Term Structure Models: IEOR E4710 Spring 2010

c 2010 by Martin Haugh

°

Continuous-Time Short Rate Models

These notes provide an overview of single- and multi-factor models of the short rate. We will begin with a
generic single-factor model where the dynamics of rt under the physical measure, P , are given. Following the
approach of Vasicek, we then derive the PDE that must be satisfied by derivative security prices. We then use
the martingale approach to give an alternative (and familiar) expression for derivative security prices. The
consistency of the two approaches is then demonstrated using the Feynman-Kac PDE representation. We show
how the Martingale Representation theorem can be used to construct hedging strategies and then discuss some
specific single-factor models. These examples include the Vasicek and CIR models, and more generally, affine
models. We then conclude with multi-factor models and describe some specific examples.
Once we have moved on to discussing specific models, we will generally describe the dynamics of rt (and other
factors in multi-factor models) under the EMM, Q, and not bother with the P -dynamics of rt . This is the
standard approach in term-structure modeling and was the approach we followed when we used binomial and
trinomial lattices to model the short rate.
Multi-factor short-rate models have to some extent been superseded by market models over the past decade.
They remain important, however, as they are often applied in risk-management settings and are also used for
pricing hybrid and credit derivatives. Moreover, they are sometimes employed alongside market models for
pricing fixed income derivatives as they are well understood and can be used to provide ‘sanity’ checks on the
more complicated models that might be used by trading desks.

1 The PDE Approach

Let us assume that the P -dynamics of the short rate are given by the SDE

drt = αp (t, rt ) dt + βp (t, rt ) dWtP . (1)

where WtP is a P -Brownian motion. In particular, rt is a Markov process. Consider now the time t price, ZtT , of
a zero-coupon bond maturing at time T . We assume that ZtT is a sufficiently smooth function of (t, rt ) so that
Itô’s Lemma implies
· T ¸
T ∂Z ∂Z T 1 2 T
2∂ Z ∂Z T
dZt = + αp (t, rt ) + βp (t, rt ) dt + β p (t, r t ) dWtP (2)
∂t ∂r 2 ∂r2 ∂r
£ ¤
= ZtT m(t, r; T ) dt + S(t, r; T ) dWtP (3)

where m(t, r; T ) and S(t, r; T ) are implicitly defined1 by (3). Now consider times t < T1 < T2 where T1 and T2
are fixed. We will first construct a self-financing portfolio (portfolio A) involving2 the cash account, Bt , and ZtT2
1 We T T
will also use mt i and St i to denote m(t, r; Ti ) and S(t, r; Ti ), respectively.
2 It T
is a slight abuse of notation to use Zt i to refer to both the zero-coupon bond price and the zero-coupon bond itself but
we do so to avoid introducing further notation!
Continuous-Time Short Rate Models 2

that replicates ZtT1 (portfolio B).

Portfolio A: Hold at units of ZtT2 and bt units of Bt at time t.

Portfolio B: Hold 1 unit of ZtT1 at time t.

We choose at and bt such that portfolio A is self-financing3 and the two portfolios have equal value, i.e.,
VtA = VtB where VtA := at ZtT2 + bt Bt and VtB := ZtT1 . We then have

at dZtT2 + bt dBt = dZtT1 . (4)

Exercise 1 Why does (4) imply that portfolio A is self-financing?

Substituting (3) into (4) we obtain

h i h i
at ZtT2 mTt 2 dt + StT2 dWtP + bt rt Bt dt = ZtT1 mTt 1 dt + StT1 dWtP . (5)

Equating drift and volatility terms then implies

StT1 ZtT1
at =
StT2 ZtT2
" #
1 T1 T1 mTt 2 ZtT1 StT1
bt = m t Zt − .
rt Bt StT2

Since VtA = VtB , we then have

" #
StT1 ZtT1 T2 1 T1 T1 mTt 2 ZtT1 StT1
Zt + m t Zt − Bt = ZtT1 . (6)
StT2 ZtT2 rt Bt StT2

Simplifying we obtain
" #
StT1 1 T1 mTt 2 StT1
+ mt − = 1
StT2 rt StT2

mTt 1 − rt mTt 2 − rt
⇒ = =: λ(t, rt ) (7)
StT1 StT2

independently of T . Note that in deriving (7) we only used the fact that ZtTi was the price of a non-dividend
paying security and did not use any other properties of zero-coupon bonds. We therefore obtain that for any
derivative security with time t price, Pt , and dynamics given by
£ ¤
dPt = Pt µ(t, rt ) dt + σ(t, rt ) dWtP

we must have λ(t, rt ) = (µ(t, rt ) − r)/σ(t, rt ).

