 13.2.1 we obtain
for t ≤ u
Q[t < T
d
≤ u [ (
∞
∨ H
t
] = 1
T
d
>t¦
e
t
0
λ
Q
(s) ds
E
Q
1
t<T
d
≤u¦
[ (
∞
= 1
T
d
>t¦
e
t
0
λ
Q
(s) ds
e
−
t
0
λ
Q
(s) ds
−e
−
u
0
λ
Q
(s) ds
= 1
T
d
>t¦
1 −e
−
u
t
λ
Q
(s) ds
,
which is the regular conditional distribution of T
d
conditional on ¦T
d
> t¦
and (
∞
∨ H
t
. Diﬀerentiation in with respect to u yields its density function
1
T
d
>t¦
λ
Q
(u)e
−
u
t
λ
Q
(s) ds
1
t≤u¦
.
Hence the arbitrage price of the recovery at default given that t < T
d
≤ T
is given by
π(t) = E
Q
e
−
T
d
t
r(s) ds
δ1
t<T
d
≤T¦
[ T
t
= E
Q
E
Q
e
−
T
d
t
r(s) ds
δ1
t<T
d
≤T¦
[ (
∞
∨ H
t
[ T
t
= δ1
T
d
>t¦
E
Q
¸
T
t
e
−
u
t
r(s) ds
λ
Q
(u)e
−
u
t
λ
Q
(s) ds
du [ T
t
= δ1
T
d
>t¦
T
t
E
Q
λ
Q
(u)e
−
u
t
(r(s)+λ
Q
(s)) ds
[ T
t
du.
For the above aﬃne model (13.7)–(13.8) this expression can be made more
explicit (→exercise). As a result, the price of the corporate bond bond price
with recovery at default is
C(t, T) = C
0
(t, T) +π(t),
where C
0
(t, T) is the bond price with zero recovery.
The above calculations and an extension to stochastic recovery go back
to Lando [16].
13.2.4 Measure Change
We consider an equivalent change of measure and derive the behavior of the
compensator process for the stopping time T
d
. Again, we take the above
13.2. INTENSITY BASED METHOD 161
stochastic setup and let (D1)–(D3) hold. So that
M(t) = H(t) −
t
0
λ(s)1
T
d
>s¦
ds
is a (P, T
t
)martingale. Now let µ be a positive ((
t
)predictable process such
that
Λ
Q
(t) :=
t
0
µ(s)λ(s) ds < ∞ a.s. for all t ∈ R
+
.
We will construct an equivalent probability measure Q ∼ P such that
Λ
Q
(t ∧ T
d
)
is the (Q, T
t
)compensator of H. This does not, however, imply that Λ
Q
(t)
is the (Q, (
t
)hazard process Γ
Q
(t) = −log Q[T
d
> t [ (
t
] of T
d
in general. A
counterexample has been constructed by Kusuoka [15], see also [1, Section
7.3].
The following analysis involves stochastic calculus for cadlag processes of
ﬁnite variation (FV), which in a sense is simpler than for Brownian motion
since it is a pathwise calculus. We recall the integration by parts formula for
two rightcontinuous FV functions f and g
f(t)g(t) = f(0)g(0) +
t
0
f(s−) dg(s) +
t
0
g(s−) df(s) + [f, g](t),
where
[f, g](t) =
¸
0<s≤t
∆f(s)∆g(s), ∆f(s) := f(s) −f(s−).
Lemma 13.2.8. The process
D(t) := C(t)V (t)
with
C(t) := exp
t
0
(1 −µ(s))λ(s)1
T
d
>s¦
ds
V (t) :=
1
T
d
>t¦
+µ(T
d
)1
T
d
≤t¦
=
1, t < T
d
µ(T
d
), t ≥ T
d
162 CHAPTER 13. DEFAULT RISK
satisﬁes
D(t) = 1 +
t
0
D(s−) (µ(s) −1) dM(s)
and is thus a positive Plocal martingale.
Proof. Notice that [C, V ] = 0 and
V (t) = 1 +
t
0
(µ(s) −1) dH(s) = 1 +
t
0
V (s−) (µ(s) −1) dH(s).
Hence
D(t) = 1 +
t
0
C(s−) dV (s) +
t
0
V (s−) dC(s)
= 1 +
t
0
C(s−)V (s−) (µ(s) −1) dH(s)
+
t
0
C(s)V (s−)(1 −µ(s))λ(s)1
T
d
>s¦
ds
= 1 +
t
0
D(s−) (µ(s) −1) dM(s).
Since D(s−) is locally bounded and Λ
Q
(t) < ∞ we conclude that D is a
Plocal martingale.
Lemma 13.2.9. Let T ∈ R
+
. Suppose E[D(T)] = 1 (hence (D(t))
t∈[0,T]
is a
martingale), so that we can deﬁne an equivalent probability measure Q ∼ P
on T
T
by
dQ
dP
= D(T).
Then the process
M
Q
(t) := H(t) −Λ
Q
(t ∧ T
d
), t ∈ [0, T], (13.9)
is a Qmartingale.
Proof. It is enough to show that M
Q
is a Qlocal martingale. Indeed, Λ
Q
is an increasing continuous (and hence predictable) process, (13.9) therefore
the unique Doob–Meyer decomposition of H under Q. Since H is uniformly
integrable, so is M
Q
([14, Theorem 1.4.10]).
13.2. INTENSITY BASED METHOD 163
From Bayes’ rule we know that M
Q
is a Qlocal martingale if and only if
DM
Q
is a Plocal martingale. Notice that
[D, M
Q
](t) = ∆D(T
d
)1
T
d
≥t¦
= D(T
d
−) (µ(T
d
) −1) 1
T
d
≥t¦
=
t
0
D(s−) (µ(s) −1) dH(s).
Integration by parts gives
DM
Q
(t) =
t
0
D(s−) dM
Q
(s) +
t
0
M
Q
(s−) dD(s) + [D, M
Q
](t)
=
t
0
D(s−) dH(s) −
t
0
D(s−)µ(s)λ(s)1
T
d
>s¦
ds
+
t
0
M
Q
(s−) dD(s) +
t
0
D(s−) (µ(s) −1) dH(s)
=
t
0
M
Q
(s−) dD(s) +
t
0
D(s−)µ(s) dM(s),
which proves the claim.
Pricing by the “Martingale Approach”
We remark again that Λ
Q
is diﬀerent from Γ
Q
in general, so that the methods
from Section 13.2.3 do not apply for the above situation. Yet, there is a way
to derive the pricing formulas from Section 13.2.3 under Assumption (D4)
for Q. The detailed analysis can be found in [1, Section 8.3].
164 CHAPTER 13. DEFAULT RISK
Bibliography
[1] T. R. Bielecki and M. Rutkowski, Credit risk: modelling, valuation and
hedging, Springer Finance, SpringerVerlag, Berlin, 2002.
[2] BIS, Zerocoupon yield curves: Technical documentation, Bank for In
ternational Settlements, Basle, March 1999.
[3] T. Bj¨ ork, Arbitrage theory in continuous time, Oxford University Press,
1998.
[4] T. Bj¨ ork, Y. Kabanov, and W. Runggaldier, Bond market structure in
the presence of marked point processes, Math. Finance 7 (1997), no. 2,
211–239.
[5] A. Brace, D. G¸ atarek, and M. Musiela, The market model of interest
rate dynamics, Math. Finance 7 (1997), no. 2, 127–155.
[6] D. Brigo and F. Mercurio, Interest rate models—theory and practice,
SpringerVerlag, Berlin, 2001.
[7] J. Cox, J. Ingersoll, and S. Ross, A theory of the term structure of
interest rates, Econometrica 53 (1985), 385–408.
[8] D. Filipovi´c, Consistency problems for Heath–Jarrow–Morton interest
rate models, SpringerVerlag, Berlin, 2001.
[9] D. Heath, R. Jarrow, and A. Morton, Bond pricing and the term struc
ture of interest rates: A new methodology for contingent claims valua
tion, Econometrica 60 (1992), 77–105.
[10] J. C. Hull, Options, futures, and other derivatives, fourth ed., Prentice
Hall International, Inc., 2000.
165
166 BIBLIOGRAPHY
[11] J. James and N. Webber, Interest rate modelling, Wiley, 2000.
[12] R. Jamshidian, Libor and swap market models and measures, Finance
and Stochastics 1 (1997), 290–330.
[13] R. Jarrow, Modelling ﬁxed income securities and interest rate options,
McGrawHill, 1996.
[14] I. Karatzas and S. E. Shreve, Brownian motion and stochastic calculus,
second ed., SpringerVerlag, New York, 1991.
[15] S. Kusuoka, A remark on default risk models, Advances in mathematical
economics, Vol. 1, Springer, Tokyo, 1999, pp. 69–82.
[16] D. Lando, On Cox processes and creditrisky securities, Rev. Derivatives
Res. 2 (1998), 99–120.
[17] R. Litterman and J.A. Scheinkman, Common factors aﬀecting bond re
turns, Journal of Fixed Income 1 (1991), 54–61.
[18] R. C. Merton, On the pricing for corporate debt: the risk structure of
interest rates, J. Finance 29 (1974), 449–470.
[19] M. Musiela and M. Rutkowski, Martingale methods in ﬁnancial mod
elling, Applications of Mathematics, vol. 36, SpringerVerlag, Berlin
Heidelberg, 1997.
[20] C. Nelson and A. Siegel, Parsimonious modeling of yield curves, J. of
Business 60 (1987), 473–489.
[21] C. R. Rao, Principal component and factor analyses, Statistical methods
in ﬁnance, NorthHolland, Amsterdam, 1996, pp. 489–505.
[22] R. Rebonato, Interestrate option models, second ed., Wiley, 1998.
[23] D. Revuz and M. Yor, Continuous martingales and Brownian motion,
Grundlehren der mathematischen Wissenschaften, vol. 293, Springer
Verlag, BerlinHeidelbergNew York, 1994.
[24] Bernd Schmid, Pricing credit linked ﬁnancial instruments, Lecture Notes
in Economics and Mathematical Systems, vol. 516, SpringerVerlag,
Berlin, 2002, Theory and empirical evidence.
BIBLIOGRAPHY 167
[25] J. Michael Steele, Stochastic calculus and ﬁnancial applications, Appli
cations of Mathematics, vol. 45, SpringerVerlag, New York, 2001.
[26] L. E. O. Svensson, Estimating and interpreting forward interest rates:
Sweden 19921994, IMF Working Paper No. 114, September 1994.
[27] R. Zagst, Interestrate management, SpringerVerlag, Berlin, 2002.
2
Contents
1 Introduction 2 Interest Rates and Related Contracts 2.1 ZeroCoupon Bonds . . . . . . . . . . . . . . . . . . . 2.2 Interest Rates . . . . . . . . . . . . . . . . . . . . . . 2.2.1 Market Example: LIBOR . . . . . . . . . . . 2.2.2 Simple vs. Continuous Compounding . . . . . 2.2.3 Forward vs. Future Rates . . . . . . . . . . . 2.3 Bank Account and Short Rates . . . . . . . . . . . . 2.4 Coupon Bonds, Swaps and Yields . . . . . . . . . . . 2.4.1 Fixed Coupon Bonds . . . . . . . . . . . . . . 2.4.2 Floating Rate Notes . . . . . . . . . . . . . . 2.4.3 Interest Rate Swaps . . . . . . . . . . . . . . 2.4.4 Yield and Duration . . . . . . . . . . . . . . . 2.5 Market Conventions . . . . . . . . . . . . . . . . . . . 2.5.1 Daycount Conventions . . . . . . . . . . . . . 2.5.2 Coupon Bonds . . . . . . . . . . . . . . . . . 2.5.3 Accrued Interest, Clean Price and Dirty Price 2.5.4 YieldtoMaturity . . . . . . . . . . . . . . . . 2.6 Caps and Floors . . . . . . . . . . . . . . . . . . . . . 2.7 Swaptions . . . . . . . . . . . . . . . . . . . . . . . . 7 9 9 11 12 12 13 14 15 16 16 17 20 22 22 23 24 25 25 29 33 33 35 36
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . .
3 Statistics of the Yield Curve 3.1 Principal Component Analysis (PCA) . . . . . . . . . . . . . . 3.2 PCA of the Yield Curve . . . . . . . . . . . . . . . . . . . . . 3.3 Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1 T Bond as Numeraire . . . . . . . . . . . .
. .
. . . . . . . . . . . .1 Vasicek Model .2 An Expectation Hypothesis . . . .
.3 Inverting the Yield Curve . . . . .2 Arbitrage and Martingale Measures . .
. . . . . . . . . . . . . . . 7 Short Rate Models 7. .
. . . . .
. . . . .3 Dothan Model . 6. . . . . . .2.6. . . . 4. . . . . . . .2. . . . . . .
. . . .5 Some Standard Models . . . . . . . . α = 0). . . .3 Problems . . .2. 7. . . . . . . . . . . . . . . . . . . . . . . . .
. . . . .3 Hedging and Pricing . . . . . . 7. . . . . . . . β const. . . . .1 Bond Markets . . . 7. .3 Option Pricing in Gaussian HJM Models . . . . . . . .5. .1 Examples . . b.1 Generalities . . . . . . . . . . . 7. . 7. . .2. . . . . . . . . 7. .
. . . . .
. . .
. . . . . . . . . . . . 8 HJM Methodology 9 Forward Measures 9. . . . . . . . . . . . 4. . . . . . . . . 4. . . .
. . .5. .2 Cox–Ingersoll–Ross Model . . . . . . .
. . .2 Diﬀusion Short Rate Models . . . . . . . . 7.
6 NoArbitrage Pricing 6. . . . . . . . . . . . . . . . . . .5 Hull–White Model . . . . . . . . . . . . . . . . . . . . . . .5. . . . . . . . . . . . . . . . .6 Option Pricing in Aﬃne Models . 9. . . . . .4 4 Estimating the Yield Curve 4. . . . .1 Example: Vasicek Model (a. . . . . . . . . .2 Money Markets . . . . . . . . . . . . . . 7. . . . . 4.
. . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . .4 Ho–Lee Model . 7. . . . . 6. . . . . . 7.1 A Bootstrapping Example . . . . .
. . . .
. . . . . . 9. .4 Aﬃne Term Structures . . . . . . . . . . . . . . . . . . .4 Parametrized Curve Families . . . . 5 Why Yield Curve Models?
CONTENTS 39 39 44 45 46 48 49 65 67 67 69 73 77 77 79 82 83 83 85 85 86 87 88 89 90 92 95 97 97 99 101
. . . . . . . . . . 4.
. . 7. . . 7. .2. . . . . . . . . . . . . . . . . . . . . . .5. .1 SelfFinancing Portfolios .2 General Case . . . . . . .
. . . . . . . . . . . . . . . .5. .
. . . .3 Polynomial Term Structures .
. . . 109 113 . .
. . . . . . . . 130 . . . . . 11. . .1 Transition and Default Probabilities . . . . . . .2.
. 129 .4 Measure Change . . . . . . 122 . . . . . . . . . . . . . . . .1. . . . . .4 ExponentialPolynomial Families 11. . . . . 13. . . .1 Discretetenor Case . .1 Historical Method . 106 . 11. . . . . . . . . . . . . 11.1 Forward Contracts . . . . 10. . . . .4.2 Futures Contracts . . . . . . . . . . . . . . . . 156 . . . . . . . . . . . .
. 148 . . . . . . . . . . . . . 10. . . . . . . . . . .
. . . . . . .2 Computation of Default Probabilities . . . . . . . .
.
5 105 . . .
. .1 NoArbitrage Condition . . 145 . . . . . . . . . . . . . . . . . . 13. . . . 118 .
. . . . 12. .
12 Market Models 12. . .2. . . . . . . . . . . . . . . . . . .1. . 160
. . . . . . . . . . . 157 . 12. . . . . . . 140 145 . . 11. . . .4 Forward vs. . . . . . . 13. . . Setup .2. . .CONTENTS 10 Forwards and Futures 10.2. . . 146 . . . . . . . . . . . . . . . 117 . . . . . .2 Continuoustenor Case . . .1 Nelson–Siegel Family . . . . . . . . . . . . . 13 Default Risk 13. . . . . . .1.1.
. . .2 Aﬃne Term Structures . . . . . . . . . .2 Intensity Based Method . 122 . 157 . .
. . .2 Svensson Family . 108 . . . . .3 Pricing Default Risk . . . . . .
. . . . .4. . . . . . . . 123 127 .3 Interest Rate Futures . . . . . 13. 115 . . . . . . . . . . . 10. . . . 105 . .2 Structural Approach . . . . . . . Futures in a Gaussian 11 MultiFactor Models 11. . . . .1 Construction of Intensity Based Models 13. . 13. . . . 13. . . . . . 150 . . . . . .1 Models of Forward LIBOR Rates . . . . .
.
6
CONTENTS
.
• The very nature of ﬁxed income instruments causes diﬃculties. yet it is diﬃcult to ﬁnd a convenient textbook for a onesemester course like this. BM[6]: Brigo–Mercurio (01) [6]. Much emphasis is on the practical implementation and calibration of selected models. 7
. I will frequently refer to the following books: B[3]: Bj¨rk (98) [3]. which usually requires constant (zero) interest rates. These issues should therefore not been left out.Chapter 1 Introduction
These notes have been written for a graduate course on ﬁxed income models that I held in the fall term 2002–2003 at Princeton University. This is a book on interest rate modelling written by two quantitative analysts in ﬁnancial institutions. Chapters 15–20 are on interest rates. A pedagogically well written introduction to matheo matical ﬁnance. other than for stock derivatives. • Interest rate theory is not standardized yet: there is no wellaccepted “standard” general model such as the Black–Scholes model for equities. in implementing and calibrating models. There are several reasons for this: • Until recently. The number of books on ﬁxed income models is growing. many textbooks on mathematical ﬁnance have treated stochastic interest rates as an appendix to the elementary arbitrage pricing theory.
R[22]: Rebonato (98) [22]. MR[19]: Musiela–Rutkowski (97) [19]. Z[27]: Zagst (02) [27].
. interest rate modelling and risk management. J[13]: Jarrow (96) [13]. A comprehensive textbook on mathematical ﬁnance. Discrete time only. An encyclopedic treatment of interest rates and their related ﬁnancial derivatives. Introduction to ﬁxedincome securities and interest rate options. INTRODUCTION
JW[11]: James–Webber (00) [11].8
CHAPTER 1. Much emphasis is on market pricing practice. Much emphasis on market practice for pricing and handling interest rate derivatives. Written by a practitionar. A comprehensive book on ﬁnancial mathematics with a large part (Part II) on interest rate modelling. I did not intend to write an entire text but rather collect fragments of the material that can be found in the above books and further references.
t1 ) + · · · + Cn−1 P (t. 9
. tn−1 ) + (1 + Cn )P (t. T ) is continuously diﬀerentiable in T . and many more
2.1
ZeroCoupon Bonds
A dollar today is worth more than a dollar tomorrow. for briefty also T bond . T ) = 1 for all T .
P(t. • P (T. T ). T ). The time t value of a dollar at time T ≥ t is expressed by the zerocoupon bond with maturity T .Chapter 2 Interest Rates and Related Contracts
Literature: B[3](Chapter 15). In theory we will assume that • there exists a frictionless market for T bonds for every T > 0. such as couponbearing bonds C1 P (t.T)  t 1  T
→ future cashﬂows can be discounted. • P (t. P (t. BM[6](Chapter 1). This is a contract which guarantees the holder one dollar to be paid at the maturity date T .
INTEREST RATES AND RELATED CONTRACTS
In reality this assumptions are not always satisﬁed: zerocoupon bonds are not traded for all maturities. T ) ≤ 1 is a decreasing curve (which is equivalent to positivity of interest rates).10
CHAPTER 2.
. T ) are not known with certainty before t. this is a good starting point for doing the mathematics. March 2002 1 0.6 0. However. More realistic models will be introduced and discussed in the sequel.4 0.8 0. Yet.8 0. T ) is a smooth curve.
US Treasury Bonds. T ) is a stochastic process since bond prices P (t. The third condition is purely technical and implies that the term structure of zerocoupon bond prices T → P (t. T ) might be less than one if the issuer of the T bond defaults.
PHt. and P (T.10L 1 0. Therefore we leave away this requirement.2 t 1 2 3 4 5 6 7 8 9 10
A reasonable assumption would also be that T → P (t. already classical interest rate models imply zerocoupon bond prices greater than 1.6 0.4 0.2 Years 1 2 3 4 5 6 7 8 9 10
Note that t → P (t.
2. T ) := R(t. T ) . t. • At S: obtain
P (t. S) − log P (t. A prototypical forward rate agreement (FRA) is a contract involving three time instants t < T < S: the current time t.S) P (t.2
Interest Rates
The term structure of zerocoupon bond prices does not contain much visual information (strictly speaking it does). INTEREST RATES
11
2. S] prevailing at t is given by eR(t.T ) P (t. A better measure is given by the implied interest rates. T ) = − log P (t. S)
• The simple spot rate for [t. There is a variety of them. T. T ) := F (t. t. P (t. T )
• The continuously compounded forward rate for [T. We are led to the following deﬁnitions.S)
Sbonds = zero net investment. T ) ⇔ R(t. S) := 1 P (t. the expiry time T > t. T ) −1 .T ) P (t.
P (t. • At t: sell one T bond and buy • At T : pay one dollar. T ) ⇔ F (t.
dollars. S] prevailing at t is given by 1+(S −T )F (t. T.2.S)
The net eﬀect is a forward investment of one dollar at time T yielding dollars at S with certainty.T ) P (t. T ) log P (t. P (t. S) S−T
• The continuously compounded spot rate for [T.T. T ] is F (t.S)(S−T ) := P (t. and the maturity time S > T . T. S) S−T P (t. S) = P (t. S) = − .
• The simple (simplycompounded) forward rate for [T. P (t. T −t
. S] is R(t. T ) = 1 T −t 1 −1 .
It is a mathematical fact that 1+
1
R m
m
→ eR
as m → ∞. the threemonths forward LIBOR for the period [T. T ) = F (t.1
Market Example: LIBOR
“Interbank rates” are rates at which deposits between banks are exchanged. T ) = exp −
f (t.2.
t
2. etc.1)
The function T → f (t. INTEREST RATES AND RELATED CONTRACTS • The instantaneous forward rate with maturity T prevailing at time t is deﬁned by f (t. ranging from overnight to 12 months. T. The most important interbank rate usually considered as a reference for ﬁxed income contracts is the LIBOR (London InterBank Oﬀered Rate) 1 for a series of possible maturities. u) du .12
CHAPTER 2. T ) := lim R(t.
To be more precise: this is the rate at which highcredit ﬁnancial institutions can borrow in the interbank market.
T ↓t
Notice that (2. ∂T
(2. T + 1/4] at time t is given by L(t. T ) . These rates are quoted on a simple compounding basis. and at which swap transactions (see below) between banks occur. For example. If the rate is compounded twice per year the terminal value is (1 + R/2)2 . T. • The instantaneous short rate at time t is deﬁned by r(t) := f (t.1) together with the requirement P (T. Continuous Compounding
One dollar invested for one year at an interest rate of R per annum growths to 1 + R. T ) = 1 is equivalent to
T
P (t.2
Simple vs. T + 1/4).2. S) = −
S↓T
∂ log P (t. T ) is called the forward curve at time t.
.
2. T ). t) = lim R(t.
• At S: pay one dollar.
2.
13
Example: e0. This makes the theory more tractable.04 = 1. This is a pure arbitrage opportunity.2. Future Rates
Can forward rates predict the future spot rates? Consider a deterministic world. Then we follow the strategy: • At t: sell one Sbond.3
Forward vs.2.
This is equivalent to f (t. (Where do we use the assumption of a deterministic world?) The net is a riskless gain of −P (t. S) < 0.2) yields
S S
f (t. S) > P (t. S)+P (t. T )P (T. S) < P (t. eR = 1 + R + o(R) for R small. systematic proﬁt) we have necessarily P (t. ∀t ≤ T ≤ S. Net cost: −P (t. S) for some t ≤ T ≤ S. T )P (T. we often consider continuously compounded interest rates. T )P (T. S) (×1/P (t. If markets are eﬃcient (i. S) = r(S). and buy P (T.2. If P (t. which contradicts the assumption. u) du =
T T
f (T. T )P (T.
∀t ≤ T ≤ S. no arbitrage = no riskless. S) dollars and buy one Sbond. receive one dollar. Suppose that P (t. INTEREST RATES Moreover. • At T : receive P (T. T )P (T. S) = f (T. S) = P (t. u) du. S)). (2. S) the same proﬁt can be realized by changing sign in the strategy. S) + P (t. Since the exponential function has nicer analytic properties than power functions. S). S) T bonds. Taking logarithm in (2.04081.e. ∀t ≤ T ≤ S
.2)
Proof.
14
CHAPTER 2. INTEREST RATES AND RELATED CONTRACTS
(as time goes by we walk along the forward curve: the forward curve is shifted). In this case, the forward rate with maturity S prevailing at time t ≤ S is exactly the future short rate at S. The real world is not deterministic though. We will see that in general the forward rate f (t, T ) is the conditional expectation of the short rate r(T ) under a particular probability measure (forward measure), depending on T . Hence the forward rate is a biased estimator for the future short rate. Forecasts of future short rates by forward rates have little or no predictive power.
2.3
Bank Account and Short Rates
The return of a one dollar investment today (t = 0) over the period [0, ∆t] is given by 1 = exp P (0, ∆t)
∆t
f (0, u) du = 1 + r(0)∆t + o(∆t).
0
Instantaneous reinvestment in 2∆tbonds yields 1 1 = (1 + r(0)∆t)(1 + r(∆t)∆t) + o(∆t) P (0, ∆t) P (∆t, 2∆t) at time 2∆t, etc. This strategy of “rolling over”2 just maturing bonds leads in the limit to the bank account (moneymarket account) B(t). Hence B(t) is the asset which growths at time t instantaneously at short rate r(t) B(t + ∆t) = B(t)(1 + r(t)∆t) + o(∆t). For ∆t → 0 this converges to dB(t) = r(t)B(t)dt and with B(0) = 1 we obtain
t
B(t) = exp
0
2
r(s) ds .
This limiting process is made rigorous in [4].
2.4. COUPON BONDS, SWAPS AND YIELDS
15
B is a riskfree asset insofar as its future value at time t + ∆t is known (up to order ∆t) at time t. In stochastic terms we speak of a predictable process. For the same reason we speak of r(t) as the riskfree rate of return over the inﬁnitesimal period [t, t + dt]. B is important for relating amounts of currencies available at diﬀerent times: in order to have one dollar in the bank account at time T we need to have T B(t) r(s) ds = exp − B(T ) t dollars in the bank account at time t ≤ T . This discount factor is stochastic: it is not known with certainty at time t. There is a close connection to the deterministic (=known at time t) discount factor given by P (t, T ). Indeed, we will see that the latter is the conditional expectation of the former under the risk neutral probability measure.
Proxies for the Short Rate → JW[11](Chapter 3.5) The short rate r(t) is a key interest rate in all models and fundamental to noarbitrage pricing. But it cannot be directly observed. The overnight interest rate is not usually considered to be a good proxy for the short rate, because the motives and needs driving overnight borrowers are very diﬀerent from those of borrowers who want money for a month or more. The overnight fed funds rate is nevertheless comparatively stable and perhaps a fair proxy, but empirical studies suggest that it has low correlation with other spot rates. The best available proxy is given by one or threemonth spot rates since they are very liquid.
2.4
Coupon Bonds, Swaps and Yields
In most bond markets, there is only a relatively small number of zerocoupon bonds traded. Most bonds include coupons.
16
CHAPTER 2. INTEREST RATES AND RELATED CONTRACTS
2.4.1
Fixed Coupon Bonds
A ﬁxed coupon bond is a contract speciﬁed by • a number of future dates T1 < · · · < Tn (the coupon dates) (Tn is the maturity of the bond), • a sequence of (deterministic) coupons c1 , . . . , cn , • a nominal value N , such that the owner receives ci at time Ti , for i = 1, . . . , n, and N at terminal time Tn . The price p(t) at time t ≤ T1 of this coupon bond is given by the sum of discounted cashﬂows
n
p(t) =
i=1
P (t, Ti )ci + P (t, Tn )N.
Typically, it holds that Ti+1 −Ti ≡ δ, and the coupons are given as a ﬁxed percentage of the nominal value: ci ≡ KδN , for some ﬁxed interest rate K. The above formula reduces to
n
p(t) =
Kδ
i=1
P (t, Ti ) + P (t, Tn ) N.
