Interest Rate Models

key developments in the

Mathematical Theory of Interest Rate Risk Management

presented by

Lane P. Hughston
Professor of Financial Mathematics
Department of Mathematics, King’s College London
The Strand, London WC2R 2LS, UK
lane.hughston@kcl.ac.uk
www.mth.kcl.ac.uk
and
Dorje C. Brody
Royal Society University Research Fellow
Theory Group, Blackett Laboratory,
Imperial College, London SW7 2BZ, UK
dorje@imperial.ac.uk
http://theory.ic.ac.uk/˜brody

1
Chapter 1
Discount bonds and interest rates. Libor and swap rates. Forward prices and
forward rates. Short rate and forward short rate. Positive interest conditions.
Interest rate derivative structures.

1.1 Discount bonds and interest rates

The formulae involved with interest rate modelling can get complicated. It is important
to use an unambiguous scheme of notation that can be carried across a range of different
models and at the same time is useful for calculations.

Time 0 denotes the present. Times a, b, c, etc., denote various future times, as do s, t,
u, and so on. Alphabetical order will often be used to suggest chronological order. Occa-
sionally, we use an upper case T to draw attention to a particular date (e.g. a termination
date).

We use the notation Pab to denote the value at time a of a discount bond maturing at
time b. At time b, the bond pays one unit of “currency”. We fix a currency throughout here.

In fact, for any class of financial assets we have a corresponding system of discount bonds.
Thus, for dollars, Pab denotes the price at time a, in dollars, of a bond that pays one dollar
at the maturity b.

Equally, we can speak of a “sterling” discount bond, or even a “gold” discount bond. In the
latter case, Pab could denote the price at time a, in ounces of gold, of a contract delivering
one ounce of gold at time b.

Occasionally, a comma will be inserted for clarity. Thus Pt,x+t denotes the value of a discount
bond at time t that matures at time x + t.

For any fixed value of t, the system of discount bond prices PtT for T ∈ [t, ∞) is called
the discount-function at that time. The present discount function is P0T .

2
Associated with any discount bond Pab there are various rates that can be quoted.

For example, the simple interest rate Lab is defined by:

1
Pab = . (1.1)
1 + (b − a)Lab

The continuously compounded rate Rab is defined by:

Pab = e−(b−a)Rab . (1.2)

The unit of time is one calendar year, and these rates are quoted in an “annualised” basis.

Inverting these relations we find that the simple rate is given by


1 1
Lab = −1 (1.3)
b − a Pab

The corresponding expression for the continuously compounded rate is

1
Rab = − log Pab . (1.4)
b−a

1.2 Libor and Swap rates

The Libor rate for a given period is usually quoted on a simple annualised basis, so some-
times we call Lab the Libor rate associated with Pab .

Note that although rates can be quoted in various ways, the discount bond price is unique (it
is a price!). That is a good reason for focusing on discount bonds. These are the fundamental
“assets” of interest rate theory, and it is their behaviour we are trying to model.

Another very important type of rate frequently quoted in the over-the-counter interest rate
markets is the swap rate.

There are various types of swap rates, and various conventions dealing with day counts,
and so on. It is best therefore to give a mathematically concise definition that can be
adapted easily to various situations.

The swap rates defined in this way are “pure” in the sense that they are based on the
basic discount function, and do not take into account credit, liquidity, and other market
factors that may affect “real” swap rates.

3
Let 0 denote the present, t some date in the future, and T1 , T2 , . . . , Tn a series of future
dates beyond t.

For each such series (T1 , T2 , . . . , Tn ) there is a unique swap rate st .

This rate is determined by the condition that if the rate of interest st is paid on a unit
principal on each of the dates T1 , T2 , . . . , Tn and if the unit principal is paid at time Tn , then
the present value at time t of this cash flow is unity.

More specifically, we have the condition

st (PtT1 + PtT2 + · · · + PtTn ) + PtTn = 1. (1.5)

Solving for st we have

1 − PtTn
st = (1.6)
PtT1 + PtT2 + · · · + PtTn

The sum

n
VtT1 ...Tn = PtTi (1.7)
i=1

is sometimes called the ‘basis point value’ (bpv) at time t associated with the date system
T1 , T2 , . . ., Tn .

We note that because st can always be expressed as a combination of various discount

bond values, it makes sense to speak of derivative payoffs based on st .

A derivative whose payoff depends on st can thus be viewed as a kind of exotic option
based on the discount bonds.

There are elements of convention involved in how real swap rates are quoted. For exam-
ple, if st is paid semi-annually (i.e. T1 , T2 , etc., are spaced at half-yearly intervals), then 2st
is the quoted swap rate. This is an artifact of market convention and need not concern us
here, but of course it should be born in mind.

4
1.3 Forward prices and forward rates
The forward price of a discount bond will be denoted by Ptab .

This is the price contracted at time t for purchase of a discount bond at time a that matures
at time b.

A standard arbitrage argument shows that

Ptb
Ptab = . (1.8)
Pta
The argument runs as follows.

Suppose at time t a ‘careless’ market maker is willing to sell me a b-maturity bond on a

forward basis at time a for a price Qtab that is less than Ptab .

I would then purchase Qtab /Pta a-maturity bonds at time t, and simultaneously short Qtab /Ptb
b-maturity bonds.

At the same time I purchase 1/Pta b-maturity bonds on a forward basis from the dealer.

At time a, the a-maturity bonds mature, leaving me with Qtab /Pta in cash, which I uses
to purchase 1/Pta b-maturity bonds (taking advantage of the forward agreement).

Then at time b, the long investment pays off 1/Pta , whereas I owe Qtab /Ptb on the ma-
turing short position.

Since 1/Pta > Qtab /Ptb , I have made a risk free profit.

A similar argument allows me to arbitrage the dealer if a forward price greater than Ptab is
made.

Thus we see that Ptab = Ptb /Pta is the correct forward price for a discount bond.

The associated forward rates are given by

1
Ptab = (1.9)
1 + (b − a)Ltab

and

Ptab = e−(b−a)Rtab . (1.10)

5
Here Ltab and Rtab are the forward rates, quoted at time t, for the period [a, b], on a simple
and on a continuously compounded basis, respectively.

We call Ltab the forward Libor rate made at time t for the period [a, b].

It also makes sense to speak of a forward swap rate.

This is the swap rate sta contracted at time t for a swap entered into at time a with the
payment dates b1 , b2 , . . . , bn . Then we have
1 − Ptabn
sta = . (1.11)
Ptab1 + Ptab2 + · · · + Ptabn
Clearly we have stt = st .

1.4 Short rates and forward short rates.

The rate rb = lima→b Lab is called the short rate.

This is the rate of interest, at time a, on a very short period loan (e.g., “overnight”),
expressed on an annualised basis.

If we assume, as seems reasonable, that Pab is differentiable in the maturity date, then
a short computation shows that

∂Pab 
ra = − . (1.12)
∂b a=b
Over the short term, “compounding” is irrelevant, and thus
lim Lab = lim Rab . (1.13)
a→b a→b

The forward short rate fta is the rate of interest contracted at time t for a very short period
loan at some later time a.

For example, I might agree today to loan you $1,000,000 for one day, one year from now, at
a rate of interest of 6% annualised. Then we would have f01 = 0.06 (a = 0, b = 1).

The forward short rate is also called the “instantaneous forward rate” (for example, in
Heath, Jarrow & Morton 1992).

We note that the forward short rate is by definition given by the limit
fta = lim Ltba . (1.14)
b→a

6
Thus we have

∂Ptab  ∂ ln Pta
fta = −  =− . (1.15)
∂b a=b ∂a

The latter relation is often effectively adopted as a definition for fta in the literature, but it
is important to see that it is not really a definition: it derives from an underlying economic
relation.

The significance of the relation

∂ ln Pta
fta = − (1.16)
∂a
is that it is invertible:

 T 
PtT = exp − ftu du . (1.17)
t

Thus, at any fixed time t, knowledge of the discount function PtT at that time, for maturity
T , is equivalent to knowledge of the system of forward short rates ftu determined (i.e. con-
tractable) at that time over the interval u ∈ [0, T ].

Note, incidentally, that (1.17) incorporates the maturity condition PT T = 1.

1.5 Positive interest conditions

For many applications we want to build in an interest rate positivity condition.

This is not automatic in the HJM framework, but later when we examine the Flesaker-
Hughston framework and its extensions we will see how this feature can be incorporated.

For positive interest we require the following two conditions valid for all 0 ≤ a ≤ b < ∞:

0 < Pab ≤ 1, (1.18)

∂Pab
< 0. (1.19)
∂b
There are various ways of ensuring these conditions are satisfied. For many models they are
not. Whether or not this is a material issue depends on the circumstances.

From a fundamental point of view, however, we require nominal interest rates to be strictly

7
positive. This is because if someone offers to loan you money at a negative rate of interest,
then you can immediately take advantage of them and effect an arbitrage.

The positive interest conditions are sufficient to ensure that all the commonly encountered
rates are positive: Libor rates, swap rates, forward Libor and swap rates, short rate, and
forward short rate.

1.6 Interest rate derivative structures

Let us now turn to the consideration of interest-rate related contingent claims.

First, we need to ask what is meant by an “interest rate derivative”.

One general mathematical way of defining a European-style interest rate derivative is to

say that the payout at time T is any random variable HT that is FT -measurable, where (Ft )
is the natural filtration of the multi-dimensional Brownian motion driving the discount-bond
system.

In practice, the payout of an interest rate derivative is specified in terms of one or more
well-defined rates associated with the given contract period.

Equivalently, we let HT be specified as a function of the values of one or more discount

bonds during the interval [0, T ]. The maturities of these discount bonds may or may not lie
in that interval.

For example, the payout

(a) HT = max (PT b − K, 0) (1.20)

defines a call option on a discount bond (b > T ).

The payout

(b) HT = X max (LT b − R, 0) (1.21)

defines a simple caplet on the Libor rate LT b , where R is the cap rate, and X is the notional
paid per interest rate point (e.g., $1,000,000 per interest rate point above R).

Normally, a caplet is paid “in arrears”, meaning the rate is set at some earlier time a,
and paid at T , so in that case, the payout is

(c) HT = X max (LaT − R, 0) , (1.22)

8
for the rate LaT set earlier at time a.

However, since LaT is known at time a, we can regard the normal caplet as a derivative
that pays the discounted value Ha = PaT HT at the earlier time a, where HT is the payout
defined in (c).

By definition, we have
1
PaT = . (1.23)
1 + (T − a)LaT

It follows, as we noted earlier, that


1 1
LaT = −1 . (1.24)
T −a PaT

Therefore, the effective payout Ha at time a is given by the following calculation:

Ha = PaT HT
= XPaT max (LaT − R, 0)


 
1 1
= XPaT max − 1 − R, 0
T − a PaT


1
= X max (1 − PaT ) − RPaT , 0
T −a
X
= max (1 − PaT − (T − a)RPaT , 0)
T −a

X 1
= [1 + R(T − a)] max − PaT , 0
T −a 1 + R(T − a)
= N max (K − PaT , 0) . (1.25)

Here the strike K is given by

1
K= (1.26)
1 + R(T − a)

and the notional N is

X[1 + R(T − a)]
N= . (1.27)
T −a
Thus we see that a position in standard caplet is equivalent to a position in N puts on the
discount bond, where the strike price K on the put is the value of a discount bond with
simple yield R.

9
There are many subtle ways of transforming one type of interest rate derivative structure
into another with the same effective payoff.

This is important both in the marketing and the risk management of such products.

As another example, suppose we consider the case of a swaption, the option to enter into a
swap at time t for the dates (T1 , T2 , · · · , Tn ) at a fixed “strike” swap-rate R.

Assuming that the option is to pay the fixed rate R, then the payoff Ht at time t is

Ht = VtT1 ...Tn Max(st − R, 0). (1.28)

n
Here VtT1 ...Tn = i=1 PtTi is the bpv at time t for the coupon dates (T1 , T2 , · · · , Tn ).

Clearly, the option is exercised iff the “actual” swap rate st observed at time t is greater
than R.

Thus an alternative way of writing the swaption payout Ht is:

 +
n
Ht = 1 − PtTn − R PtTn . (1.29)
i=1

It should be evident that an alternative interpretation of a swaption is to regard it as an

option at time t to acquire (A) a portfolio consisting of a unit of cash and a short position
in a Tn -maturity bond, in exchange for (B) a portfolio consisting of R units each of the
Ti -maturity bonds for i = 1, 2, . . . , n.

This is the economic interpretation of a swaption in terms of the exchange of actual as-
sets.

The swaption considered above is an option to pay the fixed leg of a swap, and is thus
called a payer swaption. There is an analogous structure which is an option to receive the
fixed leg of a swap, called a receiver swaption.

10
Chapter 2
Dynamical equations for a non-dividend-paying asset. Money market account
and risk premium process. Martingales, supermartingales and submartin-
gales. Martingale relations for a single asset. Transformation to the risk
neutral measure. No-arbitrage relation for derivatives. Derivative pricing.
Girsanov transformation.

2.1 Dynamical equations for a non-dividend-paying as-

set
For a single asset with limited liability and price process St , the stochastic equation for the
dynamics of St is:
dSt
= µt dt + σt dWt . (2.1)
St
This equation is defined on a probability space Π = (Ω, F, P ) with filtration (Ft ), with
respect to which Wt is a standard Brownian motion.

We assume that µt (drift) and σt (volatility) are adapted to the filtration (Ft ).

Initially, we consider the simple situation where (Ft ) is generated by Wt . Later, when
other basic assets are brought into play, we let the filtration (Ft ) be larger.

We can think of Π as representing the economy, and (Ft ) as representing the market in-
formation flow up to time t.

For many purposes we can, without serious loss of generality, assume that µt and σt are
bounded .

This will be a sufficient technical condition to ensure that the relevant stochastic integrals

11
exist, and the relevant martingale condition is satisfied when this is needed. In practice this
condition can often be relaxed in various ways.

If µ and σ are constant the solution of St is:

St = S0 exp µt + σWt − 12 σ 2 t . (2.2)

This is called the geometric Brownian motion model for St .

The geometric Brownian motion model was introduced by Paul Samuelson, and was used
by Fisher Black and Myron Scholes as an assumption in the derivation of their celebrated
option pricing formula.

More generally, for path dependent µt and σt , which for simplicity we may here assume
to be adapted and bounded, we have the following solution for the asset price in terms of µt
and σt :

 t  t  t 
1 2
St = S0 exp µs ds + σs dWs − 2 σs ds . (2.3)
0 0 0

We regard µt and σt as being specified exogenously.

We can use Ito’s lemma to verify that the stochastic equation is satisfied. First, we note
that
dSt 1 (dSt )2
d log St = − . (2.4)
St 2 St2
Thus squaring each side we have:
(dSt )2
(d log St )2 = . (2.5)
St2
So putting these two equations together we get:
dSt
= d log St + 12 (d log St )2 (2.6)
St
But taking the logarithm of (2.3) we have:
 t  t  t
log St = log S0 + µs ds + 1
σs dWs − 2 σs2 ds. (2.7)
0 0 0

So by taking the stochastic differential we obtain

d log St = µt dt + σt dWt − 12 σt2 dt. (2.8)

12
Thus by squaring and only keeping the (dWt )2 = dt term we also have:
(d log St )2 = σt2 dt. (2.9)
It follows immediately that
dSt
= µt dt + σt dWt . (2.10)
St

2.2 Money market account and risk premium process

To proceed further, we introduce a ‘risk-free’ asset, the money-market account, with price
process Bt , satisfying
dBt
= rt dt, (2.11)
Bt
Here rt is the short-term interest rate, which we also assume to be adapted to the market
filtration (Ft ).

The solution for the money market account process Bt is

 t 
Bt = B0 exp rs ds . (2.12)
0

Now we introduce the market risk premium process λt , defined for a non-dividend paying
asset by
µt = rt + λt σt . (2.13)
The process λt measures, instantaneously, the extra rate of return offered by the asset, above
the risk-free rate rt , per unit of volatility σt .

Note that in the case of a non-dividend paying asset, and in the absence of risk, the rate of
return would be rt .

In the case of a dividend paying asset, the process for µt is given by

µt = rt − δt + λt σt , (2.14)
where δt is the dividend rate.

In the case of a single asset the drift condition (2.13) merely defines λt .

In the case of multiple assets the relation gets generalised and is equivalent to the condition
of no arbitrage.

13
2.3 Martingales, supermartingales and submartingales
Now we derive an important relation that ties together the values of an asset at two different
times.

One of the central concepts in the modern theory of finance is the idea of a martingale.

The point of the martingale concept is that it gives a mathematical embodiment to the
notion of a fair game of chance.

It also helps to clarify in mathematical terms what we mean by a forecast.

In what follows we also need to know about the related concepts of supermartingale, and
submartingale.

The concept of supermartingale, in particular, plays a special role in interest rate theory.

A stochastic process M is an (Ft )-martingale if

(a) E [|Mt |] < ∞, for all t ≥ 0, (2.15)

(b) Ms = E [Mt | Fs ] , for all s < t. (2.16)

Part (b) of this definition expresses the idea that the expected value of the process at time
t, given information up to time s, is equal to the value of the process at time s.

When there is no ambiguity we sometimes write Et [X] = E[X|Ft ] for conditional expec-
tation with respect to the sigma-algebra Ft .

We can modify the definition above to account for martingales defined only for t ∈ [0, T ∗ ],
where T ∗ > 0 is a fixed time horizon.

A standard Brownian motion Wt is a martingale. So are, for example, the processes given
by
1
Mt = (Wt2 − t), (2.17)
2

1
Mt = (Wt3 − 3tWt ) (2.18)
6

1
Mt = (Wt4 − 6tWt2 + 3t2 ). (2.19)
24

14
Another example is given by

Mt = exp σWt − 12 σ 2 t , (2.20)

where σ is a constant.

To see that the process 12 (Wt2 − t) is a martingale, we observe that

Es [Wt2 − t] = Es [(Ws + (Wt − Ws ))2 − t]

= Es [Ws2 ] + Es [(Wt − Ws )2 ] − t
= Ws2 − s. (2.21)

More generally, let us define the polynomial H n (x, y) by the generating function

∞
exp ξx − 12 ξ 2 y = ξ n H n (x, y). (2.22)
n=0

Then for each value of n, the process H n (Wt , t) is a martingale, and the polynomial examples
mentioned above arise as the first few values of n.

The polynomials H n (x, y) are given by

1
n/2 
H n (x, y) = 2
y hn (x/ 2y), (2.23)

where hn (u) are the standard Hermite polynomials.

Martingales also arise as certain classes of stochastic integrals.

For example, if σt is Ft -adapted and bounded, then

 t
Mt = M 0 + σs dWs (2.24)
0

is a martingale.

So is:

 t  t 
Mt = M0 exp σs dWs − 1
2
σs2 ds . (2.25)
0 0

A process Xt is an (Ft )-supermartingale if


(c) E |Xt | < ∞, for all t ≥ 0, (2.26)
(d) Xs ≥ E [Xt | Fs ] , for all s < t. (2.27)

15
Similarly, a process Xt is an (Ft )-submartingale if

(e) E |Xt | < ∞, for all t ≥ 0, (2.28)
(f) Xs ≤ E [Xt | Fs ] , for all s < t. (2.29)

A process is a martingale iff it is both a supermartingale and a submartingale. If Xt is a

supermartingale, then −Xt is a submartingale.

Another important way of generating martingales is by taking conditional expectations.

Thus if Z is a random variable such that E[|Z|] < ∞, then

Mt = Et [Z] (2.30)

defines a martingale by virtue of the “tower property” of conditional expectation Es Et = Es

for s < t.

2.4 Martingale relations for a single asset

Returning to the case of a single asset, let us introduce the relationship µt = rt + λt σt into
the formula for St . We then have
dSt
= rt dt + σt (dWt + λt dt) . (2.31)
St
Equivalently, St is given by

 t 
 t  t 
1 2
St = S0 exp rs ds exp σs (dWs + λs ds) − 2 σ ds . (2.32)
0 0 0

It follows that

 t  t 
St 1 2
= S0 exp σs (dWs + λs ds) − 2
σ ds . (2.33)
Bt 0 0

Now suppose that we define the process Λt by

 t  t 
1 2
Λt = exp − λs dWs − 2 λs ds . (2.34)
0 0

We call Λt the risk adjustment density or risk premium density martingale. It follows from
Itô’s lemma that

dΛt = −Λt λt dWt . (2.35)

16
Equivalently, by integration of this relation, incorporating the initial condition, we have:
 t
Λt = 1 − Λs λs dWs . (2.36)
0

Thus, assuming λt is bounded, we have the martingale relation

Λs = Es Λt , for all s ≤ t, where Es Λt := E [Λt | Fs ] . (2.37)

Now we show the following important result:

Λt St
is a martingale. (2.38)
Bt
Indeed, a simple computation shows by completing the squares that:

 t  t 
Λt St 1 2
= exp (σs − λs ) dWs − 2 (σs − λs ) ds , (2.39)
Bt 0 0

and the desired property follows since σt is bounded. The martingale property for Λt St /Bt
can be written
 
Ss St
Λs = Es Λt , s < t. (2.40)
Bs Bt

This is the formula that links past and future values of St , and thus can be thought of as a
forecasting relation.

2.5 Transformation to the risk neutral measure

For any random variable Xt measurable with respect to the sigma-algebra Ft , we define a
new probability measure P λ with expectation

Es [Λt Xt ]
Eλs [Xt ] = . (2.41)
Λs
This formula explains why we call Λt a “density”.

The new probability measure (i.e. new rule for taking expectations) obtained in this way is
called the risk-neutral measure.

This terminology is reserved for the measure obtained by use of the density Λt associated
with the risk premium process λt .

17
Under the risk-neutral measure, we have
 
Ss λ St
= Es , s < t. (2.42)
Bs Bt
That is, the discounted asset price is a martingale (where the discounting is taken with re-
spect to the money market account).

Another way of putting this is that in the risk neutral measure the value of the asset is
a martingale when expressed in units of Bt , i.e., when we use Bt as a numeraire.

As we shall see, there are other measures associated with other choices of numeraire.

2.6 No-arbitrage relation for derivatives

Suppose that there is a derivative associated with St and its price process is Ht .

We assume that Ht is adapted to the filtration (Ft ) like St , and in particular that Ht is
fully characterised by an FT -measurable terminal value HT , i.e. its payoff.

This means intuitively that HT can depend in a very general way on the behaviour of
Wt (and hence St ) over the interval [0, T ].

Of course, HT might be relatively simple, like a call option HT = max (ST − K, 0) or a

short position in a forward contract HT = K − ST .

But it might be path-dependent, like a knock-out option, or an Asian option, or an American

option (exercisable at some random time τ ≤ T , with the proceeds future valued and paid
at time T ).

For the price dynamics of Ht let us write

dHt
= µH H
t dt + σt dWt . (2.43)
Ht
Then a well-known hedging argument can be used to establish that
µH
t − rt µ t − rt
H
= . (2.44)
σt σt
The hedging argument is as follows. Suppose we have a long position in the derivative, and
we wish to hedge that position with a short position in the underlying asset.

18
We form at time t the portfolio with value Ht − ∆t St where ∆t is the number of asset
units shorted.

We examine the dynamics of the portfolio over the next small interval of time. The change
in the value of the portfolio is given by dHt − ∆t dSt .

Then if
Ht σtH
∆t = , (2.45)
St σt
the “risks” (i.e. the coefficients of dWt ) cancel, and the portfolio offers an instantaneously
definite rate of return given by
Ht µH
t − ∆t St µt
. (2.46)
Ht − ∆t St
We equate this “hedged” rate of return to rt and insert the correct hedge ratio ∆t . Then
the desired no-arbitrage relation
µH
t − rt µ t − rt
H
= . (2.47)
σt σt
immediately pops out.

This relation is general, and is applicable in a fully path-dependent context.

2.7 Derivative pricing

We have assumed that (a) both the derivative and the asset price are adapted to the same
Brownian motion filtration, (b) there are no dividends, (c) there are no transaction costs,
(d) there are no constraints (e.g. limits) on the hedge position, and (e) the hedge portfolio
can be adjusted continuously.

Note that if we further assume Ht = H(St , t) for some function H(S, t) of two variables,
then the relation above becomes a PDE (the Black-Scholes equation) if µt , σt , rt and λt are
all likewise expressible as such functions.

This leads us down the “classical” path of derivative pricing, which can be highly effec-
tive when the assumptions indicated apply.

Generally, these assumptions break down if either (a) the derivative is path dependent or

19
(b) the asset price dynamics are path dependent.

The implication of the no-arbitrage condition (i.e. the general hedging argument) is that the
derivative price and the underlying asset both have the same risk premium λt .

As a consequence, defining Λt as before, it follows that Λt Ht /Bt is a martingale:

 
Λs Hs Λt Ht
= Es . (2.48)
Bs Bt
Equivalently, we have
 
Hs λ Ht
= Es , (2.49)
Bs Bt

where Eλ denotes expectation in the risk neutral measure. In particular, we have

 
ΛT HT
H0 = E . (2.50)
BT
Equivalently:
 
λ HT
H0 = E . (2.51)
BT
This is the risk-neutral valuation formula which says in words that the present value of a
derivative is equal to the risk-neutral expectation of its terminal payoff.