Remark 1 λ(t, rt ) is usually called the market-price-of risk. Since rt is stochastic4 , we sometimes refer to the
market-price-of-risk process.

Using (2), (3) and (7) we obtain5 the PDE that any derivative security price, Pt , must satisfy
∂P ∂P 1 2 ∂2P
(αp,t − λt βp,t ) + + βp,t − rt P = 0. (8)
∂r ∂t 2 ∂r2
3 Note that portfolio B is clearly self-financing.
4 Inthe Black-Scholes model for equities λt is a constant since each of µ, σ and r is also constant.
5 We could also derive a similar PDE for dividend paying securities.
Continuous-Time Short Rate Models 3

In order to solve (8) we also need to specify boundary conditions. These conditions will depend on the nature of
the derivative security. For example, if Pt is the price of a zero-coupon bond that matures at time T > t then
the boundary condition is PT ≡ 1.

Remark 2 Note that if we know the dynamics of just one traded security, i.e. we know µ(t, rt ) and σ(t, rt ) for
that security, then we know λ(t, rt ) and we can then use (8) to compute the price of any other security.

Remark 3 We sometimes call (8) the Fundamental PDE for 1-factor models. For multi-factor models we can
easily derive an analogous PDE.

2 The Martingale Approach

In the martingale approach we begin with an EMM, Q, under which all discounted security prices are
martingales. Let us assume that that the Q-dynamics of rt are given by
drt = α(t, rt ) dt + β(t, rt ) dWt (9)
where Wt is a Q-Brownian motion. Then the time t price, Pt , of a traded security (that does not pay any
intermediate cash-flows) is given by · ¸
Q PT
Pt = Bt Et (10)
BT
³R ´
t
where Bt := exp 0 rs ds is the time t value of the cash account. For example, we have
· RT ¸
− rs ds
ZtT = EQ t e t . (11)

An obvious question that comes to mind is how the PDE and martingale approaches are related. We now
discuss this issue and it will not be surprising to see that the Feynman-Kac representation and Girsanov’s
Theorem provide the link.

Connection to PDE Approach

We consider the case of a non-dividend paying security with time t price, Pt , that satisfies (10). The
Feynman-Kac formula then states that Pt also satisfies the following PDE
∂P ∂P 1 ∂2P
α(t, rt )+ + β(t, rt )2 2 − rt P = 0 (12)
∂r ∂t 2 ∂r
Since Pt satisfies both (8) and (12) it must be the case that
βp (t, rt ) = β(t, rt ) and

αp (t, rt ) − α(t, rt )
λ(t, rt ) = . (13)
βp (t, rt )

Comparing (1) and (9) we also see that

dWtP = dWt − λt dt
so that in particular, we have
µ Z t Z ¶
dQ 1 t 2
= exp − λs dWsP − λs ds . (14)
dP 0 2 0
by Girsanov’s Theorem.
Continuous-Time Short Rate Models 4

Remark 4 While we used non-dividend paying securities to derive the relationships (13) and (14), we could6
also have done so using dividend paying securities. In particular, the EMM Q may be used to price all securities
(dividend or non-dividend paying) in the economy.

Remark 5 We mention once again that it is standard in term-structure modeling to model interest rates and
security price processes directly under Q without any direct reference to the physical measure, P .

