Interest Rate Derivatives: Lecture notes

Dr. Graeme West

Financial Modelling Agency
graeme@finmod.co.za
www.finmod.co.za

December 26, 2010

Contents

1 Introduction to Interest Rates 4

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Day count conventions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Coupon Bonds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Yield-to-Maturity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.5 Term Structure of Default-Free Interest Rates . . . . . . . . . . . . . . . . . . . . . . 8
1.6 The par bond curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.7 Reminder: Forward Rate Agreements . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.8 The continuous forward curve . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
1.9 The raw interpolation method for yield curve construction . . . . . . . . . . . . . . . 12
1.10 Traditional Measures of Interest Rate Risk . . . . . . . . . . . . . . . . . . . . . . . . 13
1.10.1 (Macauley) Duration and Modified Duration . . . . . . . . . . . . . . . . . . 13
1.10.2 Convexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.10.3 Problems with these measures . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.10.4 pv01 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.11 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2 Discrete-time Interest Rate Models 17

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.2 The basic lattice construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.3 Normal Distribution (Ho-Lee model) . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.4 Formalising the Lattice Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.5 Lognormal Distribution (Black-Derman-Toy model) . . . . . . . . . . . . . . . . . . 22
2.6 Options on Zero-Coupon Bonds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.7 Forwards on Zero-Coupon Bonds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.8 Hedging Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
2.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26

1
3 Black’s Model 28
3.1 European Bond Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.1.1 Different volatility measures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2 Caplets and Floorlets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.3 Caps and Floors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
3.3.1 Valuation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3.2 A call/put on rates is a put/call on a bond . . . . . . . . . . . . . . . . . . . 31
3.3.3 Greeks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
3.4 Stripping Black caps into caplets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
3.5 Swaptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
3.5.1 Valuation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.5.2 Greeks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
3.5.3 Why Black is useless for exotics . . . . . . . . . . . . . . . . . . . . . . . . . . 39
3.6 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

4 One and two factor continuous-time interest rate models 42

4.1 Derivatives Modelled on a Single Stochastic Variable . . . . . . . . . . . . . . . . . . 42
4.2 What is the market price of risk? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.3 Exogenous (equilibrium) short rate models . . . . . . . . . . . . . . . . . . . . . . . . 44
4.3.1 GBM model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.4 A particular class of models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.4.1 Constant parameter model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.4.2 Vasicek model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.4.3 Cox-Ingersoll-Ross model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
4.5 Two factor equilibrium models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.6 Equilibrium models of the logarithm of the short rate . . . . . . . . . . . . . . . . . 52
4.6.1 The Continuous time version of the Black-Derman-Toy model . . . . . . . . . 52
4.6.2 The Black-Karasinski Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.7 No-arbitrage models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.7.1 The Continuous time version of the Ho-Lee model . . . . . . . . . . . . . . . 54
4.7.2 The Extensions of Hull & White . . . . . . . . . . . . . . . . . . . . . . . . . 55
4.7.3 The need for the convexity adjustment . . . . . . . . . . . . . . . . . . . . . . 57
4.7.4 Derivation of the convexity adjustment . . . . . . . . . . . . . . . . . . . . . . 58
4.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60

5 The LIBOR market model 62

5.1 The model for a single forward rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.2 The pricing of caplets and caps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
5.3 A common measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64
5.4 Pricing exotic instruments under LMM . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.4.1 A simple example . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65

2
5.4.2 Calibration of the parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

3
Chapter 1

Introduction to Interest Rates

1.1 Introduction
In equity option pricing we make the assumption throughout that interest rates were previsible.
This greatly simplifies discussions of hedging and replication, and allows the use of the Black-
Scholes analysis and framework. The major implication of this assumption concerns the numeraire
in the martingale pricing formula. Recall that the chosen numeraire is the money-market account
(continuous investment/borrowing at deterministic interest rates). On application of the
[ assump-
]
𝑓 (0) ℚ 𝑓 (𝑇 ) 1
tion, we could bring the numeraire 𝐵 (𝑇 ) in the martingale pricing formula: 𝐵(0) = 𝔼0 𝐵(𝑇 ) out
𝐵(0) ℚ
of the expectation to evaluate the price of the derivative as: 𝑓 (0) = 𝐵(𝑇 ) 𝔼0 [𝑓 (𝑇 )].
This equation forms the basis of the Black-Scholes solution. The measure 𝑄 is the so-called Risk
Neutral Measure, is unique, and implies that the stochastic processes for the underlying (default-
free) securities and their derivatives have an expected drift rate of 𝑟 (the prevailing risk-free rate),
under 𝑄. (In fact, under 𝑄, ALL traded securities have an expected return of 𝑟!)
Despite our reservations, the model is extensively used, as seen for example in the extensive use of
the Black-Scholes formula for short-dated options. This is a consequence of the underlying stochastic
process (for the stock) being fairly remote from the stochastic process for the interest rate.
1 Some notation: 𝐵(𝑇 ) is the amount of money that 1 at time 0 has grown to by time 𝑇 , by continual reinvestment

at the short dated rate: 𝐵(⋅) is called the money market account, or simply the bank account. Usually this short
dated rate is taken to be the overnight rate, so money is placed in the overnight account, and left there, compounding
each day. Of course, by this, we mean the usual international meaning of the overnight account, in South Africa, a
true overnight rate does not exist, as discussed in West [2009].
𝑍(𝑡, 𝑇 ) is the discount factor for time 𝑇 as observed at time 𝑡. It is known when we reach 𝑡, from the bootstrap of
the yield curve at that time. Often, 𝑡 = 0, in which case we might abbreviate 𝑍(0, 𝑇 ) to 𝑍(𝑇 ).
𝑍(0,𝑇 )
𝑍(0; 𝑡, 𝑇 ) is the forward discount factor from time 𝑡 to time 𝑇 . As proved in West [2009], it is equal to 𝑍(0,𝑡) .
𝐶(𝑡, 𝑇 ) is the capitalisation factor for time 𝑇 as observed at time 𝑡. The same comments apply. 𝐶(0; 𝑡, 𝑇 ) is the
𝐶(0,𝑇 )
forward capitalisation factor from time 𝑡 to time 𝑇 . As proved in West [2009], it is equal to 𝐶(0,𝑡) . Note that 𝑍(⋅)
and 𝐶(⋅) are inverse.
We do not know the value of 𝐵(𝑇 ) at time 0 because interest rates are stochastic. When we price equity derivatives,
we assume that this stochasticity is removed, so 𝐵(⋅) = 𝐶(⋅).

4
However, this cannot be the case when options and derivatives are written either on interest rates,
or on securities whose values are dependent on interest rates (e.g. bond options, swaps, caps, floors
etc.) In these cases, it is exactly the fluctuation of interest rates that the option buyer seeks to
hedge; so an assumption of constancy of interest rates makes little sense. In particular, the bond
option model we saw in West [2009] is very problematic.
As we saw in West [2009] vanilla type interest rate derivatives such as deposits (JIBAR deposits),
FRAs and swaps do not require interest rate modelling, as they are priced using pure no arbitrage
considerations. They can be replicated using instruments that pay certain cash flows in the future
and consequently require no statistical modelling at all. They do require an accurate yield curve,
though. Producing this is not always as simple a task as it would appear. We saw a (naively) simple
method in West [2009]. For more information on the subtle difficulties that can arise, see Hagan
and West [2006], Hagan and West [2008].
It will be assumed that you are perfectly familiar with all of the interest rate material dealt with in
West [2009].
In order to price more complicated interest rate products that include optionality, we need to arrive
at a statistical model of the evolution of the yield curve. The models we will consider in this course
model all interest rates as dependent on a single rate, often termed the short rate. The evolution
of the short rate then governs the evolution of all rates along the entire curve. However, this is
quite a task, as a change in the yield curve is a complicated phenomenon, since it may undergo
combinations of parallel shifts, slope changes and curvature changes.
These models are roughly divided into two categories: equilibrium models and no arbitrage models.
In the equilibrium approach, if the model can be trusted to give a fundamentally correct, albeit
necessarily simplified, description of economic reality then there will be discrepancies between model
and market values. According to the model these discrepancies will reflect trading opportunities!
Equilibrium Models attempt to describe the economy of interest rates as a whole. Clearly this
approach is quiet abstract, and is not much used - although there are some equilibrium models that
can be reformulated as no arbitrage models.
The pure no arbitrage approach seeks to represent the value of a complicated interest rate derivative
in terms of vanilla instruments or cash instruments. The prices of these vanilla instruments will be
taken as given and any model must recover their actual traded prices. No Arbitrage Models take the
market prices of vanilla products as basic building blocks, and infer from them more complicated
derivative prices. However, different models could give different prices of the more complicated
instruments, even though they recover the same prices for the vanilla instruments!
Another complicating feature is that we cannot buy and sell interest rates. THE UNDERLYING
OF THE MODEL IS NOT A TRADEABLE INSTRUMENT. The construction of a “replicating”
portfolio requires more thought: it isn’t just the delta and cash, as it is for equity: both of these
factors don’t perform as we would naively like. What we can buy and sell is bonds, whose prices
are themselves derivative of interest rates.
Initially, we examine default-free securities and the term structure of interest rates. Once we have
done this, we are in a position to model the movements of the yield curve.

5
1.2 Day count conventions
Denote the generic period between two payment dates as 𝛼 parts of a year. The rules could be
different for bonds and swaps, and even for the floating and fixed legs of the swaps. Moreover, the
rules differ by jurisdiction. The relevant markets are

• The bond market: the market for the issuing of treasury bonds. Day count conventions are
relevant for the accrual of interest and hence the conversions between clean and dirty price.2

• The money market: the market for FRAs and hence the market for the floating leg of swaps.

• The swap market: the market for the fixed leg of swaps.

The day count conventions in the various markets are as follows

spot/value date bond money fixed swap

RSA 𝑡, 𝑡 + 3 Actual/365 Actual/365 Actual/365
USA 𝑡+2 Actual/Actual Actual/360 30/360 3
UK 𝑡 Actual/Actual Actual/365 Actual/365
Euro 𝑡+2 Actual/Actual Actual/360 30/360
Japan 𝑡+2 Actual/365 Actual/360 Actual/365
Canada 𝑡 Actual/365 Actual/365 Actual/365
Australia 𝑡 Actual/Actual Actual/365 Actual/365

where Actual/Actual appears in International Swaps and Derivatives Association [25 November
1998]. This is rather tricky, they point out that at the time there were three conflicting interpre-
tations of this rule. The approach taken thus needs to be recorded when a deal is made. The
approaches are a follows:

(i) ISDA Actual/actual (historical). Split the period of interest into the years in which it occurs.
For each year, divide the number of actual days in the period by the number of days in that
year. Day count is equal to the sum of these fractions.
For example, if the period is from 20 September 2003 to 20 March 2004, and we are now at
13 January 2004, then the period day count is 102 80
365 + 366 and the time elapsed day count is
102 13
365 + 366 .

(ii) AFB Actual/Actual Euro. The numerator is the actual number of days, the denominator is
either 365 or 366 depending on whether or not the period includes a 29 February.
182 115
In the above example, the period day count is 366 and the elapsed day count is 366 .

(iii) ISMA Actual/Actual Bond. This is the actual number of days, divided by the product of the
number of days in the period and the number of periods in the year.
2 In fact it is unlikely that there will ever be a need for modelling the US Treasury curve. This is because then the

so-called TED spread needs to be measured i.e. the spread between the treasury and AAA curve, which is determined
by the swaps. Secondly, the US Treasury is retiring a lot of its debt, so the treasury curve is very illiquid.
3 Actual/Actual and Actual/360 also occur.

6
 1
In the above example, the period day count is 2 (as there are two periods in a year) and the
115
elapsed day count is 182⋅2 .
This may be the most common convention, and is also known as ISMA Rule 251. Thus, for
this calculation let the start date of the period of interest be 𝑡1 , the date of interest 𝑡, and the
end date of the period of interest be 𝑡2 . Then on an Actual/Actual basis
𝑡 − 𝑡1
Accrued period = (1.1)
(𝑡2 − 𝑡1 ) ⋅ round( 𝑡2365
−𝑡1 , 0)
𝑡2 − 𝑡
Period remaining = (1.2)
(𝑡2 − 𝑡1 ) ⋅ round( 𝑡2365
−𝑡1 , 0)

The EURIBOR quotation will be a representative rate for euro deposits based on quotations from
a pan-European panel of banks. The euro-LIBOR quotation will be a representative rate for euro
deposits based on quotations from a panel of 16 banks in the London market. The conventions for
both of these are the ‘Euro’ conventions above.

1.3 Coupon Bonds

Coupon-bearing bond issuers pay regular fixed interest payments to the holder of the bond on
specific dates (these are the coupons) as well as the par or face value at maturity. The bond value
at any time 𝑡 must be the present value of both its face value and its coupons (the coupon rate is
pre-fixed e.g. the r153 has a 13% annual coupon, i.e. 6.5% of par value is paid out to the holder
every 6 months).
Suppose a bond pays amounts 𝑝1 , 𝑝2 , . . . , 𝑝𝑛 at times 𝑡1 , 𝑡2 , . . . , 𝑡𝑛 . (𝑝𝑖 does not necessarily
mean coupon here, for example, 𝑐𝑛 could be the coupon and bullet.) The current bond price 𝑉 is
the sum of the present values of all payments i.e.
𝑛
∑
𝑉 = 𝑝𝑖 𝑍(0, 𝑡𝑖 ) (1.3)
𝑖=1

The act of regarding each coupon as a separate zero ensures arbitrage-free pricing. If this were not
true, the coupon bond could be synthetically replicated using zero-coupon bonds. This replication
is obviously not practically as straightforward as is made out here.

1.4 Yield-to-Maturity
Coupon-bearing bond prices are often quoted in terms of their yield-to-maturity. This is defined as
an interest rate per annum that equates the present value of the bond’s associated cash flows to the
current market price. There exists a terminology confusion here because the yield-to-maturity of a
coupon bond does not correspond directly to any value on the zero-coupon yield curve, in particular,
it is not the value at the maturity (or duration, to be defined later) on the yield curve, even taking
into account possible conversion between different NAC* quoting methods.
The yield-to-maturity is merely a convenient way of expressing the price of a coupon-bearing bond
in terms of a single interest rate. In some rough sense, the YTM represents a weighted average of
the interest rates along the current zero-coupon yield curve.

7
If payments are made semi-annually (the most common) then 𝑦 - the semi-annually compounded
yield-to-maturity - is implicitly defined by:
𝑛
∑ 𝑝𝑖
𝑉 = (1.4)
𝑖=1
(1 + 𝑦2 )2𝑡𝑖

where there are payments 𝑝1 , 𝑝2 , . . . , 𝑝𝑛 ; the payment 𝑝𝑖 occurs at time 𝑡𝑖 , and time is measured
in years.
The market observables are the bond prices 𝑉 , so the value for 𝑦 must be calculated numerically.
(1.4) cannot be inverted, but Newton’s method comes to the rescue. The yield-to-maturity is the
holding period return per annum on the coupon bond.
Using the effective yield-to-maturity as the interest rate offered by the bond, implicitly assumes
that coupon reinvestment takes place at the semi-annual yield-to-maturity over the bond life. In a
world of changing interest rates it is unlikely that this will happen: there is no guarantee that the
ytm was an actual market rate at any reinvestment date, even if the yield curve does not change.
Nevertheless, the yield to maturity is a very useful concept: it enables a first order fair comparison
of different bonds trading simultaneously. So,

• Typically, the market will trade on price, but market participants will want to know the yield
to maturity of the instruments they are trading - they require a yield-given-price algorithm.
Here they use Newton’s method.

• Less typically (such as in South Africa) the market will trade on yield. We then require a very
precise formula for converting from yield to price, in order to determine the cash flows required
at the bond exchange. For repos/carrys, the yield-given-price algorithm will be needed anyway.

1.5 Term Structure of Default-Free Interest Rates

The term structure of interest rates is defined as the relationship between the yield-to-maturity on
a zero coupon bond and the bond’s maturity. If we are going to price derivatives which have been
modelled in continuous-time off of the curve, it makes sense to use continuously-compounded rates
from the outset.
Building a yield curve from existing data is a difficult task. The liquidity of the market plays a
large part in determining whether the exercise can be done. For US Bond market it is only on-
the-run bonds (last auctioned) which will give the most accurate indication of where yields are. In
South Africa this task is doubly difficult and requires some econometric artistry. South Africa has
a sophisticated and liquid swap market. This makes swap rates a better starting point for a yield
curve model.
The results of bootstrapping will be near unique in liquid markets, but there may be significant
variation in less liquid markets or markets with fewer inputs.
In so-called normal markets, yield curves are upwardly sloping, with longer term interest rates being
higher than short term. A yield curve which is downward sloping is called inverted. A yield curve
with one or more turning points is called mixed. Constructing a yield curve consists of solving (1.3)
for the discount factors 𝑍(0, 𝑡) in a piecewise fashion starting with the shortest maturity instruments

8
and progressing to the longer-dated coupon-bearing instruments. A mixed yield curve is shown in
Figure 1.1. The discount factors are also shown.
The South African yield curve typically has two to four turning points. It is often stated that such
mixed yield curves are signs of market illiquidity or instability. This is not the case. Supply and
demand for the instruments that are used to bootstrap the curve may simply imply such shapes.
However, many of the models that we see later in this course are driven by one factor - and intuitively
this is clearly best suited to a normal or inverted curve (because, essentially, the model dictates that
when the curve moves, it more or less moves in parallel). Thus, these models need to be analysed
carefully for their suitability in the South African market: see Svoboda [2002].

Figure 1.1: Some yield curves and their discount functions

The shape of the graph for 𝑍(0, 𝑡) does not reflect the shape of the yield curve in any obvious way.
The discount factor curve must - by no arbitrage - be monotonically decreasing whether the yield
curve is normal, mixed or inverted. Nevertheless, many bootstrapping and interpolation algorithms
for constructing yield curves miss this absolutely fundamental point. See Hagan and West [2006],
Hagan and West [2008].

1.6 The par bond curve

Suppose the bond pays ANNUAL coupons of 𝑅𝑛 at times 𝑡1 , 𝑡2 , . . . , 𝑡𝑛 , with the bullet being at
time 𝑡𝑛 . The inter-coupon times are 𝛼𝑖 , so the 𝑖𝑡ℎ payment is in fact 𝑅𝑛 𝛼𝑖 . For it to be a par bond,
we must have
∑𝑛
1 = 𝑍(0, 𝑡𝑛 ) + 𝑅𝑛 𝛼𝑖 𝑍(0, 𝑡𝑖 ) (1.5)
𝑖=1

We now have a function 𝑡𝑛 → 𝑅𝑛 which maps maturity dates to the requisite coupon size. This
function is called the par bond curve.
But this should look familiar, and we didn’t choose 𝑅𝑛 as the notation for the annual coupon size
by chance. You see that this is EXACTLY the curve of fair swap rates trading in the market.

9
Remember that the value of the 𝑖𝑡ℎ floating payment is

𝑖
𝑉float = 𝑍(𝑡, 𝑡𝑖−1 ) − 𝑍(𝑡, 𝑡𝑖 )
[ ]
𝐶(0, 𝑡𝑖 )
= 𝑍(0, 𝑡𝑖 ) −1
𝐶(0, 𝑡𝑖−1 )
= 𝑍(0, 𝑡𝑖 )𝛼𝑖 𝑓𝑖𝑠

where 𝑓𝑖𝑠 is the simple forward rate for the period [𝑡𝑖−1 , 𝑡𝑖 ]. Thus
𝑛
∑
𝑉fix = 𝑅𝑛 𝑍(𝑡, 𝑡𝑖 )𝛼𝑖
𝑖=1
𝑛
∑
𝑉float = 𝑍(0, 𝑡𝑖 )𝛼𝑖 𝑓𝑖𝑠
𝑖=1

as as these are equal, we have

𝑛
∑
𝑅𝑛 = 𝑤𝑖 𝑓𝑖𝑠 (1.6)
𝑖=1
𝑍(𝑡, 𝑡𝑖 )𝛼𝑖
𝑤𝑖 = ∑𝑛 (1.7)
𝑗=1 𝑍(𝑡, 𝑡𝑗 )𝛼𝑗

Thus the par coupon rates are a weighted average of the simple forward rates along the life of the
bond, where the weights are decreasing. On the other hand, we have
𝑛
∏ 𝑐
𝑒𝑟𝜏 = 𝑒𝑓𝑖 𝛼𝑖
𝑖=1

and so
𝑛
∑ 𝛼𝑖
𝑟= 𝑓𝑖𝑐 (1.8)
𝑖=1
𝜏

where now 𝑓𝑖𝑐 is the continuous forward rate for the period [𝑡𝑖−1 , 𝑡𝑖 ], and so the zero rates are an
almost equally weighted average of the continuous forward rates along the life of the bond.
Allowing for the difference in compounding conventions, we see in general that if the yield curve is
normal then the par curve is below the zero curve which is below the forward curve, while if the
yield curve is inverted, then the par curve is above the zero curve which is above the forward curve.

