Term Structure Models: IEOR E4710 Spring 2010

c 2010 by Martin Haugh

°

Market Models
One of the principal disadvantages of short rate models, and HJM models more generally, is that they focus on
unobservable instantaneous interest rates. The so-called market models that were developed1 in the late 90’s
overcome this problem by directly modeling observable market rates such as LIBOR2 and swap rates. These
models are straightforward to calibrate and have quickly gained widespread acceptance from practitioners. The
first market models were actually developed in the HJM framework where the dynamics of instantaneous forward
rates are used via Itô’s Lemma to determine the dynamics of zero-coupon bonds. The dynamics of zero coupon
bond prices were then used, again via Itô’s Lemma, to determine the dynamics of LIBOR. Market models are
therefore not inconsistent with HJM models. In these lecture notes, however, we will prefer to specify the
market models directly rather than derive them in the HJM framework. In the process, we will derive Black’s
formulae for caplets and swaptions thereby demonstrating the consistency of these formulae with martingale
pricing theory.
Throughout these notes, we will ignore the possibility of default or counter-party risk and treat LIBOR interest
rates as the fundamental rates in the market. Zero-coupon bond prices are then computed using LIBOR rather
than the default-free rates implied by the prices of government securities. This does result in a minor
inconsistency in that we price derivative securities assuming no possibility of default yet the interest rates
themselves that play the role of “underlying security”, i.e. LIBOR and swap rates, implicitly incorporate the
possibility of default. This inconsistency actually occurs in practice when banks trade caps, swaps and other
instruments with each other, and ignore the possibility of default when quoting prices. Instead, the associated
credit risks are kept to a minimum through the use of netting agreements and by counter-parties limiting the
total size of trades they conduct with one another. This approach can also be justified when counter-parties
have a similar credit rating and similar exposures to one another. Finally, we should mention that it is indeed
possible3 , and sometimes necessary, to explicitly model credit risk even when we are pricing ‘standard’ securities
such as caps and swaps. It goes without saying of course, that default risk needs to be modeled explicitly when
pricing credit derivatives and related securities.

1 LIBOR, Swap Rates and Black’s Formulae for Caps and

Swaptions
We now describe two particularly important market interest rates, namely LIBOR and swap rates. We first
define LIBOR and forward LIBOR, and then describe Black’s formula for caplets. After defining LIBOR we then
proceed to discuss swap rates and forward swap rates as well as describing Black’s formula for swaptions. In
practice, the “underlying security” for caps and swaptions are LIBOR and LIBOR-based swap rates. Therefore
by modeling the dynamics of these rates directly we succeed in obtaining more realistic models than those
developed in the short-rate or HJM framework.
1 See Miltersen, Sandmann and Sondermann (1997), Brace, Gatarek and Musiela (1997), Jamshidian (1997) and Musiela and

Rutkowski (1997).
2 These models apply equally well to Euribor rates.
3 See chapter 11 of Cairns for a model where swaps are priced taking the possibility of default explicitly into account.
Market Models 2

LIBOR
The forward rate at time t based on simple interest for lending in the interval [T1 , T2 ] is given by4
à !
1 ZtT1 − ZtT2
F (t, T1 , T2 ) = (1)
T2 − T1 ZtT2

where, as before, ZtT is the time t price of a zero-coupon bond maturing at time T . Note also that if we
measure time in years, then (1) is consistent with F (t, T1 , T2 ) being quoted as an annual rate.
LIBOR rates are quoted as simply-compounded interest rates, and are quoted on an annual basis. The
accrual period or tenor, T2 − T1 , is usually fixed at δ = 1/4 or δ = 1/2 corresponding to 3 months and 6 months,
respectively. With a fixed value of δ in mind we can define the δ-year forward rate at time t with maturity T as
à !
1 ZtT − ZtT +δ
L(t, T ) := F (t, T, T + δ) = . (2)
δ ZtT +δ

Note that the δ-year spot LIBOR rate at time t is then given by L(t, t).

Remark 1 LIBOR or the London Inter-Bank Offered Rate, is determined on a daily basis when the British
Bankers’ Association (BBA) polls a pre-defined list of banks with strong credit ratings for their interest rates.
The highest and lowest responses are dropped and then the average of the remainder is taken to be the LIBOR
rate. Because there is some credit risk associated with these banks, LIBOR will be higher than the
corresponding rates on government treasuries. However, because the banks that are polled have strong credit
ratings the spread between LIBOR and treasury rates is generally not very large and is often less than 100 basis
points. Moreover, the pre-defined list of banks is regularly updated so that banks whose credit ratings have
deteriorated are replaced on the list with banks with superior credit ratings. This has the practical impact of
ensuring that forward LIBOR rates will still only have a very modest degree of credit risk associated with them.

