Interest Rate Modelling -

The Mathematics of the

Fixed Income World

Basic Finance
A zero coupon bond (ZCB) is a contract which pays a
known, …xed amount, called the principal at some known
date in the future called the maturity ; T . Typically the
principal is a normalised quantity of $1: Usually, but not
always, the value of the bond prior to maturity is less
than the principal. As we approach T , the value of the
bond converges to the principal.

A coupon bearing bond is similar to the ZCB, except that

in addition to the principal paid at T; it also pays out at
various dates prior to the maturity a coupon - which is a
known fraction of the principal. These special dates are
known as coupon dates. As an example consider such a
product that pays out a principal value in ten years time
and, say, 3% coupons (i.e. $0.03) every year. Clearly this
contract is more valuable than its zero coupon equivalent.
1
This coupon paying product can be regarded as a col-
lection of ZCBs. That is with a principal equal to the
coupon and maturity equal to the coupon date for each
coupon date and one with principal equal to the actual
principal and maturity equal to the actual maturity.

Yield to Maturity : Consider the price of a ZCB of ma-

turity T and principal $1 at time t < T to be Z (t; T ) :
The yield to maturity y is de…ned as
1
Z (t; T ) = e y(T t) () y = log Z:
T t
It is the constant rate of return which gives the observed
bond price (so in a sense like an implied volatility).

For a coupon bearing bond; again with principal $1and

coupons Ci paid on dates ti we can de…ne the y by
N
X
V (t; T ) = e y(T t) + Cie y(ti t);
i=1
ti>t
where N is the total number of coupons paid. We see
that the yield is a decreasing function of increasing price
and that price is a decreasing function of increasing yield.
2
The yield curve: suppose that today we are observing the
prices of ZCBs of varying maturities. So now t is …xed
but T varies. So the yield also varies as a function of T;

y = y (T ; t) ;
that is it is contingent on today being t:

The yield curve is a plot of the yield against maturity.

Typically we plot yield against the time to maturity (T t)
and call the plot the Term structure of interest rates.

The duration can be thought of in two ways. Here we

use parallels with option pricing.

One is to consider this to be the sensitivity of the bond

price to yield. In …nance terms it is the percentage change
in the bond price in response to a one percent change in
the yield. Di¤erentiating V (t; T ) above gives
N
X
dV
= (T t) e y(T t) (ti t) Cie y(ti t)
dy i=1
ti>t

3
Now consider Taylor’s theorem for V = V (y ) a small
change in yield y
dV 1 d2V 2
V (y + y ) = V (y ) + y + 2 2 y + :::
dy dy
dV
V (y + y ) V (y ) = y + O y2
dy
hence to leading order
dV
V = y
dy
so the relative change is
V 1 dV
= y:
V V dy
Of importance is the quantity
N
X
(T t) e y(T t) + (ti t) Cie y(ti t)
i=1
1 dV ti>t
=
V dy XN
e y(T t) + Cie y(ti t)
i=1
ti>t
called the duration. The larger the duration, the bigger
the response of the bond’s relative value to changes in
yield.
4
Interest Rate Modelling

Introduction
When pricing non linear derivatives products a model for
the randomness in the underlying is required.

This is as true for …xed-income instruments as it is for

equity derivatives.

There are a number of ways to model interest rates:

Monte Carlo

Trees

PDEs

5
This is an introduction to the PDE framework for interest
rate products.

The basis of this section is to derive the bond pricing

equation for a one factor model. The underlying state
variable is a non-tradeable asset. Hence when pricing
derivatives written on such securities a new term called
the market price of risk appears in the equation. The
evolution of this will be discussed in detail.

Unlike for equities, for interest rates there are numerous

models. A number of these will be presented. The focus
will be on bond prices which can be obtained analytically,
i.e. tractable and a¢ ne models.

We start by de…ning a stochastic di¤erential equations to

model the random nature of an interest rate, using one
source of randomness. This is the subject of one-factor
interest rate modelling.

We begin by having one source of randomness, the spot

interest rate.

6
The model will allow the short-term interest rate, the
spot rate, to follow a random walk.

This model leads to a parabolic partial di¤erential equa-

tion for the prices of bonds and other interest rate deriv-
ative products.

Later we will consider modelling the …xed-income world

using two sources of randomness.

7
Stochastic interest rates
The ‘spot rate’that we will be modelling is a very loosely-
de…ned quantity, meant to represent the yield on a bond
of in…nitesimal maturity.

In practice one should take this rate to be the yield on a

liquid …nite-maturity bond, say one of one month.

Bonds with one day to expiry do exist but their price is

not necessarily a guide to other short-term rates. Since
we cannot realistically forecast the future course of an in-
terest rate, it is natural to model it as a random variable.

We are going to model the behaviour of r, the inter-

est rate received by the shortest possible deposit.

8
From this we will see the development of a model for all
other rates.

The interest rate for the shortest possible deposit is com-

monly called the spot interest rate.

