om

i.c
CM2-18: The term structure of interest rates Page 1

ng
si
om
o m
as
.m
w
w
w
The term structure of
interest rates
Syllabus objectives
4.5 Models of the term structures of interest rates

1. Explain the principal concepts and terms underlying the theory of a term
structure of interest rates.
2. Describe the desirable characteristics of models for the term structure of
interest rates.
3. Apply the term structure of interest rates to modelling various cash flows,
including calculating the sensitivity of the value to changes in the term
structure.
4. Describe, as a computational tool, the risk-neutral approach to the pricing
of zero-coupon bonds and interest rate derivatives for a general
one-factor diffusion model for the risk-free rate of interest.
5. Describe, as a computational tool, the approach using state price deflators
to the pricing of zero-coupon bonds and interest rate derivatives for a
general one-factor diffusion model for the risk-free rate of interest.
6. Demonstrate an awareness of the Vasicek, Cox-Ingersoll-Ross and
Hull-White models for the term structure of interest rates.
7. Discuss the limitations of these one-factor models and show an awareness
of how these issues can be addressed.

The Actuarial Education Company © IFE: 2019 Examinations

 om
i.c
Page 2 CM2-18: The term structure of interest rates

ng
si
om
o m
0 Introduction

as
.m
w
In this chapter we will look at stochastic models for the term structure of interest rates. In

w
w
particular, we will focus on models that are framed in continuous time and which are
arbitrage-free.

Continuous-time models are generally used for their greater tractability compared to
discrete-time ones. An arbitrage-free model of the term structure is one that generates arbitrage-
free bond prices.

There are two main types of models used to describe interest rates mathematically:

1. The Heath-Jarrow-Morton approach uses an Ito process to model the forward rate for an
investment with a fixed maturity. We will not consider this approach here.

2. Short-rate models use an Ito process to model the short rate. We will look at three
specific models of this type: the Vasicek model, the Cox-Ingersoll-Ross model and the
Hull-White model.

Ito processes are a key feature of these models, so it might be helpful to review the topics of
Brownian motion, Ito’s Lemma and stochastic differential equations from earlier in the course.

Interest rate modelling is the most important topic in derivative pricing. Interest rate
derivatives account for around 80% of the value of derivative contracts outstanding, mainly
swaps and credit derivatives used to support the securitisation of debt portfolios. More
philosophically, is the fact that at any one time there will be a multitude of contracts (bonds)
written on the same underlying (an interest rate), and derivative pricing is principally
concerned with pricing derivatives coherently.

The Core Reading in this chapter is adapted from the course notes written by Timothy
Johnson.

© IFE: 2019 Examinations The Actuarial Education Company

 om
i.c
CM2-18: The term structure of interest rates Page 3

ng
si
om
o m
1 Notation and preliminaries

as
.m
w
1.1 Zero-coupon bonds

w
w
The multitude of traded instruments leads to the first challenge in interest rate modelling:
the multitude of definitions of interest rates.

Modelling interest rates is more complicated than modelling share prices because interest rates
depend not only on the current time (which we will denote by t ), but also on the term of the
investment. For example, an investor with a 10-year bond will normally earn a different rate of
interest than an investor with a 5-year bond.

The basic debt instrument is the discount bond (or, equivalently, the zero-coupon bond).
This is an asset that will pay one unit of currency at time T and is traded at time t  T . If the
interest rate, R, is constant between t and T then we can say that the price of the discount
bond purchased at t and maturing at T is given by P (t ,T ) where:

1
P (t ,T ) 
(1  R (t ,T ))(T  t )

The spot rate R(t ,T ) is the effective rate of interest applicable over the period from time t to
time T that is implied by the market prices at time t .

Observe that P (T ,T )  1 and for all t  T , P (t ,T )  P (T ,T )  1 . We define   T  t in what

follows.

The discrete bond yield calculated from discount bond prices is:

1
R (t , t   )  1
P (t , t   )1/

Alternatively we could write this as:

1
R(t ,T )  1
P(t ,T )1/(T t )

If a ‘spot’ rate is paid m times a year, then:

nm
1  R
  1 
P (0, n)  m

The limit as m   is a continuously compounded rate, r (t ,T ) (‘force of interest’), such

that:

1
e  r (t ,T )  P (t ,T ) 
(1  R (t ,T ))

The continuously compounded bond yield is calculated as:

ln P (t ,T )
r (t ,T )  


The Actuarial Education Company © IFE: 2019 Examinations

 om
i.c
Page 4 CM2-18: The term structure of interest rates

ng
si
om
m
The spot rate r(t ,T ) is the continuously compounded rate of interest applicable over the period

o
as
from time t to time T that is implied by the market prices at time t .

.m
w
w
1.2 Yield curves

w
Fixing t  0 and plotting yield, R(0,T ) or r (0,T ) , against maturity, T, gives the yield curve
which gives information on the term structure, how interest rates for different maturities are
related. Typically, the yield curve increases with maturity, reflecting uncertainty about
far-future rates. However, if current rates are unusually high, the yield curve can be
downward sloping, and is inverted.

There are various theories explaining the shape of the yield curve. The expectations theory
argues that the long-term rate is determined purely by current and future expected
short-term rates, so that the expected final value of investing in a sequence of short-term
bonds equals the final value of wealth from investing in long-term bonds.

The market segmentation theory argues that different agents in the market have different
objectives: pension funds determine longer-term rates, market makers determine short-term
rates, and businesses determine medium-term rates, which are all determined by the supply
and demand of debt for these different market segments.

The liquidity preference theory argues that lenders want to lend short term while borrowers
wish to borrow long term, and so forward rates are higher than expected future zero rates
(and yield curves are upward sloping).

Zero rate is another name for the spot rate.

The figure below shows the yield curve for the UK Government bond market on 31 December
2003. The ‘term’ plotted on the x -axis corresponds to T  t .

5.5%

5.0%

4.5%
Yield

4.0%

3.5%
0 5 10 15 20 25 30 35
Term

UK yield curves 31 December 2003

© IFE: 2019 Examinations The Actuarial Education Company

 om
i.c
CM2-18: The term structure of interest rates Page 5

ng
si
om
m
The circles show the remaining term and the gross redemption yield for each of the available

o
as
bonds. These approximate to the spot rates r(t ,T ) for zero-coupon bonds with corresponding

.m
terms. A mathematical curve has been fitted to these points. At this particular time the yield

w
w
curve had a humped shape.

w
The other curve shown is the forward rate curve f (0,t ,T ) , which is defined in the next section.

