5.

The Heath-Jarrow-Morton (HJM)

Framework

I decided I’d spent too much time philosophizing.

It is, unfortunately, one of my character flaws.
J’onn J’onnz in “Martian Manhunter annual” 2, 1999, DC Comics.
Modeling the interest-rate evolution through the instantaneous short rate
has some advantages, mostly the large liberty one has in choosing the related
dynamics. For example, for one-factor short-rate models one is free to choose
the drift and instantaneous volatility coefficient in the related diffusion dy-
namics as one deems fit, with no general restrictions. We have seen several
examples of possible choices in Chapter 3. However, short-rate models have
also some clear drawbacks. For example, an exact calibration to the initial
curve of discount factors and a clear understanding of the covariance struc-
ture of forward rates are both difficult to achieve, especially for models that
are not analytically tractable.
The first historically important alternative to short-rate models has been
proposed by Ho and Lee (1986), who modeled the evolution of the entire yield
curve in a binomial-tree setting. Their basic intuition was then translated in
continuous time by Heath, Jarrow and Morton (HJM) (1992) who developed a
quite general framework for the modeling of interest-rate dynamics. Precisely,
by choosing the instantaneous forward rates as fundamental quantities to
model, they derived an arbitrage-free framework for the stochastic evolution
of the entire yield curve, where the forward-rates dynamics are fully specified
through their instantaneous volatility structures. This is a major difference
with arbitrage free one-factor short-rate dynamics, where the volatility of
the short rate alone does not suffice to characterize the relevant interest-rate
model. But in order to clarify the matter, let us consider the Merton (1973)
toy short-rate model.
Assume we take the following equation for the short rate under the risk-
neutral measure:
drt = θdt + σdWt , r0 .
If you wish, this is a very particular toy version of the Hull-White model (3.33)
seen in Chapter 3 with constant coefficient θ. Now, for this (affine) model,
one can easily compute the bond price,
2

σ θ
P (t, T ) = exp (T − t)3 − (T − t)2 − (T − t)rt ,
6 2
184 5. The Heath-Jarrow-Morton (HJM) Framework

and the instantaneous forward rate

∂ ln P (t, T ) σ2
f (t, T ) = − = − (T − t)2 + θ(T − t) + rt .
∂T 2
Differentiate this and substitute the short-rate dynamics to obtain

df (t, T ) = (σ 2 (T − t) − θ)dt + θdt + σdWt ,

or
df (t, T ) = σ 2 (T − t)dt + σdWt .
Look at this last equation. The drift σ 2 (T − t) in the f ’s dynamics is deter-
mined as a suitable transformation of the diffusion coefficient σ in the same
dynamics. This is no mere coincidence due to the simplicity of our toy model,
but a general fact that Heath, Jarrow and Morton (1992) expressed in full
generality.
Clearly, if one wishes to directly model this instantaneous forward rate,
there is no liberty in selecting the drift of its process, as it is completely
determined by the chosen volatility coefficient. This is essentially due to the
fact that we are modeling a derived quantity f and not the fundamental
quantity r. Indeed, f is expressed in terms of the more fundamental r by
"   #
T
∂ ln Et exp − t r(s)ds
f (t, T ) = − ,
∂T
and, as you see, an expectation has already acted in the definition of f , adding
structure and taking away freedom, so to say.
More generally, under the HJM framework, one assumes that, for each T ,
the forward rate f (t, T ) evolves according to

df (t, T ) = α(t, T )dt + σ(t, T )dW (t),

where W is a (possibly multi-dimensional) Brownian motion. As we have just

seen in the above example, contrary to the short-rate modeling case, where
one is free to specify the drift of the considered diffusion, here the function α
is completely determined by the choice of the (vector) diffusion coefficient σ.
The importance of the HJM theory lies in the fact that virtually any
(exogenous term-structure) interest-rate model can be derived within such
a framework.1 However, only a restricted class of volatilities is known to
imply a Markovian short-rate process. This means that, in general, burden-
some procedures, like those based on non-recombining lattices, are needed to
price interest-rate derivatives. Substantially, the problem remains of defining
1
Even the celebrated LIBOR market model was developed starting from
instantaneous-forward-rate dynamics in Brace, Gatarek and Musiela (1997), al-
though it is possible to obtain it also through the change-of-numeraire approach,
as we will see in the next chapter.
 5.1 The HJM Forward-Rate Dynamics 185

a suitable volatility function for practical purposes. This is the reason why in
this book we do not devote too much attention to the HJM theory, preferring
to deal with explicitly formulated models.
In this chapter, we briefly review the HJM framework and explicitly
write the HJM no-arbitrage condition. We then describe some analogies
with instantaneous short-rate models. We show, in particular, that a one-
factor HJM model with deterministic volatility is equivalent to the Hull
and White (1990b) short-rate model (3.32). We then introduce the Ritchken
and Sankarasubramanian (1995) (RS) framework and briefly illustrate the
Li, Ritchken and Sankarasubramanian (1995a, 1995b) algorithm for pricing
derivatives. We finally mention the Mercurio and Moraleda (2000) humped-
volatility model as a specific example of a one-factor Gaussian model within
the HJM framework.