Let us suppose that the interest rate r is governed by a

stochastic di¤erential equation of the form

dr = u(r; t) dt + w(r; t) dX: (1)

The functional forms of u(r; t) and w(r; t) determine the

behaviour of the spot rate r.

For the present we will not specify any particular choices

for these functions.

9
The pricing equation for the general model
When interest rates are stochastic a …xed-income instru-
ment has a price of the form V (r; t). Previously we used
V (S; t) to denote an equity derivative price where the
underlying stock S was used, to hedge with. Pricing in-
terest rate products presents a di¤erent problem, which
lies in the hedging. There are now no assets with which
to hedge because r is not a traded quantity, which is what
makes interest rate derivatives harder.

To get round this problem we hedge with IR deriva-

tive securities of di¤erent maturities - here we cannot
do V (r; t) r:

We are not modelling a traded asset; the traded asset

(the bond, say) is a derivative of our independent
variable r.

Pricing a …xed-income instrument presents new technical

problems, and is in a sense harder than pricing an option
since there is no underlying asset with which to hedge.

10
 The only way to construct a hedged portfolio is by
hedging one bond with a bond of a di¤erent maturity.

We set up a portfolio containing two bonds with di¤erent

maturities T1 and T2.

The bond with maturity T1 has price V1(r; t; T1) and the
bond with maturity T2 has price V2(r; t; T2).

We hold one of the former and a number of the lat-

ter.

We have
= V1 V2: (2)

Again as with equities we exploit correlation (in this the-

oretical world). One factor modelling implies there is just
one source of randomness used for bonds, caps, ‡oors,
etc.
11
It makes no di¤erence which way around V1 and V2 are,
i.e. we can have = V2 V1: We will see later why
this is unimportant. Then in one time-step dt the change
in the portfolio is de…ned by

d = dV1 dV2: (3)

We hold …xed during the course of the time-step and

reset when we come back to hedge again.

The change in this portfolio in a time dt can be obtained

using Itô. So we have for i = 1; 2
@Vi @Vi 1 @ 2Vi 2
dVi = dt + dr + 2
dr
@t @r 2 @r
where dr2 = w2dt: So the portfolio can be written as
d =

@V1 @V1 1 2 @ 2 V1
@t dt + @r dr + 2 w @r2 dt
@V2 @V2 1 2 @ 2 V2
@t dt + @r dr + 2 w @r2 dt ; (4)

12
This is not riskless because the dr brings in the random
component dX: So if we can eliminate the coe¢ cients
of dr then the portfolio is risk-free. That is @V 1
@r and
@V2
@r ; so we set
@V1 @V2
= 0;
@r @r
which is rearranged to give

@V1 @V2
= : (5)
@r @r

This choice of eliminates all the.randomness associated

with d . So replace with @V@r
1 @V2 in (4) to give
@r

.
d = @V1
+ 1 w 2 @ 2 V1 @V1 @V2 @V2 2
+ 12 w2 @@rV22 dt
@t 2 @r2 @r @r @t
(6)

This means that the change in the portfolio is completely

riskless, hence it must grow at the risk free rate, i.e. no-
arbitrage:
d =r !d = r dt (7)
dt
13
where r is the spot rate at time t: If it was not, then we
would have two factors in our model. So d = r dt
gives
.
@V1 @V2
d = r V1 @r @r V1 dt (8)

Hence (6) and (8) are equivalent, so equate them

1 2 @ 2V1 . 1 2 @ 2V2
@V1 @V1 @V2 @V2
@t + w @r2 @r @r @t + w @r2 dt
2 . 2
@V1 @V2
= r V1 @r @r V1 dt:
Now write this equation so that the left hand and ride
hand sides are in terms of V1 and V2 respectively.

@V1
+ 1 w 2 @ 2 V1 rV1 @V2
+ 1 w 2 @ 2 V2 rV2
@t 2 @r2 = @t 2 @r2 : (9)
@V1 @V2
@r @r

The lhs depends only on V1; i.e. on T1: The rhs is de-
pendant upon V2 only, i.e. on T2: This is one equation in
two unknowns - but more importantly should hold for all
14
r and t: Hence (9) is a function of (r; t) ; i.e. a function
of 2 common variables to V1 and V2: LHS is f (r; t; T1)
and RHS is g (r; t; T2) :

So each side can be a function of r and t but must be

independent of T1 and T2: Both sides must be equal to
a universal constant. The only way for this to be possible
is for both sides to be independent of the maturity date.

Recall when we solved the simple one dimensional heat

equation using the separation of variables method
@u @ 2u
=
@t @x2
u (X; t) = X (x) T (t)
which simpli…es to
T 0 (t) X 00 (x)
= = constant
T (t) X ( x)
i.e. constant indep. of x; t:

So (9) becomes
@V1
+ 1 w 2 @ 2 V1 rV1 @V2
+ 1 w 2 @ 2 V2 rV2
@t 2 @r2 = @t 2 @r2 = constant
@V1 @V2
@r @r
15
Hedging with V1 and V2, or V3 and V4 or Vi and Vj etc.
makes no di¤erence. The nonlinear operator
2
@
@t + 12 w2 @r
@
2 r
@
@r
is a universal operator, i.e. some function for all Vi: Let’s
call this a (r; t) is a universal property of interest rates -
but currently does not have a nice …nance interpretation.