1.7 Reminder: Forward Rate Agreements

These are the simplest derivatives: a FRA is an OTC contract to fix the yield interest rate for some
period starting in the future. If we can borrow at a known rate at time 𝑡 to date 𝑡1 , and we can
borrow from 𝑡1 to 𝑡2 at a rate known and fixed at 𝑡, then effectively we can borrow at a known rate
at 𝑡 until 𝑡2 . Clearly

𝐶(𝑡, 𝑡1 )𝐶(𝑡; 𝑡1 , 𝑡2 ) = 𝐶(𝑡, 𝑡2 ) (1.9)

10
is the no arbitrage equation: 𝐶(𝑡; 𝑡1 , 𝑡2 ) is the forward capitalisation factor for the period from 𝑡1
to 𝑡2 - it has to be this value at time 𝑡 with the information available at that time, to ensure no
arbitrage.
In a FRA the buyer or borrower (the long party) agrees to pay a fixed yield rate over the forward
period and to receive a floating yield rate, namely the 3 month JIBAR rate. At the beginning of
the forward period, the product is net settled by discounting the cash flow that should occur at the
end of the forward period to the beginning of the forward period at the (then current) JIBAR rate.
This feature - which is typical internationally - does not have any effect on the pricing.

𝑡 𝑡1 𝑡2
?

Figure 1.2: A long position in a FRA.

In South Africa FRA rates are always quoted for 3 month forward periods eg 3v6, 6v9, . . . , 21v24
or even further. The quoted rates are simple rates, so if 𝑓 is the rate, then
𝐶(𝑡, 𝑡2 )
1 + 𝑓 (𝑡2 − 𝑡1 ) = 𝐶(𝑡; 𝑡1 , 𝑡2 ) = (1.10)
𝐶(𝑡, 𝑡1 )
Thus
( )
1 𝐶(𝑡, 𝑡2 )
𝑓= −1 (1.11)
𝑡2 − 𝑡1 𝐶(𝑡, 𝑡1 )

1.8 The continuous forward curve

Using continuous rates, the forward rate governing the period from 𝑡1 to 𝑡2 , denoted 𝑓 (0; 𝑡1 , 𝑡2 )
satisfies
𝑍(0, 𝑡2 )
exp(−𝑓 (0; 𝑡1 , 𝑡2 )(𝑡2 − 𝑡1 )) = 𝑍(0; 𝑡1 , 𝑡2 ) :=
𝑍(0, 𝑡1 )
Immediately, we see that forward rates are positive (this is equivalent to the discount function
decreasing). We have either of
ln(𝑍(0, 𝑡2 )) − ln(𝑍(0, 𝑡1 ))
𝑓 (0; 𝑡1 , 𝑡2 ) = − (1.12)
𝑡2 − 𝑡1
𝑟2 𝑡2 − 𝑟1 𝑡1
= (1.13)
𝑡2 − 𝑡1
Let the instantaneous forward rate for a tenor of 𝑡 be denoted 𝑓 (𝑡), that is, 𝑓 (𝑡) = lim𝜖↓0 𝑓 (0; 𝑡, 𝑡+𝜖),
for whichever 𝑡 this limit exists. Clearly then
𝑑
𝑓 (𝑡) = − ln(𝑍(𝑡)) (1.14)
𝑑𝑡
𝑑
= 𝑟(𝑡)𝑡 (1.15)
𝑑𝑡

11
So 𝑓 (𝑡) = 𝑟(𝑡) + 𝑟′ (𝑡)𝑡, so the forward rates will lie above the yield curve when the yield curve is
normal, and below the yield curve when it is inverted. By integrating,4
∫ 𝑡
𝑟(𝑡)𝑡 = 𝑓 (𝑠) 𝑑𝑠 (1.16)
0
∫ 𝑡
= 𝑟(𝑡𝑖−1 )𝑡𝑖−1 + 𝑓 (𝑠) 𝑑𝑠 (1.17)
𝑡𝑖−1
( ∫ 𝑡 )
𝑍(𝑡) = exp − 𝑓 (𝑠) 𝑑𝑠 (1.18)
0

Also
𝑡𝑖
𝑟𝑖 𝑡𝑖 − 𝑟𝑖−1 𝑡𝑖−1
∫
1
𝑓𝑖𝑑 := = 𝑓 (𝑠) 𝑑𝑠 (1.19)
𝑡𝑖 − 𝑡𝑖−1 𝑡𝑖 − 𝑡𝑖−1 𝑡𝑖−1

which shows that the average of the instantaneous forward rate over any interval [𝑡𝑖−1 , 𝑡𝑖 ] is equal
to the discrete forward rate for that interval.

1.9 The raw interpolation method for yield curve construc-

tion
This method corresponds to piecewise constant forward curves. This method is very stable, is trivial
to implement, and is usually the starting point for developing models of the yield curve. One can
often find mistakes in fancier methods by comparing the raw method with the more sophisticated
method.
By definition, raw interpolation is the method which has constant instantaneous forward rates on
every interval 𝑡𝑖−1 < 𝑡 < 𝑡𝑖 . From (1.19) we see that that constant must be the discrete forward
rate for the interval, so 𝑓 (𝑡) = 𝑟𝑖 𝑡𝑖𝑡−𝑟 𝑖−1 𝑡𝑖−1
𝑖 −𝑡𝑖−1
for 𝑡𝑖−1 < 𝑡 < 𝑡𝑖 . Then from (1.17) we have that

𝑟𝑖 𝑡𝑖 − 𝑟𝑖−1 𝑡𝑖−1
𝑟(𝑡)𝑡 = 𝑟𝑖−1 𝑡𝑖−1 + (𝑡 − 𝑡𝑖−1 )
𝑡𝑖 − 𝑡𝑖−1
By writing the above expression with a common denominator of 𝑡𝑖 − 𝑡𝑖−1 , and simplifying, we get
that the interpolation formula on that interval is
𝑡 − 𝑡𝑖−1 𝑡𝑖 − 𝑡
𝑟(𝑡)𝑡 = 𝑟𝑖 𝑡𝑖 + 𝑟𝑖−1 𝑡𝑖−1 (1.20)
𝑡𝑖 − 𝑡𝑖−1 𝑡𝑖 − 𝑡𝑖−1
which explains yet another choice of name for this method: ‘linear 𝑟𝑡’; the method is linear interpo-
lation on the points 𝑟𝑖 𝑡𝑖 . Since ±𝑟𝑖 𝑡𝑖 is the logarithm of the capitalisation/discount factors, we see
that calling this method ‘linear on the log of capitalisation factors’ or ‘linear on the log of discount
factors’ is also merited.
This raw method is very attractive because with no effort whatsoever we have guaranteed that all
instantaneous forwards are positive, because every instantaneous forward is equal to the discrete
forward for the ‘parent’ interval. This is an achievement not to be sneezed at. It is only at the
points 𝑡1 , 𝑡2 , . . . , 𝑡𝑛 that the instantaneous forward is undefined, moreover, the function jumps at
that point.
4 We
∫𝑡
𝑓 (𝑠) 𝑑𝑠, so 𝑟(𝑡)𝑡 = [𝑟(𝑠)𝑠]𝑡0 =
∫
have 𝑟(𝑠)𝑠 + 𝐶 = 0 𝑓 (𝑠) 𝑑𝑠.

12
1.10 Traditional Measures of Interest Rate Risk
Traditionally interest rate (bond) traders used two measures of interest rate risk.

1.10.1 (Macauley) Duration and Modified Duration

Bonds might alternatively be priced not off of the yield curve but using a NACS yield to maturity.
What happens then? Suppose our bond is priced NACn, then
𝑛
∑ 𝑝𝑖
𝑉 (𝑦) =
𝑖=1
(1 + 𝑛𝑦 )𝑛𝑡𝑖

This time
𝑝𝑖
( 𝑦 )𝑛𝑡𝑖
1+ 𝑛
𝑤𝑡 =
𝑉
∑𝑛
and 𝐷 = 𝑖=1 𝑡𝑖 𝑤𝑖 is as before. But now something changes. This time
𝑛
𝑑𝑉 1 ∑ 𝑡𝑖 𝑝𝑖
= − 𝑦
𝑑𝑦 1+ 𝑛 𝑖=1 (1 + 𝑛𝑦 )𝑛𝑡𝑖
𝑑𝑉
so it is not true that 𝑑𝑦 = −𝐷𝑉 . So what we do is define a new quantity, called modified duration
𝐷𝑚 , by:
𝑛
1 ∑ 𝑡𝑖 𝑝𝑖
𝐷𝑚 𝑉 = 𝑦 (1.21)
1+ 𝑛 𝑖=1 (1 + 𝑛𝑦 )𝑛𝑡𝑖
so
𝑑𝑉
= −𝐷𝑚 𝑉 (1.22)
𝑑𝑦
We also have the definition of the dollar duration

𝐷$ = 𝐷𝑚 𝑉 (1.23)

Also, note that

𝐷
𝐷𝑚 = (1.24)
1 + 𝑛𝑦
As 𝑛 → ∞, the correction between duration and modified duration disappears, and they become
the same thing for NACC rates.

1.10.2 Convexity
Of course there is a non-linear relationship between the bond price and the yield-to-maturity. For
small shifts in yield, the first order duration calculation is a good measure of the sensitivity of the
bond price. If Δ𝑦 is large, the approximation is no longer accurate.
Using a Taylor expansion to second-order we can define convexity.

Δ𝑉 = 𝑉 (𝑦 + Δ𝑦) − 𝑉 (𝑦)
𝑑𝑉 𝑑2 𝑉 2
Δ𝑦 + 21 2 (Δ𝑦) + 𝑂 Δ𝑦 3
( )
=
𝑑𝑦 𝑑𝑦

13
and now convexity 𝐶 is defined as

𝑑2 𝑉
𝐶𝑉 = (1.25)
𝑑𝑦 2
so we have
Δ𝑉
= −𝐷𝑚 Δ𝑦 + 21 𝐶Δ𝑦 2 + 𝑂 Δ𝑦 3
( )
(1.26)
𝑉

1.10.3 Problems with these measures

Because the yield-to-maturity is, in some sense, a measure of the average rate of interest represented
by the yield curve with respect to the bond in question, information about more subtle yield curve
shifts is lost. It is possible for the change in the yield-to-maturity to mask the underlying zero
coupon curve fluctuations.
By using duration measurements as interest rate risk measurement across different bonds, it is
assumed that each bond experiences the same yield-to-maturity shift as the yield curve moves. This
only occurs if the zero coupon curve moves in a parallel fashion i.e. the whole curve moves up or
down. We have already discussed how, in general, this is not the case. Furthermore, for large yield
shifts, the convexity correction will not capture the price shift, so by using this measure we are
trapped in a small-movement, parallel-shift regime. This makes for naı̈ve hedging.

1.10.4 pv01
The pv01 of an instrument is how much its price changes when the yield changes by 1 basis point.
We also discard the minus sign. Since a basis point is 0.0001, a general result is that
𝑑𝑉
pv01 = −0.0001 (1.27)
𝑑𝑦

If we are considering a bond, then using (1.23) we have

pv01 = 0.0001𝐷$ (1.28)

Alternatively, if we have a yield curve, then it is the change in price when the curve moves in parallel
by 1 basis point. Then
𝑑𝑉
pv01 = −0.0001 (1.29)
𝑑𝑟
Duration calculations can give silly results, and usually the task at hand can best be handled with
pv01 instead.
For example, what is the duration of a swap? We can calculate ∂𝑉∂𝑦 , but to get the duration we must
(negate and) divide by 𝑉 , which could be close to (or exactly!) 0, and hence we have numerical
instability. Rather we should stick to the pv01 calculation and be done with it.
Usually the duration calculation should be done at the portfolio level, and then a swap should be
split up as short a fixed coupon bond and long a floating rate note. The fixed coupon bond duration
is found as normal and the floating rate note has a duration which is at most the time to the next
reset.

14
If one wants to have a robust definition of the duration of a swap on a stand alone basis, one should
say it is the duration of the floating rate note minus the duration of the fixed coupon bond. The
pv01 of a just starting swap is equal to the duration of the short fixed coupon bond, because the
duration of a just starting floating rate note is 0.

1.11 Exercises
1. The current discount rate, assuming a 360 day year, on a 90-day bill is 3.5%. The face value
is $1 million.

(a) What is the price of the bill?

(b) If the discount rate increases by one basis point to 3.51%, what is the change in the price
of the bill?
(c) If the discount rate decreases by one basis point to 3.49% percent, what is the change in
the price of the bill?
(d) Is the price function symmetric?

2. The continuously-compounded yield on a deposit which pays one million ZAR in 363 days’
time is 8.55%.

(a) What is the current value of the deposit?

(b) What is the yield expressed as a simple interest rate, assuming a 365 day year?

3. The current discount rate on a 91-day bill is 3.68 percent, assuming a 360 day year. What is
the simple interest rate, assuming a 365 day year?

4. Use excel to find the ytm of a NACS bond. So, the inputs are a bunch of cash flows, their
dates of payment, today’s date, and the total cash price; the output needs to be the ytm.
Make sure that not only newly issued bonds can be catered for. Use an actual/365 day count.
Write a macro which, each time it is played, calculates the next estimate of the ytm, based
on the previous one, using the Newton-Bailey method.

5. Write a vba function that, given

(a) period start date;

(b) current date;
(c) period end date;
(d) day count convention (abbreviated name).

will calculate the portion of the year already elapsed.

6. Show that if a newly issued bond prices at par then the ytm is equal to the coupon (with the
frequency of the ytm being the frequency of the coupon).

7. Create a spreadsheet which, given the yield curve as input, calculates the par bond curve and
the forward curve. Graph and label all three curves.

15
8. Create a spreadsheet that given a set of semi-annual bond cash flows with their dates, the ytm
of the bond, and today’s date, can calculate the duration, modified duration, and convexity
of the bond. The bond is not necessarily newly issued. Use an actual/36 day count. Fill out
the cash flows for the r153 and compare the answers you get with the answers you get from
your SAFM bond calculator.

9. A similar exercise: create a spreadsheet that given the yield curve values, a set of semi-annual
bond cash flows with their dates, and today’s date, can calculate the duration and convexity
of the bond. Do not calculate a ytm.

16
Chapter 2

Discrete-time Interest Rate Models

2.1 Introduction
The problem of interest rate derivatives can be approached broadly in one of two ways. We can
either model the interest rate process or the bond price process. If we are modelling the interest rate
process, then we must decide which interest rate(s) it is that we will model. In either case, there
are consistency conditions that must hold because of the structure of the yield curve: in particular,
implied forward rates should be positive.
In a discrete-time framework, we can use a similar model to that used in the binomial model of stock
price movements to model the bond prices and/or the interest rates. Modelling the price of the bond
requires us to ensure that the process is arbitrage-free, consistent with the initial term-structure
and obeys the boundary condition - that the price of a maturing bond is par, plus possibly the final
coupon.
In this chapter we will model the evolution of the short-rate (defined as the continuously-compounded
rate over Δ𝑡, the discrete time period for the lattice). The yield curve evolution is then governed
by one underlying factor. Substantial literature is devoted to multifactor models, where either more
than one source of uncertainty is modelled (usually principle components or their proxies), or the
evolution of the entire forward curve is modelled.
One-factor models are simple and tractable but they are not very flexible. In particular, they will
be hampered by their inability within the model for certain yield curve shapes to occur. In the most
simple models interest rates can become negative.

2.2 The basic lattice construction

Consider the following NACC term structure, with Δ𝑡 = 1:
𝑖 1 2 3 4 5 6 7 8
𝑟𝑖 6.1982% 6.4030% 6.8721% 7.0193% 7.2000% 6.9000% 6.9000% 7.0000%
𝜎𝑖 1.7000% 1.5000% 1.1000% 1.0000% 1.0000% 1.0000% 1.1000%

17
The risk free rate 𝑟𝑖 is the NACC rate for the period [0, 𝑡𝑖 ].
The volatilities are not the bond price volatilities. They are also not the volatilities of the risk free
rates that are from my yield curve bootstrap. They are the forward volatilities of the spot rate
(over Δ𝑡) over the time period [𝑡𝑖 , 𝑡𝑖+1 ], given the spot rates at time 𝑡 = 0. In other words, they are
the volatilities for the short rate for the period [𝑡𝑖 , 𝑡𝑖+1 ] that will be observed at time 𝑡𝑖 .
Of course, there is no 𝜎0 : the spot rate for the period from time 0 = 𝑡0 to time 𝑡1 = Δ𝑡 is known.
Similarly, given that the horizon in the above example is 𝑡8 , the volatility in the period [𝑡8 , 𝑡9 ] is
irrelevant.
We use binomial trees for the Δ𝑡 rate, the general notational convention of a tree will be as in Figure
2.1.

(0, 6)

(0, 5) 
*

H
j
H
(0, 4)  (1, 6)
*

H
j
H
(0, 3) 
*

H
j
H (1, 5) 
*

H
j
H
(0, 2) 
*

H
j
H (1, 4) 
*

H
j
H (2, 6)

(0, 1) 
*

H
j
H (1, 3) 
*

H
j
H (2, 5) 
*

H
j
H
(0, 0) 
*

H
j
H (1, 2) 
*

H
j
H (2, 4) 
*

H
j
H (3, 6)

(1, 1) 
*

H
j
H (2, 3) 
*

H
j
H (3, 5) 
*

H
j
H
(2, 2) 
*

H
j
H (3, 4) 
*

H
j
H (4, 6)

(3, 3)  (4, 5) 
*
 *

H
j
H H
j
H
(4, 4) 
*

H
j
H (5, 6)

(5, 5) 
*

H
j
H
(6, 6)

Figure 2.1: General node notation for binomial trees

Using the binomial assumption the spot rate in Δ𝑡 time takes one of two values: 𝑟𝑡 (0, 1) or 𝑟𝑡 (1, 1),
corresponding are the discount factors 𝑍 𝑡 (0, 1) and 𝑍 𝑡 (1, 1).1

2.3 Normal Distribution (Ho-Lee model)

The values we use here will be seen to be consistent with assuming that the spot rate process is
normally distributed: this is the Ho and Lee model Ho and Lee [1986].
1 The superscript 𝑡 means ‘tree’, in order to avoid confusion with the original discount factors from the original

yield curve. 𝑍 𝑡 (𝑖, 𝑗) is the discount factor for the following Δ𝑡 period given that we have evolved to node (𝑖, 𝑗). So
𝑍 𝑡 (0, 0) = 𝑍(0, 𝑡1 ) (!!).