2.4.2
Floating Rate Notes
There are versions of coupon bonds for which the value of the coupon is not ﬁxed at the time the bond is issued, but rather reset for every coupon period. Most often the resetting is determined by some market interest rate (e.g. LIBOR). A ﬂoating rate note is speciﬁed by • a number of future dates T0 < T1 < · · · < Tn , • a nominal value N . The deterministic coupon payments for the ﬁxed coupon bond are now replaced by ci = (Ti − Ti−1 )F (Ti−1 , Ti )N,
. but that the cashﬂow ci is at time Ti . SWAPS AND YIELDS
17
where F (Ti−1 .3
Interest Rate Swaps
An interest rate swap is a scheme where you exchange a payment stream at a ﬁxed rate of interest for a payment stream at a ﬂoating rate (typically LIBOR). Ti ). There are many versions of interest rate swaps. Without loss of generality we set N = 1. The time t value of ci therefore is P (t. Cost: P (t. Ti−1 ). Ti )) = P (t. Ti−1 ): P (Ti−1 . Ti−1 ) − P (t. Zero net investment. for t = T0 : p(T0 ) = 1. The value p(t) of this note at time t ≤ T0 is obtained as follows.2. Ti ) Ti bonds. • At Ti : receive 1/P (Ti−1 . Ti−1 ) − P (t. . Tn ).
In particular.
2. Summing up we obtain the (surprisingly easy) formula
n
p(t) = P (t. Ti ) we then have ci = 1 P (Ti−1 . and we note that F (Ti−1 . A payer interest rate swap settled in arrears is speciﬁed by • a number of future dates T0 < T1 < · · · < Tn with Ti − Ti−1 ≡ δ (Tn is the maturity of the swap). Tn ) +
i=1
(P (t.4. • At Ti−1 : receive one dollar and buy 1/P (Ti−1 . The time t value of 1 paid out at Ti is P (t. By deﬁnition of F (Ti−1 . COUPON BONDS. Ti ) dollars.Ti ) • At t: buy a Ti−1 bond. . Ti ) − 1. T0 ). Ti ). Ti ) is the prevailing simple market interest rate. .4.
The time t value of −1 paid out at Ti is −P (t. Ti ) is determined already at time Ti−1 (this is why here we have T0 in addition to the coupon dates T1 . .
Ti ). INTEREST RATES AND RELATED CONTRACTS • a ﬁxed rate K. Ti−1 .18
CHAPTER 2.
Of course. Ti ) (F (t. . Cashﬂows take place only at the coupon dates T1 . . Ti ) − K)δN. Ti−1 . Tn . • and receives ﬂoating F (Ti−1 . we can rewrite (2. Ti ) i=1
The following alternative representation of Rswap (t) is sometimes useful. The remaining question is how the “fair” ﬁxed rate K is determined. Hence Rswap (t) = P (t. . Ti ) . δ n P (t. Ti−1 ) − P (t. Ti )δN . (2. T0 ) − P (t. Since P (t. the equidistance hypothesis is only for convenience of notation and can easily be relaxed. Ti−1 ) − P (t. Tn ) − Kδ
P (t.3) as N δP (t.
i=1
A receiver interest rate swap settled in arrears is obtained by changing the sign of the cashﬂows at times T1 . The forward swap rate Rswap (t) at time t ≤ T0 is the ﬁxed rate K above which gives Πp (t) = Πr (t) = 0. • a nominal value N . . Its value at time t ≤ T0 is thus Πr (t) = −Πp (t).
. Ti ) = F (t.3) The total value Πp (t) of the swap at time t ≤ T0 is thus
n
Πp (t) = N
P (t. Tn ) . the holder of the contract • pays ﬁxed KδN . . Tn . Ti )δP (t. Ti )). . . Ti ) − K) . T0 ) − P (t. Ti ) − KδP (t. and using the previous results we can compute the value at t ≤ T0 of this cashﬂow as N (P (t. At Ti . The net cashﬂow at Ti is thus (F (Ti−1 . .
P (t. p. and a mediating institution would also make money.248). Ti−1 . • Company B pays the intermediary ﬁxed at 5 5/16% in exchange for LIBOR (payer swap). SWAPS AND YIELDS Summing up yields
n
19
Πp (t) = N δ
i=1
P (t. Tj )
These weights are random. By agreeing to swap streams of cashﬂows both companies could be better oﬀ. Ti ) − K) . Ti ) . Swaps were developed because diﬀerent companies could borrow at different rates in diﬀerent markets. COUPON BONDS. Ti ) (F (t.
. • Company B: is borrowing ﬂoating at LIBOR plus 1%. Ti ). Ti−1 . Ti ). Example → JW[11](p.4. • Company A pays LIBOR to the intermediary in exchange for ﬁxed at 5 3/16% (receiver swap). but there seems to be empirical evidence that the variability of wi (t) is small compared to that of F (t. This is used for approximations of swaption (see below) price formulas in LIBOR market models: the swap rate volatility is written as linear combination of the forward LIBOR volatilities (“Rebonato’s formula” → BM[6]. Ti−1 .
and thus we can write the swap rate as weighted average of simple forward rates n Rswap (t) =
i=1
wi (t)F (t.11) • Company A: is borrowing ﬁxed for ﬁve years at 5 1/2%.2. but could borrow ﬂoating at LIBOR plus 1/2%.
with weights wi (t) =
n j=1
P (t. but could borrow ﬁxed for ﬁve years at 6 1/2%.
and in BM[6] it is a combination of simple spot rates (for maturities up to 1 year) and annually compounded spot rates (for maturities greater than 1 year). This risk has been partly taken up by the intermediary. There is a variety of other terminologies. T ). INTEREST RATES AND RELATED CONTRACTS
• Company A is now paying LIBOR plus 5/16% instead of LIBOR plus 1/2%.
.T )(T −t) . T ) = e−R(t. the function T → R(t. In JW[11] the yield curve is is given by simple spot rates. • Company B is paying ﬁxed at 6 5/16% instead of 6 1/2%. T ) the zerocoupon yield is simply the continuously compounded spot rate R(t. this is why Company B had higher borrowing rates in the ﬁrst place.20 Net:
CHAPTER 2.4
Yield and Duration
For a zerocoupon bond P (t. in return for the money it makes on the spread.4.
2. T ) is referred to as (zerocoupon) yield curve. zerocoupon curve (BM[6]). • The intermediary receives ﬁxed at 1/8%.
5 1/2 % Company A LIBOR 5 3/16 % Intermediary LIBOR 5 5/16 % Company B LIBOR + 1%
Everyone seems to be better oﬀ. etc. The term “yield curve” is ambiguous. such as zerorate curve (Z[27]). That is. Accordingly. But there is implicit credit risk. P (t.
p
.02 Years 1 2 3 4 5 6 7 8 9 10
21
Now let p(t) be the time t market value of a ﬁxed coupon bond with coupon dates T1 < · · · < Tn . .06 0. but • coupon payments at diﬀerent points in time from the same bond are discounted by the same rate. Ti )ci .2.
t ≤ T1 .
n
p(t) =
i=1
P (t.4. COUPON BONDS. y = y(0). . The Macaulay duration of the coupon bond is deﬁned as DM ac :=
n i=1
Ti ci e−yTi .
Again we ask for the bond’s “internal rate of interest”. .21). → R[22](p.4.1). and write p = p(0). cn and nominal value N (see Section 2.
Remark 2. It is argued by Schaefer (1977) that the yieldtomaturity is an inadequate statistics for the bond market: • coupon payments occurring at the same point in time are discounted by diﬀerent discount factors. March 2002 0.08 0. SWAPS AND YIELDS
US Yield Curve. that is. that is. Tn ]) continuously compounded rate which generates the market value of the coupon bond: the (continuously compounded) yieldtomaturity y(t) of this bond at time t ≤ T1 is deﬁned as the unique solution to n p(t) =
i=1
ci e−y(t)(Ti −t) . coupon payments c1 . the constant (over the period [t.1 0.4. etc.04 0. . For simplicity we suppose that cn already contains N .1. To simplify the notation we assume now that t = 0.
A ﬁrst order sensitivity measure of the bond price w. .22
CHAPTER 2. In fact. Ti ) i=1 Ti ci e D := = Ti .r.5. . and it provides us in a certain sense with the “mean time to coupon payment”. p p i=1 with yi := R(0. If t and T denote two dates expressed as day/month/year. INTEREST RATES AND RELATED CONTRACTS
The duration is thus a weighted average of the coupon dates T1 .t. Here are three examples of daycount conventions:
. This is shown by the obvious formula dp d = dy dy
n
ci e−yTi
i=1
= −DM ac p. The market evaluates the year fraction between t and T in diﬀerent ways. Ti ). we have d ds
n
ci e−(yi +s)Ti
i=1
s=0 = −Dp.r. changes in the yieldtomaturity (see Z[27](Chapter 6.
Hence duration is essentially for bonds (w.5
2.1
Market Conventions
Daycount Conventions
Time is measured in years.1.t. . T ) is given by the duration of the bond n n −yi Ti ci P (0. . The bond equivalent of the gamma is convexity: n n d2 −(yi +s)Ti ci e s=0 = ci e−yi Ti (Ti )2 .r. parallel shifts of the entire zerocoupon yield curve T → R(0. As such it is an important concept for interest rate risk management: it acts as a measure of the ﬁrst order sensitivity of the bond price w. it is not clear what T − t should be. Tn .3) for a thourough treatment). C := 2 ds i=1 i=1
2. The daycount convention decides upon the time measurement between two dates t and T .t. parallel shift of the yield curve) what delta is for stock options.
• 30/360: months count 30 and years 360 days. → BM[6](p. Debt securities issued by the U. • Bonds: coupon bonds (semiannual) with time to maturity between 10 and 30 years3 . Treasury are divided into three classes: • Bills: zerocoupon bonds with time to maturity less than one year.1)
2. 5. and the daycount convention for T − t is given by actual number of days between t and T . Z[27](Sect.2). it is important to realize for which daycount convention a speciﬁc interest rate is quoted.4). • Notes: coupon bonds (semiannual) with time to maturity between 2 and 10 years. m2 . 2000 and T =July 4. The daycount convention for T − t is given by min(d2 . 365 • Actual/360: as above but the year counts 360 days.5. Z[27](Sect. y1 ) and T = (d2 .2
Coupon Bonds
→ MR[19](Sect. the issuance of 30 year treasury bonds has been stopped. 30) + (30 − d1 )+ (m2 − m1 − 1)+ + + y2 − y1 . 360 12 Example: The time between t=January 4. 2002 is given by 4 + (30 − 4) 7 − 1 − 1 + + 2002 − 2000 = 2. J[13](Chapter 2) Coupon bonds issued in the American (European) markets typically have semiannual (annual) coupon payments. y2 ). Let t = (d1 . m1 .2. 5. MARKET CONVENTIONS
23
• Actual/365: a year has 365 days.5. 360 12 When extracting information on interest rates from data.
3
Recently.
.S.5.2). 11.
These are the coupons or principal (=nominal) amounts of Treasury bonds trading separately through the Federal Reserve’s bookentry system. or dirty price. This is why prices are diﬀerently quoted at the exchange. .
For t ∈ (T1 .3
Accrued Interest. . Tn and payments c1 . . t ∈ (Ti−1 . They are synthetically created zerocoupon bonds of longer maturities than a year. or clean price. . t).
. INTEREST RATES AND RELATED CONTRACTS
In addition to bills. t). . Ti ] is deﬁned by AI(i. of the coupon bond at time t is pclean (t) := p(t) − AI(i. .
t ≤ T1 . notes and bonds. Treasury securities called STRIPS (separate trading of registered interest and principal of securities) have traded since August 1985.5. Ti ). Ti ]. . . . whenever we buy a coupon bond quoted at a clean price of pclean (t) at time t ∈ (Ti−1 .
etc. . . The accrued interest at time t ∈ (Ti−1 . Tn which is due to the coupon payments.
That is. Ti ). we have to pay is p(t) = pclean (t) + AI(i. t) := ci t − Ti−1 Ti − Ti−1
(where now time diﬀerences are taken according to the daycount convention). The quoted price. the cash price. Ti ]. They were created in response to investor demands. cn the price formula
n
p(t) =
i=1
ci P (t. . T2 ] we have
n
p(t) =
i=2
ci P (t.
2. Hence there are systematic discontinuities of the price trajectory at t = T1 . Clean Price and Dirty Price
Remember that we had for the price of a coupon bond with coupon dates T1 .24
CHAPTER 2.
6
Caps and Floors
→ BM[6](Sect. Z[27](Sect.
. It thus consists of • a number of future dates T0 < T1 < · · · < Tn with Ti − Ti−1 ≡ δ (Tn is the maturity of the cap). T + δ) (e. 1.4
YieldtoMaturity
The quoted (annual) yieldtomaturity y (t) on a Treasury bond at time t = T i ˆ is deﬁned by the relationship pclean (Ti ) = and at t ∈ [Ti . A cap is a strip of caplets. T + δ) − κ)+ . and we assume here that Ti+1 − Ti ≡ 1/2
2. j−i−1+τ (1 + y (t)/2) ˆ (1 + y (t)/2)n−i−1+τ ˆ j=i+1
n
N rc N/2 + .2) Caps A caplet with reset date T and settlement date T + δ pays the holder the diﬀerence between a simple market rate F (T. N the nominal amount and τ = Ti+1 − t Ti+1 − Ti (semiannual coupons).6). Ti+1 ) pclean (t) = rc N/2 N + .5. • a cap rate κ.6.2. 5.g. LIBOR) and the strike rate κ.6. CAPS AND FLOORS
25
2.
is again given by the daycount convention. Its cashﬂow at time T + δ is δ(F (T. j−i (1 + y (Ti )/2) ˆ (1 + y (Ti )/2)n−i ˆ j=i+1
n
where rc is the (annualized) coupon rate.
and
n
Cp(t) =
i=1
Cpl(i. Floors A ﬂoor is the converse to a cap. We write Cpl(i. (1 + δκ) 1 − P (Ti−1 . It can be shown (→ exercise) that the cashﬂow (2. . A ﬂoor is a strip of ﬂoorlets. . Tn . It protects against low rates.
. . Ti ))+ . t) for the price of the ith ﬂoorlet and
n
F l(t) =
i=1
F ll(i. that is. It guarantees that the interest to be paid on a ﬂoating rate loan never exceeds the predetermined cap rate κ. t)
for the price of the ﬂoor. . Ti ) − κ)+ . i = 1. . INTEREST RATES AND RELATED CONTRACTS
Cashﬂows take place at the dates T1 . . which means that caps can be priced very easily in those models. A cap gives the holder a protection against rising interest rates. n. t)
for the time t price of the cap. Write F ll(i. . Ti ) 1 + δκ
+
.4) at time Ti is the equivalent to (1 + δκ) times the cashﬂow at date Ti−1 of a put option on a Ti bond with strike price 1/(1 + δκ) and maturity Ti−1 .4) Let t ≤ T0 . (2. t).
This is an important fact because many interest rate models have explicit formulae for bond option values. .26
CHAPTER 2.
for the time t price of the ith caplet with reset date Ti−1 and settlement date Ti . At Ti the holder of the cap receives δ(F (Ti−1 . the cashﬂow of which is – with the same notation as above – at time Ti δ(κ − F (Ti−1 .
1 and Figure 2. Ti ) (F (t. δ n P (0. T0 ) − P (0. t) = δP (t. t)) − κΦ(d2 (i. Correspondingly. t))) .1 (→ JW[11](p. Ti ) (κΦ(−d2 (i.Ti−1 . Ti ) i=1
the forward swap rate. Black’s Formula It is market practice to price a cap/ﬂoor according to Black’s formula. It is a challenge for any market realistic interest rate model to match the given volatility curve. t) :=
F (t. t))) . we have t = 0. Typically. Black’s formula for the value of the ith ﬂoorlet is F ll(i. t)) − F (t. Let t ≤ T0 .2 (i. Floors and Swaps
27
Caps and ﬂoors are strongly related to swaps. and σ(t) is the cap volatility (it is the same for all caplets). Indeed.
± 1 σ(t)2 (Ti−1 − t) 2 √ σ(t) Ti−1 − t
. Let t = 0. Ti−1 . Black’s formula for the value of the ith caplet is Cpl(i. t) = δP (t. The cap/ﬂoor is said to be atthemoney (ATM) if κ = Rswap (0) = P (0. and outofthemoney (OTM) if κ > Rswap (0) (κ < Rswap (0)).Ti ) κ
(Φ stands for the standard Gaussian cumulative distribution function).6. where log d1. The cap (ﬂoor) is inthemoney (ITM) if κ < Rswap (0) (κ > Rswap (0)). where Πp (t) is the value at t of a payer swap with rate κ. Ti−1 . Ti )Φ(d1 (i. Ti )Φ(−d1 (i. one can show the parity relation (→ exercise) Cp(t) − F l(t) = Πp (t). nominal one and the same tenor structure as the cap and ﬂoor. An example of a US dollar ATM market cap volatility curve is shown in Table 2. and T0 and δ = Ti − Ti−1 being equal to three months.49)). CAPS AND FLOORS Caps. Tn ) .2. Cap/ﬂoor prices are quoted in the market in term of their implied volatilities.
0 16.2 17.7 17.4
Figure 2.9 18.5 14.7 18.1 17.7 12. 23 July 1999
18% 16% 14% 12%
5
10
15
20
25
30
. INTEREST RATES AND RELATED CONTRACTS
Table 2.4 18.4 18.1: US dollar ATM cap volatilities.1: US dollar ATM cap volatilities.8 18.28
CHAPTER 2. 23 July 1999 Maturity (in years) 1 2 3 4 5 6 7 8 10 12 15 20 30 ATM vols (in %) 14.5 18.
Since Πp (T0 . Ti ) − K)
. Here the correlation between diﬀerent forward rates will enter the valuation procedure. Ti )δ(F (T0 .
Accordingly. ITM . the swaption maturity coincides with the ﬁrst reset date of the underlying swap. one can show (→ exercise) that the payoﬀ (2. Ti−1 . T0 . K < (>)Rswap (t).7
Swaptions
A European payer (receiver) swaption with strike rate K is an option giving the right to enter a payer (receiver) swap with ﬁxed rate K at a given future date. contrary to the cap case. K > (<)Rswap (t). SWAPTIONS
29
2. Usually. the payer (receiver) swaption with strike rate K is said to be ATM . Ti−1 . K) = N
i=1
P (T0 . OTM . Ti ) − K). respectively. Rswap (T0 )) = 0.
. if K = Rswap (t). This is a fundamental diﬀerence between caps/ﬂoors and swaptions.5) of the payer swaption at time T0 can also be written as
n
N δ(Rswap (T0 ) − K) and for the receiver swaption
+ i=1
P (T0 .5)
Notice that. Ti ). is
n
Πp (T0 . Recall that the value of a payer swap with ﬁxed rate K at its ﬁrst reset date. the swaption maturity.
Hence the payoﬀ of the swaption with strike rate K at maturity T0 is
n +
N
i=1
P (T0 . The underlying swap lenght Tn − T0 is called the tenor of the swaption. at time t ≤ T0 . this payoﬀ cannot be decomposed into more elementary payoﬀs.
(2.7.
n
N δ(K − Rswap (T0 ))
+ i=1
P (T0 .2. Ti )δ(F (T0 . Ti ).
2 and Figure 2. Ti ).
i=1 n
Swptr (t) = N δ (KΦ(−d2 (t)) − Rswap (t)Φ(−d1 (t))) with log d1.253)).2 (→ BM[6](p.30
CHAPTER 2. An x × yswaption is the swaption with maturity in x years and whose underlying swap is y years long.
± 1 σ(t)2 (T0 − t) 2 √ . An interest model for swaptions valuation must ﬁt the given today’s volatility surface. σ(t) T0 − t
. INTEREST RATES AND RELATED CONTRACTS
Black’s Formula Black’s formula for the price at time t ≤ T0 of the payer (Swptp (t)) and receiver (Swptr (t)) swaption is
n
Swptp (t) = N δ (Rswap (t)Φ(d1 (t)) − KΦ(d2 (t)))
P (t. A typical example of implied swaption volatilities is shown in Table 2.
i=1
and σ(t) is the prevailing Black’s swaption volatility.2 (t) :=
Rswap (t) K
P (t. Swaption prices are quoted in terms of implied volatilities in matrix form. Ti ).
7.1 9.7 11.9 11.9 12.6 13.5 10.4 10.3 7y 12.3.4 17.7 11.2: Black’s implied volatilities (in %) of ATM swaptions on May 16.5 11. Maturities are 1.8 14.5 14.4 11.1 11.5 13.4 6y 12.6 14.3 12.4
Figure 2.5 2y 15.5.1 13.2 11.6 8.7 11.0 11.8 8.4 9.10 years.2: Black’s implied volatilities (in %) of ATM swaptions on May 16.8 9y 12. swaps lengths from 1 to 10 years.6 10.2.9 15.6 11.8 15.7 17.4.8 5y 13.5 3y 14.9 12.1 12.1 8y 12.9 12.9 11.6 16.9 8.7 9.9 11.3 10. SWAPTIONS
31
Table 2.8 13.8 10.3 10. 2000.3 11.2.6 10y 11.0 10.3 9.4 11.4 4y 13.4 12.7.4 9.6 12.6 15.7 11. 2000.1 10.3 11.8 9.9 12.8 10.7 12.9 10. 1y 2y 3y 4y 5y 7y 10y 1y 16.9 11.5 9.
Maturity 2 4 6 8 10 16 14 Vol 12 10 2 6 4 Tenor 8 10
.1 10.
32
CHAPTER 2. INTEREST RATES AND RELATED CONTRACTS
.
• There exists a unique orthogonal matrix A = (p1 . . . N −1
N N k=1 (xi (k)
xi (k) (mean of xi ). . .
k=1
ˆ We assume that Σ is nondegenerate (otherwise we can express an xi as linear combination of the other xj s). pn ) (that is. [21] • Let x(1). . x(N ) be a sample of a random n × 1 vector x. ˆ Σij = = where 1 µ[xi ] := N − µ[xi ])(xj (k) − µ[xj ]) N −1 N k=1 xi (k)xi (k) − N µ[xi ]µ[xj ] . . .Chapter 3 Some Statistics of the Yield Curve
3. A−1 = AT and Aij = pj.2). ˆ • Form the empirical n × n covariance matrix Σ.1 Principal Component Analysis (PCA)
→ JW[11](Chapter 16. 33
.i ) consisting of orthonormal n × 1 Eigenvectors ˆ pi of Σ such that ˆ Σ = ALAT . .
l=1
T ˆ Aik Cov[xk . . • Deﬁne z := AT x.34
CHAPTER 3. . . ˜ and w = w − µ[w] (µ[w]=mean of w). . wj ] = Cov[wi . • The principal components (PCs) are the n × 1 vectors p1 . Then
n
Cov[zi . n.
• Sometimes the following view is useful (→ R[22](Chapter 3)): set σi := V ar[xi ] vi :=
1/2
ˆ = Σii
n j=1
1/2
n
1/2
=
j=1
A 2 λj ij λ j wj .
xi − µ[xi ] = σi
Aij σi
. pn : x = Az = z1 p1 + · · · zn pn . . . . wj ] = δij . zj ] =
k. λn ) with λ1 ≥ · · · ≥ λn > 0 (the Eigenvalues ˆ of Σ). . . Then ˜ ˜ ˜ ˜ µ[w] = 0. .
Aij
λ j wj . i = 1. . which indicates the amount of variance explained by pi . . . ˜ ˜
n
w = µ[x] +
j=1 n
pj
λj w j . . The key statistics is the proportion λi n j=1 the explained variance by pi . xl ]AT = AT ΣA jl
ij
= λi δij . .
Hence the zi s are uncorrelated. where L1/2 := diag( λ1 . The importance of component pi is determined by the size of the corresponding Eigenvalue. λn ). √ √ • Normalization: let w := (L1/2 )−1 z. and x = µ[x] + AL In components xi = µ[xi ] +
j=1 1/2
λj
. STATISTICS OF THE YIELD CURVE where L = diag(λ1 .
Cov[wi . λi . .
Now let x = (x1 . i. p2 = 0.358 0.329 −0. σi σj
where π is some lowdimensional parameter (this is adapted to the calibration of market models → BM[6](Chapter 6.3.61): UK market in the years 19891992 (the original maturity spectrum has been divided into eight distinct buckets.176 0. PCA typically leads to the following picture (→ R[22]p. µ[vi ] = 1 and
35
xi = µ[xi ] + σi vi .9)).2. • the second PC is upward sloping (tilt → slope). PCA OF THE YIELD CURVE
2 Then we have µ[vi ] = 0.352 • The ﬁrst PC is roughly ﬂat (parallel shift → average rate).204 −0. t + τi ). j). vj ] = ˆ Σij = σi σj
n k=1
Aik Ajk λk = ρ(π.e. Corr[xi .044 0.268 0.365 −0. .367 p1 = 0.368 0.364 .291 0.354 −0. t + τi−1 . t + ∆t + τi ) − R(t. t + ∆t + τi−1 .490 −0. i. p3 = −0.316 0. .176 .404 0.121 0.
.2
PCA of the Yield Curve
xi = R(t + ∆t.361 0.455 −0. 0. xj ] = Cov[vi . The ﬁrst three principal components are 0. say
for some maturity spectrum 0 = τ0 < · · · < τn . . . • the third PC humpshaped (ﬂex → curvature).161 . It can be appropriate to assume a parametric functional form (→ reduction of parameters) of the correlation structure of x.343 0. xn )T be the increments of the forward curve.
3. n = 8).722 0.461 0.
Princeton University).6 0.
0.17 2 6.93 3 0. STATISTICS OF THE YIELD CURVE Figure 3. Explained Variance (%) 1 92.2 0.58) A typical example of correlation among forward rates is provided by
. J.61 4 0.03 6–8 0.8 3 4 5 6 7 8
Table 3.36
CHAPTER 3.4 0.2 2 0.1: Explained Variance of the Principal Components (PCs). PCA of the yield curve goes back to the seminal paper by Litterman and Scheinkman (91) [17] (Prof.1: First Three PCs.6 0.24 5 0.01 PC
The ﬁrst three PCs explain more than 99 % of the variance of x (→ Table 3.1).
3.3
Correlation
→ R[22](p. Scheinkman is at the Department of Economics.4 0.
5. 2.6 0.95 0.95 1 shows the correlation for changes of forward rates of maturities 0. 1.5 3
→ Decorrelation occurs quickly.9 0. The data is from the US Treasury yield curve 1987–1994. The following matrix (→ Figure 3.2: Correlation between the short rate and instantaneous forward rates for the US Treasury curve 1987–1994
1 0. → Exponentially decaying correlation structure is plausible. 0.69 0.93 0.93 1 0.6 1 0.8 0.87 0.7 0.
Figure 3.2) 1 0.3.5 1 1. 1. 3 years.64 0.
.99 0.9 0.5 2 2.3.74 0. CORRELATION
37
Brown and Schaefer (1994).5.97 0.85 1 0.96 0.92 1 0.
38
CHAPTER 3. STATISTICS OF THE YIELD CURVE
.
04 5y 2.34 Jun 96 99.25 Sep 96 99. 9 January 1996.14 3y 1.53 2m 0.90 Swaps (%) 2y 1. 1996.1: Yen data.50 1m 0. The spot date t0 is 11 January.60 4y 2. 1996 (→ JW[11](Section 5. 360
Table 4.43 7y 3. δ(T. S) = actual number of days between T and S .1 A Bootstrapping Example
→ JW[11](p.Chapter 4 Estimating the Yield Curve
4.01 10y 3.4)). The daycount convention is Actual/360.36
39
. LIBOR (%) o/n 0.10 Dec 96 98.49 1w 0. The idea is to build up the yield curve from shorter maturities to longer maturities. We take Yen data from 9 January.55 3m 0.129–136) This is a naive bootstrapping method of ﬁtting to a money market yield curve.56 20 19 18 18 Futures Mar 96 99.
S5 ) 31 Then we use the relation q= P (t0 . S5 ). respectively. S5 )1−q . Ti+1 ). Ti+1 ] prevailing at t0 . . . T1 ). 33. Ti . where FF (t0 . . Ti ) 1 + δ(Ti . T1 ) = P (t0 . The zerocoupon bonds are P (t0 .