For example, if µ, r and σ are constant, and if HT is a simple call option payoff on ST ,
then this reduces to the Black-Scholes formula:

 

ln S0 erT /K + 12 σ 2 T −rT ln S0 erT /K − 12 σ 2 T
H 0 = S0 N √ −e KN √ (2.52)
σ T σ T

where
 x
1 1 2
N (x) = √ e− 2 ξ dξ (2.53)
2π −∞

is the standard normal distribution function.

2.8 Girsanov transformation∗

These results can be tied together nicely by the use of the Girsanov transformation.

20
We note that in the case of both the asset and the derivative, as a consequence of the
no-arbitrage condition, the term dWt + λt dt is common to the dynamics:
dSt
= rt dt + σt (dWt + λt dt) (2.54)
St
dHt
= rt dt + σtH (dWt + λt dt) (2.55)
Ht

Now we define a new process Wtλ according to the formula

 t
λ
Wt = Wt + λs ds. (2.56)
0

It follows that dWtλ = dWt + λt dt.

The essence of the theorem of Girsanov is that if Wt is a Brownian motion with respect
to P , then Wtλ is a Brownian motion with respect to P λ . Then we say that Wtλ is a P λ -
Brownian motion. The dynamics of St and Ht can be written
dSt
= rt dt + σt dWtλ , (2.57)
St
dHt
= rt dt + σtH dWtλ . (2.58)
Ht

In the risk neutral measure, Wtλ is a Brownian motion.

Thus we see that, as a consequence of the Girsanov transformation, the risk premium effec-
tively drops out of the dynamics for both the underlying asset as well as the derivative.

With respect to the risk neutral measure both St and Ht have a rate of return given by
rt , the rate of return offered on the locally risk-free money-market asset Bt .

A more precise account of Girsanov’s theorem is as follows.

Let (Ω, F, P ) be a probability space equipped with a filtration (Ft ). Suppose that Wt is
a n-dimensional (Ft )-Brownian motion defined on this probability space.

Let λαt be a n-dimensional, (Ft )-measurable process satisfying

 t 
2
P |λs | ds < ∞ = 1. (2.59)
0

21
Under these assumptions, the process Λt given by

  t 
1 t 2
Λt = exp − |λs | ds − λs · dWs (2.60)
2 0 0
is well defined for all t. We can verify that
 t
Λt = 1 − Λs λs · dWs , (2.61)
0
A sufficient condition for Λt to be a martingale is the Novikov condition:

 t 
1 2
E exp |λs | ds < ∞, (2.62)
2 0

in which case E ΛT = 1. This condition is satisfied, in particular, if λt is bounded.

If Λt is a martingale, then, given any fixed time T > 0, we can define a probability measure
QT on (Ω, FT ) by
QT (A) = E [ΛT 1A ] , for all A ∈ FT . (2.63)
The Girsanov theorem states that, given any fixed time T > 0, the process Wt∗ defined by
 t
∗
Wt = Wt + λs ds, t ∈ [0, T ] (2.64)
0
is a n-dimensional Brownian motion on (Ω, FT , QT ).

We can, for example, verify that Wt∗ is normally distributed with respect to the measure QT
by use of the method of characteristic functions.

Given any t ∈ [0, T ], we calculate the characteristic function of the random variable W̃t .
 ∗  ∗
EQT eizWt = EP ΛT eizWt
 ∗
= EP Λt eizWt

 t  t  t 
P 1 2
= E exp − λs dWs − 2 λs ds + izWt + iz λs ds
0 0 0

 t  t 
P 1 2 1 2
= E exp − (λs − iz) dWs − 2 (λs − iz) ds − 2 z t
0 0

 t  t 
P

= E exp − (λs − iz) dWs − 2 1

(λs − iz) ds exp − 12 z 2 t
2

1 2
0 0
= exp − 2 z t . (2.65)
This shows that the random variable Wt∗ is normally distributed, with mean 0 and variance t.

An elaboration of this argument leads to the result that Wt∗ is a QT -Brownian motion.

22
Chapter 3
Dynamical equations for multiple assets. Market completeness. Valuation
of derivatives in complete multi-asset market. Hedgeable and unhedgeable
claims in incomplete markets.

3.1 Dynamical equations for multiple assets

We model the economy by a probability space (Ω, F, P ) equipped with standard augmented
filtration {Ft } generated by a standard n-dimensional Brownian motion Wtα , α = 1, 2, · · · , n,
over the time interval 0 ≤ t ≤ T ∗ , for some terminal date T ∗ . For some applications we may
wish to take T ∗ = ∞.

According to the Ito calculus, we have dWtα dWtβ = δ αβ dt, where δ αβ is the identity ma-
trix. Note that the different components of Wtα are taken to be uncorrelated.

Let us assume we have a system of m non-dividend-paying risky assets with price processes

dSti n
i
= µt dt + σtiα dWtα . (3.1)
Sti α=1

Here, Sti (i = 1, 2, · · · , n) represents the price process for asset number i.

The drift process µit and the volatility process σtiα are assumed to be bounded and pro-
gressively measurable with respect to the filtration {Ft }.

Intuitively speaking, the latter condition means that these processes depend on the path
of the Brownian motion from 0 up to time t, but otherwise, there is no source of ‘extraneous’
randomness.

This is essentially a causality condition.

23
For the moment, we shall not fix the relation between the number of assets m and the
number of Brownian motions n.

In the case of a complete market, we normally require that m should be greater than or
equal to n.

In other words, for a complete market, there should be at least as many genuinely ‘in-
dependent’ assets as there are ‘sources of randomness’.

Otherwise, there may be more sources of randomness than there are independent means
of hedging away this randomness! That would mean an ‘incomplete’ market.

At time t, the relative magnitude of the price fluctuation of asset i due to Brownian motion
number α is given by σtiα , which we call the volatility matrix.

The exogenous specification of µit and σtiα determines the asset price processes Sti , once
initial prices have been given, according to the formula

 t  t 
i i
i 1 i2
i
St = S0 exp µs − 2 σs ds + σs dWs . (3.2)
0 0

Here we use the compact notation


n
σsi dWs = σsiα dWsα (3.3)
α=1

and

σsi2 = σsiα σsiα . (3.4)
α

For each fixed value of i, we think of σsi as a vector volatility process with n components,
one for each of the n independent Brownian motion.

3.2 Market completeness

For some considerations we impose a condition of market completeness. For market com-
pleteness we require first that the m × n rectangular matrix σtiα should be of rank n.

The interpretation of this condition is that any fluctuation in the Brownian motion is neces-
sarily realised by at least one of the assets in the form of a corresponding price fluctuation.

24
More precisely, σtiα is of maximal rank n at time t if, for any nonzero vector η α = (η 1 , η 2 , · · · , η n )
we have

n
η α σtiα = 0. (3.5)
α=1

If this holds for all η α = 0, then any fluctuation dWtα in the Brownian motion results
in a nontrivial asset price fluctuation dSti .

This is evident from the basic dynamical equations.

Additionally, we will sometimes require to impose a condition on the volatility structure,

sufficient to keep it from getting to ‘close’ to degeneracy.

This can be imposed by requiring that the symmetric matrix


m
ραβ
t = σtiα σtiβ (3.6)
i=1

satisfies the condition that there exists a number  such that

ραβ αβ
t > δ . (3.7)

In other words
 
ραβ
t − δ αβ
ηαηβ > 0 (3.8)
α,β

for any nonvanishing vector η α . This ensures that the eigenvalues of ραβ
t are bounded from
below by .

3.3 Absence of arbitrage in a multi-asset context

Now let us consider the principle of no arbitrage.

This principle implies in the case of an asset that pays no dividend that the drift is of
the form

n
µit = rt + λαt σtiα , (3.9)
α=1

25
for some progressively measurable vector process λt , independent of the value of i.

This is the market risk premium vector, which has the interpretation of being the extra
rate of return, above the interest rate, per unit of volatility in the factor α.

Hence, the no-arbitrage condition tells us that the given family of assets shares a com-
mon risk premium process λαt .

Once we deduce the existence of a market risk premium process, we obtain the following
stochastic equation for the asset dynamics:

dSti n
= r t dt + σtiα (dWtα + λαt dt). (3.10)
Sti α=1

We note the important fact that, in a complete market, the risk premium vector is uniquely
determined by the given stochastic system.

This follows from the observation that, if (3.9) were satisfied for any other choice of risk
premium vector, say, λαt + ηtα , then the market completeness would imply ηtα = 0.

In an incomplete market we can then ask whether it is appropriate to regard λαt as be-
ing exogenously specified.

3.4 Valuation of derivatives in complete multi-asset

markets
Consider now the valuation of derivatives in a complete market.

Many aspects of the present analysis have analogues in the case of a single asset, but there
are some new twists as well that carry over to interest rate theory.

First, we need to introduce the unit initialised money market account process:

 t 
Bt = exp rs ds (3.11)
0

In a complete market with risk premium vector λt , the asset price processes are

 t  t  t 
i i i 1 i 2
St = S0 exp rs ds + σs (dWs + λs ds) − 2 (σs ) ds . (3.12)
0 0 0

26
As a consequence, we see that the ratios of Sti to Bt (discounted asset prices) are given by

 t  t 
Sti i i 1 i 2
= S0 exp σs (dWs + λs ds) − 2 (σs ) ds . (3.13)
Bt 0 0

The combination dWt + λt dt appearing here suggests that, with a change of measure, the
discounted asset prices will be martingales.

To see this, we form the density martingale

 t  t 
1 2
Λt = exp − λs dWs − 2 λs ds . (3.14)
0 0

A short calculation shows that the ratio

 t  t 
Λt Sti i i 1 i 2
= S0 exp − (σs − λs )dWs − 2 (σs − λs ) ds (3.15)
Bt 0 0

is a martingale:
 
Λs Ssi Λt Sti
= Es . (3.16)
Bs Bt

This relation has to hold among all the given assets subject to a no arbitrage condition.

We may therefore consider the situation where one or more of these assets is a deriva-
tive.

Let HT denote the payoff of such a derivative, and let Ht denote the price process for
the derivative at earlier times.

It follows that the value of the derivative is given by:

 
Bt ΛT
Ht = Et HT . (3.17)
Λt BT

For the present value we then obtain the risk neutral valuation formula:
 
λ HT
H0 = E . (3.18)
BT

If dividends are paid, then we need to modify these formulae slightly.

27
In the dynamics for Sti we replace rt with rt − δti where δti is the dividend rate, and we
find that
 t
i Λt Sti Λu δui Sui
Mt = + du (3.19)
Bt 0 Bu
is a martingale.

Then we can develop pricing formulae where both the assets and the derivatives pay contin-
uous dividend.

3.5 Natural numeraire and state-price density

There is an interesting economic interpretation of the basic derivatives pricing formula (3.17).

We note that the process Λt is “dimensionless”, whereas Bt is an asset price. Thus, the
ratio Bt /Λt is also an asset price.

Writing ξt = Bt /Λt we deduce that the dynamical equation for ξt is

dξt
= (rt + λ2t )dt + λt dWt . (3.20)
ξt
We think of the process ξt as the value process for a special portfolio in the money market
account and the basic risky assets with the value process ξt .

Sometimes the value process ξt is referred to as the “natural numeraire portfolio”.

The present value of any other asset, when valued in units of the numeraire portfolio, acts
as an unbiased forecast for the future value of that asset, when expressed in units of the
numeraire portfolio at that time. In other words,
 i
Ssi S
= Es t . (3.21)
ξs ξt

Another useful way of thinking about ξt is to define the related process

1
Vt = . (3.22)
ξt
This is called the state-price density.

The state price is the value of one unit of cash in units of the natural numeraire.

28
For any non-dividend-paying asset St we have
 
VT
St = Et ST . (3.23)
Vt

Now suppose that St is a ‘derivative’ that pays one unit of cash at time T .

Then St is the price process PtT of a discount bond with maturity T . Thus:
 
VT
PtT = Et . (3.24)
Vt

3.6 Incomplete markets

We now consider more generally the case where the market is not complete.

In practice, it is common to encounter derivatives that cannot be completely hedged.

Nevertheless, we may consider a ‘decomposition’ of a given product into a ‘hedgeable’ and

‘unhedgeable’ parts.

If the market in incomplete, then typically then volatility matrix σtiα is degenerate (i.e.
it has one or more zero eigenvalues).

This implies that the risk premium vector λαt that satisfies the no arbitrage condition (3.9)
is not uniquely determined by the specification of the asset price processes.

Nevertheless, we may consider the subspace of Rn spanned by the nondegenerate components

of the volatility matrix σtiα , and construct a decomposition of the form

λαt = ψtα + ϕαt (3.25)

Here, ψtα is the vector λαt with minimum length that satisfies the condition


n
µit = rt + λαt σtiα , (3.26)
α=1

whereas ϕαt satisfies


n
ϕαt σtıα = 0. (3.27)
α=1

29
We now define the process ξt by the dynamics
dξt

= rt + ψt2 dt + ψt · dWt (3.28)

ξt
This is the unique natural numeraire process corresponding to the hedgeable part of the
portfolio.

In other words, ξt is the unique attainable numeraire process.

In a complete market, the derivative price process is given by

 
HT
G t = Et . (3.29)
ξT
However, in an incomplete market, the derivative payout HT contains unhedgeable compo-
nents.

Therefore, we consider the decomposition

H T = JT + KT . (3.30)
Here, JT corresponds to the hedgeable part of the derivative.

This is obtained by taking the conditional expectation Et [HT /ξT ], and projecting the re-
sulting martingale into the subspace spanned by the volatility vectors σtiα .

Then we let t → T and multiply by ξT to obtain JT .

For the remaining unhedgeable part KT , its expectation is given by

 
KT
Et = 0. (3.31)
ξT

α
λ

Space spanned by σi α Projection

Λα

σj α σk α
Σ

Figure 3.1: The decomposition of the risk premium vector.

30
Hence the hedgeable part of the product can be priced in essentially the conventional manner,
while the unhedgeable part can, say, be transferred to a specialist desk to deal with the
residual risk.

31
Chapter 4
Discount bond dynamics. Interest rate volatility and correlation. Short
rate and instantaneous forward rate processes. Heath-Jarrow-Morton (HJM)
framework. Valuation and hedging of interest rate derivatives.

4.1 Price processes for discount bonds

Now we turn to the modelling of interest rate dynamics.

The key idea here is to keep the discount bonds in the centre of the stage.

The short rate, forward short rates, Libor rates, forward Libor rates, and swap rates are
all subsidiary processes.

If one focuses on discount bonds, then the theory of interest rates assumes a unified, co-
herent shape, and also fits in nicely with the consideration of other asset classes, e.g., foreign
currencies, credit-risky bonds, inflation-linked bonds, equities and so on.

As indicated earlier, we write Pab for the value at time a of a discount bond that ma-
tures at time b to deliver one unit of currency. The initial discount function is given by P0b ,
and we have the maturity condition Paa = 1 for all a.

For any given value of b we regard Pab as a stochastic process in the a variable over the
interval 0 ≤ a ≤ b. Thus we have a one-parameter family of assets for which the price
processes are given by Pab .

We call a the “process index” and b the “maturity index”.

We can infer from context whether Pab refers to “the value at time a of a bond that matures
at b”, or “the whole process for fixed b”, or “the values at a fixed time a for a range of
maturity dates b”, or “the whole system of processes”.

32
We shall assume here that the market is driven by a multi-dimensional family of independent
Brownian motions Wtα .

The “factor index” α can be understood (as before) as labelling the basis for a finite di-
mensional vector space, or as an “abstract” index representing a Hilbert space element in
the infinite dimensional case.

The discount bond dynamics are then given by the stochastic equation
dPab
= µab da + Ωab dWa (4.1)
Pab
Here µab is the drift process for the b-maturity bond. Ωab is the corresponding vector volatil-
ity process. Both are assumed to be adapted to the filtration (Ft ) generated by Wtα .

We require that Ωaa = 0, corresponding to the fact that a maturing bond has zero volatility.

We also make the technically useful assumption that the process Ωab is differentiable in
the maturity index.

More specifically, we assume that there exists a process σas such that
 b
Ωab = − σas ds, (4.2)
a

where the minus sign appears as a matter of convention.

This relation enforces the constraint Ωaa = 0.

Note that in the term Ωab dWa there is, as indicated earlier, an implied summation over
the suppressed vector indices.

Now we impose the no-arbitrage condition. By the same line of argument as in the multi-
asset case this ensures the existence of a risk premium vector λαa such that the drift µab is
given by

n
µab = ra + λαa Ωαab . (4.3)
α=1

Suppressing vector indices, we write this as µab = ra + λa Ωab .

Here ra is the short rate, i.e. the rate of return on an instantaneously maturing discount bond.

33
In the discussion on multi-asset dynamics, we regarded ra as an exogenously specified process.
However, in the present consideration, the short rate is given by

∂Pab 
ra = − . (4.4)
∂b a=b

In fact, we shall show that ra can be effectively eliminated as a fundamental variable, and
an expression for the discount bonds can be derived entirely in terms of λαa and Ωαab .

Alternatively, we can eliminate Ωαab , and an expression for the discount bond can be de-
rived in terms of the martingale Λt (which incorporates λαa ) and the short rate process rt
(which can be specified arbitrarily). This will be shown later.

These diverse but ultimately equivalent ways of characterising interest rate dynamics are at
the root of the various apparently diverse approaches to modelling that have been developed.

Inserting the expression for the drift (4.3) into the dynamics (4.1) of the bond prices, we get

dPab
= ra da + Ωab (dWa + λa da) . (4.5)
Pab
Basic interest rate models usually assume the interest rate market is complete.

This means, in particular, that the process Ωαab has to satisfy a nondegeneracy
 α α sufficient to
α
ensure that there does not exist at any time a vector η such that α Ωab η = 0 for all b > a.

In essence, this is equivalent to assuming that any interest rate derivative can be hedged with
a suitable self-financing portfolio of discount bonds, together with the money market account.

Because the system of discount bonds is infinite, but each individual bond has a finite
life, there are various ways in which the completeness condition can be met.

It is important to recognise that completeness is a rather strong assumption, and there-

fore may not be realised in practice.

Even if the discount bond market is not complete, there are circumstances in which it
is appropriate to regard a definite choice of λa as being specified exogenously.

Note that the risk premium vector in the discount bond dynamics (4.5) combines sugges-
tively with the Brownian motion so as to indicate a change of measure. We shall return to
this point when we consider the valuation of interest rate derivatives.

34
4.2 Discount bond volatility and correlation
Let us now consider some “local” properties of the discount bond dynamics.

The dynamical equations under the assumption of no arbitrage are

dPab
= ra da + Ωab · (dWa + λa da) . (4.6)
Pab
It follows on account of the Ito relations

dWtα dWtβ = δ αβ dt, dWtα dt = 0, (dt)2 = 0, (4.7)

that

2
dPab
= |Ωab |2 da. (4.8)
Pab

Here |Ωab |2 = nα=1 Ωαab Ωαab is the squared magnitude of the volatility vector for the bond
with maturity b.

We refer to |Ωab | as the local volatility of the b-maturity discount bond.

If we consider bonds of two different maturities, say b and c, then the instantaneous or
local correlation for their price dynamics is given by the process
Ωtb · Ωtc
ρtbc = . (4.9)
|Ωtb | |Ωtc |

Clearly we have −1 ≤ ρtbc ≤ 1.

To work out the dynamics of Pab , we need to know the vector processes Ωαtb and λαt .

However, to work out the probability laws for Pab , we only require the scalar combinations
|Ωab |, ρtbc , |λt |, and λt · Ωtb .

4.3 Solution for the discount bond processes

The dynamical equation for the bond price involves the bond volatility, the relative risk, and
the short rate.

However, we shall show now that the short rate can be eliminated, to give a representa-
tion of the bond price process in which the exogenous variables are the volatility process and

35
the relative risk process.

The solution of the bond dynamics can be expressed in the form

 a  a 
1 2
Pab = P0b Ba exp Ωsb (dWs + λs ds) − 2 |Ωsb | ds . (4.10)
0 0

Here Ba is the unit-initialised money market account process, given as usual by

 a 
Ba = exp rs ds . (4.11)
0

We observe, on the other hand, that the maturity condition Paa = 1 allows us to solve for
Ba in (4.10).

In particular, if we set a = b, we get:

 a  a 
−1 1 2
Ba = (P0a ) exp − Ωsa (dWs + λs ds) + 2 |Ωsa | ds . (4.12)
0 0

This shows how the short rate can be expressed in terms of the risk premium vector and the
discount bond volatility.

More explicitly, by taking logarithms in (4.12), differentiating with respect to a, and us-
ing the relation Ωaa = 0, we get the following formula for ra :
 a  a
ra = −∂a ln P0a + Ωsa ∂a Ωsa ds − ∂a Ωsa (dWs + λs ds) . (4.13)
0 0

Here ∂a denotes differentiation with respect to a. Thus we have solved for ra in terms of λαa
and Ωαab .

In obtaining these expressions, suitable technical conditions are required to be satisfied

by the discount bond drift and volatility. We shall return later to address this issue more
explicitly.

Inserting formula (4.12) for the money market account into (4.10) for the discount bonds,
we then obtain the following general quotient formula for the discount bonds:
 a a

exp 0 Ωsb (dWs + λs ds) − 12 0 |Ωsb |2 ds

Pab = P0ab  a a
. (4.14)
exp 0 Ωsa (dWs + λs ds) − 12 0 |Ωsa |2 ds

Here P0ab = P0b /P0a denotes the forward value of a b-maturity bond, i.e., the value negoti-
ated today for purchase at time a of a b-maturity discount bond.

36
In the quotient formula (4.14) note that the numerator and the denominator are essentially
similar in structure, except the b in the numerator gets replaced by an a is the denominator.

The quotient formula is the desired explicit expression for the bond prices in terms of the two
exogenous variables, the volatility vector and the relative risk vector, with the elimination
of the short rate.

4.4 HJM dynamics for the forward short rate

The forward short rate process is given by

fab = −∂b ln Pab . (4.15)

From the quotient formula it follows by differentiation that these rates can be expressed as
follows:
 a  a
fab = −∂b ln P0b + Ωsb ∂b Ωsb ds + ∂b Ωsb (dWs + λs ds) . (4.16)
0 0

Heath, Jarrow & Morton (1992) take a general Itô process for the forward short rates as the
starting point, and impose appropriate no-arbitrage and market completeness conditions to
obtain an expression of the form (4.16).

We write σab = −∂b Ωab for the forward short rate (i.e. instantaneous forward rate) volatility.

It follows that
 b
Ωab = − σau du. (4.17)
a

This builds in the constraint Ωaa = 0, as we noted earlier.

Then for the forward short rate processes in terms of σab we obtain:
 a
 b   a
fab = f0b + σsb σsu du ds + σsb (dWs + λs ds) . (4.18)
0 s 0

Taking the stochastic differential of this expression on the process index we get

 b 
dfab = σab σau du da + σab · (dWa + λa da) . (4.19)
a

37
These are the dynamics of the forward short rate, sometimes called the HJM dynamics.

It should be clear that the arbitrage-free dynamics of the discount bond system and the
HJM forward short rate dynamics are for most practical purposes entirely equivalent.

Given the solution of the stochastic equation for fab , we can use the relation

 b 
Pab = exp − fau du (4.20)
a

to find the bond price.

The forward short rate processes are important from a conceptual point of view, but it
should be noted that practical applications invariably refer back to the bond price process
Pab and the short rate process ra .

4.5 Risk neutral valuation of interest rate derivatives

For the value of Ht at time t of a hedgeable interest rate derivative that pays HT at time T ,
we have the forecasting relation
 
Λt Ht ΛT HT
= Et . (4.21)
Bt BT

Equivalently, we can write this in the form

 
HT
Ht = Bt Eλt . (4.22)
BT

Here Eλt denotes conditional expectation in the risk-neutral measure induced by Λt which is
defined by the exponential martingale

 t  t 
1 2
Λt = exp − λs · dWs − 2 |λs | ds . (4.23)
0 0

One of the important features of interest rate theory is that the discount bonds themselves
can be viewed as a species of “derivative”.

The bond that matures at time T has a payoff of unity at that time.

38
As a consequence, if we set HT = 1 in (4.21), we have the following risk-neutral valua-
tion formula for the discount bond price process:
 
Bt ΛT
PtT = Et . (4.24)
Λt BT
An expression of this form was derived by Vasicek (1977).

In the risk neutral measure this can be written as follows:

 
λ 1
PtT = Bt Et
BT

 T 
λ
= Et exp − rs ds . (4.25)
t

These formulae are often used as a starting point for interest rate modelling.

This is because it is possible to specify λt and rt exogenously, without any a priori re-
lation holding between them.

In particular, it follows from the risk-neutral valuation formula that Ptt = 1, and that
for any choice of the process rt and risk premium density Λt , the ratio
Λt PtT
(4.26)
Bt
is a martingale.

This implies that the bond-price system PtT satisfies the no-arbitrage condition, and thus
qualifies as a bona-fide interest-rate model.

Thus summing up, we see that there are two apparently distinct but nevertheless entirely
equivalent ways of “covering” the entire category of interest rate models:
(a) by specifying the relative risk process and vector volatility processes,
(b) by specifying the relative risk density together with the short rate process.
We shall return later to investigate in more detail the problem of how to characterise a
general interest rate model, but let us first consider some specific interest rate models.