18
Using martingale pricing theory on bond prices (analogously to stock prices) we know that there
exists a unique probability measure 𝑄 = {𝜋} such that the bond price normalised by the money-
market account follows a martingale. Under the martingale measure, the current forward bond price
is the expected price. Then

𝑍(0, 2) = 𝔼ℚ
0 [𝑍(1, 2)𝑍(0, 1)]

= 𝑍(0, 1)𝔼ℚ
0 [0] 𝑍(1, 2)

= 𝑍(0, 1)[𝜋𝑍 𝑡 (0, 1) + (1 − 𝜋)𝑍 𝑡 (1, 1)] (2.1)

where 𝜋 denotes the probability of an up move and 1 − 𝜋 the probability of a down move.
Now the interest rates 𝑟𝑡 (1, 1) and 𝑟𝑡 (0, 1) must match the volatility term structure. Note that in
general, if a variable 𝑟 can take on two values, 𝑎 and 𝑏, with 𝑎 > 𝑏, and with probabilities 𝜋 and
1 − 𝜋 respectively, then the variance of 𝑟 is given by

variance(𝑟) = 𝔼ℚ 𝑟2 − (𝔼ℚ [𝑟])2

[ ]

= 𝜋𝑎2 + (1 − 𝜋)𝑏2 − (𝜋𝑎 + (1 − 𝜋)𝑏)2

= (𝑎 − 𝑏)2 𝜋(1 − 𝜋) (2.2)

So, it is very good idea to assume that 𝜋 = 21 . (As usual, the model has three free parameters and
is matching two characteristic equations, the mean and variance. So there is one degree of freedom,
which we now use up.) Then

𝑍(0, 2) = 𝑍(0, 1) 12 [𝑒−𝑎 + 𝑒−𝑏 ]

1
stdev(𝑟) = 2 (𝑎 − 𝑏)

which is two equations in two unknowns, and solves easily:2 𝑟𝑡 (1, 1) = 8.3223%, 𝑟𝑡 (1, 0) = 4.9223%.
Now extend the lattice from one to two years, and assume that the volatility is time-dependent
but not state-dependent (i.e. volatility varies from left to right and not from top to bottom).
Furthermore, we will require that the tree recombines. This information is enough to calibrate the
tree in closed inductive form.
But first, in order to explore a bit further, some information from heaven: guess that the values for
the short-rate at time 𝑡 = 2 are 𝑟𝑡 (2, 2) = 10.8523%, 𝑟𝑡 (1, 2) = 7.8583% and 𝑟𝑡 (0, 2) = 4.8583%.
What do we need to do to check that this heavenly information is correct? First, and easily, the
volatility structure:
1 𝑡
2 (𝑟 (2, 2) − 𝑟𝑡 (1, 2)) = 1
2 (0.108583 − 0.078583) = 0.015
1 𝑡 𝑡 1
2 (𝑟 (1, 2) − 𝑟 (0, 2)) = 2 (0.078583 − 0.048583) = 0.015

Now,

𝑍(0, 3) = 𝑍(0, 1)𝔼ℚ

0 [𝑍(1, 3)]

= 𝑍(0, 1)𝔼ℚ
0 [𝑍(1, 2)𝑍(2, 3)]

= 𝑍(0, 1)[ 12 𝑍 𝑡 (1, 1)𝔼ℚ 1 𝑡 ℚ

0 [𝑍(2, 3)∣(1, 1)] + 2 𝑍 (0, 1)𝔼0 [𝑍(2, 3)∣(0, 1)]]

= 𝑍(0, 1)[ 12 𝑍 𝑡 (1, 1) 12 [𝑍 𝑡 (2, 2) + 𝑍 𝑡 (1, 2)] + 12 𝑍 𝑡 (0, 1) 12 [𝑍 𝑡 (1, 2) + 𝑍 𝑡 (0, 2)]
2 It doesn’t solve easily because it is two equations in two unknowns; that fact means that there should be a

solution. The ease of the solution comes from convenient facts of the exponential.

19
Exercise 2.3.1. We know from the previous steps the values of 𝑍(0, 1), 𝑍 𝑡 (1, 1) and 𝑍 𝑡 (0, 1). Check
that the heavenly values work in the above equation.

This makes the model parameters consistent with all the market information provided, and hence
the model is arbitrage free.

2.4 Formalising the Lattice Construction

We will have the binomial lattice of discount factors 𝑍 𝑡 (𝑖, 𝑗), which are the discount factors over the
following period if at time 𝑗 we are in state 𝑖. Typically the periods are three months apart. Here
𝑖 is a state (vertical) index and 𝑗 is a time (horizontal) index.
Let 𝜎(𝑗) be the annualised volatility of the Δ𝑡 month forward rates for the period [𝑡𝑗 , 𝑡𝑗+1 ]. In the
absence of implied information, the volatilities can only be calibrated via an historical analysis of
the yield curve. This volatility will be calibrated under the Ho-Lee type assumption of normality of
interest rates i.e. one takes differences of rates rather than log differences of rates in the calculation
estimating standard deviations. Nevertheless, the square root of time rule for volatilities still applies,
because we are assuming a Brownian evolution of the rate.
In the Ho-Lee model, there are equal probabilities for evolving to the two subsequent nodes in the
tree, and furthermore by a normality assumption the difference in adjacent interest rates at time 𝑗
for the period from then to time 𝑗 + 1 is independent of the state 𝑖, so
√
𝑟𝑡 (𝑖 + 1, 𝑗) = 𝑟𝑡 (𝑖, 𝑗) + 2𝜎(𝑗) Δ𝑡 (2.3)

Hence

𝑍 𝑡 (𝑖 + 1, 𝑗) =𝑍 𝑡 (𝑖, 𝑗)𝐸(𝑗) (1 ≤ 𝑖 ≤ 𝑗 − 1) (2.4)

[ ]
𝐸(𝑗) := exp −2𝜎(𝑗)Δ𝑡3/2 (2.5)

To proceed, define new variables 𝜆(𝑖, 𝑗). This variable is called the Arrow-Debreu price, and is the
price at time 0 of a security that pays off exactly one if we pass through the node (𝑖, 𝑗), the payoff
occurring at the moment of passing through. Thus

𝜆(0, 0) = 1 (2.6)
1 𝑡
𝜆(0, 𝑗) = 2 𝜆(0, 𝑗 − 1)𝑍 (0, 𝑗 − 1) (2.7)
1 𝑡 1 𝑡
𝜆(𝑖, 𝑗) = 2 𝜆(𝑖 − 1, 𝑗 − 1)𝑍 (𝑖 − 1, 𝑗 − 1) + 2 𝜆(𝑖, 𝑗 − 1)𝑍 (𝑖, 𝑗 − 1) (0 < 𝑖 < 𝑗) (2.8)
1 𝑡
𝜆(𝑗, 𝑗) = 2 𝜆(𝑗 − 1, 𝑗 − 1)𝑍 (𝑗 − 1, 𝑗 − 1) (2.9)

Now, by no arbitrage we have

𝑗
∑
𝑍(0, (𝑗 + 1)Δ𝑡) = 𝜆(𝑖, 𝑗)𝑍 𝑡 (𝑖, 𝑗) (2.10)
𝑖=0

and now recalling (2.4) we have

𝑗
∑ 𝑗
∑
𝑍(0, (𝑗 + 1)Δ𝑡) = 𝜆(𝑖, 𝑗)𝑍 𝑡 (𝑖, 𝑗) = 𝑍 𝑡 (0, 𝑗) 𝜆(𝑖, 𝑗)𝐸(𝑗)𝑖
𝑖=0 𝑖=0

20
 𝜆(𝑖, 𝑗) 0 1 2 3 4 5 6 7 8
0 1.0000 0.4699 0.2237 0.1065 0.0511 0.0245 0.0122 0.0060 0.0030
1 0.4699 0.4399 0.3099 0.1963 0.1170 0.0692 0.0396 0.0223
2 0.2162 0.3003 0.2830 0.2230 0.1636 0.1113 0.0724
3 0.0970 0.1813 0.2126 0.2061 0.1737 0.1342
4 0.0435 0.1013 0.1460 0.1627 0.1553
5 0.0193 0.0552 0.0914 0.1151
6 0.0087 0.0285 0.0533
7 0.0038 0.0141
8 0.0016
𝑍 𝑡 (𝑖, 𝑗) 0 1 2 3 4 5 6 7
0 0.9399 0.9520 0.9526 0.9586 0.9606 0.9946 0.9891 0.9971
1 0.9201 0.9244 0.9377 0.9416 0.9749 0.9695 0.9754
2 0.8971 0.9173 0.9229 0.9556 0.9503 0.9541
3 0.8974 0.9046 0.9366 0.9315 0.9334
4 0.8867 0.9181 0.9130 0.9131
5 0.8999 0.8949 0.8932
6 0.8772 0.8738
7 0.8547
𝑡
𝑟 (𝑖, 𝑗) 0 1 2 3 4 5 6 7
0 6.198% 4.922% 4.858% 4.231% 4.023% 0.545% 1.100% 0.295%
1 8.322% 7.858% 6.431% 6.023% 2.545% 3.100% 2.495%
2 10.858% 8.631% 8.023% 4.545% 5.100% 4.695%
3 10.831% 10.023% 6.545% 7.100% 6.895%
4 12.023% 8.545% 9.100% 9.095%
5 10.545% 11.100% 11.295%
6 13.100% 13.495%
7 15.695%

Figure 2.2: The Ho-Lee trees associated with the given data

𝑍(0, (𝑗 + 1)Δ𝑡)
𝑍 𝑡 (0, 𝑗) = ∑𝑗 (2.11)
𝑖
𝑖=0 𝜆(𝑖, 𝑗)𝐸(𝑗)

Thus our recursion is as follows, at time step 𝑗:

(a) Calculate 𝜆(𝑖, 𝑗) for 0 ≤ 𝑖 ≤ 𝑗 from (2.6), (2.7), (2.8) and (2.9).

(b) Calculate 𝑍 𝑡 (0, 𝑗) from (2.11).

(c) Calculate 𝑍 𝑡 (𝑖, 𝑗) for 1 ≤ 𝑖 ≤ 𝑗 from (2.4).

Exercise 2.4.1. Verify Figure 2.2 for the given data.

21
2.5 Lognormal Distribution (Black-Derman-Toy model)
Discrete time models of interest rates where the forward rates are lognormally distributed are usually
taken to be some variation of Black et al. [1990]. The problem with the normality assumption in the
previous section is that it is possible for interest rates to become negative (even if a mean reversion
factor is introduced). Clearly, a lognormal assumption precludes this.
Assume now that the logarithm of the spot rate is normally distributed, and the volatility parameter
pertains to the logarithms of the interest rates rather than the interest rates themselves. So now
√
ln 𝑟𝑡 (𝑖 + 1, 𝑗) = ln 𝑟𝑡 (𝑖, 𝑗) + 2𝜎(𝑗) Δ𝑡
√
𝑟𝑡 (𝑖 + 1, 𝑗) = 𝑟𝑡 (𝑖, 𝑗) exp(2𝜎(𝑗) Δ𝑡) (2.12)

Hence

𝑍 𝑡 (𝑖 + 1, 𝑗) = 𝑍 𝑡 (𝑖, 𝑗)𝐸(𝑗) (1 ≤ 𝑖 ≤ 𝑗 − 1) (2.13)

[ √ ]
𝐸(𝑗) := exp 2𝜎(𝑗) Δ𝑡 (2.14)

which are calculated in advance. To proceed, again define new variables 𝜆(𝑖, 𝑗) exactly as before:
(2.6), (2.7), (2.8), (2.9) and (2.10) are unchanged. Thus
𝑗 𝑗
∑ ∑ 𝑖
𝑍(0, (𝑗 + 1)Δ𝑡) = 𝜆(𝑖, 𝑗)𝑍 𝑡 (𝑖, 𝑗) = 𝜆(𝑖, 𝑗)𝑍 𝑡 (0, 𝑗)𝐸(𝑗)
𝑖=0 𝑖=0

which does not have a closed form solution. Hence, we need to solve this for 𝑍 𝑡 (0, 𝑗) numerically.
Let 𝑥 = 𝑍 𝑡 (0, 𝑗). Using Newton’s method, we solve

𝑥1 = 𝑍 𝑡 (0, 𝑗 − 1)
∑𝑗 𝐸(𝑗)𝑖
𝜆(𝑖, 𝑗)𝑥𝑛 − 𝑍(0, (𝑗 + 1)Δ𝑡)
𝑥𝑛+1 = 𝑥𝑛 − 𝑖=0∑𝑗 𝑖
𝑖 𝐸(𝑗) −1
𝑖=0 𝜆(𝑖, 𝑗)𝐸(𝑗) 𝑥𝑛

Convergence here is extremely rapid: to double precision in 3 or 4 iterations. The Newton function
above is near linear (in a very wide range). Moreover, the iteration can be performed simultaneously
across all 𝑗.
Note that the term structure of volatilities in this lognormal model will reflect higher values than
those under the normality assumption. We see this via the following:
Δ𝑟
ln[𝑟 + Δ𝑟] − ln[𝑟] = ln[𝑟(1 + 𝑟 )] − ln[𝑟]
Δ𝑟
= ln[1 + 𝑟 ]
≃ 3 Δ𝑟
𝑟

for small Δ. This implies that,

𝜎normal
𝜎lognormal = 4 (2.15)
𝑟

that the Taylor series of ln(1 + 𝑥) is 𝑥 − 21 𝑥2 + 13 𝑥3 + ⋅ ⋅ ⋅ .

3 Recall
4 So,
the classic rule of thumb in South Africa: to get from market volatilities to ballpark Ho-Lee volatilities: knock
off a decimal place. Quoted volatilities will be lognormal because use of Black’s model is the default: see Chapter 3.

22
Exercise 2.5.1. Assume the following initial data, with Δ𝑡 = 14 :
𝑖 1 2 3 4 5 6 7 8
𝑟𝑖 6.1982% 6.4030% 6.8721% 7.0193% 7.1000% 7.2021% 7.3120% 7.3000%
𝜎𝑖 20.0000% 18.0000% 17.0000% 17.0000% 17.0000% 17.0000% 17.0000%
Check Figure 2.1.

𝜆(𝑖, 𝑗) 0 1 2 3 4 5 6 7 8
0 1.0000 0.4923 0.2425 0.1193 0.0588 0.0290 0.0143 0.0071 0.0035
1 0.4923 0.4842 0.3568 0.2342 0.1443 0.0854 0.0492 0.0278
2 0.2417 0.3556 0.3496 0.2870 0.2121 0.1465 0.0966
3 0.1181 0.2319 0.2851 0.2806 0.2419 0.1912
4 0.0577 0.1416 0.2087 0.2396 0.2365
5 0.0281 0.0827 0.1422 0.1870
6 0.0136 0.0468 0.0923
7 0.0066 0.0260
8 0.0032
𝑍 𝑡 (𝑖, 𝑗) 0 1 2 3 4 5 6 7
0 0.9846 0.9852 0.9839 0.9858 0.9871 0.9877 0.9883 0.9903
1 0.9820 0.9808 0.9832 0.9847 0.9854 0.9862 0.9885
2 0.9771 0.9801 0.9819 0.9827 0.9836 0.9864
3 0.9765 0.9785 0.9796 0.9806 0.9839
4 0.9746 0.9758 0.9771 0.9810
5 0.9714 0.9729 0.9775
6 0.9680 0.9734
7 0.9685
𝑡
𝑟 (𝑖, 𝑗) 0 1 2 3 4 5 6 7
0 6.198% 5.950% 6.473% 5.723% 5.213% 4.961% 4.696% 3.894%
1 7.267% 7.750% 6.783% 6.179% 5.880% 5.566% 4.616%
2 9.278% 8.041% 7.325% 6.970% 6.598% 5.471%
3 9.530% 8.682% 8.261% 7.820% 6.485%
4 10.291% 9.792% 9.270% 7.687%
5 11.606% 10.987% 9.111%
6 13.023% 10.799%
7 12.800%

Table 2.1: The Black-Derman-Toy trees associated with the given data

2.6 Options on Zero-Coupon Bonds

The procedure that we follow is the usual martingale pricing approach. We generate an (arbitrage-
free) tree of discount factors and then use risk neutral valuation. Using this basis we can also price
complicated interest rate derivatives. In general, the option will have maturity 𝑇 and will be written

23
on an instrument with maturity 𝑇1 where (𝑇1 > 𝑇 ).
As a simple example, consider such an option on a zero coupon bill. By arbitrage considerations,
the value of the bill at maturity of the option 𝑇 will be 𝑍(𝑇, 𝑇1 ). The value of such a bill today
is 𝑍(0; 𝑇, 𝑇1 ) ie. the forward discount value. But to value the option, we need to consider the
volatility. Then the European call and put boundary conditions are:
{
𝑍(𝑇, 𝑇1 ) − 𝐾 for 𝑍(𝑇, 𝑇1 ) > 𝐾
𝑐[𝑍(𝑇, 𝑇1 ), 0; 𝐾] = (2.16)
0 for 𝑍(𝑇, 𝑇1 ) ≤ 𝐾
{
0 for 𝑍(𝑇, 𝑇1 ) ≥ 𝐾
𝑝[𝑍(𝑇, 𝑇1 ), 0; 𝐾] = (2.17)
𝐾 − 𝑍(𝑇, 𝑇1 ) for 𝑍(𝑇, 𝑇1 ) < 𝐾

respectively.
As an example we will consider a 18 month option on a six month zero-coupon bond i.e. after 18
months we decide whether or not to buy/sell a zero-coupon bond at the strike price; the zero coupon
bond pays 1 after 2 years.
We use the Black-Derman-Toy model with the previous data. Given the tree of prices, there is very
little to do. All we need to realise is that at time 𝑡6 the bond has a value of

𝑉 (𝑖, 6) = 𝑍 𝑡 (𝑖, 6) 21 𝑍 𝑡 (𝑖, 7) + 𝑍 𝑡 (𝑖 + 1, 7)

[ ]
(2.18)

Thus, the payoff of the option is

𝑉 𝑡 (𝑖, 6) = max(𝜂(𝑉 (𝑖, 6) − 𝐾, 0))

The value of the option now is

6
∑
𝑉 (0) = 𝜆(𝑖, 6)𝑉 𝑡 (𝑖, 6)
𝑖=0

Alternatively we can induct backwards through the tree as usual. There are benefits to the latter
approach as it will enable us to derive hedge ratios.
Verify that, for a call option on a six month zero coupon bond, strike 0.95, we get the tree of option
values
0 1 2 3 4 5 6
0.0117 0.0146 0.0175 0.0203 0.0230 0.0255 0.0279
0.0091 0.0121 0.0152 0.0183 0.0211 0.0238
0.0065 0.0094 0.0127 0.0160 0.0191
0.0039 0.0065 0.0099 0.0134
0.0016 0.0033 0.0068
0.0000 0.0000
0.0000
and so a price of 0.0117.
Put-Call parity for bond options is analogous to that of European options written on non-dividend
paying stocks. Consider European put and call options written on zero-coupon bonds with 𝑝𝑎𝑟 = 1.
The option maturity is 𝑇 and the bond maturity is 𝑇1 , where 𝑇1 > 𝑇 . Then

𝑝[𝑍(𝑡, 𝑇1 ), 𝑇 − 𝑡; 𝐾] + 𝑍(𝑡, 𝑇1 ) = 𝑐[𝑍(𝑡, 𝑇1 ), 𝑇 − 𝑡; 𝐾] + 𝐾𝑍(𝑡, 𝑇 ) (2.19)

24
2.7 Forwards on Zero-Coupon Bonds
The forward value of the bond is
6
∑
𝑉 (0) = 𝑝𝑖 𝑉 (𝑖, 6)
𝑖=0

where 𝑝𝑖 is the probability of arriving at the node 𝑖. These probabilities may be found using Pascal’s
triangle, or directly via combinatorial arguments; 𝑝𝑖 = 2−6 6𝑖 . Hence the forward value is
()

6 ( )
∑ 6
𝑉 (0) = 2−6 𝑉 𝑡 (𝑖, 6)
𝑖=0
𝑖

2.8 Hedging Options

Replicating portfolios for options can be constructed using bonds or forwards. Consider the BDT
example as previously. The hedging portfolio must contain two instruments (because of the binomial
tree structure). We can choose any two of the zero coupon bonds that can be found in the market.
Let us choose bonds maturing at time 𝑡1 and 𝑡2 ; we seeking hedge quanta 𝑛1 and 𝑛2 .