1.3 Short rate and forward rates

The short or instantaneous rate, r (t ) , is the interest rate charged today for a very short
period (ie overnight). This is defined (equivalently) as:

r (t )  r (t , t   )  R (t , t   )

where  is a small positive quantity. So the short rate r (t) is the force of interest that applies in
the market at time t for an infinitesimally small period of time  . Using the relationship
developed in the opening section we have:


r (t )   ln P (t , t   )


The short rate is often the basis of some interest rate models; however, it will not generate,
on its own, discount bond prices.

The forward rate, F (0, t ,T ) if discretely compounded and f (0, t ,T ) if continuously

compounded, relates to a loan starting at time t, for the fixed forward rate, the forward rate,
repaid at maturity, T. It involves three times, the time at which the forward rate agreement is
entered into (typically 0), the start time of the forward rate, t and the maturity of the forward
rate agreement, T.

The law of one-price/the no-arbitrage principle, implies:

1
 P (0, t )  T t
F (0, t ,T )    1
 P (0,T ) 

At time t we can consider the market prices of two investments, one maturing at time t and one
at a later time T . These two prices will imply a certain rate of interest applicable between time t
and time T . This is F (0,t ,T ) . In an arbitrage-free market, it represents the rate of interest at
which we can agree at time 0 to borrow or lend over the period from t to T .

F(0,t,T)

0 t T

The Actuarial Education Company © IFE: 2019 Examinations

 om
i.c
Page 6 CM2-18: The term structure of interest rates

ng
si
om
o m
Question

as
.m
Suppose that the current time corresponds to t  5 and that the force of interest has been a

w
w
w
constant 4% pa over the last 5 years. Suppose also that the force of interest implied by current
market prices is a constant 4% pa for the next 2 years and a constant 6% pa thereafter.

If T  10 and S  15 , write down or calculate each of the four quantities P(t ,T ) , r (t) , f (t ,T , S)
and r (t ,T ) using the notation above.

Solution

4% 6%

t=5 T=10 S=15

 P(5,10)  e(20.04  30.06)  e 0.26  0.771

 r(5)  0.04

 f (5,10,15)  0.06

1
 r (5,10)  (2  0.04  3  0.06)  0.052
5

For continuously compounded forward rates:

r (0,T )T  r (0, t )t
f (0, t ,T ) 
T t
(r (0,T )  r (0, t ))T
 r (0, t ) 
T t

The forward rate is related to the zero-coupon bond price as follows:

1 P(0 t)
f (0 tT )  log for t  T
T t P(0T )

Question

Derive this relationship.

© IFE: 2019 Examinations The Actuarial Education Company

 om
i.c
CM2-18: The term structure of interest rates Page 7

ng
si
om
o m
Solution

as
.m
w
One way to see this is to say that the price for a bond purchased at time 0 and maturing at time T

w
should equal the price for a similar bond maturing at the earlier time t but discounted for the

w
extra period T  t at the forward rate applicable over that period. In other words:

P(0T )  P(0 t)exp  f (0 tT )(T  t)

Rearranging this equation gives the required result:

1 P(0 t)
f (0 tT )  log
T t P(0T )

Alternatively we could use the Core Reading results from above:

r (0,T )T  r (0,t)t
f (0,t ,T ) 
T t
 log P(0,T )  T  log P(0,t)  t

 T t
T t
log P(0,t)  log P(0,T )

T t
1 P(0 t)
 log
T t P(0T )

The instantaneous rate f (0,T ) is defined as lim f (0,t ,T ) .

t T

From the Core Reading above we have that:

(r (0,T )  r (0,t))T
f (0,t ,T )  r (0,t) 
T t

So, in the limit t  T , we get the instantaneous forward rate:

r (0,T )
f (0,T )  r (0,T )  T
T

 ln(P (0,T ))
T

Question

 1  
Show that r (0T )    log P(0T )  r (0T )  .
T T  T 

The Actuarial Education Company © IFE: 2019 Examinations

 om
i.c
Page 8 CM2-18: The term structure of interest rates

ng
si
om
o m
Solution

as
.m
w
1
By applying the product rule to the equation r (0T )   log P(0T ) , we have:

w
w
T

   1 
r (0T )    log P(0T ) 
T T  T 
1  
log P(0T )  log P(0T )   
1

T T T  T 

1   1 
 log P(0T )  log P(0T )  2 
T T T 
1  
   log P(0T )  r (0T ) 
T  T 

So f (0,T ) , the instantaneous forward rate, is the force of interest at future time T implied by the
current market prices at time 0. Then the short rate r(T ) is given by r (T )  f (T ,T ) .


We can generalise lim f (0,t ,T )  f (0,T )   ln(P(0,T )) to give:
t T T


f (t ,T ,T )  f (t ,T )   ln(P(t ,T ))
T

Hence, the fundamental theorem of calculus tells us that:

 T 
P (t ,T )  exp    f (t , u, u )du 
 t 

 T 
 exp    f (t ,u)du 
 t 

We can also deduce the following relationship, which shows that the spot rate is an average of
the forward rates:

1 1 T
T  t t
r (tT )   log P(tT )  f (t ,u u)du
T t

Question

Under one particular term structure model:

f (t ,T )  0.03e 0.1(T t)  0.06(1  e 0.1(T t ) ) .

Sketch a graph of f (t ,T ) as a function of T , and derive expressions for P(t ,T ) and r (t ,T ) .

© IFE: 2019 Examinations The Actuarial Education Company

 om
i.c
CM2-18: The term structure of interest rates Page 9

ng
si
om
o m
Solution

as
.m
The factor e 0.1(T t ) equals 1 when T  t , but then decreases exponentially to zero as T   .

w
w
w
So f (t ,T ) is a weighted average of 0.03 and 0.06, and its graph will increase from 0.03 (ie a force
of interest of 3%) to 0.06.