Both sides of (9) can be functions of the ‘variables’ in

the model, r and t, since these are common to all instru-
ments.

Dropping the subscript from V , we have

@V + 1 w2 @ 2V rV
@t 2 @r2 = a(r; t)
@V
@r

for some function a(r; t).

16
We can write this w.lo.g as

a (r; t) = (r; t) w (r; t) u (r; t)

for a given u(r; t) and non-zero w(r; t) this is always

possible.

The …xed-income pricing equation or bond pricing equa-

tion is therefore

@V 1 2 @ 2V @V
+ w 2
+ (u w) rV = 0: (10)
@t 2 @r @r

This is another parabolic partial di¤erential equation, very

similar to the Black-Scholes equation.

As with equity derivatives, until a …nal condition is speci-

…ed the PDE does not know what kind of product is being
priced. So we need a time condition corresponding to the
payo¤ at maturity, T .

17
For example, the …nal condition for a zero-coupon is

V (r; T ) = 1:

So it is a normalised condition called the redemption

value.

18
The market price of risk?

The term (r; t) introduced earlier is called the market

price of risk. This arises because we have to hedge with
another bond rather than r itself. This also appears when
pricing volatility derivatives. We begin by supposing that
we hold an unhedged position in one bond with maturity
date T .

We saw earlier from Itô that in a time step dt this bond

changes in value by

!
@V @V 1 @ 2V
@V
dV = w dX + + w2 2 + u dt: (11)
@r @t 2 @r @r

Having derived the bond pricing equation (bpe) we know

from (10)

@V 1 2 @ 2V @V @V
+ w + u w rV = 0
@t 2 @r2 @r @r

19
which can be arranged to give
@V 1 2 @ 2V @V @V
+ w 2
+ u = w + rV: (12)
@t 2 @r @r @r

So substitute (12) in to (11) to give

@V @V
dV = w dX + w + rV dt;
@r @r
or
@V
dV rV dt = w (dX + dt): (13)
@r

We can interpret (13) …nancially as the di¤erence be-

tween

1. dV - return on an unhedged bond (contains risk)

and

2. rV dt - risk free return, i.e. the money invested from

selling a bond which earns at the risk free rate, i.e.
dV = rV
dt

20
which gives w @V
@r (dX + dt):

Since the coe¢ cient of dX 6= 0; the portfolio is not

riskless and the term dt is the extra return on the port-
folio per unit of risk dX: It is the amount by which the
market wishes to be compensated, for hedging with a
non-tradeable asset.

The sign of is generally negative because as r increases

the bond price decreases =) @V@r < 0:

For equities, since S is a tradeable asset

r
=

More on this later.

21
We now approach the market price of risk di¤erently, by
not assuming the actual form of a (r; t) : Start by writing
the BPE as
@V 1 2 @ 2V @V
+ w a(r; t) rV = 0
@t 2 @r2 @r
@V 1 2 @ 2V @V
+ w = a(r; t) + rV
@t 2 @r2 @r

add u (r; t) @V
@r to both sides

@V 1 2 @ 2V @V @V @V
+ w +u = a(r; t) + rV + u (14)
@t 2 @r2 @r @r @r

and from earlier the unhedged bond is given by (11) : So

we substitute (14) into (11) ; which gives
@V @V @V
dV = w dX + a(r; t) + rV + u (r; t) dt:
@r @r @r

Now do the (unhedged bond) (risk free return)

@V @V
dV rV dt = w dX + (a + u) dt
@r @r
@V a+u
= w dX + dt
@r w
22
so put
a+u
= (r; t) :
w
Rearranging gives

a(r; t) = (r; t) w (r; t) u (r; t)

and justi…es our earlier de…nition of a(r; t):

23
If we compare the BPE with the BSE
@V 1 2 @ 2V @V
+ w 2
+ (u w) rV = 0:
@t 2 @r @r
@V 1 2 2 @ 2V @V
+ S + rS rV = 0;
@t 2 @S 2 @S

The four terms in the equation represent, in order as

written,

time decay @V
@t ,

di¤usion 12 w2,

drift (u w) and

discounting rV .

The equation is similar to the backward equation for a

probability density function except for the …nal discount-
ing term.
24
 As such we can interpret the solution of the bond
pricing equation as the expected present value of all
cash‡ows.

This idea should be familiar from equity derivatives.

As with equity options, this expectation is not with re-

spect to the real random variable, but instead with re-
spect to the risk-neutral variable.

There is this di¤erence because the drift term in the

equation is not the drift of the real spot rate u, but
the drift of another rate, called the risk-neutral spot
rate. This rate has a drift of u w.

Recall that for equities we have a real drift rate denoted

by . However, we price as if the asset grows with a rate
r, the risk-free rate.