5.1 The HJM Forward-Rate Dynamics

Heath, Jarrow and Morton (1992) assumed that, for a fixed a maturity T , the
instantaneous forward rate f (t, T ) evolves, under a given measure, according
the following diffusion process:

df (t, T ) = α(t, T )dt + σ(t, T )dW (t),

(5.1)
f (0, T ) = f M (0, T ),

with T → f M (0, T ) the market instantaneous-forward curve at time t =

0, and where W = (W1 , . . . , WN ) is an N -dimensional Brownian motion,
σ(t, T ) = (σ1 (t, T ), . . . , σN (t, T )) is a vector of adapted processes and α(t, T )
is itself an adapted process. The product σ(t, T )dW (t) is intended to be the
scalar product between the two vectors σ(t, T ) and dW (t).
The advantage of modeling forward rates as in (5.1) is that the current
term structure of rates is, by construction, an input of the selected model.
Remember, in fact, that the following relations between zero-bond prices and
forward rates hold (see (1.23)):

∂ ln P (t, T )
f (t, T ) = − ,
∂T
T
P (t, T ) = e− t f (t,u)du .

The dynamics in (5.1) is not necessarily arbitrage-free. Following an approach

similar to that of Harrison and Pliska (1981) (see Chapter 2) Heath, Jarrow
and Morton proved that, in order for a unique equivalent martingale mea-
sure to exist, the function α cannot be arbitrarily chosen, but it must equal a
quantity depending on the vector volatility σ and on the drift rates in the dy-
namics of N selected zero-coupon bond prices. In particular, if the dynamics
(5.1) are under the risk-neutral measure, then we must have
186 5. The Heath-Jarrow-Morton (HJM) Framework
 T 
N  T
α(t, T ) = σ(t, T ) σ(t, s)ds = σi (t, T ) σi (t, s)ds, (5.2)
t i=1 t

so that the integrated dynamics of f (t, T ) under the risk-neutral measure are
 t  T  t
f (t, T ) = f (0, T ) + σ(u, T ) σ(u, s)ds du + σ(s, T )dW (s)
0 u 0
N 
 t  T N 
 t
= f (0, T ) + σi (u, T ) σi (u, s)ds du + σi (s, T )dWi (s)
i=1 0 u i=1 0

and are fully specified once the vector volatility function σ is provided. Given
this dynamics of the instantaneous forward rate f (t, T ), application of Ito’s
lemma gives the following dynamics of the zero-coupon bond price P (t, T ):
   
T
dP (t, T ) = P (t, T ) r(t)dt − σ(t, s)ds dW (t) ,
t

where r(t) is the instantaneous short term interest rate at time t, that is
 t  t  t
r(t) = f (t, t) = f (0, t) + σ(u, t) σ(u, s)ds du + σ(s, t)dW (s)
0 u 0
N 
 t  t N 
 t
= f (0, t) + σi (u, t) σi (u, s)ds du + σi (s, t)dWi (s).
i=1 0 u i=1 0
(5.3)

5.2 Markovianity of the Short-Rate Process

The short-rate process (5.3) is not a Markov process in general. Notice in
fact that the time t appears in the stochastic integral both as extreme of
integration and inside the integrand function. However, there are suitable
specifications of σ for which r is indeed a Markov process. As proven by
Carverhill (1994), this happens, for example, if we can write, for each i =
1, . . . , N ,
σi (t, T ) = ξi (t)ψi (T ), (5.4)
with ξi and ψi strictly positive and deterministic functions of time. Under
such a separable specification, the short-rate process becomes
N 
 t  t N 
 t
r(t) = f (0, t) + ξi (u)ψi (t) ξi (u)ψi (s)ds du + ξi (s)ψi (t)dWi (s)
i=1 0 u i=1 0