𝑉 (0) = 𝑍(0, 𝑡1 )𝑛1 + 𝑍(0, 𝑡2 )𝑛2

𝑉 𝑡 (0, 1) = 𝑛1 + 𝑍 𝑡 (0, 1)𝑛2
𝑉 𝑡 (1, 1) = 𝑛1 + 𝑍 𝑡 (1, 1)𝑛2

in other words

0.0117 = 0.9846𝑛1 + 0.9685𝑛2

0.0146 = 𝑛1 + 0.9852𝑛2
0.0091 = 𝑛1 + 0.9820𝑛2

Although this appears to be an over-specified system, it isn’t really, as the first equation follows
from the others, so we discard it. Thus we have a 2 × 2 system:
[ ][ ] [ ]
1 0.9852 𝑛1 0.0146
=
1 0.9820 𝑛2 0.0091

which solves as 𝑛1 = −1.64, 𝑛2 = 1.6793 by using Cramer’s rule, say.

What can you say about the effectiveness of this hedge? How can it be modified to be more robust?
Hedging can also be done with 2-period forwards. The tree of forwards prices is:
0 1 2 3 4 5 6
0.9627 0.9658 0.9686 0.9712 0.9736 0.9758 0.9779
0.9596 0.9629 0.9660 0.9688 0.9714 0.9738
0.9562 0.9598 0.9632 0.9662 0.9691
0.9526 0.9565 0.9601 0.9634
0.9487 0.9529 0.9568
0.9444 0.9490
0.9398

25
Here, the last column is just the values of (2.18). We then induct to the left: the value in cell (𝑖, 𝑗) is
just the average of cells (𝑖, 𝑗 + 1) and (𝑖 + 1, 𝑗 + 1). As rates evolve, so the forward price ‘converges’
to the actual price after 18 months. Define the delta of the option with respect to the forwards
contract in the usual fashion:
𝑉 𝑡 (0, 1) − 𝑉 𝑡 (1, 1) 0.0146 − 0.0091
Δ= = = 0.8763
𝑉𝑓𝑡 (0, 1) − 𝑉𝑓𝑡 (1, 1) 0.9658 − 0.9596
Then replicate the option with a forward position and a cash account, with quanta 𝑚1 and 𝑚2
respectively:

𝑉 (0) = 𝑚1 0 + 𝑚2 1
𝑡
𝑉 (0, 1) = 𝑚1 [𝑉𝑓𝑡 (0, 1) − 𝑉𝑓𝑡 (0, 0)] + 𝑚2 𝑍(0, 1)−1
𝑉 𝑡 (1, 1) = 𝑚1 [𝑉𝑓𝑡 (1, 1) − 𝑉𝑓𝑡 (0, 0)] + 𝑚2 𝑍(0, 1)−1

in other words

0.0117 = 𝑚2
0.0146 = [0.9658 − 0.9627]𝑚1 + 1.0156𝑚2
0.0091 = [0.9596 − 0.9627]𝑚1 + 1.0156𝑚2

which implies that 𝑚1 = 0.8763, 𝑚2 = 0.0117. As usual, the hedge ratio is dynamic and we need
to re-balance the hedge portfolio at 𝑡 = 1, depending on up/down movement from 𝑡 = 0.

2.9 Exercises
1. Consider the following quarterly data, satisfying the conditions of the Black-Derman-Toy
model:

1 2 3 4 5 6 7 8
zero 6.00% 6.50% 6.30% 7.00% 7.00% 7.00% 7.00% 7.00%
vol 20.00% 18.00% 17.00% 16.00% 15.00% 15.00% 14.00%

(1) Create arrays in vba that will store these rates.

(2) Type these rates into excel, and write vba code to read them (using a loop) into the two
arrays.
(3) Create in the same way as above arrays ArrowDebreu(0 to 8,0 to 8), Bt(0 to 7,0 to 7),
and ratet(0 to 7, 0 to 7).
(4) Create the time index j and the state index i. By looping through time, populate (the
upper triangle) of all of these matrices. Here you use Newton iteration.
Remark: all your dims should be together, at the start of your code. It is bad form to
dim things only as you need them. However, the redims can be in the heart of the matter.
(A more sophisticated approach will have the number of data points as a variable and the
redims can involve this variable. So, by necessity, the redim can only occur after some
calculation work has been done.)

26
(5) Use loops to again ‘print’ the output values in excel.
(6) Price the options and forwards seen in class in vba.

27
Chapter 3

Black’s Model

We consider the Black Model for futures/forwards which is the market standard for quoting prices
(via implied volatilities). Black [1976] considered the problem of writing options on commodity
futures and this was the first “natural” extension of the Black-Scholes model. This model also is
used to price options on interest rates and interest rate sensitive instruments such as bonds. Since
the Black-Scholes analysis assumes constant (or deterministic) interest rates, and so forward interest
rates are realised, it is difficult initially to see how this model applies to interest rate dependent
derivatives.
However, if 𝑓 is a forward interest rate, it can be shown that it is consistent to assume that

• The discounting process can be taken to be the existing yield curve.

• The forward rates are stochastic and log-normally distributed.

The forward rates will be log-normally distributed in what is called the 𝑇 -forward measure, where
𝑇 is the pay date of the option. This model is consistent is within the domain of the LIBOR market
model. We can proceed to use Black’s model without knowing any of the theory of the LMM;
however, Black’s model cannot safely be used to value more complicated products where the payoff
depends on observations at multiple dates.

3.1 European Bond Options

The clean (quoted) price for a bond is related to the all-in (dirty, cash) price via:

𝔸 = ℂ + 𝕀𝐴 (𝑡) (3.1)

where the accrued interest 𝕀𝐴 (𝑡) is the accrued interest as of date 𝑡, and is non-zero between coupon
dates. The forward price is a carried all-in price, not a clean price. The option strike price 𝕂 might
be a clean or all-in strike; usually it is clean. If so, we change it to a all-in price by replacing 𝕂 with
𝕂𝔸 = 𝕂 + 𝕀𝔸 (𝑇 ).

28
Applying Black’s model to the price of the bond, the value of the bond option per unit of nominal
is [Hull, 2005, §26.2]

𝑉𝜂 = 𝜂𝑍(0, 𝑇 )[𝔽𝔸 𝑁 (𝜂𝑑1 ) − 𝕂𝔸 𝑁 (𝜂𝑑2 )] (3.2)

𝔽𝔸 1 2
ln 𝕂𝔸 ± 2𝜎 𝜏
𝑑1,2 = √ (3.3)
𝜎 𝜏
where 𝜂 = 1 stands for a call, 𝜂 = −1 for a put. Here 𝜎 is the volatility measure of the fair forward
all in price, and 𝜏 as usual is the term of the option in years with the relevant day-count convention
applied.

Example 3.1.1. Consider a 10𝑚 European call option on a 1, 000, 000 bond with 9.75 years to
maturity. Suppose the coupon is 10% NACS. The clean price is 935,000 and the clean strike price
is 1,000,000. We have the following yield curve information:
Term Rate
3m 9%
9m 9.5%
10m 10%
The 10 month volatility on the bond price is 9%.
Firstly, 𝕂ℂ = 1, 000, 000 so 𝕂𝔸 = 𝕂ℂ + 𝕀𝔸 (𝑇 ) = 1, 008, 333.33. (There will be one month of accrued
interest in 10 months time.)
Secondly, 𝔸 = [ℂ + 𝕀𝔸 (0) = 960, 000.
[ (There is currently
]] three months of accrued interest.)
3 9 10
Thirdly, 𝔽𝔸 = 960, 000 − 50, 000 𝑒− 12 9% + 𝑒− 12 9.5% 𝑒 12 10% = 939, 683.97.
Hence 𝑉𝑐 = 7, 968.60 and 𝑉𝑝 = 71, 129.06.

3.1.1 Different volatility measures

The volatility above is a price volatility measure. However, quoted volatilities are often yield volatil-
ity measures. The relationship between the various volatilities of the bond is given via Ito’s lemma
as
𝜎𝑦 𝑦Δ
𝜎𝔸 = − (3.4)
𝔸
𝜎𝔸 𝔸
𝜎𝑦 = − (3.5)
𝑦Δ
ℂ
𝜎𝔸 = 𝜎ℂ (3.6)
𝔸
How do we see this? Note that
𝑑𝑦 = 𝜇𝑦 𝑑𝑡 + 𝜎𝑦 𝑦 𝑑𝑍
is the geometric Brownian motion for the yield 𝑦. Now 𝔸 = 𝑓 (𝑦) and so
Δ𝜎𝑦 𝑦
𝑑𝔸 = ⋅ ⋅ ⋅ 𝑑𝑡 + 𝑓 ′ (𝑦)𝜎𝑦 𝑦 𝑑𝑍 := ⋅ ⋅ ⋅ 𝑑𝑡 + 𝔸 𝑑𝑍
𝔸
But, also,
𝑑𝔸 = 𝜈𝔸 𝑑𝑡 + 𝜎𝔸 𝔸 𝑑𝑍
and so the result follows - except for a missing minus sign. Why?

29
3.2 Caplets and Floorlets
See [Hull, 2005, §26.3]. Suppose that a market participant loses money if the floating rate falls.
For example, they are long a FRA. The floorlet pays off if the floating rate decreases below a
predetermined minimum value, which is of course called the strike. Thus, the floorlet ‘tops-up’
the payment received in the FRA, if required, so that the net rate received is at least the strike.
The floating rate will be the prevailing 3-month LIBOR rate, which is set at the beginning of each
period, with settlement at that time by discounting, much like a FRA. Indeed, it is a FRA position
that this floorlet is hedging; the FRA date schedule is being applied.
Let the floorlet rate (strike) be 𝑟𝐾 .
Let the 𝑇0 × 𝑇1 FRA period be of length 𝛼 as usual, where day count conventions are observed.
Suppose the LIBOR rate for the period, observed at the determination date is 𝐽1 . Then the payoff
at the determination date is 1+𝐽1 1 𝛼 𝛼 max(𝑟𝐾 − 𝐽1 , 0) per unit notional; a floorlet is like a put on the
interest rate. Of course the valuation is the same whether settled in advance or arrears.
In an analogous fashion to a floorlet, we have a caplet, where the payout occurs when the floating
rate rises above 𝑟𝐾 , the cap rate.

3.3 Caps and Floors

A cap is like a strip of caplets which will be used to hedge a swap. However, because swaps are
settled in arrears so too is each payment in the cap strip, unlike an individual caplet. Moreover, the
cap strip will have the swap date schedule applied and not the FRA date schedule. So, caps stand
to swaps exactly as caplets stand to FRAs.
A cap might be forward starting or spot starting (that is, starting immediately). However, in the
latter case, there is no payment in 3 months time - it is excluded from the computations and from any
payments because there is no optionality. Thus, for example, a 2y cap actually has seven payments,
not eight. Alternatively, one might consider that a spot starting cap is actually a forward starting
cap starting in 3m time.
Let the cap rate (strike) be 𝑟𝐾 . Let the 𝑖𝑡ℎ reset period from 𝑡𝑖−1 to 𝑡𝑖 be of length 𝛼𝑖 as usual,
where day count conventions are observed. Suppose the LIBOR rate for the period, observed at
time 𝑡𝑖−1 , is 𝐽𝑖 . Then the payoff at time 𝑡𝑖 is 𝛼𝑖 max(𝐽𝑖 − 𝑟𝐾 , 0) per unit notional.
In an analogous fashion to writing a cap, we can write a floor, where the payout occurs when the
floating rate drops below 𝑟𝐾 , the floor rate.

30
3.3.1 Valuation
We value each caplet or floorlet separately off the yield curve using the implied forward rates at
𝑡 = 0, for each time period 𝑡𝑖 . Then,
𝑛
∑
𝑉 = 𝑉𝑖 (3.7)
𝑖=1
𝑍(0, 𝑡𝑖 )𝛼𝑖 𝜂 𝑓 (0; 𝑡𝑖−1 , 𝑡𝑖 )𝑁 𝜂𝑑𝑖1 − 𝑟𝐾 𝑁 𝜂𝑑𝑖2
[ ( ) ( )]
𝑉𝑖 = (3.8)
𝑓 (0;𝑡𝑖−1 ,𝑡𝑖 ) 1 2
ln 𝑟𝐾 ± 2 𝜎𝑖 𝑡𝑖−1
𝑑𝑖1,2 = √ (3.9)
𝜎𝑖 𝑡𝑖−1

where 𝜂 = 1 stands for a cap(let), 𝜂 = −1 for a floor(let), and where the floating rate (swap) curve
is being used. 𝑓 (0; 𝑡𝑖−1 , 𝑡𝑖 ) is the simple forward rate for the period from 𝑡𝑖−1 to 𝑡𝑖 . So we apply
(1.11). Cap prices are quoted with 𝜎𝑖 = 𝜎 for 𝑖 = 1, 2, . . . , 𝑛. On the other hand, if we are
assembling a set of caplets into a cap, then the 𝜎𝑖 will be different.
Note that the 𝑖𝑡ℎ caplet is being valued in the 𝑡𝑖 forward measure.

3.3.2 A call/put on rates is a put/call on a bond

A caplet (a call on an interest rate) is actually a put on a floating-rate bond whose yield is the LIBOR
floating rate (a put option because of the inverse relationship between yield and bond price).
To see this, each caplet has a payoff in arrears of 𝛼𝑖 max(𝑟𝑖 − 𝑟𝐾 , 0). The value of this in advance is

𝑉 (𝑡𝑖−1 ) = (1 + 𝛼𝑖 𝑟𝑖 )−1 𝛼𝑖 max(𝑟𝑖 − 𝑟𝐾 , 0)

( )
𝛼𝑖 𝑟𝑖 − 𝛼𝑖 𝑟𝐾
= max ,0
1 + 𝛼𝑖 𝑟𝑖
( )
1 + 𝛼𝑖 𝑟𝐾
= max 1 − ,0
1 + 𝛼𝑖 𝑟𝑖
( )
1 1
= (1 + 𝛼𝑖 𝑟𝐾 ) max − ,0
1 + 𝛼𝑖 𝑟𝐾 1 + 𝛼𝑖 𝑟𝑖
which at time 𝑡 = 𝑡0 is 1 + 𝛼𝑖 𝑟𝐾 many puts on a zero coupon bond maturing at time 𝑡𝑖 with the
option exercise at 𝑡𝑖−1 , strike 1+𝛼1𝑖 𝑟𝐾 .
Likewise, floors - which are European put options on rates - are actually European call options on
the (underlying) floating-rate bond.

Example 3.3.1. Suppose we have a given term structure

Term Rate 𝑍
0.75 11.00% 0.92081
1 11.30% 0.89315
and consider a 9 × 12 caplet with strike 𝑟𝐾 = 12.1818% and yield volatility 𝜎𝑦 = 10%. Then
( 0.92081 )
the forward period is 𝛼 = 0.25 and the forward NACQ rate is given by 𝑟𝐹 = 0.89315 − 1 0.25 =
12.3880%. We find 𝑑1 = 0.23708 and 𝑑2 = 0.15048 using (3.9), and calculate 𝑁 (𝑑1 ) = 0.59370 and
𝑁 (𝑑2 ) = 0.55981. Thus, using (3.8), we obtain the value of the caplet as 𝑉𝑐 = 0.0011953.
1
Now we consider the put on a bond. We find the strike 𝐾 = 1+𝛼𝑟 𝐾
= 0.97045, price volatility
𝜎 = 𝜎𝑦 𝑟𝐹 𝛼 = 0.3097% (using (3.4) - 𝛼 is the duration of the forward bond). From (3.3), calculate

31
𝑑1 = −0.18509 and 𝑑2 = −0.18778, and hence find 𝑁 (𝑑1 ) = 0.57342 and 𝑁 (𝑑2 ) = 0.57447. Then
the value of the bond option from (3.2) is 𝑉𝑏 = 0.00116. However, we still need to take into account
the size which is given by 1 + 𝛼𝑟𝐾 = 1.03045. Multiplying the size with 𝑉𝑏 yields 0.0011952.

Differences between these two values typically occur at the 5𝑡ℎ decimal place. It is impossible,
mathematically, to have these two values equal; in the caplet model, rates are lognormal and in the
bond option model, bond prices are lognormal. These two models cannot be made compatible.

3.3.3 Greeks
Let 𝑟 denote the entire yield curve.
We need the following preliminary calculations:
∂
𝑓 (0; 𝑡𝑖−1 , 𝑡𝑖 ) = 1 (3.10)
∂𝑟
∂
𝑍(0, 𝑡𝑖 ) = −𝑡𝑖 𝑍(0, 𝑡𝑖 ) (3.11)
∂𝑟
∂
𝑍(0, 𝑡𝑖 ) = −𝑟𝑖 𝑍(0, 𝑡𝑖 ) (3.12)
∂𝜏
∂ 𝑖 ∂ 𝑖 1
𝑑 = 𝑑 = √ (3.13)
∂𝑟 1 ∂𝑟 2 𝑓 (0; 𝑡𝑖−1 , 𝑡𝑖 )𝜎𝑖 𝑡𝑖−1
∂ 𝑖 ∂ 𝑖 √
𝑑 = 𝑑 + 𝑡𝑖−1 (3.14)
∂𝜎 1 ∂𝜎 2
∂ ∂ 𝜎
𝑑𝑖 = 𝑑𝑖 + √ (3.15)
∂𝑡𝑖−1 1 ∂𝑡𝑖−1 2 2 𝑡𝑖−1

pv01

[
∂𝑉𝑖 ∂
= −𝑡𝑖 𝑉𝑖 + 𝑍(0, 𝑡𝑖 )𝛼𝑖 𝜂 𝑓 (0; 𝑡𝑖−1 , 𝑡𝑖 )𝑁 (𝜂𝑑𝑖1 )
∂𝑟 ∂𝑟
]
∂ ∂
+𝑓 (0; 𝑡𝑖−1 , 𝑡𝑖 )𝑁 ′ (𝜂𝑑𝑖1 ) 𝑑𝑖1 − 𝑟𝑋 𝑁 ′ (𝜂𝑑𝑖2 ) 𝑑𝑖2
∂𝑟 ∂𝑟
= −𝑡𝑖 𝑉𝑖 + 𝑍(0, 𝑡𝑖 )𝛼𝑖 𝜂𝑁 (𝜂𝑑𝑖1 ) [1 + 𝛼𝑖 𝑓 (0; 𝑡𝑖−1 , 𝑡𝑖 )] (3.16)
𝑛
∂𝑉 ∑ ∂𝑉𝑖
= (3.17)
∂𝑟 𝑖=1
∂𝑟

1 ∂𝑉
Then pv01 = 10000 ∂𝑟 .

Vega

Vega is ∂𝑉∂𝜎 . It is the Greek w.r.t. the cap volatility; it is not a Greek with respect to the caplet
volatilities, so 𝜎𝑖 = 𝜎 for 𝑖 = 1, 2, . . . , 𝑛.

32
 [ ]
∂𝑉𝑖 ′ 𝑖 ∂ 𝑖 ′ 𝑖 ∂ 𝑖
= 𝑍(0, 𝑡𝑖 )𝛼𝑖 𝑓 (0; 𝑡𝑖−1 , 𝑡𝑖 )𝑁 (𝑑1 ) 𝑑1 − 𝑟𝑋 𝑁 (𝑑2 ) 𝑑2
∂𝜎 ∂𝜎 ∂𝜎
′ 𝑖
√
= 𝑍(0, 𝑡𝑖 )𝛼𝑖 𝑓 (0; 𝑡𝑖−1 , 𝑡𝑖 )𝑁 (𝑑1 ) 𝑡𝑖−1 (3.18)
𝑛
∂𝑉 ∑ ∂𝑉𝑖
= (3.19)
∂𝜎 𝑖=1
∂𝜎

Bucket (Caplet/Floorlet) Vega

For sensitivities to caplet volatilities, we would be looking at a bucket risk type of scenario. This
∂𝑉
would be ∂𝜎 𝑖
, and this only makes sense of course in the case where we have individual forward-
forward volatilities rather than just a flat cap volatility.
∂𝑉𝑖
𝑍(0, 𝑡𝑖 )𝛼𝑖 𝑓 (0; 𝑡𝑖−1 , 𝑡𝑖 )𝑁 ′ (𝑑𝑖1 ) 𝑡𝑖−1
√
=
∂𝜎𝑖
where now it is the caplet volatility 𝜎𝑖 being used, not the cap volatility 𝜎.