3 Specific Models
The Vasicek Model
The Vasicek model assumes that the short rate, rt , follows a Gaussian model where

drt = α(µ − rt ) dt + σ dWt (15)

and where as usual, Wt is a Q-Brownian motion. We can then solve (15) by applying Itô’s Lemma to
Yt := exp(αt)rt . We obtain
Z t Z t
Yt = Y0 + αµ eαu du + σ eαu dWu
0 0
so that Z
¡ ¢ t
−αt −αt −αt
rt = e r0 + µ 1 − e + σe eαu dWu . (16)
0
In order to compute the term-structure in Vasicek’s model we have to compute
· RT ¸
T T − rs ds
Z0 = E 0 e 0 . (17)

RT
We therefore need to find the distribution of the random variable X := 0 rs ds. To do this we apply Theorem
7 from the lecture notes, Introduction to Stochastic Calculus and Mathematical Finance, to see that
X ∼ N(a, b2 ) where
(µ − r0 ) ¡ −αT ¢
a = e − 1 + µT
α
Z ÃZ !2 µ ¶
T T
2 2 2αv −αt dt σ2 4e−αT − e−2αT − 3
b = σ e e dv = T+ .
0 v 2α2 2α

If we now observe that E[exp(φX)] = exp(φa + φ2 b2 /2) when X ∼ N(a, b2 ) we then find

Z0T = exp(A(0, T ) − B(0, T )r0 ) (18)

where
1 − e−α(T −t)
B(t, T ) =
µ α ¶
σ2 σ2
A(t, T ) = µ − 2 [B(t, T ) − (T − t)] − B(t, T )2 .
2α 4α
More generally, it is easy to see (either by repeating the above analysis or using the fact that rt is a Markov
process) that
ZtT = exp(A(t, T ) − B(t, T )rt ). (19)

6 The PDEs in (8) and (12) would change slightly to reflect dividend payments.
Continuous-Time Short Rate Models 5

Remark 6 Note that ZtT depends on t and T only through their difference, T − t.

Exercise 2 Find the SDE satisfied by ZtT . What do you notice about its behavior as t → T ?

Example 1 (Derivatives with ZCB as Underlying Security in Vasicek’s Model)

Suppose we wish to price a security that expires at time T with payoff CT := C(ZTU ) where ZTU is the value at
time T of a zero-coupon bond maturing at time U > T . For example, if the security in question is a call option
then C(x) = (x − K)+ . Martingale pricing tell us that the time t security price, Ct , satisfies
· ¸
Q CT
Ct = B t Et . (20)
BT

In order to evaluate the expectation on the right-hand-side of (20) it would appear that we need to find the joint
distribution of ZTU and BT . While this is feasible, we can also compute Ct by changing to the forward measure,
P T . Equation (17) in the lecture notes, Introduction to Stochastic Calculus and Mathematical Finance, tells us
that we can also represent Ct as
T £ ¤
Ct = ZtT EP t C(ZTU ) . (21)
To take advantage of (21), we need to find the P T -dynamics of ZtU . Towards this end, we take the following
steps:

1. Apply Itô’s Lemma to (19) and find that ZtT satisfies

dZtT = rt ZtT dt − ZtT B(t, T )σdWt (22)

= rt ZtT dt − ZtT S(t, T )dWt (23)

where Wt is a Q-Brownian motion and S(t, T ) := σB(t, T ).

2. Now define YtU := ZtU /ZtT and again apply Itô’s Lemma to find that

dYtU = YtU S(t, T ) [S(t, T ) − S(t, U )] dt + YtU [S(t, T ) − S(t, U )] dWt (24)

3. Note that by definition of P T , it must be that YtU is a P T -martingale. Girsanov’s Theorem then tells us
T
that WtP is a P T Brownian motion where
Z t
PT
Wt := Wt + S(v, T ) dv.
0

Note also that we then have T

dYtU = YtU [S(t, T ) − S(t, U )] dWtP (25)

4. Note that since ZTT ≡ 1 we may write (21) as

T £ ¤
Ct = ZtT EP
t C(YTU ) . (26)

Therefore to compute Ct we need the P T -distribution of YTU .

Exercise 3 What is the P T -distribution of YTU ? (Hint: Note that the volatility term, S(t, U ) − S(t, T ), is a
deterministic function of t!)
Continuous-Time Short Rate Models 6

Exercise 4 Give a Black-Scholes-like expression for the time t price of a European call option that expires at
time T with payoff (ZTU − K)+ .

Remark 7 Note that the analysis of Example 1 is not specific to the Vasicek model. In fact, whenever the
volatility process, S(t, T ), for the time T -maturing zero-coupon bond is deterministic, we can obtain a similar
Black-Scholes-like formula.