7%
6%
5%
4%
3%
2%
1%
0%

We can find P(t ,T ) from the relationship P(tT )  exp    f (t u)du  :
T
 t 

P(tT )  exp    f (t u)du 

T
 t 

 t
T

 exp    0.03e 0.1(ut )  0.06(1  e 0.1(ut ) ) du 


 t
T
 
 exp    0.06  0.03e 0.1(ut ) du 


 T
 exp   0.06u  0.3e 0.1(ut)  
  t 


 exp 0.06(T  t)  0.3e 0.1(T t )  0.3 
1
We can then find r(t ,T ) from the relationship r (tT )  log P(tT ) :
T t

1
r (tT )  log P(tT )
T t


1
T t

0.06(T  t)  0.3e 0.1(T t)  0.3 
 1  e 0.1(T t ) 
 0.06  0.3  
 T t 

The Actuarial Education Company © IFE: 2019 Examinations

 om
i.c
Page 10 CM2-18: The term structure of interest rates

ng
si
om
o m
2 Desirable characteristics of a term structure model

as
.m
w
Question

w
w
What do you think term structure models are used for?

Solution

The main uses of term structure (interest rate) models are:

 by bond traders looking to identify and exploit price inconsistencies
 for calculating the price of interest rate derivatives
 by investors with a portfolio involving bonds or loans who want to set up a hedged
position
 for asset-liability modelling.

Equilibrium models start with a theory about the economy, such that interest rates revert to
some long-run average, are positive or their volatility is constant or geometric. Based on
the model for (typically) the short rate, the implications for the pricing of assets is worked
out. Examples of equilibrium models are Rendleman and Bartter, Vasicek and Cox-
Ingersoll-Ross.

Being based on ‘economic fundamentals’, equilibrium models rarely reproduce observed

term structures. This is unsatisfactory.

No-arbitrage models use the term structure as an input and are formulated to adhere to the
no-arbitrage principle. An example of a no-arbitrage model is the Hull-White (one- and two-
factor).

The implication here is that the Vasicek and the Cox-Ingersoll-Ross models permit arbitrage
opportunities. This is not actually true; both of these models result in no-arbitrage bond price
formulae, even though their underlying construction principles are based in economic theory.

We will now discuss characteristics of a term structure model that are regarded as desirable
features.

 The model should be arbitrage free.

In very limited circumstances this is not essential, but in the majority of modern actuarial
applications, it is essential. Most obviously, anything involving dynamic hedging would
immediately identify and exploit any arbitrage opportunities.
The markets for government bonds and interest rate derivatives are generally assumed to
be pretty much arbitrage-free in practice.

© IFE: 2019 Examinations The Actuarial Education Company

 om
i.c
CM2-18: The term structure of interest rates Page 11

ng
si
om
m


o
Interest rates should ideally be positive. Banks have to offer investors a positive return to

as
prevent them from withdrawing paper cash and putting it ‘under the bed’. This might be

.m
impractical for a large life office or pension fund but, nevertheless, it typically holds in

w
w
practice.

w
Some term structure models do allow interest rates to go negative.
One such example is the Vasicek model we will see later in this chapter.
Whether or not this is a problem depends on the probability of negative interest rates
within the timescale of the problem in hand and their likely magnitude if they can go
negative. It also depends on the economy being modelled, as negative interest rates have
been seen in some countries.

 r (t) and other interest rates should exhibit some form of mean-reverting behaviour.
Again this is because the empirical evidence suggests that interest rates do tend to mean
revert in practice.
This might not be particularly strong mean reversion but it is essential for many actuarial
applications where the time horizon of a problem might be very long.

 How easy is it to calculate the prices of bonds and certain derivative contracts?
This is a computational issue. It is no good in a modelling exercise to have a wonderful
model if it is impossible to perform pricing or hedging calculations within a reasonable
amount of time.
This is because we need to act quickly to identify any potential arbitrage opportunities or
to rebalance a hedged position.
Thus we aim for models that either give rise to simple formulae for bond and option
prices, or that make it straightforward to compute prices using numerical techniques.

 Does the model produce realistic dynamics?

For example, can it reproduce features that are similar to what we have seen in the past
with reasonable probability? Does it give rise to a full range of plausible yield curves,
ie upward-sloping, downward-sloping and humped?

 Does the model, with appropriate parameter estimates, fit historical interest rate data
adequately?
 Can the model be calibrated easily to current market data?
If so, is this calibration perfect or just a good approximation? This is an important point
when we are attempting to establish the fair value of liabilities. If the model cannot fit
observed yield curves accurately then it has no chance of providing us with a reliable fair
value for a set of liabilities.

 Is the model flexible enough to cope properly with a range of derivative contracts?

The Actuarial Education Company © IFE: 2019 Examinations

 om
i.c
Page 12 CM2-18: The term structure of interest rates

ng
si
om
o m
3 Models for the term structure of interest rates

as
.m
w
3.1 The risk-neutral approach to pricing

w
w
We will assume that the short rate is driven by an Ito diffusion:

drt   (t , rt )dt   (t , rt )dW t

where:

   t ,rt  is the drift parameter

   t ,rt  is the volatility parameter

 W t is a Wiener process under the martingale measure.

In actuarial work we are used to assuming a fixed rate of interest in calculations. However, in this
chapter we are considering stochastic models where future interest rates behave randomly. This
means that we need to specify which probability measure we are using.

If we are to have a model that is arbitrage-free then we need to consider the prices of tradeable
assets, with the most natural of these being the zero-coupon bond prices P(t ,T ) .

Modelling the short rate r (t) does not tell us directly about the prices of the assets traded in the
market. To see whether arbitrage opportunities exist or not, we need to examine these prices.

We can use an argument similar to the derivation of the Black-Scholes model using the martingale
approach to demonstrate that:

 
P(tT )  EQ exp    rudu  rt 
T

  t  

where Q is called the risk-neutral measure.

Since we are considering Markov models, any information about the value of the short rate
before time t is irrelevant, and so we can replace rt with its filtration Ft . We can then write this
equation in the equivalent form:

 
P(tT )  EQ exp    rudu   1 Ft 
T

  t  

Since the payoff from the bond at maturity is P(T ,T )  1 , we can see that this equation is
analogous to the valuation formula Vt  EQ e r (T t ) X Ft  , which we met in an earlier chapter for
 
derivatives based on shares.

© IFE: 2019 Examinations The Actuarial Education Company

 om
i.c
CM2-18: The term structure of interest rates Page 13

ng
si
om
m
If we had a derivative whose payoff X was dependent on the future value of the bond, then its

o
as
value at time t would be:

.m
w
w
 
Vt  EQ exp    rudu  X Ft 
T

w
  t  

This formula is interesting because it tells us that, if we assume that the market is arbitrage-free,
we can work out the value of the bond at the current time t based on the current short rate rt
and the possible future values of ru for t  u  T . We don’t need to know the entire term
structure, ie the values of f (t ,u) for t  u  T , at the current time.