. 11/3/96. . 19/3/97}. P (t0 . Si )
To calculate zerocoupon bond from futures prices we need P (t0 . T5 ). we set F (t0 . 1 . . Ti+1 ) ≡ 91/360. . T5 } = {20/3/96. . S5 } = {12/1/96. 60 and 91 days to maturity. Ti+1 ) = P (t0 . We treat futures rates as if they were simple forward rates. 18/12/96. Ti . S5 ) = = 0. which is equivalent to using linear interpolation of continuously compounded spot rates R(t0 . δ(S4 . S4 )q P (t0 . ESTIMATING THE YIELD CURVE • The ﬁrst column contains the LIBOR (=simple spot rates) F (t0 . We use geometric interpoliation P (t0 . 1 + F (t0 . hence δ(Ti . 22 δ(T1 . Ti .709677. T1 ) = q R(t0 . 13/2/96. 11/4/96} hence for 1. Si ) = • The futures are quoted as futures price for settlement day Ti = 100(1 − FF (t0 . Ti+1 )).40
CHAPTER 4. . Ti . Ti+1 ) is the futures rate for period [Ti . that is. 7. and {T1 . . Ti . Si ) for maturities {S1 . T2 ). . . Ti+1 ) F (t0 . Ti+1 ) where
to derive P (t0 . S4 ) + (1 − q) R(t0 . . Si ) δ(t0 . Ti+1 ) = FF (t0 . 18/9/96. 19/6/96. 18/1/96.
12/7/99. We obtain P (t0 . Ui ) P (t0 . P (t0 .
. U2 ) + Rswap (t0 . . Un ) . 11/1. T5 ). 11/1/06 Rswap (t0 . U2 ) = R(t0 . Un )
This gives P (t0 .1. {U1 . . 20. U3 ) = (Rswap (t0 . 12/1/04. T2 ) + 91 65 R(t0 . U1 ). 11/1/02. . For maturity U3 this is 1 Rswap (t0 . T4 ) + 91 R(t0 . Un ) for n = 3. U4 )). 11/1/01. U2 ) (and hence Rswap (t0 . U20 } = 11/7/01.
From the data we have Rswap (t0 .
For a swap with maturity Un the swap rate at t0 is given by
n i=1
(U0 := t0 ). 13/1/03. 05.4. A BOOTSTRAPPING EXAMPLE • Yen swaps have semiannual cashﬂows at dates 11/7/96. 11/1/99. Ui ) for i = 4. . 13/7/98. 11/7/03. 12/7/04. Un ) = 1 − P (t0 . 91
All remaining swap rates are obtained by linear interpolation. Ui )
41
. 13/1/97. 12/1/98. Un ) = . U1 ). Ui ) P (t0 . 8. . 6. 11/1/00. T3 ) 91 26 R(t0 . 20. 11/7/05. Un ) i=1 δ(Ui−1 . 11/7/02. 11/7/00. 10. . 2 We have (→ exercise)
n−1 1 − Rswap (t0 . . δ(Ui−1 . Un )δ(Un−1 . 11/7/97. Rswap (t0 . . U2 )) by linear interpolation of the continuously compounded spot rates 69 R(t0 . U1 ) = 22 R(t0 . Ui ) P (t0 . 14. 1 + Rswap (t0 .
ti ).
→ The entire yield curve is constructed from relatively few instruments. ti . 29. ti ) = −
i = 1. . The method exactly reconstructs market prices (this is desirable for interest rate option traders).3 shows the spot rate curves from LIBOR. ti+1 ) = − δ(ti . ti ) log P (t0 . Spot and forward rates are continuously compounded log P (t0 .1: Zerocoupon bond curve
1 0. It is evident that the three curves are not coincident to a common underlying curve. reﬂecting the derivative of T → − log P (t0 . nonsmooth forward curve. ti+1 ) R(t0 .6 0. ti+1 ) − log P (t0 . ESTIMATING THE YIELD CURVE Figure 4.2. But it produces an unstable. 29
(we have 29 points and set P (t0 .4 0. t0 ) = 1).42
CHAPTER 4. R(t0 . ti ) δ(t0 . .1 is the implied zerocoupon bond price curve P (t0 .
.2 2 4 6 8 Time to maturity 10
In Figure 4. . . .8 0. Figure 4. Our naive method made no attempt to meld the three curves together.
The forward curve. ti ) . . . i = 0. futures and swaps. . T ). is very unsmooth and sensitive to slight variations (errors) in prices. The spot and forward rate curves are in Figure 4.
A BOOTSTRAPPING EXAMPLE Figure 4.5 Time to maturity 2
→ Another method would be to estimate a smooth yield curve parametrically from the market rates (for fund managers.06 0.01 0.008 0.012 0.05 0.5 0.006 0.04 0.3: Comparison of money market curves
0.4.02 0. In reality futures rates are greater than forward rates.009 0.011 0. forward rates (upper curve)
0. which is
.2: Spot rates (lower curve).007 0. long term strategies). The amount by which the futures rate is above the forward rate is called the convexity adjustment.005 1 1.03 0.01 2 4 6 8 Time to maturity 10
43
Figure 4. The main diﬃculties with our method are: • Futures rates are treated as forward rates.1.
. D(xN )) with cashﬂow dates t0 < T1 < · · · < TN . the segments of LIBOR. ESTIMATING THE YIELD CURVE model dependend. . in some markets intermediate swaps are indeed priced as if their prices were found by linear interpolation.2
General Case
The general problem of ﬁnding today’s (t0 ) term structure of zerocoupon bond prices (or the discount function) x → D(x) := P (t0 . • liquidity eﬀects. . and a vector of pricing errors. 2 where τ is the time to maturity of the futures contract. Reasons for including errors are • prices are never exactly simultaneous. In general. • Swap rates are inappropriately interpolated. at S5 = 11/4/96) are ignored once P (t0 . • allows for smoothing.
.44
CHAPTER 4. p = C ·d+ . Ti − t 0 = x i . . etc). • LIBOR rates beyond the “stup date” T1 = 20/3/96 (that is. and d = (D(x1 ). low coupons). • tax eﬀects (high coupons. However. t0 + x) can be formulated as where p is a vector of n market prices. T1 ) is found. • roundoﬀ errors in the quotes (bidask spreads.
4. The linear interpolation produces a “sawtooth” in the forward rate curve. and σ is the volatility parameter. futures and swap markets overlap. C the related cashﬂow matrix. An example is 1 forward rate = futures rate − σ 2 τ 2 .
p.75 9 next coupon 15/11/96 19/01/97 26/09/96 03/03/97 06/11/96 27/02/97 07/12/96 08/03/97 13/10/96 maturity dirty price date (pi ) 15/11/96 103. .15 27/08/02 111. and the daycount convention is actual/365.28 06/11/01 101. T2 = 13/10/96.06 07/12/05 106.75 8.
. The coupon payments are semiannual.
1≤j≤N
Example (→ JW[11]. coupon (%) 10 9.44 03/03/00 106.49 13/10/08 110. . GENERAL CASE
45
4.4.5 7.2.2: Market prices for UK gilts. • dates of all cashﬂows t0 < T1 < · · · < TN . forming the n × N cashﬂow matrix C = (ci.
Table 4.82 19/01/98 106. The spot date is 4/9/96.426): UK government bond (gilt) market.j at time Tj (may be zero). 1996. pn ). . T3 = 06/11/97. .04 26/03/99 118. .2.75 12.j ) 1≤i≤n .1
Data:
Bond Markets
• vector of quoted/market bond prices p = (p1 . . T1 = 26/09/96.24 08/09/06 98. .25 9 7 9.87
bond bond bond bond bond bond bond bond bond
1 2 3 4 5 6 7 8 9
Hence n = 9 and N = 1 + 3 + 6 + 7 + 11 + 12 + 19 + 20 + 25 = 104. • bond i with cashﬂows (coupon and principal payments) ci. 4/9/96. selection of nine gilts. . September 4.
c1 = · · · = cn−1 = δK. 0 0 0 0 0 0 0 4. FRAs and forward swaps have notional price 0.. The 9 × 104 cashﬂow matrix 0 0 0 105 0 0 0 0 0 0 .. 0 3. 0 0 0 0 4. c0 = −1. c1 = −1 at T1 = T . 0 0 0 0 0 0 0 0 3. S]): p = 0. c2 = 1+(S−T )F at T2 = S. c1 = · · · = cn−1 = δK.5 0 0 0 0 0 0 0 . Swap (receiver.2
Money Markets
Money market data can be put into the same price–cashﬂow form as above. LIBOR is for o/n (1/365)...125 0 0 0 0 0 0 0 0 6.875 0 . FRA (forward rate F for [T.5 0 0 . cn = 1 + δK.
• if T0 = t0 : p = 1. The spot date t0 is 8/10/97.. C= 0 0 0 0 0 0 0 4...25 0 0 0 0 0 . 6..2. 1997.. p. cn = 1 + δK. LIBOR (rate L. 0 0 0 0 0 4. 1m (33/360).. maturity T ): p = 1 and c = 1 + (T − t0 )L at T .5 0 0 0 0 0 0 0 0 .
. swap rate K. ESTIMATING THE YIELD CURVE is
No bonds have cashﬂows at the same date. Example (→ JW[11]. Ti − Ti−1 ≡ δ): since
n
0 = −D(T0 − t0 ) + δK
j=1
D(Tj − t0 ) + (1 + δK)D(Tn − t0 ). .46
CHAPTER 4.428): US money market on October 6. tenor t0 ≤ T0 < · · · < Tn .875 0 0 0 0 .
4..125 .875 0 0 0 ... . 0 4.... and 3m (92/360). The daycount convention is Actual/360.
• if T0 > t0 : p = 0. → at t0 : LIBOR and swaps have notional price 1.
GENERAL CASE
47
Futures are three month rates (δ = 91/360).12 16/9/98 94 16/12/98 6.24 17/12/97 94. T + δ)). . .56 6. We take them as forward rates.2.26 19/11/97 94.18 17/6/98 94.42 6.71875 8/1/98 94. The ﬁrst payment date is 8/10/98.625 10/11/97 5. Swaps are annual (δ = 1).56 6. The ﬁrst 14 columns of
.23 18/3/98 94.32 6.01253 6. T1 = 9/10/97.22 6. 1997. That is.16 6.27 15/10/97 94.4. October 6. . .10823 6. the quote of the futures contract with maturity date (settlement day) T is 100(1 − F (t0 .
Table 4. N = 3 + 14 + 30 = 47.3: US money market.59375 9/10/97 5. LIBOR Period o/n 1m 3m Oct97 Nov97 Dec97 Mar98 Jun98 Sep98 Dec98 2 3 4 5 7 10 15 20 30 Rate Maturity Date 5. T. T2 = 15/10/97 (ﬁrst future).56
Futures
Swaps
Here n = 3 + 7 + 9 = 19. T3 = 10/11/97.
10 = 1. 0 0 0 0 0 0 0 0 0 0 c11.14 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
4.060125.12 c14.0616.12 c16.12 = c18.12 = 0. This makes ordinary least square (OLS) regression
d∈RN
min {
2
 = p − C · d} (⇒ C T p = C T Cd∗ )
unfeasible.00516.12 = 0.01456.12 c12.12 = 0.01517 c11.2.48
CHAPTER 4.12 c13. we have n N .10 0 0 0 0 0 0 0 0 0 0 0 −1 c8.11 = 1.
. c14. ESTIMATING THE YIELD CURVE
the 19 × 47 cashﬂow matrix C are c11 0 0 0 0 0 0 0 0 0 0 0 0 c23 0 0 0 0 0 0 0 0 0 0 0 0 0 c36 0 0 0 0 0 0 −1 0 0 0 0 c47 0 0 0 0 0 0 0 −1 0 0 0 c58 0 0 0 0 0 0 0 −1 0 0 0 c69 0 0 0 0 0 0 0 0 0 −1 c7.01486.0632. c23 = 1.0656.01451.12 = 0.061082. c47 = 1. Moreover.01448.12 = 0.11 0 0 0 0 0 0 0 0 0 0 −1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 with c11 = 1.01461.13 −1 c10.0642. c69 = 1.01459.12 = c19. c58 = 1.12 c15.12 c18. many entries of C are zero (diﬀerent cashﬂow dates). c8.00016. c13.12 c19.14 = 1.01471.12 c17.3
Problems
Typically. c15.12 = 0.12 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
c9. c17.0622. c10. c12.13 = 1. c36 = 1.12 = 0. c7. c9. c16.
2. and let φ(x. z) = z1 ψ1 (x) + · · · + zm ψm (x). and even worse for forward rates. . .
This leads to a nonlinear optimization problem min p − C · d(z) .
with state space Z ⊂ Rm .
0
z ∈ Z. ψm (preferably with compact support). As a result this leads to a ragged spot rate (yield) curve. . z). Still regression only yields values D(xi ) at the payment dates t0 + xi → interpolation technics necessary.4. z) = exp −
φ(u. GENERAL CASE
49
One could chose the data set such that cashﬂows are at same points in time (say four dates each year) and the cashﬂow matrix C is not entirely full of zeros (Carleton–Cooper (1976)). z) du . But there is nothing to regularize the discount factors (discount factors of similar maturity can be very diﬀerent). t0 + x) = φ(x) = φ(x. . For regularity reasons (see below) it is best to estimate the forward curve R+ x → f (t0 . z) du
for some payment tenor 0 < x1 < · · · < xN .
z∈Z
with di (z) = exp −
0
xi
φ(u.
4. Linear Families Fix a set of basis functions ψ1 .2.4
Parametrized Curve Families
Reduction of parameters and smooth yield curves can be achieved by using parametrized families of smooth curves
x
D(x) = D(x.
.
a4 . +
k = −3. ESTIMATING THE YIELD CURVE
Cubic Bsplines A cubic spline is a piecewise cubic polynomial that is everywhere twice diﬀerentiable. . 1.50
CHAPTER 4. +
hence it has m + 3 parameters {a0 .
0.06 0. . .05 0.01 2 4 6 8 10 12
.02 0. Its general form is
3 m−1
σ(x) =
i=0
a i xi +
j=1
bj (x − ξj )3 . m − 1. . A basis for the cubic splines on [ξ0 . . . b1 .04 0.03 0. Figure 4. . ξk+4 ]. 11}. . bm−1 } (a kth degree spline has m + k parameters).i=j
1 ξi − ξ j
(x − ξj )3 . It interpolates values at m + 1 knot points ξ0 < · · · < ξm . 6. Introduce six extra knot points ξ−3 < ξ−2 < ξ−1 < ξ0 < · · · < ξm < ξm+1 < ξm+2 < ξm+3 . 8. The spline is uniquely characterized by speciﬁcation of σ or σ at ξ0 and ξm .4: Bspline with knot points {0. . ξm ] is given by the m + 3 Bsplines
k+4 k+4
ψk (x) =
j=k i=k. . . .
The Bspline ψk is zero outside [ξk .
8. The estimation with all 8 Bsplines leads to
z∈R8
min p − CΨz = p − CΨz ∗ = 0. .197
• As a basis we use the 8 (resp. z) = ψ1 (0)z1 + · · · + ψm (0)zm = 1. Its exact yield is y=− 365 103. D(x. 20. Notice that the ﬁrst bond is a zerocoupon bond. 11. 1. d(z) = · = .2). Example We take the UK government bond market data from the last section (Table 4. 15. zm
this leads to the linear optimization problem
z∈Rm
If the n × m matrix A := CΨ has full rank m. is 12.5). z1 . 30} (see Figure 4. the unique unconstrained solution is z ∗ = (AT A)−1 AT p. . The maximum time to maturity. 6. z) min p − CΨz . . 0.23
.989 = 0.11 [years]. . A reasonable constraint would be D(0. ﬁrst 7) Bsplines with the 12 knot points {−20. . ψ1 (xN ) · · · ψm (xN ) D(xN . −2. z) .822 1 log =− log 0. =: Ψ · z . −5.4. . . GENERAL CASE
51
Estimating the Discount Function Bsplines can also be used to estimate the discount function directly (Steeley (1991)).2. x104 . With ψ1 (x1 ) · · · ψm (x1 ) D(x1 . . z) = z1 ψ1 (x) + · · · + zm ψm (x). 72 105 0.0572. 25.
23383 −4.38766 4. 8. and forward curve (cont. −5. 3monthly forward rates) shown in Figure 4.9717 855.8641 11.56562 7.
0. 25.01 5 10 15 20 25 30
with
and the discount function. The estimation with only the ﬁrst 7 Bsplines leads to
z∈R7
∗ z =
13.98066 6.57992 7.32 ∗ z = 17. spot rates).69741 6.5: Bsplines with knots {−20. 11. yield curve (cont.
with
. comp.07 0.49629 7.074
.4665 8. −2.9919 . 6.04 0. 15.05 0. comp.52
CHAPTER 4. 0. 20.06 0. 30}.03 0.7.8019 11.
min p − CΨz = p − CΨz ∗ = 0.02 0. 1.28853 5. ESTIMATING THE YIELD CURVE
Figure 4.3603 8.
4.9. 15.39 5 15. 3month forward rates) shown in Figure 4. −5.
0. yield curve (cont. comp. −5. comp.035 0. comp. 0. 30} (see Figure 4.23152
with
∗
and the discount function.2.
. 25.02 0.015 0.6: Five Bsplines with knot points {−10. 4. 30}.03 0. • Next we use only 5 Bsplines with the 9 knot points {−10.40296 6. spot rates).01 0.4385 12.005 5 10 15 20 25 30
The estimation with this 5 Bsplines leads to
z∈R
min p − CΨz = p − CΨz ∗ = 0. Figure 4. 4. 0. yield curve (cont.652 19. and forward curve (cont. 20.9886 7.
z =
. 3monthly forward rates) shown in Figure 4. −2. GENERAL CASE
53
and the discount function. 15. spot rates).8.025 0. 25. comp.6). −2. and forward curve (cont. 20.
6 0.1 0.15 0.12 0.2 2 0. ESTIMATING THE YIELD CURVE
Figure 4. The dot is the exact yield of the ﬁrst bond.1 0.08 0.7: Discount function.8 0.04 0.02 2 0.25 0.06 0.4 0.
1 0.14 0. yield and forward curves for estimation with 8 Bsplines.2 0.54
CHAPTER 4.05 2 4 6 8 10 12 4 6 8 10 12 4 6 8 10 12
.
15 0.6 0.06 0.05 2 4 6 8 10 12 4 6 8 10 12 4 6 8 10 12
.14 0.4 0.25 0.4.1 0.04 0.02 2 0.2 2 0.2. The dot is the exact yield of the ﬁrst bond. GENERAL CASE
55
Figure 4.2 0.8 0.1 0.8: Discount function. yield and forward curves for estimation with 7 Bsplines.08 0.12 0.
1 0.
04 0.05 2 4 6 8 10 12 4 6 8 10 12 4 6 8 10 12
. yield and forward curves for estimation with 5 Bsplines.1 0.6 0.9: Discount function.02 2 0.8 0. ESTIMATING THE YIELD CURVE
Figure 4.2 0.
1 0.4 0.56
CHAPTER 4. The dot is the exact yield of the ﬁrst bond.25 0.1 0.14 0.15 0.2 2 0.12 0.06 0.08 0.
TN ) at 0 < T1 < · · · < TN ≡ T . Y N . Example: Lorimier (95).2.32 and 0.9 are more regular than those in Figure 4. T1 ). → Direct estimation of the yield or forward curve. and consequently the N yields Y1 . For ease of notation we set t0 = 0 (today). • The Bspline ﬁts are extremely sensitive to the number and location of the knot points. Sabine Lorimier suggests a spline method where the number and location of the knots are determined by the observed data itself. In her PhD thesis 1995. . P (0.4. → Smoothing splines. The data is given by N observed zerocoupon bonds P (0. but their correctness criteria (0.23).7. . . The curves in Figures 4. → Need criterions asserting smooth yield and forward curves that do not ﬂuctuate too much and ﬂatten towards the long end. GENERAL CASE Discussion • In general. • There is a tradeoﬀ between the quality (or regularity) and the correctness of the ﬁt. → Optimal selection of number and location of knot points for splines. Smoothing Splines The least squares criterion min p − C · d(z)
z 2
has to be replaced/extended by criterions for the smoothness of the yield or forward curve. . . P (0.39) are worse than for the ﬁt with 8 Bsplines (0. • Splines are not useful for extrapolating to long term maturities. Problems mostly at short and long term maturities. .
. Ti ) = exp(−Ti Yi ).8 and 4. . . splines can produce bad ﬁts.
57
• Estimating the discount function leads to unstable and nonsmooth yield and forward curves.
T ] is a second order polynomial on [0.1)
0
with an arbitrary constant α > 0.58
CHAPTER 4.
The optimization problem then is min F (f ). h
H
= g(0)h(0) +
0
g (u)h (u) du.4)
.
f ∈H
(*)
The parameter α tunes the tradeoﬀ between smoothness and correctness of the ﬁt. hk (0) = Tk . Theorem 4. ESTIMATING THE YIELD CURVE
Let f (u) denote the forward curve. . (4. .
and the nonlinear functional on H
T N
F (f ) :=
0
(f (u)) du + α
i=1
2
Ti
2
Yi Ti −
f (u) du
0
. . The ﬁtting requirement now is for the forward curve Ti √ f (u) du + i / α = Ti Yi . N.2)
(f (u))2 du. (4. Problem (*) has a unique solution f .1. . T ]} with scalar product
T
g. Tk ] with hk (u) = (Tk − u)+ . k = 1.3)
where hk ∈ C 1 [0. which is a second order spline characterized by
N
f (u) = f (0) +
k=1
ak hk (u)
(4. The aim is to minimize smoothness criterion
T
2
as well as the (4.2.
0
Introduce the Sobolev space H = {g  g ∈ L2 [0.
k=1 N
(4. .
∀g ∈ H. Integration by parts yields
Tk 0 Tk
g(u) du = Tk g(Tk ) − = Tk g(0) + Tk
ug (u) du
0 Tk 0 T Tk
g (u) du −
ug (u) du
0 H.
Tk
hl du = hl .6)
α Yk Tk − f (0)Tk −
a l h l . hk
l=1
H
= ak . . (4. .2. g
for all g ∈ H. GENERAL CASE and f (0) and ak solve the linear system of equations
N
59
ak Tk = 0. hk
T
H
= 0 we obtain
f − f (0).
Proof. N. g Hence
H
=
0
f (u)g (u) du = 0. In particular.
= Tk g(0) +
0
(Tk − u)+ g (u) du = hk . . .
(4. .4. for all g ∈ H with g.
A (local) minimizer f of F satisﬁes d F (f + g) =0 = 0 d or equivalently
T N Tk Tk
f g du = α
0 k=1
Yk Tk −
f du
0 0
g du.
f − f (0) ∈ span{h1 .5) k = 1. . hk
0
H. hN }
.7)
In particular. .
. (4. It remains to show that f exists and is unique.5)–(4.7) is equivalent to (4. h1 H α h N .8) . hl
H λl
+ λk = 0. N. ESTIMATING THE YIELD CURVE
what proves (4.7)
((g − f ) + f ) du + α
T 2 0
2
Tk
2
k=1 N
Yk Tk −
Tk 0
g du
0 Tk 2
= F (f ) +
(g − f ) du + α
k=1
f du −
g du
0
≥ F (f ). . . .6). Let g ∈ H.5) (set u = 0). hN H + 1
N
Let λ = (λ0 .4) and (4.3).7) with g − f ∈ H. hk
l=1
H 0
g(u) du = 0
for all g ∈ H.7) is a suﬃcient condition for f to be a global minimizer of F .60
CHAPTER 4. . that the linear system (4. . hN H A= . h2 H · · · α h 1 . . . . . . h1 H + 1 α h1 . ..
. where we used (4.6) has a unique solution (f (0). Next we show that (4. (4. that is. . Thus we have shown that (4. aN ). . . h2 H · · · α h N . . . αTN α h N . . .6) holds. . .3)–(4. .. The corresponding (N + 1) × (N + 1) matrix is 0 T1 T2 ··· TN αT1 α h1 . . Tk λk = 0
k=1 N
αTk λ0 + α
l=1
h k . that is.7) can be rewritten as
N N Tk
k=1
ak − α Yk Tk − f (0)Tk −
a l h l . . This is true if and only if (4. λN )T ∈ RN +1 such that Aλ = 0. then
T N
F (g) =
0 (4. Hence we have
T N Tk N Tk
f (u)g (u) du =
0 k=1
ak −Tk g(0) +
g(u) du =
0 k=1
ak
0
g(u) du. a1 .
k = 1.
and (4.
9)
a perfect ﬁt. .
(4. . . if coupon bond prices are given. k = 1.8)). N . . . l = 1. That is. . This and much more is carried out in Lorimier’s thesis. The role of α is as follows: • If α → 0 then by (4. .7) implies that
Tk
f (u) du = Yk Tk . .9). N. . To estimate the forward curve from N zerocoupon bonds—that is. GENERAL CASE Multiplying the latter equation with λk and summing up yields
N 2 N
61
α
k=1
λ k hk
H
+
k=1
λ2 = 0. maximal regularity
T
(f (u))2 du = 0
0
but no ﬁtting of data. f minimizes (4. (f )2 du + α log pk − log ckl exp − min f ∈H 0 0
k=1 l=1
If the coupon payments are small compared to the nominal (=1).2) subject to the constraints (4. see (4. YN )T —one has to solve the linear system A· f (0) a = 0 Y (see (4.
. That is.3) and (4. .6) we have f (u) ≡ f (0). then the above method has to be modiﬁed and becomes nonlinear. whence A is nonsingular. a constant function. k
Hence λ = 0. yields Y = (Y1 . • If α → ∞ then (4.1). this reads 2 n N T Tl f du .
Of course. .
0
k = 1.4. . . then this problem has a unique solution. . n. With p ∈ Rn denoting the market price vector and ckl the cashﬂows at dates Tl . .2. .
z) = z1 + (z2 + z3 x)e−z4 x + z5 e−z6 x . including 6 parameters z1 .5 and 7 diﬀerent values for z3 = 1.13.21. E.76. Svensson (94) [26] (Prof. z4 = 0. Svensson is at the Economics Department. .19. z2 = −4.
8 6 4 2
5
10
15
20
Table 4. O.
.2. .4 is taken from a document of the Bank for International Settlements (BIS) 1999 [2].62
CHAPTER 4. . . L. φS (x. . −4.77. ESTIMATING THE YIELD CURVE
ExponentialPolynomial Families Exponentialpolynomial functions p1 (x)e−α1 x + · · · + pn (x)e−αn x (pi =polynomial of degree ni )
form nonlinear families of functions. −2.22. −3.18.
Figure 4. . 0. Princeton University) This is an extension of Nelson–Siegel. z) = z1 + (z2 + z3 x)e−z4 x . z4 and φN S (x. . .69. Popular examples are: Nelson–Siegel (87) [20] There are 4 parameters z1 . −0. −1. z6 .10: Nelson–Siegel curves for z1 = 7.
BIS 1999 [2]. wp for weighted prices.
.4.2. NS is for Nelson–Siegel. GENERAL CASE
63
Table 4. • Consistency: do the curve families go well with interest rate models? → this point will be exploited in the sequel.4: Overview of estimation procedures by several central banks. S for Svensson. Central Bank Belgium Canada Finland France Germany Italy Japan Norway Spain Sweden UK USA Method S or NS S NS S or NS S NS smoothing splines S S S S smoothing splines Minimized Error wp sp wp wp yields wp prices yields wp yields yields bills: wp bonds: prices
Criteria for Curve Families • Flexibility (do the curves ﬁt a wide range of term structures?) • Number of factors not too large (curse of dimensionality). • Regularity (smooth yield or forward curves that ﬂatten out towards the long end).
64
CHAPTER 4. ESTIMATING THE YIELD CURVE
.
. → Correlation structure among observable quantities can now be obtained analytically or numerically. PCA).g. a 10y swap with semiannual payments) or more. → The valuation of such products requires the modelling of the entire covariance structure. ﬂoors. This is exactly what interest rate (curve) models oﬀer: • reduction of ﬁtting degrees of freedom → makes problem manageable. 65
. . swaptions consist of a sequence of cashﬂows at T1 . . → Specify the dynamics of a small number of variables (e.Chapter 5 Why Yield Curve Models?
→ R[22](Chapter 5) Why modelling the entire term structure of interest rates? There is no need when pricing a single European call option on a bond. . But: the payoﬀs even of “plainvanilla” ﬁxed income products such as caps. Historical estimation of such large covariance matrices is statistically not tractable anymore. → Simultaneous pricing of diﬀerent options and hedging instruments in a consistent framework. Tn .g. =⇒ It is practically and intellectually rewarding to consider noarbitrage conditions in much broader generality. → Need strong structure to be imposed on the comovements of ﬁnancial quantities of interest. where n may be 20 (e.
66
CHAPTER 5. WHY YIELD CURVE MODELS?
.
[25]. .
6. Wd ). and many more. Financial Market We consider a ﬁnancial market with n traded assets. This is not a restriction since always one can set a stochastic process to be zero after a ﬁnite time T if this were the ultimate time horizon (as in the Black–Scholes model). S0 (0) = 1 67 ⇔ S0 (t) = e
t 0
r(s) ds
. Reference is B[3]. P). The background for stochastic analysis can be found in many textbooks. . and do not a priori ﬁx a ﬁnite time horizon. . etc.