Theta

Theta can only be calculated w.r.t. some explicit assumptions about yield curve evolution. The
assumption can be that when time moves forward, the bootstrapped continuous curve will remain
constant. By full revaluation we can then calculate the theta of the position with one day of time
decay.

Delta hedging caps/floors with swaps

We require the ‘with delta’ value of a cap or floor. Often the client wants to do the deal ‘with
delta’, which means that the ‘linear’ hedge comes with the option trade. Thus we quote delta based
on hedging the cap/floor with a (forward starting) swap with the same dates, basis and frequency.
Then
∂𝑉 (cap)
∂𝑟 pv01(cap)
Δ= ∂𝑉 (swap)
= (3.20)
pv01(swap)
∂𝑟

The numerator is found in (3.17). For the denominator: the value of the swap fixed receiver, fixed
rate 𝑅 having been set, is
𝑛
∑
𝑉 (swap) = 𝑅 𝛼𝑖 𝑍(0, 𝑡𝑖 ) − 𝑍(0, 𝑡0 ) + 𝑍(0, 𝑡𝑛 ) (3.21)
𝑖=1

as seen in (3.30), (3.31). So the derivative is

𝑛
∂𝑉 (swap) ∑
= −𝑅 𝛼𝑖 𝑡𝑖 𝑍(0, 𝑡𝑖 ) + 𝑡0 𝑍(0, 𝑡0 ) − 𝑡𝑛 𝑍(0, 𝑡𝑛 ) (3.22)
∂𝑟 𝑖=1

33
 Figure 3.1: A cap with its forward swap hedge

3.4 Stripping Black caps into caplets

Since each caplet is valued separately we expect a different volatility measure for each. Cap volatil-
ities are always quoted as flat volatilities where the same volatility is used for each caplet, which
in some sense will be a weighted average of the individual caplet volatilities. Thus, we use 𝜎 = 𝜎𝑖
for 𝑖 = 1, 2, . . . , 𝑛. Most traders work with independent volatilities for each caplet, though, and
these are called forward-forward vols. There exists a hump at about 1 year for the forward-forward
vols (and, consequently, also for the flat vol, which can be seen as a cumulative average of the
forward-forward vols). This can be observed or backed out of cap prices: see Figure 3.2.

Figure 3.2: Caplet (fwd-fwd) and cap (flat fwd-fwd) volatility term structure

Thus, from a set of caplet volatilities 𝜎1 , 𝜎2 , . . . , 𝜎𝑛 we may need to determine the corresponding
cap volatility 𝜎. This of course is a uniquely determined implied volatility problem. It only makes
sense if the strikes of all the caplets are equal to the strike of the cap; this won’t be the case in

34
general (typically quoted volatilities will be at the money).
A far more difficult problem is the specification of the term structure of caplet volatilities given an
incomplete term structure of cap volatilities. This is a type of bootstrap problem: there is insufficient
information to determine a unique solution. For example, only 3x6, 6x9 and 9x12 caplets and 1y,
2y, 3y , 4y caps might be available. This critical point is generally not really well dealt with in the
textbooks.
Instantaneous forward rate volatilities will be specified. Suppose the instantaneous volatility of
𝐹𝑘 (𝑡) is modelled as 𝜎𝑘 (𝑡). Having done so, one now has the implied volatility of the 𝑇𝑘−1 × 𝑇𝑘
caplet given by
√
∫ 𝑇𝑘−1
1
𝜎𝑘,imp = 𝜎𝑘 (𝑡)2 𝑑𝑡 (3.23)
𝑇𝑘−1 0
Some forms will allow us to calibrate to given cap or caplet term structures exactly. Clearly,
and as emphasised in [Brigo and Mercurio, 2006, §6.2], the pricing of caps is independent of the
joint dynamics of forward rates. However, that does not mean that calibration should also be an
independent process. There are such straightforward formulations of the calibration of caps, so
then the only parameters left to tackle swaptions calibration are the instantaneous correlations of
forward rates, and this will typically be inadequate for use in the LMM.
[Brigo and Mercurio, 2006, §6.3.1] discuss seven different formulations for calibrating cap term
structures which allow for more or less flexibility later in the swaption calibration.
The approach suggested in [Rebonato, 2002, Chapter 6], [Rebonato, 2004, Chapter 21] is to specify
a parametric form for the instantaneous volatility such as

𝜎𝑘 (𝑡) = (𝑎 + 𝑏(𝑇𝑘−1 − 𝑡))𝑒−𝑐(𝑇𝑘−1 −𝑡) + 𝑑 (3.24)

This form is flexible enough to reproduce the typical shapes that occur in the market. It can
accommodate either a humped form or a monotonically decreasing volatility. The model is time
homogeneous, and the parameters have some economic interpretation, as described in Rebonato
[2002]. For example, the time-0 volatility is 𝑎 + 𝑑 and the long run limit is 𝑑. Furthermore, within
the context of models such as LMM, calculus is easy enough: for example
∫ ( )( )
(𝑎 + 𝑏(𝑇𝑖 − 𝑡))𝑒−𝑐(𝑇𝑖 −𝑡) + 𝑑 (𝑎 + 𝑏(𝑇𝑗 − 𝑡))𝑒−𝑐(𝑇𝑗 −𝑡) + 𝑑 𝑑𝑡
𝑎𝑑 ( 𝑐(𝑡−𝑇𝑖 ) ) 𝑏𝑑 ( )
= 𝑒 + 𝑒𝑐(𝑡−𝑇𝑗 ) + 𝑑2 𝑡 − 2 𝑒𝑐(𝑡−𝑇𝑖 ) [𝑐(𝑡 − 𝑇𝑖 ) − 1] + 𝑒𝑐(𝑡−𝑇𝑗 ) [𝑐(𝑡 − 𝑇𝑗 ) − 1]
𝑐 𝑐
𝑐(2𝑡−𝑇𝑖 −𝑇𝑗 ) [
𝑒
2𝑎2 𝑐2 + 2𝑎𝑏𝑐(1 + 𝑐(𝑇𝑖 + 𝑇𝑗 − 2𝑡)) + 𝑏2 (1 + 2𝑐2 (𝑡 − 𝑇𝑖 )(𝑡 − 𝑇𝑗 ) + 𝑐(𝑇𝑖 + 𝑇𝑗 − 2𝑡))
]
+ 3
4𝑐
=: 𝐼(𝑡, 𝑇𝑖 , 𝑇𝑗 ) (3.25)

as in [Rebonato, 2002, §6.6 - correcting for the typo], [Jäckel, 2002, (12.13)]. Then as in (3.23)
√
1
𝜎𝑘,imp = (𝐼(𝑇𝑘−1 , 𝑇𝑘−1 , 𝑇𝑘−1 ) − 𝐼(0, 𝑇𝑘−1 , 𝑇𝑘−1 )) (3.26)
𝑇𝑘−1

Now the problem has gone from being under-specified to over-specified; an error minimisation
algorithm will be used. Financial constraints are that 𝑎 + 𝑑 > 0, 𝑐 > 0, 𝑑 > 0. What we do here is,
for any choices of 𝑎, 𝑏, 𝑐 and 𝑑,

35
 • determine the caplet volatilities for every caplet using (3.26).

• find the model price of all the caplets using the caplet pricing formula.

• find the model price of the caps that are trading in the market (for example, the 1y, 2y, ...
caps).

• We can then formulate an error function which measures the difference between model and
market prices. This function will be something like
∑
err𝑎,𝑏,𝑐,𝑑 = ∣𝑉𝑖 (𝑎, 𝑏, 𝑐, 𝑑) − 𝑃𝑖 ∣ (3.27)
𝑖

The 𝑖 varies only over those caps that actually trade (are quoted) in the market.

• Minimise the error. Solver might be used, but it might need to be trained to find a reasonable
solution. Use of Nelder-Mead is suggested.

Some care needs to be taken here. The inputs will be at the money cap volatilities. For each cap,
the at the money level (the forward swap rate) will probably be different. These are the forwards
that need to be used in the cap pricing formula. The output is a parametric form for at the money
caplet volatilities, and for each of these the at the money level will be different. This difference is
usually ignored, as we are pricing without any skew anyway.
Having found the parameters, one still wants to price instruments that trade in the market exactly.
Thus, after the parameters 𝑎, 𝑏, 𝑐, 𝑑 have been found, the model is re-specified as
[ ]
𝜎𝑘 (𝑡) = 𝐾𝑘 (𝑎 + 𝑏(𝑇𝑘−1 − 𝑡))𝑒−𝑐(𝑇𝑘−1 −𝑡) + 𝑑 (3.28)

Equivalently,
√
1
𝜎𝑘,imp = 𝐾𝑘 (𝐼(𝑇𝑘−1 , 𝑇𝑘−1 , 𝑇𝑘−1 ) − 𝐼(0, 𝑇𝑘−1 , 𝑇𝑘−1 )) (3.29)
𝑇𝑘−1

We assume that 𝐾𝑘 is a piecewise constant function, changing only at the end of each cap i.e. as
a cap terminates and a new calibrating cap is applied. For example, if we have a 1y and a 2y cap
(and others of later tenor) then 𝐾2 = 𝐾3 = 𝐾4 , and 𝐾5 = 𝐾6 = 𝐾7 = 𝐾8 .
With these assumptions, 𝐾𝑘 is found uniquely. For caplets, the value of 𝐾𝑘 is found directly. For
caps, we note that there is one root find for each set of equal 𝐾𝑘 ’s; we proceed from smallest to
largest 𝑘. Thus

• For any given 𝐾𝑘 , calculate the volatility using (3.29).

• price all the caplets using the caplet pricing formula.

• find the model price of the cap.

• vary 𝐾𝑘 to match this model price with the market price. As the model price is an increasing
function of 𝐾𝑘 , the root is unique. We use a root finder such as Brent’s method.

The model is no longer time homogeneous, and the deviation from being so is in some sense measured
by how far the 𝐾𝑘 deviate from 1. The better the fit of the model, the closer these values are to 1,
one would hope to always have values of 𝐾𝑘 between 0.9 and 1.1 say. This correction is discussed
in [Rebonato, 2004, §21.4].

36
3.5 Swaptions
A swaption is an option to enter into a swap. We consider European swaptions. (Bermudan
swaptions also exist.) Thus, at a specified time 𝑡0 , the holder of the option has the option to enter
a swap which commences then (the first payment being one time period later, at 𝑡1 , and lasts until
time 𝑡𝑛 ).
Of course, we have two possibilities

(a) a payer swaption, which gives the holder the right but not the obligation to receive floating,
and pay a fixed rate 𝑟𝐾 (a call on the floating rate).

(b) a receiver swaption, which gives the holder the right but not the obligation to receive a fixed
rate 𝑟𝐾 , and pay floating (a put on the floating rate).

Let 𝑓 be the fair (par) forward swap rate for the period from 𝑡0 to 𝑡𝑛 . The date schedule for swaptions
is the swap schedule. The time of payments of the forward starting swap are 𝑡1 , 𝑡2 , . . . , 𝑡𝑛 , where
𝑡0 , 𝑡1 , . . . , 𝑡𝑛 are successive observation days, for example, quarterly, calculated according to the
relevant day count convention and modified following rules. As usual, let 𝑡𝑖 − 𝑡𝑖−1 = 𝛼𝑖 , measured
in years, for 𝑖 = 1, 2, . . . , 𝑛.
Note that if 𝑛 = 1 then we have a one period cap (a payer swaption) or a one period floor (a receiver
swaption). Thus, modulo the date schedule and the advanced/arrears issue, a caplet or floorlet.
Note that in general a swap (forward starting or starting immediately; in the later case 𝑡0 = 0) with
a fixed rate of 𝑅 has the fixed leg payments worth
𝑛
∑
𝑉fix = 𝑅 𝛼𝑖 𝑍(0, 𝑡𝑖 ) (3.30)
𝑖=1

while the floating payments are worth

𝑉float = 𝑍(0, 𝑡0 ) − 𝑍(0, 𝑡𝑛 ) (3.31)

Hence the fair forward swap rate, which equates the fixed and floating leg values, is given by

𝑍(0, 𝑡0 ) − 𝑍(0, 𝑡𝑛 )
𝑓 = ∑𝑛 (3.32)
𝑖=1 𝛼𝑖 𝑍(0, 𝑡𝑖 )

Of course, these values are derived from the existing swap curve. Thus, the fair forward swap rate
is dependent upon the bootstrap and interpolation method associated with the construction of the
yield curve. Nevertheless, empirically it is found that the choice of interpolation method will only
affect the result to less than a basis point, and typically a lot less. Also, let
𝑛
∑
𝐿= 𝛼𝑖 𝑍(0, 𝑡𝑖 ) (3.33)
𝑖=1

𝐿 is called the level, or the annuity.

37
Figure 3.3: A Reuter’s page for at the money volatility quotes for caps/floors, caplet/floorlets, and
swaptions. In the swaption table 3mnth, 6mnth, 1year, 2year refers to the expiry date of the option;
1yr, 2yr, 3yr, 5yr refers to the tenor of the swap.

3.5.1 Valuation
The value of the swaption per unit of nominal is [Hull, 2005, §26.4]

𝑉𝜂 = 𝐿𝜂[𝑓 𝑁 (𝜂𝑑1 ) − 𝑟𝐾 𝑁 (𝜂𝑑2 )] (3.34)

𝑓 1 2
ln 𝑟𝐾 ± 2 𝜎 𝑡0
𝑑1,2 = √ (3.35)
𝜎 𝑡0

where 𝜂 = 1 stands for a payer swaption, 𝜂 = −1 for a receiver swaption. Here 𝜎 is the volatility of
the fair forward swap rate, and is an implied variable quoted in the market.

3.5.2 Greeks
Delta

Δ = 𝜂𝑁 (𝜂𝑑1 ) (3.36)

pv01
This is given by
[ ]
∂𝑉 ∂𝐿 ∂𝑓 ′ ∂𝑑1 ′ ∂𝑑2
=𝜂 [𝑓 𝑁 (𝜂𝑑1 ) − 𝑟𝐾 𝑁 (𝜂𝑑2 )] + 𝜂𝐿 𝑁 (𝜂𝑑1 ) + 𝑓 𝑁 (𝜂𝑑1 )𝜂 − 𝑟𝐾 𝑁 (𝜂𝑑2 )𝜂
∂𝑟 ∂𝑟 ∂𝑟 ∂𝑟 ∂𝑟
∂𝐿 ∂𝑓
=𝜂 [𝑓 𝑁 (𝜂𝑑1 ) − 𝑟𝐾 𝑁 (𝜂𝑑2 )] + 𝜂𝐿 𝑁 (𝜂𝑑1 ) (3.37)
∂𝑟 ∂𝑟

38
We now apply several times the fact that
∂
𝑍(0, 𝑡) = −𝑡𝑍(0, 𝑡) (3.38)
∂𝑟
Firstly
𝑛
∂𝐿 ∑
= −𝑡𝑖 𝛼𝑖 𝑍(0, 𝑡𝑖 ) (3.39)
∂𝑟 𝑖=1

(3.40)
𝑔(𝑟,𝜏 )
Also, if we have a function ℎ(𝑟,𝜏 ) , then
( 𝑔 )′
= 𝑔 ′ ℎ−1 − 𝑔ℎ−2 ℎ′ (3.41)
ℎ
(3.42)

where the differentiation is with respect to 𝑟. We apply (3.38) and this to (3.32).
1 ∂𝑉
Then pv01 = 10000 ∂𝑟 .
Vega

[ ]
∂𝑉 ′ ∂𝑑1 ′ ∂𝑑2
= 𝐿𝜂 𝑓 𝑁 (𝜂𝑑1 )𝜂 − 𝑟𝐾 𝑁 (𝜂𝑑2 )𝜂
∂𝜎 ∂𝜎 ∂𝜎
′
√
= 𝐿𝑓 𝑁 (𝑑1 ) 𝜏 (3.43)

Theta
As before.

3.5.3 Why Black is useless for exotics

In each Black model for maturity date 𝑇 , the forward rates will be log-normally distributed in what
is called the 𝑇 -forward measure, where 𝑇 is the pay date of the option. The fact that we can ‘legally’
discount the expected payoff under this measure using today’s yield curve is a consequence of some
profound academic work of Geman et al. [1995], which establishes the existence of alternative pricing
measures, and the ways that they are related to each other. This paper is very significant in the
development of the Libor Market Model.
We can use Black’s model without knowing any of the theory of Geman et al. [1995]; however, the
Black model cannot safely be used to value more complicated products where the payoff depends
on observations at multiple dates. For this an alternative model which links the behaviour of the
rates at multiple dates will need to be used.
For this, the most extensive approach is the Libor Market Model. Here inputs are all the Black
models as well as a correlation structure between all the forward rates. From a properly calibrated
LMM, one recaptures (up to the calibration error) the prices of traded caplets, caps, and swaptions.
However, this calibration can be difficult. It then requires Monte Carlo techniques to value other
derivatives.
An intermediate approach is the use of a more parsimonious model with just a few driving factors
- the so-called single-factor models.

39
3.6 Exercises
1. A company caps three-month JIBAR at 9% per annum. The principal amount is R10 million.
On a reset date, namely 20 September 2004, three-month JIBAR is 10% per annum.

(a) What payment would this lead to under the cap?

(b) When would the payment be made?
(c) What is the value of the payment on the reset date?

2. Use Black’s model to value a one-year European put option on a bond with 9 years and 11
months to expiry. Assume that the current cash price of the bond is R105, the strike price
(clean) is R110, the one-year interest rate is 10%, the bond’s price volatility measure is 8%
per annum, and the coupon rate is 8% NACS.

3. Consider an eight-month European put option on a Treasury bond that currently has 14.25
years to maturity. The current cash bond price is $910, the exercise price is $900, and the
volatility measure for the bond price is 10% per annum. A semi-annual coupon of $35 will be
paid by the bond in three months. The risk-free interest rate is 8% for all maturities up to
one year.

(a) Use Black’s model to determine the price of the option. Consider both the case where the
strike price corresponds to the cash price of the bond and the case where it corresponds
to the clean price.
(b) Calculate delta, gamma (both with respect to the bond price 𝐵) and vega in the above
problem when the strike price corresponds to the quoted price. Explain how they can be
interpreted.

4. Using Black’s model calculate the price of a caplet on the JIBAR rate. Today is 20 September
2004 and the caplet is the one that corresponds to the 9x12 period.
The caplet is struck at 9.4%. The current JIBAR rate is 9.3%, the 3x6 FRA is 9.4% and the
6x9 FRA is 9.34%, and the 9x12 FRA is 9.20%.
Interest-rate volatility is 15%.
Also calculate delta, gamma (both with respect to the 9x12 forward rate) and vega in the
above problem.

5. Suppose that the yield, 𝑅, on a discount bond follows the process

𝑑𝑅 = 𝜇(𝑅, 𝑡)𝑑𝑡 + 𝜎(𝑅, 𝑡)𝑑𝑧

where 𝑑𝑧 is a standard Wiener process under some measure. Use Ito’s Lemma to show that
the volatility of the discount bond price declines to zero as it approaches maturity, irrespective
of the level of interest rates.

6. The price of a bond at time 𝑇, measured in terms of its yield, is 𝐺 (𝑦𝑇 ) . Assume geometric
Brownian motion for the forward bond yield, 𝑦, in a world that is forward risk-neutral with
respect to a bond maturing at time 𝑇. Suppose that the growth rate of the forward bond yield
is 𝛼 and its volatility is 𝜎𝑦 .

40
 (a) Use Ito’s Lemma to calculate the process for the forward bond price, in terms of 𝛼, 𝜎𝑦 ,
𝑦 and 𝐺 (𝑦) .
(b) The forward bond price should follow a martingale in the world we are considering. Use
this fact to calculate an expression for 𝛼.
(c) Assume an initial value of 𝑦 = 𝑦0 . Now show that the expected value of 𝑦 at time 𝑇 can
be directly calculated from the above expression.