Black’s Formula for Caplets

Consider now a caplet with payoff δ(L(T, T ) − K)+ at time T + δ. The time t price, Ct , is given by
· +
¸
Q δ(L(T, T ) − K)
Ct = Bt Et
BT +δ
P £ ¤
= δZtT +δ Et T +δ (L(T, T ) − K)+ .

where (Bt , Q) is an arbitrary numeraire-EMM pair and (ZtT +δ , PT +δ ) is the forward measure-numeraire pair.
The market convention is to quote caplet prices using Black’s formula which equates Ct to a Black-Scholes like
formula so that
· µ ¶ µ ¶¸
T +δ log(L(t, T )/K) + σ 2 (T − t)/2 log(L(t, T )/K) − σ 2 (T − t)/2
Ct = δZt L(t, T )Φ √ − KΦ √ (3)
σ T −t σ T −t

where Φ(·) is the CDF of a standard normal random variable. Note that (3) is what you would get for Ct if you
assumed that
dL(t, T ) = σL(t, T ) dW T +δ (t)
where W T +δ (t) is a PT +δ -Brownian motion and σ is an ‘implied’ volatility that is used to quote prices.
Black’s formula for caps is to equate the cap price with the sum of caplet prices given by (3) but where a
common σ is assumed. Similar formulae exist for floorlets and floors.
4 This follows from a simple arbitrage argument. Prove it!
Market Models 3

Swap Rates
Consider a payer forward start swap where the swap begins at some fixed time Tn in the future and expires at
time TM ≥ Tn . We assume the accrual period is of length δ. Since payments are made in arrears, the first
payment occurs at Tn+1 = Tn + δ and the final payment at TM +1 = TM + δ. Then martingale pricing implies
that the time t < Tn value, SWt , of this forward start swap is
 
XM
B t
SWt = EQ t
δ (L(Tj , Tj ) − R)
j=n
B Tj+1

where R is the fixed (annualized) rate specified in the contract. A standard argument using the properties of
floating-rate bond prices implies that
M
X +1
T T
SWTn = 1 − ZTnM +1 − Rδ ZTnj . (4)
j=n+1

Exercise 1 Prove (4).

Equation (4) in turn easily implies (why?) that for t < Tn we have
M
X +1
T T
SWt = ZtTn − Zt M +1 − Rδ Zt j .
j=n+1

Definition 1 The forward swap rate at time t is the value R = R(t, Tn , TM ) for which SWt = 0. In particular,
we obtain
T
ZtTn − Zt M +1
R = R(t, Tn , TM ) = P M +1 T
. (5)
δ j=n+1 Zt j
The spot swap rate is then obtained by taking t = Tn in (5).

Now consider the time t price5 of a payer-swaption that expires at time Tn > t and with payments of the
b (annualized) and a notional
underlying swap taking place at times Tn+1 , . . . , TM +1 . Assuming a fixed rate of R
principle of $1, the value of the option at expiration is given by the positive part of (4). It satisfies
 +
M
X +1
T b T
CTn = 1 − ZTnM +1 − Rδ ZTnj  . (6)
j=n+1

T
Using (5) at t = Tn we can substitute for 1 − ZTnM +1 in (6) and find that
 +
h i M
X +1
b
δ R(Tn , Tn , TM ) − R T
CTn = ZT n 
j

j=n+1
 
M
X +1 h i+
δ T b .
= ZTnj  R(Tn , Tn , TM ) − R (7)
j=n+1

5 Note that in (6) we have implicitly assumed that the strike is k = 0.

Market Models 4

Therefore we see that the swaption is like a call option on the swap rate. The time t value of the swaption, Ct ,
is then given by the Q-expectation of the right-hand-side of (7), suitably deflated by the numeraire.

Black’s Formula for Swaptions

Market convention, however, is to quote swaption prices via Black’s formula which equates Ct to a
Black-Scholes-like formula so that
 " Ã !
M
X +1 b 2
T log(R(t, T n , T M )/ R) + σ (Tn − t)/2
Ct = δ Zt j  R(t, Tn , TM )Φ √ −
j=n+1
σ Tn − t
à !#
b − σ 2 (Tn − t)/2
log(R(t, Tn , TM )/R)
b
RΦ √ (8)
σ Tn − t

where again σ is an ‘implied’ volatility that is used to quote prices. Note that the expression in (8) is what we
would obtain for the expectation of
 
M
X +1 h i+
δ T b
Zt j  R(Tn , Tn , TM ) − R
j=n+1

if dR(t, Tn , TM ) = σR(t, Tn , TM ) dWt .

It should be stated that Black’s formulae for caps and swaptions did not originally correspond to prices that arise
from the application of martingale pricing theory to some particular model. As originally conceived, they merely
provided a framework for quoting market prices. The market models of these lecture notes will provide a belated
justification for these formulae. We shall see that the justifications are mutually inconsistent, however, in that it
is impossible for both formulae to hold simultaneously within the one model.

2 The Term Structure of Volatility

The term structure of volatility6 is a graph of volatility plotted against time to maturity, τ . There are of course
many definitions of volatility and care is needed in specifying which definition is intended. Some commonly used
definitions of the term structure of volatility at time t include:
1. The volatility of spot rates Ytt+τ as a function of τ . Depending on the model under consideration, this
volatility may be available in closed form and the model calibrated to historical or implied rates.
2. The volatility, σ(t, t + τ ), of instantaneous forward rates, f (t, t + τ ).
3. The implied volatility, σ, given by Black’s formula for caplets. This will vary with time to maturity and
can be computed at any time from caplet prices in the market. The implied volatility also varies with the
strike of the caplet, i.e. there is a volatility skew for each maturity. The term structure of caplet
volatilities if therefore strike dependent.
4. The implied volatility, σ, given by Black’s formula for caps. Again this will vary with time to maturity and
strike. It can can be computed at any time from market prices for caps.
When calibrating term structure models it is common to calibrate using both market prices or rates, and the
term structure of volatility. As a result we often want to work with models that allow for a rich variety of term
structures of volatility as well of course, as a rich variety of term structures of interest rates.