.
Si > 0. . . . n
and the riskfree asset dS0 (t) = r(t)S0 (t) dt. . F . following strictly positive Itˆ processes o
d
dSi (t) = Si (t)µi (t) dt +
j=1
Si (t)σij (t) dWj (t). a ddimensional Brownian motion W = (W1 . We shall assume that F = F∞ = ∨t≥0 Ft .Chapter 6 NoArbitrage Pricing
This chapter brieﬂy recalls the basics about pricing and hedging in a Brownian motion driven market. and the ﬁltration (Ft )t≥0 generated by W . [23]. such as [14]. MR[19](Chapter 10).1
SelfFinancing Portfolios
The stochastic basis is a probability space (Ω.
i = 1. From time to time we recall some of the fundamental results without proof. .
n
. . volatility σ = (σij ). and short rates r are assumed to form adapted processes which meet the required integrability conditions such that all of the above (stochastic) integrals are welldeﬁned.
The portfolio φ is called selfﬁnancing (for S) if the stochastic integrals
t
φi (u) dSi (u). . .s. . . . Remark 6. Let h = (h1 .1.68
CHAPTER 6. is any adapted process φ = (φ0 . . . .
j=1 0
E
h(s)
2
ds < ∞
(the class of such processes is denoted by L2 ) then h · W is a martingale and the Itˆ isometry holds o
t 2 t
E
h(s) dW (s)
0
=E
h(s)
0
2
ds . . .1. It is always understood that for a random variable “X ≥ 0” means “X ≥ 0 a. .” (that is. Its corresponding value process is
n
V (t) = V (t. µn ). φ) :=
i=0
φi (t)Si (t).1. Theorem 6.
0
i = 0.2 (Stochastic Integrals). If
t
h(s)
0
2
ds < ∞ for all t > 0
(the class of such processes is denoted by L) one can deﬁne the stochastic integral
t d t
(h · W )t ≡ If moreover
0
h(s) dW (s) ≡
∞ 0
hj (s) dWj (s). . φn ). NOARBITRAGE PRICING
The drift µ = (µ1 . P[X ≥ 0] = 1). . etc. or trading strategy. . hd ) be a measurable adapted process.
Selfﬁnancing Portfolios A portfolio. .
6.2. ARBITRAGE AND MARTINGALE MEASURES are well deﬁned and dV (t; φ) =
i=0 n
69
φi (t) dSi (t).
Numeraires All prices are interpreted as being given in terms of a numeraire, which typically is a local currency such as US dollars. But we may and will express from time to time the prices in terms of other numeraires, such as Sp for some 0 ≤ p ≤ n. The discounted price process vector Z(t) := implies the discounted value process
n
S(t) Sp (t)
˜ V (t; φ) :=
i=0
φi (t)Zi (t) =
V (t; φ) . Sp (t)
Up to integrability, the selfﬁnancing property does not depend on the choice of the numeraire. Lemma 6.1.3. Suppose that a portfolio φ satisﬁes the integrability conditions for S and Z. Then φ is selfﬁnancing for S if and only if it is selfﬁnancing for Z, in particular
n n
˜ dV (t; φ) =
i=0
φi (t) dZi (t) =
i=0 i=p
φi (t) dZi (t).
(6.1)
Since Zp is constant, the number of terms in (6.1) reduces to n. Often (but not always) we chose S0 as the numeraire.
6.2
Arbitrage and Martingale Measures
Contingent Claims Related to any option (such as a cap, ﬂoor, swaption, etc) is an uncertain future payoﬀ, say at date T , hence an FT measurable random variable X (a contingent (T )claim). Two main problems now are: • What is a “fair” price for a contingent claim X? • How can one hedge against the ﬁnancial risk involved in trading contingent claims?
70
CHAPTER 6. NOARBITRAGE PRICING
Arbitrage An arbitrage portfolio is a selfﬁnancing portfolio φ with value process satisfying V (0) = 0 and V (T ) ≥ 0 and P[V (T ) > 0] > 0 for some T > 0. If no arbitrage portfolios exist for any T > 0 we say the model is arbitragefree. An example of arbitrage is the following. Lemma 6.2.1. Suppose there exists a selfﬁnancing portfolio with value process dU (t) = k(t)U (t) dt, for some measurable adapted process k. If the market is arbitragefree then necessarily r = k, dt ⊗ dPa.s. Proof. Indeed, after discounting with S0 we obtain U (t) ˜ = U (0) exp U (t) := S0 (t) Then (→ exercise) ψ(t) := 1{k(t)>r(t)} yields a selfﬁnancing strategy with discounted value process ˜ V (t) =
0 t t 0 t 0
(k(s) − r(s)) ds .
˜ ψ(s) dU(s) =
˜ 1{k(s)>r(s)} (k(s) − r(s))U (s) ds ≥ 0.
Hence absence of arbitrage requires ˜ 0 = E[V (T )] = ˜ 1{k(t,ω)>r(t,ω)} (k(t, ω) − r(t, ω))U (t, ω) dt ⊗ dP
>0 on N
N
where N := {(t, ω)  k(t, ω) > r(t, ω)} is a measurable subset of [0, T ] × Ω. But this can only hold if N is a dt ⊗ dPnullset. Using the same arguments with changed signs proves the lemma.
6.2. ARBITRAGE AND MARTINGALE MEASURES
71
Martingale Measures We now investigate when a given model is arbitragefree. To simplify things in the sequel • we ﬁx S0 as a numeraire, and ˜ • V will express the discounted value process V /S0 . But the following can be made valid for any choice of numeraire. An equivalent probability measure Q ∼ P is called an equivalent (local) martingale measure (E(L)MM) if the discounted price processes Zi = Si /S0 are Q(local) martingales. Theorem 6.2.2 (Girsanov’s Change of Measure Theorem). Let Q ∼ P be an equivalent probability measure. Then there exists γ ∈ L such that the density process dQ/dP is the stochastic exponential E(γ · W ) of γ · W dQ F = Et (γ · W ) := exp dP t Moreover, the process ˜ W (t) := W (t) −
t t 0
1 γ(s) dW (s) − 2
t
γ(s)
0
2
ds .
(6.2)
γ(s) ds
0
(6.3)
is a QBrownian motion. Conversely, if γ ∈ L is such that E (γ · W ) is a uniformly integrable martingale with E∞ (γ · W ) > 0 — suﬃcient is the Novikov condition E exp 1 2
∞ 0
γ(s)
2
ds
<∞
(6.4)
(see [23, Proposition (1.26), Chapter IV]) — then (6.2) deﬁnes an equivalent probability measure Q ∼ P. Market Price of Risk Let Q be an ELMM and γ (the stochastic logarithm ˜ of the density process) and W given by (6.2) and (6.3). Integration by parts yields the Zdynamics dZi (t) = Zi (t) (µi (t) − r(t)) dt + Zi (t)σi (t) dW (t) ˜ = Zi (t) (µi (t) − r(t) + σi (t) · γ(t)) dt + Zi (t)σi (t) dW (t).
V (t. and vice versa. This would certainly be a moneymaking machine. . . Admissible Strategies In the presence of local martingales one has to be alert to pitfalls. Both conditions 1 and 2. . M can be chosen to be of the form t M (t) =
0
where 1 := (1. φ) ≥ −a for some a ∈ R. This is why −γ is called the market price of risk . Zi can be written as stochastic exponential o ˜ Zi = E(σi · W ). V (t.
φ(s) dW (s)
(Dudley’s Representation Theorem). “suicide strategies” remain (however. . say arbitrage. φ) is a true Qmartingale. . they do not introduce arbitrage).5)
Hence if σi satisﬁes the Novikov condition (6. For example it is possible to construct a local martingale M with M (0) = 0 and M (1) = 1. 1)T . n.5) has a solution γ ∈ L such that E(γ · W ) is a uniformly integrable martingale (the Novikov condition (6. φ) is a Qsupermartingale for every ELMM Q. . . NOARBITRAGE PRICING
If σ is nondegenerate (in particular d ≤ n and rank[σ] = d) then γ is uniquely speciﬁed by −γ = σ −1 · (µ − r1)
µi − r + σi · γ = 0 dt ⊗ dQa. n then the ELMM Q is in fact an EMM.2) deﬁnes an ELMM Q. OR ˜ 2. which looks like the (discounted) value process of a selfﬁnancing strategy. if (6.4) is suﬃcient) then (6.g.s. . Condition 1 is more universal (it does not depend on a particular Q) and implies that V (t. by Itˆ’s formula. for all i = 1. . Conversely. for some ELMM Q. In the same way “suicide strategies” (e. are sensitive with respect to the choice of numeraire!
.
(6. Yet. . Even worse. however. . . To rule out such examples we have to impose additional constraints on the choice of strategies.72 Hence necessarily γ satisﬁes
CHAPTER 6. M (0) = 1 and M (1) = 0) can be constructed.4) for all i = 1. If γ is unique then Q is the unique ELMM. Here are two typical examples: A selfﬁnancing strategy φ is admissible if ˜ 1. There are several ways to do so. Notice that.
φ) = X. It has become a custom (and we will follow this tradition) to consider the existence of an ELMM as synonym for the absence of arbitrage: absence of arbitrage = existence of an ELMM. HEDGING AND PRICING
73
The Fundamental Theorem of Asset Pricing The existence of an ELMM rules out arbitrage. we have ˜ ˜ 0 ≤ EQ [V (T )] ≤ V (0) = 0. etc) that also the converse holds true: if arbitrage is deﬁned in the right way (“No Free Lunch with Vanishing Risk”). Lemma 6. in the sense that there exists no admissible (either Condition 1 or 2) arbitrage strategy.3. A simple example: suppose S1 is the price process of the T bond. that is. It is folklore (Delbaen and Schachermayer 1994. → the existence of an ELMM is now a standing assumption. Since V is a Qsupermartingale in any case (for some ELMM Q). ˜ whence V (T ) = 0. This is called the Fundamental Theorem of Asset Pricing. then its absence implies the existence of an ELMM Q. let V be the discounted value process of an admissible strat˜ ˜ ˜ egy.2.3
Hedging and Pricing
Attainable Claims A contingent claim X due at T is attainable if the exists an admissible strategy φ which replicates/hedges X.6.
. V (T . Then the contingent claim X = 1 due at T is attainable by an obvious buy and hold strategy with value process V (t) = S1 (t). Then the model is arbitragefree. Suppose there exists an ELMM Q.
6.3. with V (0) = 0 and V (T ) ≥ 0. ˜ Proof. Indeed.
Q) (6.74
CHAPTER 6.
Applying Itˆ’s formula yields o d and dY = d (Y D) = = 1 D = YDd dW − ˜ dW . T ]. Suppose that σ is nondegenerate.5) yields a uniformly integrable martingale E(γ · W ) and hence a unique ELMM Q.
with the density process D(t) = dQ/dPFt = Et (γ · W ). Hence Y D is a Pmartingale and by the representation theorem 6. 1 1 1 + d(Y D) + d Y D.1. Then
Bayes
t ∈ [0.3 we can ﬁnd ψ ∈ L such that
t
Y (t)D(t) = D(t)EQ [Y (T )  Ft ] = E[Y (T )D(T )  Ft ]. (6. Lemma 6. Proof.6)
and that the unique market price of risk −γ given by (6. NOARBITRAGE PRICING
Complete Markets The main problem is to determine which claims are attainable.
1 ψ−Yγ D 1 ψ−Yγ D
˜ =:ψ
.3. Then the model is complete in the sense that any contingent claim X with X/S0 (T ) ∈ L1 (FT . that is d ≤ n and rank[σ] = d. D D D 1 ψ − Y γ · γ dt D 1 D =− 1 1 γ dW + γ D D
2
dt.3. Deﬁne the Qmartingale Y (t) := EQ [X/S0 (T )  Ft ] . This is most conveniently carried out in terms of discounted prices.7) is attainable.
Y (t)D(t) = Y (0) +
0
ψ(s) dW (s).
8)
φi dZi =
i=1 i=1
˜ ˜ ˜ ˜ ˜ ˜ ˜ φi Zi σi dW = (σ −1 )T ψ · σ dW = ψ · σ −1 σ dW = ψ dW = dY. Theorem 6.) Pricing In the above complete model the fair price prevailing at t ≤ T of a T claim X which satisﬁes (6. models can be generically incomplete (as real markets are). σ is nondegenerate. (6.
(This theorem requires the ﬁltration (Ft ) to be generated by W . Zi
(6. These conditions are in fact equivalent (see for example MR[19](Chapter 10)). 3. However. Theorem 6. HEDGING AND PRICING Now deﬁne φi = then it follows that
n n
75 ˜ ((σ −1 )T ψ)i . and then the pricing becomes a difﬁcult issue. and the topic beyond the scope of this course. the model is complete. see (6.6).3. The following are equivalent: 1.3.
Hence φ yields an admissible strategy with discounted value process satisfying
n
˜ V (T .3. φ) = S0 (t)EQ [X/S0 (T )  Ft ] . The literature on incomplete markets is huge. 2.9)
0
Hence nondegeneracy of σ (see (6. there exists a unique ELMM Q.6.9) ˜ V (t. φ) = Y (T ) = EQ [X/S0 (T )] +
i=1
T
φi (s) dZi(s) = X/S0 (T ).8)) implies uniqueness of Q and completeness of the model.10)
We shall often encounter complete models.6) and (6.3 (Representation Theorem). Every Plocal martingale M has a continuous version and there exists ψ ∈ L such that
t
M (t) = M (0) +
0
ψ(s) dW (s).7) is given by (6. (6. φ) = S0 (t)V (t.2 (Completeness).
.
. in particular. for the price at t = 0 Y (0) = E[Xπ(T )].10) price of X at t =: Y (t) = S0 (t)EQ [X/S0 (T )  Ft ] .g. T ) = E π(T ) S0 (t)  Ft = E Q  Ft . Sn is still arbitragefree (why?). . . S0 (t) dP t
 Ft
and.76
CHAPTER 6. the market price of risk) and then price a T claim X satisfying (6. π(t) S0 (T )
Also one can check (→ exercise) that if Q is an EMM then Si π are Pmartingales. This is a consistent pricing rule in the sense that the enlarged market Y. S0 . Now deﬁne π(t) := By Bayes formula we then have Y (t) = S0 (t)EQ [X/S0 (T )  Ft ] = S0 (t) = E [Xπ(T )  Ft ] . → exercise) P (t. . π(t) E
X dQ  S0 (T ) dP FT dQ  dP Ft
1 dQ F . The price of a T bond for example is (if 1/S0 (T ) ∈ L1 (Q). for short rate models) to exogenously specify a particular ELMM Q (or equivalently. NOARBITRAGE PRICING
Stateprice Density It is a custom (e. This is why π is called the stateprice density process. .7) according to (6.
T ) . MR[19](Chapter 12). T ))t∈[0. 77 t ∈ [0.
. The ﬁltration (Ft )t≥0 is generated by a ddimensional Brownian motion W .
• all zerocoupon bond prices (P (t.Chapter 7 Short Rate Models
→ B[3](Chapters 16–17). • noarbitrage: there exists an EMM Q.T ] are adapted processes (with P (T. such that P (t. T ) = 1 as usual). We assume that • the short rates follow an Itˆ process o dr(t) = b(t) dt + σ(t) dW (t) determining the savings account B(t) = exp
t 0
r(s) ds . P). F . As in the last section we ﬁx a stochastic basis (Ω. where P is considered as objective probability measure. B(t) is a Qmartingale for all T > 0. T ].1
Generalities
Short rate models are the classical interest rate models. etc
7.
1). (7. B(t)
Proof. T ). T ]. T1 ). Under the above assumptions. P (t. To be more general one would have to consider strategies involving a continuum of bonds. T ) dW (t) P (t.1)
(compare this to the last section).3. T ) This illustrates again the role of the market price of risk −γ as the excess of instantaneous return over r(t) in units of volatility. P (t. Tn ).3) that the T bond price satisﬁes under the objective probability measure P dP (t. Let −γ denote the corresponding market price of risk dQ F Et (γ · W ) = dP t ˜ and W = W − γ dt the implied QBrownian motion. . T ) (7. say P (t.3)
and hence
P (t. T ) = EQ e−
T t
r(s) ds
 Ft
(7.1. the process r satisﬁes under Q ˜ dr(t) = (b(t) + σ(t) · γ(t)) dt + σ(t) dW (t).78
CHAPTER 7. T )Et σ γ · W . T ) = (r(t) − γ(t) · σ γ (t. .1. t ∈ [0. SHORT RATE MODELS
According to the last chapter. such that dP (t. Exercise (proceed as in the Completeness Lemma 6. the existence of an ELMM for all T bonds excludes arbitrage among every ﬁnite selection of zerocoupon bonds. for any T > 0 there exists an adapted Rd valued process σ γ (t. It follows from (7. Proposition 7. T ) ˜ = P (0. T )) dt + σ γ dW (t). .2)
Moreover. T ) ˜ = r(t) dt + σ γ (t. This can be done (see [4] or Mike Tehranchi’s PhD thesis 2002) but is beyond the scope of this course.
. For convenience we require Q to be an EMM (and not merely an ELMM) because then we have P (t. .
Ft )Brownian motion. Let Z ⊂ R be a closed interval. r(t)) dW (t) (7. Q). this is also reﬂected by the nonuniqueness of the EMM (the market price of risk). where now Q is considered as martingale measure. Q can be any equivalent probability measure Q ∼ P.4) admits a unique Zvalued solution r = r ρ with
t t
r(t) = ρ +
0
b(u.2.1). We let W denote a ddimensional (Q. According to the Completeness Theorem 6. and b and σ continuous functions on R+ × Z. see also (7.3. However. A short rate model is not fully determined without the exogenous speciﬁcation of the market price of risk. All contingent claims can be priced by taking Qexpectations of their discounted payoﬀs. r(t)) dt + σ(t. DIFFUSION SHORT RATE MODELS
79
In a general equilibrium framework.2.5)
. F . (Ft )t≥0 . we are unable to identify the market price of risk.7. In other words. Ingersoll and Ross (85) [7]). We assume that for any ρ ∈ Z the stochastic diﬀerential equation (SDE) dr(t) = b(t. r(u)) dW (u)
and such that exp −
t
T
r(u) du ∈ L1 (Q)
(7. It is custom (and we follow this tradition) to postulate the Qdynamics (Q being the EMM) of r which implies the Qdynamics of all bond prices by (7.3). and hence the savings account B(t). the savings account alone cannot be used to replicate bond payoﬀs: the model is incomplete.2
Diﬀusion Short Rate Models
We ﬁx a stochastic basis (Ω. the market price of risk is given endogenously (as it is carried out in the seminal paper by Cox. we started by specifying the Pdynamics of the short rates. r(u)) du +
0
σ(u.
7. The market price of risk (and hence the objective measure P) can be inferred by statistical methods from historical observations of price movements. A priori. Since our arguments refer only to the absence of arbitrage between primary securities (bonds) and derivatives.
T ) = EQ exp −
t
r(u) du  Ft . SHORT RATE MODELS
for all 0 ≤ t ≤ T . usually referred to as Feynman–Kac formula.
(7. Let T > 0 and Φ be a continuous function on Z. r) + b(t. or
2.
It turns out that P (t. In the following we write a(t. r) := for the diﬀusion term of r(t). and
T
F (t. Condition (7. r)∂r F (t. M is uniformly bounded.2. and assume that F = F (t.
σ(t.5) is always satisﬁed if Z ⊂ R+ . r)∂r F (t. T ] × Z
2 ∂t F (t. Lemma 7.7)
. A good reference for SDEs is the book of Karatzas and Shreve [14] on Brownian motion and stochastic calculus. Notice that (7.80
CHAPTER 7. r) ∈ C 1. T ]×Z) is a solution to the boundary value problem on [0. uniformly in t. r(t)) ∈ L2 [0. t and T . r) in r. Suﬃcient for the existence and uniqueness is Lipschitz continuity of b(t. ∂r F (t. This is a general property of certain functionals of Markov process. r) 2
2
(7. T ]. r(t))e−
t 0
r(u) du
.5) allows us to deﬁne the T bond prices
T
P (t. then M is a true martingale. T ]. r(t)) = EQ exp −
t
r(u) du Φ(r(T ))  Ft . r) and σ(t. r) + a(t. r(t))e−
t 0
r(u) du
σ(t. r) − rF (t. If in addition either 1. uniformly in t.6)
Then M (t) = F (t. T ) can be written as a function of r(t).
is a local martingale.
t ≤ T.
t ∈ [0. If d = 1 then H¨lder continuity of order 1/2 o of σ in r. r) = Φ(r). is enough. r) = 0 F (T.1.2 ([0.
2. r(t). One can also show the converse that the expectation on the right hand side of (7.
Multiplying with e
t 0
r(u) du
We call (7.
0
r(u) du Φ(r(T ))  Ft .7)) equals F (t. We can apply Itˆ’s formula to M and obtain o dM (t) = ∂t F (t. r(t)).6) the term structure equation for Φ. r)∂r F (t. r(t))e−
t 0
t 0
r(u) du
σ(t.2. r(t)) dW (t)
r(u) du
σ(t. T ] × Z).2 ([0. we have only shown that if a smooth solution F of (7. for Φ ≡ 1 we get the T bond price P (t. T ) = F (t.7. Is it computationally eﬃcient? Solving PDEs numerically in more than three dimensions causes diﬃculties. we have found a pricing algorithm. Since T M (T ) = Φ(r(T ))e− 0 r(u) du we get F (t. T ) as a function of t. DIFFUSION SHORT RATE MODELS Proof. It is now clear that either Condition 1 or 2 imply that M is a true martingale. In particular. Strictly speaking. r) where F solves the term structure equation (7. r(t)) − r(t)F (t.2. r(t)) e−
t 0
81
r(u) du
dt
+ ∂r F (t. r(t))
2 + a(t. Its solution F gives the price of the T claim Φ(r(T )).7) conditional on r(t) = r can be written as F (t. This is general Markov theory and we will not prove this here.
Hence M is a local martingale. In any case. which in particular means that F may not be in C 1. r(t))e− = ∂r F (t. r(t))e
−
t 0
T r(u) du
= M (t) = EQ exp − yields the claim. The nuisance is that we have to solve a
. PDEs in less than three space dimensions are numerically feasible. r(t))∂r F (t.6) but usually only in a weak sense. T ). Remark 7. and the dimension of Z is one. r(t) (and T ) P (t. r(t)) dW (t).6) exists and satisﬁes some additional properties (Condition 1 or 2) then the time t price of the claim Φ(r(T )) (which is the right hand side of (7. r(t)) + b(t.
Cox–Ingersoll–Ross (CIR. b(t) ≥ 0. r(t) dW (t). Hence short rate models that admit closed form solutions to the term structure equation (7. If not otherwise stated.1
Examples
This is a (far from complete) list of the most popular short rate models. Dothan (1978): Z = R+ . SHORT RATE MODELS
PDE for every single zerocoupon bond price function F (·. Ho–Lee (1986): Z = R. Hull–White (extended CIR. 8. dr(t) = (b(t) + β(t)r(t)) dt + σ(t) dW (t). 1. d (t) = (b(t) + β(t) (t)) dt + σ(t) dW (t). Black–Karasinski (1991): Z = R+ . dr(t) = βr(t) dt + σr(t) dW (t). dr(t) = (b(t) + β(t)r(t)) dt + σ(t) r(t) dW (t). Black–Derman–Toy (1990): Z = R+ . From that we might want to derive the yield or even forward curve. 2. at least for Φ ≡ 1. 7.82
CHAPTER 7. For all examples we have d = 1. dr(t) = (b + βr(t)) dt + σ dW (t). the parameters are realvalued. are favorable. dr(t) = (b + βr(t)) dt + σ 3. If we do not impose further structural assumptions we may run into regularity problems.
7. 1985): Z = R+ .
. dr(t) = b(t) dt + σ dW (t). (t) = log r(t). 1990): Z = R. T > 0. 5. 1990): Z = R+ . Vasicek (1977): Z = R. ·. 4.2. 6. dr(t) = β(t)r(t) dt + σ(t)r(t) dW (t). T ).6). Hull–White (extended Vasicek. b ≥ 0.
β and σ P (0. β. β. T ) = F (0. T.4
Aﬃne Term Structures
Short rate models that admit closed form expressions for the implied bond prices F (t. T ) = 1 implies A(T.7.
. T ) = B(T.6). T ) − B(t. Conversely. r. Then the implied T bond price is a function of the current short rate level and the three model parameters b. r. r(0). the functions b(t) etc are fully determined by the empirical initial yield curve.6) to match a given initial yield curve.3. The most tractable models are those where bond prices are of the form F (t. Therefore the class of timeinhomogeneous short rate models (such as the Hull–White extensions) was introduced. But F (0. It turns out that it is often too restrictive and will provide a poor ﬁt of the current data in terms of accuracy (least squares criterion).
7. r. σ). one may want to invert the term structure equation (7. r(0). Notice that F (T. r(0). b. T ) = 0. T )r). T )
Once the short rate model is chosen. T. T ) are favorable. T ) = exp(−A(t. σ) is just a parametrized curve family with three degrees of freedom. Say we have chosen the Vasicek model. INVERTING THE YIELD CURVE
83
7. T ) = F (0. Such models are said to provide an aﬃne term structure (ATS).3
Inverting the Yield Curve
T → P (0. the initial term structure
and hence the initial yield and forward curve are fully speciﬁed by the term structure equation (7. for some smooth functions A and B. The nice thing about ATS models is that they can be completely characterized. Usually. By letting the parameters depend on time one gains inﬁnite degree of freedom and hence a perfect ﬁt of any given curve. b.
r) = a(t) + α(t)r and b(t. Hence we can solve (7. T ) − b(t)B(t. In fact.8). T ). α. T ) + α(t)B 2 (t. r) and b(t. r)B(t. This is the (Hull–White extension of the) CIR model. T2 ) is invertible. This is the (Hull–White extension of the) Vasicek model. β. ·) and B 2 (t.84
CHAPTER 7. B(T. ∂t B(t. Replace a(t. A(T. r) ≥ 0 and r(t) does not leave the state space Z.8) can be further speciﬁed. T ) = a(t)B 2 (t. T ) = exp(−A(t. r) by (7.8). T ) − B(t. α(t) ≥ 0 and b(t) ≥ 0 (otherwise the process would cross zero). so the left hand side of (7. T ) + (∂t B(t. ·) are linearly independent since otherwise B(t. The functions A and B in turn satisfy the system ∂t A(t.
.4) provides an ATS only if its diﬀusion and drift terms are of the form a(t. Black–Derman–Toy and Black–Karasinski models have an ATS. r). T2 ) −B(t. Looking at the list in Section 7. Z = R+ : necessarily a(t) = 0.2. T1 ) B 2 (t. t) = 0. it can be shown that every ATS model can be transformed via aﬃne transformation into one of the two cases 1.11) reads a(t)B 2 (t.6) and obtain a(t.11) for a(t. Hence we can ﬁnd T1 > T2 > t such that the matrix B 2 (t. T ) − b(t.1 we see that all short rate models except the Dothan. The functions a. T ) = α(t)B 2 (t. T ) − β(t)B(t. r) = b(t) + β(t)r. T ) = 0. T ) − β(t)B(t. Terms containing r must match.11)
The functions B(t. This proves the claim.8)
for some continuous functions a. which yields (7. T ) = 0. T ) − 1. α. r)B 2 (t. Z = R: necessarily α(t) = 0 and a(t) ≥ 0. T ) − b(t)B(t. β are arbitrary. T )r) in the term structure equation (7. SHORT RATE MODELS
Proposition 7. β in (7. and b. b. T1 ) −B(t. T ) = ∂t A(t. which trivially would lead to be above results with a(t) = α(t) ≡ 0.9) (7.1. and β is arbitrary. They have to be such that a(t. r) and b(t. (7.10)
Proof. (7. 2. (7. b. r. The short rate model (7. T ) + 1)r. T ) r.4. We insert F (t. ·) ≡ B(t.
T ).5
Some Standard Models
We discuss some of the most common short rate models. Equations (7. BM[6](Chapter 3)
7. The explicit solution is B(t.9)–(7. see Figure 7. so that also the objective Pdynamics of r(t) is of the above form.5. 2β
V ar[r(t)] = σ e Hence
2 2βt 0
e−2βs ds =
Q[r(t) < 0] > 0. → B[3](Section 17.1
Vasicek Model
dr = (b + βr) dt + σ dW
The solution to is explicitly given by (→ exercise) r(t) = r(0)eβt + b βt e − 1 + σeβt β
t 0
e−βs dW (s).10) become σ2 2 ∂t A(t. T ) = 0. B(T. which is not satisfactory (although this probability is usually very small). If β < 0 then r(t) is meanreverting with mean reversion level b/β. T ) = 0. T ) = 1 β(T −t) e −1 β
.