4.6 Market Models∗

A good example of an important spin-off of the HJM approach, which has enjoyed consider-
able popularity as a basis for applications, is the so-called ”market model” methodology.

39
There are a number of different variations on this approach–too many to attempt to survey
here–according to which the forward Libor rates and/or swap rates associated the discount
bond system are regarded as the “fundamental” dynamical entities.

In its simplest form, the idea of the market model is as follows. The forward Libor rates Ltab
are defined in a standard way by the relation
1
Ptab = , (4.27)
1 + (b − a)Ltab

where Ptab = Ptb /Pta denotes the forward price made at time t for purchase of a b-maturity
discount bond at time a.

For convenience we introduce a “tenor” parameter δ = b − a, and write Lδta = Ltab .

It is then a straightforward exercise in Ito calculus to work out the dynamics of Lδta , starting
from the bond price dynamics given by
dPtT
= rt dt + ΩtT (dWt + λt dt). (4.28)
PtT
The result is a relation of the following form:

dLδta 1 + δLδta
= (Ωt,a − Ωt,a+δ )(dWt + λt dt − Ωt,a+δ ). (4.29)
Lδta δLδta
The key observation that follows is that if ωt,a is a prescribed deterministic volatility process
for a given fixed tenor then we can solve the equation

1 + δLδta
(Ωt,a − Ωt,a+δ ) = ωt,a (4.30)
δLδta
for the bond volatility in terms of the forward Libor rates and ωt,a .

This shows that there exists an HJM model with the prescribed deterministic volatility
for the given forward Libor rate.

The next step is to change measure so as to eliminate the drift, which can clearly be carried
out since now we know the bond volatility process.

As a consequence, we are left with a log-normal process for the forward Libor rate in the
new measure.

40
It is generally recognised that the market model framework has probably been the single
most influential development in interest rate theory in the post-1992 years following the
advent of the HJM approach.

Many authors have contributed, in one way or another, and to varying degrees, to its orig-
ination and promulgation, and it would be impossible here to attempt with any success an
objective account of the development of the market models and their various extensions,
with all the relevant attributions.

41
Chapter 5
General theory of short rate diffusion models. Diffusion processes. The
Feynman-Kac formula. Derivation of the discount bond pricing equation.

5.1 General theory of short rate diffusion models

An interesting and important class of interest rate models can be obtained by assuming:

(a) the short rate is a diffusion process,

(b) the discount bonds depend on rt as a state variable.

Such models are often called “short rate” models.

Many of the most well-known interest rate models fall into this category, including for exam-
ple the Vasicek model, the CIR model, the Black-Karazinski model, the Black-Derman-Toy
model, the Hull-White model, and the rational lognormal model.

More specifically, we consider a family of discount bond price processes PtT such that

PtT = P (t, rt , T ). (5.1)

Here P (t, r, T ) is a function of three variables. Thus the short rate acts as a “state variable”
for this family of models.

The short rate process rt (t ≥ 0) is assumed to satisfy a stochastic differential equation

of the form

drt = µ(t, rt ) dt + σ(t, rt ) dWt . (5.2)

Each of µ(t, r) and σ(t, r) is a function of two variables.

42
The process Wt is a standard one-dimensional Brownian motion with respect to the nat-
ural probability measure.

Our goal is to derive a partial differential equation satisfied by the function P (t, r, T ) that
arises as a consequence of the dynamical equations for PtT .

This is called the ‘bond pricing equation’.

5.2 Diffusion processes

Before embarking on a derivation of the bond-pricing equation, we digress briefly to say a
few words about diffusions.

This is a topic of great interest in its own right, with many applications.

A process Xt satisfying a stochastic differential equation of the form

dXt = a(t, Xt ) dt + b(t, Xt ) dWt (5.3)
where a(t, x) and b(t, x) are deterministic functions is called a time inhomogeneous diffusion.

If a(t, x) and b(t, x) do not depend explicitly on t, then Xt is a time homogeneous diffu-
sion.

Generalisations of (5.3) can also be considered for which Xt and Wt are both multi-dimensional.

Early interest rate models (e.g., the Vasicek and CIR models) were based on homogeneous
diffusions, but later it was recognised that inhomogeneous diffusions added flexibility to the
models for fitting initial data, in particular initial yield curve and implied volatility data.

Homogeneous diffusions are more appropriate to equilibrium models, but these are not so
useful in a banking context.

If f (t, x) is a smooth function of two variables, then by Ito’s formula we have


∂f 1 2 ∂ 2f ∂f ∂f
df (t, Xt ) = + b 2
+a dt + b dWt . (5.4)
∂t 2 ∂x ∂x ∂x
Here, of course, the derivatives ∂f /∂t, ∂f /∂x and ∂ 2 f /∂x2 are valued at x = Xt .

The second order differential operator

1 ∂2 ∂
L = b2 2 + a (5.5)
2 ∂x ∂x

43
is called the generator of the diffusion.

The generators of diffusions arise naturally in connection with elliptic and parabolic par-
tial differential equations, and in certain cases there are natural probabilistic interpretations
of the solutions of these equations.

These results form the basis of the importance of PDE methods in finance, and have numer-
ous practical applications. We give a few examples.

(a) Consider the parabolic equation

∂ψ
= Lψ, (5.6)
∂t
subject to the initial condition ψ(0, x) = f (x), for some prescribed continuous function f .

Now let Ex denote the expectation operator. Here, the superscript x indicates that we
assume that initially X0 = x.

Under appropriate technical assumptions, the solution of (5.6) is given by

ψ(t, x) = Ex [f (Xt )]. (5.7)

(b) A more general result is the following. Consider the partial differential equation
∂ψ
= Lψ − gψ + h, (5.8)
∂t
subject to the initial condition ψ(0, x) = f (x).

Here, g(x) ≥ 0, h(t, x) and f (x) are prescribed continuous functions.

Then, under suitable technical assumptions, the solution of (5.8) can be expressed in the
form
 t
 u 
 t  
x
ψ(t, x) = E exp − g(Xs ) ds h(u, Xu ) du + exp − g(Xs ) ds f (Xt ) . (5.9)
0 0 0

This result is known as the Feynman-Kac formula.

(c) We now turn to a different kind of result, involving stopping times. Let f (x) be a
smooth function, and let τ be a stopping time such that Ex [τ ] < ∞.

We recall that τ is a stopping time relative to the filtration {Ft } if for every t the event

44
{τ ≤ t} is Ft -measurable.

Intuitively, this means that once we reach any given time t, we can determine whether
the even has occurred or not.

Then Dynkin’s formula says that

 τ 
x x
E [f (Xτ )] = f (x) + E Lf (Xs ) ds . (5.10)
0

(d) Another very useful result is the so-called Kolmogorov forward equation, also known as
the Fokker-Planck equation.

This is a partial differential equation for the probability density function ρ(t, x) of Xt :

∂ρ 1 ∂2 2 ∂
= 2
(b ρ) − (aρ). (5.11)
∂t 2 ∂x ∂x
Given an initial distribution ρ(0, x) for X0 , we can work out the distribution of Xt at later
times t by solving this equation.

We see, for example, that if a = 0 and b = 1 is a constant, then the diffusion is a Brownian
motion and the Fokker-Planck equation reduces to the heat equation.

These results have various multi-dimensional generalisations.

5.3 Derivation of the bond-pricing equation

Let us return to the bond-price process PtT = P (t, rt , T ), t ∈ [0, T ].

For each fixed value of T , we can use Itô’s lemma to obtain

2


∂P ∂P 1 2∂ P ∂P
dP (t, rt , T ) = +µ + 2σ dt + σ dWt . (5.12)
∂t ∂r ∂r2 ∂r

The bracketed expressions are valued at (t, rt ).

In the case of an interest rate system driven by a single Brownian motion, the no-arbitrage
dynamics of PtT are given by

dPtT = (rt + λt ΩtT )PtT dt + ΩtT PtT dWt . (5.13)

45
Equating these two relations we see that the volatility process is given by
1 ∂P (t, rt , T )
ΩtT = σ(t, rt ) . (5.14)
P (t, rt , T ) ∂r
Comparison of the drift terms appearing in (5.12) and (5.13) then allows us to deduce that
the risk premium process λt must be of the form

λt = λ(t, rt ), (5.15)

for some function λ(t, r) of two variables.

In the light of these observations, we can equate the drifts in (5.12) and (5.13) to obtain the
following PDE:
∂P ∂P ∂2P ∂P
+µ + 12 σ 2 2 = rP + λσ . (5.16)
∂t ∂r ∂r ∂r
This is the bond-pricing equation for the short-rate state variable models.

We require a solution of this PDE subject to the terminal condition PT T = 1, or equiv-

alently, P (T, r, T ) = 1 for all r.

5.4 Solution and calibration of the bond pricing equa-

tion
To obtain an interest rate model by this technique, we first need to specify the functions
µ(t, r), σ(t, r) and λ(t, r).

Then we solve the stochastic differential equation (5.2) for the process rt , t ≥ 0.

Ideally, we want diffusions such that rt > 0, for all t ≥ 0, but some models, particularly
older models such as the Vasicek model, do not necessarily have this property.

We then solve the partial differential equation (5.16) for P (t, r, T ) subject to the termi-
nal maturity condition P (T, r, T ) = 1, which must hold for all values of the variable r.

Finally, we may also wish to impose an initial condition P (0, r0 , T ) = P0T , where

r0 = − (∂T P0T )|T =0 . (5.17)

This condition incorporates the initial discount function P0T into the dynamics.

46
The “terminal” condition P (T, r, T ) = 1 is typical of the sort of information one needs
to obtain a unique solution of a parabolic equation such as (5.16), if we are told the func-
tions µ, σ and λ.

This is related to the fact that, subject to some technical considerations, the ordinary heat
equation
∂φ 1 ∂2φ
= (5.18)
∂t 2 ∂x2
has a unique solution φ(t, x) if we supply the initial condition φ(0, x) = f (x).

It will thus not always be possible to impose an initial condition such as (5.17) as well.

This point is illustrated, for example, in the so-called “equilibrium” or “stationary” models,
for which µ, σ and λ do not depend on the variable t, and are functions of r alone.

In this case rt , t ≥ 0, is a stationary diffusion process.

Some well-known examples of models of this type are:

(a) the Vasicek model, for which µ(r) = k(θ − r), σ(r) = σ, and λ(r) = λ where k, θ, σ and
λ are constants;
√ √
(b) the CIR model, for which µ(r) = k(θ − r), σ(r) = σ r and λ(r) = ξ r/σ, where k, θ,
σ and ξ are constants.

In both cases, k and θ are taken to be positive.

In the CIR model we also require kθ > 12 σ 2 , which ensures that rt > 0, for all t ≥ 0, if
we assume that r0 > 0.

It should be noted that the CIR volatility parameter has different “dimensions” from the
Vasicek volatility parameter.

In these examples, the specification of the parameters k, θ, σ, λ, ξ and r0 is sufficient

to completely determine the initial discount function.

In other words, in a stationary short rate model we cannot expect to be able to incorpo-
rate an arbitrary initial yield curve.

Historically, this is one of the reasons why the “extended” models were developed. Origi-
nally, the goal of interest rate modelling was to determine, by equilibrium conditions, a finite

47
dimensional ‘class’ of yield curves to which the actual yield curve would have to belong.

Thus, one way to incorporate initial conditions is to regard the functions µ, σ, λ as only
partially specified. For example, if we set

µ(t, r) = k(t)(θ(t) − r) and σ(t, r) = σ(t), (5.19)

then we obtain the so-called extended Vasicek model , due to Hull and White, and to Jamshid-
ian.

In this case, the disposable functions k(t), θ(t) and σ(t) are chosen so that the initial condi-
tion is satisfied for the discount function.

That only “uses up” one of the three functions (θ as it happens), so it is possible (in princi-
ple) to fix some other initial conditions as well, e.g. implied volatility data for certain classes
of interest rate options.

If sufficient initial “market” data is specified to fix all three functions, then we say that
the model has been calibrated .

There is no known general a priori principle that dictates which initial data should be
incorporated into an interest rate model.

The banking industry is still experimenting with this issue.

A more useful point of view, perhaps, is based on the idea of conditioning.

That is to say, the interest rate model is always conditioned on the data available, and
likewise the pricing of derivatives is always conditional on the information supplied regard-
ing the prices of other derivatives.

Of course, we might try to calibrate an equilibrium model to the initial yield curve, e.g.
by choosing the parameters k, θ, σ, λ and r0 in the case of Vasicek or CIR to fit the initial
term structure.

In practice, one expects this to be difficult since the initial (i.e. current) yield curve can
exhibit a good measure of highly tuned microstructure (e.g. with some rates specified down
to a basis point).

Also, due to spreads and variations between deals there is necessarily some ‘fuzziness’ in
the specification.

48
It should be clear that a short-rate model is a fortiori a discount bond model, and is there-
fore an HJM model.

Thus, mathematically speaking, short rate models constitute a subset of the HJM class,
not a distinct class.

Sometimes it is said that ‘all’ HJM models are short rate models. This is true, but only if
one gives a rather different interpretation to the meaning of ‘short rate’ model. Usually it is
clear from context.

49
Chapter 6
Theory of affine term structure models, including the Vasicek model and the
Cox-Ingersoll-Ross (CIR) model.

6.1 The Vasicek model

Let us consider in more detail the model of Vasicek (1977).

The short rate is assumed to follow a dynamical relation of the form

drt = k(θ − rt ) dt + σ dWt . (6.1)

The constants k, θ and σ are taken to be positive, and have the following interpretation:

σ is the absolute volatility of the rate rt ,

θ is the mean reversion level ,

k is the mean reversion rate.

Clearly, rt is a diffusion process.

The dynamics of rt are exactly solvable in the case of the Vasicek model. The result is
called an Ornstein-Uhlenbeck process, and the solution for rt is as follows:
 t
−kt
rt = θ + (r0 − θ)e + σ ek(s−t) dWs . (6.2)
0

One easily checks that (6.2) satisfies (6.1), subject to the initial condition r0 .

The technique we use to solve (6.1) is to multiply each side of the equation by the inte-
grating factor ekt , and the result drops out quickly.

50
The theory of this process is described in a well-known article by Doob.

From the formula for rt we can read off a number of qualitative features of the process.
 
t ks
For example, since E 0 e dWs = 0, we see that

E [rt ] = θ + (r0 − θ)e−kt . (6.3)

This result shows that the mean of rt starts at r0 , and then over time reverts to θ.

The ‘speed’ of this movement is governed by the constant k.

Also, we have
σ2

var [rt ] = 1 − e−2kt . (6.4)

2k
Thus the variance of rt is initially zero, and it increases to a maximum level given by 12 σ 2 /k.
The calculation of the variance involves a simple application of the Ito isometry.

In particular, if f (s, t) is deterministic, then


 2   t
t
E f (s, t)dWs = f 2 (s, t)ds. (6.5)
0 0

There is a characteristic time-scale 1/k associated with the mean-reversion rate. This de-
termines the time scale over which rt moves from r0 towards θ.

The bond pricing equation is exactly solvable in the Vasicek model if we assume the rel-
ative risk λ is a constant.

The solution can be written as follows:


1 −k(T −t)

∞ ∞ σ2

−k(T −t) 2
P (t, rt , T ) = exp 1−e (R − rt ) − (T − t)R − 3 1 − e . (6.6)
k 4k
The constant R∞ here is defined by
λσ σ2
R∞ = θ − − 2. (6.7)
k 2k
The significance of R∞ is that it represents the continuously compounded rate of interest
(yield) on a bond of very long maturity. This is seen as follows.

51
We recall that, given the bond price PtT , the continuously compounded yield RtT is

PtT = exp (−(T − t)RtT ) . (6.8)

Inverting this relation we have

1
RtT = − ln PtT . (6.9)
T −t
Hence in the present example we have
1
σ2
2
RtT = R + (rt − R) 1 − e−k(T −t) + 3 1 − e−k(T −t) . (6.10)
k(T − t) 4k (T − t)
Thus for fixed t, we have RtT → R as T → ∞.

On the other hand, we can also check that Rtt = rt . In other words, the yield on a very
short maturity bond is the short rate.

6.2 Affine models

The expression for the bond price in the Vasicek model can be simplified if we introduce the
variable u = T − t for the time to maturity, and set

Btu = Pt,t+u = P (t, rt , t + u). (6.11)

Here Btu represents the price at time t of a bond with u years until maturity.

We call u the tenor of the bond.

For convenience, let us define the function

1

f (u) = 1 − e−ku . (6.12)

k
Then for the Vasicek bond price we have


σ2 2
Btu = exp −f (u)rt + (f (u) − u)R − f (u) . (6.13)
2k
We note that the exponent is a linear function of rt , and the coefficients are functions of u.

The class of interest rate models for which Btu = Pt,t+u can be put into the form

Btu = e−f (u)rt −g(u) (6.14)

52
is of special interest. These are called the stationary affine models.

A short rate model generates a stationary discount bond system if rt is stationary and
Btu can be expressed in the form Btu = B(rt , u), where B(r, u) is a function of two variables.

If the bond price takes the more general form

PtT = e−F (t,T )rt −G(t,T ) (6.15)

for deterministic functions F (t, T ) and G(t, T ), then we have an extended affine model.

For example, it is an interesting exercise to show that in the case of an extended Vasicek
model , for which rt follows a process of the form

drt = k(t)(θ(t) − rt ) dt + σ(t) dWt , (6.16)

where k(t), θ(t) and σ(t) are positive, deterministic functions, then the bond price system is
of the extended affine type.

In the extended Vasicek model (also sometimes known as the Hull-White model), it is a
straightforward exercise to show that
 t  t
−β(t) −β(t) β(s) −β(t)
rt = r 0 e +e e θ(s) ds + e eβ(s) σ(s) dWs , (6.17)
0 0

where
 t
β(t) := k(s) ds (6.18)
0

The relevant calculations follow the arguments given earlier.

In particular, for constant parameters, this reduces to the previous expression given for
the short rate process rt .

6.3 The CIR model

In this important and rather more complicated model (Cox, Ingersoll & Ross 1985), the
short rate is assumed to be a mean reverting square-root process, for which the dynamics are
given by:
√
drt = k(θ − rt ) dt + σ rt dWt . (6.19)

53
The solution of this equation is not as easy as in the Vasicek model.

Nevertheless, its essential features can be revealed by writing the dynamics in integral form
 t
−kt √
rt = θ + (r0 − θ)e + σ e−k(t−s) rs dWs . (6.20)
0

The mean reverting property of rt is apparent.

We immediately infer the mean of rt :

E [rt ] = θ + (r0 − θ)e−kt . (6.21)

By use of the Itô isometry we then obtain


var [rt ] = E (rt − E [rt ])2
 t 
2 −2kt 2ks
= σ e E e rs ds (6.22)
0

Substituting the expression for E [rs ] into this and integrating, we get:

σ2θ
2 σ 2

var [rt ] = 1 − e−kt + r0 e−kt 1 − e−kt . (6.23)

2k k
Thus for small t we have var [rt ] σ 2 tr0 .

Whereas, for large t we have var [rt ] → σ 2 θ/2k.

It is a subtle result due to Feller that:

(a) if r0 > 0, then rt ≥ 0, and

(b) if r0 > 0 and kθ > 12 σ 2 , then rt > 0 (strictly positive interest rates).

Now so far we have not yet considered the market price of risk.

To obtain a solution for the bond pricing equation it turns out that we need to assume
that the relative risk process is of the special form
√
ξ rt
λt = (6.24)
σ
where ξ is a constant.

54
Then we can solve for PtT as a function P (t, rt , T ).

If we write PtT in the affine form

PtT = e−F (t,T )rt −G(t,T ) (6.25)

then the solution for F (t, T ) and G(t, T ) can be found in terms of k, θ, σ and ξ.

In particular, writing

P (t, r, T ) = e−F (t,T )r−G(t,T ) , (6.26)

we see that the bond-pricing equation reduces to two conditions, namely

∂F
1+ = (k + ξ)F + 12 σ 2 F 2 (6.27)
∂t

and

∂G
= −kθF. (6.28)
∂t
We need to solve these subject to the boundary conditions

∂F 
F (T, T ) = 0, G(T, T ) = 0, and = 1. (6.29)
∂t t=T

For convenience, we define the constants

√
ν := k + ξ, γ := ν 2 + 2σ 2 (6.30)

Then the solution is given by

2(eγx − 1)
F (t, T ) = (6.31)
(γ + ν)(eγx − 1) + 2γ
  2kθ
2γe(γ+ν)x σ2
e−G(t,T ) = (6.32)
(γ + ν)(eγx − 1) + 2γ

where x = T − t.

55
Chapter 7
Overview of term structure frameworks. Admissible term structures and
term structure comparison. Dynamics of the term structure density. Positive
interest HJM volatility structure.

7.1 Overview of term structure frameworks

Dynamical models for interest rates suffer from the fact that it is difficult to isolate the
independent degrees of freedom in the evolution of the term structure.

The question is, which ingredients in the determination of an interest rate model can and
should be specified independently and exogenously?

We shall consider briefly two examples of this can be done for general interest rate models,
indicating as well the associated drawbacks.

Example 1. Dynamic models for the short rate. The independent degrees of freedom are
given by:

(a) the specification of the short rate rt as an essentially arbitrary Ito process, and

(b) a market risk premium λαt (α = 1, 2, · · · , n).

The model for the discount bonds is


 T 
1
PtT = Et ΛT exp − rs ds . (7.1)
Λt t

Here Et denotes conditional expectation with respect to the filtration Ft . The density mar-
tingale Λt is defined by

 t  t 
1 2
Λt = exp − λs dWs − 2 λs ds . (7.2)
0 0

56
where

n
λs dWs = λαs dWsα (7.3)
α=1

An advantage of this general model is that rt and λαt can be specified exogenously, and for
interest rate positivity it suffices to let the process rt be positive.

There are two disadvantages to this approach.

Firstly, the model is specified implicitly: the conditional expectation is generally difficult
to calculate. Secondly, the initial term structure is not fed in directly.

A further simplification can be achieved by introducing the state price density:

 t 
Vt = Λt exp − rs ds . (7.4)
0

It follows that
Et [VT ]
PtT = . (7.5)
Vt

Then it is sufficient to specify the state price density Vt alone, and we recover rt and λαt
from the relation
dVt
= −rt dt − λt dWt . (7.6)
Vt

Example 2. The Heath-Jarrow-Morton framework. In this case the independent dynam-

ical degrees of freedom consist of:

(a) the initial term structure P0T ,

(b) the market risk premium process λαt , and

α
(c) the forward short rate volatility process σtT for each maturity T .

The model for the discount bonds is

 T 
PtT = exp fts ds . (7.7)
t

57
The forward short rates are
 t  t
∂
ftT =− ln P0T − σsT ΩsT ds + σsT (dWs + λs ds), (7.8)
∂T s=0 s=0

where
 T
ΩαtT =− α
σtu du. (7.9)
u=t

The advantage of the HJM framework is that it allows a direct input of the initial term
structure, as well as control over the volatility structure of the discount bonds.

A disadvantage of the HJM approach is that there is no guarantee of interest rate posi-
α
tivity, and it is not easy to impose a condition on σtT to achieve this.

Now we consider an alternative framework for isolating the independent degrees of free-
dom in interest rate dynamics that has the virtue of retaining the desirable features of both
examples cited above, while eliminating the undesirable features.

The key idea is the introduction of a term structure density process ρt (x) defined by

∂
ρt (x) = − Btx . (7.10)
∂x
Here Btx denotes the system of bond prices at time t when we parameterise the bonds by
the tenor variable x = T − t (Musiela parameterisation), so

Btx = Pt,t+x . (7.11)

We make the assumption that Btx → 0 for large x.

It is then a straightforward exercise to verify that the interest rate positivity conditions
∂
0 < Btx ≤ 1, and Btx < 0 (7.12)
∂x
are equivalent to the following relations on ρt (x):
 ∞
ρt (x) > 0, and ρt (x)dx = 1. (7.13)
0

We therefore conclude that any positive interest rate model can be regarded as a random
process on the space of density functions on the positive real line.

58
The idea is that we treat the yield curve as a mathematical object in its own right, identified
as a “point” ρ lying in the space M of all possible yield curves.

With the specification of an initial yield curve ρ0 we model the resulting dynamics as a
random trajectory ρt in M. By bringing the structure of M into play it is possible both to
clarify the status of existing interest models, and also to devise new interest rate models.

7.2 Admissible term structures and term structure com-

parison.
There is a natural ‘information geometry’ associated with the space of yield curves.

Let t = 0 denote the present, and P0x a family of discount bond prices satisfying P00 = 1,
where x is the tenor (0 ≤ x < ∞).

We impose the condition that interest rates should always be positive with the following
criterion:
Definition. A term structure is said to be admissible if the discount function P0x is of
class C ∞ and satisfies 0 < P0x ≤ 1, ∂x P0x < 0, and limx→∞ P0x = 0.
An admissible discount function can be viewed as a complementary probability distribution.
In other words, we can think of the tenor date as an abstract random variable X, and for
its distribution write

Pr[X < x] = 1 − P0x . (7.14)

The associated density function ρ(x) = −∂x P0x satisfies ρ(x) > 0 for all x, and
 ∞
ρ(u)du = P0x .
x

We say that a density function is smooth if it is of class C ∞ on the positive half-line

R1+ = [0, ∞).