Remark 8 We could also have solved the problem of Example 1 without switching to the measure, P T , by
RT
showing that ( 0 rs ds, rT ) has a bivariate normal distribution under Q . (See Cairns for a proof based on
RT
computing the joint moment generating function of ( 0 rs ds, rT ).)

The Cox-Ingersoll-Ross Model

The Cox-Ingersoll-Ross (CIR) model assumes that the short-rate, rt , satisfies
√
drt = α(µ − rt ) dt + σ rt dWt ; r0 > 0. (27)

where α and µ are positive constants. The principal advantage of the CIR model over the Vasicek model is that
the short rate is guaranteed to remain non-negative.7 Unlike the Vasicek model, however, the CIR model is not
Gaussian and is therefore considerably more difficult8 to analyze. Perhaps surprisingly, zero-coupon bond prices
can still be computed analytically and are given by
· RT ¸
− rs ds
ZtT = EQ t e t = exp(A(T − t) − B(T − t)rt ). (28)

where
µ ¶
2αµ 2γe(γ+α)τ /2
A(τ ) = log
σ2 (γ + α)(eγτ − 1) + 2γ
γτ
2(e − 1)
B(τ ) =
(γ + α)(eγτ − 1) + 2γ
p
γ = α2 + 2σ 2

Remark 9 The expression for ZtT in (28) can be verified by checking that it satisfies the PDE (12) with
√
PT ≡ 1, α(t, rt ) = α(µ − rt ) and β(t, rt ) = σ rt .

Closed form solutions for option prices on zero-coupon bonds can also be found in this model. In general,
derivatives prices can be estimated by either numerically solving the PDE in (12) with appropriate boundary
conditions or by using Monte-Carlo methods.

Exercise 5 What is the time 0 forward price for delivery at time τ of a zero-coupon bond that matures at time
T > τ in the CIR model?

7 Indeed, it can be shown that if 2αµ > σ 2 then the short rate will remain strictly positive. See Cairns for a proof of this

fact.
8 See Cairns or Shreve for detailed treatments of the CIR model.
Continuous-Time Short Rate Models 7

4 Generalization to Time-Varying Parameters

The ability to calibrate models to market data is a desirable feature of any model and this points to one of the
main drawbacks9 of the Vasicek and CIR models. These models only have a finite number of free parameters10
and so it is not possible to specify these parameter values in such a way that model prices (e.g. for zero-coupon
bonds of different maturities) coincide with observed market prices. This problem is overcome11 by allowing the
parameters to vary deterministically with time. Some of the more well-known models with time-varying
parameters are described below.
Example 2 (The Ho-Lee Model)
We assume that the short-rate satisfies
drt = θ(t) dt + σdWt
where θ(t) is a deterministic function of time. It is possible to extend the Ho-Lee model and also make σ(t) a
deterministic function of time.

Example 3 (The Black-Derman-Toy Model)

This model assumes that Yt := log(rt ) satisfies

dYt = θ(t) dt + σdWt

where again θ(t) is a deterministic function of time. Itô’s Lemma may be applied to see that rt is a geometric
Brownian motion. This model was originally specified as a lattice model.

Example 4 (The Black-Karansinski Model)

The Black-Karansinski model is a generalization of the Black-Derman-Toy model where we assume that
Yt := log(rt ) satisfies
dYt = α(t)(log µ(t) − Yt ) dt + σ(t)dWt
where α(t), µ(t) and σ(t) are deterministic functions of time. Itô’s Lemma can be applied to find the dynamics
of rt .

Example 5 (Hull and White)

The Hull-White model generalizes the Vasicek model and assumes

drt = α(µ(t) − rt ) dt + σdWt

where µ(t) is a deterministic function of time that may be interpreted as a local mean-reversion level. It is also
possible for α and σ to be deterministic functions of time. The Hull-White model is a Gaussian model and so it
is straightforward to price derivatives using the same techniques we used for the Vasicek model.