It follows that r(t) is a Gaussian process with mean E [r(t)] = r(0)eβt + and variance
t
b βt e −1 β σ 2 2βt e −1 . T ) − bB(t.4). and r(t) converges to a Gaussian random variable with mean b/β and variance σ 2 /(2β). 2 ∂t B(t.7. Vasicek assumed the market price of risk to be constant. SOME STANDARD MODELS
85
7. T ) − 1. for t → ∞. A(T. T ) = B (t. T ) = −βB(t.5.1.
09 (mean reversion level).
Short Rates 0.2
Cox–Ingersoll–Ross Model
dr(t) = (b + βr(t)) dt + σ r(t) dW (t).
. for b ≥ 0.86
CHAPTER 7.5. T ) ds 2 t t σ 2 4eβ(T −t) − e2β(T −t) − 2β(T − t) − 3 eβ(T −t) − 1 − β(T − t) +b .6). T ) − B(t. T ) − =−
∂s A(s. T ) = A(T. σ = 0.08 0.1 0. T )r(t)) .86.07 Time in Months
50
100
150
200
250
300
350
and A is given as ordinary integral
T
A(t.08. = 4β 3 β2
We recall that zerocoupon bond prices are given in closed form by P (t. r(0) ≥ 0.
It is worth to mention that.12 0.1: Vasicek short rate process for β = −0.0148 and r(0) = 0.
7. b/β = 0.11 0. T ) ds + b B(s. T ) = exp (−A(t. SHORT RATE MODELS
Figure 7. It is possible to derive closed form expression also for bond options (see Section 7.09 0. T ) ds
t
T σ2 T 2 B (s.
T ) − 1. The ATS equation (7. T ) = (γ − β) (eγ(T −t) − 1) + 2γ ∂t B(t. T ) = −
. T ) = where γ := β 2 + 2σ 2 . Moreover.
Hence also in the CIR model we have closed form expressions for the bond prices.7. B(T. 2 This is called a Riccati equation. as it is the case for the Hull–White extension. This is easily integrated r(t) = r(s) exp β − σ 2 /2 (t − s) + σ(W (t) − W (s)) . This also holds when the coeﬃcients depend continuously on t. for every r(0) ≥ 0. T ) = 0. this is mainly the reason why the CIR model is so popular.10) now becomes nonlinear σ2 2 B (t.
7.5. it can be shown that also bond option prices are explicit(!) Together with the fact that it yields positive interest rates. s ≤ t. SOME STANDARD MODELS
87
has a unique strong solution r ≥ 0.5.
Thus the Fs conditional distribution of r(t) is lognormal with mean and variance (→ exercise) E[r(t)  Fs ] = r(s)eβ(t−s)
2 (t−s)
V ar[r(t)  Fs ] = r 2 (s)e2β(t−s) eσ
−1 . It is good news that the explicit solution is known 2 eγ(T −t) − 1 B(t. T ) − βB(t. Even more. if b ≥ σ 2 /2 then r > 0 whenever r(0) > 0. The market price of risk is chosen to be constant.
. Integration yields 2b log σ2 2γe(γ−β)(T −t)/2 (γ − β) (eγ(T −t) − 1) + 2γ
A(t.3
Dothan Model
Dothan (78) starts from a driftless geometric Brownian motion under the objective probability measure P dr(t) = σr(t) dW P(t). which yields dr(t) = βr(t) dt + σr(t) dW (t) as Qdynamics.
.10) become σ2 2 B (t. The idea of lognormal rates is taken up later by Sandmann and Sondermann (1997) and many others. T ) − b(t)B(t. T ) = 0. T ). T ) = T − t.4
Ho–Lee Model
For the Ho–Lee model dr(t) = b(t) dt + σ dW (t) the ATS equations (7. Let ∆t be small. 2 ∂t B(t.
We face an expectation of the type E[exp(exp(Y ))] where Y is Gaussian distributed. Similarly. SHORT RATE MODELS
The Dothan and all lognormal short rate models (Black–Derman–Toy and Black–Karasinski) yield positive interest rates. T ) = 0. T ) = Hence B(t. But no closed form expressions for bond prices or options are available (with one exception: Dothan admits an “semiexplicit” expression for the bond prices. But such an expectation is inﬁnite.
σ2 (T − t)3 + 6
T t
b(s)(T − s) ds. one shows that the price of a Eurodollar future is inﬁnite for all lognormal models.
7. T ) = −1. A(t. This means that in arbitrarily small time the bank account growths to inﬁnity in average. see BM[6]). then
∆t
E[B(∆t)] = E exp
r(s) ds
0
≈ E exp
r(0) + r(∆t) ∆t 2
. ∂t A(t. B(T. which ﬁnally led to the so called market models with lognormal LIBOR or swap rates.9)–(7.88
CHAPTER 7.5. T ) = − A(T. A major drawback of lognormal models is the explosion of the bank account.
We can also integrate this expression to get P (t.
t
Let f ∗ (0. t) +
σ 2 t2 + σW (t). T ) = ∂T A(t. t) + σ 2 t(T − t) + r(t). T ) + ∂T B(t.5. However.5
Hull–White Model
The Hull–White (1990) extensions of Vasicek and CIR can be ﬁtted to the initial yield and volatility curve. r(t) ﬂuctuates along the modiﬁed initial forward curve. T ) = e−
T t 2
f ∗ (0. In this model we choose the constants β and σ to obtain a nice volatility structure whereas b(t) is chosen in order to match the initial yield curve. T ) − f ∗ (0. t) = E[r(t)] − σ 2 t2 . gives a perfect ﬁt of f ∗ (0. T )r(t) = − σ2 (T − t)2 + 2
T
89
b(s) ds + r(t). Equation (7. T ) − 1. Plugging this back into the ATS yields f (t. B(T. T ) = −βB(t.
It is interesting to see that
t
r(t) = r(0) +
0
b(s) ds + σW (t) = f ∗ (0. We therefore restrict ourself to the following extension of the Vasicek model that was analyzed by Hull and White 1994 dr(t) = (b(t) + βr(t)) dt + σ dW (t).5. T ) be the observed (estimated) initial forward curve. T ). 2
That is. T ) is just as in the Vasicek model ∂t B(t.t)(T −t)− σ2 t(T −t)2 −(T −t)r(t)
. and we have f ∗ (0. this ﬂexibility has its price: the model cannot be handled analytically in general.10) for B(t. 2
7. T ) = f ∗ (0. SOME STANDARD MODELS The forward curve is thus f (t. s) + σ 2 s. T ) = 0
.7.s) ds+f ∗ (0. Then b(s) = ∂s f ∗ (0.
T ) = ∂T A(0.90 with explicit solution
CHAPTER 7. β
Equation (7. T ) + ∂T B(0. T ) ds +
0 2 0 2 T
b(s)∂T B(s.
7. T ) =
1 β(T −t) e −1 . Plugging in and performing performing some calculations eventually yields f (t. T )r(0) b(s)eβ(T −s) ds + eβT r(0) . T ) = f ∗ (0.
=:φ(T )
σ = − 2 eβT − 1 2β
=:g(T )
+
0
The function φ satisﬁes ∂T φ(T ) = βφ(T ) + b(T ). T )) f ∗ (0. T ) ds +
t t
b(s)B(s. T ) + ∂T B(0.6
Option Pricing in Aﬃne Models
We show how to price bond options in the aﬃne framework. T ) − eβ(T −t) f ∗ (0. T ) + g(T )) − β(f ∗ (0.
. T ) + g(T )). T ) = − σ2 2
T T
B 2 (s. The procedure has to be carried out rigorously from case to case. The discussion is informal. SHORT RATE MODELS
B(t. It follows that b(T ) = ∂T φ(T ) − βφ(T ) = ∂T (f ∗ (0. we do not worry about integrability conditions.9) for A(t. T ) now reads A(t. t) − + eβ(T −t) r(t). σ 2 β(T −t) e −1 2β 2 eβ(T −t) − eβ(T +t) φ(0) = r(0). T ) ds
We consider the initial forward curve (notice that ∂T B(s. T ) = −∂s B(s. T )r(0) = σ2 2
T T
∂s B 2 (s.
λ)r(t) .
.S))r(T )
e
= e−A(T. λ) φ(T. λ) = λ. Since discounted zerocoupon bond prices are martingales we obtain for T ≤ S (→ exercise) E e−
S 0 T t
r(s) ds −λr(T )
e
 Ft = e−φ(t.S)−φ(0. λ) − β(t)ψ(t.S)r(T ) −λr(T )
e
e
= e−A(T. T. T. T.6.
r(s) ds −λr(T )
e
= E e−
T 0
r(s) ds −A(T. and indeed.S))−ψ(0. T.S)−φ(0. the so called Sforward measure. T ) = e−A(t. λ) − 1 ψ(T. Let λ ∈ C. we have φ(t.T. λ) − b(t)ψ(t.S)−ψ(0. T.T )−B(t.6). T.λ+B(T.T.S))r(0) .S)))r(0) . T.T. λ) = α(t)ψ 2 (t.7.λ+B(T.T )r(t) .9)–(7. S)
S
deﬁnes an equivalent probability measure QS ∼ Q on FS . λ) = 0 ∂t ψ(t. T. Now let t = 0 (for simplicity only). λ) = a(t)ψ 2 (t. Hence we have shown that the (extended) Laplace transform of r(T ) with respect to QS is EQS e−λr(T ) = eA(0. T.
91
This looks much like the ATS equations (7.T. T ).λ)−ψ(t. and φ and ψ be given as solutions to ∂t φ(t. one sees that E e− In fact. T ) and ψ(t. 0) = B(t.λ+B(T. 0) = A(t. But e− 0 r(s) ds dQS = dQ P (0.T.T.S) E e−
T 0
r(s) ds −(λ+B(T. T.S)−B(T. diﬀusion term a(t) + α(t)r and ATS P (t.S))+(B(0. OPTION PRICING IN AFFINE MODELS Let r(t) be a diﬀusion short rate model with drift b(t) + β(t)r.S)−A(T. by plugging the right hand side below in the term structure equation (7.λ+B(T.10).
S)r(T ) 1{r(T )≤r∗ } − K1{r(T )≤r∗ } A(T. α = 0). S. SHORT RATE MODELS
By Laplace (or Fourier) inversion.6.and T forward measure. r) 1 (T. Its price today (t = 0) is π = E e−
T 0
r(s) ds
e−A(T. For β = −0. S) + log K . S)Φ where
1 (T. one gets the distribution of r(T ) under QS .S)r(T ) − K
+
. S)
T 0
r(s) ds
= P (0. We now consider a European call option on a Sbond with expiry date T < S and strike price K. Vasicek or CIR) this distribution is explicitly known (e.S)r(T ) − K where r ∗ = r ∗ (T. A similar closed form expression is available for the price of a put option.1
Example: Vasicek Model (a. In some cases (e. B(T. b/β = 0.g. In general. The pricing of the option boils down to the computation of the probability of the event {r(T ) ≤ r ∗ } under the S. T.
1{r(T )≤r∗ } − KE e−
r(s) ds
1{r(T )≤r∗ }
7. and hence an explicit price formula for caps. b. r(0)) 2 (T )
1 β eβT (b + βr) − b − a 2 − eβ(S−T ) − 2eβT + eβ(S+T ) β2 a 2βT e −1 2 (T ) := β :=
and Φ(x) is the cumulative standard Gaussian distribution function.g. S. Gaussian or chisquare). T )Φ
1 (T.
r∗ − r∗ −
We obtain (→ exercise) π = P (0.86.92
CHAPTER 7. K) := − Hence π = E e−
S 0
+
= e−A(T.S)−B(T.S)−B(T.
The payoﬀ can be decomposed according to e−A(T. S)QS [r(T ) ≤ r ∗ ] − KP (0. this is done numerically. T )QT [r(T ) ≤ r ∗ ].09
.S)−B(T. r(0)) 2 (T )
− KP (0. β const. S.
Table 7.2 (→ exercise).0689651 0.0121906 0.025847 0.01503 0.0815358 0. σ = 0. the Vasicek model cannot produce humped volatility curves.0402203
Figure 7. one gets the ATM cap prices and Black volatilities shown in Table 7.0148 and r(0) = 0. OPTION PRICING IN AFFINE MODELS
93
(mean reversion level).00907115 0.14 0.0326962 0.129734 0.1: Vasicek ATM cap prices and Black volatilities.1.0613624 0.12 0.08 0.0562337 0.7. In contrast to Figure 2.0743607 0.0199647 0.6.0525296 0.0443967 0.1.1 0.
0.0416089 ATM vols 0.028963 0.0370565 0. as in Figure 7.017613 0.2: Vasicek ATM cap Black volatilities.08.0915455 0.106348 0.1 and Figure 7.00567477 0.0647515 0.0485755 0.06 0.04 5 10 15 20 25 30
.0221081 0.00215686 0. Maturity 1 2 3 4 5 6 7 8 10 12 15 20 30 ATM prices 0.
SHORT RATE MODELS
.94
CHAPTER 7.
95
. 1992) [9]. Jarrow and Morton (HJM.Chapter 8 Heath–Jarrow–Morton (HJM) Methodology
→ original article by Heath.
96
CHAPTER 8. HJM METHODOLOGY
.
It is called the T forward measure. T )B(T ) For t ≤ T we have dQT P (t. t ∈ [0. dQ P (0. T ) . B(t) are strictly positive martingales. T ) dQT F t = E Q  Ft = . Since
we can deﬁne an equivalent probability measure QT ∼ Q on FT by dQT 1 = . 97
.6.Chapter 9 Forward Measures
We consider the HJM setup (Chapter 8) and directly focus on the (unique) EMM Q ∼ P under which all discounted bond price processes P (t.
9. T ].1
T Bond as Numeraire
1 1 > 0 and EQ =1 P (0. T )B(T ) P (0. T )B(t) This probability measure has already been introduced in Section 7. T )B(T )
Fix T > 0. dQ dQ P (0.
T ) B(s)
=
P (s. namely the T bond. FORWARD MEASURES
t ∈ [0. which in most cases turns out to be rather hard work.T )B(t) P (t.
T
To compute π(t) we have to know the joint distribution of exp − t r(s) ds and X. T ) is a QT martingale. S)  Fs = P (t.T )B(s)
 Fs
P (s.S) B(s) P (s.98 Lemma 9. FT ).T ) P (s. Since Q is related to the riskfree asset. if r is deterministic) we would have π(t) = P (t.S) P (0.T ) P (0.1. S) .1) B(T ) Its fair price at time t ≤ T is then given by π(t) = EQ e−
T t
r(s) ds
X  Ft . Thus we have to compute a double integral. T )
We thus have an entire collection of EMMs now! Each QT corresponds to a diﬀerent numeraire. P (t. for instance. T ) =
P (t. Bayes’ rule gives EQ T EQ P (t. S) .T ) P (t. let X be a T claim such that X ∈ L1 (Q.
Proof. since • we only have to compute the single integral EQ [X  Ft ]. and integrate with respect to that distribution. (9. one usually calls Q the risk neutral measure. a much nicer formula. Let s ≤ t ≤ S ∧ T . For any S > 0.1. P (t. T forward measures give simpler pricing formulas.
. S ∧ T ]. If B(T )/B(t) and X were independent under Q (which is not realistic! it holds. P (s. T )EQ [X  Ft ] .
CHAPTER 9. Indeed.
T ) dW (s).2. but under QT : Proposition 9. T ) as conditional expectation of the future short rate r(T ). T ) = f (0.2. The good news is that the above formula holds — not under Q though.1)).2). T )EQT [X  Ft ] .3) (9. To see that. T ) +
0
σ(s.2)
which proves (9. AN EXPECTATION HYPOTHESIS
99
• the bond price P (t.
9.4)
. Proof.3). (9. u) du ds +
s 0
σ(s. T )B(T ) (9. T ) can be observed at time t and does not have to be computed. T )B(T ) P (t. Then EQT [X] < ∞ and π(t) = P (t. That is.9. T )EQT [X  Ft ] . T ) EQT [X  Ft ] = P (0. X < ∞ (by (9. T )B(t)EQ X  Ft P (0. T ) ·
σ(s.1. And π(t) = P (0. The forward rates then follow the dynamics
t T t
f (t. P (0. the expression of the forward rates f (t.2
An Expectation Hypothesis
Under the forward measure the expectation hypothesis holds. T )B(t) P (0. Bayes’s rule yields EQT [X] = EQ whence (9. Let X be a T claim such that (9. T )B(t) = P (t.1) holds. we write W for the driving QBrownian motion.
t ∈ [0. the expectation hypothesis holds under the forward measures f (t.
−
·
σ(·. Equation (9. T ) = f (0. T )Et B(t) We thus have dQT F = E t dQ t Girsanov’s theorem applies and
t T T T
−
·
σ(·.1. T ) = P (0. u) du · W
. Under the above assumptions. FORWARD MEASURES
The Qdynamics of the discounted bond price process is P (t. T ) = EQT [r(T )  Ft ] .
. T ) dW T (s). T ) = P (0. u) du ds. T ))t∈[0.
Hence. u) du · W
.2. T ) − B(s)
T
σ(s. u) du dW (s). T ]. T ) +
0
σ(s.
(9.
s
(9.4) now reads
t
f (t. if
T
EQ T then
σ(s.6)
W T (t) = W (t) +
0 s
σ(s.T ] Summarizing we have thus proved
is a QT martingale.
Lemma 9.
is a QT Brownian motion.5)
This equation has a unique solution P (t. T ) + B(t)
t 0
P (s.100
CHAPTER 9. T )
0
2
ds < ∞
(f (t.
Its price at time t = 0 (for simplicity only) is π = E Q e−
T 0
r(s) ds
(P (T. T ) exp
0
v(s. S) − K)+ . T ) and hence
t t
P (t.
T
(9. S) 1(P (T. S) P (0. T ) = P (t.T )
(9. T ) := − We also know that
P (t. S) ≥ K) − KEQ B(T )−1 1(P (T.3
Option Pricing in Gaussian HJM Models
We consider a European call option on an Sbond with expiry date T < S and strike price K. S)
σT. S)QS [P (T. S) ≥ K] − KP (0. u) du.T ) P (t. T ) P (0.3.S) t∈[0. T ) = r(t) dt + v(t.S (s) dW S (s) −
1 2
S
t
σT. In fact (→ exercise) P (t.S (s) dW (s) − 2
t 0
t
v(s.8)
.S) P (t.6 and decompose π = EQ B(T )−1 P (T. T )
0
2
− v(s.9. T ) = P (0. S)
t
t∈[0. T )QT [P (T. We already know that dP (t.S (s) := v(s. u) du.S (s)
0
ds
σT. S) ≥ K] .T ] t
σ(t. S) ≥ K) = P (0.7) is a
is a QS martingale and
QT martingale.T ]
× exp = where
0
1 σT.
We proceed as in Section 7. T ) 2
2
ds
where v(t. T ) exp P (0. S)
2
2
ds
P (0. This option pricing formula holds in general.
P (t. T ) dW (t) P (t. OPTION PRICING IN GAUSSIAN HJM MODELS
101
9. T ) dW (s) +
0 T
r(s) −
1 v(s. T ) − v(s. S) =
σ(s.
S)Φ[d1 ] − KP (0.3. FORWARD MEASURES
× exp − =
0
1 σT. S) = P (t.8) and Φ is the standard Gaussian CDF.
σT. S) ≥ K] = QT 1 P (T. T ) = (σ1 (t. the option price is π = P (0. T )
This suggests to look at those models for which σT.S is deterministic. P (T. S) ≥ K] = QS QT [P (T. . T )
2
2
ds
P (0.S (s)
ds
. Proof.S (s) is given in (9.S) KP (0.S)
T 0 1 2 T 0 2
σT. Under the above Gaussian assumption. Proposition 9. T ).S) forward measures. . S)
0
2
− v(s.S (s)
2
ds
σT. T ) are Gaussian distributed. σd (t.T ) are lognormally distributed under the respective P (T.S) P (0.
Now observe that QS [P (T. log d1.T ) − log P (0. . and hence forward rates f (t. . S) K P (T.S (s) dW (s) − 2
t 0
t
v(s. T )
t
CHAPTER 9. We thus assume that σ(t. T )Φ[d2 ].2 =
P (0.T ) T 0
where
±
1 2
T 0
σT. We obtain the following closed form option price formula.S (s)
2
2
ds . T )
σT.S (s)
ds
σT. S) ≥K . and P (T.S) hence P (T. S) exp − P (0. T ) ≤ P (T. It is enough to observe that log P (T.T ) + P (T. T )) are deterministic functions of t and T . S) P (t.S (s)
0
ds .S (s) dW T (s) −
1 2
t
σT.1.T ) and P (T. T ) P (0.102 and P (0.
S) log P (T.
. respectively.3.1 (→ exercise).S (s)
2
ds
σT.3.S (s)
ds
are standard Gaussian distributed under QS and QT . OPTION PRICING IN GAUSSIAN HJM MODELS and
T 0 2
103
P (0.9.S) P (T.T ) + T 0
1 2
σT. the Vasicek option price formula from Section 7.6.1 can now be obtained as a corollary of Proposition 9. Of course.T ) − log P (0.
FORWARD MEASURES
.104
CHAPTER 9.
thus EQ e − This is equivalent to f (t.
r(s) ds
Y  Ft
. The underlying is in both cases a T claim Y. T. with time of delivery T > t. an interest rate. an index. any traded or nontraded asset. 105
T t T t
r(s) ds
(Y − f (t.1
Forward Contracts
A forward contract on Y. This can be an exchange rate.
10. the forward price is chosen such that the present value of the forward contract is zero. or Hull (2002) [10] We discuss two common types of term contracts: forwards. Y) = 1 EQ e − P (t. • at t. and with the forward price f (t. and futures. for some ﬁxed future date T . which are actively traded on many exchanges. contracted at t. a commodity such as copper. T. Y) is deﬁned by the following payment scheme: • at T . T ) = EQT [Y  Ft ] . which are mainly traded OTC.Chapter 10 Forwards and Futures
→ B[3](Chapter 20). the holder of the contract (long position) pays f (t. etc. Y) and receives Y from the underwriter (short position). Y))  Ft = 0. T. T.
They are required to keep a certain amount of money as a safety margin. which makes the futures contract on Y. t] the holder of the contract receives (or pays. The volumes in which futures are traded are huge.
.2
Futures Contracts
A futures contract on Y with time of delivery T is deﬁned as follows: • at every t ≤ T . T. if negative) the amount F (t. • at T . One of the reasons for this is that in many markets it is diﬃcult to trade (hedge) directly in the underlying object. an Sbond delivered at T ≤ S is
P (t. Holding a (short position in a) futures does not force you to physically deliver the underlying object (if you exit the contract before delivery date). gas or electricity. there is a market quoted futures price F (t. the holder of the contract (long position) pays F (T . and selling short makes it possible to hedge against the underlying. etc. T. P (t. T. or a commodity such as copper. This might be an index which includes many diﬀerent (illiquid) instruments. Y). T. T. any traded asset S delivered at T is
The forward price f (s. Y) − F (s.S) . a dollar delivered at T is 1.T ) S(t) .T )
3. T. which is EQ e −
T s
r(u) du
(Y − f (t.
10. • during any time interval (s. Y) (this is called marking to market). if entered at t. Y))  Fs = E Q e−
T s
r(u) du
Y  Fs − P (t. Y). So there is a continuous cashﬂow between the two parties of a futures contract.106
CHAPTER 10. Y) and receives Y from the underwriter (short position). FORWARDS AND FUTURES
Examples The forward price at t of 1. 2. T. equal to zero. P (t. T )f (t. Y) has to be distinguished from the (spot) price at time s of the forward contract entered at time t ≤ s.
Y) is a continuous semimartingale (or Itˆ process). . . .F B
T
=
s t
1 dF (s). N . Y) = EQ [Y  Ft ] . T ]. is given by EQ [Σ  Ft ] where
N
Σ :=
i=1
1 (F (ti ) − F (ti−1 )) . our model economy is driven by Brownian motion and changes in a continuous way. T. Indeed. The present value of the corresponding cashﬂows F (ti ) − F (ti−1 ) at ti . B(s) t
. (10. T ]. T. Then the futures price process is given by the Qmartingale F (t. We can rewrite Σ as
N
i=1
1 (F (ti ) − F (ti−1 )) + B(ti−1 )
N
i=1
1 1 − B(ti ) B(ti−1 )
(F (ti ) − F (ti−1 )) . Under the appropriate integrability assumptions (uniform integrability) we conclude that T 1 EQ dF (s)  Ft = 0. . this is just how futures prices are deﬁned . We now give a heuristic argument for (10. i = 1.2. First.1) based on the above characterization of a futures contract.
If we let the partition become ﬁner and ﬁner this expression converges in probability towards
T t
1 dF (s) + B(s)
T
d
t
1 . FUTURES CONTRACTS
107
Suppose Y ∈ L1 (Q). let t = t0 < · · · < tN = t be a partition of [t.1) Often. B(s)
since the quadratic variation of 1/B (ﬁnite variation) and F (continuous) is zero. Now suppose we enter the futures contract at time t < T . hence EQ [Σ  Ft ] = 0.10. We face a continuum of cashﬂows in the interval (t. B(ti )
But the futures contract has present value zero. Hence there is no reason to believe that futures prices evolve discontinuously. and we may assume that o F (t) = F (t. This has to hold for any partition (ti ).
T ) [100 per cent] where LF (t. T ]. F
T]
<∞
F (t) =
0
1 dM (s). The Eurodollar futures contract is tied to the LIBOR.2.4) Interest rate futures contracts may be divided into futures on short term instruments and futures on coupon bonds. Fix a maturity date T and let L(T ) denote the 3months LIBOR for the period [T. is deﬁned by 1 F (t. L(T )) = 1 − LF (t. prevailing at T .108 and that
t
CHAPTER 10. LIBOR is the interbank rate of interest for Eurodollar loans. June. T ) tends to L(T ). B(s)
t ∈ [0.1). T ]. T. T ) is the corresponding futures rate (compare with the example in Section 4. M B(s)2
t
s
= EQ [ F.3
Interest Rate Futures
→ Z[27](Section 5. moreover
T
EQ then
0
1 d M. B(s)
t ∈ [0.2). If. FORWARDS AND FUTURES
M (t) =
0
1 dF (s) = EQ B(s)
T 0
1 dF (s)  Ft . which implies (10. T + 1/4].
10. The futures price. dollars]. The market quote of the Eurodollar futures contract on L(T ) at time t ≤ T is 1 − LF (t. As t tends to T . T ) [Mio. used for the marking to market. The maturity (delivery) months are March. 4
.
is a Qmartingale. Eurodollars are deposits of US dollars in institutions outside of the US. September and December. LF (t.
is a Qmartingale. We only consider an example from the ﬁrst group. It was introduced by the International Money Market (IMM) of the Chicago Mercantile Exchange (CME) in 1981. and is designed to protect its owner from ﬂuctuations in the 3months (=1/4 years) LIBOR.
At maturity there is no physical delivery. On the other hand. T.4
Forward vs. S(T )) = EQ [S(T )  Ft ]. settlement is made in cash. S(T )) = S(t) . Futures in a Gaussian Setup
Let S be the price process of a traded asset.4. T. P (t. T. T.4. L(T )) = 1 − 1 L(T ) = Y is not 4 P (T. T ) leads to a cashﬂow of 106 × 10−4 × 1 = 25 [dollars]. L(T )) 4
1 = EQ [F (T . Fix a delivery date T . FUTURES IN A GAUSSIAN SETUP
109
Consequently. the 4 underlying Y is a synthetic value. T ) are deterministic functions in t. a change of 1 basis point (0. T + 1/4) = 1 − 1 L(T )P (T. In fact. 4
We also see that the ﬁnal price F (T . T ) = −
σ(t. Hence the Qdynamics of S is of the form dS(t) = r(t) dt + ρ(t) dW (t). S(t) for some volatility process ρ. Proposition 10. 4
we obtain an explicit formula for the futures rate LF (t.01%) in the futures rate LF (t. since 1 1 − LF (t.
Under Gaussian assumption we can establish the relationship between the two prices.1. Suppose ρ(t) and v(t. T ) = F (t. T + 1/4) as one might suppose. T. FORWARD VS. where
T
v(t. Instead. T ) = EQ [L(T )  Ft ] . T ) F (t.
10. u) du
t
. The forward and futures prices of S for delivery at T are f (t.10. L(T ))  Ft ] = 1 − EQ [L(T )  Ft ] .