6 ‘Quants’ in the fixed-income industry commonly refer to the ‘term-structure of volatility’ when discussing fixed-income

derivatives and models. In this section we briefly give some possible definitions of the ‘term-structure of volatility’ but we will
not need these definitions elsewhere in the course.
Market Models 5

3 Numeraires and Zero-Coupon Bond Prices

³ R ´
t
While the cash account with Bt := exp − 0 rs ds has been the default numeraire to date, we will not work
with the cash account as our numeraire in the context of market models. The reason is clear: in market models
we take LIBOR rates (or swap rates) with a fixed tenor, δ, in mind, as our fundamental interest rates. It would
therefore be very inconvenient (as well as defeating the purpose) if we had to determine the instantaneous short
rate at each point in time. As a result we will generally work with other numeraire-EMM pairs as described
below.
First, however, we will fix the maturities or tenor dates to which our market models will apply. At time t we
could in principal have LIBOR rates, L(t, T ), available for all T > t. This is unnecessary, however, as the prices
of most important securities, e.g. caps, floors, swaps, swaptions, Bermudan swaptions, etc., are determined by
the rates (LIBOR or swap) applying to only a finite set of maturities. We therefore fix in advance a set of tenor
dates7

0 := T0 < T1 < T2 < . . . < TM < TM +1 with

δi := Ti+1 − Ti , i = 0, 1, . . . , M.

While the δi ’s are usually nominally equal, e.g. 1/4 or 1/2, day-count conventions will results in slightly different
values for each δi . We let Ztn denote the time t price of a zero-coupon bond maturing at time Tn > t for
n = 1, . . . , M . Similarly, we use Ln (t) to denote the time t forward rate applying to the period [Tn , Tn+1 ] for
n = 0, 1, . . . , M . In particular, (2) then states

Ztn − Ztn+1
Ln (t) = , for 0 ≤ t ≤ Tn , n = 0, 1, . . . , M. (9)
δn Ztn+1

With some work we can invert (9) to obtain an expression for bond prices in terms of LIBOR rates. We find
n−1
Y 1
ZTni = for n = i + 1, . . . , M + 1. (10)
j=i
1 + δj Lj (Ti )

Equation (10) only determines the bonds prices at the fixed maturity dates. However, for an arbitrary date t we
can easily check that
n−1
Y
φ(t) 1
Ztn = Zt for 0 ≤ t ≤ Tn . (11)
1 + δj Lj (t)
j=φ(t)

where we define φ(t) to be next tenor date after time t. That is,

φ(t) := min {i : t < Ti }.

i=1,...,M +1

φ(t)
Remark 2 The presence of Zt in (11) suggests that it may not be sufficient to model only the dynamics of
the forward LIBOR rates, Ln (t), when we specify a market model since they are not sufficient to determine
φ(t)
Zt at an arbitrary time t. However, as we shall see below, this will not prove to be a problem as the φ(t)
factor vanishes upon deflating by the numeraire.

Exercise 2 Prove equations (10) and (11).

7 The notation and setup in this section and the next will borrow heavily from Section 3.7 in Monte Carlo Methods in

Financial Engineering by Glasserman.

Market Models 6

Numeraire-EMM Pairs
The following numeraire-EMM pairs are commonly used in market models:
1. The spot measure, Q, assumes that Bt∗ is the numeraire where Bt∗ is defined as follows.
• start with $1 at t = 0 and then purchase 1/Z01 of the zero-coupon bonds maturing at time T1
• at time T1 reinvest the funds in the zero-coupon bond maturing at time T2
• by continuing in this way, we see that at time t the spot numeraire will be worth
φ(t)−1
φ(t)
Y
Bt∗ = Zt [1 + δj Lj (Tj )]. (12)
j=0

Note the similarity between this numeraire and our usual cash account.
2. The forward measure, P T , takes the zero-coupon bond maturing at time T as numeraire. We have seen
this numeraire-EMM pair already.
3. The swap
PM measure, P X , is useful for pricing swaptions analytically. It takes the numeraire to be
Xt = δ k=1 Ztk , which is indeed a positive security price process.

Deflating Zero-Coupon Bond Prices by the Spot Numeraire

Equations (11) and (12) show that deflated8 zero-coupon bond prices, Dn (t), satisfy
 
φ(t)−1 n−1
Y 1 Y 1
Dn (t) =   for 0 ≤ t ≤ Tn . (13)
j=0
1 + δj Lj (Tj ) 1 + δj Lj (t)
j=φ(t)

φ(t)
In particular, we see that the factor, Zt , has vanished.