Single-Factor Affine Term-Structure Models

Some of the models we have described thus far (e.g. Vasicek, CIR, Hull and White) have term structures of the
form
ZtT = exp (a(t, T ) + b(t, T )rt ) . (29)
¡ ¢
If we define the yield as YtT := − log(ZtT )/(T − t) so that ZtT = exp −(T − t)YtT , then we see that these
models have yields that are affine in rt . A model with this property is called an affine term-structure model. We
have the following result.
9 Of
course models with too many free parameters are often guilty of over-fitting.
10 Three,to be specific.
11 Though new problems associated with over-fitting can then arise!
Continuous-Time Short Rate Models 8

Theorem 1 Consider a 1-factor model of the form drt = α(t, rt ) dt + β(t, rt ) dWt where Wt is a
Q-Brownian motion. Then for all t ∈ [0, T ],

ZtT = exp (a(t, T ) + b(t, T )rt ) (30)

if and only if α(rt , t) and β(rt , t)2 are affine in rt , i.e.

α(t, r) = α1 (t) + α2 (t)r (31)

β(t, r)2 = β1 (t) + β2 (t)r (32)

Proof:
(i) Suppose (30) holds. Then we may apply Itô’s Lemma to obtain
· ¸
∂a ∂b 1
dZtT = ZtT + rt + b(t, T )α(t, rt ) + b(t, T )2 β(t, rt )2 dt + ZtT β(t, rt ) dWt . (33)
∂t ∂t 2

However under Q it must also be the case that the instantaneous drift of dZtT is given by rt ZtT dt. This implies
∂a ∂b 1
+ rt + b(t, T )α(t, rt ) + b(t, T )2 β(t, rt )2 − rt = 0. (34)
∂t ∂t 2
Now if we differentiate twice with respect to rt across (34) we see that

∂2 1 2 ∂
2
b(t, T ) α(t, rt ) + b(t, T ) (β(t, rt )2 ) = 0. (35)
∂r2 2 ∂r2
Since (35) must hold for all T , it must be the case that

∂2 ∂2
α(t, rt ) = 0 and (β(t, rt )2 ) = 0
∂r2 ∂r2
implying (31) and (32) as desired.

(ii) For the opposite direction, suppose now that (31) and (32) hold. We first substitute (31) and (32) into
(12) to obtain

∂ZtT ∂ZtT 1 ∂ 2 ZtT

[α1 (t) + α2 (t)r] + + [β1 (t) + β2 (t)r] − rt P = 0 (36)
∂r ∂t 2 ∂r2
ZTT = 1

and then use (30) to substitute expressions for the partial derivatives in (36). The resulting PDE may then be
reduced to a pair of ODE’s in a(t, T ) and b(t, T ) for which solutions may be shown to exist subject to technical
conditions.

Other Models
As we saw in (9) a generic single-factor model for rt is given by

drt = α(t, rt ) dt + β(t, rt ) dWt (37)

where Wt is a Q-Brownian motion. Many single-factor models can be therefore be analyzed once we specify the
functional forms of α(t, rt ) and β(t, rt ) in (37). In general, care is needed in specifying α(t, rt ) and β(t, rt ) so
that: (i) (37) has a solution and (ii) the solution implies behavior for rt that is satisfactory. Once these
conditions are satisfied, the principal concerns are whether or not the model is tractable and can easily be
calibrated to market data, and whether or not the empirical properties of the model are consistent with those
observed in the market-place.
Continuous-Time Short Rate Models 9

5 Hedging in Single-Factor Models

We now describe how to hedge a derivative security in a generic 1-factor model where

drt = α(t, rt ) dt + β(t, rt ) dWt .

and where Wt is a Q-Brownian motion. We assume the derivative expires at some time τ > 0 and that it does
not pay intermediate cash-flows between 0 and τ . Its time t value is denoted by Ct so that its payout upon
expiration is Cτ . Martingale pricing then tells us that Ct /Bt is a Q-martingale. The Martingale Representation
Theorem then gives the existence of an adapted process, φt , such that
Z t
Ct
= C0 + φs dWs , (38)
Bt 0
h Rt i
assuming as usual that B0 = 1. In particular, we have Ct = Bt C0 + 0 φs dWs so Itô’s Lemma then implies
· Z t ¸
dCt = C0 + φs dWs rt Bt dt + Bt φt dWt
0
= rt Ct dt + Bt φt dWt . (39)

In order to hedge the derivative, we need to choose our hedging securities. We will use the cash account, Bt ,
and some other12 security, Pt say, with Q-dynamics given by

dPt = Pt [rt dt + σt dWt ] (40)

where σt is an adapted process that is strictly greater than 0. We can now rewrite (39) as