We define the bank, or money-market, account process as:

dBt  rt Bt dt , B0  1

and:

 t 
Bt  exp   ru du 
 0 

According to the standard theory, all discounted assets must be martingales under the
martingale measure, or:

P (t ,T )  1 
 EQ  Ft 
Bt B
 T 

Since Bt is known at time t, then we can re-write this as:

B 
P (t ,T )  EQ  t Ft 
B
 T 

The martingale measure (ie the risk-neutral probability measure) is chosen so that this
relationship holds by definition. It is the set of probabilities such that the expected future value of
the payout (which is guaranteed to be P(T ,T )  1 ) discounted from time T to time t (ie Bt / BT ) is
equal to the value of the contract at time t, P(t ,T ) .

Then we have:

  T  
 
P (t ,T )  EQ  exp   ru du  Ft 

  t  

Compare with:

 T 
P (t ,T )  exp    f (t , u )du 
 t 

from earlier.

The Actuarial Education Company © IFE: 2019 Examinations

 om
i.c
Page 14 CM2-18: The term structure of interest rates

ng
si
om
m
The difference between P (t ,T ) and Bt is captured by noting:

o
as
.m
 T   t 

w
P (t ,T )  exp    f (t , u )du  while Bt  exp   ru du 

w
w
 t   0 

So the value of Bt , the cash account at time t, can be calculated using the known past short rates.
In contrast, the price of the bond is the expectation of unknown future values of the short rate.
However, given the known bond prices at time t, we can derive a set of consistent forward rates.

By considering the price of the bond at time t  0 we can see that:

 t   t 

 0

P (0, t )  exp   f (0, u )du 

while 
Bt  exp  ru du  and f (u , u )  ru
 0 

We now assume that the discount bond price, P (t ,T ) , is some deterministic function of the
short-rate process, rt :

P (t ,T )  g (t , rt )

By using the stochastic product rule from the stochastic calculus chapter we have:

 P (t ,T ) 
d
 B t 
 1
  d g (t , rt )Bt 
 d  g(t ,rt ) Bt1  g(t ,rt )d Bt1  
  t  
 d  g(t ,rt ) Bt1  g(t ,rt )d  exp    rudu  
  0  

  t  
 d  g(t ,rt ) Bt1  g(t ,rt )rt  exp    rudu   dt
 
  0  

 Bt1 d  g (t , rt )   rt g (t , rt )dt 

By applying Ito’s Lemma to the function g (t , rt ) we have:

 P (t ,T )  
1 g (t , rt ) g (t , rt ) 1  2 g (t , rt ) 2 
d   Bt    (t , rt )   (t , rt )  rt g (t , rt )  dt
 Bt   t  r 2  r 2 
 t t 
g (t , rt )
 Bt1  (t , rt )dW t
rt

Question

Verify the above result by using Taylor’s theorem.

© IFE: 2019 Examinations The Actuarial Education Company

 om
i.c
CM2-18: The term structure of interest rates Page 15

ng
si
om
o m
Solution

as
.m
w
By Taylor’s theorem in two variables we have:

w
w
 P(t ,T )  1  g(t , rt ) g(t ,rt ) 1 2g(t ,rt ) 2
d   Bt  dt  drt   dr  1
t   Bt rt g(t , rt )dt
B
 t    t rt 2 rt
2


But drt   (t , rt )dt   (t ,rt )dW t so  drt    2 (t ,rt )dt and:

2

 P(t ,T )  1  g(t , rt ) g(t ,rt ) 1 2g(t ,rt ) 2 

d   Bt  dt    (t ,rt )dt   (t ,rt )dW t    (t , rt )dt 

 Bt   t rt 2 rt2 

Bt1rt g(t , rt )dt

 g(t ,rt ) g(t ,rt ) 1 2g(t ,rt ) 2 

 Bt1    (t ,rt )   (t , r )  r g(t , r )  dt
 t r 2 r 2 t t t 
 t t 
g(t ,rt )
Bt1  (t ,rt )dWt
rt

In order for this to be a martingale, we require that:

g (t , rt ) g (t , rt ) 1  2 g (t , rt ) 2
  (t , rt )   (t , rt )  rt g (t , rt )  0
t rt 2 rt2

with the boundary condition that P (T ,T )  g (T , rT )  1 for all rT .

Question

P(t ,T )
Why does the partial differential equation above need to equal zero in order for to be a
Bt
martingale?

Solution

P(t ,T )
Martingale processes have zero drift. In order for to have zero drift the coefficient of the
Bt
dt term in the stochastic differential equation must be zero.

The Actuarial Education Company © IFE: 2019 Examinations

 om
i.c
Page 16 CM2-18: The term structure of interest rates

ng
si
om
o m
3.2 The Vasicek model (1977)

as
.m
Vasicek assumes that:

w
w
w
drt   (  rt )dt   dW t

for constants   0 ,  and,  .

Here  represents the ‘mean’ level of the short rate. If the short rate grows (driven by the
stochastic term) the drift becomes negative, pulling the rate back to  . The speed of the
‘reversion’ is determined by  . If  is high, the reversion will be very quick.

The graph below show a simulation of this process based on the parameter values   0.1 ,
  0.06 and   0.02 .

12%

10%

8%

6%

4%

2%

0%

-2%

Example simulation of short rate from the Vasicek model

This yields the partial differential equation:

g (t , rt ) g (t , rt ) 1  2 g (t , rt ) 2
  (  rt )    rt g (t , rt )  0
t rt 2 rt2

by making the substitution (t ,rt )   (  rt ) and letting  (t ,rt ) be constant.