Proposition 1. The system of admissible term structures is isomorphic to the convex

space D(R1+ ) of everywhere positive smooth density functions on the positive real line.
The requirement that P0x should be of class C ∞ can be weakened, but in practice any term
structure can be approximated arbitrarily closely by a ‘nearby’ term structure with a smooth
density.

It is reasonable to insist that the forward short rate curve f0x = −∂x ln P0x is piecewise

59
continuous and nonvanishing for all x < ∞.

Given a pair of term structure densities ρ1 (x) and ρ2 (x) we can define a distance function
φ12 on M by
 ∞
−1
φ12 = cos ξ1 (x)ξ2 (x)dx, (7.15)
0

where ξi (x) = ρi (x). We call this angle the Bhattacharyya distance between the given
yield curves.

The geometrical interpretation of φ12 arises from the fact that the map ρ(x) → ξ(x) as-
sociates to each point of M a point in the positive orthant S + of the unit sphere in the
Hilbert space L2 (R1+ ), and φ12 is the resulting spherical angle on S + .

Note that 0 ≤ φ < 12 π and that orthogonality can never be achieved if forward rates are
nonvanishing.

As a simple illustration we consider the family of discount bonds given by

−κ
Rx
P0x = 1 + , (7.16)
κ
where R and κ are constants.

In this case we have a flat term structure, with a constant annualised rate of interest R
assuming compounding at the frequency κ over the life of each bond.

For κ = 1 this reduces to the case of a flat rate on the basis of a simple yield, and in
the limit κ → ∞ we recover the case of a flat rate on the basis of continuous compounding.

For the density function ρ(x) = −∂x P0x associated with (7.16) we obtain

−(κ+1)
Rx
ρ(x) = R 1 + . (7.17)
κ
Let us write ρi (x) for the density corresponding to R = Ri (i = 1, 2) for a fixed value of κ.
A direct calculation of the integral (7.15) for κ = 1 gives

√ 
−1 R1 R2 R1
φ12 = cos log . (7.18)
R 1 − R2 R2
In the limit κ → ∞ (continuous compounding) we find that

√ 
−1 2 R1 R2
φ12 = cos . (7.19)
R1 + R2

60
Note that the bracketed term in (7.19) is the ratio of the geometric and arithmetic means of
the two rates. In this limit we have ρ(x) → Re−Rx .

7.3 Dynamics of the term structure density

Now let us consider the evolution of the term structure density.

We write PtT for the random value at time t of a discount bond that matures at time
T , where T ∈ R1+ and 0 ≤ t ≤ T , and assume, for each T , that PtT is an Ito process on the
interval t ∈ [0, T ]:

dPtT = mtT dt + ΣtT dWt . (7.20)

The absolute drift mtT and the absolute volatility process ΣtT are assumed to satisfy regu-
larity conditions sufficient to ensure that ∂T PtT is also an Ito process.

For interest rate positivity we require 0 < PtT ≤ 1 and ∂T PtT < 0.

Additionally we impose the asymptotic conditions limT →∞ PtT = 0, and limT →∞ ∂T PtT = 0.

Because PtT is positive, the forward short rate process ftT is an Ito process iff −∂T PtT
is an Ito process.

For no arbitrage we require the existence of an exogenous market risk premium process
λt such that

mtT = rt PtT + λt ΣtT . (7.21)

We do not assume the bond market is complete. If the bond market is complete, however,
then λt is determined endogenously by the bond price system.

We introduce the Musiela parameterisation x = T − t, and write Btx = Pt,t+x for the
price at time t of a bond for which the time to maturity is x.

We have the following dynamics for Btx :

dBtx = (rt − ft,t+x )Btx dt + Σt,t+x (dWt + λt dt). (7.22)

Now consider the time dependent term structure density ρt (x) defined by (7.10), for which
we have the normalisation condition
 ∞
ρt (x)dx = 1, (7.23)
x=0

61
or equivalently
 ∞
ρt (u − t)du = 1. (7.24)
u=t

The relation
ρt (T − t) = ftT PtT (7.25)
allows us to deduce an interpretation of the normalisation condition. In particular, the
formula
 ∞
Ptu ftu du = 1 (7.26)
t

says that the value at time t of a continuous cash flow in perpetuity that pays the small
amount ftu du at time u is always unity.

Thus we can think of ftu as defining the ‘convenience yield’ associated with a position in cash.

An analogous calculation shows that

 ∞
κ 1
Ptu ftu du = (7.27)
t κ
for any positive value of the exponent κ.

This relation can be interpreted by saying that if we ‘fix’ the convenience yield (e.g., by
swapping the unit of cash for the corresponding future cash flow), and then rescale all the
interest rates Rtu by the same factor κ, so Rtu → κRtu for all u ≥ t, then the value of the
promised cash flow scales inversely with respect to κ.
Returning now to the evolutionary equation we write
ωtx = −∂x Σt,t+x . (7.28)
Then we obtain the following dynamics for ρt (x):
dρt (x) = (rt ρt (x) + ∂x ρt (x))dt + ωtx (dWt + λt dt). (7.29)
∞
The process ωtx is subject to the constraint 0 ωtx dx = 0, which implies that ωtx is of the
form
ωtx = ρt (x)(νt (x) − ν̄t ), (7.30)
where νt (x) is unconstrained, and
 ∞
ν̄t = ρt (u)νt (u)du. (7.31)
0

62
It follows from equation (7.28) that the absolute discount bond volatility ΣtT is given in the
Musiela parameterisation by
 ∞
Σt,t+x = ωtu du
u=x
 ∞
= ρt (u)νt (u)du − ν̄t Btx . (7.32)
u=x

This relation has an interesting probabilistic interpretation. Suppose, in particular, we write

Ix (u) for the indicator function

Ix (u) = χ(u ≥ x), (7.33)

where χ(A) is unity if A is true and vanishes otherwise.

Then the bond price Btx can be written in the form of an abstract ‘expectation’:
 ∞
Btx = ρt (u)Ix (u)du = Mt [Ix ] , (7.34)
u=0

where
 ∞
Mt [g] = ρt (u)g(u)du (7.35)
u=0

for any function g(x).

The absolute discount bond volatility can then be expressed as an abstract covariance of
the form

Σt,t+x = Mt [Ix νt ] − Mt [Ix ] Mt [νt ]. (7.36)

We see that the bond volatility structure Σt,t+x is invariant under the transformation νt (x) →
νt (x) + αt , where αt is independent of x.

This ‘gauge’ freedom can be used to set λt = −ν̄t . Then λt and Σtx are both determined by
νt (x).
Proposition 2. The general admissible term structure evolution based on the filtration
generated by a Brownian motion Wt on H is a measure valued process ρt (x) on D(R1+ ) that
satisfies

dρt (x) = (rt ρt (x) + ∂x ρt (x)) dt + ρt (x) (νt (x) − ν̄t ) (dWt − ν̄t dt) , (7.37)

63
 ∞
where ν̄t = 0 ρt (u)νt (u)du. The volatility structure νt (x) can be specified exogenously along
with the initial term structure density ρ0 (x). The associated short rate process rt = ρt (0)
satisfies

drt = rt2 + ∂x ρt (x)|x=0 dt + rt (νt (0) − ν̄t )(dWt − ν̄t dt). (7.38)

The dynamical equation for the term structure density can be solved exactly as follows:
Proposition 3. The solution of the dynamical equation for ρt (x) in terms of the volatility
structure νt (x) and the initial term structure density ρ0 (x) is
  
t 1 t 2
exp s=0 VsT dWs − 2 s=0 VsT ds
ρt (T − t) = ρ0 (T )  ∞    , (7.39)
t 1 t
u=t
ρ 0 (u) exp s=0
Vsu dW s − V
2 s=0 su
2 ds du

where
Vtu = νt (u − t). (7.40)

Proof∗ . The second term in the drift on the right of (7.37) can be eliminated by setting
x = T − t, which gives us
dρt (T − t) = rt ρt (T − t)dt + ρt (T − t) (νt (T − t) − ν̄t ) (dWt − ν̄t dt) . (7.41)
Integrating this relation and separating out the terms involving ν̄t we obtain
 t  
t 1 t 2
exp s=0 rs ds + s=0 νs (T − s)dWs − 2 s=0 νs (T − s)ds
ρt (T − t) = ρ0 (T )    . (7.42)
t 1 t
exp s=0 ν̄s dWs − 2 s=0 ν̄s ds
2

It follows by use of the definition (7.40) that

 t t 
t
exp s=0 rs ds + s=0 VsT dWs − 12 s=0 VsT 2
ds
ρt (T − t) = ρ0 (T )  t  . (7.43)
t
exp s=0 ν̄s dWs − 12 s=0 ν̄s2 ds

Then with an application of the normalisation condition (7.24) we deduce as a consequence

of (7.43) that

 t  t  t 
1 2
exp − rs ds + ν̄s dWs − 2 ν̄s ds
s=0 s=0 s=0
 ∞
 t  t 
1 2
= ρ0 (u) exp Vsu dWs − 2 Vsu ds du. (7.44)
u=t s=0 s=0

When this relation is inserted in the denominator of (7.43), we immediately obtain (7.39).
♦

64
7.4 Positive interest HJM volatility structure.
It is interesting in this connection to note, by setting T = t in (7.39), that the short rate
process is given by
  
t 1 t 2
exp s=0 Vst dWs − 2 s=0 Vst ds
rt = ρ0 (t)  ∞    . (7.45)
t 1 t
u=t
ρ 0 (u) exp s=0
Vsu dW s − V
2 s=0 su
2 ds du

We observe, in particular, that in a deterministic model, with Vst = 0, this formula reduces
∞
to rt = ρ0 (t)/ t ρ0 (u)du, or, in other words, rt = f0t .

For the market risk premium process it follows from (7.31) together with the relation λt = −ν̄t
that
∞   
α t 1 t 2
ρ
u=t 0
(u)Vtu exp V
s=0 su
dW s − V
2 s=0 su
ds du
λαt = −  ∞ 
t   . (7.46)
1 t
ρ
u=t 0
(u) exp s=0 su
V dW s − V 2 ds du
2 s=0 su

These formulae show that, given the initial term structure density ρ0 (x) and the volatility
structure νt (x), we can reconstruct the short rate process and the market risk premium pro-
cesses.

We deduce from (7.39) that the corresponding formula for the bond price process is
∞   
t 1 t 2
ρ (u) exp s=0 Vsu dWs − 2 s=0 Vsu ds du
u=T 0
PtT =  ∞    . (7.47)
t 1 t
u=t
ρ 0 (u) exp s=0
Vsu dW s − V 2 ds du
2 s=0 su

For the unit-initialised money market account Bt , satisfying dBt = rt Bt dt and B0 = 1, we

have
 t 
t
exp s=0 ν̄s dWs − 12 s=0 ν̄s2 ds
Bt =  ∞    , (7.48)
t 1 t
u=t
ρ 0 (u) exp s=0
Vsu dW s − 2 s=0
V 2
su ds du

which follows directly from (7.44).

The density martingale Λt is given by

 t  t 
Λt = exp 1
ν̄s dWs − 2 ν̄s2 ds , (7.49)
s=0 s=0

65
For the state price density we have
 ∞
 t  t 
1 2
Zt = ρ0 (u) exp Vsu dWs − 2
Vsu ds du. (7.50)
u=t s=0 s=0

As a consequence we can then check that Zt = Λt /Bt .

If we divide (7.39) by (7.47) we are led to a recipe for constructing the general positive
interest HJM forward short rate system ftT in terms of freely specified data:
 t 
t
exp s=0 VsT dWs − 12 s=0 VsT 2
ds
ftT = ρ0 (T )  ∞    . (7.51)
t 1 t
ρ
u=T 0
(u) exp V
s=0 su
dW s − V
2 s=0 su
2 ds du

Note that when T = t this expression reduces to formula (7.45). A short calculation then
allows us to deduce the following result:
Proposition 4. The general positive interest HJM forward short rate volatility structure
is

σtT = ftT (VtT − UtT ) (7.52)

where ftT is given by (7.51), and

∞  t 
t 1
u=T
ρ0 (u)Vtu exp Vsu dWs −
s=0 2
du V 2 ds
s=0 su
UtT = ∞  t   . (7.53)
1 t
u=T
ρ 0 (u) exp s=0
Vsu dW s − V 2 ds du
2 s=0 su

The initial term structure density ρ0 (x) and the volatility structure Vtu (u ≥ t) are freely
specifiable.
In other words, in the HJM theory the forward short rate volatility is not freely specifiable
if the interest rates are to be positive.

Instead it must be of the form (7.52) where VtT is freely specifiable, along with the ini-
tial term structure.

This result establishes a connection between the present approach and Example 2, and
resolves the outstanding difficulty associated with that example.

66
 ξ
D (R1+ ) ρ S+
2
1
0
1
0 1
0
0
1 0
1
1
0 ρ
0
1
ρ
1 Αρ1+Βρ 2
S
H

Figure 7.1: The system of admissible term structures. A smooth positive interest term
structure can be regarded as a point in D(R1+ ), the convex space of smooth and everywhere
positive density functions on the positive half-line R1+ . Associated with each point ρ ∈ D(R1+ )
there is a ray ξ lying in the positive orthant S+ of the unit sphere S in the Hilbert space
H = L2 (R1+ ). A dynamical trajectory on D(R1+ ) can then be mapped to a corresponding
trajectory in S+ .

67
Chapter 8
Construction of admissible models. Moment analysis and the role of per-
petual annuity. The information content of the term structure. Entropic
calibration. Canonical term structures.

8.1 Construction of admissible models

As a consequence of Proposition 3 we see that the general term structure density can also
be expressed in the form

ρ0 (t + x)Mt,t+x
ρt (x) =  ∞ , (8.1)
ρ (t + u)Mt,t+u du
u=0 0

or equivalently

ρ0 (T )MtT
ρt (T − t) =  ∞ , (8.2)
ρ (u)Mtu du
u=t 0

where for each T the process MtT is a martingale (0 ≤ t ≤ T < ∞) such that MtT > 0 and
M0T = 1.

The process MtT is the exponential martingale associated with VtT .

This expression for ρt (T − t) arises also in the Flesaker and Hughston framework, in which
the discount bond system has the representation
∞
ρ0 (u)Mtu du
PtT = u=T∞ . (8.3)
ρ (u)Mtu du
u=t 0

Quasi-lognormal models. An interesting class of specific models is obtained if we restrict

the Brownian motion to be one-dimensional and let the volatility structure Vtu = νt (u − t)
appearing in (7.39) be deterministic.

68
Then Vtu is a function of two variables defined on the region 0 ≤ t ≤ u < ∞. The result-
ing term structure model has a good deal of tractability and exhibits some desirable features.

In particular, the function Vtu has the right structure for allowing a calibration of the model
to a family of implied caplet volatilities for a fixed strike (e.g., at-the-money).

If the dimensionality of the Brownian motion is increased then other strikes can be in-
corporated as well.

Semi-linear models∗ . Another interesting special case can be obtained if we write

 ∞
ρ0 (u) = e−uR φ(R)dR, (8.4)
0

for the initial term structure density, where φ(R) is the inverse Laplace transform of ρ0 (u).

Then for certain choices of the martingale family Mtu the integration in (8.3) can be carried
out explicitly.

An example can be obtained as follows. Let Mt be a martingale (0 ≤ t < ∞) and Qt

the associated quadratic variation satisfying (dMt )2 = dQt , and set

MtT = exp (α + βT )Mt − 12 (α + βT )2 Qt . (8.5)

This model arises if we put

νt (T − t) = (α + βT )σt (8.6)

in Proposition 2, where the process σt is defined by dMt = σt dWt . Then the u-integration
can be carried out explicitly in the expressions for ρt (x) and PtT , and the results can be
expressed in closed form:
∞  ∞

R=0
φ(R) u=T e−uR Mtu du dR
PtT =  ∞  ∞
. (8.7)
φ(R) e −uR M du dR
R=0 u=t tu

Here the bracketed expression in the integrand in the numerator is given by:
 ∞

−uR 1 (Mt − R/β)2
e Mtu du = √ exp + αR/β
u=T |β| Qt 2Qt


Mt − R/β 
×N ± √ ∓ (α + βT ) Qt , (8.8)
Qt

69
where

1 x

N (x) = √ exp − 12 ξ 2 dξ (8.9)

2π −∞

is the normal distribution function, and the ± sign is chosen in accordance with the sign of β.

For example, in the case of an initial term structure with a constant continuously com-
pounding rate r, corresponding to the choice φ(R) = δ(R − r), we obtain
 √ 
N ± M√ t −r/β
Qt
∓ (α + βT ) Qt
PtT =  √ . (8.10)
Mt −r/β
N ± Qt ∓ (α + βt) Qt
√

8.2 Moment analysis and the role of the perpetual an-

nuity
Some interesting aspects of the term structure dynamics are captured in the properties of
the moments of ρt (x), defined by
 ∞  ∞
(n)
x̄t = xρt (x)dx, x̄t = (x − x̄t )n ρt (x)dx (8.11)
0 0

where n ≥ 2.

For example, in the case of a continuously compounded flat yield curve given at t = 0
(2) (3)
by the density function ρ0 (x) = Re−Rx , we have x̄0 = R−1 , x̄0 = R−2 , x̄0 = 3R−3 and
(4)
x̄0 = 9R−4 .

The first four moments, if they exist, are the mean, variance, skewness and kurtosis of
the distribution of the ‘abstract’ random variable X characterising the term structure.

The mean x̄t is a characteristic time-scale associated with the yield curve, and its inverse
1/x̄t is an associated characteristic yield. The financial significance of x̄t will be discussed
shortly.

For simplicity we introduce the following notation for the variance process:
 ∞
vt = x2 ρt (x)dx − (x̄t )2 . (8.12)
0

We assume that ρt (x) and the discount bond volatility Σt,t+x fall off sufficiently rapidly to
n
(x) = 0 and limx→∞ xn Σt,t+x = 0 for n = 1, 2, and that the integrals
 ∞ n that limx→∞x∞ ρtn−1
ensure
0
x ρt (x)dx and 0 x Σt,t+x dx exist for n = 1, 2.

70
 Proposition 5. The mean x̄t of an admissible arbitrage-free term structure satisfies
dx̄t = (rt x̄t − 1)dt + Σ̄t (dWt + λt dt), (8.13)
∞
where Σ̄t = 0
Σt,t+x dx.
There is a critical value x̄∗t for the first moment given by
1
x̄∗t = (1 − λt Σ̄t ), (8.14)
rt
such that when x̄t > x̄∗t the drift of x̄t is positive, and the drift increases as x̄ increases.

When x̄t < x̄∗t , the drift of x̄t is negative, and the drift decreases further as x̄t decreases.

The first moment x̄t has the natural financial interpretation of being the value at time t
of a perpetual annuity paid on a continuous basis.

In particular, an integration by parts shows that

 ∞
x̄t = Btx dx, (8.15)
0

corresponding to an annuity of one unit of cash per year paid continuously in perpetuity.

Higher moments of the term structure density can then be interpreted in terms of the du-
ration, convexity, etc., of the annuity—in other words, as a measure of the sensitivity of the
value of the annuity to an overall change in interest rate levels.

For example, let us write

Btx = e−xrt (x) , (8.16)
where rt (x) is the continuously compounded rate at time t for tenor x, then under a small
parallel shift ∆r in the yield curve given by
rt (x) −→ rt (x) + ∆r, (8.17)
we have, to first order,
Btx −→ (1 − x∆r)Btx . (8.18)

Therefore, to first order the value of the annuity changes by the amount
 ∞
1
x̄t −→ x̄t − 2 ∆r x2 ρt (x)dx, (8.19)
0

71
where in obtaining the second term we use an integration by parts.
Proposition 6. Under a parallel shift in the yield curve the change ∆x̄t in the value of
the perpetual is

∆x̄t = −Dt x̄t ∆r, (8.20)

where the duration Dt of the perpetual annuity is given by


1 ∞ 2
x ρt (x)dx
Dt = 2  0∞ . (8.21)
0
xρt (x)dx

8.3 The information content of the term structure

Now we introduce another important example of a functional of the term structure, the
Shannon entropy of the density function ρt (x). This is defined by
 ∞
St [ρ] = − ρt (x) ln ρt (x)dx. (8.22)
0

Because ρt (x) has dimensions of inverse time, St [ρ] is defined only up to an overall addi-
tive constant. The difference of the entropies associated with two yield curves therefore has
an invariant significance.

One can think of St [ρ] as being a measure of the ‘information content’ of the term structure
at time t. In particular, the higher the value of St [ρ], the lower the information content.

Since ρt (x) is subject to a dynamical law, we can infer a corresponding dynamics for the
entropy.
Proposition 7. The entropy associated with an admissible arbitrage-free term structure
dynamics obeys the evolutionary law

 ∞ 

1
dSt = rt (St + ln rt − 1) + 2 Γt dt + νt (x)st (x)dx − ν̄t St dWt∗ (8.23)
0

where dW ∗ = dWt + λt dt, st (x) = −ρt (x) ln ρt (x) is the entropy density, and the process Γt
is defined by
 ∞
Γt = (νt (x) − ν̄t )2 ρt (x)dx. (8.24)
0

72
8.4 Entropic calibration
The principle of entropy maximisation can be used as the basis for a new yield curve cali-
bration methodology.

In particular, given a set of data points on a yield curve, the ‘least biased’ term structure
can be determined by maximising the Shannon entropy subject to the given data constraints.

The general idea behind the maximisation of entropy under constraints can be sketched
as follows.

Suppose that, given a function H(X) of a random variable X, we are told that the ex-
pectation of H(X) with respect to an unknown distribution with density ρ(x) is U , i.e.,
 ∞
H(x)ρ(x)dx = U. (8.25)
0

The aim then is to find the density ρ(x) that is least biased and yet consistent with the
information (8.25).

In other words, we wish to eliminate any superfluous information in ρ(x).

We also have the normalisation condition

 ∞
ρ(x)dx = 1. (8.26)
0

Subject to the constraints (8.25) and (8.26) we then determine the density ρ(x) that max-
imises the entropy.

This is carried out by introducing Lagrange multipliers, and considering the variational
relation
δ
(−ρ ln ρ − λρH − νρ) = 0. (8.27)
δρ
The solution is

ρ(x) = exp (−λH(x) − ν − 1) , (8.28)

where λ and ν are determined implicitly.

Let us illustrate the idea by considering the situation in which we are given a set of data
points on the yield curve together with the value of a perpetual annuity.

73
The problem is to calibrate the initial term structure to the given data.

This example is interesting because if we are given only the value x̄0 of the perpetual annuity,
then the maximum entropy term structure is
ρ0 (x) = Re−Rx , (8.29)
where R = 1/x̄0 , and thus P0x = e−Rx for the discount function.

Therefore, we see that it is the annuity constraint that leads to the desired exponential
‘die-off’ of the discount function.

This feature is preserved in the more elaborate examples we discuss below, where bond
data-points are introduced as well.

In the more general situation, the bond prices with a given set of tenors xi (i = 1, 2, · · · , r)
are observed to be B0xi = ηi .

In addition, we have the initial value x̄0 = ξ of the perpetual annuity.

Subject to these constraints, the maximum entropy term structure is determined by the
variational principle
 
δ r
−ρ(x) ln ρ(x) − λρ(x)x − µi ρ(x)Ixi (x) − νρ(x) = 0, (8.30)
δρ i=1

where Ixi (x) = 1 for x ≥ xi and vanishes otherwise.

The parameters λ, µi and ν are determined by the normalisation condition and data con-
straints
 ∞  ∞
xρ(x)dx = ξ, and Ixi (x)ρ(x)dx = ηi . (8.31)
x=0 x=0

The solution is
 
1 r
i
ρ(x) = exp −λx − µ Ixi (x) , (8.32)
Z(λ, µ) i=1

where
  
∞ 
r
Z(λ, µ) = exp −λx − µi Ixi (x) dx. (8.33)
0 i=1

74
The Lagrange multipliers are then determined implicitly by
∂ ln Z ∂ ln Z
− =ξ and − = ηi . (8.34)
∂λ ∂µi

As a consequence of (8.32) we see that pointwise calibration to the discount bond prices,
along with the information of the price of the annuity, gives a piecewise exponential term
structure density function.

If there is further information at our disposal, then that can also be included in the system
of constraints so that all available information is used efficiently in the calibration procedure.

We now consider in more detail the simple case where the observed data consist of two pieces
of information—the bond price P0T1 for a fixed maturity date T1 , and the value ξ = x̄0 of
the perpetual annuity.

This is a rather artificial example; nevertheless it serves to illuminate the main points of
the procedure.

The variational problem implies the existence of three rates r0 , r1 , and R such that the
term structure density is

r0 e−Rx for 0 ≤ x < T1
ρ(x) = (8.35)
r1 e−Rx for T1 ≤ x < ∞.

The constraints are given by:

 T1
ρ(x)dx = 1 − P0T1 (8.36)
0
for the bond price;
 ∞
ρ(x)dx = 1 (8.37)
0
for the normalisation; and
 ∞
xρ(x)dx = ξ (8.38)
0
for the perpetual annuity.

A short calculation shows that these relations reduce to:

r0

1− 1 − e−RT1 = P0T1 , (8.39)

R

75
 r1 − r0 −RT1 r0
e + = 1, (8.40)
R R


r1 − r0 −RT1 1 r0
e T1 + + 2 = ξ. (8.41)
R R R
Clearly, given P0T1 and ξ, we can proceed to infer values of r0 , r1 , and R.