4 The LIBOR Market Model

Dynamics under the Spot Measure
We assume that the dynamics of the LIBOR rates satisfy

dLn (t) = µn (t)Ln (t) dt + Ln (t)σn (t)T dW (t), 0 ≤ t ≤ Tn , n = 1, . . . , M (14)

where W (t) is a d-dimensional Brownian motion, and µn (t) and σn (t) are adapted processes that may depend
on the current vector of interest rates L(t) := (L1 (t), . . . , LM (t)). The assumption of no arbitrage and the
positivity of deflated bond prices implies the existence of an Rd -valued process νn (t) such that

dDn (t) = Dn (t)νnT (t) dW (t). (15)

We could apply Itô’s Lemma directly to our expression for Dn (t) in (13) but this would be awkward. Instead we
will apply Itô’s Lemma to Yn (t) := log Dn (t). We see from (15) that
1
dYn (t) = − ||νn (t)||2 dt + νnT (t) dW (t) (16)
2
8 We will take the spot numeraire to be the default numeraire.
Market Models 7

We can also find an alternative expression for dYn (t) using (13). In particular, noting that the first factor in
(13) is constant between maturities, we obtain via Itô’s Lemma
n−1
X
dYn (t) = − d log (1 + δj Lj (t))
j=φ(t)
à !  
n−1
X n−1
X
δj µj (t)Lj (t) δj2 Lj (t)2 σjT (t)σj (t) δj Lj (t)σjT (t)
= − − 2 dt −   dW (t).(17)
1 + δj Lj (t) 2 (1 + δj Lj (t)) 1 + δj Lj (t)
j=φ(t) j=φ(t)

Comparing the volatility terms in (16) and (17) then gives us

n−1
X δj Lj (t)σj (t)
νn (t) = − . (18)
1 + δj Lj (t)
j=φ(t)

We would now like to find an expression for the µj ’s. Towards this end, we could compare the drift terms in (16)
and (17), and this is easy to do when n = 2 and φ(t) = 1. After some straightforward algebra, we easily find9

µ1 (t) = −σ1T (t)ν2 (t), 0 ≤ t ≤ T1 .

More generally, we obtain

n
X δj Lj (t)σnT (t)σj (t)
µn (t) = −σnT (t) νn+1 (t) = . (19)
1 + δj Lj (t)
j=φ(t)

We could have obtained (19) by again comparing the drift terms in (16) and (17) but this appears to be very
cumbersome. Exercise 3 instead provides a more elegant approach.

Exercise 3 Use induction to establish10 that the drifts, µn (t), must satisfy (19) under the no-arbitrage
assumption. In particular, first assume µ1 , . . . , µn−1 have been chosen in a manner that is consistent with the
Q-martingale assumption on D1 , . . . , Dn . Then show that Dn+1 is a martingale if and only if Ln Dn+1 is a
martingale. Finally, apply Itô’s Lemma to Ln Dn+1 and use the martingale property to obtain (19).

We therefore obtain that the arbitrage free Q-dynamics of the forward LIBOR rates are given by
 
Xn T
δ j L j (t)σ (t)σj (t)
dLn (t) =  n  Ln (t) dt + Ln (t)σn (t)T dW (t), 0 ≤ t ≤ Tn , n = 1, . . . , M. (20)
1 + δj Lj (t)
j=φ(t)

Dynamics under the Forward Measure

Consider now the case where we use the forward measure, PM +1 , and the associated numeraire, ZtM +1 . We use
b n (t) to denote zero-coupon bonds prices that have been deflated by ZtM +1 . Equation (11) with n replaced by
D
M + 1 then implies that
YM
Db n (t) = (1 + δj Lj (t)) . (21)
j=n

We would like to find the market-price-of-risk process, η M +1 (t) ∈ Rd , that relates the Q-Brownian motion W (t)
to the the PM +1 Brownian motion, WM +1 (t), so that

dW (t) = dWM +1 (t) − η(t) dt. (22)

9 Note that L1 (t), and therefore µ1 (t), do not have any meaning for t > T1 .
10 See also Glasserman, page 170.
Market Models 8

There are a number of ways to do this but perhaps the easiest is the approach we followed with the Vasicek
model when we switched to the forward measure. Equation (21) implies D b M (t) = 1 + δM LM (t) so that
b M (t) = δM dLM (t).
dD (23)
We now substitute for dLM (t) in (23) using (20) evaluated at n = M , and then substitute for W (t) using (22).
b M (t) is a PM +1 -martingale we find that
Since D
M
X δj Lj (t)σj (t)
η(t) = .
1 + δj Lj (t)
j=φ(t)

In particular, we obtain the arbitrage-free PM +1 -dynamics of the forward LIBOR rates are given by
 