To establish the form of g (t , rt )  P (t ,T ) , recall that for the Ornstein-Uhlenbeck process:

t
rt    (r0   )e  t   e  (t u )dW u

0

and so, integrating again:

T T

 rsds  T   (r0   ) 1  e
1  T
 

  1  e
 (T  u )
 dW
u
0 0

© IFE: 2019 Examinations The Actuarial Education Company

 om
i.c
CM2-18: The term structure of interest rates Page 17

ng
si
om
m
T

o
as
This implies that  rsds is normally distributed with conditional mean:

.m
0

w
w
w
T 
0  
1
E  rs ds   T  (r0   ) 1  e T

 
 

Using Ito isometry:

T  T
VarQ   ( X t , t )dWt   E  2 ( X t , t ) dt
 
0  0  
 

we can see that:

T  2  2 1 
VarQ  rs ds   2  T  (1  e T ) 
 (1  e 2T ) 
0     2 
 

T
Since  rsds is normally distributed, using the moment generating function of a normal,
0


N  , 2
 , random variable:
 t  21 2t 2
E e tX   e
 

by putting t  1 , we can say that:

  T 
P (0,T )  EQ exp   rs ds  

   
  0 

   1
2 

1


 exp    T  (r0   ) 1  e T    2
 2  
 
 T 
2

(1  e T
) 
1
2
(1  e 2T  
)  
  
  

In general, by setting:

b(t ,T ) 

1
1  e  (T  t )

and:

 2  2 2
a(t ,T )       b(t ,T )  T  t   b (t ,T )
 2 2  4
 

we have:

P (t ,T )  e a (t ,T )  b (t ,T )rt

The Actuarial Education Company © IFE: 2019 Examinations

 om
i.c
Page 18 CM2-18: The term structure of interest rates

ng
si
om
m
By defining   T  t , then equivalently we have:

o
as
.m
P(t ,T )  ea( )b( )rt

w
w
w
where:

b( ) 
1

1  e 
and:

 2  2 2
a( )     2   b( )     b ( )
 2  4


Question

Show that the instantaneous forward rate for the Vasicek model can be expressed as:

 2  2
f (t ,T )  r(t)e      2  (1  e  )  2 (1  e  )e 
 2  2


where   T  t .

Solution

We have a formula for P(t ,T ) under this model. So the instantaneous forward rate can be
derived from this using the relationship:


f (tT )   log P(tT )
T


By use of the chain rule, and noting that  1 , this gives:
T


f (tT )   [a( )  b( )r (t)]
T
 
 [a( )  b( )r (t)]
 T
 
 [a( )  b( )r (t)]  a( )  b( )r (t)
T 

© IFE: 2019 Examinations The Actuarial Education Company

 om
i.c
CM2-18: The term structure of interest rates Page 19

ng
si
om
m
From the definitions of b( ) and a( ) , we find that:

o
as
.m
d  1  e   

w
b( )   e

w
d   

w
 

and:

d   2  2 
a( )  (b( )   )    2   b( )2 
d   2  4 
 

 2  2
 (b( )  1)    2    2b( )b( )
 2  4
 
  2   2  1  e   
 (e   1)    2    e
 2  2   
 
 2  2
 (1  e  )    2   2 (1  e  )e 
 2  2


Substituting these expressions into the general formula for f (t ,T ) gives the required answer.

Question

Write down an expression in terms of the model parameters for the long rate, ie the
instantaneous forward rate corresponding to T  t   , according to the Vasicek model.

Solution

Letting T   (and hence    ) in the equation for f (t ,T ) gives:

 2  2 2
f (t , )  r(t)  0     2  (1  0)  2 (1  0)  0    2
 2  2 2


The curves shown on the graph of gilt yields in Section 1 were fitted using a Vasicek model with
parameter values   0.131 ,   0.083 and   0.037 .

Question

‘The particular model used for the graph implies that interest rates are mean-reverting to the
value   0.083 .’

True or false?

The Actuarial Education Company © IFE: 2019 Examinations

 om
i.c
Page 20 CM2-18: The term structure of interest rates

ng
si
om
o m
Solution

as
.m
w
The dynamics of r (t) for this particular Vasicek model are:

w
w
dr (t)  0.131 [r (t)  0.083]dt  0.037dW (t)

under the risk-neutral probability measure Q . Under this measure W (t) is standard Brownian
motion and therefore has zero drift and the process mean-reverts to the value 0.083.

However, under the real-world probability measure P , W (t) would have non-zero drift and the
process will mean-revert to a different value. In fact, although we will not prove it here, the long-
term rate in the real world can be found from the formula derived in the previous question,
namely:

2
  0.0431 ie 4.31%
2 2

3.3 The Cox-Ingersoll-Ross (CIR) model (1985)

In Vasicek’s model (and Hull-White, below) interest rates are not strictly positive. This
assumption is not ideal for a short-rate model. CIR use the Feller, or square root mean
reverting process which is positive (it can instantaneously touch 0 but immediately re-
bounds):

drt   (  rt )dt   rt dW t

for constants   0 ,   0 and,  .

We can see that the form of the drift of rt is the same as for the Vasicek model. The critical
difference between the two models occurs in the volatility, which is increasing in line with the
square root of rt . Since this diminishes to zero as rt approaches zero, and provided  2 is not too
large ( rt will never hit zero provided  2  2 ), we can guarantee that rt will not hit zero.
Consequently all other interest rates will also remain strictly positive.

The graph below shows a simulation of this process based on the parameter values   0.1 ,
  0.06 and   0.1 .

© IFE: 2019 Examinations The Actuarial Education Company

 om
i.c
CM2-18: The term structure of interest rates Page 21

ng
si
om
o m
as
.m
20%

w
18%

w
w
16%
14%
12%
10%
8%
6%
4%
2%
0%

Simulation from Cox-Ingersoll-Ross model

The associated PDE is:

g (t , rt ) g (t , rt ) 1  2 g (t , rt ) 2
  (  rt )   rt  rt g (t , rt )  0
t rt 2 rt2

The only difference between this and the Vasicek PDE is the inclusion of an rt in the second
derivative term.

Again, P (t ,T )  e a (t ,T )  b(t ,T )rt , with:

2(e   1)
b( ) 
(   )(e   1)  2

and:

2  1 (  )
2 e 2 
a( )  ln  
2  (   )(e  1)  2 
 

 1 (  )

2   e2 
 2 ln b( ) 
  (e  1) 
 

where:

   2  2 2

It turns out that these values for a and b are not that different from those in Vasicek’s
model.

The distribution of rt is given by a ‘non-central chi-squared’ distribution. This is a

‘fat-tailed’ distribution.

The Actuarial Education Company © IFE: 2019 Examinations

 om
i.c
Page 22 CM2-18: The term structure of interest rates

ng
si
om
m
If X1 , X2 ,, Xn are independent random variables, each with a N(0,1) distribution, then

o
as
Y  X12  X22    Xn2 has a chi-square distribution with n degrees of freedom.