In particular, equation (8.39) allows us to deduce the bond price P0T1 if we are given r0
and R, whereas we can use (8.40) to eliminate r1 in (8.41) to obtain
1  r0 
+ T1 1 − =ξ (8.42)
R R
for the value of the perpetual in terms of r0 and R.

Alternatively, given the initial short rate r0 and the value of the perpetual ξ we have
1 − r0 T1
R= . (8.43)
ξ − T1
This value of R can then be inserted in (8.39) to determine the bond price.

The scale factor r1 is given by

r1 = RP0T1 eRT1 . (8.44)

Thus we obtain

r0 e−Rx (0 ≤ x < T1 )
ρ(x) = (8.45)
RP0T1 e−R(x−T1 ) (T1 ≤ x < ∞),

for the term structure density, and



1 − rR0 1 − e−Rx (0 ≤ x < T1 )

P0x = (8.46)
P0T1 e−R(x−T1 ) (T1 ≤ x < ∞),

for the discount function, from which yield curve R0x can be constructed via the standard
prescription
ln P0x
R0x = − , (8.47)
x
and it should be evident by inspection that R0x is continuous in x.

In this example we can alternatively regard the short rate r0 and the bond price P0T1 as

76
the actual ‘independent’ data. Then (8.39) can be used to deduce R, which allows us to
infer the annuity price ξ by use of (8.42).

This illustrates the point that, although we assume from the outset the existence of a per-
petual, we can infer an implied value of that instrument by the use of other market data
(e.g., the short rate).

The same idea carries forward in the case where we have multiple data points for the bond
prices, for a given set of n maturity dates Tj (j = 1, 2, · · · , n), and we are led to a simple
iterative algorithm for determining the term structure in terms of the short rate and the
specified bond data points.
Proposition 8. Given a set of bond prices P0Tj (j = 1, 2, · · · , n) and the existence of
the value of the perpetual annuity, the maximum entropy term structure density function is

n
ρ(x) = ITk Tk+1 (x)rk e−Rx . (8.48)
k=0

Here T0 = 0, Tn+1 = ∞, ITk Tk+1 (x) = 1 if x ∈ [Tk , Tk+1 ) and vanishes otherwise, r0 is the
short rate, and
P0Tk − P0Tk+1
rk = R . (8.49)
e−RTk − e−RTk+1
The value of R is determined from equation (8.39).

The corresponding discount function P0x is given by

rk −RTk
P0x = P0Tk − (e − e−Rx ) (8.50)
R
for x ∈ [Tk , Tk+1 ).
Proof. To see this, we insert the piecewise exponential density function (8.48) into a
series of constraints of the form (8.36) for the bond prices, together with the normalisation
constraint (8.37) and the perpetual constraint (8.38).

Then the bond price constraints give rise to a set of relations of the form
rk −RTk
(e − e−RTk+1 ) = P0Tk − P0Tk+1 , (8.51)
R
for k = 0, 1, · · · , n − 1.

In particular, for k = 0, we recover (8.39), which can be used to solve for R in terms of

77
the short rate r0 and the bond price P0T1 .

Then, by substitution of this in (8.51) for general k, and the use of further bond price
data, we obtain the other rates rk (k = n).

As for rn , we note that if we divide the integration range in (8.36) into two regions [0, Tn ]
and [Tn , ∞], then the normalisation condition becomes
rn −RTn
e = P0Tn , (8.52)
R
which determines rn in terms of R and P0Tn .

Substitution of these results in the perpetual constraint


1
n
−RTk 1 r0
(rk − rk−1 )e Tk + + 2 =ξ (8.53)
R k=1 R R

allows the implied value ξ of the perpetual annuity to be determined from the short rate r0
and the bond price data P0Tj .

The discount function can be determined by use of the fact that

 x
1 − P0x = ρ(u)du
0
 Tk  x
= ρ(u)du + rk e−Ru du (8.54)
0 Tk
 x
= 1 − P0Tk + rk e−Ru du,
Tk

when x ∈ [Tk , Tk+1 ). ♦

Next, we turn to the problem: Given an existing term structure ρ2 (x) and a set of new
data points, how does one determine the new term structure that is ‘closest’ to the previous
one?

This can be addressed by use of the statistical J-divergence:

J(ρ1 , ρ2 ) = S 12 (ρ1 + ρ2 ) − 12 (S(ρ1 ) + S(ρ2 )) . (8.55)

Here, as before, the entropy is defined by

 ∞
S(ρ) = − ρ(x) ln ρ(x)dx. (8.56)
0

78
Statistical J-divergence defines the ‘separation’ between ρ1 and ρ2 .

The solution to the problem here is given by the ρ1 that minimises the J-divergence, subject
to the constraints:
 ∞
ρ1 (x) = 1 (8.57)
0

and
 ∞
ρ1 (x)dx = P0Ti . (8.58)
Ti

Introducing Lagrange multipliers µi , we must solve for

  ∞ 
δ n
J(ρ1 , ρ2 ) − µi ITi (x)ρ1 (x)dx = 0. (8.59)
δρ1 i=0 0

The solution can be written as:


n
1
ρ1 (x) = ITk Tk+1 (x) ρ2 (x). (8.60)
k=0
2 exp(δk ) − 1

Here, T0 = 0, Tn+1 = ∞, and


k
δk = −2 µi (8.61)
i=0

Let us write
 ∞
ρ2 (x)dx = Q0Ti , (8.62)
Ti

This is just the bond prices in the ‘old’ term structure.

To eliminate Lagrange multipliers, we note that

 Ti
ρ1 (x)dx = 1 − P0Ti , (8.63)
0

which implies
i−1 
 Tj+1
1 − P0Ti = ρ1 (x)dx
j=0 Tj


i−1  Tj+1
1
= ρ2 (x)dx. (8.64)
j=0
2 exp(δj ) − 1 Tj

79
If we also recall that
 Tj+1
ρ2 (x)dx = Q0Tj − Q0Tj+1 (8.65)
Tj

then the solution for this calibration problem can be summarised in the following form.
Proposition 9. Given a set of bond prices P0Tj (j = 1, 2, · · · , n) and an existing term
structure density ρ̂(x), the minimum J-divergence term structure density function is
 n 

ρ(x) = ITk Tk+1 (x)∆k ρ̂(x) (8.66)
k=0

Here T0 = 0, Tn+1 = ∞, ITk Tk+1 (x) = 1 if x ∈ [Tk , Tk+1 ) and 0 otherwise, and
P0Tk − P0Tk+1
∆k = (8.67)
Q0Tk − Q0Tk+1
The corresponding discount function P0x is
P0Tk − P0Tk+1
P0x = P0Tk − (Q0Tk − Q0x ) (8.68)
Q0Tk − Q0Tk+1

for x ∈ [Tk , Tk+1 ).

8.5 Canonical term structures∗

As an interesting example of a class of models that arises as a consequence of the maximisa-
tion of an entropy functional under constraints, we let the term structure density be of the
form
exp (−gtT − θt htT )
ρt (T − t) =  ∞ , (8.69)
u=t
exp (−gtu − θt htu ) du
where θt is a one-dimensional Ito process, and the functions gtT and htT are deterministic,
defined over the range 0 ≤ t ≤ T < ∞.

At each time t the term structure density thus defined belongs to an exponential family
parameterised by the value of θt . If we set
 ∞
Z(θ) = exp (−gtu − θt htu ) du, (8.70)
u=t

then we find that all the moments of the function htT can be determined from the generating
function Z(θ) by formal differentiation.

80
For example, for the first moment of htT we have
 ∞
∂ ln Z(θ)
htu ρt (u − t)du = − . (8.71)
u=t ∂θ
The corresponding bond price system can then be written in the Flesaker-Hughston form
∞
Ntu du
PtT = u=T
∞ , (8.72)
u=t
N tu du

where NtT = exp(−gtT − θt htT ). By Ito’s lemma, it follows that NtT satisfies

dNtT  
= − ġtT + θt ḣtT dt − htT dθt + 12 h2tT (dθt )2 , (8.73)
NtT

where the dot indicates partial differentiation with respect to t, so ġtT = ∂t gtT and ḣtT =
∂t htT .

We assume that the trajectory θt of the canonical parameter satisfies a stochastic equa-
tion of the form

dθt = αt dt + βt dWt . (8.74)

The no-arbitrage condition implies that NtT is a positive martingale. Therefore, the drift of
NtT vanishes for all T :

ġtT + ḣtT θt + αt htT = 12 βt2 h2tT . (8.75)

This relation implies that the processes αt and βt determining the dynamics of θt are of the
form
1 2
αt = At θt + Bt , and β
2 t
= Ct θt + Dt (8.76)

where the functions At , Bt , Ct and Dt are deterministic.

It follows that θt is a square-root process. Substitution of these equations into (8.75) gives
a set of Bernoulli equations of the form

ḣtT + At htT − Ct h2tT = 0 (8.77)

for htT and

ġtT + Bt htT − Dt h2tT = 0 (8.78)

81
for gtT .

The general solution of (8.77) is

 t 
 t 
1 Cu
= − exp Au du u du + ET , (8.79)
htT 0 0 exp( 0 Av dv)

where ET is an function of T , determined by the initial term structure.

To proceed further, let us consider the special case where Dt = 0 and θt is positive and
mean-reverting. Then Bt and Ct are both positive and At is negative, and for gtT we have
 t
gtT = − Bu huT du + FT , (8.80)
0

where FT is another arbitrary function. In the elementary case where At , Bt and Ct are
constants, the functions htT and gtT are given by
A
htT = (8.81)
C − GT eAt
and


B GT − Ce−At
gtT = ln − FT , (8.82)
C GT − C

where GT = AET + C. The condition that htT should be positive ensures that GT is of the
form GT = CHT e−AT where the function HT satisfies HT > 1 but is otherwise arbitrary.

For NtT we then obtain:

 BC

HT − eAT Aθt
NtT = exp − FT . (8.83)
HT − eA(T −t) C(HT e−A(T −t) − 1)

The function FT is then determined by the specification of the initial term structure for t = 0.

In particular, because N0T = ρ0 (T ), we obtain

 BC

HT − eAT Aθt Aθ0
NtT = ρ0 (T ) exp − . (8.84)
HT − eA(T −t) C(HT e −A(T −t) − 1) C(HT e−AT − 1)

82
Chapter 9
Review of the Flesaker-Hughston framework. Integral formulae for discount
bonds. Supermartingales and potentials. Rational log-normal model.

9.1 Risk-adjusted discount bond volatility

We return now to the general theory of interest rate dynamics, and establish another ex-
pression for the discount bonds, which we call the integral representation.

This representation has the advantage of bringing out the positive interest condition.

Recall that for the general arbitrage-free dynamics of a discount bond system we have the
following dynamics:
dPab
= ra da + Ωab (dWa + λa da). (9.1)
Pab
Here Pab is the value of a discount bond at time a that matures at time b, ra is the short
rate, Ωab is the bond volatility vector, and λa is the relative risk vector.

The economy is modelled by a probability space (Ω, F, P ) with filtration (Ft ). We as-
sume that (Ft ) is generated in a standard way by a multi-dimensional Brownian motion.

We recall the fact that under suitable technical conditions the solution to the dynamical
equation is

 a  a 
1 2
Pab = P0b Ba exp Ωsb (dWs + λs ds) − 2 Ωsb ds (9.2)
0 0
Here Ba is the unit-initialised money market account process.

The solution for Ba , obtained by setting Paa = 1, is

 a  a 
−1
Ba = (P0a ) exp − 1
Ωsa (dWs + λs ds) + 2 Ω2sa ds . (9.3)
0 0

83
For the short-term interest rate ra , we have
 a  a
ra = −∂a ln P0a + Ωsa ∂a Ωsa ds − ∂a Ωsa (dWs + λs ds). (9.4)
0 0

Putting these ingredients together (inserting (9.3) into (9.2)) we have the formula
 a a

exp 0 Ωsb (dWs + λs ds) − 12 0 Ω2sb ds

Pab = P0ab  a a
. (9.5)
exp 0 Ωsa (dWs + λs ds) − 12 0 Ω2sa ds

Here, P0ab = P0b /P0a denotes the forward value of a b-maturity bond.

Recall that P0ab is the value negotiated today for purchase at time a of a b-maturity bond.

It will be useful to build an analogy with the single asset situation.

In that case we recall that for the dynamics of a non-dividend paying asset St we have

 t  t 
1 2
St = S0 Bt exp σs (dWs + λs ds) − 2 σs ds . (9.6)
0 0

Then, introducing the density martingale, we deduce, under suitable technical conditions,
that the following ratio is a martingale:

 t  t 
Λt St 1 2
= exp (σs − λs ) dWs − 2 (σs − λs ) ds . (9.7)
Bt 0 0

In the case of interest rate dynamics, σs gets replaced by Ωsb , and a result similar to (9.7)
holds for each discount bond.

More specifically, we have

 a  a 
Λa Pab
= P0b exp Vsb dWs − 1
2
Vsb2 ds (9.8)
Ba 0 0

where

Vab := Ωab − λa . (9.9)

The quantity Vab , which we call “risk-adjusted volatility”, plays a useful role in the theory
of interest rates.

84
Note that Vab contains the information of both the discount bond volatility and the in-
terest rate market price of risk.

This is because of the constraint Ωaa = 0 (a maturing bond has zero volatility), which
implies that λa = −Vaa and that Ωab = Vab − Vaa .

Note that “risk premium” and “volatility” have the same units (inverse square-root time),
so it makes sense to combine them additively.

Now, setting Paa = 1, we obtain a formula for Λa /Ba , and for the discount bonds we get
 a a

P0b exp 0 Vsb dWs − 12 0 Vsb2 ds

Pab =  a a
. (9.10)
P0a exp 0 Vsa dWs − 12 0 Vsa2 ds

In this expression we note that, for each fixed value of b, the numerator is an exponential
martingale.

9.2 Integral representation for discount bonds

We have thus represented Pab as a quotient of the form
∆ab
Pab = , (9.11)
∆aa
where ∆ab is a one-parameter family of positive martingales.

Here the martingale property holds with respect to the “natural” probability measure P .

We make technical assumptions sufficient to ensure that the bond price goes to zero for
large values of the maturity, and that the martingale property of ∆ab is preserved under
differentiation with respect to the maturity parameter.

We find then that ∆ab can be expressed in the form

 ∞
∆ab = (−∂s P0s )Mas ds. (9.12)
b

Here Mas is a one-parameter family of martingales, initialised to unity at time zero (M0s = 1)
to ensure satisfaction of the initial condition ∆0b = P0b .

The argument that establishes the integral representation is as follows.

85
Since limb→∞ Pab = 0 by assumption, we have limb→∞ ∆ab = 0, and thus
 ∞
∆ab = − ∂s ∆as ds. (9.13)
b

By assumption, ∂s ∆as is a martingale.

Since ∆0b = P0b , it follows that

∂s ∆as
Mab = (9.14)
∂s P0s
is a unit-initialised martingale, and thus we obtain (9.12).

In particular, for a positive interest rate model it is necessary and sufficient that initial
interest rates are positive, and that the martingale family Mas should be positive.

With these ingredients in place, we see that the discount bond process can be written in the
form:
∞
(−∂s P0s )Mas ds
Pab = b∞ . (9.15)
a
(−∂s P0s )Mas ds
This is the “positive interest” integral representation for the general interest rate model.

By a “Flesaker-Hughston” model we usually mean any representation of the discount bonds

in the form (9.15) for some choice of the martingale family Mas .

One can verify by inspection that if initial interest rates satisfy the positivity conditions

0 < P0b ≤ 1 and ∂b P0b < 0. (9.16)

If the martingale family Mas is positive, then the positive interest conditions

0 < Pab ≤ 1 and ∂b Pab < 0 (9.17)

are satisfied for future valuation dates, and for bonds of all maturities.

9.3 Integral representations in the risk-neutral mea-

sure
A representation of the form (9.15) exists for any measure equivalent to the natural measure.

86
That is to say, in the absence of arbitrage, and with some technical conditions, given a
probability measure P̂ equivalent to the natural economic measure P , there exists a martin-
gale family Mas such that the discount bonds Pab are given by an integral representation of
the form (9.15).

In the case of a complete market the representation thus obtained is unique.

We note that if Mas represents the martingale family with respect to the natural proba-
bility measure P , then
Mas
M̂as = (9.18)
Λa
is the appropriate new martingale family, with respect to a new measure P̂ , where Λa is the
change-of-measure density martingale.

This is because if Mt is any martingale with respect to P , then Mt /Λt is a martingale

with respect to P̂ , where P̂ is defined in terms of conditional expectation by
Ea [Λb Xb ]
Êa [Xb ] = (9.19)
Λa
for any random variable which is measurable with respect to Fb .

Now by use of the risk neutral measure we have the bond valuation formula
 
1
Pab = Ba Êa . (9.20)
Bb
 a

Here Ba = exp 0 rs ds is the money market account.

By inspection we evidently have

 
1
∆ab = Êa . (9.21)
Bb
Therefore we deduce that the martingale family for the risk-neutral measure is
 
rs
M̂as = Êa . (9.22)
(−∂s P0s ) Bs
As a consequence, we see that for the natural measure we have
Mas = Λa M̂as
 
rs
= Λa Êa
(−∂s P0s ) Bs
 
Λs rs
= Ea (9.23)
(−∂s P0s ) Bs

87
This gives us a construction for the martingale family Mas , given rs and Λs together with
initial bond data.

9.4 Potentials and positive supermartingales

Let us return now to the representation of the discount bonds given by
∆ab
Pab = . (9.24)
∆aa
For ∆ab here, we have

 a  a 
∆ab = P0b exp Vsb dWs − 1
2
Vsb2 ds , (9.25)
0 0

where

Vsb = Ωsb − λs . (9.26)

For ∆aa we have

Λa
∆aa = . (9.27)
Ba
We note that ∆ab = Ea [∆bb ].

The quantity Vt = ∆tt is the state-price density.

The state-price density satisfies the following differential equation:

dVt
= −rt dt − λt dWt (9.28)
Vt
Thus we see that if rt is positive, then Vt is a positive supermartingale.

Now as a consequence of (9.24) we have

E[Vt ]
P0t = . (9.29)
V0
Thus to ensure that the initial discount function vanishes asymptotically, we require

lim E[Vt ] = 0. (9.30)

t→∞

88
A positive supermartingale Vt that satisfies (9.30) is called a potential.

We see therefore that the concept of a potential is mathematically very natural as a ba-
sis for interest rate theory.

In the potential method we represent the bond price by a formula of the following form:
Ea [Vb ]
Pab = . (9.31)
Va
The potential method can be used to generate a number of new and potentially interesting
interest rate models.

There are several ways of representing potentials.

One method is to introduce strictly increasing adapted process At defined for all time
0 ≤ t ≤ ∞, and write
Vt = Et [A∞ ] − At (9.32)

If we write At in the form

 t
At = ηs ds (9.33)
0

where ηs is positive, then clearly

 ∞   t
Vt = Et ηs ds − ηs ds
0 0
 ∞ 
= Et ηs ds
 ∞ t
= Et [ηs ]ds. (9.34)
t

Now, if we define P0t according to (9.29), clearly we have

∞
E[ηs ]ds
P0t = t∞ , (9.35)
0
E[ηs ]ds
Therefore, for the derivative of P0t we obtain
E[ηt ]
−∂t P0t =  ∞ . (9.36)
0
E[η s ]ds

89
If we define
Et [ηs ]
Mts = . (9.37)
(−∂t P0t )

we obtain
 ∞
Vt = (−∂s P0s )Mts ds (9.38)
t

where Mts is a unit-initialised positive martingale family, and we are back to the Flesaker-
Hughston representation.

Another useful way to represent potentials is by the introduction of a square-integrable

random variable X∞ satisfying
 2
E X∞ < ∞. (9.39)

Then we define the martingale

Xt = Et [X∞ ] (9.40)

and write

Vt = Et (X∞ − Xt )2 (9.41)

In other words, Vt is defined to be the conditional variance of X∞ , given information up

to t.

We recall that for any random variable X, the conditional variance of X with respect to Ft
is defined by

Vart [X] = Et [(X − Et [X])2 ]. (9.42)

One can check that Vt is a supermartingale, and that E[Vt ] → 0 as t → ∞.

In this approach the entire interest rate framework is captured in the specification of a
single random variable X∞ . We shall have more to say about such conditional variance
framework shortly.

90
9.5 Rational Models
Now suppose we let the positive martingale family Mab be of the form

Mab = αb + βb Ma (9.43)

where αb and βb are positive deterministic functions satisfying αb + βb = 1, and Ma is any

positive martingale, normalised so that initially we have M0 = 1.

Then a short calculation shows that

Fb + Gb Ma
Pab = , (9.44)
Fa + Ga Ma
where Fb and Gb are positive decreasing functions, satisfying

Fb + Gb = P0b (9.45)

where P0b is the initial discount function.

Inspection shows that Pbb = 1, 0 < Pab ≤ 1, and ∂b Pab < 0, the positive interest condi-
tions.

This is the so-called rational model (Flesaker & Hughston 1996).

If Ma is chosen, for example, to be a geometric Brownian motion, then we obtain the rational
log-normal model.

In the extended rational log-normal model we have

 a  a 
1 2
Ma = exp σ(s) dWs − 2 σ (s) ds (9.46)
0 0

where σ(s) is deterministic.

This model is one of the simplest of all interest rate models.

It admits completely analytic formulae for the valuation of caps, floors and swaptions of
all maturities.

A short calculation shows that the short rate, in the case of a general rational model, is
given by
F  (t) + G (t)Mt
rt = − . (9.47)
F (t) + G(t)Mt

91
It is not difficult to show then, in the case of the extended rational log-normal model , that
rt is a diffusion.

In other words, in the RLN model rt satisfies a stochastic equation of the form

drt = δ(t, rt ) dt + γ(t, rt ) dWt (9.48)

where δ(t, r) and γ(t, r) are each deterministic functions of two variables.

It is an interesting exercise to show in this case that γ(t, r) is a quadratic polynomial in

the short rate.

The two positive roots to this equation correspond to (time dependent) upper and lower
bounds on the interest rate process.

The RLN model is an important example of a completely tractable system of interest rate
dynamics exhibiting many desirable qualitative features.

A relatively complete analysis of the valuation of caps and swaptions in the rational log-
normal model has been given by Musiela & Rutkowski (1997).

92
Chapter 10
Multi-currency interest rate dynamics. Compatible exchange rate systems.
Geometric analysis of foreign exchange volatility and correlation. Quanto
effects. International models for interest rates and foreign exchange.

10.1 Interest rate and foreign exchange dynamics

Let us consider the problem of constructing an extension of the basic HJM framework suit-
able for the valuation of interest rate and foreign exchange derivatives.

We consider an international economy consisting of a set of n currencies, and for each

currency a family of discount bonds denominated in that currency.

For such an economy it is possible to deduce a set of formulae for the price processes of
these discount bonds and the associated exchange rates, subject to the conditions of no ar-
bitrage.

In the multi-currency situation we do not wish to single out any preferred currency.

So we work with the natural measure P , and transform to the risk neutral measure as-
sociated with a choice of currency only for special applications.

In the multi-currency situation there is a numeraire process associated with each currency,
and these are all related to one another via the exchange rate process.

If Stij denotes the price of one unit of currency i in units of currency j (e.g., the price
of one pound sterling in dollars), then the relation is given more specifically by
ξti Stij = ξtj , (10.1)
where ξti denotes the price process for the numeraire asset, expressed in units of currency i.

93
The effect of the no arbitrage condition on the international interest rate and foreign ex-
change markets is to ensure the existence of a “global” numeraire asset, the value of which
can be expressed in any currency.

If the market is complete, then the global numeraire is completely determined by the given
asset processes. More generally, we simply assume the existence of a global pricing kernel.

The global numeraire has the property that the ratio of the value (in a given currency)
of any nondividend-paying tradable asset to the value (in the same currency) of the global
numeraire is a martingale with respect to natural measure.

An arbitrage-free complete system of interest rates and foreign exchange is called an “Amin-
Jarrow” economy.

i
Let us write Pab for the value (in units of currency i) at time a (time 0 is the present)
of a default-free discount bond that matures at time b to deliver one unit of currency i.

We shall write Bai for the value at time a (in units of currency i) of a money market account
for currency i, initialised to one unit of currency at time 0.

The money market account for currency i can be expressed in terms of the short rate rsi
for the currency by the formula

 a 
i i
Ba = exp rs ds . (10.2)
0

We shall write λia for the risk premium vector for currency i.

Thus λia determines the excess rate of return (above the short rate in currency i), per unit
of volatility, for assets denominated in that particular currency.

For the discount bonds we have the following dynamics:

i
dPab
i
= rai da + Ωiab (dWa + λia da). (10.3)
Pab

Here Ωiab is the discount bond volatility vector for currency i.

For any given value of i, the dynamics (10.3) look much like the bond dynamics we have
already considered.

However, in the present context, the multi-dimensional Brownian motion Wtα drives the

94
whole economy, and λiα
t is the risk premium vector for any asset denominated in that cur-
rency.

Now let us consider the process Saij for the exchange rate. This must of course be of the
form
dSaij ij ij
ij = µa da + νa dWa , (10.4)
Sa
where µij ij
a is the drift and νa is the volatility vector.