Bt φt
dCt = rt Ct dt + σt Ct dWt . (41)
σ t Ct
Recalling equation (15) in the Introduction to Stochastic Calculus and Mathematical Finance lecture notes, we
see that the Q-dynamics of the wealth process, Vt , associated with a self-financing trading strategy are given by

dVt = rt Vt dt + θt σt Vt dWt (42)

where θt and (1 − θt ) are the fractions of time t wealth, Vt , invested in the risky security, Pt , and cash account,
respectively, at time t.
Now if we compare (41) and (42), we see that the self-financing strategy that replicates Cτ is a portfolio that at
time t invests θt := Bt φt /σt Ct in Pt and $(1 − θt )Ct in the cash account.

Exercise 6 Suppose we choose the cash account and a zero-coupon bond maturing at time T as our hedging
instruments. What must the relationship between τ and T be?

Exercise 7 How would you go about hedging a European option on a zero-coupon bond in the CIR model?

12 P might, for example, represent the price of a particular zero-coupon bond. Note also that the Q-dynamics of any traded
t
security must have a drift equal to rt .
Continuous-Time Short Rate Models 10

6 Multi-Factor Models
In the single-factor affine term-structure models we saw that the zero-coupon bonds prices were given by

ZtT = exp (a(t, T ) + b(t, T )rt ) (43)

for some deterministic functions a(t, T ) and b(t, T ). This then implies that

dZtT
= rt dt + b(t, T )σ(t, rt ) dWt (44)
ZtT

where σ(t, rt ) is the volatility coefficient for rt . Equation (44) implies that the returns on zero-coupon bonds of
different maturities are instantaneously perfectly correlated since

b(t, T1 )σ(t, rt ) · b(t, T2 )σ(t, rt )

p p = 1.
b(t, T1 )2 σ(t, rt )2 b(t, T2 )2 σ(t, rt )2

Indeed, this is the reason why we only need the cash account and one other security to hedge derivative
securities in these models. Depending on the application of interest, this can be a very unsatisfactory property
of single-factor models. For example, a single factor model would be entirely inappropriate for pricing a
slope-of-the-yield-curve option with a time T payoff given by

h(rT , T ) := max (S(rT , T ) − K, 0)

YTT2 − YTT1
where S(rT , T ) = ; T < T 1 < T2 .
T2 − T1
More generally, it would be unwise to use a single-factor model for pricing fixed-income derivatives that do not
mature in the relatively near future, e.g. 1 or 2 years. A similar comment applies to a derivative written on an
underlying security that does not mature in the near future, regardless of whether or not the derivative13 itself
matures in the near future. The simple reason for this is the well-recognized fact that one factor can not
adequately explain movements in the entire term structure. For example, we often see yields at opposite ends of
the term structure move in opposite directions. This behavior is more easily explained with multi-factor models.
Multi-factor models were introduced primarily to overcome these problems. For example, in our brief discussion
of Gaussian multi-factor models below we will see that less than perfect instantaneous correlations between
bond returns are possible.

Gaussian Multi-Factor Models

Pn (i)
We can specify a Gaussian multi-factor model for the short rate by setting rt = i Xt and assuming that the
Q-dynamics of Xt ∈ Rn are given by

dXt = A(µ − Xt ) dt + C dWt (45)

where Wt is an n-dimensional Q-Brownian motion, µ ∈ Rn and A and C are n × n matrices. It can then be
shown that the solution to (45) is a Gaussian process given by
Z t Z t
−At −A(t−s)
Xt = e X0 + e µ ds + e−A(t−s) C dWs . (46)
0 0

Zero-coupon bond prices can again be shown to be affine and have the form
¡ ¢
ZtT = exp a(t, T ) + b(t, T )Xt
13 A 1-year-10-year swaption is such an example.
Continuous-Time Short Rate Models 11

where a(t, T ) is a scalar function and b(t, T ) is an Rn -valued function. Itô’s Lemma then implies that the
Q-dynamics of ZtT are given by
dZtT
= rt dt + b(t, T )σ dWt (47)
ZtT
where σ is the n × n volatility matrix for rt .

Exercise 8 Compute the instantaneous correlation coefficient, ρ, between returns on zero-coupon bonds of
maturities T1 and T2 . Note that ρ need not equal 1.