N  t  t 
N  t
= f (0, t) + ψi (t) ξi2 (u) ψi (s)ds du + ψi (t) ξi (s)dWi (s).
i=1 0 u i=1 0
 5.3 The Ritchken and Sankarasubramanian Framework 187

Notice that in the one-factor case (N = 1), if we define the (strictly-positive)

deterministic function A by
 t  t
2
A(t) := f (0, t) + ψ1 (t) ξ1 (u) ψ1 (s)ds du,
0 u

and assume its differentiability, we can write

 t
dr(t) = A (t)dt + ψ1 (t) ξ1 (s)dW1 (s) + ψ1 (t)ξ1 (t)dW1 (t)
0

r(t) − A(t)
= A (t) + ψ1 (t) dt + ψ1 (t)ξ1 (t)dW1 (t)
ψ1 (t)
= [a(t) + b(t)r(t)] dt + c(t)dW1 (t),
with obvious definition of the coefficients a, b and c.
We therefore end up with the general short-rate dynamics proposed by
Hull and White (1990b), see (3.32), thus establishing an equivalence between
the HJM one-factor model for which (5.4) holds and the general formula-
tion of the Gaussian one-factor short-rate model of Hull and White (1990b).
In particular, we can easily derive the HJM forward-rate dynamics that is
equivalent to the short-rate dynamics (3.33). To this end, let us set
σ1 (t, T ) = σe−a(T −t) ,
where a and σ are now real constants, so that
ξ1 (t) = σeat ,
ψ1 (T ) = e−aT ,
σ2  2
A(t) = f (0, t) + 1 − e−at .
2a2
The resulting short-rate dynamics is then given by

∂f σ 2  −at 
dr(t) = (0, t) + e − e−2at
∂T a



σ2  
−at 2
−a r(t) − f (0, t) − 2 1 − e dt + σdW1 (t)
2a

∂f σ2  −2at

= (0, t) + af (0, t) + 1−e − ar(t) dt + σdW1 (t),
∂T 2a
which is equivalent to (3.33) when combined with (3.34).

5.3 The Ritchken and Sankarasubramanian Framework

It is now clear that an arbitrary specification of the forward-rate volatility
will likely lead to a non-Markovian instantaneous short-rate process. In such
188 5. The Heath-Jarrow-Morton (HJM) Framework

a case, we would soon encounter major computational problems when dis-

cretizing the dynamics (5.3) for the pricing of a general derivative. In fact, the
approximating lattice will not be recombining, and the number of nodes in
the tree will grow exponentially with the number of steps. This will make the
numerical procedure quite difficult to handle, especially as far as execution
time (combined with a pricing accuracy) is concerned.
These pricing problems can be addressed by noting that, even though the
short-rate process is not Markovian, there may yet exist a higher-dimensional
Markov process having the short rate as one of its components. Exploiting
such intuition, Ritchken and Sankarasubramanian (1995) have identified nec-
essary and sufficient conditions on the volatility structure of forward rates
for capturing the path dependence of r through a single sufficient statistic.
Precisely, they proved the following.
Proposition 5.3.1. Consider a one-factor HJM model. If the volatility func-
tion σ(t, T ) is differentiable with respect to T , a necessary and sufficient con-
dition for the price of any (interest-rate) derivative to be completely deter-
mined by a two-state Markov process χ(·) = (r(·), φ(·)) is that the following
condition holds:
T
σ(t, T ) = σRS (t, T ) := η(t)e− t
κ(x)dx
, (5.5)

where η is an adapted process and κ is a deterministic (integrable) function.

In such a case, the second component of the process χ is defined by
 t
2
φ(t) = σRS (s, t)ds.
0

Accordingly, zero-coupon-bond prices are explicitly given by

 
P (0, T ) 1 2
P (t, T ) = exp − Λ (t, T )φ(t) + Λ(t, T )[f (0, t) − r(t)] ,
P (0, t) 2
where  T
u
Λ(t, T ) = e− t
κ(x)dx
du.
t

Differentiation of equation (5.3) shows that, under the RS class of volatil-

ities (5.5), the process χ, and hence the instantaneous short-rate r, evolve
according to



dr(t) µ(r,
 t)dt + η(t)dW(t)
dχ(t) = = (5.6)
dφ(t) η 2 (t) − 2κ(t)φ(t) dt
with
∂
µ(r, t) = κ(t)[f (0, t) − r(t)] + φ(t) +
f (0, t).
∂t
We can now see that η is nothing but the instantaneous short-rate volatility
process.
 5.3 The Ritchken and Sankarasubramanian Framework 189

The yield curve dynamics described by (5.6) can be, therefore, discretized
in a Markovian (recombining) lattice in terms of the two variables r and φ.
This was suggested by Li, Ritchken and Sankarasubramanian (1995a, 1995b)
(LRS), who developed an efficient lattice to approximate the processes (5.6).
Their tree construction procedure is briefly outlined in the following, under
the particular case where

η(t) = σ̂(r(t))
σ̂(x) := vxρ

with v and ρ positive constants, ρ ∈ [0, 1], so that

T
σRS (t, T ) = v[r(t)]ρ e− t
κ(x)dx
.