In a complete market with no arbitrage, the drift and volatility take the following remarkable
form:

µij i j j i j
a = ra − ra + (λa − λa ) · λa (10.5)

and

νaij = λja − λia . (10.6)

10.2 Compatible exchange rate systems∗

Suppose we have an n-by-n matrix Saij of positive Itô processes based on a multi-dimensional
Brownian motion Waα .

We assume that Saij satisfies the compatibility conditions

Saij Sajk = Saik , (10.7)

and that Saii = 1.

It follows that Saij = 1/Saji , and that Saii = 1.

We call such a set of processes a “compatible exchange rate system”.

What constraints does the form (10.7) place on the resulting exchange rate dynamics?

For any compatible exchange rate system there exists a set of positive processes ξai such
that

Saij = ξaj /ξai . (10.8)

The proof of this follows if we write (10.7) in the form in the form Saij = Sakj /Saki . Fixing a
value k, we define ξai = Saki for all i.

95
This is well-defined procedure since Saij is always positive, and shows that Saij splits into
a quotient.

The split (10.8) is not unique, since it is invariant under the transformation ξai → πa ξai
for any positive process πa .

We shall investigate the consequences of this freedom later.

Suppose we therefore write

dξai
= Rai da + λia dWa (10.9)
ξai

for the stochastic equation satisfied by ξai , for some given choice of ξai .
2
Without loss of generality we define the process rai by setting Rai = rai + λix . Then we
have
dξai
= rai da + λia (dW + λia da). (10.10)
ξai

A short calculation making use of the Itô quotient relation shows that

dSaij j i ij j
ij = (ra − ra ) da + νa (dWa + λa da). (10.11)
Sa
The exchange rate volatility νaij is thus given as indicated earlier by

νaij = λja − λia . (10.12)

Clearly we have νaij = −νaji . Thus we see that the splitting of νaij to a difference of two vector
processes arises from the compatibility condition.

Recall that Saij is the price of one unit of currency i in units of currency j.

Thus the volatility vector for the price of sterling in dollars is minus the volatility vec-
tor for the price of one dollar in sterling.

Now a currency is not a non-dividend paying asset.

The “dividend” earned by currency i is the interest it continuously accumulates in a money

market account.

96
Thus in (10.11) the overall drift in the value of currency i is of the form

µij j i ij j
a = r a − ra + νa · λ a . (10.13)

This is given by the risk-free rate on the valuation currency j, less the “dividend yield” rai ,
plus the excess rate of return.

The excess rate of return is given by the inner product of the volatility vector for cur-
rency i (when priced in units of currency j) times the relative risk vector for the valuation
currency.

10.3 Geometric analysis of FX volatility

The volatility vectors for a compatible foreign exchange rate system have to fit together to
form a polytope in the multidimensional Euclidean space in which the Wiener process takes
its values.

This is on account of the relation νij + νjk + νki = 0.

Thus for three currencies we have a triangle, four give a tetrahedron, and so on.

In that picture the vertices of the figure correspond to currencies, and the length of the
edge joining two given vertices is the magnitude of the instantaneous volatility of the asso-
ciated exchange rate.

The cosine of the angle between two edges (whether they intersect or not a common vertex)
measures the instantaneous correlation between the movements in the given exchange rates.

The relation νtij = λjt − λit allows one to take this set of ideas a step further, incorpo-
rating the relative risk into the picture.

In particular, if we fix an origin, then the relation νtij = λjt − λit shows us that the sys-
tem of risk premium vectors for the various currencies, viewed as emanating from the origin,
determines the location and structure of the volatility ‘polytope’.

10.4 Scale transformations∗

Now suppose that we are just given the process Saij . To what extent does this process deter-
mine rai and λia in a complete market free of arbitrage?

97
 Y

θ
$
DM

L
Figure 10.1: Four currency tetrahedron. The six edge-length correspond to the volatilities of
the six exchange rates for the four given currencies. The angles between edges determine the
corresponding correlations.

The exchange rate volatility νaij given by (10.12) is invariant under the transformation

ξai → πa ξai , (10.14)

since the exchange rate Saij itself is left unchanged.

Under the scale transformation (10.14) we find, after a short calculation, that the risk pre-
mium and short rate transform as follows:

λia → λia + Ψa (10.15)

rai → rai + Φa − Ψ2a − λia Ψa . (10.16)

Here the vector process Ψa and the scalar process Φa are defined by
dπa
= Φa da + Ψa dWa . (10.17)
πa
The process λia is thus determined by the exchange rate system up to a transformation of
the form (10.15) for an arbitrary vector process Ψa .

Geometrically, this can be pictured as a translation of the entire volatility polytope in the
direction given by Ψa .

One can think of such transformations as representing “global” change in the international
economy. For example, one might have an overall drop in interest rates coupled with a
general change in risk aversion as regards some particular source of risk.

Once λia is fixed, then the interest rates are determined up to an overall change of level
Φa .

98
10.5 “Quanto” effects∗
It is worth noting the effect that a transformation to the risk neutral measure associated
with a given “domestic” currency j has on the bond price process for a “foreign” currency.

The process for the domestic bond price, when we transform to the risk neutral measure,
becomes
j
dPab j j j
j = ra da + Ωab dWa , (10.18)
Pab

where dWaj = dWa + λja da.

The bond process for the “foreign” currency i, which in the original measure is given by
i
dPab
i
= rai da + Ωiab (dWa + λia da), (10.19)
Pab

transforms to
i
dPab
i
= rai da + Ωiab (dWaj − νaij da), (10.20)
Pab

when expressed in terms of Waj , which is a Brownian motion in the risk neutral measure
associated with currency j.

Note the appearance of νaij = λja − λia , the foreign exchange volatility vector, in this for-
mula.

The “quanto” correction term appearing here involves the inner product of the foreign dis-
count bond volatility vector and the exchange rate volatility vector.

This can be re-expressed in more familiar terms as a product of the bond volatility level, the
foreign exchange volatility level, and a correlation factor.

10.6 Martingale representation for FX and interest rate

systems∗
An Amin-Jarrow economy is completely characterised by a set of n one-parameter families
i
of unit-initialised martingales denoted Mas , along with a set of initial term structure data
P0s for each currency, and a set of initial exchange rates S0ij .
i

99
We require that the initial exchange rates are compatible in the sense that S0ij S0jk = S0ik
(e.g. the price of sterling in dollars times the price of one dollar in yen gives the price of
sterling in yen).

i
For positive interest rates we require, in addition to the above, that the martingales Mas are
i
strictly positive, and that the initial discount functions P0s exhibit positive interest in the
sense that
i i
0 < P0b ≤ 1 and ∂b P0b <0 (10.21)

for all maturities. Here ∂b denotes differentiation with respect to b.

For the discount bonds in currency i we have an integral representation of the form
∞ i i
i (−∂s P0s )Mas ds
Pab = b∞ i i
. (10.22)
a
(−∂s P0s )Mas ds
i
Here again Pab denotes the value (in units of currency i) at time a (time 0 is the present) of
a discount bond that matures at time b to deliver one unit of currency i.

Note that each discount bond is valued in its “own currency”.

The system of exchange rates is then given by

∞ i i
ij ij a
(−∂s P0s )Mas ds
Sa = S0  ∞ j j . (10.23)
a
(−∂s P0s )Mas ds

Taking into account the given initial conditions, it follows that the compatibility conditions

Saij Sajk = Saik (10.24)

are satisfied.

The numeraire process, which in currency i has the value ξai , is given by

ξ0i
ξai =  ∞ i i ds
, (10.25)
a
(−∂s P0s )Mas

where initial values ξ0i are such that

S0ij = ξ0j /ξ0i . (10.26)

The existence of such a system of initial values is ensured by the initial compatibility condi-
tions on the exchange rates.

100
 i
The basic martingales Mas are defined for all s ≥ a ≥ 0 (up to some time horizon), and
satisfy
i i
Ea Mbs = Mas . (10.27)

There is an explicit formula for the risk premium vector for each currency, given by
∞ i i i
(−∂s P0s )Mas σas ds
λia = − a ∞ i i ds
. (10.28)
a
(−∂s P0s )Mas
i
Here the vector process σas is defined by
i i i
dMas = Mas σas dWa , (10.29)

The discount bond volatilities are given in terms of the basic martingales according to the
scheme

Ωiab = Vabi − Vaa

i
, (10.30)

where the “risk adjusted” volatility Vabi is given by

∞
i ∂ P M i σ i ds
b s 0s as as
Vab = ∞ i ds
. (10.31)
b
∂s P0s Mas

Thus λia = −Vaa

i
, consistent with equation (10.28).

The short rate is given by

i i
∂a P0a Maa
rai =  ∞ i i ds
. (10.32)
a
(−∂s P0s )Mas

With this information at hand we can verify again that the cross-currency process Stij satisfies

dSaij j i j i j
ij = (ra − ra ) da + (λa − λa ) (dWa + λa da). (10.33)
Sa
We shall return to the matter of international interest rate and foreign exchange systems in
greater depth in due course.

101
Chapter 11
Axiomatic framework for continuous asset price dynamics. Perpetual floating
rate notes. Price processes for discount bonds. Dynamics of the state price
density.

11.1 Axiomatic framework for continuous asset price

dynamics
The idea now is to develop an axiomatic scheme that will ensure the existence of an arbitrage-
free system of discount bonds over all time horizons, but that is general enough also to allow
a place for other systems of assets.

The methodology that we propose, which in effect unifies a number of important features
of the theory of interest rate modelling and the theory of volatility modelling, is based on a
conditional variance representation for the state price density, and makes use of the Wiener
chaos expansion technique in a novel way.

We model the unfolding of random market events in the usual way with the specification of
a fixed probability space (Ω, F, P ) which we denote as Π.

We assume that the economy Π is equipped with the standard augmented filtration Φ =
(Ft )0≤t≤T ∗ generated by a system of one or more independent Wiener processes (Wtα )0≤t≤T ∗
(α = 1, · · · , k).

Here T ∗ represents a fixed time horizon, which for the moment we leave unspecified but
eventually will be assumed to be infinite.

The probability measure P is to be interpreted as the “natural” measure, and filtration-

dependent concepts (such as adaptedness or the martingale property) are defined relative to
Φ.

102
We assume in this investigation that the random processes on Π followed by asset prices
are continuous semimartingales adapted to Φ.

The absence of arbitrage in the economy will be characterised according to the following
scheme.

We assume the existence of a continuous semimartingale ξt , adapted to Φ, which we call

the “natural numeraire” process, satisfying ξt > 0 for all t ∈ [0, T ∗ ], such that the following
three axioms hold:

(A1) There exists a strictly increasing (and hence “risk-free”) asset with price process Bt
(the money-market account).

(A2) If St is the price-process of any asset, and Dt is the adapted dividend rate for that asset,
so that Dt dt represents the small random dividend paid at time t, then the process Mt
defined by  t
St Ds
Mt = + ds
ξt 0 ξs
ia a martingale.

(A3) There exists an asset (a floating rate note) that offers a dividend rate sufficient to
ensure that the value of the asset remains constant.

Now let us examine some of the consequences of these axioms.

Existence of risk adjustment density

Since the process Bt introduced in (A1) is by assumption continuous and strictly increasing,
there exists an adapted process rt > 0 such that

 t 
Bt = B0 exp rs ds . (11.1)
0

Because the money market account is a non-dividend paying asset, it follows as a consequence
of (A1) and (A2) that there exists a positive martingale Λt such that

Bt
= Λt . (11.2)
ξt
Since Λt is positive, there exists an adapted vector-valued process λt such that

dΛt = −Λt λt dWt , (11.3)

103
where here, and similarly elsewhere, we use the shorthand


k
λt dWt = λαt dWtα . (11.4)
α=1

As a consequence of (11.3), we then have

 t  t 
1 2
Λt = ρ0 exp − λs dWs − 2 λs ds . (11.5)
0 0

Uniqueness of the money market account

At most one process Bt can exist satisfying axioms (A1) and (A2). For if Bt∗ were another
such increasing price process, then we would have
Λt ρ∗
= t∗ (11.6)
Bt Bt
for some positive martingale ρ∗t . But this relation implies that
dΛt dρ∗
= (rt − rt∗ )dt + ∗t (11.7)
Λt ρt
which shows that for Λt and ρ∗t both to be martingales we have rt = rt∗ .

Dynamic equations for risky-assets

Axiom (A2) implies, in the case of a non-dividend-paying asset, that St can be written in
the form
Bt Mt
St = (11.8)
Λt
where Mt is a martingale.

Thus, if we write dMt = θt dWt it is a straightforward exercise to verify that

dSt = (rt St + λt ψt )dt + ψt dWt , (11.9)

where the vector-valued process ψt is defined by

Bt θt
ψt = + λ t St . (11.10)
Λt
In particular, if the asset price St is positive, then Mt is positive, and we can write θt =
(σt − λt )Mt for some vector-valued process σt , from which it follows that ψt = σt St .

104
In that case the dynamical equation satisfied by St can be written in the form
dSt
= (rt + λt σt )dt + σt dWt , (11.11)
St
where σt is the adapted vector-valued volatility process for the given asset, and λt has the
interpretation of the market risk premium.

We recognise (11.11) as the dynamics of a risky asset with limited liability in a market
with no arbitrage.

However the dynamical equation

dSt = (rt St + λt ψt )dt + ψt dWt , (11.12)

has the advantage of holding in the more general situation for assets such as portfolio po-
sitions including borrowing, short sales, or derivatives, where the value of the position may
swing into the red as well as the black.

Risky assets with dividend

In the case of a dividend paying asset these formulae need to be modified slightly, and in
place of (11.12) we obtain

dSt = (rt St − Dt + λt ψt )dt + ψt dWt (11.13)

as a consequence of (A2), with ψt defined as before according to ψt = Bt θt /Λt + λt St .

Then if St is positive we can introduce a proportional dividend rate δt by the relation

Dt = δt St , and we obtain the simplified expression
dSt
= (rt − δt + λt σt )dt + σt dWt , (11.14)
St
where σt is defined as before by ψt = σt St .

Clearly, (11.14) conforms to the familiar dynamics of a dividend or interest paying asset
with limited liability.

For example, if St is the price of a foreign currency, then δt corresponds to the overnight rate
for that currency. We consider the case of a foreign currency in greater detail later.

105
Assets of constant value
Now let us examine axiom (A3) more closely. Such a “cash” asset that maintains a constant
value has the interpretation of being a floating rate note.

Equation (11.14) shows that if we set St = 1 for all t ∈ [0, T ∗ ] then the “dividend” rate
offered for such an instrument must be rt . It follows that
 t
1 rs
+ ds is a martingale. (11.15)
ξt 0 ξs

In particular since rt and ξt are positive we deduce that

 
1
E <∞ (11.16)
ξt

and
 t 
rs
E ds < ∞ (11.17)
0 ξs

for all t ∈ [0, T ∗ ].

11.2 Price processes for discount bonds

To proceed further we introduce a system of discount bonds on the economy Π.

More precisely, this will be the discount bond system associated with the base currency
in terms of which the other assets on Π are priced and with respect to which the money
market process Bt is defined.

The discount bond price processes will be denoted PtT , where 0 ≤ t ≤ T ≤ T ∗ ≤ ∞.

We shall as usual regard the zero-coupon bond for a given value of T as a default-free
contract that pays one unit of the base currency at time T .

Then PtT denotes the price of the bond at time t, and by the definition of the contract
we require that PT T = 1 for all T ∈ [0, T ∗ ].

For the moment we make no other assumptions concerning the discount bond processes
other than those properties applicable to all assets implicit in axioms (A1), (A2), and (A3),
though later we add a further important assumption concerning the asymptotic behaviour
of the bond prices in the case of an infinite time horizon.

106
Since PtT represents the price process of a non-dividend-paying asset for each value of
T ∈ [0, T ∗ ], it follows from axiom (A2) that PtT /ξt is a martingale, and hence that there
exists a family of positive martingales MtT such that
Bt MtT
PtT = . (11.18)
Λt
Because MtT is a positive martingale for each bond maturity date T ∈ [0, T ∗ ], there exists a
vector-valued process ΩtT such that
dMtT
= (ΩtT − λt )dWt . (11.19)
MtT
We thus that the dynamics of the discount bond system are given by
dPtT
= (rt + λt ΩtT )dt + ΩtT dWt . (11.20)
PtT
We recognise ΩtT as being the T -maturity discount bond vector relative volatility process.

It then follows, by integrating (11.20), if we make use of the relation Ptt = 1, that the
discount bond price processes can be represented in the form
 t t 
t
exp 0 λs ΩsT ds + 0 ΩsT dWs − 12 0 Ω2sT ds
PtT = P0tT  t t  , (11.21)
t
exp 0 λs Ωst ds + 0 Ωst dWs − 12 0 Ω2st ds

and that the money market account process is given by a corresponding expression of the
form
B0
Bt =  t t . (11.22)
t 1
P0t exp 0
λs Ωst ds + 0
Ωst dWs − 2 0
Ω2st ds

Here we have used the notation P0tT = P0T /P0t for the t-forward price made at time 0 for a
T -maturity discount bond.

The volatility structure approach

An interesting feature of the expressions (11.21) and (11.22) is that the discount bond sys-
tem and the money market account can be represented directly in terms of the market risk
premium process λt and the bond volatility process ΩtT , together with the initial discount
function P0t , without direct reference to the short rate rt .

107
It is therefore legitimate to regard λt and ΩtT as being subject to an exogenous specifi-
cation.

Indeed, historically this observation is of considerable significance since it forms the ba-
sis of the approach to interest rate derivatives pricing frequently used in practice according
to which one “models the volatility structure”.

In such an approach one typically assumes market completeness, then transforms to the risk
neutral measure to eliminate the market risk premium, and then models the bond volatility
process exogenously, calibrating it to a suitable given set of market interest rate option data.

It has been a problematic feature of the volatility approach, however, that if λt and ΩtT
are specified exogenously, then there is no guarantee that axiom (A1) is satisfied—that is to
say, the resulting interest rates need not be positive.

Additionally, there is no reason to suppose, a priori, that the bond volatilities will take
on a given form in the risk neutral measure.

Let us therefore put to one side the “volatility structure” approach, and return to the
consideration of the assumptions (A1), (A2), and (A3) in the context of a term structure
model.

Martingale relations
Because the discount bonds are non-dividend-paying assets, it follows as a consequence of
(A2) that the martingale relations
 
PtT
E <∞ (11.23)
ξt
and
 
PtT PuT
= Et (11.24)
ξt ξu
hold for all 0 ≤ t ≤ u ≤ T ≤ T ∗ .

Here Et [−] denotes as usual the conditional expectation with respect to the σ-algebra Ft .

It follows from (11.23) by setting t = T that the existence of the discount bond system
implies that the inequality
 
1
E <∞ (11.25)
ξt

108
holds for all t ∈ [0, T ∗ ].

It is interesting to note, as was shown by Baxter (1997), that the inequality (11.17) is
the additional assumption required to ensure the differentiability of the bond price system
with respect to the maturity date.

Instantaneous forward rates

In other words, as a consequence of (11.17) there exists a family of continuous semimartin-
gales ftu , adapted to Φ, for all 0 ≤ t ≤ u ≤ T ∗ , such that

 T 
PtT = exp − ftu du . (11.26)
t

It then follows that

−∂T ln PtT = ftT , (11.27)
where ∂T denotes differentiation with respect to T , and also that
lim ftT = rT (11.28)
t→T

and
lim ΩtT = 0. (11.29)
t→T

The importance of the existence of the instantaneous forward rates is that the class of
interest rate models under consideration here is equivalent to the family of all positive in-
terest HJM models (Heath, Jarrow and Morton 1992) defined over the relevant time horizon.

We take the view here nevertheless that the instantaneous forward rates are in some sense
secondary, and that primary significance should be attached to modelling the natural nu-
meraire process ξt .

Risk neutral valuation formula

In particular, setting u = T in (11.24) we obtain the pricing formula
 
1
PtT = ξt Et . (11.30)
ξT
Thus, once axioms (A1), (A2), and (A3) have been specified, the associated discount bond
system is also determined.

We note that PtT is unchanged if we multiply ξt by a positive constant.

109
11.3 Dynamics of the state price density
It is be useful now to introduce the related process Vt = 1/ξt which has the interpretation
of being the state price density.

It follows from equation (11.1) that Vt = Λt /Bt , and from (11.16) we have E[Vt ] < ∞
for all t ∈ [0, T ∗ ].

In particular, since Bt is Ft -measurable and increasing we deduce that

   
ΛT ΛT Et [ΛT ] Λt
Et [VT ] = Et < Et = = = Vt , (11.31)
BT Bt Bt Bt
for t < T .

In other words, we have Et [VT ] < Vt , and thus we see that Vt is a supermartingale.

Now writing the risk neutral valuation formula in the form

 
VT
PtT = Et , (11.32)
Vt
we see that PtT < 1 for all t < T .

Pricing kernel
The quotient KtT = VT /Vt can be regarded as a “pricing kernel” for derivatives (Constan-
tinides 1992).

In particular, suppose that Ht is for t ∈ [0, T ] the price process of a derivative asset on
Π with a European-style payoff HT at time T .

Then by (A2) we have

Ht = Et [KtT HT ] , (11.33)
a relation that remains valid independently of any hedgeability considerations.

Note that no assumption of market completeness is made in our axiomatic scheme.

Properties of the state price density

It follows from the dynamical equations for Bt and Λt that the dynamics of Vt are given by
dVt = −rt Vt dt − λt Vt dWt . (11.34)

110
Therefore, given Vt we can recover the short rate rt and the market risk premium process λt .

Integrating (11.34) from t to T we get

 T  T
VT = Vt − rs Vs ds − λs Vs dWs . (11.35)
t t

Taking the conditional expectation of each side of (11.35) we obtain

 T 
Et [VT ] = Vt − Et rs Vs ds . (11.36)
t

Dividing by Vt we then arrive at the formula

 T 
PtT = 1 − Et Kts rs ds , (11.37)
t

which has a natural economic interpretation from which a number of interesting consequences
can be deduced.

It follows for example as a corollary of (11.37) that for any two maturity dates T1 and
T2 we have
 T2 
PtT1 − PtT2 = Et Kts rs ds . (11.38)
T1

Therefore if T2 > T1 , we deduce that PtT2 < PtT1 , and hence that the random forward price
PtT2
PtT1 T2 = , (11.39)
PtT1
made at time t for purchase at time T1 of a T2 -maturity discount bond satisfies

0 < PtT1 T2 ≤ 1 (11.40)

for all 0 ≤ t ≤ T1 ≤ T2 < ∞.

This in turn implies the positivity of all forward rates.

Interpretation of the instantaneous forward rates

Another interesting corollary of (11.37) follows if we differentiate each side of this equation
with respect to T , from which we deduce that

ftT PtT = Et [KtT rT ] . (11.41)

111
This relation shows that the instantaneous forward rates can be interpreted as the value, at
time t, future-valued to time T , of the contingent claim that pays the short rate rT at time
T on a unit principal.

It follows that the term structure density ρt (x) for tenor x = T − t is the value at time
t of an instrument that pays the rate rT at time T on a unit principal.

Equation (11.37) says that ownership of a T -maturity discount bond is equivalent to own-
ership of one unit of the cash asset, but without the right to the dividend flow of the cash
asset from time t to time T .

To put the matter in another way, a money-lender will be willing at time t to part with
one unit of cash in exchange for a discount bond maturing at time T together with a con-
tinuous flow of interest from time t to time T .

Equivalently, to hold a T -maturity floating-rate note is the same as holding a T -maturity

discount bond together with the right to a continuous stream of interest from time t to T .

112
Chapter 12
The conditional variance representation for the state price density. Interest
rate models as elements of L2 (Ω, F, P ). Elements of Wiener chaos. First
chaos models.

12.1 The conditional variance representation

Now suppose we consider the case of an interest rate system with an infinite time horizon
T ∗ = ∞. It follows from (11.37) that
 T 
P0T = 1 − E K0s rs ds . (12.1)
0

This relation can be interpreted as saying that the value of a T -maturity discount bond at
time 0 is one unit of cash less the present value of the interest stream from time 0 to time
T.

The idea is that by holding the discount bond one forgoes the dividends associated with
the cash until the maturity date of the bond—at which point one acquires the cash.

On the role of potential

The ownership of a discount bond that never matures (i.e. matures at T = ∞) is equivalent
to ownership of a unit of floating rate note stripped of its interest stream for all time—in
other words, the ownership of nothing.

As a consequence we conclude that

lim P0T = 0, (12.2)
T →∞

or equivalently
 ∞ 
V0 = E rs Vs ds . (12.3)
0

113
Indeed, we shall now take it as part of the definition of a discount bond system that T ∗ = ∞
and that the natural numeraire ξt and the interest rate rt are such that (12.2) holds, or
equivalently
 ∞ 
rs
ξ0 E ds = 1. (12.4)
0 ξs

Alternatively, it follows from (11.32) that the asymptotic condition (12.2) holds if and only
if

lim E [VT ] = 0. (12.5)

T →∞

This is the condition that the process Vt is a “potential”, i.e. a positive supermartingale
with the property that its expectation vanishes in the limit.

Thus, as was pointed out by Rogers (1997), it should be regarded as an essential element of
interest rate theory that the state price density should have this property.