.m
w
w
w
If X1 , X2 ,, Xn are independent random variables and Xi  N(di ,1) , then Yd  X12  X22    Xn2 is
said to have a non-central chi-squared distribution with n degrees of freedom and non-centrality
n
parameter d   di2 .
i 1

So the non-central chi-squared distribution can be thought of as a lopsided version of the ordinary
chi-square distribution.

Question

What is the mean of the non-central chi-squared distribution with n degrees of freedom and
non-centrality parameter d ?

Solution

Since Xi  N(di ,1) , we find that:

E[ X i2 ]  Var ( Xi )  [E ( X i )]2  1  di2

It follows that:

n
E[Yd ]  E[ X12  X22    Xn2 ]   (1  di2 )  n  d
i 1

3.4 Vasicek and CIR yield curves

Recall:

ln P (t ,T )
r (t ,T )  
T t

and:

P(t ,T )  ea(t ,T )b(t ,T )rt

so we have:

a(t ,T )  b(t ,T )rt

r (t ,T )  
T t

Alternatively:

a( )  b( )rt

r( )  


© IFE: 2019 Examinations The Actuarial Education Company

 om
i.c
CM2-18: The term structure of interest rates Page 23

ng
si
om
m
Then the yield curves coming out of the Vasicek model are of three (related) types:

o
as
.m
w
w
w
These yield curves are generally described as being:
 normal, with short-term yields being lower than long-term yields
 inverted, with short-term yields being higher than long-term yields
 humped, ie a yield curve with a turning point.

Recall that the duration, D, of an asset, whose interest rate dependent price is given by B, is
defined by:

B
  D y
B

where y is the instrument’s yield. In the context of interest rate models, this is equivalent
to:

P (t ,T )
  DP (t ,T )  b(t ,T )P (t ,T )
rt

and there is a connection between this duration D , and the function b.

The relationship above comes from the fact that:

B B
 D y   BD
B y

B
and is the rate at which the bond price changes with respect to a change in its yield, ie
y
P(t ,T )
.
rt

As P(t ,T )  ea(t ,T )b(t ,T )rt , then:

rt

rt
e 
P(t ,T )  a(t ,T )b(t ,T )rt

 b(t ,T )ea(t ,T )b(t ,T )rt  b(t ,T )P(t ,T )

The Actuarial Education Company © IFE: 2019 Examinations

 om
i.c
Page 24 CM2-18: The term structure of interest rates

ng
si
om
o m
3.5 The Hull-White model (1990)

as
.m
The Hull-White model is an extension of Vasicek where the mean-reversion level,  , is a

w
w
deterministic function of time:

w
drt   ( (t )  rt )dt   dW t

for constants   0 and  .

In some representations, the parameter  is also allowed to be a function of time. This is known
as the extended Vasicek model.

This yields a PDE similar to Vasicek, and so we can start by ‘guessing’ that
P (t ,T )  e a(t ,T )  b(t ,T )rt  g (t , rt ) and so:

g (t , rt ) g (t , rt ) 1  2 g (t , rt ) 2
  ( (t )  rt )    rt g (t , rt )  0
t rt 2 rt2

By noting that:

g(t ,rt )  a(t ,T ) b(t ,T ) 

 g(t ,rt )   rt 
t  t t 

g(t ,rt )
  g(t ,rt )b(t ,T )
rt

 2g(t ,rt )
 g(t ,rt )b2 (t ,T )
rt2

Then the Hull-White PDE becomes:

 a(t ,T ) b(t ,T )  1
g(t ,rt )   rt    g(t ,rt )b(t ,T )( (t)  rt )  g(t , rt )b2 (t ,T ) 2  rt g(t , rt )  0
  t t  2

 a(t ,T ) b(t ,T ) 1 
 g(t ,rt )   rt   b(t ,T )( (t)  rt )  b2 (t ,T ) 2  rt   0
 t t 2 

  b(t ,T )  a(t ,T ) 1 
 g(t ,rt )  rt   b(t ,T )  1    b(t ,T ) (t)  b2 (t ,T ) 2   0
  t  t 2 

a(t ,T ) b(t ,T )
By letting a(t ,T )  and b(t ,T )  then we have:
t t

 1 
g (t , rt )  rt   b(t ,T )  b(t ,T )  1  a (t ,T )   b(t ,T ) (t )  b2 (t ,T ) 2   0
 2 

© IFE: 2019 Examinations The Actuarial Education Company

 om
i.c
CM2-18: The term structure of interest rates Page 25

ng
si
om
m
This is essentially Vasicek but we have:

o
as
.m
T  s 
 

w
b(t ,T )  exp    (u )du  ds

w
 
 t 

w
t

As  has been taken as a constant in the model above we have:

T  s 
b(t ,T )   exp    du  ds
 
t  t 
T
  exp    s  t   ds
t


1
 1  e  T t 

So b(t ,T ) is the same as for the Vasicek model.

T
a(t ,T )     (s)b(s,T )  21 
2 2

b (s,T ) ds
t

Question

Show that a(t ,T ) and b(t ,T ) satisfy the equation

 1 
g(t ,rt )  rt  b(t ,T )  b(t ,T )  1   a(t ,T )   b(t ,T ) (t)  b2 (t ,T ) 2   0
 2 

Solution

By differentiating a(t ,T ) and b(t ,T ) we have:

a(t ,T )
a(t ,T )    (t)b(t ,T )  12  2b2 (t ,T )
t

and

b(t ,T )
 e  
 T t
b(t ,T ) 
t

The Actuarial Education Company © IFE: 2019 Examinations

 om
i.c
Page 26 CM2-18: The term structure of interest rates

ng
si
om
m
Substituting these into the given equation gives:

o
as
 

.m
g(t ,rt ) rt  b(t ,T )  b(t ,T )  1   a(t ,T )   b(t ,T ) (t)  12 b2 (t ,T ) 2

w
w
w
t
 g(t ,rt ) 
 
 r  b(t ,T )  e  T t   1   (t)b(t ,T ) 


  1  2b2 (t ,T )   b(t ,T ) (t)  1 b2 (t ,T ) 2 
 2 2 
 g(t ,rt )  rt  0  0 

0

as required.