Li, Ritchken and Sankarasubramanian considered the following transforma-

tion, which yields a process with constant volatility
 
1 
Y (t) = dx
σ̂(x) x=r(t)

where the right-hand side denotes a primitive of 1/σ̂(x) calculated in x = r(t)

and where the constant in the primitive is set to zero. This is exactly the
transformation of r needed to have a unit diffusion coefficient when applying
Ito’s formula to compute its differential, as we shall see in a moment. Notice
that, by substituting the expression for σ̂ and by integrating, we have

1
ln(r(t)) if ρ = 1
Y (t) = Ȳ (r(t)) := v 1 (5.7)
v(1−ρ) [r(t)]
1−ρ
if 0 ≤ ρ < 1

with the function Y defined in a suitable domain Dρ depending on the value

of ρ. Denoting by x = ϕ(y) the inverse function of y = Ȳ (x) on Dρ , we have

evy if ρ = 1
ϕ(y) = (5.8)
(v(1 − ρ)y) 1/(1−ρ)
if 0 ≤ ρ < 1

Application of Ito’s lemma and straightforward algebra show that

dY (t) = m(Y, φ, t)dt + dW (t)

dφ(t) = [σ̂(r(t)) − 2κ(t)φ(t)] dt

where indeed the diffusion coefficient in the Y dynamics is one, and where

κ(t)[f (0, t) − ϕ(Y (t))] + φ(t) + ∂f

∂t (0, t) vρ
m(Y, φ, t) = − (5.9)
v[ϕ(Y (t))]ρ 2[ϕ(Y (t))]1−ρ

for any 0 ≤ ρ ≤ 1.
190 5. The Heath-Jarrow-Morton (HJM) Framework

If, for instance, ρ = 0.5, we have that, for r(t) > 0, Y (t) = 2 r(t)/v,
2 2
ϕ(y) = v 4y , and m(Y, φ, t) in (5.9) becomes

κ(t)[f (0, t) − 14 v 2 (Y (t))2 ] + φ(t) + ∂f

∂t (0, t) 1
m(Y, φ, t) = 1 2 − ,
2 v Y (t)
2Y (t)
whereas, if ρ = 1, we have that, for r(t) > 0, Y (t) = ln[r(t)]/v, ϕ(y) = evy
and
κ(t)[f (0, t) − evY (t) ] + φ(t) + ∂f
∂t (0, t) 1
m(Y, φ, t) = − v.
vevY (t) 2
Building a lattice for Y is eased by the presence of a unit diffusion coefficient,
which is the reason for adopting such a transformation in the first place. Also,
since φ’s dynamics (5.6) has no diffusion part, the related lattice component
need not branch. The approximating lattice is then constructed as follows.
Divide the given time horizon into intervals of equal length ∆t, and suppose
that, at the beginning t of some time interval, the state variables are y and φ,
and that, in the next time period, they move either to (y + , φ∗ ) or to (y − , φ∗ )
where
√
y + = y + (J(y, φ) + 1) ∆t,
√
y − = y + (J(y, φ) − 1) ∆t,
φ∗ = φ + [σ̂ 2 (ϕ(y)) − 2κ(t)φ]∆t
" √ #
Setting Z(y, φ) = int m(y, φ, t) ∆t , where int[x] denotes the largest integer
smaller or equal than the real x, the function J is defined by

|Z(y, φ)| if Z(y, φ) is even
J(y, φ) =
Z(y, φ) + 1 otherwise.
The branching probabilities, p for an up-move and 1 − p for a down-move,
are then derived by solving
p(y + − y) + (1 − p)(y − − y) = m(y, φ, t)∆t,
thus obtaining
m(y, φ, t)∆t + y − y −
p= .
y+ − y−
The above choices ensure that
y + ≥ y + m(y, φ, t)∆t ≥ y − ,
and hence that the probabilities of moving from one node to another in the
lattice lie always in [0, 1].2
Once the LRS tree has been built, derivatives prices can then be calculated
in a quite standard way.
2
We refer to Li, Ritchken and Sankarasubramanian (1995a) for a detailed descrip-
tion of all the calculations above.
 5.4 The Mercurio and Moraleda Model 191