25
For …xed-income products the real growth of the spot
interest rate may be u(r; t) but we price as if it were
u(r; t) (r; t)w(r; t).

The latter is the risk-adjusted drift rate.

When pricing interest rate derivatives it is important to

model, and price, using the risk-neutral rate.

The risk-neutral spot rate evolves according to

dr = (u w)dt + w dX: (15)

We can use Monte Carlo simulations for pricing …xed-

income products, but we ensure that we simulate the
risk-neutral spot rate process.

26
The relationship between prices and ex-
pectations

The basis of the Monte Carlo method is to price options

using expectations. In the equity world we can write

V (S; t) = e r(T t)EQ[Payo¤];

when interest rates are constant.

We can write

RT
V (S; t) = e t r( ) d
EQ[Payo¤];

when rates are deterministic, r(t),

27
When we have a …xed-income product and rates are sto-
chastic the present value term must go inside the expec-
tation. . .

RT
V (S; t) = EQ e t r( ) d
[Payo¤] :

The value of a derivative is the risk-neutral expectation

of the present value of the payo¤.

If we wish to price an interest rate derivative then the

payo¤ will be in terms of r; i.e. Payo¤(r) :

28
Rule of Thumb:

The new term, market-price-of-risk arises because our

modelled underlying variable, r, is not traded.

The general rule is

If a modelled quantity is traded then the risk-neutral

growth rate is r

r=

If a modelled quantity is not traded then the risk-

neutral growth rate is

real growth market price of risk volatility

If we start with a model for Geometric Brownian Motion

dS = Sdt + SdX
then the pricing equation is
@V 1 2 2 @ 2V @V
+ S 2
+( )S rV = 0 (16)
@t 2 @S @S
29
but S is traded which means it satis…es (16) ; substituting
V = S gives

( )S 1 rS = 0
r =
which is the risk-neutral growth (called r) =) S =
r
which is the Sharpe ratio.

For an interest rate we have

a+u
r =
w

30
Tractable models and solutions of
the pricing equation
This forms the basis of obtaining analytical/closed form
solutions of the bond pricing equation.

We have built up the pricing equation for an arbitrary

model.

That is, we have not speci…ed the precise form of the

risk-neutral drift, u w, or the volatility, w.

How can we choose these functions to give us a good

model?

There are two ways to proceed:

Choose a model that matches reality as closely as

possible - use numerical schemes for PDEs
31
 Choose a model which is easy to solve - these are
called tractable models

Certain choices for risk-neutral drift and di¤usion give

analytical solutions of the BPE. Let’s examine some, to-
gether with the conditions on drift and volatility that lead
to tractable models, that is, models for which the solu-
tion of the pricing equation for zero-coupon bonds can
be found analytically.

32
Named models

Unlike equities, there are numerous models for interest

rates. We have already seen that they possess the prop-
erty to become negative and a mean reverting behaviour.
There are many interest rate models, associated with the
names of their inventors. here are just a few

Vasicek

Cox, Ingersoll & Ross

Ho & Lee

Hull & White

and many more.

33
Vasicek

The Vasicek model, developed in 1977 (for the risk-neutral

spot rate) takes the form

dr = ( r )dt + 1=2dX:

Thus the pricing equation is

@V 1 @ 2V @V
+ 2
+( r) rV = 0:
@t 2 @r @r

And the …nal condition for a zero-coupon bond is

V (r; T ) = 1:

And look for a solution of the form

34
 V (r; t; T ) = eA(t) rB(t):

A = A (t; T )
B = B (t; T )

So substitute this expression for V into the BPE. We

need three derivative terms

@ A(t) rB(t)
e = A_ (t) rB_ (t) V
@t

where _ is a time dependent derivative, i.e. d=dt.

@
eA(t) rB(t) = B (t)V;
@r

and
@2 A(t) rB(t) = B (t)2 V:
e
@r2
35
Substituting these expressions into the pricing equation
for the Vasicek model and dividing through by V we get

1
A_ (t) rB_ (t) + B (t)2 ( r )B (t) r = 0:
2

Grouping the terms, we have an expression that is linear

in r:

1
A_ (t) + B (t)2 B (t) +r B_ (t) + B (t) 1 = 0:
2

Both of the expressions in parentheses must be zero.