Recursive relation of the state price density

We see therefore that once an appropriate asymptotic condition has been placed on the
discount bond system we have the key relation
 ∞ 
Vt = Et rs Vs ds . (12.6)
t

This formula has the economic interpretation that a floating rate note that promises to pay
the rate rt on a unit principal in perpetuity necessarily has the value unity.

An alternative expression for Vt can be deduced from (12.6) if we define the increasing
process
 t
At = rs Vs ds. (12.7)
0

Then we obtain the relation Vt = Et [A∞ ] − At as discussed earlier.

This forms the basis of the Flesaker-Hughston framework and its extensions (see, e.g., Fle-
saker and Hughston 1996, 1997, 1998, Rutkowski 1997, Musiela and Rutkowski 1997, Rogers
1997, James and Webber 2000, Hunt and Kennedy 2000, Jin and Glasserman 2001).

In the present investigation, we take an alternative point of view and emphasize a rather
different feature of the state price density that emerges in this context: namely, that Vt can

114
be interpreted as a conditional variance.

This makes use of an idea appearing in Meyer (1966). More precisely, let σt be a vector
process satisfying

σt2 = rt Vt . (12.8)

Then we can define a random variable X∞ by the formula

 ∞
X∞ = σs dWs . (12.9)
0

The existence of X∞ is guaranteed by virtue of axiom (A3) which implies that

 ∞ 
E rs Vs ds < ∞. (12.10)
0

It follows immediately then by virtue of the Ito isometry that

 ∞ 
2
Vt = Et σs ds
t

 2 
∞
= Et σs dWs
t

  2 
∞ t
= Et σs dWs − σs dWs . (12.11)
0 0

However, because
 t
Et [X∞ ] = σs dWs , (12.12)
0

we deduce that

Vt = Et (X∞ − Et [X∞ ])2 , (12.13)

which we recognise as the conditional variance of X∞ with respect to the σ-algebra Ft .

In particular we note that X∞ ∈ L2 (Ω, F, P ).

We shall take the view that the random variable X∞ should in some sense be regarded
as the “primitive” in the construction of the associated interest rate system.

115
12.2 Interest rate models as elements of L2(Ω, F, P )
Let us now recapitulate what we have learned so far.

The market is characterised by a probability space Π = (Ω, F, P ) which we can assume

to be the classical Wiener space associated with a system of n independent Brownian mo-
tions.

If we assume the existence of an arbitrage-free system of discount bonds on Π then it fol-

lows from the considerations of the previous sections that there exists a random variable
X∞ ∈ L2 (Π) with zero mean such that the state price density Vt is given by the conditional
variance

Vt = Et (X∞ − Et [X∞ ])2 (12.14)
and the discount bond system is given by
Et [VT ]
PtT = . (12.15)
Vt
The state-price density is fully determined by the random variable X∞ .

Conversely, given the state-price density process, we can determine the short rate process
and then use the relation σt2 = rt Vt to construct the integrand in the expression for the
corresponding asymptotic random variable X∞ .

We therefore have a correspondence between arbitrage-free positive interest rate models

and square-integrable zero-mean random variables on the Wiener space Π.

Interestingly, this space has a very rich natural structure that can be exploited in the anal-
ysis of the associated interest rate systems.

The key point is that we can represent X∞ , and therefore characterise the corresponding
interest rate system, by use of a Wiener chaos expansion.

In particular, the integrand σs in the defining equation (12.9) can be expanded in a unique
way in a series of the form
 s  s  s1
σs = φs + φss1 dWs1 + φss1 s2 dWs2 dWs1 + · · · . (12.16)
0 0 0

Inserting this expression into (12.9) we then obtain the following representation for the
random variable X∞ :
 ∞  ∞ s
X∞ = φs dWs + φss1 dWs1 dWs + · · · . (12.17)
0 0 0

116
The integrands φs = φα (s), φss1 = φαα1 (s, s1 ), φss1 s2 = φαα1 α2 (s, s1 , s2 ), and so on, appearing
here are deterministic tensor-valued functions, where s ≥ s1 ≥ s2 ≥ · · · .

Then for the expectation of the square of the random variable X∞ we have
 ∞  ∞ s
 2 2
E X∞ = φs ds + φ2ss1 ds1 ds + · · · . (12.18)
0 0 s1 =0

It should be evident by consideration of formula (12.13) that for each choice of X∞ we obtain
a specific interest rate model.

Nesting of interest rate models

In addition, the different models thus arising are nested in a natural way.

To be precise, by an interest rate model we mean the filtered probability space Π together
with the pair (Vt , PtT ).

We shall call an interest rate model that only contains terms up to order n in the ex-
pansion of X∞ an nth -order chaos model.

If X∞ contains only the nth order term we shall call the resulting interest rate model a
“pure” chaos model of order n. It should be evident that the nth -order chaos models are
contained as a subset of the mth -order chaos models, for all n < m.

Despite the relatively high level of abstraction in the overall framework, the inputs of such
models are simply the deterministic functions φs , φs,s1 , φs.s1 ,s2 and so on.

It follows that interest rate models can be classified according to their chaos structure, and
indeed all positive interest HJM models based on a Brownian filtration can be systematically
built up in this way.

12.3 Elements of Wiener chaos

Before we embark upon the analysis of specific interest rate models it will be helpful first if
we review briefly in a little more detail the basics of the Wiener chaos technique.

This will also give us the opportunity to develop the notation further. The material discussed
in this section is for the most part well established, and we refer the reader for example to
Nualart (1995), Øksendal (1997) or Teichmann (2002) for further details. The foundations
of the chaos technique can be found in Wiener (1938) and Ito (1951).

117
The applications of Wiener chaos to problems in finance were pioneered by Lacoste (1996).

Let H be a real Hilbert space with scalar product ·, · .

Given an element h ∈ H, its norm will be denoted h. We introduce a field of ran-
dom variables W = {Wh , h ∈ H}.

We say that W is a Gaussian field if W is a Gaussian family of random variables with

zero mean such that E[Wg Wh ] = g, h for all g, h ∈ H.

Under this definition the map h → Wh is a linear isometry of the space H onto a closed
subspace of L2 (Ω, F, P ), which we denote by H1 .

It follows immediately that W(ag+bh) = aWg + bWh for any a, b ∈ R and g, h ∈ H.

The elements of H1 are zero-mean Gaussian random variables.

Next we introduce the Hermite polynomials Hn (x), defined by the formula

n
1 1 2 d 1 2
Hn (x) = (−1)n e 2 x n
(e− 2 x ), n ≥ 1, (12.19)
n! dx
and H0 (x) = 1.

These polynomials play a fundamental role in the Wiener chaos expansion.

The Hermite polynomials of degree one, two, three and four are H1 (x) = x, H2 (x) = 12 (x2 −1),
H3 (x) = 16 (x3 − 3x), and H4 (x) = 24
1
(x4 − 6x2 + 3) respectively.

Let X and Y be random variables with a jointly Gaussian distribution such that E[X] =
E[Y ] = 0, and E[X 2 ] = E[Y 2 ] = 1.

Then for all n, m ≥ 0 we have

1
E[Hn (X)Hm (Y )] = δnm (E [XY ])n . (12.20)
n!
For each n ≥ 1 we denote by Hn the linear subspace of L2 (Ω, F, P ) generated by the random
variables {Hn (Wh ), h ∈ H, h = 1}, with the convention that H0 denotes the constants.

For n = 1, we recover the space H1 of zero mean Gaussian random variables.

118
It should be evident from (12.20) that Hn and Hm are orthogonal for n = m.

The subspace Hn is called the Wiener chaos of order n.

If we denote by G the σ-field generated by the random variables {Wh , h ∈ H}, then the
space L2 (Ω, G, P ) can be decomposed into the following infinite orthogonal sum of the sub-
spaces Hn :

L2 (Ω, G, P ) = ⊕∞
n=0 Hn . (12.21)

This fundamental decomposition of L2 (Ω, G, P ) leads to the representation of any element

of this space by series of terms resulting from the orthogonal projection of the given element
on to the various chaos subspaces.

Now let us reduce the generality of the underlying Hilbert space and consider the case
H = L2 (R+ , B, µ), where B denotes the Borel σ-algebra on R+ and µ is the Lebesgue mea-
sure.

In this case any element of the nth -order Wiener chaos can be represented as an Ito in-
tegral of a square integrable function.

More precisely, let us consider the subspace ∆n of R+

n
defined by

∆n = {(s, s1 , · · · , sn−1 ) ∈ R+
n
; 0 ≤ sn−1 ≤ · · · ≤ s1 ≤ s ≤ ∞}. (12.22)
n
Also, let the function φn : R+ → R, be square integrable in the sense that
 ∞ s  sn−1
··· φ2n (s, s1 , · · · , sn−1 )dsn−1 · · · ds1 ds < ∞. (12.23)
0 0 0

Then if we let Wt denote a one-dimensional Brownian motion, we can verify that the random
variable In (φn ) defined by the multiple Ito integral
 ∞ s  sn−1
In (φn ) = ··· φn (s, s1 , · · · , sn−1 )dWsn−1 · · · dWs1 dWs (12.24)
0 0 0

is an element of the nth Wiener chaos subspace Hn .

Indeed, the integral on the right hand side of the equation above is an Ito integral on
∆n since the integrand is adapted and square integrable.

W
Now let us write F∞ for the σ-field generated by Wt over the totality of the infinite time
horizon.

119
By combining expression (12.24) with the decomposition (12.21), one is led to the result
that any square integrable random variable X ∈ L2 (Ω, F∞
W
, P ) can be expressed as a chaos
expansion according to the scheme

∞
X= In (φn ), (12.25)
n=0

where the deterministic functions φn ∈ L2 (R+

n
) are uniquely determined by the random
variable X (see, e.g., Revuz and Yor 2001).

Inner product formulae for L2 (Π)

It is a straightforward exercise to verify explicitly by use of the Ito isometry and the stochas-
tic Fubini theorem (interchange of integration and expectation) that elements of distinct
chaos spaces are orthogonal.

For example, if X ∈ H1 , and Y ∈ H2 we have

 ∞  ∞  s
X= φ(s)dWs , and Y = φ(s, s1 )dWs1 dWs , (12.26)
0 0 0

for some choice of φ(s) ∈ L2 (R+

1
) and φ(s, s1 ) ∈ L2 (R+2
), and thus
 ∞  ∞ s 
E [XY ] = E φ(s)dWs φ(s, s1 )dWs1 dWs
0 0 0
 ∞  s 
= E φ(s)φ(s, s1 )dWs1 ds
0 0
 ∞  s 
= E φ(s)φ(s, s1 )dWs1 ds
0 0
= 0. (12.27)

On the other hand, if A, B ∈ H2 are two elements of the same chaos, e.g.,
 ∞ s  ∞ s
A= α(s, s1 )dWs1 dWs , B= β(s, s1 )dWs1 dWs , (12.28)
0 0 0 0

120
then their inner product is given by
 ∞  s  ∞ s 
E [AB] = E α(s, s1 )dWs1 dWs β(s, s1 )dWs1 dWs
0 0 0 0
 ∞
 s  s  
= E α(s, s1 )dWs1 β(s, s1 )dWs1 ds
0 0 0
 ∞  s  s 
= E α(s, s1 )dWs1 β(s, s1 )dWs1 ds
0 0 0
 ∞ s
= α(s, s1 )β(s, s1 )ds1 ds. (12.29)
0 0

Thus the random variables A and B are orthogonal in H2 if and only if the corresponding
elements of L2 (R+
2
) are orthogonal.

Factorisable chaos elements

Another useful result arises in the case for which φn (t1 , t2 , · · · , tn ) is “factorisable” in the
special form

φn (s, s1 , · · · , sn−1 ) = h(s)h(s1 ) · · · h(sn−1 ), (12.30)

for some element h(t) ∈ L2 (R+

1
) with unit norm.

Then for this choice of φn we have the relation In (φn ) = Hn (Wh ), where Hn (Wh ) is the
nth Hermite polynomial formed from the unit-norm Gaussian random variable Wh defined
by
 ∞  ∞
Wh = h(s)dWs , h2 (s)ds = 1. (12.31)
0 0

We note, in particular, that

  ∞
1 2
exp αWh − α = αn Hn (Wh ). (12.32)
2 n=0

The formulae presented in this section apply in the case of the Wiener chaos based on a
standard one-dimensional Brownian motion.

The extension to the general case of a multidimensional Brownian motion is straightforward,

and consists of replacing the deterministic coefficients φs , φss1 , φss1 s2 , etc., with tensorial ex-
pressions of the form φα (s), φαα1 (s, s1 ), φαα1 α2 (s, s1 , s2 ), and so on.

121
12.4 First chaos models
Now we proceed to consider in more detail the structure and classification of interest rate
models according to the scheme outlined in the previous sections.

The first Wiener chaos offers the simplest application of the method and gives rise to a
deterministic interest rate model.

One should remember that the majority of the applications of interest rate theory start
from the deterministic case, so this case should not be regarded as trivial.

Indeed, the chaos framework offers new insights into the relation between deterministic
models and their stochastic generalisations.

It is interesting to note in this connection that even in the case of a deterministic inter-
est rate model there is still a random variable underpinning the dynamics.

For simplicity we shall assume that the dimension of the Brownian motion is one.

In the case of a first chaos model we then write

 ∞
X∞ = φs dWs , (12.33)
0

where φs is a deterministic function of one variable.

A straightforward calculation by use of the Ito isometry confirms that the corresponding
expression for the potential is given by
 ∞
Vt = φ2s ds. (12.34)
t

This is clearly a positive supermartingale that tends to zero in expectation, and it is evident
that the interest rate model that arises is deterministic.

The corresponding expression for the discount bonds is

∞ 2
φ ds
PtT = T∞ s2 . (12.35)
t
φs ds

Thus, the first chaos is sufficient to characterise a deterministic interest rate structure.

In other words, we can identify the space of positive interest yield curves with the first chaos.

122
As a simple example, suppose we take
√ 1
φs = Re− 2 Rs (12.36)

for the first chaos expansion.

Then the associated discount bond becomes

PtT = e−R(T −t) . (12.37)

We remark that there is a direct link between the chaos structure presented here and the
applications of information geometry considered earlier in our discussion of the space of
admissible yield curves.

123
Chapter 13
Second chaos models. Factorisable second chaos models. Foreign exchange
systems.

13.1 Second chaos models

The second chaos models are the simplest models that introduce stochasticity.

In a single-factor second chaos model the random variable X∞ can be represented in the
form
 ∞
X∞ = σs dWs , (13.1)
0

with the adapted process σs given by

 s
σs = φs + φss1 dWs1 . (13.2)
0

Here φs = φ(s) is a deterministic function of one variable, and φss1 = φ(s, s1 ) is a determin-
istic function of two variables.

The second chaos representation for X∞ is then given by

 ∞  ∞ s
X∞ = φs dWs + φss1 dWs1 dWs . (13.3)
0 0 0

In the case of a second chaos model we can think of the deterministic coefficients φ(s) and
φ(s, s1 ) as supplying just enough freedom to allow for calibration to the initial yield curve
and a complete set of caplet prices for all tenors and maturities.

It is a straightforward exercise to show as a consequence of equation


Vt = Et (X∞ − Et [X∞ ])2 , (13.4)

124
that we are then led to the following expression for the state price density:
 ∞
 t 2  ∞ s
Vt = φs + φss1 dWs1 ds + φ2ss1 ds1 ds. (13.5)
t 0 t t

The derivation of formula (13.5) can be established most directly if we write

 ∞
Vt = Mts ds, (13.6)
t

where the positive martingale family Mts is defined for 0 ≤ t ≤ s ≤ ∞ by the relation

Mts = Et σs2 . (13.7)

The fact that Vt can be represented in this way follows as a consequence of

 ∞ 
2
Vt = Et σs ds . (13.8)
t

Then a short calculation making use of the relation (13.2) and the conditional Ito isometry
gives

 t 2  s
Mts = φs + φss1 dWs1 + φ2ss1 ds1 . (13.9)
0 t

To check that the expression appearing on the right hand side of (13.9) is indeed a martingale
we note that
2
Mts = Rts − Qts + Qss , (13.10)

where, for each value of s, Rts is the martingale

 t
Rts = φs + φss1 dWs1 (13.11)
0

and Qts is the associated quadratic variation:

 t
Qts = φ2ss1 ds1 . (13.12)
0

2
If Rts is a martingale and Qts is its quadratic variation, then Rts − Qts is also a martingale,
and hence so is Mts since Qss is deterministic and independent of t.

On the other hand Qss is just the extra term required to ensure Mts is positive for all
0 ≤ t ≤ s ≤ ∞, as is clear from expression (13.9).

125
The discount bond system can then be put into the Flesaker-Hughston form
∞
Mts ds
PtT = T∞ , (13.13)
t
Mts ds

and the initial term structure that corresponds to this system is given by
∞
M0s ds
P0T = T∞ . (13.14)
0
M0s ds
s
More explicitly, we have M0s = φ2s + 0 φ2ss1 ds1 and hence:
∞ 2 s 2

φs + φss ds 1 ds
P0T = T∞ 2 0s 2 1
. (13.15)
0
φs + 0 φss1 ds1 ds

Clearly, by an overall adjustment of the scale of X∞ we can set the denominator in (13.15)
to unity.

With this choice of normalisation the corresponding term structure density is given by
ρ(T ) = M0T .

Expressions for the discount bond volatility and the market price
of risk arising in the case of a general second chaos model
Making use of the Ito quotient identity

d (At /Bt ) dAt dBt (dBt )2 dAt dBt

= − + − , (13.16)
(At /Bt ) At Bt Bt2 At Bt

we deduce that the discount bond volatility is given by

∞ ∞
Uts ds Uts ds
ΩtT =  ∞ T
−  t∞ , (13.17)
T
M ts ds t
M ts ds

and that the market risk premium vector is given by

∞
Uts ds
λt = −  t∞ . (13.18)
t
M ts ds

Here for convenience we have introduced the vector-valued process Uts defined by Uts =
2Rts φst . We note that the constraint ΩT T = 0 is automatically satisfied.

126
The instantaneous forward rate process ftT can be calculated by use of the formula ftT =
−∂T ln PtT and we find
MtT
ftT =  ∞ . (13.19)
T
Mts ds
The short rate process is given analogously by the formula
Mtt
rt =  ∞ , (13.20)
t
Mts ds

which is equivalent to the relation σt2 = rt Vt .

At first glance, the expressions related to the second chaos might look complicated.

However the only exogenously specified ingredients are the deterministic functions φs and
φss1 .

In fact, all the formulae above can be expressed in terms of the underlying Gaussian random
variables Rts .

Option pricing in a second chaos model

We observe that for fixed values of t and s the random variable Mts defined by (13.9) is given
by the square of a Gaussian random variable, plus a constant.

Therefore, for fixed t and T the random variable

 ∞
ZtT = Mts ds, (13.21)
T

can be understood as the integral of a parametric family of squared Gaussian random vari-
ables, plus a constant.

The next step is to define the joint distribution function of the random variables ZtT1 , and
ZtT2 by

FtT1 T2 (x, y) = Prob [ZtT1 ≤ x and ZtT2 ≤ y] . (13.22)

We denote the corresponding joint density function by ftT1 T2 (x, y).

Now the payoff for a call option that expires at time t and is written on a T -maturity
discount bond is

Ht = (PtT − K)+ , (13.23)

127
for some strike K.

Therefore, according to (11.33) the price of this instrument is


H0 = E Vt (PtT − K)+ . (13.24)

By virtue of (13.13) this is evidently equivalent to


H0 = E (ZtT − KZtt )+ , (13.25)

which can be written in terms of the density function f (x, y) in the form
 ∞ ∞
H0 = f (x, y) (x − Ky)+ dxdy. (13.26)
0 0

Analogous formulae can be derived for other types of options.

13.2 Factorisable second chaos models

A considerable simplification can be achieved when the second chaos coefficient φss1 sepa-
rates, that is to say, when φss1 can be written as a finite sum of products of functions of one
variable.

In this situation we obtain a model characterised by a finite set of state variables.

We shall examine in some detail the case where there is a single such term, and set

φs = αs (13.27)

and

φss1 = βs γs1 , (13.28)

where αs , βs and γs1 are deterministic functions of one variable.

The resulting “factorisable” second chaos model then depends on a single state variable.
This model is completely tractable in the sense that it leads to closed-form expressions both
for bond prices and various types of options on bond prices, which we discuss at greater
length below.

First we observe that in the factorisable case we have

 t
φs + φss1 dWs1 = αs + βs Rt , (13.29)
0

128
where the Gaussian martingale Rt is defined by
 t
Rt = γs1 dWs1 . (13.30)
0
At any given time t, the random variable Rt is the sole state variable that characterises the
interest rate system in this model.

If we define the corresponding quadratic variation process Qt by

 t
Qt = γs2 ds, (13.31)
0

then it follows that the process Rt2 − Qt is also a martingale, and the positive martingale
family Mts defined by (13.9) reduces to expression

Mts = αs2 + βs2 Qs + 2αs βs Rt + βs2 Rt2 − Qt . (13.32)

Clearly, Qs ≥ Qt for all s ≥ t, so Mts > 0 for all values of Rt .

For the integral of Mts we can write

 ∞

Mts ds = AT + BT Rt + CT Rt2 − Qt , (13.33)

T
where for convenience in what follows we define the following processes:
 ∞
2

At = αs + βs2 Qs ds,
t
 ∞
Bt = 2 αs βs ds,
t
 ∞
Ct = βs2 ds. (13.34)
t
Setting T = t in (13.33) we see that the state price density is given by

Vt = At + Bt Rt + Ct Rt2 − Qt , (13.35)
and thus that the discount bond price can be written as the ratio of a pair of quadratic
polynomials in the state variable Rt :
AT + BT Rt + CT (Rt2 − Qt )
PtT = . (13.36)
At + Bt Rt + Ct (Rt2 − Qt )
Given these expressions, it is then a straightforward exercise to work out formulae for the
bond volatility, the market price of risk, the short rate, and the instantaneous forward rates,
all of which depend upon Rt .

Because Rt is a Gaussian martingale, it is in principle straightforward to simulate the dy-

namical trajectories of these quantities.

129
Valuation of options in second chaos models
The present value H0 of a European-style call option with strike K exercisable at time t on
a discount bond with maturity T is given by

H0 = E Vt (PtT − K)+ . (13.37)

Now clearly, according to

Vt = At + Bt Rt + Ct Rt2 − Qt , (13.38)

and
AT + BT Rt + CT (Rt2 − Qt )
PtT = , (13.39)
At + Bt Rt + Ct (Rt2 − Qt )
we have

Vt (PtT − K) = (AT − KAt ) − (CT − KCt ) Qt

+ (BT − KBt ) Rt + (CT − KCt ) Rt2 . (13.40)

To proceed let us therefore now fix

√ t, T and K, and introduce the standard normally dis-
tributed random variable Z = Rt / Qt .

Then (13.40) above can be written in the form

Vt (PtT − K) = A + BZ + CZ 2 . (13.41)

Here the quantities A, B and C are defined by:

A = (AT − KAt ) − (CT − KCt )Qt ,

1/2
B = (BT − KBt )Qt ,
C = (CT − KCt )Qt . (13.42)

Therefore if we construct the polynomial P(z) = A + Bz + Cz 2 , we see that the value of the
call option is given by

1 1 2
H0 = √ P(z)e− 2 z dz, (13.43)
2π P(z)≥0

which by an analysis of the roots of P(z) can be reduced to a simple explicit expression
involving the normal distribution function and its density.

Analogous formulae can then be deduced for various other types of options, as we shall
indicate shortly.

130
Explicit formulae for options on discount bonds
Let us proceed then case by case to examine the behaviour of the polynomial P(z) more
closely.

First we distinguish the cases C = 0 and C = 0. If C = 0 then P(z) is linear, and for
the value of the call option we obtain

H0 = AN (−z0 ) + Bρ(z0 ) (13.44)

when B > 0, and

H0 = AN (z0 ) − Bρ(z0 ) (13.45)

when B < 0. Here z0 = −A/B is the single root of P(z), N (z) is the standard normal
distribution function, and ρ(z) is the standard normal density function.

If C = 0, then we need to consider the sign of the discriminant ∆ = B 2 − 4AC.

If ∆ ≤ 0 then for C > 0 the option is guaranteed to expire in the money, and we have
H0 = P0T − KP0t .

If C < 0 then the option will expire out of the money and H0 = 0.

If ∆ > 0 then, again, we have to consider the cases C > 0 and C < 0.

Let us write
√ √
−B − ∆ −B + ∆
z1 = , z2 = (13.46)
2C 2C
for the roots of P(z). Then if C > 0 we obtain

H0 = (P0T − KP0t ) (N (z1 ) + N (−z2 ))

1 √  1 √ 
− B − ∆ ρ(z1 ) + B + ∆ ρ(z2 ), (13.47)
2 2
and if C < 0 we obtain

H0 = (P0T − KP0t ) (N (z1 ) − N (z2 ))

1 √  1 √ 
− B − ∆ ρ(z1 ) + B + ∆ ρ(z2 ). (13.48)
2 2
Thus we see that in the factorisable second-chaos framework the pricing of options on dis-
count bonds is completely tractable.

131
More generally, the value of an option on any predesignated set of deterministic cash-flows
is also tractable, for example an option on a coupon bond.

To obtain the above formulae, we have set A0 = 1. This can be achieved without loss
of generality by changing the scale of X∞ .