XM T
δ j L j (t)σ (t)σ j (t)
dLn (t) = −  n  Ln (t) dt + Ln (t)σn (t)T dWM +1 (t), 0 ≤ t ≤ Tn , n = 1, . . . , M.
j=n+1
1 + δ j L j (t)
(24)
Black’s Formula for Caplets
We are now in a position to derive Black’s formula (see (3)) for caplet prices. If we take n = M in (24), then
we obtain
dLM (t) = LM (t)σM (t)T dWM +1 (t) (25)
implying in particular11 that LM (t) is a PM +1 -martingale. If we assume that σM (t) is a deterministic function,
then we easily see that LM (t) is log-normally distributed. In particular, we obtain
µ Z Z t ¶
1 t 2 2
log LM (t) ∼ N log(LM (0)) − ||σM (s)|| ds , ||σM (s)|| ds . (26)
2 0 0
RT
We can now obtain (3) if we let TM = T and set σ 2 = 0 M ||σM (s)||2 ds / TM .
Note also that there is no problem when we take σM (t) to be deterministic in (25) which contrasts with the
HJM framework. This is because while the numerators in the drift of (20) are quadratic in Lj (t), the 1 + δj Lj (t)
term in the denominator ensures that there is no possibility of explosion in the SDE. This is a further advantage
of the market model framework where we model simple LIBOR rates rather than instantaneous forward rates.

Remark 3 Note that under the EMM PM +1 , the LIBOR rates Ln for n < M are not martingales. However, if
we approximate the Lj (t) term with Lj (0) in the drift component of (24), then we would have log-normal
dynamics for Ln (t), assuming that all the σj (t)’s are deterministic. For reasonably short maturities, this
approximation is sometimes used to construct approximate analytic prices for some derivative securities.

The Caplet Volatility Surface

For a given market caplet price, Cpt(TM , K) say, we can compute the implied volatility, σ(K, TM ), and plot it
as a function of the strike, K, keeping the time-to-maturity, TM , fixed. If the market model with a deterministic
σM (t) is correct then we should obtain σ(K, TM ) = σ(TM ) for all K. Until approximately the mid 1990’s this
was more or less the case but soon afterwards a noticeable skew started to appear in the market-place. This led
to the development of models where the σj (t)’s were themselves stochastic. Examples12 of such models are the
CEV13 model, shifted log-normal model and mixture of log-normals model. Other approaches to capturing the
skew allow for jumps in the market rates. The success of these models depend to the extent that they can price
caplets, floorlets and swaptions analytically and whether or not they are reasonably straightforward to calibrate.

BGM’s Approximation for Swaption Prices

In their original paper, Brace, Gatarek and Musiela (BGM) succeeded in deriving Black’s formula for caplets and
thereby demonstrated its consistency with martingale pricing. Their framework did not enable them to derive
11 Subject, as usual, to technical conditions.
12 These are all described in Chapter 6 of the 2001 edition of Brigo and Mercurio. Rebonato’s “Volatility and Correlation”
devotes several chapters to modeling the smile.
13 i.e. the constant elasticity of variance model.
Market Models 9

Black’s formula for swaptions, however. Instead they provided an analytic approximation for swaption prices that
we will not describe14 here. It is worth mentioning, however, that their approximation works well in practice and
provides swaption prices that are very close to those obtained via Monte Carlo simulation.

Calibration
A good analytic approximation to swaption prices is very valuable as it allows for the possibility of calibrating
the model to both cap and swaption prices simultaneously. Note that calibrating a market model amounts to
selecting the parameters of the σj processes (or functions if they are are assumed to be deterministic functions
of time). If swaption prices could only be computed or estimated via Monte-Carlo simulation then calibrating to
swaption prices in this model would be very time-consuming. It is worth pointing out that caplet prices depend
only on the level of volatility in the forward rates. This is clear from (26). Swaption prices, however, also depend
on the correlations between the forward rates. So one common approach to calibration is the following:

1. Use market caplets to determine the level of forward rate volatility

2. Use swaption prices or historical forward rate correlations to determine the individual σj ’s subject to the
caplet constraint in 1.
As with most models, calibration typically includes minimizing a (possibly weighted) sum of squares over the
parameters of the processes (or functions), σj . This is what we do in Step 2 above although in this case we
have the additional constraints of ensuring that caplets are priced correctly by the model. The ability to
compute the model prices quickly is important for a successful calibration algorithm. Indeed this factor is a big
influence on the choice of calibration securities in practice. As usual, however, the sum of squares is generally a
non-convex function of the decision parameters and so there will typically be many local-minima, an issue which
also needs to be handled carefully. We will not say any more about calibration of market models in these notes
other than to point out that is a very important topic and is still a subject of ongoing research.

5 A Swap Market Model for Pricing Swaptions

Consider a payer-swaption that expires at time Tn > t and with payments of the underlying swap taking place at
b (annualized) and a notional principle of $1, we showed in (7)
times Tn+1 , . . . , TM +1 . Assuming a fixed rate of R
that the time Tn price of the swaption is given by
 
M
X +1 h i+
T b .
CTn = δ ZTnj  R(Tn , Tn , TM ) − R (27)
j=n+1

This implies that the time t price of the swaption, Ct , satisfies

³ ´h i+ 
PM +1 Tj b
δ Z R(Tn , T n , TM ) − R
x  
j=n+1 Tn
Ct = Xt EP t   (28)
XTn

where Xt is the time t price of the chosen numeraire security and Px is the corresponding EMM. A particularly
convenient choice of numeraire that we will adopt is the portfolio15 consisting of δ units of each of the
PM +1 T
zero-coupon bonds maturing at times Tn+1 , . . . , TM +1 . Then Xt = δ j=n+1 Zt j and we find
 