The advantage of this model over Vasicek is that  (t ) can be chosen to reproduce (as
closely as possible) the exact yield curve, rather than the restricted forms of the Vasicek
model.

© IFE: 2019 Examinations The Actuarial Education Company

 om
i.c
CM2-18: The term structure of interest rates Page 27

ng
si
om
o m
4 Summary of short-rate modelling

as
.m
w
4.1 Summary of models

w
w
The properties of these models are summarised below:

Model Dynamics rt  0 for all t Distribution of rt

Vasicek drt   (  rt )dt   dW t No Normal

Non-central chi-
CIR drt   (  rt )dt   rt dW t Yes
squared

Hull-White –
drt   ( (t )  rt )dt   dW t No Normal
Vasicek

Non-central chi-
Hull-White – CIR drt   ( (t )  rt )dt   rt dW t Yes
squared

The last row of the table introduces an alternative form of the Hull-White model that extends the
CIR model.

There are analytic solutions for P (t ,T ) and option prices for each of these four models.

4.2 Limitations

Question

What is a one-factor term structure model?

Solution

A one-factor model is a model in which interest rates are assumed to be influenced by a single
source of randomness.

Bearing in mind that the purpose of interest rate models is to price interest rate derivatives,
there are some short-comings of short-rate models:

 Single factor short-rate models mean that all maturities behave in the same way -
there is no independence. This essentially means they are useless for pricing
swaptions (but OK for caps/floors).

 There is little consistency in valuation between the models.

 They are difficult to calibrate.

One-factor models, such as Vasicek and CIR, have certain limitations with which it is
important to be familiar.

The Actuarial Education Company © IFE: 2019 Examinations

 om
i.c
Page 28 CM2-18: The term structure of interest rates

ng
si
om
m
First, if we look at historical interest rate data we can see that changes in the prices of

o
as
bonds with different terms to maturity are not perfectly correlated as one would expect to

.m
see if a one-factor model was correct. Sometimes we even see, for example, that

w
short-dated bonds fall in price while long-dated bonds go up. Recent research has

w
w
suggested that around three factors, rather than one, are required to capture most of the
randomness in bonds of different durations.

Second, if we look at the long run of historical data we find that there have been sustained
periods of both high and low interest rates with periods of both high and low volatility.
Again, these are features which are difficult to capture without introducing more random
factors into a model.

This issue is especially important for two types of problem in insurance: the pricing and
hedging of long-dated insurance contracts with interest rate guarantees; and asset-liability
modelling and long-term risk management.

One-factor models do, nevertheless have their place as tools for the valuation of simple
liabilities with no option characteristics; or short-term, straightforward derivatives
contracts.

For other problems it is appropriate to make use of models which have more than one
source of randomness: so-called multi-factor models. Hull-White is not really a multi-factor
model, the  and  parameters are deterministic and aid fitting. A multi-factor version of
Vasicek would involve a multidimensional driving Wiener process and possibly
stochastic  .

Question

Summarise the characteristics of the Vasicek, Cox-Ingersoll-Ross and Hull-White models.

© IFE: 2019 Examinations The Actuarial Education Company

 om
i.c
CM2-18: The term structure of interest rates Page 29

ng
si
om
o m
Solution

as
.m
w
The following table summarises the characteristics of the Vasicek, Cox-Ingersoll-Ross (CIR) and

w
Hull-White models.

w
Vasicek CIR Hull-White

Arbitrage-free Yes Yes Yes

Positive interest rates No Yes No

Mean-reverting interest rates Yes Yes Yes

Easy to price bonds and derivatives Yes Yes(1) Yes

Realistic dynamics No(2) No(2) No(2)

Adequate fit to historical data No No Yes

Easy to calibrate to current data No No Yes(3)

Can price a wide range of derivatives No(4) No(4) No(4)

Notes:

(1) Although the CIR model is harder to use than the other two models, it is more tractable
than models with two or more factors.

(2) All three models produce perfectly correlated changes in bond prices, which is inconsistent
with the empirical evidence, and fail to model periods of high and low interest rates and
high and low volatility.

(3) Whilst one-factor models are generally difficult to calibrate, the Hull-White model is easier
than the other two because its time-varying mean-reversion function aids fitting.

(4) All three models can be used to price short-term, straightforward derivatives, but not
complex derivatives.

The Actuarial Education Company © IFE: 2019 Examinations

 om
i.c
Page 30 CM2-18: The term structure of interest rates

ng
si
om
o m
5 State-price deflator approach to pricing

as
.m
w
5.1 Short-rate models

w
w
Other modellers prefer to take a different approach to model-building using state price deflators.

With this approach we first specify a strictly positive diffusion process A(t) with SDE under P :

dA(t)  A(t)   A (t)dt   A (t)dW (t)

where  A (t) and  A (t) will be stochastic.

Note that, with this approach, the dynamics of the state price deflator A(t) are defined in terms
of the real-world probability measure.

We define zero-coupon bond prices according to the formula

E  A(T )  Ft 
P(tT )  P
A(t)

It can be shown, although it isn’t done here, that this method gives rise to an arbitrage-free model
for bond prices.

Question

Check that this formula works when t  T .

Solution

When t  T we have:

E  A(T )  FT 
P(T T )  P
A(T )

The expectation here is conditional on the history of the process A(t) up to time T .
Consequently, there is no additional randomness in A(T ) , which can be considered to be a known
value. So we have:

A(T )
P(T T )  1
A(T )

This is exactly what we would expect, since the bond matures at time T and has a value of 1 at
that time.

© IFE: 2019 Examinations The Actuarial Education Company

 om
i.c
CM2-18: The term structure of interest rates Page 31

ng
si
om
o m
Question

as
.m
Suppose that A(t) is a deterministic process with  A (t)   and  A (t)  0 , where  is a

w
w
w
constant. Show that A(t)  ke t and evaluate P(t ,T ) .

Solution

Here the SDE for A(t) becomes dA(t)   A(t) dt , which is an ordinary differential equation. This
can be written in the form:

1 dA(t) d
  or log A(t)  
A(t) dt dt

Integrating gives:

log A(t)   t  constant

So: A(t)  ke t , where k is a constant.

EP  A(T )  Ft 
P(t ,T ) can now be calculated from the formula P(tT )  , ie:
A(t)

EP ke  T  Ft 
P(tT )    t 
ke

Since A(t) is not random here, we can simplify this to get:

ke  T
P(tT )   t
 e  (T t )
ke

This question illustrates that the process A(t) acts like a discount factor that applies to payments
at time t . It is a generalisation of the familiar vt factor.