T. S(T )) = EQ [f (T . FORWARDS AND FUTURES
is the volatility of the T bond (see (9. T )
t t 0
µ(s) dW (s) −
1 2
t
µ(s)
0
2
ds
× exp and hence
0
µ(s) · v(s. T. T ) is negative d S. T ) − ρ(s)) · v(s. T ) ds
for t ≤ T . T ) dt
t
= S(t)P (t. P (·. T. S(T )) = f (t. if the instantaneous correlation of dS(t) and dP (t. T. T ) ds . S(T )) = S(0) exp − P (0. Hence. S(T ))  Ft ]
T
= f (t. T )ρ(t) · v(t. It is clear that f (t. S(T )) = f (t. T ) ds .
By assumption µ(s) is deterministic. T. S(T )) exp
t
µ(s) · v(s.
as desired.
T
EQ exp − and
t
µ(s) dW (s) −
1 2
T
µ(s)
t
2
ds
 Ft = 1
F (t. S(T )) exp −
T
t
1 µ(s) dW (s) − 2
T
µ(s)
t
2
ds
× exp
t
µ(s) · v(s. Consequently. S(T )) exp
t
(v(s. Write µ(s) := v(s. T. T. T.7)). T ) ds .
T
f (T . T ) − ρ(s). Proof.110
CHAPTER 10.
. T ) ≤ 0
then the futures price dominates the forward price. Then
T
F (t.
In a Gaussian HJM framework (σ(t. S) for t ≤ T < S.u) du)ds
−1
σ(s.v) dv·
S s
σ(s. if
σ(s. FUTURES IN A GAUSSIAN SETUP Similarly. one can show (→ exercise)
111
Lemma 10. u) du ds.4. v) then futures rates are always greater than the corresponding forward rates. T ) et( − (S − T )P (t. S)  Ft ] T P (t. T ) = EQ [r(T )  Ft ] −
t
σ(s. FORWARD VS. u) ≥ 0 for all s ≤ min(u.
s
F (t. v) · σ(s. Hence. T ) ·
S T
σ(s. S) = EQ [F (T.10.4. T ) deterministic) we have the following relations (convexity adjustments) between instantaneous and simple futures and forward rates
T T
f (t. T.2.
.
FORWARDS AND FUTURES
.112
CHAPTER 10.
T ). This is too restrictive from two points of view: • statistically: the evolution of the entire yield curve is explained by a single variable. S). A (m)factor model is an interest rate model of the form f (t. u)du σ(t. S) d P (·. S) =
T t T t
σ(t. u)du
= 1. P (·. since the driving (Markovian) factor. T ) = H(T − t. P (·. r(t)). T ). Fix m ≥ 1 and a closed set Z ⊂ Rm (state space). The inﬁnitesimal increments of all bond prices are perfectly correlated d P (·. is onedimensional.
• analytically: the family of attainable forward curves H = {H(·. r)  r ∈ R} is only onedimensional. u)du
σ(t.Chapter 11 MultiFactor Models
We have seen that every timehomogeneous diﬀusion short rate model r(t) induces forward rates of the form f (t. u)du σ(t. P (·. Z(t)) 113
. we now allow for multiple factors. This a onefactor model. To gain more ﬂexibility. T ) = H(T − t. for some deterministic function H. r(t). T )
t t t S t S t
d P (·.
F . z) ds . where
x
Π(x. Z(t)). Here W is a ddimensional Brownian motion deﬁned on a ﬁltered probability space (Ω. . (Ft ). . . Timeinhomogeneous models are included in the above setup. z) := exp −
H(s.
0
Notice that the short rates are now given by r(t) = H(0. for all z0 ∈ Z.Z z0 (s)) ds
t∈[0. Q). dZ(t) = b(Z(t)) dt + ρ(Z(t)) dW (t) Z(0) = z0 . b1 ≡ 1 and ρ1j ≡ 0 for j = 1.114
CHAPTER 11. (A3) the above SDE has a unique Zvalued solution Z = Z z0 . for every z0 ∈ Z. Z z0 (t)) e
t 0
H(0. We assume that (A1) H ∈ C 1. (A2) b : Z → Rm and ρ : Z → Rm×d are continuous functions. MULTIFACTOR MODELS
where H is a deterministic function and Z (state process) is a Zvalued diﬀusion process. for all z0 ∈ Z. d). Simply set Z1 (t) = t (that is. T ) = Π(T − t. Z z0 (t)). (A4) Q is the risk neutral local martingale measure for the induced bond prices P (t.T ]
is a Qlocal martingale.
. Hence the assumption (A4) is equivalent to (A4’) Π(T − t. satisfying the usual conditions.2 (R+ × Z). .
Z(t)) du. Z(t)) +
m
i=1
bi (Z(t))∂zi H(T − t.s.2 (R+ × Z) we can apply Itˆ’s o formula and obtain
m
df (t.l=1
akl (Z(t))∂zi H(T − t. T ) =
− ∂x H(T − t. for all t ≤ T and initial points z0 = Z(0). Z(t))ρij (Z(t)).
. . . Z(t))
+
d m
1 aij (Z(t))∂zi ∂zj H(T − t.
where a(z) := ρ(z)ρT (z). z) → H(x. NOARBITRAGE CONDITION
115
11.
The HJM drift condition now reads
m
− ∂x H(T − t. d.11.1)
σj (t.
j = 1. z) is in C 1. Hence the induced forward rate model is of the HJM type with
m
(11.
This has to hold a.l=1 m
ρkj (Z(t))ρlj (Z(t))∂zi H(T − t. T ) =
i=1
∂zi H(T − t.1
NoArbitrage Condition
Since the function (x. Z(t)) 2 i. Z(t)) du
=
k. Z(t)) +
m
i=1
bi (Z(t))∂zi H(T − t.j=1
d
+
i=1 j=1
∂zi H(T − t. Z(t))
+
m
1 aij (Z(t))∂zi ∂zj H(T − t.j=1
T
=
j=1 k. Z(t))
∂zi H(u − t. Z(t))ρij (Z(t)) dWj (t). Z(t))
T t
t
∂zi H(u − t.1. . . Letting t → 0 we thus get the following result. Z(t)) dt 2 i.
z) du
m x x
1 = ∂x aij (z) 2 i. Then there exists a sequence 0 ≤ x1 < · · · < xM such that the M × M matrix with kth row vector built by ∂zi H(xk . First. This is an inverse problem. Let z ∈ D. the last expression in (11. z) du
0 0
∂zj H(u. z) 2 i j
x
∂zi H(u. z) du. one takes b and ρ (and hence a) as given and looks for a solution H for the PDE (11. Or. z)
m
+
i. one takes H as given (an estimation method for the yield curve) and tries to ﬁnd b and a such that (11. z) 2 i j
· 0
∂zi H(u. z) and 1 ∂z ∂z H(xk .j=1
aij (z)
1 ∂z ∂z H(x.1.1.1. It turns out that the latter approach is quite restrictive on possible choices of b and a. Proof.116
CHAPTER 11.2). are linearly independent for all z in some dense subset D ⊂ Z. Suppose that the functions ∂zi H(·. z) ∈ R+ × Z. Then b and a are uniquely determined by H.2) is satisﬁed for all (x.
0
. z) du. there is equivalence between (A4) and
m
∂x H(x. MULTIFACTOR MODELS
Proposition 11. z) − ∂zi H(·.1). z) and 1 ∂z ∂z H(·. z)
i. Proposition 11. z) − ∂zi H(·.2) can be written as
m x
aij (z)∂zi H(x.2). by symmetry.j=1
∂zi H(u. z).
There are two ways to approach equation (11. z) du
0
(11.1 (Consistency Condition).2. where a is deﬁned in (11. z) − ∂zi H(x. Set M = m + m(m + 1)/2. the number of unknown functions bk and akl = alk . z) =
i=1
bi (z)∂zi H(x. z) du . Notice that. z) 2 i j
xk
∂zi H(u. Remark 11.2) for all (x.j=1 0
∂zi H(u.
for 1 ≤ i ≤ j ≤ m. Under the above assumptions (A1)–(A3).3.
3) yields
m m
g0 (x)−g0 (0)+
i=1
zi (gi (x)−gi (0)) =
i=1
bi (z)Gi (x)−
1 aij (z)Gi (x)Gj (x). Remark 11.
We ﬁrst look at the simplest. a(z).
i.j=1
(11. b(z) and a(z) are uniquely determined by (11. is invertible.2
Aﬃne Term Structures
H(x. z)  z ∈ Z} is used for daily estimation of the forward curve in terms of the state variable z.2.3) where Gi (x) :=
0
gi (u) du. statistical calibration is only possible for the drift of the model (or equivalently. AFFINE TERM STRUCTURES
117
for 1 ≤ i ≤ j ≤ m.
Integrating (11.1. and (11. of Z is not aﬀected by any Girsanov transformation. then the diﬀusion coeﬃcient. 2 i. Thus. Consequently. since the observations of z are made under the objective measure P ∼ Q. any Qdiﬀusion model Z for z is fully determined by H.2). This holds for each z ∈ D.j=1 (11. If Ft = FtW is the Brownian ﬁltration. By continuity of b and a hence for all z ∈ Z. Suppose that the the parametrized curve family H = {H(·.4)
m
. where dQ/dP is left unspeciﬁed by our consistency considerations.11. for the market price of risk). z) = g0 (x) + g1 (x)z1 + · · · gm (x)zm . namely the aﬃne case:
Here the second order zderivatives vanish. under the stated assumption.2) reduces to
m m
∂x g0 (x) +
i=1
zi ∂x gi (x) =
i=1
1 bi (z)gi (x) − ∂x 2
x
m
aij (z)Gi (x)Gj (x) . Then the above proposition tells us that.4.
11.
4 for the onefactor case. whence Z1 (t) is the (nonMarkovian) short rate process. .
(11. . . Notice that we have the freedom to choose g0 (0). .4) for b and a. This extends what we have found in Section 7.7)
.ij zk .4) and matching constant terms and terms containing zk s we obtain a system of Riccati equations
m
∂x G0 (x) = g0 (0) +
i=1 m
1 aij Gi (x)Gj (x) bi Gi (x) − 2 i. . 2 i. .3
Polynomial Term Structures
n
We extend the ATS setup and consider polynomial term structures (PTS) H(x. β.j=1 1 βki Gi (x) − αk. Gm . A typical choice is g1 (0) = 1 and all the other gi (0) = 0.6)
∂x Gk (x) = gk (0) +
i=1
with initial conditions G0 (0) = · · · = Gm (0) = 0. G1 G1 . . . Gm Gm are linearly independent functions. G1 G2 . .ij Gi (x)Gj (x). . Since the left hand side is aﬃne is z. a and αk .118 Now if
CHAPTER 11.j=1
m
m
(11. Plugging this back into (11. .5) (11. which are related to the short rates by r(t) = f (t. . t) = g0 (0) + g1 (0)Z1 (t) + · · · + gm (0)Zm (t).
11.
k=1
aij (z) = aij +
for some constant vectors and matrices b. gm (0). MULTIFACTOR MODELS
G1 . z) =
i=0
gi (x) (Zt )i . we can invert and solve the linear equation (11. we obtain that also b and a are aﬃne
m
bi (z) = bi +
j=1 m
βij zj αk.
we denote the integral of gi by
x
where N := I =
1.1 (Maximal Degree Problem I). iN be a numbering of the set of multiindices
n
I = {i = (i1 . POLYNOMIAL TERM STRUCTURES
119
where we use the multiindex notation i = (i1 . . . that is. Here n denotes the degree of the PTS. . we now shall show the amazing result that n > 2 is not consistent with (11. 2 k. i2 . . . . For n = 2 we have a quadratic term structure (QTS). . As above. . 2}. im )  i ≤ n}. and deg a(z) = 0 if n = 2 (QTS) and deg a(z) ≤ 1 if n = 1 (ATS). 1 ≤ µ ≤ ν ≤ N . and that ρ ≡ 0. Proof. . . .3. Suppose that Giµ and Giµ Giν are linearly independent functions.
i=0
Gi (x) :=
0
gi (u) du.l=1 ∂zk ∂zl 1 ∂z i ∂z j akl (z) Aij (z) = Aji (z) := . . Do we gain something by looking at n = 3 and higher degree PTS models? The answer is no.3. Then necessarily n ∈ {1.
µ. there exists an index i with i = n and gi = 0. . i = i1 + · · · + im and i im z i = z11 · · · zm .2) can be rewritten
N N N m m
(11. n} and k ∈ {1. . b(z) and a(z) are polynomials in z with deg b(z) ≤ 1 in any case (QTS and ATS). .10)
.
Theorem 11.8) (11. which has also been studied in the literature. Thus for n = 1 we are back to the ATS case. Deﬁne the functions 1 ∂2zi ∂z i akl (z) + Bi (z) := bk (z) ∂zk 2 k. .9)
µ=1
giµ (x) − giµ (0) z iµ =
µ=1
Giµ (x)Biµ (z) −
Giµ (x)Giν (x)Aiµ iν (z). Moreover.ν=1
(11. . For µ ∈ {1. . Let i1 . im ). m} we write (µ)k for the multiindex with µ at the kth position and zeros elsewhere. . In fact.2).11. .l=1 ∂zk ∂zl Equation (11.
for some constant C ∈ R+ . and thus Bi (z) and Aij (z) are polynomials in z of order less than or equal n.
k ∈ {1. m}.3. l ∈ {1.
k.12)
We may assume that akk ≡ 0. If n = 1 there is nothing more to prove. deg a(z) ≤ n. m}. Now let n = 2. 2A(1)k (1)l (z) = akl (z). m}.13) k. . l ∈ {1.
. MULTIFACTOR MODELS
By assumption we can solve this linear equation for B and A. where degµ denotes the degree of dependence on the single component zµ . . . and hence deg a(z) = 0. Hence degl akl (z) ≤ 1. In particular. . . and b and ρ satisfy a linear growth condition b(z) + ρ(z) ≤ C(1 + z ). since ρ ≡ 0.12) yields deg k akk (z) = 0. We ﬁnally have B(1)k +(1)l (z) = bk (z)zl + bl (z)zk + akl (z). . . l ∈ {1. . Equation (11. ∀z ∈ Z.12) cannot be a polynomial in z of order less than or equal n unless n ≤ 2. .(1)k +(1)l (z) = akk (z)zl2 + 2akl (z)zk zl + all (z)zk .
k. from which we conclude that deg b(z) ≤ 1. . . m}.
(11. Consider
2 2A(1)k +(1)l .11)
hence b(z) and a(z) are polynomials in z with deg b(z). we have B(1)k (z) = bk (z). This proves the ﬁrst part of the theorem. . We can relax the hypothesis on G in Theorem 11. .
(11. (11. . .
From the preceding arguments it is now clear that also deg l akk (z) = 0. But then the right hand side of (11.1 if from now on we make the following standing assumptions: Z ⊂ Rm is a cone. An easy calculation shows that
2n−2 2A(n)k (n)k (z) = akk (z)n2 zk . Notice that by deﬁnition degµ akl (z) ≤ (degµ akk (z) + degµ all (z))/2. .120
CHAPTER 11.
11.3. POLYNOMIAL TERM STRUCTURES Theorem 11.3.2 (Maximal Degree Problem II). Suppose that a(z)v, v ≥ k(z) v 2 , for some function k : Z → R+ with
z∈Z, z →∞
121
∀v ∈ Rm ,
(11.14)
lim inf k(z) > 0.
(11.15)
Then necessarily n ∈ {1, 2}. Conditions (11.14) and (11.15) say that a(z) becomes uniformly elliptic for z large enough. Proof. We shall make use of the basic inequality z i  ≤ z This is immediate, since z i  = z i Now deﬁne
N i
,
∀z ∈ Rm .
im
(11.16)
z1  z
i1
zm  ··· z
≤ 1,
∀z ∈ Rm \ {0}.
Γk (x, z) :=
µ=1 N
Giµ (x) Giµ (x)
µ=1
∂z iµ ∂zk ∂ 2 z iµ . ∂zk ∂zl
(11.17)
Λkl (x, z) = Λlk (x, z) :=
(11.18)
Then (11.2) can be rewritten as (integration)
n m
i=0
(gi (x) − gi (0)) z =
i
bk (z)Γk (x, z)
k=1
+
1 akl (z) (Λkl (x, z) − Γk (x, z)Γl (x, z)) , 2 k,l=1 (11.19)
m
Suppose now that n > 2. We have from (11.17) Γk (x, z) =
i=n
Gi (x)ik z i−(1)k + · · · =: Pk (x, z) + · · · ,
122
CHAPTER 11. MULTIFACTOR MODELS
where Pk (x, z) is a homogeneous polynomial in z of order n − 1, and · · · stands for lower order terms in z. By assumptions there exist x ∈ R+ and k ∈ {1, . . . , m} such that Pk (x, ·) = 0. Choose z ∗ ∈ Z \{0} with Pk (x, z ∗ ) = 0 and set zα := αz ∗ , for α > 0. Then we have zα ∈ Z and Γk (x, zα ) = αn−1 Pk (x, z ∗ ) + · · · , where · · · denotes lower order terms in α. Consequently, Γk (x, zα ) Pk (x, z ∗ ) = = 0. α→∞ zα n−1 z ∗ n−1 lim (11.20)
Combining (11.14) and (11.15) with (11.20) we conclude that L := lim inf
α→∞
1 zα
2n−2
a(zα )Γ(x, zα ), Γ(x, zα ) ≥ lim inf k(zα )
α→∞
Γ(x, zα ) 2 > 0. (11.21) zα 2n−2
On the other hand, by (11.19),
n
L≤
i=0
gi (x) − gi (0)
zα
i zα 
2n−2
+
b(zα ) zα
Γ(x, zα ) 1 a(zα ) + 2n−3 zα 2 zα 2
Λ(x, zα ) , zα 2n−4
for all α > 0. In view of (11.17), (11.18), (11.13) and (11.16), the right hand side converges to zero for α → ∞. This contradicts (11.21), hence n ≤ 2.
11.4
ExponentialPolynomial Families
We consider the Nelson–Siegel and Svensson families. For a discussion of general exponentialpolynomial families see [8].
11.4.1
Nelson–Siegel Family
GN S (x, z) = z1 + (z2 + z3 x)e−z4 x .
Recall the form of the Nelson–Siegel curves
11.4. EXPONENTIALPOLYNOMIAL FAMILIES
123
Proposition 11.4.1. There is no nontrivial diﬀusion process Z that is consistent with the Nelson–Siegel family. In fact, the unique solution to (11.2) is a(z) = 0, b1 (z) = b4 (z) = 0, b2 (z) = z3 − z2 z4 , b3 (z) = −z3 z4 .
The corresponding state process is Z1 (t) ≡ z1 , Z2 (t) = (z2 + z3 t) e−z4 t , Z3 (t) = z3 e−z4 t , Z4 (t) ≡ z4 , where Z(0) = (z1 , . . . , z4 ) denotes the initial point. Proof. Exercise.
11.4.2
Svensson Family
GS (x, z) = z1 + (z2 + z3 x)e−z5 x + z4 xe−z6 x .
Here the forward curve is
Proposition 11.4.2. The only nontrivial HJM model that is consistent with the Svensson family is the Hull–White extended Vasicek short rate model dr(t) = z1 z5 + z3 e−z5 t + z4 z −2z5 t − z5 r(t) dt + √ z4 z5 e−z5 t dW ∗ (t),
where (z1 , . . . , z5 ) are given by the initial forward curve f (0, x) = z1 + (z2 + z3 x)e−z5 x + z4 xe−2z5 x and W ∗ is some Brownian motion. The form of the corresponding state process Z is given in the proof below. Proof. The consistency equation (11.2) becomes q1 (x) + q2 (x)e−z5 x + q3 (x)e−z6 x + q4 (x)e−2z5 x + q5 (x)e−(z5 +z6 )x + q6 (x)e−2z6 x = 0, (11.22)
z GS (x. q6 .2) are ∂x GS (x. . .
.124
CHAPTER 11. 1 e−z5 x −z5 x xe . . 6. z5 z2 q3 (x) = a66 (z) 4 x4 + · · · . z) = 6 −x e 0 3 −z6 x z4 x e x − z15 e−z5 x + z15 − zx − z12 e−z5 x + z12 5 5 5 . j ≤ 4. z) = (−z2 z5 + z3 − z3 z5 x)e−z5 x + (z4 − z4 z6 x)e−z6 x . z) = 0 (z2 x2 + z3 x3 )e−z5 x 0
x 0
∂zi ∂zj GS (x. . . (11. z2 q2 (x) = a55 (z) 3 x4 + · · · . z) du =
0 0 0 2 −z x . we assume for the moment that z5 = z 6 . z) = −z6 x xe (−z2 x − z3 x2 )e−z5 x −z4 x2 e−z6 x 0 −xe−z5 x 2 −z5 x −x e . z ∂z5 GS (x.
Straightforward calculations lead to
z GS (u. MULTIFACTOR MODELS
for some polynomials q1 . deg q5 deg q6 ≤ 3. z) = 0 for 1 ≤ i. z ∂z6 GS (x. . . .23)
Then the terms involved in (11. Indeed. z6 deg q4 . 1 1 x −z6 x + z2 − z6 − z 2 e 6 6 2z3 z3 z3 2 z2 z2 z2 2z3 −z5 x − z2 − z3 x + z5 + z 2 x + z 2 + z 3 e z5 5 5 5 5 5 z4 2 z x + 2z24 x + 2z34 e−z6 x − z4 3 z6 z z
6 6 6
q1 (x) = −a11 (z)x + · · · . z5 + z6 = 0 and zi = 0 for all i = 1.
z5 1 1 q5 (x) = a34 (z) x2 + · · · . + z5 z6 1 q6 (x) = a44 (z) x2 + · · · . b3 (z) = −z5 z3 .11.
. The remaining terms are q2 (x) = (b3 (z) + z3 z5 )x + b2 (z) − z3 − a22 (z) + z 2 z5 . deg q2 (x). expression (11. Because of (11.23) we know that the exponents −2z5 . Only a22 (z) is left as strictly positive candidate among the components of a(z).4. z5 while q1 = q5 = q6 = 0. If 2z5 = z6 then also a22 (z) = 0. Taking this into account. We are left with q1 (x) = b1 (z). . 1 q4 (x) = a22 (z) . Hence b1 (z) = a3j (z) = aj3 (z) = a4j (z) = aj4 (z) = 0 ∀j = 1. But a is a positive semideﬁnite symmetric matrix. b2 (z) = z3 + z4 − 25 z2 . deg q3 ≤ 1. b4 (z) = −2z5 z4 . 6. z6
Because of (11.23) we conclude that a11 (z) = a55 (z) = a66 (z) = 0. . . . z5
q3 (x) = (b4 (z) + z4 z6 )x − z4 . . EXPONENTIALPOLYNOMIAL FAMILIES
125
where · · · stands for lower order terms in x. . If 2z5 = z6 then the condition q3 + q4 = q2 = 0 leads to a22 (z) = z4 z5 . 6.22) simpliﬁes considerably.
q4 (x) = a33 (z)
1 2 x +··· . −(z5 + z6 ) and −2z6 are mutually diﬀerent. Hence a1j (z) = aj1 (z) = a5j (z) = aj5 (z) = a6j (z) = aj6 (z) = 0 ∀j = 1 .
Z(t)) = z1 + Z2 (t) satisﬁes dr(t) = z1 z5 + z3 e−z5 t + z4 z −2z5 t − z5 r(t) dt + √ z4 z5 e−z5 t dW ∗ (t). Z1 . Thus. the above results thus extend for all z ∈ Z. In that case we have. By continuity of a(z) and b(z) in z. Z4 (t) = z4 z −2z5 t and
d
dZ2 (t) = z3 e−z5 t + z4 z −2z5 t − z5 Z2 (t) dt +
ρ2j (t) dWj (t).23) holds is dense Z. Hence the corresponding short rate process r(t) = GS (0. In particular.23). Z3 (t) = z3 e−z5 t . we only have a nontrivial process Z if Z6 (t) ≡ 2Z5 (t) ≡ 2Z5 (0). all Zi ’s but Z2 are deterministic. writing shortly zi = Zi (0). Z1 (t) ≡ z1 . MULTIFACTOR MODELS
We derived the above results under the assumption (11. Z5 and Z6 are even constant. since a(z) = 0 if 2z5 = z6 . But the set of z where (11.126
CHAPTER 11. 2j
j=1
By L´vy’s characterization theorem we have that e
d
W ∗ (t) :=
j=1 0
t
√
ρ2j (s) dWj (s) z4 z5 e−z5 s
is a realvalued standard Brownian motion (→ exercise).
.
j=1
where ρ2j (t) (not necessarily deterministic) are such that
d
ρ2 (t) = a22 (Z(t)) = z4 z5 e−2z5 t .
the forward δperiod LIBOR for the future date T prevailing at time t is the simple forward rate L(t. One may want to model other rates. who succeeded to ﬁnd a HJM type model inducing lognormal LIBOR rates. T + δ) = P (t. continuously compounded rates. as in Chapter 9. as we have seen. T ) P (t. Both approaches lead to Black’s formula for either caps (LIBOR models) or swaptions (swap rate models). such as LIBOR. T + δ) = 1 δ P (t. T + δ) is a martingale for the (T + δ)forward measure QT +δ . directly. T ) −1 .Chapter 12 Market Models
Instantaneous forward rates are not always easy to estimate. The principal idea of both approaches is to chose a diﬀerent numeraire than the riskfree account (the latter does not even necessarily have to exist). T. T ) = F (t. who developed a framework for arbitragefree LIBOR and swap rate models not based on HJM. The breakthrough came 1997 with the publications of Brace–Gatarek– Musiela [5] (BGM). for a ﬁxed δ (typically 1/4 = 3 months). To start with we consider the HJM setup. P (t. Because of this they are usually referred to as “market models”. T ) σT. and Jamshidian [12]. There has been some eﬀort in the years after the publication of HJM [9] in 1992 to develop arbitragefree models of other than instantaneous. Recall that.8)) d P (t. T )/P (t. T + δ)
We have seen in Chapter 9 that P (t.T +δ (t) dW T +δ (t). In particular (see (9. P (t. T + δ) 127
.
T )
2 ds
1 2
2
ds . T )λ(t. Hence the QT +δ distribution of log L(T.T ) κ T t
δ(L(T. (12. T ) =
CHAPTER 12. T + δ)EQT +δ δ(L(T. T ) exp
s
λ(u. T )
s
2
du . δ
Now suppose there exists a deterministic Rd valued function λ(t. where log d1. T ) − κ)+  Ft = δP (t. T )
2
ds. which is equivalent to
t
L(t. T + δ) 1 = (δL(t. T ) + 1)σT. T ))) . T ) dW T +δ (u) −
1 2
t
λ(u. T ) = L(s. T ) 1 1 P (t. we get dL(t. T ) dW T +δ (t). T ) σT.128 Hence dL(t. T )
. T )) − κΦ(d2 (t. T ) conditional on Ft is Gaussian with mean 1 log L(t. T + δ) (L(t.T +δ (t) = λ(t.2 (t.1) δL(t.T +δ (t) dW T +δ (t). T ). T ) + 1 Plugging this in the above formula. T ) such that δL(t. T ) − κ)+  Ft
±
1 2
T t
λ(s. T + δ) δ P (t. T )
t
2
ds
λ(s.T +δ (t) dW T +δ (t) = δ P (t. T ) − 2 and variance
t T T
λ(s. settlement date T + δ and strike rate κ is thus EQ e −
T +δ 0
r(s) ds
= P (t. T ) :=
L(t.
λ(s. MARKET MODELS
P (t. T ) d σT.
The time t price of a caplet with reset date T . T )Φ(d1 (t. T ) = L(t.
for s ≤ t ≤ T .
This is just Black’s formula for the caplet price with σ(t)2 set equal to 1 T −t
T
λ(s.
Diﬀerentiating in T gives σ(t. as (→ exercise)
T +δ T
σ(t.7) There is a more direct approach to LIBOR models without making reference to continuously compounded forward and short rates.u) du
T +δ T
λ(t. u) du = 1 − e−
T +δ T
f (t. T ) + (f (t.6]. But do such HJM models exist? The answer is yes. T + δ) = σ(t. T )
t
2
ds.1. T ) = 0 for T ∈ [0. T ). once σ(t. This all has been carried out by BGM [5].12. Section 5. δ)).
12. In a sense. the numeraires accordingly being bond price processes.u) du
This is a recurrence relation that can be solved by forward induction.1) yields Black’s formula for caplet prices.