M
X +1 ·h i+ ¸
T b
Ct = δ Zt j  EP t
x
R(Tn , Tn , TM ) − R (29)
j=n+1

14 See Chapter 9 of Cairns for a derivation.

15 There is no difficulty taking a portfolio of securities rather than a fixed individual security as the numeraire. More generally
in fact, we could take a dynamic self-financed portfolio as the numeraire security, assuming of course that it has strictly positive
value at all times.
Market Models 10

Jamshidian (1997) developed a term structure framework where at any time t the current term structure was
given in terms of the forward swap rates, R(t, Ti , TM ) for i = φ(t), . . . , M . In particular, he showed that it was
possible to assume that the Px -dynamics of R(t, Tn , TM ) satisfy

dR(t, Tn , TM ) = R(t, Tn , TM )σ(t)T dW x (t) (30)

where σ(t) is a deterministic vector of volatilities. This implies that the forward swap rate is log-normally
distributed so we can obtain16 Black’s formula for swaption prices (8).

Remark 4 When we model swap rates directly as in (30) we say that we have a swap market model. This
contrasts with the LIBOR market models of Section 4.

Remark 5 The advantage of Black’s swaption formula is that it is elegant and exact, whereas the BGM
formula is cumbersome and only an approximation. However, the BGM approximation is consistent with Black’s
formulae for caplets and caps whereas Black’s swaption formula is not. Indeed, it may be shown17 that if
forward LIBOR rates have deterministic volatilities then it it is not possible for swap rates to also have
deterministic volatilities. Therefore Black’s formulae for caplets and swaptions cannot both hold within the same
model. That said, within the LIBOR market framework with deterministic volatilities, it can be argued that
forward swap rates are approximately log-normally distributed.

6 Monte-Carlo Simulation
While it is possible to price many commonly traded derivative securities such as caps, floors and swaptions in
the market model framework, it is in general necessary to use Monte Carlo methods to price other securities.
Indeed, if our market model has stochastic volatility functions then it will typically be necessary to also use
Monte Carlo methods to price even caps, floors and swaptions.
The typical approach is to use some discretization scheme such as the Euler scheme when performing the Monte
Carlo simulation. This does not create too much of a computational burden as we will only need to simulate the
SDE’s describing the forward LIBOR dynamics for a finite number of maturities. This contrasts with the HJM
framework where we had infinitely many maturities which meant it was practically infeasible to use a very fine
discretization. This in turn prompted the development of the discrete-time HJM framework with the resulting
discrete-time arbitrage-free restriction on the drift.
It is also possible to develop discrete-time arbitrage-free market models in a manner that is analogous to our
discrete-time HJM development. As described above, however, the need to do so is not as urgent as it is
practically feasible to simulate the market model SDE’s on a sufficiently fine grid and this is what is typically
done in practice.
Nonetheless, Glasserman’s Monte Carlo Methods for Financial Engineering describes how to build discrete-time
arbitrage-free market models. It turns out to be inconvenient to choose the LIBOR rates as the fundamental
variables that we choose to discretize. Instead it is more convenient to directly model deflated bond prices as
discrete-time Q-martingales18 and to define LIBOR rates in terms of these bond prices. Other choices of
discretization variable are also possible. As usual, we can choose to simulate under any EMM that we prefer and
all of the usual variance reduction techniques may be employed.

16 Ofcourse we need to reinterpret σ in (8) in terms of the deterministic function σ(t) in (30).
17 This is done by applying Itô’s Lemma to the forward swap rate given in (5).
18 This ensures the discrete-time model is+ arbitrage-free.
Market Models 11

7 Hedging and the Greeks

For the swap market model with deterministic volatilities, it is straightforward to construct self-financing
hedging strategies for swaptions (and therefore caplets and floorlets as they are simply one-period swaptions).
But these hedging strategies rely on the assumption that the volatilities are deterministic. In practice, users of
both the LIBOR and swap market models will want to hedge against changes in volatility. How they hedge will
typically be very model dependent and ir probably as much of an art as a science. It also requires a very good
understanding of the strengths and weaknesses of the model or models that they are using to hedge.

8 Pricing Bermudan Swaptions

Perhaps the most commonly traded exotic interest rate derivative is the Bermudan swaption. We already saw
how to price these options using lattice models earlier in the course but they are considerably more difficult to
handle in the context of market models. This is because market models are high-dimensional, with a separate
state variable for each forward rate. Lattice or finite difference methods do not work well in high-dimensions due
to the so-called curse of dimensionality and so other solution techniques are required. In the earlier part of this
decade simulation methods based on cross-path regressions were developed and these have proved extremely
successful at generating good approximate prices to Bermudan / American options. We will first describe the
cross-path regression technique which is used to generate a good feasible exercise strategy. Because it results in
a feasible exercise strategy, the fair value of this strategy constitutes a lower bound on the true price of the
Bermudan option. The better the exercise strategy the closer the corresponding lower bound will be to the true
option price.