5.4 The Mercurio and Moraleda Model

We conclude the chapter by briefly reviewing the Mercurio and Moraleda

(2000) model, which explicitly assumes a humped volatility structure in the
instantaneous-forward-rate dynamics. Such assumption is motivated by the
fact that the volatility structure of forward rates, as implied by market quotes,
is commonly humped. We also refer to Section 3.6 for a detailed explanation
of the opportunity and practical relevance of such an assumption.
Mercurio and Moraleda (2000) proposed a one-factor Gaussian model in
the HJM framework by considering the following form for the volatility of
instantaneous forward rates:

σ(t, T ) = σ[γ(T − t) + 1]e− 2 (T −t) ,

λ
(5.10)

where σ, γ and λ are non-negative constants. Under such specification, (5.1)

becomes

df (t, T ) = ᾱ(T − t)dt + σ[γ(T − t) + 1]e− 2 (T −t) dW (t),

λ

 

2σ 2 2γ
[γτ + 1] e−λτ γτ +
λ
ᾱ(τ ) = − + 1 1 − e2τ ,
λ λ

which implies that instantaneous (forward and spot) rates are normally dis-
tributed.
The choice of the volatility function (5.10) is motivated by the following
features.
(a) It provides a humped volatility structure for strictly positive σ, γ and λ,
and 2γ > λ;
(b) It depends only on the “time to maturity” T − t rather than on time t
and maturity T separately;
(c) It leads to analytical formulas for European options on discount bonds;
(d) It generalizes the volatility specification of the Hull and White model
(3.33), in that for γ = 0 and λ = 2a we get the volatility of forward rates
as implied by (3.33).3
It is obvious that (b) and (d) hold. Basic calculus shows that (a) is also
true. In fact, for strictly positive σ, γ and λ, the function f (x) = σ[γx +
1] exp{− λ2 x} has the following features: (i) it is strictly positive for x ≥ 0;
(ii) it is increasing and concave in the interval [0, (2γ − λ)/(γλ)]; (iii) it has a
maximum in (2γ − λ)/(γλ) whose value is (2σγ/λ) exp{(−2γ + λ)/(2γ)}; (iv)
it is decreasing and concave in the interval [(2γ − λ)/(γλ), (4γ − λ)/(γλ)];
3
When modeling humped volatility structures, many other specifications can of
course be considered. For example the term between square brackets in (5.10)
can be generalized to be any polynomial in (T − t). It is disputable however
whether there exists a simpler characterization than (5.10) and for which (a),
(b), (c) and (d) hold.
192 5. The Heath-Jarrow-Morton (HJM) Framework

(v) it is decreasing and convex from (4γ − λ)/(γλ) onwards; (vi) it tends
asymptotically to zero.
As for property (c), Mercurio and Moraleda (2000) used the results of
Merton (1973) to prove that the time t-price of a European call option with
maturity T and strike price X on a pure discount bond with maturity S is
given by

ZBC(t, T, S, X) = P (t, S)Φ(d1 (t)) − XP (t, T )Φ(d2 (t)),

where
ln(P (t, S)/(XP (t, T )) + 12 vt2
d1 (t) := ,
vt
d2 (t) := d1 (t) − vt ,

and
4σ 2 2 2  λT  8σ 2 B  
vt2 = 7
(A λ + 2ABλ + 2B 2
) e − eλt
− 6
(Aλ + B) T eλT − teλt
λ λ
4σ 2 B 2  2 λT 
+ T e −t e ,
2 λt
λ5
where
 λ   
A := (λ + 2γ) e− 2 S − e− 2 T + γλ Se− 2 S − T e− 2 T ,
λ λ λ

 λ 
B := γλ e− 2 S − e− 2 T .
λ

However, no analytical formula for pure discount bonds can be derived. No-
tice, in fact, that the instantaneous-short-rate process is not Markovian since
(5.10) does not belong to the RS class (5.5).
Mercurio and Moraleda also tested empirically their model. They consid-
ered a time series of market cap prices and compared the fitting quality of
their model with that implied by the Hull and White model (3.33). Using the
Schwarz-Information-Criterion test, they concluded that, in most situations,
their humped-volatility model is indeed preferable. We refer to Mercurio and
Moraleda (2000) for a detailed description of the calibration results.