We have two …rst order ordinary di¤erential equations for

A(t) and B (t):

B_ (t) B (t) = 1
1
A_ (t) + B (t)2 B (t) = 0
2

36
In order for the …nal condition at t = T to be satis…ed
we need

eA(T ) rB(T ) = 1 =)
A(T ) rB (T ) = 0 8r

and so
A(T ; T ) = B (T ; T ) = 0:

So start by solving for B (t): This can be tackled as a

linear equation with integrating factor e t; so multiply
through

e tB
_ (t) e t B (t) = e t
d
e tB (t) = e t
Z T
dt Z T
d e s B (s ) = e s ds
t t
s B (s ) T 1 sT
e = e
t t
T B (T ) t B (t) 1 T t
e e = e e
37
The solution is

1 (T t) )
B= (1 e

To obtain A(t) :
dA 1
= B (t)2 + B (t)
Z T
dt 2 Z
1 T
dA (s) = (1 e 2 (T s) 2e (T s) )ds
t 2 2 t
Z T
+ (1 e (T s) )ds
t
!T
1 2
A (t) = s e 2 (T s) e (T s)
2 2 2 t
!T
1 (T s)
+ s e
t

and

1 1 B (t; T )2
A(t; T ) = 2 (B (t; T ) T + t)( ) :
2 4

38
Cox, Ingersoll & Ross

The CIR model is very similar to Vasicek but has a vari-

able di¤usion term to address the Vasicek problem, where
rates can become negative. Drift structures for both
SDEs are the same

p
dr = ( r )dt + r dX:

The spot rate is mean reverting and if > =2 the spot

rate stays positive.

The value of a zero-coupon bond is again of the form

V = eA(t) rB(t);

for di¤erent (and more complicated) functions A and B .

The BPE becomes

39
 @V 1 @ 2V @V
+ r 2 +( r) rV = 0:
@t 2 @r @r
giving two ODEs
1
B_ (t) = B (t)2 + B 1
2
A_ (t) = B (t)
Although the …rst ODE is a Riccati type equation,
dy
= a ( x ) y 2 + b ( x ) y c ( x)
dx
we don’t have a particular solution from which to obtain
a more general one. The following integrals should be
used as a hint (although the working is a little messy):
Z
dx 1 c x
= ln
x2 c 2 2c c+x !
Z
dx 1 eax
ax
= ln
b + ce ab b + ceax

40
Ho & Lee

Ho & Lee introduced some time dependence into the co-

e¢ cient of their random walk. This allows the yield curve
to be …tted.

dr = (t)dt + 1=2dX:

The value of zero-coupon bonds is again given by

V = eA(t) rB(t)
The BPE becomes

B_ (t) = 1
1
_
A(t) = B 2 + (t)B (t)
2
which we integrate over t and T
Z T Z T
dB = ds !
t t
B (T ) B (t) = (T t)
| {z }
=0
41
so
B (t) = (T t)
and for A (t)
Z T T Z T Z
1 2
dA(s) = B ds + (s)B (s)ds
t 2 t t
A (T ) A (t)
Z T Z T
1
= (T s)2 ds + (s) (T s) ds
2 t t
Z T
1
A (t) = (T s )3 + (s) (T s) ds
6 t
Z T
1
A (t) = (s)(T s)ds + (T t)3:
t 6

This model was the …rst ‘no-arbitrage model’of the term

structure of interest rates.

By this is meant that the careful choice of the function

(t) will result in theoretical zero-coupon bonds prices,
output by the model, which are the same as market prices.

This technique is also called yield curve …tting.

42
This careful choice is

@2
(t) = log ZM (t ; t) + (t t )
@t2

where today is time t = t .

In this ZM (t ; T ) is the market price today of zero-

coupon bonds with maturity T .

43
Hull & White

This process of generalisation (i.e. time dependency)

continues with the Hull & White. They extended both
the Vasicek and the CIR models to incorporate time-
dependent parameters:

1) dr = ( (t) (t)r)dt + (t)1=2dX

q
2) dr = ( (t) (t)r)dt + (t)r dX

This time dependence again allows the yield curve (and

even a volatility structure) to be …tted.

In general consider we can write:

dr = ( (t) (t)r)dt + (t)r dX

We then have the following models

= 1 : Black Derman & Toy

44
 = 12 : CIR

= 0 : Extended Vasicek (contains Ho and Lee as a

special case)

45
Black, Derman & Toy

This model is popular because it is relatively easy to cali-

brate and, with two functions to be …tted - this allows the
initial yield curve volatilities to also be calibrated. The
risk-neutral random walk is
!
d (log (t))
d (log r) = (t) log r dt + (t) dX
dt
This BDT model is not in the a¢ ne class.

46
A more general model
We have seen A¢ ne Models, i.e. ones which give ZCB
prices of the form:

V (r; t; T ) = exp (A (t; T ) rB (t; T ))

What are the conditions of the drift and di¤usion for an
analytical solution of the above form to exist?

The most general form of risk-adjusted random walk for

which this is possible has structure given by

u(r; t) (r; t)w(r; t) = (t) (t)r;

q
w(r; t) = (t)r + (t):

So this choice of u and w ensures a "nice" form for the

price.

Before we derive these conditions let us consider some

important properties:
47
 Interest rates should remain positive

Large rates eventually decrease: mean reversion

The model should be relatively easy to calibrate and

use.

Substitute V (r; t; T ) into the BPE to give

@A @B 1 2 2
r + w B (u w )B r = 0:
@t @t 2

Some of these terms are functions of t and T (i.e. A and

B ) and others are functions of r and t (i.e. u and w).