7. Consider the following quarterly data:

1 2 3 4 5 6 7 8
zero 8.00% 8.50% 8.30% 8.00% 8.00% 8.20% 8.25% 8.40%
vol 17.00% 16.00% 17.00% 16.00% 15.00% 15.00% 14.00%
Time here is modified following quarterly.
The volatility data represents the volatility of the all in price for a bond option on the r153
that expires at that time.
The details of the r153 are as follows:
Bond Name Maturity Coupon BCD1 BCD2 CD1 CD2
R153 2010/08/31 13.00% 821 218 831 228
(The BCD details of all r bonds have changed.)
Today is 29 August 2004 and the all in price is 1.1010101.
Price a vanilla European bond option, with a clean price strike, with expiry 30 May 2005,
according to Black’s model.
Construct your yield curve using raw interpolation.

41
Chapter 4

One and two factor

continuous-time interest rate
models

This chapter is derived from components of [Wilmott, 2000, Chapters 40, 41], Svoboda [2002],
Svoboda [2003]; amongst others.
The fundamental complicating factor in interest rate models is the non-traded nature of rates.
Coupled to this is the non-linear, inverse relationship between bond prices and yields. If a derivative
is dependent on one or more interest rates, the rather neat consequences of the Black-Scholes model
for equity derivatives, where the expected rate of return 𝜇 drops out of the pricing formula and is
‘replaced’ by risk-neutral valuation and a return of 𝑟, will almost certainly not be valid. What is
more, in that Black-Scholes differential equation it was exactly that 𝑟 which was a constant, it is
now a variable.
We develop models of the short rate. The short rate will be denoted 𝑟. The short rate itself is
quite a theoretical concept: at time 𝑡, 𝑟(𝑡) is the yield on a bond which matures at time 𝑡 + 𝑑𝑡. In
practice, a rate that truly exists, such as the overnight, one month or even three month rate, will
be used as a surrogate for this rate.
We will now have a new variable, known as the market price of risk; knowing the market price of
risk is equivalent to knowing the expected rate of return.

4.1 Derivatives Modelled on a Single Stochastic Variable

Assume that the short rate can be modelled as an Itô process of the following kind:

𝑑𝑟 = 𝜇(𝑟, 𝑡)𝑑𝑡 + 𝜎(𝑟, 𝑡)𝑑𝑧 (4.1)

where 𝑑𝑧 has the usual properties. (4.1) is sufficiently general for a broad spectrum of possible
models.

42
Recall that if 𝑉 = 𝑉 (𝑟, 𝑡) is a sufficiently well behaved function then Itô’s lemma tells us
2
( )
∂𝑉 ∂𝑉 1 2∂ 𝑉
𝑑𝑉 = 𝑑𝑟 + + 2𝜎 𝑑𝑡
∂𝑟 ∂𝑡 ∂𝑟2
∂2𝑉
( )
∂𝑉 ∂𝑉 ∂𝑉
= + 𝜇 + 12 𝜎 2 2 𝑑𝑡 + 𝜎 𝑑𝑧
∂𝑡 ∂𝑟 ∂𝑟 ∂𝑟
:= 𝜇𝑉 𝑉 𝑑𝑡 + 𝜎𝑉 𝑉 𝑑𝑧

Assume that 𝑉1 and 𝑉2 are sufficiently well behaved. The underlying stochastic process 𝑟 is non-
traded, so the only possible way of creating a riskless portfolio is by using both 𝑉1 and 𝑉2 . Let Π
be the portfolio which is long 1 of instrument 1 and is short 𝜎𝜎12 𝑉𝑉12 of instrument 2. Then, the risky
component (the part with 𝑑𝑧) has been eliminated, so as usual
( )
𝜎1 𝑉1 𝜎1 𝑉 1
𝜇1 𝑉1 − 𝜇2 𝑉2 = 𝑟 𝑉1 − 𝑉2
𝜎2 𝑉2 𝜎2 𝑉 2
( )
𝜎1 𝜎1
𝜇1 − 𝜇2 =𝑟 1−
𝜎2 𝜎2
𝜇1 − 𝑟 𝜇2 − 𝑟
= (4.2)
𝜎1 𝜎2
and so the quantity
𝜇𝑉 − 𝑟
𝜆 := (4.3)
𝜎𝑉
is independent of the choice of bond, and called the market price of risk. Now

𝜇𝑉 𝑉 − 𝜆𝜎𝑉 𝑉 = 𝑟𝑉

and so
∂𝑉 ∂𝑉 ∂2𝑉 ∂𝑉
+ 𝜇 + 12 𝜎 2 2 − 𝜆 𝜎 = 𝑟𝑉
∂𝑡 ∂𝑟 ∂𝑟 ∂𝑟
∂𝑉 ∂𝑉 ∂2𝑉
+ (𝜇 − 𝜆𝜎) + 21 𝜎 2 2 − 𝑟𝑉 = 0 (4.4)
∂𝑡 ∂𝑟 ∂𝑟
Note well - the drift, volatility here are those of (4.1), and not of (4.2).
This is the Black-Scholes type equation for interest rate derivatives when the yield curve is modelled
by a one-factor process of the type (4.1). Note that a vanilla bond is a derivative; options on bonds
are derivatives of derivatives. The underlying here is the short rate.
It is possible to apply (4.4) to bonds that have coupons by adding 𝐾(𝑟, 𝑡) to the left-hand side, where
𝐾(𝑟, 𝑡) 𝑑𝑡 represents the amount of coupon received in the period 𝑑𝑡. (This may be continuous or
discrete; in the latter case we will be using a finite difference scheme with specified jumps in value
to the bond as it goes ex coupon.)
We require two “spatial” boundary conditions and one final boundary condition to fully specify the
model. For example, we know that if 𝑉 (𝑟, 𝑡, 𝑇 ) is the price of a bond with par value 1, 𝑉 (𝑟, 𝑇, 𝑇 ) = 1.
Other examples would be the terminal intrinsic value of a vanilla option, etc.
Using the Feynman-Kac theorem (4.4) is the solution to the expectation of the 𝑟-discounted value
of 𝑉 (𝑇 ) where 𝑟 is subject to the Itô process

𝑑𝑟 = (𝜇 − 𝜆𝜎)𝑑𝑡 + 𝜎𝑑𝑧 (4.5)

43
The Brownian motion here is not the same as the original one: 𝑑𝑧 = 𝜆𝑑𝑡 + 𝑑𝑧. The drift here is
𝜇 − 𝜆𝜎; the original (real world) drift was of course 𝜇.

4.2 What is the market price of risk?

The market price of risk is the return in excess of the risk-free rate that the market wants as
compensation for taking risk.
In classical economic theory no rational person would invest in a risky asset unless they expect to
beat the return from holding a risk-free asset. Classically risk is measured by volatility. The market
price of risk thus has the form of (4.3). This quantity is not affected by leverage.
If the modelled quantity is directly traded, then one can continuously and perfectly hedge a position,
and in so doing eliminate risk totally. Thus the market price of risk is a quantity that does not need
to appear in the model (or the differential equation that describes it).
If the modelled quantity is not directly traded, then there will be an explicit reference in the option-
pricing model to the market price of risk. This is because associated risk cannot be hedged, and so
how much extra return is needed to compensate for taking this unhedgeable risk must be modelled.
When a quantity that is not traded is modelled stochastically then the equation governing the
pricing of derivatives is usually of diffusion form, with the market price of risk appearing in the drift
term with respect to the non-traded quantity.
An example is the interest rate market. Since interest rates are not tradable, you will not be able to
create a risk-free portfolio. In fact you will trade one or more assets that depends on that quantity,
rather than the quantity itself (e.g. you want to take a position on an interest rate, you might want
to trade bonds since their value depends on interest rates). The same happens for hedging, since
you will need again to use another instrument similar to the original one to cover your position, but
is not the underlying quantity.
The result is that there is never have a portfolio that completely eliminates risk, and so an agent
will require a premium to balance his diminished utility function resulting from taking risk.
Another example is a stochastic volatility model in the equity market, such as Heston. Here the
volatility is a risk that is not a traded asset. However, the market can be completed by trading in
any vanilla option. This was observed in Hagan et al. [2002], and formalised in Davis and Ob̷lój
[2007]. It can also be completed by trading in a variance swap.

4.3 Exogenous (equilibrium) short rate models

These early equilibrium models are based on a mathematical model of the economy. They focus on
describing and explaining the interest rate term structure. However, a fundamental problem with
the equilibrium approach is that the models may not be arbitrage free, in other words, they fail to
price even the vanilla inputs trading in the market. This is to be expected: these instruments are
not inputs to the models, they are outputs. As a consequence it is unlikely that these models will
be used nowadays.

44
4.3.1 GBM model
Dothan [1978], Rendelman and Bartter [1980]
This is simply
𝑑𝑟 = 𝜇𝑟𝑑𝑡 + 𝜎𝑟𝑑𝑧 (4.6)
which is the same as the usual log-normal stock price model. However, an obvious problem that
arises is that of mean-reversion: interest rates appear to be pulled back to some long-run average
level over time.

4.4 A particular class of models

(4.4) is the bond pricing equation for an arbitrary model. In addition there will be the boundary
conditions as already mentioned. In order to progress, we want to specify the risk neutral drift
𝜇 − 𝜆𝜎 and the volatility 𝜎 to arrive at tractable models i.e. models where (at the very least) the
price of zero coupon bonds can be found analytically.1 In other words, we will attempt to solve
prices under the process (4.5).
Time will be denoted by the variable 𝑡; it starts at time 0 and runs to terminal time 𝑇 . Calibration
takes place at time 0; 𝑇 is the maturity of the longest dated bond under consideration (or even
further, for that matter). 𝑍(𝑟, 𝑡, 𝑇 ) is the value at time 𝑡 of a zero coupon bond maturing at time
𝑇 . The only stochastic variable is 𝑟 = 𝑟(𝑡) i.e. the yield curve at time 𝑡 is determined entirely by
the value of the short rate 𝑟(𝑡).
The class of models we consider is as follows: the zero-coupon bond value has dependency on the
one-factor short rate via:
𝑍(𝑟, 𝑡, 𝑇 ) = exp(𝐴(𝑡, 𝑇 ) − 𝑟𝐵(𝑡, 𝑇 )) (4.7)
The 𝐴 and 𝐵 functions are functions of 𝑡; they are determined at time 0. The class of models that
result from this assumption are referred to as affine. The different affine models will have different
formulae for 𝐴(𝑡, 𝑇 ) and 𝐵(𝑡, 𝑇 ). The continuous yield curve at time 𝑡 is determined as follows:
𝐴(𝑡, 𝑇 ) − 𝑟𝐵(𝑡, 𝑇 )
r(𝑟, 𝑡, 𝑇 ) = − (4.8)
𝑇 −𝑡
So, if the short rate 𝑟 varies instantaneously, the entire yield curve r(𝑟, 𝑡, 𝑇 ) varies.
Note that
( )
∂𝑍 ∂𝐴 ∂𝐵
=𝑍 −𝑟
∂𝑡 ∂𝑡 ∂𝑡
∂𝑍
= −𝐵(𝑡, 𝑇 )𝑍
∂𝑟
∂2𝑍
= 𝐵(𝑡, 𝑇 )2 𝑍
∂𝑟2
The bond described in (4.7) is a derivative of the short rate, so it satisfies (4.4). Dividing by 𝑍 we
get
∂𝐴 ∂𝐵
−𝑟 + 12 𝜎 2 𝐵 2 − (𝜇 − 𝜆𝜎)𝐵 − 𝑟 = 0. (4.9)
∂𝑡 ∂𝑡
1 This will give us a check on the model, by comparing these results to our bootstrap, and in fact, might turn the

whole problem inside-out, and become the calibration procedure.

45
If we differentiate this equation twice with respect to 𝑟, and divide through by 𝐵, we arrive at

1 ∂2 2 ∂2
2 ∂𝑟 2 𝜎
𝐵 − (𝜇 − 𝜆𝜎) = 0.
∂𝑟2
In this the expiry date 𝑇 is arbitrary, so the coefficients must be separately zero:
∂2 2
𝜎 =0
∂𝑟2
∂2
(𝜇 − 𝜆𝜎) = 0
∂𝑟2
and so we can find functions 𝛼(𝑡), 𝛽(𝑡), 𝜃(𝑡) and 𝛾(𝑡) which are function of 𝑡 alone (and not of 𝑟)
such that

𝜎 2 (𝑟, 𝑡) = 𝛼(𝑡)𝑟 + 𝛽(𝑡) (4.10)

𝜇(𝑟, 𝑡) − 𝜆(𝑟, 𝑡)𝜎(𝑟, 𝑡) = 𝜃(𝑡) − 𝛾(𝑡)𝑟 (4.11)

Substituting (4.10) and (4.11) into (4.9) we have

∂𝐴 ∂𝐵
−𝑟 + 21 (𝛼(𝑡)𝑟 + 𝛽(𝑡))𝐵 2 − (𝜃(𝑡) − 𝛾(𝑡)𝑟)𝐵 − 𝑟 = 0
∂𝑡 ∂𝑡
and so
∂𝐴
= 𝜃(𝑡)𝐵 − 21 𝛽(𝑡)𝐵 2 (4.12)
∂𝑡
∂𝐵
= −1 + 𝛾(𝑡)𝐵 + 21 𝛼(𝑡)𝐵 2 (4.13)
∂𝑡
(4.13) is what is known as a Ricatti equation, it is an equation in 𝐵 which does not involve 𝐴.
Having solved (4.13) we then insert this solution into (4.12) and do straightforward integration to
obtain 𝐴. In general the system will be solved numerically; however, sometimes it is possible to find
closed form solutions, and we will focus on these cases.
The boundary conditions must be:

𝐴(𝑇, 𝑇 ) = 0
𝐵(𝑇, 𝑇 ) = 0

The process in (4.5) has become

√
𝑑𝑟 = (𝜃(𝑡) − 𝛾(𝑡)𝑟)𝑑𝑡 + 𝛼(𝑡)𝑟 + 𝛽(𝑡)𝑑𝑧 (4.14)

4.4.1 Constant parameter model

Let us start with the simplest case: all of 𝛼, 𝛽, 𝜃, 𝛾 are constants. Then
𝑑𝐵
= −1 + 𝛾𝐵 + 21 𝛼𝐵 2
𝑑𝑡
and so (after some unpleasant calculus)
2(exp(𝜓1 (𝑇 − 𝑡)) − 1)
𝐵(𝑡, 𝑇 ) =
(𝛾 + 𝜓1 )(exp(𝜓1 (𝑇 − 𝑡)) − 1) + 2𝜓1
√
𝜓1 = 𝛾 2 + 2𝛼

46
Now, we can write
𝑑𝐴 𝜃𝐵 − 21 𝛽𝐵 2
= 1 2
2 𝛼𝐵 + 𝛾𝐵 − 1
𝑑𝐵
and so

1 𝐵+𝑏 1
2 𝛼𝐴 = 𝑎𝜓2 ln(𝑎 − 𝐵) + (𝜓2 + 12 𝛽)𝑏 ln − 2 𝛽𝐵 − 𝑎𝜓2 ln 𝑎
𝑏
𝜃 − 12 𝑎𝛽
𝜓2 =
𝑎+𝑏
√
±𝛾 + 𝛾 2 + 2𝛼
𝑏, 𝑎 =
𝛼
Note that here we have incorporated the final conditions.
Obviously 𝐴 and 𝐵 are functions of 𝜏 = 𝑇 − 𝑡, and not functions of 𝑡 and 𝑇 separately.

4.4.2 Vasicek model

Vasicek [1977] models the short rate as an Ornstein-Uhlenbeck process; the mean-reversion property
is modelled. All four parameters are constant, with 𝛼 = 0, and (of necessity) 𝛽 > 0. Thus the
√
volatility is time and state independent and equal to 𝛽. So
√
𝑑𝑟 = (𝜃 − 𝛾𝑟)𝑑𝑡 + 𝛽𝑑𝑧
( )
𝜃 √
=𝛾 − 𝑟 𝑑𝑡 + 𝛽𝑑𝑧
𝛾
:= 𝛾 (𝜇 − 𝑟) 𝑑𝑡 + 𝜎𝑑𝑧 (4.15)

The short rate mean reverts to 𝜇 at a ‘speed’ of 𝛾. This model is very tractable, and there are
explicit solutions for a number of derivatives based on it. Unfortunately, the Vasicek model permits
negative interest rates.
Now (4.12) and (4.13) become

𝑑𝐴
= 𝜃𝐵 − 21 𝛽𝐵 2
𝑑𝑡
𝑑𝐵
= −1 + 𝛾𝐵
𝑑𝑡
(there is no dependency on 𝑟 in the parameters - 𝑡 is the only variable in play, and partial differen-
tiation is the same as total differentiation) and so (again, after some work)
1
𝐵(𝑡, 𝑇 ) = (1 − exp(−𝛾(𝑇 − 𝑡)))
𝛾
( 1 𝛽 − 𝜃𝛾)((𝑇 − 𝑡) − 𝐵) 𝛽𝐵 2
𝐴(𝑡, 𝑇 ) = 2 −
𝛾2 4𝛾
It is possible to generate increasing, decreasing or humped curves under this model. See Figure 4.1.
We can solve for the distribution of the short rate. First, put 𝑓 (𝑟, 𝑡) = 𝑟𝑒𝛾𝑡 - here we are using an

47
 Figure 4.1: Possible yield curves under the Vasicek model, varying only 𝜃.

integrating factor. Using Itô’s lemma, we have

𝑑𝑓 = 𝑒𝛾𝑡 𝑑𝑟 + 𝛾𝑒𝛾𝑡 𝑟 𝑑𝑡
= 𝑒𝛾𝑡 [𝛾 (𝜇 − 𝑟) 𝑑𝑡 + 𝜎𝑑𝑧] + 𝛾𝑒𝛾𝑡 𝑟 𝑑𝑡
= 𝑒𝛾𝑡 𝛾𝜇𝑑𝑡 + 𝜎𝑒𝛾𝑡 𝑑𝑡
∫ 𝑡 ∫ 𝑡
𝑓 (𝑡) − 𝑓 (0) = 𝛾𝜇 𝑒𝛾𝑠 𝑑𝑠 + 𝜎 𝑒𝛾𝑠 𝑑𝑧(𝑠)
0 0
∫ 𝑡
𝑟𝑡 𝑒𝛾𝑡 − 𝑟0 = 𝜇 𝑒𝛾𝑡 − 1 + 𝜎 𝑒𝛾𝑠 𝑑𝑧(𝑠)
[ ]
0
∫ 𝑡
−𝛾𝑡
+ 𝜇 1 − 𝑒𝛾𝑡 + 𝜎 𝑒𝛾(𝑠−𝑡) 𝑑𝑧(𝑠)
[ ]
𝑟𝑡 = 𝑟0 𝑒
0

and so
𝜎2 (
( )
𝑟𝑡 ∼ 𝜙 𝜇 + 𝑒−𝛾𝑡 (𝑟0 − 𝜇), 1 − 𝑒−2𝛾𝑡
)
(4.16)
2𝛾
In particular, it is possible to obtain negative rates under this model.
Under the Vasicek model we can value European options on default-free bonds exactly Jamshidian
[1989]. Consider a call option on an 𝑇2 −maturity discount bond with exercise price 𝐾 and expiration
𝑇1 < 𝑇2 . Using risk-neutral valuation, the vanilla option price is:

𝑉± [𝑟, 𝑡, 𝑇1 , 𝑇2 ; 𝐾] = 𝑍(𝑟, 𝑡, 𝑇1 )𝔼ℚ

𝑡 [max(0, ±(𝑍 − 𝐾))]

where 𝑍 is the zero coupon price for expiry 𝑇2 as observed at time 𝑇1 . It has current forward value
𝑍(𝑟, 𝑡, 𝑇2 )
𝑓 (𝑍) = .
𝑍(𝑟, 𝑡, 𝑇1 )

48
It turns out that we obtain a ‘Black-like’ option pricing formula:

𝑐[𝑟, 𝑡, 𝑇1 , 𝑇2 ; 𝐾] = 𝑍(𝑟, 𝑡, 𝑇2 )𝑁 (𝑑+ ) − 𝐾𝑍(𝑟, 𝑡, 𝑇1 )𝑁 (𝑑− ) (4.17)