Gaussian multi-factor models in general have been very useful in practice owing to their tractability and ability
to capture many different empirical aspects of the data. Brigo and Mercurio14 , for example, describe the G2++
model where the short rate satisfies rt = xt + yt + φt with xt and yt modeled as dependent Gaussian processes.
Here φt is a deterministic function of time that is used to calibrate the model to the initial term structure in the
market place. Brigo and Mercurio devote substantial time to this model and describe how, for example,
zero-coupon bonds, options on zero-coupon bonds, caps, floors, futures and constant-maturity swaps can all be
priced analytically in this model. The pricing of swaptions simply amounts to numerically computing a
one-dimensional integral. This broad tractability as well as our ability to simulate it easily, has made it one of
the work-horses of the fixed-income derivatives world.

The Multi-Factor CIR Model

(i)
The multi-factor CIR model15 builds upon the single-factor CIR model by assuming that each of n factors, Xt
for i = 1, . . . , n, follows CIR-type processes that are Q-independent. If the short rate, rt , then satisfies
Pn (i)
rt = i Xt we can use the results from the single-factor CIR model to determine the term structure.
(i)
For each i = 1, . . . , n we therefore assume that Xt satisfies
³ ´ q
(i) (i) (i) (i) (i)
dXt = ai µi − Xt dt + ci Xt dWt ; X0 > 0 (48)

(i)
where the Wt ’s are Q-independent Brownian motions and ai , ci and µi are positive constants. Since each
(i)
Xt is a single-factor CIR process, we also know that there exist functions Ai (t) and Bi (t) such that
" Ã Z !#
T ³ ´
Q (i) (i)
Et exp − Xu du = exp Ai (T − t) + Bi (T − t)Xt . (49)
t

P (i)
If rt = R(t, Xt ) := i Xt then we have enough information to compute the term structure. In particular
" Ã Z !#
T
T Q
Zt = Et exp − ru du
t
" Ã Z n
!#
T X
= EQ
t exp − Xu(i) du
t i=1
" Ã n Z
!#
X T
= EQ
t exp − Xu(i) du
i=1 t
" n
à Z !#
Y T
= EQ
t exp − Xu(i) du
i=1 t
n
" Ã Z !#
Y T
= EQ
t exp − Xu(i) du
i=1 t

14 “Interest Rate Models: Theory and Practice” published by Springer-Verlag.

15 See Duffie’s Dynamic Asset Pricing for a more complete discussion.
Continuous-Time Short Rate Models 12
n
Y ³ ´
(i)
= exp Ai (T − t) + Bi (T − t)Xt
i=1
à n ³
!
X ´
(i)
= exp Ai (T − t) + Bi (T − t)Xt (50)
i=1

Remark 10 There are two methods by which we can make the multi-factor CIR model operable:
(i)
(i) We identify the factors, Xt for i = 1, . . . , n, as economically relevant and measurable variables such as
inflation rates, bond yields, economic indicators etc. We then calibrate the model parameters to these variables
and the prices of interest-rate dependent securities. This would appear to be a challenging task.
(i)
(ii) As in (i), we also assume that the factors, Xt for i = 1, . . . , n, represent economically relevant variables.
However, instead of actually trying to identify and measure these variables we make a change-of-variable
substitution so that bond yields of n different maturities now make up the state variables. Note that bond yields
(i)
at time t can be expressed in terms of the Xt ’s. This is clear from (50) and Itô’s Lemma then enables us to
write the dynamics satisfied by our new state variables, i.e. the n bond yields. This model, sometimes called a
yield-factor model, is now much easier to calibrate.

A multi-factor CIR model is superior in some respect to the multi-factor Gaussian mode. For example, interest
rates cannot go negative in the multi-factor CIR model. However, the CIR model is not as tractable as the
Gaussian model. Indeed, in deriving (50) it was necessary to assume that the two Brownian motions were
Q-independent. This was not necessary in the Gaussian case. One of the down-sides of having to assume a zero
correlation is that it limits the range of behaviors that are possible. For example, the term structure of
volatilities of the instantaneous forward rates must be downward sloping in such a CIR model whereas in
practice, a more humped shape is often observed. The multi-factor Gaussian model can capture this behavior.