Valuation of swaptions
Now we shall demonstrate that in the factorisable second-chaos framework we can also derive
explicit results for a swaption that pays (Stn − K)+ at a series of future dates Ti , for some
strike K, where i = 1, · · · , n, and Stn is the swap rate
1 − PtTn
Stn = n . (13.49)
i=1 PtTi

The effective payoff at expiry t is therefore equal to

 +
 n
Ht = 1 − PtTn − K PtTi , (13.50)
i=1

and the price for this instrument at present is

  + 

n
H0 = E Vt 1 − PtTn − K PtTi . (13.51)
i=1

The analysis turns out to be quite similar to the bond option case.

In the case of a swaption we define the quantities

   
n 
n
A∗ = At − ATn − K ATi − Ct − CT − K CTi Qt
i=1 i=1
 

n
1/2
B∗ = Bt − BTn − K BTi Qt
i=1
 

n
C∗ = Ct − CTn − K CTi Qt . (13.52)
i=1

for fixed t and Ti .

The value of the swaption is then given by


1 1 2
H0 = √ P(z)e− 2 z dz, (13.53)
2π P(z)≥0

132
where in the present case the polynomial P(z) is given by P(z) = A∗ + B ∗ z + C ∗ z 2 .

When C ∗ = 0 we have

H0∗ = A∗ N (−z0∗ ) + B ∗ ρ(z0∗ ) (13.54)

for B ∗ > 0, and

H0∗ = A∗ N (z0∗ ) − B ∗ ρ(z0∗ ) (13.55)

for B ∗ < 0. Here z0∗ = −A∗ /B ∗ .

When C ∗ = 0 then we have to consider the discriminant ∆∗ = B ∗ 2 − 4A∗ C ∗ .

For ∆∗ ≤ 0 we have that, for C ∗> 0 the contract is guaranteed to pay off and the value at
present is H0∗ = P0t − P0Tn − K ni=1 P0Ti .

On the other hand in the case that C ∗ < 0 the contract will expire worthless and H0∗ = 0.

Finally, when ∆∗ > 0 we define the two roots of P(z) by

√ √
∗ −B ∗ − ∆∗ ∗ −B ∗ + ∆
z1 = , z2 = . (13.56)
2C ∗ 2C ∗
The value of the swaption contract is then given by
 
 n
H0∗ = P0t − P0Tn − K P0Ti (N (z1∗ ) + N (−z2∗ ))
i=1
1  ∗ √ ∗ 1  ∗ √ ∗
− B − ∆ ρ(z1∗ ) + B + ∆ ρ(z2∗ ), (13.57)
2 2
when C ∗ > 0; whereas if C ∗ < 0 we get
 

n
H0∗ = P0t − P0Tn − K P0Ti (N (z1∗ ) − N (z2∗ ))
i=1
1  ∗ √ ∗ 1  ∗ √ ∗
− B − ∆ ρ(z1∗ ) + B + ∆ ρ(z2∗ ). (13.58)
2 2
It is a remarkable feature of the factorisable second chaos models that they admit tractable
closed-form expressions for both options and swaptions.

133
13.3 Foreign exchange systems
In conclusion we consider how the framework presented here generalises to the situation
where there is a foreign exchange system, with a family of discount bonds associated to each
currency.

It will be demonstrated that a chaotic representation exists for the entirety of such an
international system of interest rates and foreign exchange.

As a byproduct of this result, we are also led to a simple class of stochastic volatility models
for general asset pricing dynamics.

For convenience we shall write Stij for the price of one unit of currency i in units of currency j.

Here i, j = 0, 1, · · · , N , and we may think of the case i = 0 as referring to the particu-

lar base currency with respect to which the axioms (A1), (A2), and (A3) are framed.

In fact, there is ultimately no special significance to the choice of base currency: the entire
system is symmetrical in the ensemble of currencies.

We shall assume in the present investigation, as before, that the foreign exchange market is
“frictionless” in the sense that

Stij Stjk = Stik (13.59)

for all i, j, k.

Let us write Bti for the value in units of currency i of a money-market account in that
currency, initialised to one unit of currency i.

We assume that for each currency there exists a strictly increasing money-market asset,
with a corresponding strictly positive short rate process rti such that

 t 
i i i
Bt = B0 exp rs ds . (13.60)
0

Constant value assets

We also assume the existence of a floating rate note in each currency.

That is to say, for each i we assume the existence of an asset of constant value in units
of currency i, paying a dividend at the rate rti .

134
Derivative of the exchange rate process as a ratio system
Writing Sti0 for the value of one unit of currency i in units of the base currency, we see that
the product Sti0 Bti represents the base-currency price of a non-dividend paying asset.

Therefore by axiom (A2) we deduce for each value of i that

Sti0 Bti
Mti = (13.61)
ξt
is a martingale, from which it follows that the process Vti defined by
Sti0
Vti = , (13.62)
ξt
is a supermartingale.

Since Stij Stj0 = Sti0 for all i, j, we thus deduce that

Vti
Stij = . (13.63)
Vtj
This gives us a general expression for the exchange-rate process as a ratio of supermartingales.

As a consequence we deduce that the dynamics of Stij are given by

dStij  j j
j

i i
ij = rt − rt + λt λt − λt dt + λjt − λit dWt , (13.64)
St
where λit is the market price of risk process associated with assets that are denominated in
currency i.

The derivation of (13.64) follows directly from the relation

dVti = −rti Vti dt − λit Vti dWt (13.65)

together with the Ito quotient rule.

It is interesting to note that in the general arbitrage-free exchange rate dynamics the FX
volatility is completely determined by the associated market price of risk processes.

Foreign discount bonds

i
Let us consider the discount bond system for foreign currency number i. We denote by PtT
the value at time t of a bond that pays one unit of currency i at time T .

135
In this case Sti0 PtT
i
is the base-currency price of a non-dividend paying asset, and there-
i0 i
fore St PtT /ξt is a martingale by (A2).

It follows that Sti0 PtT

i
/ξt = Et [STi0 PTi T /ξT ].

Thus, from Sti0 /ξt = Vti and PTi T = 1, we deduce from this line of argument that

i Et [VTi ]
PtT = . (13.66)
Vti

Asymptotic behaviour
i
Now we make the additional assumption that limT →∞ P0T = 0 for all i.

It follows that a conditional variance representation exists for the state-price density as-
sociated with each currency.

i
In other words, there exists a set of random variables X∞ ∈ L2 (Ω, F, P ) for i = 0, 1, · · · , N
such that
  i
2 
i i
Vt = Et X∞ − Et X∞ . (13.67)

These random variables then each admit a chaos representation in terms of the vector Wiener
process Wtα (α = 1, · · · , k).

i
We see that once the random variables X∞ have been specified for i = 0, 1, · · · , N then
the international system of interest and foreign exchange is completely determined by the
relations
Vti
Stij = j, (13.68)
Vt

i Et [VTi ]
PtT = (13.69)
Vti
and
  i
2 
Vti = Et i
X∞ − Et X∞ . (13.70)

i
We can refer to the random variables X∞ as the generators of the corresponding interest
rate and foreign exchange system.

136
It should be evident that although we have consistently used the language of foreign ex-
change in the discussion above, the matrix process Stij can be used to characterise the price
of any asset in terms of another, providing that these prices are always positive and that we
interpret the associated short-rate systems as continuous dividend streams.

As a consequence we see that the generic model for such a “basic” asset price is a pro-
cess of the form

Et (Y∞ − Et [Y∞ ])2
St =  , (13.71)
Et (X∞ − Et [X∞ ])2

that is, a ratio of conditional variances, where X∞ and Y∞ are elements of L2 (Ω, F, P ).

For example, if we think of St as a dollar-valued share price (and we approximate the

dividend flow as continuous–an equity index might work better for that!) then X∞ carries
the information of the dollar risk premium, and the dollar interest rate, whereas Y∞ carries
the information that is more specific to the particular stock.

The simplest models leading to a nontrivial asset price stochasticity are those for which
at least one of X∞ or Y∞ is an element of the second chaos.

137
Chapter 14
Real and nominal interest rates. Models for inflation. Valuation of index-
linked bonds and other inflation related products. General principles for the
design of inflation-linked products.

14.1 Inflation linked bonds∗

Now we consider a general model of inflation and inflation-linked derivatives.

The idea is to formulate an approach to the valuation of inflation derivatives that is as

close as possible to the methodologies for valuing foreign exchange and interest rate deriva-
tives.

The theory of inflation has aspects that relate to both interest rates and foreign exchange.
In particular, a useful way of thinking about inflation is to treat the consumer price index
(CPI) as if it were the price of a foreign currency.

We begin by considering an economy consisting of discount bonds and index-linked dis-

count bonds.

The indexing of the index-linked discount bonds is with respect to the consumer price index
which at time a has the value Ca .

We think of Ca as representing the value, in units of the domestic currency (henceforth,

dollars) of a typical basket of goods and services at that time.

An increase in Ca over an interval of time then indicates that there has been inflation
over that period.

We shall define an inflation linked discount bond to be a bond which pays out Cb at the
maturity date b. In other words, the inflation linked bond pays out enough in dollars to buy

138
a unit of goods and services at that time.

Our problem is to formulate a general theory for the price processes of the consumer price
index and index-linked bonds, and tie this in with the HJM theory of interest rate derivatives.

Indexation is debt is not a new idea. An early example occurs in 1742 when Massachusetts
issued bills linked to the price of silver on the London Exchange. The risks in indexation to
a single commodity became apparent a few years later when the price of silver rose in excess
over general prices.

As a consequence, a law was passed in Massachusetts requiring a wider base of commodities

for indexation.

In 1780 notes were issued again, indexed this time with the intention of preserving the
value of notes issued as wages to soldiers in the American Revolution.

In this case, both the principal and the interest of the notes were indexed to the com-
bined market value of five bushels of corn, sixty-eight and four-sevenths pounds of beef, ten
pounds of sheep wool, and sixteen pounds of sole leather.

14.2 Payout structures for inflation-linked products∗

N
We denote by Pab the value of a nominal discount bond at time a with maturity at time b.
At maturity the nominal discount bond pays one dollar.

Then a typical inflation-linked derivative has a payout or payouts given by functions of

nominal discount bonds (at various times and of various maturities) and the consumer price
index (at various times). Some examples are as follows.

(a) Inflation cap. This pays out if inflation (as measured by percentage appreciation in
the CPI) exceeds a certain threshold K over a given period.

Thus if the period in question is the interval (a, b), then the payout Hb at time b is given by:

 
Cb
Hb = X max − 1 − K, 0 , (14.1)
Ca

where X is some dollar notional.

139
In practice the payout would have to be delayed to some still later date c (to allow for
official publication of the relevant CPI figure), so the effective payout is

 
N Cb
Hb = XPbc max − 1 − K, 0 . (14.2)
Ca

(b) Inflation swap. For a succession of intervals (ai , bi ) (i = 1, . . . , n) we receive the inflation
rate
Cb
Iab = −1 (14.3)
Ca
for that interval (with payment delayed to some slightly later time ci ), and pay a fixed rate,
all on a fixed notional.

(c) Zero strike floors on inflation. Here the idea is to protect the receiver of the infla-
tion leg in an inflation swap against a deflation scenario.

Thus instead of simply receiving Iab , which can go negative (deflation), one receives max[Iab , 0].

(d) Inflation swaption. This confers the right to enter into an inflation swap (e.g., as a
payer of the fixed rate) at some specified future time, with a given “strike” fixed rate.

(e) Inflation protected annuity. This pays a fixed “real” annuity on the future dates ai :
f N Cai
Hai = . (14.4)
C0
Here f is the nominal annuity rate (e.g., 5%), N is the notional.

The effect of the CPI is to inflate the actual payment appropriately.

(f) Knockout option. A typical structure, for example, might pay if the total inflation
exceeds a certain threshold K at time T .

Knockout would occur if the total inflation drops below a certain specified critical level
K  between time t and T .

 

CT CT
HT = N max − 1 − K, 0 unless − 1 − K  ≤ 0 at some time a
Ca Ca
in the interval t ≤ a ≤ T, in which case HT = 0. (14.5)

There are many variations on this kind of structure.

140
The basic idea is to make the option premium cheaper by having the contract specify a
cancelling of the structure in the event of certain circumstances.

(g) Cap on “real” interest rates. This might, for example, pay off

Hb = X max[LR
ab − K, 0]. (14.6)

Real rates are not necessarily available as a basis for contract specification. Instead we can
use a proxy.

(h) Proxy cap on “real” interest rates. This instead would pay

Hb = X max[LN
ab − Iab − K, 0], (14.7)

where Lab is the relevant per-period Libor rate.

Then if the Libor rate exceeds the inflation rate over the given interval by more than a
specified amount, there is a payoff.

Here we have used the difference between the Libor rate and the inflation rate as a con-
venient proxy for the “real” interest rate over the given interval.

Clearly, more “exotic” structures can also easily be represented. Analogues both from the
FX world (treating Ca as a foreign exchange rate), and the interest rate world (treating Iab
as a kind of “rate”) can be formulated.

14.3 General theory of inflation∗

N
There are three ingredients: the “nominal” discount bonds Pab , the “real” discount bonds
R
Pab , and the consumer price index Ca .

The real discount bonds are defined as follows.

R
By Pab we mean intuitively the price at time a, in units of goods and services, for one
unit of goods and services to be delivered at time b.

R
Thus Pab is the discount function that characterises “real” interest rates. If we lived in
R
a pure barter economy, with no money, then Pab would define the term structure of interest
rates.

For example, if the price of bread happened to be a good proxy for goods and services
in general, then one “unit” of goods and services could be represented by 100 loaves of bread.

141
The real term structure of interest rates would then supply information like how many
loaves of bread you should in principle be willing to part with today in exchange for a sure
delivery of 100 loaves one year from now.

The answer might be, say, 97 loaves, and that enables us to define the one-year real in-
terest rate.

Associated with the system of real discount bonds we have a corresponding system of real
interest rates. We denote a typical real rate with the notation LR
ab .

The index-linked discount bonds are related to the real discount bonds by the consumer
price index, which acts as a kind of exchange rate.

R
In other words, if we multiply the Pab by Ca , that gives us the dollar value of the b-maturity
real discount bond at time a.

In the foreign exchange analogy, we think of the nominal (dollar) discount bonds as the
“domestic” bonds. We think of the real discount bonds as “foreign” discount bonds, and the
CPI plays the role of the exchange rate.

Note that the actual inflation rate Iab for the period (a, b) is not strictly analogous to an
interest rate in the usual sense – it is only known at time b (or later!).

It is thus best thought of as an appreciation in an asset price.

But in that case what is the relation between “real” rates, “nominal” rates, and “infla-
tion” rates?

Clearly care is required, and we must not confuse categories just because these are all loosely
referred to as “rates”.

Part of the goal is to gain some insight into the relation between these various “rates”.

14.4 Price processes for nominal discount bonds∗

As usual in an HJM type framework, we assume an economy where uncertainty in the future
is modelled by a multi-dimensional Brownian motion defined with respect to the natural
probability measure.

142
Assuming no arbitrage, and thus the existence of a risk premium vector, we can write the
dynamics for the price processes of the nominal discount bonds in the form
N
dPab
N
= (raN + λN N N
a Ωab ) da + Ωab dWa . (14.8)
Pab

Here raN is the nominal short rate, λN N

a is the nominal risk premium vector, Ωab is the nominal
vector volatility, and Wa is the Brownian motion vector.

By analogy, for the real discount bonds we have

R
dPab
R
= (raR + λR R R
a Ωab ) da + Ωab dWa . (14.9)
Pab

It then follows by virtue of the foreign exchange analogy that the price dynamics for the
consumer price index are
dCa
= [raN − raR + λN N R N R
a (λa − λa )] da + (λa − λa ) dWa . (14.10)
Ca
We note that the CPI volatility vector can be expressed as the difference between the nom-
inal and real risk premium vectors.

Thus we can write

dCa
= (raN − raR + λN
a νa )] da + νa dWa , (14.11)
Ca

where νa = λN R
a − λa is the CPI volatility.

In the absence of a risk premium, we see that the drift of the CPI is given by the dif-
ference between the nominal short rate and the real short rate.

In reality, the drift of the CPI contains another term, given by the product of the nomi-
nal risk premium vector and the CPI volatility vector.

Thus if by the instantaneous rate of inflation Ia we mean the drift process for the consumer
price index, we have:

Ia = raN − raR + λN
a νa . (14.12)

This is an expression of the so-called “Fisher equation”, which relates the inflation rate to
the nominal interest rate minus the real interest rate plus a risk premium term.

143
14.5 Transfer to the nominal risk neutral measure∗
For the valuation of derivatives we want to introduce a change of measure such that the ratio
N
of any of the nominal bonds Pab to the nominal dollar money market account is a martingale.

Suppose we write BaN for the nominal money market account, which satisfies

dBaN = raN BaN da. (14.13)

Then we introduce a new probability measure P N as usual according to the scheme

Ea [Λb Xb ]
EN
a [Xb ] = , (14.14)
Ea [Λb ]

where ENa denotes conditional expectation with respect to the measure P

N
given the filtra-
tion up to time a, and where Xb is any random variable adapted to time b.

We call P N the nominal (or dollar) risk neutral measure.

Here the change of measure density process Λa is defined by

 a  
N 1 a N
2
Λa = exp − λs dWs − λs ds . (14.15)
0 2 0

With respect to P N the process WaN defined by

dWaN = dWa + λN
a da (14.16)

is a Brownian motion.

N R
Then for the processes Pab and Pab we can write
N
dPab
N
= raN da + ΩN N
ab dWa (14.17)
Pab

and
R
dPab
R
= (raN − νa ΩR R N
ab ) da + Ωab dWa . (14.18)
Pab

We note that the process (14.18) for the real discount bonds picks up a “quanto” term in
the drift in the nominal risk neutral measure.

This is appropriate since the real discount bonds are not denominated in dollars.

144
The process for the consumer price index in the risk neutral measure is:
dCa
= (raN − raR ) da + νa dWaN . (14.19)
Ca
Thus in the risk neutral measure the nominal risk premium term disappears, and we see that
the drift on the CPI is given by the difference between the nominal and real interest rates.

The process is like that of a foreign currency, and we can think of the real interest rate
as playing the role of the “foreign” interest rate.

Normally we expect raN and raR both to be positive.

There are good economic arguments to support the idea that both nominal and real in-
terest rates should be positive.

N
We note that by construction the ratio process Pab /Ba is a martingale in the nominal risk
neutral measure.

R R
So is Ca Pab /Ba , where Ca Pab is the (dollar) value of an index linked discount bond.

14.6 Valuation of inflation linked derivatives∗

Now let HT be a random variable corresponding to the payout of an inflation linked deriva-
tive.

We can think of HT as depending in a general way of the values of nominal discount bonds,
and the consumer price index at times between the present and the maturity date T .

There are many examples of inflation linked derivatives for which the payout depends in
a direct way only on the nominal discount bonds and the consumer price index, but not on
the real discount bonds.

These we shall call “index linked” derivatives, and it should be noted that these structures
are in principle more straightforward to value and hedge than inflation linked derivatives,
that also involve real interest rates.

The basic derivatives valuation formula is given in the risk neutral valuation scheme by
 
N HN
H0 = E . (14.20)
BT

145
In particular, we can consider the case where HT is the payout CT of an index linked discount
bond, normalised by the value of today’s CPI. Then we have
 
R N CT /C0
P0T = E (14.21)
BT

which shows that today’s market for index linked bonds tells us the initial real discount
function.

In reality, we have to work a bit harder, on account of the lagging effect, and the fact
that we generally have to work with coupon bonds.

Finally, by use of the foreign exchange analogy let us consider a simple Black-Scholes type
model for the valuation of index derivatives.

Let us assume deterministic interest rates (nominal and real), and a deterministic CPI
volatility, with a prescribed local volatility function νt . Then for the CPI process we can
write:

 t  
C0 P0tR 1 t 2
Ct = exp νs dWs − ν ds , (14.22)
P0tN 0 2 0 s

where the expression C0 P0tR /P0tN is the forward value for the CPI.

In this case the situation is entirely analogous to the corresponding problem for foreign
exchange, and by use of the Black-Scholes formula we can get a crude valuation for some
products in this way, though of course care is required in the case of longer dated structures.

146
Bibliography

[1] Bhattacharyya, A. 1943 On a measure of divergence between two statistical populations

defined by their probability distributions Bull. Calcutta Math. Soc. 35, 99–109.

[2] Björk, T. & Christensen, B. J. 1999 Interest rate dynamics and consistent forward rate
curves. Mathematical Finance 9, 323–348.

[3] Björk, T. & Gombani, A. 1999 Minimal realizations of interest rate models. Finance
and Stochastics 3, 413–432.

[4] Björk, T. 2001 A geometric view of interest rate theory. In Option pricing, interest
rates and risk management, Handb. Math. Finance, Cambridge: Cambridge University
Press.

[5] Brody, D. C. 2000 Modern Mathematical Theory of Finance, Tokyo: Nippon-

Hyoronsya.

[6] Brody, D. C. & Hughston, L. P. 2001a Interest rates and information geometry. Proc.
Roy. Soc. London A457, 1343–1364.

[7] Brody, D. C. & Hughston, L. P. 2001b Applications of information geometry to interest

rate theory, In Disordered and Complex Systems, P Sollich, ACC Coolen, LP Hughston,
RF Streater (eds), New York: AIP Publishing.

[8] Brody, D. C. & Hughston, L. P. 2002 Entropy and Information in the Interest Rate
Term Structure. Quantitative Finance 2, 70-80.

[9] Brody, D. C. & Hughston, L. P. (2003) Risk (to appear).

[10] Brody, D. C. & Hughston, L. P. (2003) Phil. Trans. R. Soc. London (to appear).

[11] Cover, T. M. & Thomas, J. A. 1991 Elements of Information Theory, New York: John
Wiley & Sons.

[12] Filipović, D. 2001 Consistency Problems for Heath-Jarrow-Morton Interest Rate Mod-
els, Lecture Notes in Mathematics 1760, Berlin: Springer-Verlag.

147
[13] Flesaker, B. & Hughston, L. P. 1996 Positive Interest Risk Magazine 9, 46–49; reprinted
in Vasicek and Beyond, L.P. Hughston (ed), London: Risk Publications (1996).

[14] Flesaker, B. & Hughston, L. P. 1997 International models for interest rates and for-
eign exchange Net Exposure 3, 55–79; reprinted in The New Interest Rate Models,
L.P. Hughston (ed), London: Risk Publications (2000).

[15] Flesaker, B. & Hughston, L. P. 1998 Positive Interest: An Afterword, in Hedging with
Trees, Broadie, M. & Glasserman, P. (eds), London: Risk Publications.

[16] Heath, D., Jarrow, R. & Morton, A. 1992 Bond pricing and the term structure of
interest rates: a new methodology for contingent claims valuation. Econometrica 60,
77–105.

[17] Hughston, L. P. & Rafailidis, A. (2002) King’s College Preprint.

[18] Hunt, P. J. & Kennedy, J. E. 2000 Financial Derivatives in Theory and Practice,
Chichester: John Wiley & Sons.

[19] Ikeda, N. & Watanabe, S. (1981) Stochastic Differential Equations and Diffusion Pro-
cesses Amsterdam: North-Holland.

[20] Ito, K. (1951) J. Math. Soc. Japan 3, 157-169.

[21] James, J. & Webber, N. (2000) Interest rate modelling Chichester: Wiley.

[22] Jamshidian, F. (1997) Finance and Stochastics 1, 293.

[23] Janson, S. (1997) Gaussian Hilbert Spaces Cambridge: Cambridge University Press.

[24] Jaynes, E. T. 1982 On the rationale of maximum entropy methods. Proc. IEEE 70,
939–952.

[25] Jaynes, E. T. 1983 Papers on probability, statistics and statistical physics: edited and
with an introduction by R. D. Rosenkrantz, Dordrecht: D. Reidel Publishing Co.

[26] Kennedy, D. (1994) Math. Finance 4, 247.

[27] Long, J. L. (1990) J. Financial Economics 26, 29.

[28] Lipton, A. (2001) Mathematical methods for foreign exchange Singapore: World Sci-
entific.

[29] Meyer, P. (1996) Probability and potentials Massachusetts: Blaisdell Publising Com-
pany 1966

148
[30] Musiela, M. & Rutkowski, M. 1997 Martingale Methods in Financial Modelling, Berlin:
Springer-Verlag.

[31] Nualart, D. (1995) The Malliavin calculus and related topics Berlin: Springer.

[32] Øksendal, B. (1997) An introduction to Malliavin calculus with applications to eco-

nomics Lecture notes, University of Oslo.

[33] Revuz, D. & Yor, M. (2001) Continuous Martingales and Brownian Motion (3rd ed.,
Corrected 2nd print) Berlin: Springer.

[34] Rogers, L. C. G. 1997 The potential approach to the term structure of interest rates
and foreign exchange rates Math. Finance 7, 157–176; reprinted in The New Interest
Rate Models, L.P. Hughston (ed), London: Risk Publications (2000).

[35] Rutkowski, M. 1997 A note on the Flesaker-Hughston model of the term structure of
interest rates Applied Math. Finance 4, 151–163; reprinted in The New Interest Rate
Models, L.P. Hughston (ed), London: Risk Publications (2000).

[36] Wiener, N. (1938) Amer. J. Math. 60, 897-936.

149