8.1 Computing Lower Bounds using Cross-Path Regressions

The general19 Bermudan option pricing problem at time t = 0 is to compute
· ¸
hτ
V0 := sup EQ0 (31)
τ ∈T B τ

where T = {0 ≤ t1 , . . . , tn = T } is the set of possible exercise dates, Bt is the value of the cash account at
time t and ht = h(Xt ) is the payoff function if the option is exercised at time t. Xt represents the time t
(vector) value of the state variables in the model. In the case of a Bermudan swaption in the LIBOR market
model, for example, Xt would represent the time t value of the various forward LIBOR rates. In theory (31) is
easily solved using value20 iteration. In particular, we would obtain

VT = h(XT ) and
µ · ¸¶
Bt
Vt = max h(Xt ), EQ
t V t+1 (Xt+1 ) .
Bt+1

The price of the option is then given by V0 (X0 ) where X0 is the initial state vector. As an alternative to value
iteration we could use Q-value iteration. If the Q-value function is defined to be the value of the option
conditional on it not being exercised today, i.e. the continuation value of the option, then we also have
· ¸
Q Bt
Qt (Xt ) = Et Vt+1 (Xt+1 ) . (32)
Bt+1
19 We will specialize to the Bermudan swaption later.
20 The term “value iteration” refers to the dynamic programming approach of computing the value function iteratively by
working backwards in time. This is how we compute the price of American options in the binomial model.
Market Models 12

The value of the option at time t + 1 is then

Vt+1 (Xt+1 ) = max(h(Xt+1 ), Qt+1 (Xt+1 )) (33)
so that if we substitute (33) into (32) we obtain
· ¸
Q Bt
Qt (Xt ) = Et max (h(Xt+1 ), Qt+1 (Xt+1 )) . (34)
Bt+1
Equation (34) clearly gives a natural analog to value iteration, namely Q-value iteration. If Xt is high
dimensional, then both value iteration and Q-value iteration are not feasible in practice. However, we could
perform an approximate and efficient version of Q-value iteration, and we now describe how to do this using
cross-path regressions.
The first step is to choose a set of basis functions, φ1 (X), . . . , φm (X). These basis functions define the linear
architecture that will be used to approximate the Q-value functions. In particular, we will approximate Qt (Xt )
with
Qe t (Xt ) = r1 φ1 (Xt ) + . . . + rm φm (Xt )
t t
where rt := (rt1 , . . . , rtm ) is a vector of time t parameters that is determined by the algorithm which proceeds as
follows:
Cross-Path Regression Algorithm for Approximate Q-Value Iteration

generate N independent paths of the state vector, Xt for t = 1, ..., T

set Qe T (X i ) = 0 for all i = 1 to N
T
for t = T − 1 Down to 1
Estimate rt = (rt1 , . . . , rtm )
set Qe t (Xti ) = P rtk φk (Xti ) for all i
k
end for ³ ´
set Ve0 (X0 ) = max h(X0 ), Q e 0 (X0 )

Two steps require further explanation.

³ ´
e t+1 ) on (φ1 (Xt ), . . . , φm (Xt )) where
1. First, we estimate rt by regressing α max h(Xt+1 ), Q(X
α = Bt /Bt+1 is the discount factor for moving from t + 1 to t. We have N observations for this
regression and N is usually taken to be somewhere between 10, 000 and 50, 000, though it will of course
depend on the problem at hand.
e 0 (X0 ) by averaging and
2. Second, since all N paths have the same starting point, X0 , we can estimate Q
e 1 (·) evaluated at the N successor points of X0 .
discounting Q
Remark 6 Obviously many more details are required to fully specify the algorithm. In particular, the parameter
values N and m, and the basis functions need to be chosen. Specific implementation details can also vary. For
example, if the option is out-of-the-money at some time t and in some simulated state Xt , then you could
choose to omit this sample from the regression at time t. In practice the selection of basis functions is vital.

In practice, it is quite common for an alternative estimate, V 0 , of V0 to be obtained by simulating the exercise
strategy that is defined implicity by the sequence of Q-value function approximations. That is, we define
τe = min{t ∈ T : Q e t ≤ ht } and
· ¸
he
V 0 = EQ0
τ
.
Be
τ
V 0 is then an unbiased lower bound on the true value of the option as it is the price that corresponds to a
feasible exercise strategy.
Market Models 13

Exercise 4 Given how the lower bound, V 0 , is computed, can you guess why we prefer to do an approximate
Q-value iteration instead of an approximate value-iteration?

These algorithms21 have performed surprisingly well on realistic high-dimensional problems and there has also
been considerable theoretical work explaining why this is so. The quality of V 0 , for example, can be explained in
part by noting that exercise errors are never made as long as Qt (·) and Qe t (·) lie on the same side of the optimal
exercise boundary. This means in particular, that it is possible to have large errors in Q e t (·) that do not impact
the quality of V 0 .

8.2 Back to Bermudan Swaptions

We can now use the algorithm of Section 8.1 to price Bermudan swaptions in the LIBOR market model. But
first, we will define again the Bermudan swaption contract.