+ 1 − e−
f (t.u) du
λ(t. but the construction and proof are not easy. T ) ∂T λ(t. δ) (typically.
as introduced in Section 2. using the deﬁnition of σT. The idea is to rewrite (12. Z[27](Section 4. see also [8. ·) is determined on [0. σ(t.T +δ (t).
. we place ourselves outside of the HJM framework (although HJM is often implicitly adopted). MODELS OF FORWARD LIBOR RATES
129
and Φ is the standard Gaussian CDF.6! We have thus shown that any HJM model satisfying (12. T + δ) − f (t. This gives a complicated dependence of σ on the forward curve. Now it has to be proved that the corresponding HJM equations for the forward rates have a unique and wellbehaved solution.1). T ))e−
T +δ T
f (t. T ). Instead of the risk neutral martingale measure we will work under forward measures.1
Models of Forward LIBOR Rates
→ MR[19](Chapter 14).
TM −1 ) −1 L(0. TM ) which is of course equivalent to L(t. . TM −1 ) = δ P (0. TM −1 ) = 1 δ P (0. δL(t. MARKET MODELS
12. Tm ).
We proceed by backward induction and postulate ﬁrst that dL(t. M. deterministic function λ(t. TM −1 )λ(t. . . TM −1 ) − 1 Et λ(·. 1 P (0. . • an initial strictly positive and decreasing discrete term structure P (0.1.
T
We are going to construct a model for the forward LIBOR rates with maturities T1 .130
CHAPTER 12. .TM ] . .TM (t) := δL(t. Tm ).1
Discretetenor Case
We ﬁx a ﬁnite time horizon TM = M δ. Tm ). (Ft )t∈[0. where Ft = FtW M is the ﬁltration generated by a ddimensional Brownian motion W TM (t).
Now deﬁne the bounded (why?) Rd valued process σTM −1 . M − 1. . TM ) t ∈ [0. . TM −1 ) dW TM (t). TM ]. m = 0. t ∈ [0. Tm ]. TM −1 ) + 1 t ∈ [0. Tm+1 ) m = 0.
and hence strictly positive initial forward LIBOR rates L(0. m = 0. . Tm ) −1 . TM −1 ]. TM −1 ) · W TM . . P (0. . . M. TM −1 ). . P (0. Write Tm := mδ. . QTM ). and a probability space (Ω. . for some M ∈ N. Tm ) = 1 δ P (0. . TM −1 ) = L(t. which represents the volatility of L(t. bounded. The notation already suggests that QTM will play the role of the TM forward measure. We take as given: • for every m ≤ M − 1. an Rd valued.
. TM −1 . TM −1 ]. TM −1 ) λ(t. t ∈ [0. F .
12. TM −2 ) · W TM −1 . δL(t.TM · W TM . Hence we can postulate
0
σTM −1 . TM −2 ) = −1 . Tm ))t∈[0.1. dQTM −1 and the QTM −2 Brownian motion
t
W TM −2 (t) := W TM −1 (t) −
0
σTM −2 .
P (0. TM −2 ) dW TM −1 (t). TM −2 ) = 1 δ
t ∈ [0. P (0. This induces an equivalent probability measure QTM −1 ∼ QTM on FTM −1 via dQTM −1 = ETM −1 σTM −1 . δ P (0. TM −2 )λ(t.TM −1 (s) ds.Tm ] under their respective measures QTm . TM −2 ]. MODELS OF FORWARD LIBOR RATES
131
compare with (12. TM −1 ]. 1 P (0. L(t.TM −1 · W TM −1 .
Repeating this procedure leads to a family of lognormal martingales (L(t. TM −1 ) that is. TM −2 ]. TM −2 ]. dQTM and by Girsanov’s theorem
t
W TM −1 (t) := W TM (t) − is a QTM −1 Brownian motion.1). TM −2 ) + 1 t ∈ [0. TM −2 ) λ(t.
dL(t.TM −1 (t) := δL(t.
t ∈ [0.
t ∈ [0. TM −2 ) − 1 Et λ(·.
.TM (s) ds.
yielding an equivalent probability measure QTM −2 ∼ QTM −1 on FTM −2 via dQTM −2 = ETM −2 σTM −2 . TM −2 ). TM −1 )
and deﬁne the bounded Rd valued process σTM −2 . TM −2 ) L(0. TM −2 ) = L(t.
it is not possible to uniquely determine the continuous time dynamics of a bond price P (t. Tm−1 ) m=i+1 δL(Ti . Tm ) = δ dL(t. Tm ) = . Tm ) P (0. for 0 ≤ i < j ≤ m. Tm−1 ]. Tm−1 ) = Et σTm−1 . .
(12. Tm ) under any of the forward measures QTk . The knowledge of forward LIBOR rates for all maturities T ∈ [0. M . Tm−1 ) P (0.
. Tm−1 ]. Tm ) which is a QTm martingale. Tm ) in the discretetenor model of forward LIBOR rates.2)
However. LIBOR Dynamics under Diﬀerent Measures We are interested in ﬁnding the dynamics of L(t. TM −1 ] is necessary. Tm−1 ) = δL(t.132 Bond Prices
CHAPTER 12. P (t.Tm · W Tm . P (Ti . P (t. . Tm−1 ) σTm−1 . Tm−1 )λ(t. 1 P (Ti .
P (t. From this we can derive. Tj ) = P (Ti . we then can deﬁne the forward price process P (t. Tm−1 ) dW Tm (t) = we get that P (t. Tm−1 ) P (t. Tm−1 ) + 1. . Tm ) Since d P (t. Tm−1 ) := δL(t. Tm−1 ) + 1 m=i+1
j j
t ∈ [0. . MARKET MODELS
What about bond prices? For all m = 1. Tm )
t ∈ [0.Tm (t) dW Tm (t) P (t.
Tm )
m
σTl . Tm ) dL(t. Tm ) under QTk is given according to the three cases k <m+1: k =m+1: k >m+1: dL(t.12.Tk+1 · W Tk+1 = . Tm ) = λ(t. Tm ) = λ(t. Tm ) · L(t. Ti ]. L(t. Then its price π(t) at t ≤ Tm is given by π(t) = P (t. Lemma 12.
l=m+1
for t ∈ [0. Let X ∈ L1 (QTm ) be a Tm contingent claim. Tm )
k−1
σTl .1. Let 0 ≤ m ≤ M − 1 and 0 ≤ k ≤ M .2).
Here is a useful formula. This follows from the equality
j−1 t
W (t) = W (t) − for all 1 ≤ i < j ≤ M . Then the dynamics of L(t. Tm ) · L(t. m ≤ M . Tm )dW Tk (t).Tl+1 (t)dt + λ(t. Tk ∧ Tm ].
l=i 0
t ∈ [0. Tm ) dW Tm+1 (t).Tl+1 (s) ds. Proof.
l=k
dL(t.1. Tk ) P (t. Tm )EQTm [X  Ft ] X = P (t. Tn ) only for such t). Tk+1 Ft dQ P (0. since we know P (t. 0 ≤ j ≤ m. Tk+1 ) P (t.1.2. P (Tm . Notice that P (0. Tm ) = −λ(t. Tn )EQTn  Ft . Tm ) dW Tk (t). Tk+1 ) t ∈ [0. Tn ) for all m < n ≤ M (strictly speaking.
. Proof.1. Tk ) dQTk  = Et σTk .Tl+1 (t) dt + λ(t. Tk ]. MODELS OF FORWARD LIBOR RATES
133
Lemma 12. this formula makes sense only for t = Tj . Derivative Pricing
Ti
Tj
σTl . which can be combined with (12.
for some positive integers µ < ν ≤ M . Tk+1 ) k=m k=m = P (0. Tl ) − λ(t.3) is the arbitragefree price of X). MARKET MODELS
dQTm dQTk P (0. Tν (Tµ is the ﬁrst reset date and Tν the maturity of the underlying swap). Tm ) δL(t.1. This cannot be done analytically anymore. Tm )(L(Tµ . Tm ) . L(Tµ . the reasoning in Section 9. Tl ) + 1 2
2
dt
+ λ(t. Tm ) dW Tµ (t). Swaptions Consider a payer swaption with nominal 1. Tµ )EQTµ
m=µ
P (Tµ . . so one has to resort to numerical procedures.134 Hence
CHAPTER 12. Tl ) λ(t.
. strike rate K. since the ﬁrst equality was derived in Proposition 9.Tl+1 (t) we have o
m
d log L(t. Tµ+1 ). Tν−1 )
π(0) = δP (0. Tm ) =
λ(t.
under the measure QTµ . . Tm ) − K)
. Tk )  =  = Tk+1 Ft Tn F t dQ dQ P (0. Tk ) P (t. Tm ) ·
l=µ
1 δL(t. Notice that by Lemma 12. But even if there is no risk neutral but only forward measures. Its payoﬀ at maturity is
ν−1 +
δ
m=µ
P (Tµ . Tm ) − K)
. Itˆ’s formula and the deﬁnition of σTl . Tn )
n−1
n−1
Bayes’ rule now yields the assertion. . we assumed there the existence of a savings account.2 (strictly speaking. Tµ ). L(Tµ .
The swaption price at t = 0 (for simplicity) therefore
ν−1
+
To compute π(0) we thus need to know the joint distribution of L(Tµ . Tk+1 ) P (t. We sketch here the Monte Carlo method. P (0.1. .1 makes it clear that (9. Tm ) P (t. . . Tm )(L(Tµ .1. Tn ) P (t. . maturity Tµ and underlying tenor Tµ . Tµ+1 . .
12.1. MODELS OF FORWARD LIBOR RATES
135
for t ∈ [0, Tµ ] and m = µ, . . . , ν − 1. Write α(t, Tm ) for the above drift term, i and let ti = n Tµ , i = 0, . . . , n, n ∈ N large enough, be a partition of [0, Tµ ]. Then we can approximate
ti+1 ti+1
log L(ti+1 , Tm ) = log L(ti , Tm ) +
ti
α(s, Tm ) ds +
ti
λ(s, Tm ) dW Tµ (s)
≈ log L(ti , Tm ) + α(ti , Tm ) where ζm (i) :=
ti ti+1
1 + ζm (i), n
λ(s, Tm ) dW Tµ (s),
such that ζ(i) = (ζµ (i), . . . , ζν−1 (i)), i = 0, . . . , n − 1, are independent Gaussian (ν − µ)vectors with mean zero and covariance matrix
ti+1
Cov[ζk (i), ζl (i)] =
ti
λ(s, Tk ) · λ(s, Tl ) ds,
which can easily be simulated. Forward Swap Measure We consider the above payer swap with reset dates Tµ , . . . , Tν−1 and cashﬂow dates Tµ+1 , . . . , Tν (= maturity of the swap). The corresponding forward swap rate at time t ≤ Tµ is
P (t,Tν 1 − P (t,Tµ ) P (t, Tµ ) − P (t, Tν ) ) Rswap (t) = . = ν P (t,Tk ) δ ν P (t, Tk ) δ k=µ+1 P (t,Tµ ) k=µ+1
(12.3)
Since for any 0 ≤ l < m ≤ M P (t, Tl ) P (t, Tl ) P (t, Tm−1 ) = ··· = P (t, Tm ) P (t, Tl+1 ) P (t, Tm )
m−1
(1 + δL(t, Ti )) ,
i=l
Rswap (t) is given in terms of the above constructed LIBOR rates. Deﬁne the positive QTµ martingale D(t) := P (t, Tk ) , P (t, Tµ ) k=µ+1
ν
t ∈ [0, Tµ ].
136
CHAPTER 12. MARKET MODELS
This induces an equivalent probability measure Qswap ∼ QTµ , the forward swap measure, on FTµ by D(Tµ ) dQswap = . Tµ dQ D(0) Lemma 12.1.3. The forward swap rate process Rswap (t), t ∈ [0, Tµ ], is a Qswap martingale. Proof. Let 0 ≤ m ≤ M and 0 ≤ s ≤ t ≤ Tm ∧ Tµ . Then EQswap P (t, Tm ) 1 P (t, Tm )  Fs = E Q Tµ D(t)  Fs P (t, Tµ )D(t) D(s) P (t, Tµ )D(t) 1 P (s, Tm ) = . D(s) P (s, Tµ )
Now the lemma follows (set m = 0, µ) by (12.3). The payoﬀ at maturity of the above swaption can be written as δD(Tµ ) (Rswap (T0 ) − K)+ . Hence the price is π(0) = δP (0, Tµ )EQTµ D(Tµ ) (Rswap (T0 ) − K)+ = δP (0, Tµ )D(0)EQswap (Rswap (T0 ) − K)+
ν k=µ+1
=δ
P (0, Tk )EQswap (Rswap (T0 ) − K)+ .
Lemma 12.1.3 tells us that Rswap is a positive Qswap martingale and hence of the form dRswap (t) = Rswap (t)ρswap (t) dW swap(t), t ∈ [0, Tµ ],
for some Qswap Brownian motion W swap and some swap volatility process ρswap . Hence, under the hypothesis (H) ρswap (t) is deterministic,
12.1. MODELS OF FORWARD LIBOR RATES
137
we would have that log Rswap (Tµ ) is Gaussian distributed under Qswap with mean 1 Tµ swap ρ (t) 2 dt log Rswap (0) − 2 0 and variance
0 Tµ
ρswap (t)
2
dt.
The swaption price would then be
ν
π(0) = δ
k=µ+1
P (0, Tk ) (Rswap (0)Φ(d1 ) − KΦ(d2 )) ,
Rswap (0) K Tµ 0 Tµ 0
with log d1,2 :=
±
1 2
ρswap (t)
2
2
dt .
ρswap (t)
dt
1 2
This is Black’s formula with volatility σ 2 given by 1 Tµ
Tµ 0
ρswap (t)
2
dt.
However, one can show that ρswap cannot be deterministic in our lognormal LIBOR setup. So hypothesis (H) does not hold. For swaption pricing it would be natural to model the forward swap rates directly and postulate that they are lognormal under the forward swap measures. This approach has been carried out by Jamshidian [12] and others. It could be shown, however, that then the forward LIBOR rate volatility cannot be deterministic. So either one gets Black’s formula for caps or for swaptions, but not simultaneously for both. Put in other words, when we insist on lognormal forward LIBOR rates then swaption prices have to be approximated. One possibility is to use Monte Carlo methods. Another way (among many others) is now sketched below. We have seen in Section 2.4.3 that the forward swap rate can be written as weighted sum of forward LIBOR rates
ν
Rswap (t) =
m=µ+1
wm (t)L(t, Tm−1 ),
Tm−1 ) dW Tµ . Tk−1 )L(t.Tj−1 )
. Tm ) = wm (t) = D(t)P (t. Tk−1 )L(0. MARKET MODELS
1 1 · · · 1+δL(t. the variability of the wm ’s is small compared to the variability of the forward LIBOR rates. such that the quadratic variation of log Rswap (t) becomes approximatively deterministic ρ
swap
(t)
2
≈
wk (0)wl (0)L(0. ρswap (s)
t ∈ [0. log Rswap t dt ν wk (0)wl (0)L(t.Tm−1 ) 1+δL(t. Tl−1 ) . Tm−1 )λ(t. Tl−1 )λ(t. under the Tµ forward measure QTµ
ν
dRswap (t) ≈ (· · · ) dt +
wm (0)L(t.
According to empirical studies. and deﬁne the ˜ swap Q Brownian motion (L´vy’s characterization theorem) e W ∗ (t) :=
t d
0 j=1
ρswap (s) j dWjswap (s). Tµ )
CHAPTER 12.Tµ ) ν 1 1 j=µ+1 1+δL(t. Tk−1 ) · λ(t. Tk−1 ) · λ(t.138 with weights P (t.
m=µ+1
t ∈ [0.Tµ ) · · · 1+δL(t. Tl−1 ) . Tl−1 )λ(t.
.l=µ+1
In a further approximation we replace all random variables by their time 0 values. Tm−1 ). 2 Rswap (0) k. Tµ ]. Tµ ]. So that
ν
Rswap (t) ≈
wm (0)L(t. ≈ 2 Rswap (t) k.l=µ+1
ν
Denote the square root of the right hand side by ρswap (t).
m=µ+1
and hence.
We obtain for the forward swap volatility ρswap (t)
2
=
d log Rswap . We thus approximate wm (t) by its deterministic initial value wm (0).
B ∗ is a strictly increasing and predictable process with respect to the discretetime ﬁltration (FTm ). M − 1. for all i = 0. . M. implied savings account process B ∗ (0) := 1. . . Implied Savings Account Given the LIBOR L(Ti . M . . By construction. that is.1. B ∗ (Tm ) := (1 + δL(Tm−1 . Ti ) for period [Ti . Tk−1 ) · λ(t. since it originally appears in his book R[22].l=µ+1
wk (0)wl (0)L(0. Tm−1 ))B ∗ (Tm−1 ). Tk+1 )
m < n ≤ M. For all 0 ≤ m ≤ M we have EQTM [B ∗ (TM )  FTm ] = B ∗ (Tm ) .
n−1
m = 1. MODELS OF FORWARD LIBOR RATES Then we have dRswap (t) = Rswap (t) ρswap (t) dW ∗ (t) ≈ Rswap (t)˜swap (t) dW ∗ (t). TM )
. Lemma 12. They conclude that “the approximation is satisfactory in general”. Tl−1 )λ(t. .12. . . see BM[6](Chapter 8). . Tl−1 ) dt. . Tk−1 )L(0. 2 Rswap (0)
This is “Rebonato’s formula”. .
B ∗ (Tn ) = B ∗ (Tm )
k=m ∗
1 .
Hence B (Tm ) can be interpreted as the cash amount accumulated up to time Tm by rolling over a series of zerocoupon bonds with the shortest maturities available. P (Tm . we can deﬁne the discretetime. .1. ρ
139
Hence we can approximate the swaption price in our lognormal forward LIBOR model by Black’s swaption price formula where σ 2 is to be replaced by 1 Tµ
Tµ 0 ν
k. that is. B ∗ (Tm ) is FTm−1 measurable. P (Tk . The goodness of this approximation has been tested numerically by several authors. .4. Ti+1 ]. for all m = 1.
for T ∈ [0.1.1. Each forward LIBOR rate L(t. we now need a continuum of initial dates:
. Exercise.l
12. MARKET MODELS
EQTM [B ∗ (TM )P (0. TM )
Hence (Bayes again) EQ ∗ E Q TM B ∗ (Tk )  F Tk = B ∗ (Tl )
B ∗ (Tk ) B ∗ (Tl ) B ∗ (Tl ) P (Tl . TM ) > 0. Tl ) B ∗ (Tk ) is a Q∗ martingale with respect to (FTk ). T ) will follow a lognormal process under the forward measure for the date T + δ..TM )
k) P (Tk .
k=0.TM )
B ∗ (T
 F Tk
= P (Tk . so that we can deﬁne the equivalent probability measure Q∗ ∼ QTM on FTM by dQ∗ = B ∗ (TM )P (0. it is enough to ﬁll the gaps between the Tj s..4).. TM ). Given the discretetenor skeleton constructed in the previous section. = B ∗ (Tm ) TM F T m dQ P (Tm .4)
Indeed. Tl ). (12.4 we have for m ≤ M
dQ∗ P (0. The stochastic basis is the same as before.2
Continuoustenor Case
We now specify the continuum of all forward LIBOR rates L(t. Tl ) = EQ∗ B ∗ (Tk )  F Tk . B ∗ (Tl ) 0 ≤ k ≤ l ≤ M.4 yields in particular
CHAPTER 12.1. Put in other words. Lemma 12. dQTM Q∗ can be interpreted as risk neutral martingale measure since P (Tk . in view of Lemma 12.. TM )] = 1 and B ∗ (TM )P (0. In addition. TM −1 ]. TM )  . (12.140 Proof. T ).4) shows that for any 0 ≤ l ≤ M the discretetime process P (Tk .
which proves (12.
we focus on the forward measures for dates T ∈ [TM −1 . T ]. P (0. By monotonicity there exists a unique deterministic increasing function α : [TM −1 . Tm ). Second. T ) −1 . T + δ) T ∈ [0. MODELS OF FORWARD LIBOR RATES
141
• for every T ∈ [0. and hence an initial strictly positive forward LIBOR curve L(0. • an initial strictly positive and decreasing term structure P (0. TM ]. TM ]. TM −1 ]. M − 1. we construct a discretetenor model for L(t. T ∈ [0. However. T ) Then we have dQT dQ∗ B ∗ (TM )P (0. such that log B ∗ (T ) := (1 − α(T )) log B ∗ (TM −1 ) + α(T ) log B ∗ (TM ) satisﬁes P (0. an Rd valued. we are given the values of the implied savings account B ∗ (TM −1 ) and B ∗ (TM ) and the probability measure Q∗ . . T ). T ) = 1 δ P (0. . deterministic function λ(t. which represents the volatility of L(t. We do not have to take into account forward LIBOR rates for these dates. as in the previous section. T ) = EQ∗ 1 B ∗ (T ) . TM ]. T )
. TM ) dQT = = . bounded.12. t ∈ [0.
First. ∀T ∈ [TM −1 . m = 0. T ). 1] with α(TM −1 ) = 0 and α(TM ) = 1. TM ] → [0. since they are not deﬁned there. . T ) B we can deﬁne the T forward measure QT ∼ Q∗ on FT by dQT 1 . dQTM dQ∗ dQTM B ∗ (T )P (0. TM ].1. T ). strictly positive and 1 EQ ∗ =1 ∗ (T )P (0. Since (→ exercise) B ∗ (T ) is FT measurable. TM −1 ].
Let T ∈ [TM −1 . = ∗ ∗ dQ B (T )P (0. .
T ) = L(t. T + δ) This in turn deﬁnes the positive and bounded process σT. T ]. TM ) F t = E Q T M  Ft dQTM B ∗ (T )P (0. 1 P (0.TM · W TM . T ]. T ). T +δ dQ Hence we have (→ exercise) dQT dQT +δ dQT F t = F t F dQTM dQT +δ dQTM t = Et σT. we can now deﬁne the forward LIBOR process L(t. Third. Girsanov’s theorem tells us that
t
W T (t) := W TM (t) −
σT. δL(t. T ) λ(t. T ) L(0. T ) for any T ∈ [TM −2 . since T ∈ [TM −1 . δ P (0.
.
is a QT Brownian motion.TM (s) dW TM (s) −
1 2
t
σT. T ) dW T +δ (t). T ].142
CHAPTER 12. T )λ(t.TM · W TM . TM −1 ] as dL(t. for t ∈ [0.T +δ · W T +δ Et σT +δ.
for any T ∈ [TM −2 . MARKET MODELS
By the representation theorem for QTM martingales there exists a unique σT. T ) + 1 t ∈ [0.TM · W TM = Et σT. T ].T +δ (t) := δL(t. t ∈ [0.T +δ · W T +δ .TM ∈ L such that (→ exercise) dQT B ∗ (TM )P (0. TM −1 ] are now given by dQT = ET σT. TM ] was arbitrary. The forward measures for T ∈ [TM −2 . T ) = −1 .TM (s)
0
2
ds
= Et σT. T )
t
= exp
0
σT.
0
t ∈ [0.TM (s) ds. TM −1 ].
143
Proceeding by backward induction yields the forward measure QT and Q Brownian motion W T for all T ∈ [0. T ) are positive.S · W T .12. S) P (0. we obtain the zerocoupon bond prices for all maturities. TM ]. unless S − T = mδ for some integer m. T ) dQTM dQT t P (0. S) dQS := F P (t.T +δ + σT +δ. TM −1 ].1. it is reasonable to deﬁne (why?) the forward price process
T
P (t. S) Et −σT. This way. T )
Notice that now P (T.TM . P (0. = P (0. T ) for all T ∈ [0. T ].S · W T . S) dQS = F t F P (0. Indeed. MODELS OF FORWARD LIBOR RATES for any T ∈ [TM −2 .TM := σT. S) may be greater than 1.TM − σS. T ) where (→ exercise) σT. T ) P (0. and forward LIBOR rates L(t. TM −1 ]. for t = T we get P (T.TM . S) = P (0. where σT. there may be negative interest rates for other than δ periods. t ∈ [0.
. Hence even though all δperiod forward LIBOR rates L(t. for any 0 ≤ T ≤ S ≤ TM . In particular. T ) dQT t dQTM P (0.S := σT. S) ET −σT.
MARKET MODELS
.144
CHAPTER 12.
Usually one has to model objective (for the rating) and riskneutral (for the pricing) probabilities. the payoﬀ was certain. 145
. That is.Chapter 13 Default Risk
→ [24. • Transition probabilities: between credit ratings (credit migration). • Recovery rates: proportion of value delivered after default has occurred.1
Transition and Default Probabilities
There are three main approaches to the modelling of transition and default probabilities: • Historical method: rating agencies determine default and transition probabilities by counting defaults that actually occurred in the past for diﬀerent rating classes. [1]. which is reﬂected by a higher yield on the bond. etc.
13. Investors should be adequately compensated by a risk premium. So far bond price processes P (t. there was no risk of default of the issuer. This may be the case for treasury bonds. T ) had the property that P (T. T ) = 1. Corporate bonds however may bear a substantial risk of default. For the modelling of credit risk we have to consider the following risk elements: • Default probabilities: probability that the debtor will default on its obligations to repay its debt. Chapter 2].
R (y) . Default is when this value hits a certain lower bound. We brieﬂy discuss the ﬁrst two approaches in this section. 2. The formal deﬁnition of default and transition rates is the following. • Intensity based method: default is speciﬁed exogenously by a stopping time with given intensity process. objective information and credit analysis of obligors. and MR (y) is the number of issuers with rating R at beginning of year y which defaulted in that year. Usually they operate without government mandate and are independent of any investment banking ﬁrm or similar organization. Y1 ].R :=
Y1 y=Y0 MR. Deﬁnition 13. Goes back to Merton (1974) [18]. the obligor is set on the Rating Review List (Moody’s) or the Credit Watch List (S&P). for an Rrated issuer is dR :=
Y1 y=Y0 MR (y) . After issuance and assignment of the initial obligor’s rating.1. based on the time frame [Y0 .
13. 1.1. The number of Moody’s rated obligors has increased from 912 in 1960 to 3841 in 1997. Y1 NR (y) y=Y0
where NR (y) is the number of issuers with rating R at beginning of year y.2 below. The historical oneyear transition rate from rating R to R . The historical oneyear default rate. is trR. Y1 y=Y0 NR (y)
.1
Historical Method
Rating agencies provide timely. The intensity based method is treated in more detail in Section 13.146
CHAPTER 13. If there is a tendency observable that may aﬀect the rating. Y1 ]. based on the time frame [Y0 .1. Among the biggest US agencies are Moody’s Investors Service and Standard&Poor’s (S&P). the rating agency regularly checks and adjusts the rating. DEFAULT RISK
• Structural approach: models the value of a ﬁrm’s assets.
and MR.
.1.13. Transition rates are gathered in a transition matrix as shown in Table 13. extremely strong Aa1 Aa2 High quality Aa3 A1 A2 Strong payment capacity A3 Baa1 Baa2 Adequate payment capacity Baa3 Speculativegrade ratings Ba1 Likely to fulﬁll obligations Ba2 ongoing uncertainty Ba3 B1 B2 High risk obligations B3 Caa1 Caa2 Current vulnerability to default Caa3 Ca In bankruptcy or default or other marked shortcoming
147
where NR (y) is as above. depending on economic conditions.1: Rating symbols. S&P AAA AA+ AA AAA+ A ABBB+ BBB BBBBB+ BB BBB+ B BCCC+ CCC CCCCC C D Moody’s Interpretation Investmentgrade ratings Aaa Highest quality. TRANSITION AND DEFAULT PROBABILITIES Table 13.R (y) is the number of issuers with rating R at beginning of year y and R at the end of that year. Transition and default probabilities are dynamic and vary over time.2. The historical method has several shortcomings: • It neglects the default rate volatility.
09 0.01 0.03 0.84 5.29 0.81 0.49 0. based on the time frame [1980. The obligor (=the ﬁrm) defaults by T if the total market value of its assets V (T ) at T is less than its liabilities X.72 83.16 0.66 5.07 2.24 0.46 5. see http://ﬁnancialcounsel. F . The dynamics of V (t) is modelled as geometric Brownian motion dV (t) = µ dt + σ dW (t).25 91. V (t) t ∈ [0.26 4.23 7.26
• It neglects crosscountry diﬀerences and business cycle eﬀects. and hence underestimation of trR. Rating at end of year (R ) A BBB BB B
Initial rating (R) AAA AA A BBB BB B CCC
AAA
AA
CCC
D
93. • Rating agencies react too slow to change ratings.00 0. There is a systematic overestimation of trR.49 0.09 0.05 1.R for some R = R .83 89.03 0.01 0. T ) = P [V (T ) < X  Ft ] .00 0.15 0.2: S&P’s oneyear transition and default rates.62 3.