Since we require a rule for r we di¤erentiate with respect

to r once:

48
 @B 1 2 @ @
+ B (w 2 ) B (u w) 1 = 0:
@t 2 @r @r

Di¤erentiate again with respect to r and divide through

by B :

1 @2 @2
B 2 (w 2 ) (u w ) = 0:
2 @r @r2

In this, only B can depend on the bond maturity T ,

therefore we must have each term equal to zero:

@2 2) = 0
( w
@r2
@2
2
(u w ) = 0:
@r

R @2 2 )dr = 0 R @2
@r 2 ( w @r2
(u w)dr = 0
@ (w 2 ) = (t) @ (u w) = (t)
R @@r 2 R R @@r R
@r (w ) = (t) dr @r (u w) = (t) dr
w2 = (t) r + (t) (u w) = (t) r + (t)
49
Therefore u w and w2 must be linear in r.

The equations for A and B are

@A 1
= (t)B (t)B 2
@t 2

and
@B 1
= (t)B 2 + (t)B 1:
@t 2

In order to satisfy the …nal data that V (r; T ; T ) = 1 we

must have

A(T ; T ) = 0 and B (T ; T ) = 0:

50
Popular one-factor spot-rate models
The real spot rate r satis…es the stochastic di¤erential
equation
dr = u(r; t)dt + w(r; t)dX:

Model u(r; t) (r; t)w(r; t) w(r; t)

Vasicek a br c
CIR a br cr 1=2
Ho & Lee a(t) c
Hull & White I a(t) b(t)r c(t)
Hull & White II a(t) b(t)r c(t)r 1=2
General a¢ ne a(t) b(t)r (c(t)r d(t))1=2

Here (r; t) denotes the market price of risk. The func-

tion u w is the risk-adjusted drift.

For all of these models the zero-coupon bond value is of

the form Z (r; t; T ) = eA(t;T ) rB(t;T ).

The time-dependent coe¢ cients in all of these models

allow for the …tting of the yield curve and other interest-
rate instruments.
51
Multi-factor interest rate modelling
The simple one-factor stochastic spot interest rate mod-
els cannot hope to capture the rich yield-curve structure
found in practice: from a given spot rate at a given time
they will predict the whole yield curve.

Generally speaking, the one source of randomness, the

spot rate, may be good at modelling the overall level of
the yield curve but it will not necessarily model shifts in
the yield curve that are substantially di¤erent at di¤erent
maturities.

For some instruments this may not be important. For

example, for instruments that depend on the level of the
yield curve it may be su¢ cient to have one source of ran-
domness, i.e. one factor.

More sophisticated products depend on the di¤erence be-

tween yields of di¤erent maturities and for these products
52
it is important to model the tilting of the yield curve.

One way to do this is to invoke a second factor, a second

source of randomness.

53
Theoretical framework

We now use two state equations to model simple interest

rate. One captures the evolution of the spot interest rate
r, and another independent variable l. We use

dr = u dt + w dX1
and
dl = p dt + q dX2:

Simple instruments will then have prices which are func-

tions of r, l and t, V (r; l; t).

All of u, w, p and q are allowed to be functions of r,

l and t. The correlation coe¢ cient between dX1 and
dX2 may also depend on r, l and t.

Possible choices for l :

54
 Another interest rate

long rate

volatility of spot rate

stochastic MPOR

Since we have two sources of randomness now, in pricing

one bond we must hedge with two others to eliminate
the risk:

= V (r; l; t; T ) 1 V1 (r; l; t; T1 ) 2 V2 (r; l; t; T2 ):

As before the distinguishing features of V1 and V2 are the

maturities T1 and T2; in turn.

The change in the value of this portfolio is given by

d = dV (r; l; t; T ) 1 dV (r; l; t; T1 ) 2 dV (r; l; t; T2 ):

55
We know from an extension of Itô that
@V @V @V
dV = dt + dr + dl +
@t @r @l
1 @ 2V 2 1 @ 2V 2 @ 2V
dr + dl + drdl
2 @r2 2 @l2 @r@l
Now

dr2 = w2dt
dl2 = q 2dt
drdl = ?
For drdl we use the correlation
h p p i
E [dX1dX2] = E 1 dt 2 dt = dtE [ 1 2]
= dt
So
drdl = w qdt
Substituting all into dV and rearranging gives
!
@V 1 @ 2V
1 @ 2V @ 2V
dV = + w2 2 + q 2 2 + wq dt +
@t 2 @r 2 @l @r@l
@V @V
dr + dl:
@r @l

56
We will de…ne the coe¢ cient of dt as
@V 1 2 @ 2V 1 2 @ 2V @ 2V
L(V ) = + w 2
+ q 2
+ wq
@t 2 @r 2 @l @r@l
where
@ 1 2 @2 1 2 @2 @2
L + w 2
+ q 2
+ wq
@t 2 @r 2 @l @r@l
is the Black-Scholes operator, therefore
@V @V
dV = L(V ) + dr + dl
@r @l
Similarly for V1 and V2 we have
@V1 @V1
dV1 = L(V1) + dr + dl
@r @l
@V2 @V2
dV2 = L(V2) + dr + dl
@r @l
Hence
@V @V
d = L(V )dt + dr + dl
@r @l
@V1 @V1
1 L ( V1 ) dt + dr + dl
@r @l
@V2 @V2
2 L ( V2 ) dt + dr + dl
@r @l
and grouping similar terms together gives
d = (L(V ) 1 L(V1 ) 2 L(V2 )) dt
57
 @V1 @V2 @V1 @V2
+ @V
@r 1 @r
@V
2 @r dr + @l 1 @l 2 @l dl:
(17)