Multi-Factor Affine Models

More generally, a multi-factor affine model16 for the short-rate is specified by assuming that rt = a + b · Xt and
that the Q-dynamics of Xt ∈ Rn are given by

dXt = (c + DXt ) dt + Sσ(Xt ) dWt (51)

where Wt is an n-dimensional Q-Brownian motion, c is an n-dimensional constant vector and D and S are
n × n constant matrices. We also take σ(Xt ) to be an n × n diagonal matrix with (i, i)th element given by
p
γiT Xt + δi where δi is a constant and γi an n-dimensional constant vector. Duffie and Kan (1996) then
showed that if the term structure has the form

ZtT = exp (A(t, T ) + B(t, T )Xt )

then it must be the case that Xt satisfies17 (51).

Exercise 9 Is the multi-factor CIR model an affine model?

Derivatives Pricing and Hedging in Multi-Factor Models

Pricing and hedging in multi-factor proceeds along the same lines as in single-factor models. If closed form
solutions are not available for derivative prices, then they may be computed numerically either using Monte
Carlo simulation or by solving the associated Feynman-Kac PDE. For example, suppose Xt ∈ Rn satisfies

dXt = α(t, Xt ) dt + β(t, Xt ) dWt (52)

16 For further details on affine models see James and Webber who devote an entire chapter to the subject.
17 Additional restrictions need to be placed on the parameters to ensure that rt remains non-negative.
Continuous-Time Short Rate Models 13

where Wt is an n-dimensional Q-Brownian motion. Then the time t price, Ct , of a security that has a dividend
rate function, h(Xt , t), and terminal payment function, g(XT ), is given by
"Z R R #
T s T
− ru du − ru du
Ct = E Q
t e t h(Xs , s) ds + e t g(XT ) .
t

The associated Feynman-Kac PDE is then

· ¸
∂C ∂C 1 ∂2C
+ µ(t, x) + tr σ(t, x)σ(t, x)T − r(x, t)C + h(x, t) = 0, (x, t) ∈ R × [0, T )
∂t ∂x 2 ∂x2
C(x, T ) = g(x), x∈R
∂2C ∂C
where ∂x2 is the matrix of second derivatives and ∂x is the vector of first derivatives.

In order to hedge a derivative security in a model driven by n Brownian motions it is necessary, in general, to use
n + 1 hedging securities. The hedging securities are often taken to include the cash account and n zero-coupon
bonds of different maturities. The replicating portfolio may be constructed in a manner analogous to that of
Section 5.

7 Strengths and Weaknesses of Short Rate Models

Short-rate models (including lattice models of the short rate) have a number of strengths and weaknesses:
Strengths

• The models are generally tractable and very amenable to numerical and Monte-Carlo simulation methods.
• Derivatives prices can be computed quickly. This is very important for risk-management purposes when
many securities need to be priced frequently.
• They are parsimonious and can provide “sanity checks” on more sophisticated models that can often be
calibrated to “fit everything”. Models that are calibrated to “fit everything” can be unreliable due to the
problems associated with over-fitting. Of course short rate models with deterministic time-varying
parameters (e.g. Ho-Lee, Hull and White) are also susceptible to over-fitting.

Weaknesses

• The one-factor short-rate models imply that movements in the entire term-structure can be hedged with
only two securities. Equivalently, instantaneous returns on zero-coupon bonds of different maturities are
perfectly correlated in single-factor models. Neither of these features is realistic but these problems can be
overcome by using multi-factor models. In fact models with just 2 or 3 factors can afford considerably
more modeling flexibility. Moreover, they generally retain their numerical tractability. Models with 3 or
more factors, however, tend to suffer from the curse-of-dimensionality in which case Monte-Carlo
simulation becomes18 the only practical pricing technique.
• They are not as ”close to reality” as the LIBOR market models. The latter class of models directly model
observable market quantities, i.e., LIBOR rates, and this feature makes these models relatively
straightforward to calibrate. Moreover, many securities of interest can easily be priced in these models by
making appropriate distributional assumptions about the evolution of particular LIBOR rates.
In practice, short-rate models are often used as a complement and ”sanity check” for more sophisticated models.
Moreover, their tractability means that they will continue to be used for risk-management purposes.
18 Unless analytic solutions are available.