.2
Structural Approach
Merton [18] proposed a simple capital structure of a ﬁrm consisting of equity and one type of zero coupon debt with promised terminal constant payoﬀ X > 0 at maturity T .T ] .
13.94 0.06 0.00 0.83 0.40 0.08 0.66 91.03 0.
with respect to some stochastic basis (Ω.23 25.06 0.46 1.00 0.148
CHAPTER 13.72 6. (Ft )t∈[0.18 0.20 0.10 0.02 0. Ratings Performance 2000.1.76 5.44 0.00 0.44 6. T ]. DEFAULT RISK
Table 13.R and dR .91 10. Thus the probability of default by time T conditional on the information available at t ≤ T is pd (t.com/Articles/Investment/ARTINV00000692000Ratings.28 61.88 1.04 0. P).pdf).32 0.66 83.2000] (Standard&Poor’s.94 0.24 4.
A dynamic description of V is
t N (t)
V (t) = V (0) +
0
V (s) (µ ds + σ dW (s)) +
j=1
V (τj −) eZj − 1 .
If the ﬁrm value process V (t) is continuous. τ2 . is a sequence of i. It is clear that the distribution of log V (T ) conditional on Ft and N (T ) − N (t) = n is Gaussian with mean log V (t) + mn + µ − and variance nρ2 + σ 2 (T − t). .1. . To include “unexpected” defaults one has to consider ﬁrm value processes with jumps. .d. as in the Merton approach. Z2 . T ]. σ2 2 (T − t)
. Zhou (1997) models V (t) as jumpdiﬀusion process
N (T )
V (T ) = V (t)
j=N (t)+1
e Zj e
µ− σ 2
2
(T −t)+σ(W (T )−W (t))
. T ) = Φ log
X V (t)
149
t ∈ [0. 2 Then we have pd (t. T )T =t ) is zero. TRANSITION AND DEFAULT PROBABILITIES that is 1 V (T ) = V (t) exp σ(W (T ) − W (t)) + µ − σ 2 (T − t) .i. . Gaussian N (m. N and Zj are mutually independent. + the instantaneous probability of default (∂T pd (t. ρ2 ) distributed random variables. . It is assumed that W .
1 − µ − 2 σ 2 (T − t) .
where N (t) is a Poisson process with intensity λ and Z1 . . T ]. are the jump times of N .
where τ1 .13. √ σ T −t
t ∈ [0.
The precise conditions under which it must occur (such as hitting a barrier) are easily misspeciﬁed. T ].2
Intensity Based Method
Default is often a complicated event. The ﬂow of the complete market information is represented by a ﬁltration (Ft ) satisfying the usual conditions. t ∈ [0. t ∈ [0. P). such that Td = inf{t  V (t) ≤ X(t)}.
. T ]. N (T ) − N (t) = n] P [N (T ) − N (t) = n] Φ log
X V (t)
=
− mn − µ −
σ2 2
(T − t)
nρ2 + σ 2 (T − t)
First passage time models make this approach more realistic by admitting default at any time Td ∈ [0. The default time Td is assumed to be an (Ft )stopping time. F . In this case the conditional default probability is pd (t. In this section we focus directly on describing the evolution of the default probabilities pd (t. T ) = E [H(T )  Ft ] .150 Hence the conditional default probability pd (t.
13.
n e−λ(T −t) (λ(T − t)) n!
which can be determined by Monte Carlo simulation. T ) = P [log V (T ) < log X  Ft ] =
∞ n=0 ∞ n=0
CHAPTER 13. DEFAULT RISK
P [log V (T ) < log X  Ft . we ﬁx a probability space (Ω. T ]. T ) without deﬁning the exact default event. Formally. and not just at maturity T . bankruptcy occurs if the ﬁrm value V (t) hits a speciﬁed stochastic boundary X(t). That means. The above structural approach has the additional deﬁciency that it is usually diﬃcult to determine and trace a ﬁrm’s value process. The Ft conditional default probability is now pd (t. T ) = P [Td ≤ T  Ft ] . hence the rightcontinuous default process H(t) := 1{Td ≤t} is (Ft )adapted.
For every A ∈ Ft there exists B ∈ Gt such that A ∩ {Td > t} = B ∩ {Td > t}. Intuitively speaking.4. Let t ∈ R+ . H is a uniformly integrable submartingale. T ) = 1{Td ≤t} + E [A(T ) − A(t)  Ft ] .1.10]) there exists a unique (Ft )predictable1 increasing process A(t) such that M (t) := H(t) − A(t) is a (uniformly integrable) martingale (notice that A(t) = A(t ∧ Td )).2. By the Doob–Meyer decomposition ([14. t] × B where s < t and B ∈ Fs . or equivalently. A market participant with access to the partial market information Gt cannot observe whether default has occurred by time t (Td ≤ t) or not (Td > t).
1
(13. events in Ft are Gt observable given that Td > t. by the sets {0} × B where B ∈ F0 and (s. (D1) There exists a strict subﬁltration (Gt ) ⊂ (Ft ) (partial market information) and a (Gt )adapted process Λ such that A(t) = Λ(t ∧ Td ) and Ft = Gt ∨ Ht . where Ht := σ(H(s)  s ≤ t) and Gt ∨ Ht stands for the smallest σalgebra containing Gt and Ht .1)
. The formal statement is as follows. This nicely reﬂects the aforementioned diﬃculties to determine the exact default event in practice. Lemma 13.13. In other words. This formula is the best we can hope for in general. Hence pd (t. We proceed in several steps towards an explicit expression for pd (t.2. Proof. INTENSITY BASED METHOD
151
Obviously. Theorem 1.1)} . T ) by making more and more restrictive assumptions (D1)–(D4).
The (Ft )predictable σalgebra on R+ × Ω is generated by all leftcontinuous (Ft )adapted processes. Let Ft∗ := {A ∈ Ft  ∃B ∈ Gt with property (13. Td is not a stopping time for (Gt ).
see e. and thus Γ(t).g. This proves the lemma. we conclude that Ft ⊂ Ft∗ . Let A ∈ Ft . for every A ∈ Ht the intersection A ∩ {Td > t} is either ∅ or {Td > t}. Indeed. (D2) The default probability by t as seen by a Gt informed observer satisﬁes 0 < P [Td ≤ t  Gt ] < 1. 1{Td >t} Y P [Td > t  Gt ] dP = =
B
A
B
1{Td >t} Y P [Td > t  Gt ] dP E 1{Td >t} Y  Gt P [Td > t  Gt ] dP 1{Td >t} E 1{Td >t} Y  Gt dP 1{Td >t} E 1{Td >t} Y  Gt dP.3. Theorem I.1). A consequence of the next lemma is that for any Ft measurable random ˜ ˜ variable Y there exists an Gt measurable random variable Y such that Y = Y on {Td > t}.1 there exists a B ∈ Gt with (13. Moreover Ht ⊂ Ft∗ . (13.152
CHAPTER 13. Td ]. Hence X(t).2.2) Proof. admits a rightcontinuous modiﬁcation. Simply take B = A. Since Ft∗ is a σalgebra (→ exercise) and Ft is deﬁned to be the smallest σalgebra containing Gt and Ht .2. By Lemma 13. Hence we can deﬁne the positive (Gt )adapted hazard process Γ by e−Γ(t) := P [Td > t  Gt ] . DEFAULT RISK
Clearly Gt ⊂ Ft∗ . so we can take for B either ∅ or Ω.4) the rather surprising fact that if Γ is regular enough then it coincides with Λ on [0. Hence. Let t ∈ R+ and Y a random variable. Then E 1{Td >t} Y  Ft = 1{Td >t} eΓ(t) E 1{Td >t} Y  Gt .2. by the very deﬁnition of the Gt conditional expectation. Lemma 13. Notice that X(t) := P [Td > t  Gt ] is a (Gt )supermartingale and E[X(t)] is rightcontinuous in t (→ exercise).2. [14. We show below (Lemma 13.13].
=
B
=
A
. and so 1A 1{Td >t} = 1B 1{Td >t} .
= 1{Td >t} eΓ(t) E 1{Td >T } eΓ(T )  Gt = L(t). the processes L(t) := 1{Td >t} eΓ(t) = (1 − H(t))eΓ(t) is an (Ft )martingale. Let t ≤ T . Then 1{Td >T } = 1{Td >t} 1{Td >T } . For any t ≤ T we have P [t < Td ≤ T  Ft ] = 1{Td >t} E 1 − eΓ(t)−Γ(T )  Gt .4) follows since 1{t<Td ≤T } = 1{Td >t} − 1{Td >T } . Moreover.2. For the second statement it is enough to consider E [L(T )  Ft ] = E 1{Td >t} 1{Td >T } eΓ(T )  Ft since by deﬁnition of Γ E 1{Td >T } eΓ(T )  Gt = E E 1{Td >T }  GT eΓ(T )  Gt = 1. Lemma 13.2.3. (13.4)
= 1{Td >t} eΓ(t) E E 1{Td >T }  GT  Gt = 1{Td >t} eΓ(t) E e−Γ(T )  Gt . Using this and (13.2) we derive P [Td > T  Ft ] = E 1{Td >t} 1{Td >T }  Ft = 1{Td >t} eΓ(t) E 1{Td >T }  Gt P [Td > T  Ft ] = 1{Td >t} E eΓ(t)−Γ(T )  Gt .3) (13.13.
153
As a consequence of the preceding lemmas we may now formulate the following results.
which proves (13. which proves the lemma.
. Equation (13. Proof.3). which contain an expression for the aforementioned default probabilities. INTENSITY BASED METHOD This implies E 1{Td >t} Y P [Td > t  Gt ]  Ft = 1{Td >t} E 1{Td >t} Y  Gt .
Here is the announced result for Γ. measurable.
Proof.4.154
CHAPTER 13. In view of (13. Let t ≤ T .2. Hence.
. we have
t
Λ(t ∧ Td ) =
0
λ(s)1{Td >s} ds = Γ(t ∧ Td ). DEFAULT RISK
(D3) There exists a positive.4) gives λ(t). by the uniqueness of the predictable Doob– Meyer decomposition. (Gt )adapted process λ such that
t
Γ(t) =
0
λ(s) ds.
=:I
I=
t
1{Td >t} e
t
t 0
λ(u) du
E λ(s)E 1{Td >s}  Gs  Gt ds
s t
T
= 1{Td >t} E
λ(s)e−
T t
λ(u) du
ds  Gt
= 1{Td >t} E 1 − e−
λ(u) du
 Gt . Lemma 13. Hence we refer to λ(t) as default intensity.3) we have
t
E [N (T )  Ft ] = 1 − E 1{Td >T }  Ft −
T
0
λ(s)1{Td >s} ds
−
t
E λ(s)1{Td >s}  Ft ds
T t
t λ(u) du
= 1 − 1{Td >t} E e−
T
 Gt −
0
λ(s)1{Td >s} ds
− We have further
T
t
1{Td >t} e
t 0
λ(u) du
E λ(s)1{Td >s}  Gt ds . The process
t
N (t) := H(t) −
0
λ(s)1{Td >s} ds
is an (Ft )martingale.
Taking (formally) the righthand T derivative at T = t in (13.
P [Td > t  G∞ ] = e−Γ(t) . (D2) and (D4).2. Lemma 13. Chapter 6]. Moreover. Suppose Γ is continuous. Then φ := Γ(Td ) is an exponential random variable with parameter 1 and independent of G∞ . This proves that φ = Γ(Td ) is an exponential random variable with parameter 1 and independent of G∞ . We can deﬁne its right inverse C(s) := inf{t  Γ(t) > s}.13.5. By assumption.
The next and last assumption leads the way to implement a default risk model. (D4) P [Td > t  G∞ ] = P [Td > t  Gt ] ∀t ∈ R+ . Moreover. For more details we refer to [1.2. INTENSITY BASED METHOD hence E [N (T )  Ft ] = 1 − 1{Td >t} −
0 t
155
λ(s)1{Td >s} ds = N (t). so P [Γ(Td ) > s  G∞ ] = P [Td > C(s)  G∞ ] = e−Γ(C(s)) = e−s . Td = inf{t  Γ(t) ≥ Γ(Td )} = inf{t  Γ(t) ≥ φ}.
It can be shown that (D4) is equivalent to the hypothesis (H) Every square integrable (Gt )martingale is an (Ft )martingale. Proof.
. Hence Γ(t) is nondecreasing and continuous. Then Γ(t) > s ⇔ t > C(s) and Γ(C(s)) = s. For the next lemma we only assume (D1). Td = inf {t  Γ(t) ≥ φ} .
measurable. we have for t ≤ T
T
P [Td > T  Gt ] = P φ >
0
λ(u) du  Gt
T 0
=E P φ> = E e−
T 0
λ(u) du  GT  Gt  Gt .
We ﬁnally deﬁne Ft := Gt ∨ Ht .s. DEFAULT RISK
13.1
Construction of Intensity Based Models
The construction of a model that satisﬁes (D1)–(D4) is straightforward. ∞]. Let λ(t) be a positive. We start with a ﬁltration (Gt ) satisfying the usual conditions and G∞ = σ(Gt  t ∈ R+ ) ⊂ F . and deﬁne the random time
t
Td := inf t 
0
λ(s) ds ≥ φ
with values in (0.
As for (D4) we notice that
t
P [Td > t  G∞ ] = P φ >
0
λ(u) du  G∞ = e−
t 0
λ(u) du
= P [Td > t  Gt ] .
λ(u) du
by the independence of φ and GT (this is a basic lemma for conditional expectations).
We then ﬁx an exponential random variable φ with parameter 1 and independent of G∞ .2. for all t ∈ R+ . where Ht = σ(H(s)  s ≤ t). (Gt )adapted process with the property
t 0
λ(s) ds < ∞ a.156
CHAPTER 13. Consequently. And it is an easy exercise to show that 0 < P [Td > t  Gt ] = e−
t 0
λ(u) du
Td
< 1 and φ =
0
λ(u) du.
. Conditions (D1)–(D3) are obviously satisﬁed for
t
Λ(t) = Γ(t) :=
0
λ(s) ds.
For the intensity process (13.2. T ) = P [Td ≤ T  Ft ] = where 2γe(γ−β)u/2 2b A(u) := − 2 log σ (γ − β) (eγu − 1) + 2γ 2 (eγu − 1) B(u) := . λ(0) ≥ 0.3) looks just like what we had for the riskneutral valuation of zerocoupon bonds in terms of a given short rate process (Chapter 7). these conditions are not preserved under an
. if Td > t else. (Gt )adapted short rate process r(t). respectively (unfortunately.3) we need a tractable model for the intensity process λ.5)
The proof of the following lemma is left as an exercise. (13. Notice that λ ≥ 0 is essential. But the righthand side of (13. (γ − β) (eγu − 1) + 2γ γ := β 2 + 2σ 2 . Moreover. (Gt )adapted process λQ such that
t
ΓQ (t) :=
0
λQ (s) ds < ∞ a.2. for all t ∈ R+ .
13. .13. we assume that there exists a positive. β ∈ R and σ > 0 some constants. ΛQ := ΓQ and ΓQ replacing P.5) the conditional default probability is pd (t. Λ and Γ. 1 − e−A(T −t)−B(T −t)λ(t) . INTENSITY BASED METHOD
157
13. In addition. and let dλ(t) = (b + βλ(t)) dt + σ λ(t) dW (t). An obvious and popular choice for λ is thus a square root (or aﬃne) process.
and (D1)–(D3) are satisﬁed for Q.6.s. 0.2.2
Computation of Default Probabilities
When it comes to computations of the default probabilities (13. we suppose now that we are given a riskneutral probability measure Q ∼ P and a measurable. b ≥ 0. Lemma 13.2.3
Pricing Default Risk
The stochastic setup is as above. So let W be a (Gt )Brownian motion. measurable.
T ) of a corporate zerocoupon bond with maturity T .8)
. We will determine the price C(t. δ at Td if Td ≤ T . So that Lemmas 13.4 apply. β ∈ R. DEFAULT RISK
equivalent change of measure in general). Gt )Brownian motion. As for the recovery we ﬁx a constant recovery rate δ ∈ (0. b ≥ 0. A tractable (hence aﬃne) model is easily found.7)
For the intensity process we chose the aﬃne combination λQ (t) = c0 + c1 r(t). • Partial recovery at default: the cashﬂow is ZeroRecovery The arbitrage price of C(t. (13.1–13.2. σ > 0 constant parameters and dr(t) = (b + βr(t)) dt + σ r(t) dW (t). T ) = 1{Td >t} e = 1{Td >t} e
t 0 t 0 T t
1 at T if Td > T .2 this is C(t.
(13. T ) = EQ e− In view of Lemma 13.6)
λQ (s) ds λQ (s) ds
EQ e − EQ e −
T t T t
r(s) ds r(s) ds
= 1{Td >t} EQ e−
T t (r(s)+λQ (s))ds
 Gt .2. 1{Td >T }  Gt EQ 1{Td >T }  GT  Gt (13. For the short rates we chose CIR: let W be a (Q.2. c1 ≥ 0. Pricing a corporate bond boils down to the pricing of a nondefaultable zerocoupon bond with the short rate process replaced by r(s) + λQ (s) ≥ r(s). • Partial recovery at maturity: the cashﬂow at T is 1{Td >T } + δ1{Td ≤T } . for two constants c0 . which may default.
r(s) ds
1{Td >T }  Ft . r(0) ≥ 0.158
CHAPTER 13.
Notice that this is a very nice formula. 1) and distinguish three cases: • Zero recovery: the cashﬂow at T is 1{Td >T } . T ) is C(t.
T ) is the CIR price of a defaultfree zerocoupon bond.2 shows that EQ 1{Td >t} Y  G∞ ∨ Ht = 1{Td >t} e
t 0
λQ (s) ds
EQ 1{Td >t} Y  G∞
. Partial Recovery at Default A straightforward modiﬁcation of the proofs of Lemmas 13. T ).
 Gt + δP (t. Proof. T ) stands for the price of the defaultfree zerocoupon bond.13.7. In view of (13. where 2γe(γ−β)u/2 2b(1 + c1 ) log A(u) := c0 u − σ2 (γ − β) (eγu − 1) + 2γ 2 (eγu − 1) (1 + c1 ). T ) = 1{Td >t} e−A(T −t)−B(T −t)r(t) . Exercise. T ).2.6) hence C(t. For the above aﬃne model we have C(t.2. B(u) := (γ − β) (eγu − 1) + 2γ γ := β 2 + 2(1 + c1 )σ 2 .
where P (t.2.2. T ) = 1{Td >t} e−c0 (T −t) P (t. Here we have C(t. A special case is c1 = 0 (constant intensity). T ) = (1 − δ) 1{Td >t} EQ e−
T t (r(s)+λQ (s))ds
159
. where P (t. INTENSITY BASED METHOD Lemma 13.1 and 13. Partial Recovery at Maturity This is an easy modiﬁcation of the preceding case since 1{Td >T } + δ1{Td ≤T } = (1 − δ) 1{Td >T } + δ.
where C0 (t. Combining this with Section 13.2.160
CHAPTER 13. Again.2.8) this expression can be made more explicit (→ exercise).
13.4
Measure Change
We consider an equivalent change of measure and derive the behavior of the compensator process for the stopping time Td . T ) = C0 (t. As a result. Diﬀerentiation in with respect to u yields its density function 1{Td >t} λQ (u)e−
u t
λQ (s) ds
1{t≤u} .
which is the regular conditional distribution of Td conditional on {Td > t} and G∞ ∨ Ht . T ) is the bond price with zero recovery.7)–(13. we take the above
. The above calculations and an extension to stochastic recovery go back to Lando [16].
For the above aﬃne model (13. DEFAULT RISK
for every random variable Y .1 we obtain for t ≤ u Q[t < Td ≤ u  G∞ ∨ Ht ] = 1{Td >t} e = 1{Td >t} e
t 0 t 0
λQ (s) ds λQ (s) ds
EQ 1{t<Td ≤u}  G∞ e−
t 0
λQ (s) ds
− e−
u 0
λQ (s) ds
= 1{Td >t} 1 − e−
u t
λQ (s) ds
. T ) + π(t). the price of the corporate bond bond price with recovery at default is C(t.
Hence the arbitrage price of the recovery at default given that t < Td ≤ T is given by π(t) = EQ e−
Td t
r(s) ds
Td t
δ1{t<Td ≤T }  Ft δ1{t<Td ≤T }  G∞ ∨ Ht  Ft
u t
= E Q EQ e − = δ1{Td >t} EQ
T
r(s) ds T
e−
t
r(s) ds
λQ (u)e−
u t
λQ (s) ds
du  Ft
= δ1{Td >t}
t
EQ λQ (u)e−
u (r(s)+λQ (s)) ds t
 Ft du.
∆f (s) := f (s) − f (s−).
Lemma 13. This does not. Section 7. see also [1. imply that ΛQ (t) is the (Q. which in a sense is simpler than for Brownian motion since it is a pathwise calculus.
We will construct an equivalent probability measure Q ∼ P such that ΛQ (t ∧ Td ) is the (Q.13. Gt )hazard process ΓQ (t) = − log Q[Td > t  Gt ] of Td in general. however. The following analysis involves stochastic calculus for cadlag processes of ﬁnite variation (FV). So that
t
161
M (t) = H(t) −
0
λ(s)1{Td >s} ds
is a (P. A counterexample has been constructed by Kusuoka [15]. Ft )compensator of H. Now let µ be a positive (Gt )predictable process such that
t
ΛQ (t) :=
0
µ(s)λ(s) ds < ∞ a.
where [f. for all t ∈ R+ . t ≥ Td
V (t) := 1{Td >t} + µ(Td )1{Td ≤t} =
. t < Td µ(Td ).s.3]. INTENSITY BASED METHOD stochastic setup and let (D1)–(D3) hold.2. The process D(t) := C(t)V (t) with
t
C(t) := exp
0
(1 − µ(s))λ(s)1{Td >s} ds 1.8. g](t).2. We recall the integration by parts formula for two rightcontinuous FV functions f and g
t t
f (t)g(t) = f (0)g(0) +
0
f (s−) dg(s) +
0
g(s−) df (s) + [f. Ft )martingale. g](t) =
0<s≤t
∆f (s)∆g(s).
9) therefore the unique Doob–Meyer decomposition of H under Q. Proof. so is MQ ([14. Suppose E[D(T )] = 1 (hence (D(t))t∈[0. T ].T ] is a martingale).9. It is enough to show that MQ is a Qlocal martingale. Let T ∈ R+ . dP Then the process MQ (t) := H(t) − ΛQ (t ∧ Td ).2. (13. Theorem 1. Since H is uniformly integrable. so that we can deﬁne an equivalent probability measure Q ∼ P on FT by dQ = D(T ).
Since D(s−) is locally bounded and ΛQ (t) < ∞ we conclude that D is a Plocal martingale.9)
. is a Qmartingale. (13.10]). V ] = 0 and
t t
V (t) = 1 +
0
(µ(s) − 1) dH(s) = 1 +
0
V (s−) (µ(s) − 1) dH(s). ΛQ is an increasing continuous (and hence predictable) process. Proof.4.
Hence
t t
D(t) = 1 +
0 t
C(s−) dV (s) +
0
V (s−) dC(s)
=1+
0 t
C(s−)V (s−) (µ(s) − 1) dH(s)
+
0
C(s)V (s−)(1 − µ(s))λ(s)1{Td >s} ds
t 0
=1+
D(s−) (µ(s) − 1) dM (s).162 satisﬁes D(t) = 1 +
0
CHAPTER 13. DEFAULT RISK
t
D(s−) (µ(s) − 1) dM (s)
and is thus a positive Plocal martingale. Notice that [C. t ∈ [0. Indeed. Lemma 13.
. INTENSITY BASED METHOD
163
From Bayes’ rule we know that MQ is a Qlocal martingale if and only if DMQ is a Plocal martingale.3 do not apply for the above situation. so that the methods from Section 13.3].
Integration by parts gives
t t
DMQ (t) = =
0 t 0
D(s−) dMQ (s) + D(s−) dH(s) −
t 0 t
0 t 0
MQ (s−) dD(s) + [D. MQ ](t) = ∆D(Td )1{Td ≥t} = D(Td −) (µ(Td ) − 1) 1{Td ≥t}
t
=
0
D(s−) (µ(s) − 1) dH(s). MQ ](t)
D(s−)µ(s)λ(s)1{Td >s} ds
t 0 t
+ =
0
MQ (s−) dD(s) +
D(s−) (µ(s) − 1) dH(s)
MQ (s−) dD(s) +
D(s−)µ(s) dM (s). The detailed analysis can be found in [1. there is a way to derive the pricing formulas from Section 13.
0
which proves the claim. Section 8.2.2.13. Pricing by the “Martingale Approach” We remark again that ΛQ is diﬀerent from ΓQ in general. Notice that [D.3 under Assumption (D4) for Q. Yet.2.
DEFAULT RISK
.164
CHAPTER 13.
G¸tarek. Filipovi´. Berlin. Musiela. 211–239. C. no. The market model of interest rate dynamics. and A. Runggaldier. Jarrow. and S. March 1999. Options. R. no. R. PrenticeHall International. Basle. valuation and hedging. [2] BIS. J. [8] D. A theory of the term structure of interest rates. [9] D. and M. Hull. [6] D. Credit risk: modelling. Econometrica 60 (1992). 127–155. o [4] T. and W. Brigo and F. Bielecki and M. Brace. Zerocoupon yield curves: Technical documentation. 2000. Interest rate models—theory and practice. o 1998. Mercurio. Ingersoll. Bj¨rk. Bj¨rk. Rutkowski. Math. 165
. Arbitrage theory in continuous time. Berlin. Heath. Consistency problems for Heath–Jarrow–Morton interest c rate models. 2002. D. fourth ed.. Oxford University Press. Springer Finance. Kabanov. Math. Cox. Finance 7 (1997). [3] T. 2. Y. Bank for International Settlements. 2001. [10] J. Econometrica 53 (1985). 77–105. Bond pricing and the term structure of interest rates: A new methodology for contingent claims valuation. SpringerVerlag. Bond market structure in the presence of marked point processes. and other derivatives. futures.. 385–408. Morton. Inc. a [5] A. Berlin. [7] J. Ross. 2. SpringerVerlag. Finance 7 (1997).Bibliography
[1] T. SpringerVerlag. 2001.
[20] C. Continuous martingales and Brownian motion. [19] M. NorthHolland. 2 (1998). On the pricing for corporate debt: the risk structure of interest rates. Lando. Webber. 1994. Shreve. 290–330. J. Musiela and M. 36. Tokyo.. Amsterdam. 1998. Parsimonious modeling of yield curves. Grundlehren der mathematischen Wissenschaften. Lecture Notes in Economics and Mathematical Systems. 99–120.A. Vol. Nelson and A. On Cox processes and creditrisky securities. McGrawHill. Berlin. Common factors aﬀecting bond returns. SpringerVerlag. second ed. vol. [12] R. Brownian motion and stochastic calculus. second ed. Pricing credit linked ﬁnancial instruments. Interestrate option models. 54–61. A remark on default risk models. Jarrow. 2002. SpringerVerlag. 69–82. Rev. of Business 60 (1987). E. BerlinHeidelberg. 473–489. Wiley. vol..166
BIBLIOGRAPHY
[11] J. [23] D. Principal component and factor analyses. Revuz and M. Jamshidian. Interest rate modelling. Finance and Stochastics 1 (1997). 516. [14] I. Finance 29 (1974). SpringerVerlag. R. [13] R. J. Scheinkman. [21] C.
. C. Libor and swap market models and measures. [15] S. 293. Advances in mathematical economics. [22] R. Journal of Fixed Income 1 (1991). Derivatives Res. James and N. 1996. Statistical methods in ﬁnance. 449–470. [24] Bernd Schmid. 1996. BerlinHeidelbergNew York. 1999. Yor. Kusuoka. [16] D. Merton. Applications of Mathematics. 1997. Wiley. Springer. SpringerVerlag. [17] R. 2000. Karatzas and S. Siegel. Rebonato. New York. pp. 1991. pp. Rutkowski. Modelling ﬁxed income securities and interest rate options. Martingale methods in ﬁnancial modelling. Rao. Litterman and J. Theory and empirical evidence. 1. 489–505. [18] R. vol.
SpringerVerlag. Svensson.BIBLIOGRAPHY
167
[25] J. 2001. Michael Steele. Stochastic calculus and ﬁnancial applications. September 1994. Estimating and interpreting forward interest rates: Sweden 19921994. [26] L. Interestrate management. O. 114. E. New York. 45.
. Applications of Mathematics. SpringerVerlag. [27] R. Zagst. IMF Working Paper No. vol. 2002. Berlin.