𝑝[𝑟, 𝑡, 𝑇1 , 𝑇2 ; 𝐾] = 𝐾𝑍(𝑟, 𝑡, 𝑇1 )𝑁 (−𝑑− ) − 𝑍(𝑟, 𝑡, 𝑇2 )𝑁 (−𝑑+ ) (4.18)
√
√ 1 − exp(−2𝛾(𝑇1 − 𝑡))
𝜎𝑝 = 𝛽 𝐵(𝑇1 , 𝑇2 ) (4.19)
2𝛾
ln 𝑓 (𝑍) 1 2
𝐾 ± 2 𝜎𝑝
𝑑± = (4.20)
𝜎𝑝

Remarkably, the valuation of options on coupon-bearing bonds is also possible. Although the model
was originally posited with the underlying stochastic process being Vasicek’s model, it can be applied
to any derivative where the value of zero-coupon bonds is dependent on the short rate only. The
intuition is that a set of mini-options on each of the coupons and the bond can be constructed which
will have the same value as the total option. Because the movements of the entire yield curve are
perfectly correlated there is currently no additional value to having this decomposition. Thus, the
value of the option on the coupon bearing bond can be decomposed into a portfolio of options on
zero coupon bonds.
This procedure only works if the option is European.
Consider a European call with exercise price 𝐾 and maturity 𝑇 , on a coupon-bearing bond. Suppose
that the bond provides cashflows 𝑐1 , 𝑐2 , . . . , 𝑐𝑛 at times 𝑡1 , 𝑡2 , . . . , 𝑡𝑛 after the maturity 𝑇 of the
option. The payoff of the option at time 𝑇 is
[ 𝑛 ]
∑
max 0, 𝑐𝑖 𝑍(𝑟, 𝑇, 𝑡𝑖 ) − 𝐾 (4.21)
𝑖=1

where 𝑟 is the short rate as observed at time 𝑇 ; of course unknown at time 𝑡. Let 𝑟∗ be the value of
the short rate at time 𝑇 which causes the coupon-bearing bond price to equal the strike price. 𝑟∗
is the solution of
∑𝑛
𝐾= 𝑐𝑖 𝑍(𝑟∗ , 𝑇, 𝑡𝑖 ) (4.22)
𝑖=1

We calculate 𝑟∗ now using Newton’s method. (Just using goalseek in excel will suffice for simple
problems.) Since zero prices are decreasing functions of 𝑟, it follows that the option expires in the
money for all 𝑟 < 𝑟∗ , and out of the money for all 𝑟 > 𝑟∗ . Then
[ 𝑛 ]
∑
payoff = max 0, 𝑐𝑖 𝑍(𝑟, 𝑇, 𝑡𝑖 ) − 𝐾
𝑖=1
[ 𝑛 𝑛
]
∑ ∑
= max 0, 𝑐𝑖 𝑍(𝑟, 𝑇, 𝑡𝑖 ) − 𝑐𝑖 𝑍(𝑟∗ , 𝑇, 𝑡𝑖 )
𝑖=1 𝑖=1
𝑛
∑
= 𝑐𝑖 max[0, 𝑍(𝑟, 𝑇, 𝑡𝑖 ) − 𝑍(𝑟∗ , 𝑇, 𝑡𝑖 )]
𝑖=1

Thus the call on the coupon-bearing bond is reduced to the sum of separate calls on the underlying
zeros. This technique is called the ‘Jamshidian trick’.

49
 √
Example 4.4.1. Suppose that 𝜎 = 𝛽 = 0.02, 𝛾 = 0.1, 𝜃 = 0.01 and the short rate be 10%.
Consider a 3-year European put option on a 5-year bond with a coupon of 10% NACS, strike price
𝐾 = 98, par 100.
At the end of 3 years there will be four cashflows remaining, namely

time amount discount factor

3.5 5 exp(𝐴(3; 3.50) − 𝑟(3)𝐵(3; 3.5))
4 5 exp(𝐴(3; 4) − 𝑟(3)𝐵(3; 4))
4.6 5 exp(𝐴(3; 4.5) − 𝑟(3)𝐵(3; 4.5))
5 105 exp(𝐴(3; 5) − 𝑟(3)𝐵(3; 5))

where the functional form of 𝐴(𝑡, 𝑇 ) and 𝐵(𝑡, 𝑇 ) are known - the only unknown here is 𝑟(3), the
value of the short rate that will actually be observed at time 3.
Solving (4.22) for 𝑟∗ = 𝑟(3), where 𝐾 = 98 gives 𝑟∗ = 0.1095222.

time A B Z Cash Value

0 0 0 1.000000
0.5 -0.001 0.4877 0.946830 5 4.734149
1 -0.005 0.9516 0.896731 5 4.483653
1.5 -0.011 1.3929 0.849538 5 4.247691
2 -0.018 1.8127 0.805091 105 84.534506
98

The option is the sum of four options:

time Cash Strike

3.5 5 4.734149
4 5 4.483653
4.5 5 4.247691
5 105 84.53451

We value each as follows:

time A B Z Cash Strike 𝜎𝑝 𝑑1 𝑑2 𝑉

3.5 -0.052 2.9531 0.706252 5 4.734149 1.465% 0.376364 0.361713 0.012449
4 -0.067 3.2968 0.672465 5 4.483653 2.859% 0.3903 0.361713 0.022830
4.5 -0.083 3.6237 0.640436 5 4.247691 4.184% 0.403556 0.361713 0.031429
5 -0.101 3.9347 0.610074 105 84.53451 5.445% 0.416166 0.361713 0.808417
0.875125

4.4.3 Cox-Ingersoll-Ross model

The Cox et al. [1985] model has all four parameters constant, but this time 𝛽 = 0 and 𝛼 > 0. In
this case (4.4) becomes:
√
𝑑𝑟 = (𝜃 − 𝛾𝑟)𝑑𝑡 + 𝛼𝑟𝑑𝑧 (4.23)

50
The spot rate again has a mean reversion level of 𝛾𝜃 , with reversion occurring at a rate of 𝛾 but,
unlike the Vasicek model, the steady-state distribution is not a special case of the normal distribution
(since the volatility is now rate dependent). The CIR model does not permit negative interest rates,
provided the technical condition 𝜃 > 12 𝛼 holds. CIR give explicit solutions for some interest rate
derivatives but the solutions are complicated and sometimes involve integrals that do not have exact
solutions (and have to be estimated numerically).
For the pricing of discount bonds under the affine model (4.7) the bond pricing equations (4.13)
and (4.12) reduce to:

𝑑𝐵
= 12 𝛼𝐵 2 + 𝛾𝐵 − 1
𝑑𝑡
𝑑𝐴
= 𝜃𝐵
𝑑𝑡
which has the solution:
2(exp(𝜓1 (𝑇 − 𝑡)) − 1)
𝑍(𝑡, 𝑇 ) =
(𝛾 + 𝜓1 )(exp(𝜓1 (𝑇 − 𝑡)) − 1) + 2𝜓1
√
𝜓1 = 𝛾 2 + 2𝛼
2𝜃 2𝜓1 exp((𝛾 + 𝜓1 )(𝑇 − 𝑡)/2)
𝐴(𝑡, 𝑇 ) = ln
𝛼 (𝜓1 + 𝛾)(exp(𝜓1 (𝑇 − 𝑡)) − 1) + 2𝜓1

For specific values of 𝛼, 𝛾 and 𝜃 we can generate yield curves of various shapes. Again, 𝐴 and 𝐵
are functions only of 𝑇 − 𝑡. We can use a similar approach to Jamshidian’s to value European style
options on coupon-bearing bonds.
Consider a call option on an 𝑇2 −maturity discount bond with exercise price 𝐾 and expiration
𝑇1 < 𝑇2 . Using risk-neutral valuation, the vanilla option price is:

𝑉± [𝑟, 𝑡, 𝑇1 , 𝑇2 ; 𝐾] = 𝑍(𝑟, 𝑡; 𝑇1 )𝔼ℚ

𝑡 [max(0, ±(𝑍 − 𝐾))]

where
𝑍(𝑟, 𝑡, 𝑇2 )
𝑓 (𝑍) =
𝑍(𝑟, 𝑡, 𝑇1 )
Then,
𝑉𝑐 [𝑟, 𝑡, 𝑇1 , 𝑇2 ; 𝐾] = 𝑍(𝑟, 𝑡, 𝑇1 ) 𝑓 (𝑍)𝜒2 (ℎ1 ) − 𝐾𝜒2 (ℎ2 )
[ ]

where 𝜒2 is the non-central chi-squared distribution. The parameters ℎ1 and ℎ2 have a complicated
dependence on 𝛼, 𝛾 and 𝜃.

4.5 Two factor equilibrium models

In Brennan and Schwartz [1979], Brennan and Schwartz [1982] the process for the short rate reverts
to a long rate, which is turn follows a stochastic process. The long rate is chosen as the yield on a
perpetual bond.
Longstaff and Schwartz [1992a], Longstaff and Schwartz [1992b] starts with a general equilibrium
model of the economy and derives a term structure model where is there is stochastic volatility.
This model is analytically quite tractable.

51
4.6 Equilibrium models of the logarithm of the short rate

4.6.1 The Continuous time version of the Black-Derman-Toy model

Again, we can use Rebonato [1998] for the derivation of the continuous equivalent of the Black et al.
[1990] model.
In the Black et al. [1990] model, a lognormal distribution of the short term interest rate is assumed
i.e. ln 𝑟(𝑡) is normally distributed. At each node (𝑖, 𝑗 − 1) we have 2 possible states of the world and
interest rates denoted 𝑟(𝑖, 𝑗) and 𝑟(𝑖 + 1, 𝑗) that we can evolve to. The mean short term interest
rate at this time may then be calculated as:
1
ln 𝑟𝑚 (𝑖, 𝑗) = 2 [ln 𝑟(𝑖, 𝑗) + ln 𝑟(𝑖 + 1, 𝑗)]

Also [ √ ]
𝑟(𝑖 + 1, 𝑗) = 𝑟(𝑖, 𝑗) exp 2𝜎(𝑗) Δ𝑡
and hence, we find each of the two possible rates as an offset from the median rate of interest,
𝑟𝑚 (𝑖, 𝑗)
[ √ ]
𝑟(𝑖 + 1, 𝑗) = 𝑟𝑚 (𝑖, 𝑗) exp 𝜎(𝑗) Δ𝑡
[ √ ]
𝑟(𝑖, 𝑗) = 𝑟𝑚 (𝑖, 𝑗) exp −𝜎(𝑗) Δ𝑡

So, the correct continuous analogue of this is as follows

𝑟(𝑡) = 𝑢(𝑡) exp (𝜎(𝑡)𝑧(𝑡)) (4.24)

where

𝑢(𝑡) ∼ time 𝑡 median of the short term interest rate distribution,

𝜎(𝑡) ∼ short term interest rate volatility at time 𝑡,
𝑧(𝑡) ∼ standard Brownian motion.

To examine the nature of the stochastic process driving the short term interest rate we must examine
the evolution of 𝑓 (𝑧, 𝑡) = ln 𝑟(𝑡) = ln(𝑢(𝑡)) + 𝜎(𝑡)𝑧(𝑡). So

𝑑𝑓 = 𝑑 ln 𝑢(𝑡) + 𝑑(𝜎(𝑡)𝑧(𝑡))
∂ ln 𝑢(𝑡)
= 𝑑𝑡 + 𝜎 ′ (𝑡)𝑧(𝑡) 𝑑𝑡 + 𝜎(𝑡) 𝑑𝑧(𝑡)
∂𝑡
𝑢′ (𝑡) ln 𝑟(𝑡) − ln 𝑢(𝑡)
= 𝑑𝑡 + 𝜎 ′ (𝑡) 𝑑𝑡 + 𝜎(𝑡) 𝑑𝑧(𝑡)
𝑢(𝑡) 𝜎(𝑡)
[ ′
𝑢 (𝑡) 𝜎 ′ (𝑡)
]
= + (ln 𝑟(𝑡) − ln 𝑢(𝑡)) 𝑑𝑡 + 𝜎(𝑡) 𝑑𝑧(𝑡)
𝑢(𝑡) 𝜎(𝑡)
𝜎 ′ (𝑡)
= [ln 𝑟(𝑡) − 𝜃(𝑡)] 𝑑𝑡 + 𝜎(𝑡) 𝑑𝑧(𝑡)
𝜎(𝑡)
𝑢′ (𝑡) 𝜎(𝑡)
𝜃(𝑡) := − + ln 𝑢(𝑡)
𝑢(𝑡) 𝜎 ′ (𝑡)

52
See [Cairns, 2004, §5.2.3].
It is tempting to work one-factor models in order to match not only the term structure but also
to match the price of caps (liquid instruments) and swaptions. However, this amounts to over-
parameterisation and leads to a non-stationary volatility structure. The corresponding volatility
term structure implied by the model is unlikely to be anything like the existing volatility structure
and can create derivative mispricing.

4.6.2 The Black-Karasinski Model

Black and Karasinski [1991] is the general version of the above:

𝑑 ln 𝑟(𝑡) = 𝛼(𝑡)(ln 𝜇(𝑡) − ln 𝑟(𝑡))𝑑𝑡 + 𝜎(𝑡)𝑑𝑧(𝑡) (4.25)

See [Cairns, 2004, §5.2.3]. Now let 𝑌 (𝑡) = ln 𝑟(𝑡) in the equation above so that

𝑑𝑌 (𝑡) = 𝛼(𝑡)(ln 𝜇(𝑡) − 𝑌 (𝑡))𝑑𝑡 + 𝜎(𝑡)𝑑𝑧(𝑡).

Then 𝑟 = 𝑒𝑌 so

𝑑𝑟 = 𝑒𝑌 𝑑𝑌 + 21 𝑒𝑌 ⟨𝑑𝑌, 𝑑𝑌 ⟩
= 𝑟(𝑡) [𝛼(𝑡)(ln 𝜇(𝑡) − 𝑌 (𝑡))𝑑𝑡 + 𝜎(𝑡)𝑑𝑧(𝑡)] + 21 𝑟(𝑡)𝜎 2 (𝑡) 𝑑𝑡
= 𝑟(𝑡) 𝛼(𝑡)(ln 𝜇(𝑡) − ln 𝑟(𝑡)) + 21 𝜎 2 (𝑡) 𝑑𝑡 + 𝑟(𝑡)𝜎(𝑡)𝑑𝑧(𝑡)
[ ]

Alternatively, we can solve for the dynamics of 𝑌 . This time the integrating factor is 𝑒𝐴(𝑡) where
∫𝑡
𝐴(𝑡) = 0 𝛼(𝑠) 𝑑𝑠. So let 𝑓 (𝑦, 𝑡) = 𝑦𝑒𝐴(𝑡) .

𝑑𝑓 (𝑡) = 𝑌 (𝑡)𝑒𝐴(𝑡) 𝛼(𝑡) 𝑑𝑡 + 𝑒𝐴(𝑡) 𝑑𝑌

= 𝑌 (𝑡)𝑒𝐴(𝑡) 𝛼(𝑡)𝑑𝑡 + 𝑒𝐴(𝑡) [𝛼(𝑡)(ln 𝜇(𝑡) − 𝑌 (𝑡))𝑑𝑡 + 𝜎(𝑡)𝑑𝑧(𝑡)]
= 𝑒𝐴(𝑡) [𝛼(𝑡) ln 𝜇(𝑡)𝑑𝑡 + 𝜎(𝑡)𝑑𝑧(𝑡)]
[ ∫ 𝑡 ∫ 𝑡 ]
𝑌 (𝑡) = 𝑒−𝐴(𝑡) 𝑌 (0) + 𝑒𝐴(𝑠) 𝛼(𝑠) ln 𝜇(𝑠)𝑑𝑠 + 𝑒𝐴(𝑠) 𝜎(𝑠)𝑑𝑧(𝑠)
0 0
∫ 𝑡 ∫ 𝑡
−𝐴(𝑡) 𝐴(𝑠)−𝐴(𝑡)
𝑌 (𝑡) = 𝑒 𝑌 (0) + 𝑒 𝛼(𝑠) ln 𝜇(𝑠)𝑑𝑠 + 𝑒𝐴(𝑠)−𝐴(𝑡) 𝜎(𝑠)𝑑𝑧(𝑠)
0 0

which shows that

( ∫ 𝑡 ∫ 𝑡 )
−𝐴(𝑡) 𝐴(𝑠)−𝐴(𝑡) 2(𝐴(𝑠)−𝐴(𝑡)) 2
𝑌 (𝑡) ∼ 𝜙 𝑒 𝑌 (0) + 𝑒 𝛼(𝑠) ln 𝜇(𝑠)𝑑𝑠, 𝑒 𝜎 (𝑠) 𝑑𝑠
0 0

This allows for a fairly rich calibration mechanism, using inter alia the caplet volatilities.

4.7 No-arbitrage models

We now consider no-arbitrage models. These will be generalisations of equilibrium models where
the existing term structures are taken as input to the model and parameters are chosen so that
those same term structures are outputs of the model.

53
4.7.1 The Continuous time version of the Ho-Lee model
Rebonato [1998] presents a simple analysis by which the continuous time equivalent of any discrete
time model, modelled within a binomial lattice, may be found. See also [Svoboda, 2003, §10.4.1].
Given that we are in state (𝑖, 𝑗 − 1), we can only move in the next time step to state (𝑖, 𝑗) or state
(𝑖 + 1, 𝑗). Given the assumption that the short term interest rate follows a Gaussian process we
have
√
𝑟(𝑖 + 1, 𝑗) = 𝑟(𝑖, 𝑗) + 2𝜎(𝑗) Δ𝑡 (4.26)
where 𝜎(𝑗) is the volatility of the one period rate that is earned in the period [𝑗 Δ𝑡, (𝑗 + 1) Δ𝑡].
Let 𝑟𝑚 (𝑖, 𝑗) be the expected interest rate at time 𝑗 given where we are at time 𝑗 − 1, hence:
1
𝑟𝑚 (𝑖, 𝑗) = [𝑟(𝑖 + 1, 𝑗) + 𝑟(𝑖, 𝑗)]
2
√
⇒ 𝑟(𝑖 + 1, 𝑗) = 𝑟𝑚 (𝑖, 𝑗) + 𝜎(𝑗) Δ𝑡
√
𝑟(𝑖, 𝑗) = 𝑟𝑚 (𝑖, 𝑗) − 𝜎(𝑗) Δ𝑡

and so, in continuous time, we may write:

𝑟(𝑡) = 𝜇(𝑡) + 𝜎(𝑡)𝑧(𝑡)

We may apply Itô’s Lemma to determine the stochastic process for the short term interest rate, and
if we assume that the volatility is also not time dependent then
∂𝑟 ∂𝑟 ∂2𝑟
𝑑𝑟 = 𝑑𝑡 + 𝑑𝑧 + 12 2 (𝑑𝑧)2
∂𝑡 ∂𝑧 ∂𝑧
∂𝑟
= 𝑑𝑡 + 𝜎𝑑𝑧
∂𝑡
and so have our favoured affine form: we have 𝛽 > 0 constant, 𝛼 = 0 = 𝛾, but 𝜃 a function of time:
√
𝑑𝑟 = 𝜃(𝑡)𝑑𝑡 + 𝛽𝑑𝑧 (4.27)

There is no mean reversion in the model which means that the process for 𝑟 has unbounded variance.
The function 𝜃(𝑡) can be determined analytically and calibrated from the initial term structure of
interest rates. (4.12) and (4.13) have become
∂𝐴
= 𝜃(𝑡)𝐵 − 12 𝛽𝐵 2 (4.28)
∂𝑡
∂𝐵
= −1 (4.29)
∂𝑡
and so