The formula for P(t ,T ) is a very simple looking formula, but any potential difficulty comes in
working out EP  A(T )  Ft  as we need ‘interesting’ models for A(t) in order to get interesting and
useful models for r (t) and P(t ,T ) .

Question

Suppose that  A (t)   and  A (t)   , where  and  are constants.

Show that A(t)  A(0)e( ½ )t Wt and P(tT )  e  (T t ) .

2

The Actuarial Education Company © IFE: 2019 Examinations

 om
i.c
Page 32 CM2-18: The term structure of interest rates

ng
si
om
o m
Solution

as
.m
w
The SDE for A(t) now becomes:

w
w
dA(t)  A(t)  dt   dWt 

This has the same form as the SDE for the lognormal model for share prices. So its solution is:

A(t)  A(0)e( ½ )t Wt

2

EP  A(T )  Ft 
P(t ,T ) can now be calculated from the formula P(tT )  :
A(t)

EP  A(0)e( ½ )T WT  Ft 

2

 
P(tT )  
A(0)e( ½ )t Wt
2

e( ½ )(T t)

2

 EP e WT  Ft 
eWt  

 e( ½ )(T t)EP e (WT Wt )  Ft 

2

 

This expectation is conditional on the history up to time t . But WT  Wt relates to the ‘future’
time interval (t ,T ) , and so is statistically independent. So we can write:

P(tT )  e( ½ )(T t)EP e (WT Wt ) 

2

 

Since WT  Wt  N(0,T  t) , the expectation is now just the MGF of a normal distribution, so that:

P(tT )  e( ½ )(T t)e½ (T t )

2 2

 e  (T t )

Using the state price deflator approach the risk-free rate of interest can be shown to be simply:

r(t)    A (t)

Again, this result is not derived here.

It follows that our model will have positive interest rates if and only if  A (t) is negative for all t
with probability 1. This means that A(t) must be a strictly positive supermartingale (that is,
EP  A(T )  Ft   A(t) for all t , T ).

Recall that martingales have the property that E  Xt |Fs   X s whenever t  s , ie the process has
zero drift. A supermartingale has the property that E  Xt |Fs   X s , ie the drift is negative or zero.

© IFE: 2019 Examinations The Actuarial Education Company

 om
i.c
CM2-18: The term structure of interest rates Page 33

ng
si
om
o m
Question

as
.m
w
How do we know that A(t) is a supermartingale?

w
w
Solution

A(t) takes positive values and  A (t) is negative. Therefore the drift coefficient A(t) A (t) in the
SDE for A(t) is negative. So the expected value of dA(t) is negative, ie the process has negative
drift, which means that it is a supermartingale.

The state price deflator approach looks like it is quite different from the risk-neutral approach to
pricing. However, the two approaches are, in fact, exactly equivalent: any model developed in
one framework has an equivalent form under the other framework.

5.2 Market models

Today, ‘market models’ have superseded short-rate models for situations where the
correlation between different maturity rates is critical, such as the pricing of swaptions.
These treat each maturity instrument (such as forward rate) as a distinct object, correlated
to other similar assets using a multi-dimensional Wiener process in a no-arbitrage set-up
based on the ‘state price deflator’ approach.

Define a traded asset, based on the (traded) discount bonds P (s, t ) and P (s,T ) :

1
X (s )   P (s, t )  P (s,T ) 


The price of a traded asset divided by another traded asset must be a martingale under the
measure associated with the numeraire.

A numeraire is a quantity whose values are used as the units for expressing the price of a security.
Commonly used numeraires are:
 the accumulated value of a risk-free cash account
 the price of a zero-coupon bond
 the price of a foreign currency.

So, labelling QT as the measure associated with using P (s,T ) as the numeraire, we have:

X (s )  X (t ) 
F (s, t ,T ) 
P (s,T )
 EQT  P (t ,T )   EQT  F (t , t ,T )
 

Under QT , the forward rates, F (s, t ,T ) are martingales, and:

T
dF (s, t ,T )   (s, t )F (s , t ,T )dWsQ

Note that we use t as the parameter for the volatility, rather than T, because the forward
matures at t and does not exist in (t ,T ] .

The Actuarial Education Company © IFE: 2019 Examinations

 om
i.c
Page 34 CM2-18: The term structure of interest rates

ng
si
om
o m
as
.m
w
w
w
This page has been left blank so that you can keep the chapter
summaries together for revision purposes.

© IFE: 2019 Examinations The Actuarial Education Company

 om
i.c
CM2-18: The term structure of interest rates Page 35

ng
si
om
o m
as
Chapter 18 Summary

.m
w
w
Notation

w
 t current time
 P(t ,T ) zero-coupon bond price

 r (t) instantaneous risk-free rate

 f (t ,T , S) continuously compounded forward rate for the period (T , S)

 f (t ,T ) instantaneous forward rate

 r (t ,T ) continuously compounded spot rate

Relationships
1 1 T
T  t t
 r (tT )   log P(tT )  f (t u)du
T t

 P(t ,T )  exp[r (t ,T )(T  t)]

1 P(tT )
 f (tT  S)  log
S T P(t S)


 f (t ,T )  lim F (t ,T , S)   log P(t ,T )
T S T

P(t ,T )  exp    f (t ,u)du 

T

 t 

Desirable characteristics of a term structure model

 The model should be arbitrage-free.
 Interest rates should ideally be positive.
 Interest rates should be mean-reverting.
 Bonds and derivative contracts should be easy to price.
 It should produce realistic interest rate dynamics.
 It should fit historical interest rate data adequately.
 It should be easy to calibrate to current market data.
 It should be flexible enough to cope with a range of derivatives.

The Actuarial Education Company © IFE: 2019 Examinations

 om
i.c
Page 36 CM2-18: The term structure of interest rates

ng
si
om
o m
as
Examples of one-factor models

.m
w
Vasicek model

w
w
dr(t)      r(t)dt   dW (t) under Q

Cox-Ingersoll-Ross model

dr (t)      r (t)dt   r (t)dW (t) under Q

Hull-White model

dr (t)    (t)  r(t)dt   dW (t) under Q

where  (t) is a deterministic function.

© IFE: 2019 Examinations The Actuarial Education Company