Definition 2 Let T = {0 ≤ t1 , . . . , tn−1 } be the set of possible exercise dates. Then the holder of a
Bermudan swaption has the right to enter at any time t ∈ T into an interest-rate swap with fixed rate K that
expires with last payment at time tn = T . If the swaption is exercised at time tl , then the first reset point is tl .

A payer Bermudan swaption is a a swaption where the holder of the option will pay the fixed rate if the option is
exercised. A receiver Bermudan swaption, on the other hand, will receive the fixed rate if the option is exercised.
Note, for example, that a 2 − 8 Bermudan swaption is an option where the first exercise date occurs in 2 years22
time and, regardless of when (if ever) the option is exercised, the underlying swap will expire in 2 + 8 = 10 years
time.
In order to price the Bermudan swaption, we need to specify the parameters for the cross-path regression
algorithm and in particular, the basis functions. We could, for example, select

Lj (ti ), . . . , Ln−1 (ti ), Sn (ti ), Sn (ti )2 and Sn (ti )3

as our time ti basis functions where

1 − Zin
Sn (ti ) := Pn .
δ j=i+1 Zij
is the underlying swap rate at time ti . Note that under this specification, the number of basis functions
decreases by one as you move from one period to the next. Once we specify the number of paths, N , to
simulate we can apply the cross-path regressions algorithm and obtain a lower bound on the price of the
Bermudan swaption. the quality of this lower bound, will depend mainly on the chosen set of basis functions.

Appendix: Computing Upper Bounds to Bermudan Option Prices

While the cross-path regression algorithm has been very successful in practice, an important weakness is the
inability to determine how far the solution, i.e. estimated option price, is from the true option price in any given
problem. Dual-based methods have recently been developed23 for evaluating any approximate solution by using
it to construct an upper bound on the option price. This is of value since we can then compute lower and upper
bounds on the optimal price of the Bermudan option and therefore determine how far the approximate solution
is from optimality. We now describe these dual based methods.
For an arbitrary supermartingale, πt , the value of an American option, V0 , satisfies
21 They were introduced originally by Longstaff and Schwartz (2001) and Tsitsiklis and Van Roy (2001). But see also Carriere

(1996).
22 Or months, depending on the context.
23 The dual methods were introduced independently by Haugh and Kogan (2004) and Rogers (2002).
Market Models 14

· ¸ · ¸
hτ hτ
V0 = sup EQ
0 = sup EQ
0 − πτ + πτ
τ ∈T Bτ τ ∈T Bτ
· ¸
Q hτ
≤ sup E0 − πτ + sup EQ 0 [πτ ]
τ ∈T Bτ τ ∈T
· ¸
hτ
≤ sup EQ 0 − πτ + π0
τ ∈T B τ
· µ ¶¸
Q ht
≤ E0 max − πt + π0 (35)
t∈T Bt

where the second inequality follows from the optional sampling theorem for supermartingales. Taking the
infimum over all supermartingales, πt , on the right hand side of (35) implies
· µ ¶¸
Q ht
V0 ≤ U0 := inf E0 max − πt + π0 (36)
π t∈T Bt

On the other hand, it is a known fact that the process Vt /Bt is itself a supermartingale, which implies
· ¸
Q
U0 ≤ E0 max (ht /Bt − Vt /Bt ) + V0 .
t∈T

Since Vt ≥ ht for all t, we conclude that U0 ≤ V0 . Therefore, V0 = U0 , and equality is attained when
πt = Vt /Bt .
This shows that an upper bound on the price of the American option can be constructed simply by evaluating
the right-hand-side of (35) for a given supermartingale, πt . In particular, if such a supermartingale satisfies
πt ≥ ht /Bt , the option price V0 is bounded above by π0 .
When the supermartingale πt in (35) coincides with the discounted option value process, Vt /Bt , the upper
bound on the right-hand-side of (35) equals the true price of the American option. This suggests that a tight
upper bound can be obtained by using an accurate approximation, Vet , to define πt . One possibility24 is to define
πt as a martingale:

π0 = Ve0 (37)
" #
Vet+1 Vet Vet+1 Vet
πt+1 = πt + − − Et − . (38)
Bt+1 Bt Bt+1 Bt

Let V 0 denote the upper bound we get from (35) corresponding to our choice of supermartingale in (37) and
(38). Then it is easy to see that the upper bound is explicitly given by
" Ã · ¸!#
h e
V X
t
e
V e
V
V 0 = Ve0 + EQ
t t j j−1
0 max − + EQj−1 − . (39)
t∈T Bt Bt Bj Bj−1
j=1

As may be seen from (39), obtaining an accurate estimate of V 0 is computationally demanding. First, a number
of sample paths must be simulated to estimate the outermost expectation on the right-hand-side of (39). While
this number need not be large in practice, we also need to accurately estimate a conditional expectation at each
time period along each simulated path. This requires some effort and clearly variance reduction methods would
be useful in this context. Low discrepancy sequences are also very useful for this task. Variations and extensions
of these algorithms have also been developed recently and are a subject of ongoing research.25

24 See Haugh and Kogan (2002) and Andersen and Broadie (2002) for further comments related to the choice of πt .
25 See Chapter 8 of Glasserman (2003) for a detailed treatment of Monte Carlo methods for pricing American options.