To eliminate risk, we want the coe¢ cients of dr and dl

to be zero (as they contain the uncertainty). That is

d = (L(V ) 1 L(V1 ) 2 L(V2 )) dt

which means no arbitrage, hence the change in the port-

folio occurs at the risk-free rate

d = (L(V ) 1 L(V1 ) 2 L(V2 )) dt

= r dt
= r (V 1 V{z
1 2 V2 )}dt:
|
=

Rearranging

(L(V ) 1 L(V1 ) 2 L(V2 )) = r (V 1 V1 2 V2 )

gives

(L(V ) rV ) 1 (L(V1 ) rV1) 2 (L(V2 ) rV2) = 0

Now choose 1 and 2 to make the coe¢ cients of dr
and dl in (17) equal to zero. The corresponding portfolio
58
is risk free and should earn the risk-free rate of interest, r.

We thus have the three equations

@V @V1 @V2
1 2 = 0;
@r @r @r
@V @V1 @V2
1 2 =0
@l @l @l
and
L0(V ) 0
1 L (V1 )
0
2 L (V2 ) =0
where
L0(V ) = L(V ) rV:

N.B The primes are not derivatives.

So we have 3 equations in 2 unknowns.

N equations in (N + 1) unknowns is called an over-

prescribed system. The linear system can be expressed

59
as a matrix problem
0 10 1
L0(V )
L0(V1) 0
L (V2) 1
B CB C
@ @V =@r @V1=@r @V2=@r A @ 1 A= 0
@V =@l @V1=@l @V2=@l 2
| {z }
=M
and for the system to be consistent (i.e. solution exists)
we require
det M = 0

If det M = 0 then the three rows are linearly dependent

so we can write the …rst row as a linear combination of
the other two, i.e.
@V @V
L0(V ) = 1 + 2
@r @l
for each V; V1; V2:

1; 2 are arbitrary functions of r; l; t but not T; T1; T2.

Put

1 = rw u
2 = lq p
60
We can therefore write

@V @V
L0(V ) = ( rw u) + ( lq p)
@r @l

where the two functions r (r; l; t) and l (r; l; t) are the

market prices of risk for r and l respectively, and are again
independent of the maturity of any bond.

In full, we have

@V 1 2 @ 2V @ 2V 1 2 @ 2V
+ w 2
+ wq + q 2
+
@t 2 @r @r@l 2 @l

@V @V
(u r w) + (p lq) rV = 0: (18)
@r @l

The model for interest rate derivatives is de…ned by the

choices of w, q , , and the risk-adjusted drift rates u
r w and p lq.

This is yet another parabolic partial di¤erential equation,

a multi-factor model
61
General a¢ ne model
If r and l satisfy the following:

the risk-adjusted drifts of both r and l are linear in

r and l (but can have an arbitrary time dependence)

the random terms for both r and l are both square

roots of functions linear in r and l (but can have an
arbitrary time dependence)

the stochastic processes for r and l are uncorrelated

then the two-factor pricing equation (18) has a solution

for a zero-coupon bond of the form

V (r; l; t; T ) = eA(t) B(t)r C(t)l :

Substituting this into the 2 factor BPE gives a set of

ordinary di¤erential equations for A, B and C
62
 @V
= A(t) B (t)r C (t)l V
@t
@V @ 2V 2V
= BV ! = B
@r @r2
@V @ 2V 2V
= CV ! = C
@l @l2
@ 2V @ 2V
= BCV =
@r@l @l@r

1
A(t) B (t)r C (t)l + w2B 2 + wqBC
2
1 2 2
+ q C B (u r w) C (p lq) r
2
= 0

B (t) 1 r C (t) l
1 2 2 1 2 2
+A(t) + w B + wqBC + q C
2 2
B (u r w) C (p lq)
= 0

63
Calibration

Introduction

The common approach used is to choose a favourite

model and pick parameters so that the price of ZCBs
are correct. The theoretical price can be wrong so we
go back and change the parameters to obtain the correct
value from market traded instruments. The methodology
is called CALIBRATION or FITTING. This is what banks
do!

People like calibration because we choose a well known

model and ensure that theory and market prices match
up, then use it for the pricing of other instruments. So in
other words …rstly do a …t so that the theoretical model
gives the correct ZCB price and then use it for pricing
other complex interest rate products.

In simple terms take a popular model and ensure the

parameter which was/were previously constant is/are now
time dependent - thus introducing more ‡exibility. So
64
every time you see a constant, make it a function of time.
This gives an extra degree of freedom - so introduce time
dependency in the model itself.

Easiest to do with the Ho and Lee model. Historically

this was the …rst …nance model speci…cally developed for
…tting. This class of model is called No Arbitrage as
opposed to equilibrium models.

The beauty is its simple form as a SDE.

So in using any model we have to decide how to choose

the parameters.

Should the parameters be chosen to match

the market yield curve? Or

historical interest rate data?

65
The former is calibration to a snapshot of the market at
one instant in time.

The latter is …tting to time series data.

66
Because of this need to correctly price liquid instruments,
the idea of yield curve …tting or calibration has become
popular.

When stochastic models are used in practice they are al-

most always …tted.

To match a theoretical yield curve to a market yield curve

requires a model with enough degrees of freedom. (You
are matching a curve, i.e. an ‘in…nite’number of points,
so you need in…nite degrees of freedom!)

This is done by making one or more ‘parameters’ time

dependent.

This functional dependence on time is then carefully

chosen to make an output of the model, the price of
zero-coupon bonds, exactly match the market prices
for these instruments.
67
Ho & Lee

The Ho & Lee spot interest rate model is the simplest

that can be used to …t the yield curve.

We saw earlier that the process for the risk-neutral spot

rate evolves according to

dr = (t)dt + cdX:

This is an a¢ ne model with solution of the BPE of the

form

Z (r; t; T ) = eA(t;T ) r(T t);

where we calculated

B (t; T ) = (T t)
Z T
1
A(t; T ) = (s)(T s)ds + c2(T t)3:
t 6

68
(Note that the variables are r and t, but we are also ex-
plicitly referring to the parameter T , the bond maturity.)

The normal way of progressing is that knowing (t) gives

A(t; T ); hence we can determine Z (r; t; T ):

We are not going to do this.

Working forwards: If we know (t) then the above

gives us the theoretical value of zero-coupon bonds of all
maturities. That is start with model ( (t)) and …nd an-
swer (Z ).

An inverse problem: But what if we know Z from the

market, but don’t know the unobservable ? Turn this
relationship around and ask the question

‘What functional form must we choose for (t) to

make the theoretical value of the discount rates for
all maturities equal to the market values?’
69
That is calibration.

Suppose we want to calibrate our model today, time t .

Today’s spot interest rate is r (which tomorrow will
be di¤erent) and the discount factors in the market are
ZM (t ; T ). Suppose today is t =Wednesday 20 May
2008.

Call the special, calibrated, choice for , (t).

We look at the prices of ZCBs on our screens on date t .

Each ZCB has a certain maturity T:

Today we can say what the price is of a one year bond

(T = 1) ; or 10 year bond (T = 10) ; etc. These are all
prices on 20 May 2008.

To match the market and theoretical bond prices, we

must solve

70
 ZM (t ; T ) = eA(t ;T ) r (T t ): (1)

Taking logarithms of this

log(ZM (t ; T )) = A rB
We know the forms of A and B
Z T
1
log ZM = c2(T t )3 (s)(T s)ds r (T t )
6 t
rearranging gives

Z T
(s)(T s)ds
t
1
= log(ZM (t ; T )) r (T t ) + c2 ( T t )3 :
6

We know everything on the right-hand side. So this is an

integral equation for (t).

It is called an integral equation because the unknown

term (which we are solving for) is under the integral
sign.
71
Fortunately it is a simple equation and can be solved by
a technique called di¤erentiation under the integral sign.
The method is known as Leibniz Rule.

The shortened version of this is

Z Z
@ x x @F (y; x)
F (y; x)dy = F (x; x) + dy
@x a a @x

Note that we di¤erentiate w.r.t. the upper limit.

Observe what happens if we di¤erentiate the integral term

with respect to T .

)
x T
F (y; x) (T s) (s);
y s
@F @F
= (s )
@x @T
F (x; x) (T T ) (T ) = 0

72
First di¤erentiate the l.h.s once with respect to T

Z Z T
d T
(s)(T s)ds = (s)ds + 0
dT t t

Di¤erentiate again

Z Z
d2 T d T
(s)(T s)ds = (s)ds
dT 2 t dT t
= (T ):

So, di¤erentiating the rhs twice with respect to T we get

@2 2 (T
log(ZM ( t ; T )) + c t ):
@t2
The parameters are not functions of the derivative matu-
rity. Interest rates can depend on time, but cannot have
(T ) in the SDE for the spot rate. That is we can’t have
a short term interest rate depend on the maturity of a
bond. Hence the solution is
73
 @2 2 (t
(t) = log(ZM ( t ; t)) + c t ):
@t2

With this choice for the time-dependent parameter (t)

the theoretical and actual market prices of zero-coupon
bonds are the same.

Notes:

Now that we know (t) we can price other …xed

income instruments.

We say that our prices are consistent with the

yield curve.

The same idea can be applied to other spot interest

rate models.

This is an inverse problem, and will typically be sen-

sitive to input date (the Z ).

74