𝐵(𝑡, 𝑇 ) = 𝑇 − 𝑡 (4.30)
∫ 𝑇
𝐴(𝑡, 𝑇 ) = − 𝜃(𝑠)(𝑇 − 𝑠) 𝑑𝑠 + 61 𝛽(𝑇 − 𝑡)3 (4.31)
𝑡

We can calibrate 𝜃(⋅) so that the observed yield curve (found from our yield curve bootstrap) fits
exactly. The best way to work with this is to think of 𝑡 as today, so 𝑡 = 0 and that the variable in
play is 𝑇 . Thus
−𝑇 r(𝑟, 0, 𝑇 ) = 𝐴(0, 𝑇 ) − 𝑟𝐵(0, 𝑇 )

54
We use (4.30) and (4.31) in this, and then differentiate twice w.r.t. 𝑇 .2 The result is

𝜃(𝑇 ) = 𝑓 ′ (0, 𝑇 ) + 𝛽𝑇 (4.32)

where 𝑓 (0, 𝑇 ) refers to the continuous forward rate for time 𝑇 , as seen at time 0 (with the world in
state 𝑟 = 𝑟(0)). Of course, here the interpolated curve needs to be twice differentiable. Most cubic
spline interpolation schemes satisfy this property, but of course there are plenty of problems with
such interpolation schemes. The best approach might be to use the method of Svensson [1994].
Under the Ho-Lee model we can value European options on default-free bonds in exactly the same
way as before. Consider a call option on an 𝑇2 −maturity discount bond with exercise price 𝐾 and
expiration 𝑇1 < 𝑇2 . Using risk-neutral valuation, the vanilla option price is:

𝑉± [𝑟, 𝑡, 𝑇1 , 𝑇2 ; 𝐾] = 𝑍(𝑟, 𝑡, 𝑇1 )𝔼ℚ

𝑡 [max(0, ±(𝑍 − 𝐾))]

where 𝑍 is the zero coupon price for expiry 𝑇2 as observed at time 𝑇1 . It has current forward value
𝑍(𝑟, 𝑡, 𝑇2 )
𝑓 (𝑍) = .
𝑍(𝑟, 𝑡, 𝑇1 )
It turns out that we obtain a ‘Black-like’ option pricing formula:

𝑐[𝑟, 𝑡, 𝑇1 , 𝑇2 ; 𝐾] = 𝑍(𝑟, 𝑡, 𝑇2 )𝑁 (𝑑+ ) − 𝐾𝑍(𝑟, 𝑡, 𝑇1 )𝑁 (𝑑− ) (4.33)

𝑝[𝑟, 𝑡, 𝑇1 , 𝑇2 ; 𝐾] = 𝐾𝑍(𝑟, 𝑡, 𝑇1 )𝑁 (−𝑑− ) − 𝑍(𝑟, 𝑡, 𝑇2 )𝑁 (−𝑑+ ) (4.34)
𝜎𝑝2 = 𝛽(𝑇2 − 𝑇1 ) 𝑇1 2
(4.35)
𝑓 (𝑍)
ln 𝐾 ± 12 𝜎𝑝2
𝑑± = (4.36)
𝜎𝑝

4.7.2 The Extensions of Hull & White

Hull and White [1990], Hull and White [1994] extended the Vasicek and CIR models to include
time-dependency in the drift and volatility parameters. The extended Vasicek model is
√
𝑑𝑟 = (𝜃(𝑡) − 𝛾𝑟(𝑡))𝑑𝑡 + 𝛽𝑑𝑧 (4.37)

and the extended CIR model is

√
𝑑𝑟 = (𝜃(𝑡) − 𝛾𝑟(𝑡))𝑑𝑡 + 𝛼𝑟(𝑡)𝑑𝑧 (4.38)

The extended Vasicek model with constant volatility is analytically tractable and results in a richer
volatility structure for forward rates, spot rates and discount bonds.
(4.12) and (4.13) have become
∂𝐴
= 𝜃(𝑡)𝐵 − 12 𝛽𝐵 2
∂𝑡
∂𝐵
= −1 + 𝛾𝐵
∂𝑡
2 Here we are using the Leibnitz rule, for differentiation of a definite integral with respect to a parameter [National

Institute of Standards and Technology, 2010, §1.5(iv)]:

∫ 𝜙(𝜉) ∫ 𝜙(𝜉)
𝑑 𝑑𝜙 (𝜉) 𝑑𝜓 (𝜉) 𝑑
𝑓 (𝜓, 𝜉) 𝑑𝜓 = 𝑓 (𝜙 (𝜉) , 𝜉) − 𝑓 (𝜓 (𝜉) , 𝜉) + 𝑓 (𝜓, 𝜉) 𝑑𝜓
𝑑𝜉 𝜓(𝜉) 𝑑𝜉 𝑑𝜉 𝜓(𝜉) 𝑑𝜉

55
and so

𝐵(𝑡, 𝑇 ) = 𝛾1 (1 − 𝑒−𝛾(𝑇 −𝑡) ) 3

∫ 𝑇 ∫ 𝑇
𝐴(𝑡, 𝑇 ) = − 1
𝜃(𝑠)𝐵(𝑠, 𝑇 ) 𝑑𝑠 + 2 𝛽 𝐵(𝑠, 𝑇 )2 𝑑𝑠
𝑡 𝑡

In other words,
∫ 𝑇 ∫ 𝑇
𝜃(𝑠)𝐵(𝑠, 𝑇 ) 𝑑𝑠 = 12 𝛽 𝐵(𝑠, 𝑇 )2 𝑑𝑠 − 𝑟𝐵(0, 𝑇 ) + r(𝑟, 0, 𝑇 )𝑇
0 0

We need to derive the function 𝜃(𝑇 ) at time 𝑡 = 0, today.

We now differentiate twice w.r.t. 𝑇 . Of course, we are able the plug in the known values of 𝐵(𝑠, 𝑇 ),
∂ ∂2
∂𝑇 𝐵(𝑠, 𝑇 ) and ∂𝑇 2 𝐵(𝑠, 𝑇 ). After some work, the result is

𝛽
𝜃(𝑇 ) = 𝑓 ′ (0, 𝑇 ) + 𝛾𝑓 (0, 𝑇 ) + (1 − 𝑒−2𝛾𝑇 ) (4.39)
2𝛾

Option pricing formulae are the same as in §4.4.2. Note that 𝜃 was not a parameter in the option
price there and 𝜃(𝑡) is not here - this information is reflected in the formulae through the zero
coupon bond prices.

Calibrating the Hull-White (Vasicek) model

In order to calibrate this model, we still need to find the short rate volatility 𝜎, and the mean
reversion speed 𝛾.

• Strip caps into caplets as in §3.4.

• As in §3.3.2, a caplet is actually a put on a bond.

• For any given value of 𝜎 and 𝛾, the puts on bonds have prices as functions of 𝛽 and 𝛾 and the
current yield curve, which is given and fixed. These formulae appear earlier in this section,
and refer back to the formulae of §4.4.2.

• We can then formulate a function an error function which measures the difference between
model and market (as stripped) prices. This function will be something like
∑
err𝜎,𝛾 = ∣𝑉𝑖 (𝜎, 𝛾) − 𝑃𝑖 ∣ (4.40)
𝑖

• We minimise this function using the Nelder-Mead algorithm.

One will often find that a stable minimum is not easily found, and so one can fix one of the
parameters (such as putting 𝛾 = 0.03; apparently the Bloomberg solution) and then solve for just
the one free parameter 𝜎.
These are also the parameters one will need to use in the market standard convexity adjustment
that makes futures prices like forward prices.
3 In showing this, one must be careful to distinguish the cases 𝛾𝐵 − 1 <> 0.

56
4.7.3 The need for the convexity adjustment
The forward price of a contract is not the same as the futures price. As the underlying rate is
positively correlated with spot interest rates, futures prices are higher than forward prices. To
understand the nature of the bias between forward and futures contracts when the underlying is an
interest rate, consider a money market futures contract and a forward rate agreement (FRA) on a
certain forward rate.
An investor holding futures contracts realises the gains/losses resulting from a change in the po-
sition’s value daily. On the other hand, an investor holding forward rate agreements realises the
gains/losses only at the maturity of the underlying interest rate; the value of these gains/losses is
the present value thereof. The gains realised on a long FRA position when the forward rate increases
will be discounted at a higher interest rate. The losses realised on a long FRA position when the
forward rate decreases will be discounted at a lower interest rate. This has the effect of decreasing
the gains and increasing the losses for a long FRA position compared to the long futures position.
If the discount rate was unaffected by the change in the FRA rate, then a long position in a
future could be perfectly hedged by a short position in 𝐶(0, 𝑡) many positions in a FRA, where 𝑡
is the commencement date of the FRA. However, as the spot rate (and hence the factor 𝐶(0, 𝑡)) is
positively correlated with the FRA rate, the actual p&l of this position should be positive (in the
simplest case of perfect correlation, the p&l is positive for all moves in the rate. See Figure 4.2.)

Figure 4.2: The p&l of a futures position hedged with short forwards, assuming the rates are equal
and correlation between rates is perfect.

An alternative intuitive explanation for this phenomenon is as follows: the long party to the futures
contract will tend to receive margin payments on days when interest rates rise, and make margin
payments on days when interest rates decline. That investor will then invest profits they receive at
higher interest rates, and fund losses they make at lower interest rates.
In either case we see that if futures and FRAs had the same rate, the future is far more attractive.
Thus the fixed rate in the FRA should be lower in order to restore the attractiveness of the FRA.
However, recent research shows that this factor might not be taken into account in general i.e. that

57
futures and forwards do not trade relative to each other in this way - see Poskitt [2008].

4.7.4 Derivation of the convexity adjustment

To bootstrap the current yield curve, forward rates are required. Typically, a futures rate will be
input, an adjustment to make the rate ‘forward like’ will be made, and that rate will be used. This
adjustment is model dependent. The models is use are the Ho-Lee model (perhaps the market
standard) and the Hull-White model (the only one offered by Bloomberg, so will be the market
standard soon enough). The difference between the futures rate and the forward rate is termed the
convexity adjustment.
The theoretical futures rate 𝐹 (𝑇1 , 𝑇2 ), is calculated using the following equation

𝐹 (𝑡; 𝑇1 , 𝑇2 ) = 𝔼ℚ
𝑡 [𝑓 (𝑇1 ; 𝑇1 , 𝑇2 )] (4.41)

where 𝔼ℚ𝑡 [⋅] is the expectation under the risk-neutral measure and 𝑓 (𝑡; 𝑇1 , 𝑇2 ) is the forward rate
at time 𝑡, for the time interval [𝑇1 , 𝑇2 ]. We can only solve this equation by introducing an interest
rate model. These models will require some differentiability of the curve.

Ho-Lee Model

The derivation here is based on Hull.

The dynamics of the short rate under the Ho-Lee model are

𝑑𝑟(𝑡) = 𝜃(𝑡)𝑑𝑡 + 𝜎𝑑𝑊 (𝑡) (4.42)

where 𝜎 is the volatility of the short rate, and 𝑊 (𝑡) is standard Brownian motion under the risk-
neutral measure. The 𝜃(𝑡) function is determined from the initial term structure as

𝜃(𝑡) = 𝑓 ′ (0, 𝑡) + 𝜎 2 𝑡 (4.43)

where 𝑓 ′ (0, 𝑡) is the derivative with respect to 𝑡 of the instantaneous forward rate for maturity 𝑡, as
seen today. Under the Ho-Lee model, the price of a zero-coupon bond, 𝑍(𝑡, 𝑇 ) is given by

𝑍(𝑡, 𝑇 ) = 𝑒𝐴(𝑡,𝑇 )−𝑟(𝑡)(𝑇 −𝑡)

where
𝑍(0, 𝑇 ) ∂ ln 𝑍(0, 𝑡) 1 2
𝐴(𝑡, 𝑇 ) = ln − (𝑇 − 𝑡) − 𝜎 𝑡(𝑇 − 𝑡)2
𝑍(0, 𝑡) ∂𝑡 2
Using Itô’s lemma one shows that

𝑑𝑍(𝑡, 𝑇 )
2
= 𝑍𝑡 (𝑡, 𝑇 )𝑑𝑡 + 𝑍𝑟 (𝑡, 𝑇 )𝑑𝑟(𝑡) + 12 𝑍𝑟𝑟 (𝑡, 𝑇 ) (𝑑𝑟(𝑡))
[ ]
= 𝑒𝐴(𝑡,𝑇 ) 𝑒−𝑟(𝑡)(𝑇 −𝑡) 𝐴𝑡 (𝑡, 𝑇 ) + 𝑒𝐴(𝑡,𝑇 ) 𝑒−𝑟(𝑡)(𝑇 −𝑡) 𝑟(𝑡) 𝑑𝑡 − 𝑒𝐴(𝑡,𝑇 ) 𝑒−𝑟(𝑡)(𝑇 −𝑡) (𝑇 − 𝑡)𝑑𝑟(𝑡)

+ 21 𝑒𝐴(𝑡,𝑇 ) 𝑒−𝑟(𝑡)(𝑇 −𝑡) (𝑇 − 𝑡)2 𝜎 2 𝑑𝑡

= 𝑍(𝑡, 𝑇 ) [𝐴𝑡 (𝑡, 𝑇 ) + 𝑟(𝑡)] 𝑑𝑡 − 𝑍(𝑡, 𝑇 )(𝑇 − 𝑡) [𝜃(𝑡)𝑑𝑡 + 𝜎𝑑𝑊 (𝑡)] + 12 𝑍(𝑡, 𝑇 )(𝑇 − 𝑡)2 𝜎 2 𝑑𝑡
= 𝑍(𝑡, 𝑇 ) 𝐴𝑡 (𝑡, 𝑇 ) + 𝑟(𝑡) + 21 (𝑇 − 𝑡)2 𝜎 2 − (𝑇 − 𝑡)𝜃(𝑡) 𝑑𝑡 − 𝑍(𝑡, 𝑇 )(𝑇 − 𝑡)𝜎𝑑𝑊 (𝑡)
[ ]

58
and since

𝐴𝑡 (𝑡, 𝑇 ) = (𝑇 − 𝑡)𝜃(𝑡) − 12 (𝑇 − 𝑡)2 𝜎 2 (4.44)

it follows that

𝑑𝑍(𝑡, 𝑇 ) = 𝑟(𝑡)𝑍(𝑡, 𝑇 ) 𝑑𝑡 − 𝜎𝑍(𝑡, 𝑇 ) 𝑑𝑊 (𝑡)

Now using (1.12) we can again use Itô’s lemma to find

𝑑𝑓 (𝑡; 𝑇1 , 𝑇2 )
= 𝑓𝑡 (𝑡; 𝑇1 , 𝑇2 )𝑑𝑡 + 𝑓𝑟 (𝑡; 𝑇1 , 𝑇2 )𝑑𝑟(𝑡) + 21 𝑓𝑟𝑟 (𝑡; 𝑇1 , 𝑇2 )(𝑑𝑟(𝑡))2
[ ]
1 𝐴𝑡 (𝑡, 𝑇1 )𝑍(𝑡, 𝑇1 ) + 𝑟(𝑡)𝑍(𝑡, 𝑇1 ) 𝐴𝑡 (𝑡, 𝑇2 )𝑍(𝑡, 𝑇2 ) + 𝑟(𝑡)𝑍(𝑡, 𝑇2 )
= − 𝑑𝑡
𝑇2 − 𝑇1 𝑍(𝑡, 𝑇1 ) 𝑍(𝑡, 𝑇2 )
[ ]
1 (𝑇1 − 𝑡)𝑍(𝑡, 𝑇1 ) (𝑇2 − 𝑡)𝑍(𝑡, 𝑇2 )
+ − + (𝜃(𝑡)𝑑𝑡 + 𝜎𝑑𝑊 (𝑡)) + 0𝑑𝑡
𝑇2 − 𝑇1 𝑍(𝑡, 𝑇1 ) 𝑍(𝑡, 𝑇2 )
1
= [𝐴𝑡 (𝑡, 𝑇1 ) − 𝐴𝑡 (𝑡, 𝑇2 )] 𝑑𝑡 + 𝜃(𝑡)𝑑𝑡 + 𝜎𝑑𝑊 (𝑡)
𝑇2 − 𝑇1

and again using (4.44) we have that

[ ]
𝐴𝑡 (𝑡, 𝑇1 ) − 𝐴𝑡 (𝑡, 𝑇2 )
𝑑𝑓 (𝑡; 𝑇1 , 𝑇2 ) = + 𝜃(𝑡) 𝑑𝑡 + 𝜎 𝑑𝑊 (𝑡)
𝑇2 − 𝑇1
𝜎2
= (𝑇2 + 𝑇1 − 2𝑡) 𝑑𝑡 + 𝜎 𝑑𝑊 (𝑡)
2
Thus

𝐹 (0; 𝑇1 , 𝑇2 ) = 𝔼ℚ𝑡 [𝑓 (𝑇1 ; 𝑇1 , 𝑇2 )]

∫ 𝑇1 2
𝜎
= 𝑓 (0; 𝑇1 , 𝑇2 ) + (𝑇2 + 𝑇1 − 2𝑠) 𝑑𝑠
0 2
1
= 𝑓 (0; 𝑇1 , 𝑇2 ) + 𝜎 2 𝑇1 𝑇2
2

Hull-White Model

In the Hull-White model

𝑑𝑍(𝑡, 𝑇 ) = 𝑟(𝑡)𝑍(𝑡, 𝑇 ) 𝑑𝑡 − 𝐵(𝑡, 𝑇 )𝜎𝑍(𝑡, 𝑇 ) 𝑑𝑊 (𝑡)

and after some work we get

𝐵(𝑇2 − 𝑇1 ) 𝜎 2 [
𝐵(𝑇2 − 𝑇1 )𝐵(2𝑇1 ) + 2𝐵(𝑇1 )2
]
𝐹 (0; 𝑇1 , 𝑇2 ) = 𝑓 (0; 𝑇1 , 𝑇2 ) +
𝑇2 − 𝑇1 4

59
4.8 Exercises
1. In the course of deriving the solution to the bond equation we arrive at two ordinary differential
equations for the functions 𝐴(𝑡, 𝑇 ) and 𝐵(𝑡, 𝑇 ): equations (4.19) and (4,20). Assume 𝛼(𝑡) and
𝛾(𝑡) are constants. Integrate the equation (4.20) with boundary condition 𝐵(𝑇, 𝑇 ) = 0 to
obtain a closed form solution for 𝐵(𝑡, 𝑇 ).

2. Write down the form of the Bond-pricing equation when the short rate satisfies the following
Vasicek model
√
𝑑𝑟 = (𝜃 − 𝛾𝑟) 𝑑𝑡 + 𝛽𝑑𝑧.
Solve the resulting ordinary differential equations for 𝐴(𝑡, 𝑇 ) and 𝐵(𝑡, 𝑇 ). Express the solution
for 𝐴(𝑡, 𝑇 ) in terms of 𝐵(𝑡, 𝑇 ).

3. (a) What is the difference between an equilibrium short-term interest rate model and a no-
arbitrage model?
(b) If a stock price followed a mean-reverting or path-dependent process, there would be
market inefficiency. Why is this NOT the case when short-term interest rate models are
mean-reverting or path-dependent?
(c) Suppose that the short rate is currently 3.5% and its volatility measure is 0.95% per
annum. What happens to the volatility measure when the short rate increases to 10.5%
in (i) Vasicek’s model; (ii) Rendleman and Bartter’s model; and (iii) the Cox, Ingersoll,
Ross model?

4. (a) Given the following parametrisations of the Vasicek and CIR models:
√
𝑑𝑟 = (𝜃 − 𝛾𝑟)𝑑𝑡 + 𝛽𝑑𝑧
√
𝑑𝑟 = (𝜃 − 𝛾𝑟)𝑑𝑡 + 𝛼𝑟 𝑑𝑧,

respectively. Suppose that 𝛾 = 0.15 and 𝜃 = 0.015, and that the initial short rate is
𝑟 = 10%. Let the initial volatility measure for the short rate be 2%. Compare the two
different values given by the models for the 8-year discount factor. What is the associated
8-year, continuously-compounded interest rate in each instance?
(b) What happens to the values above when 𝑟 = 9.99%? Calculate the present value of a one
basis point shift (𝑃 𝑉 01) of a par 100 zero-coupon bond in each instance.

5. Perform the derivation of the 𝜃(𝑇 ) function in the Ho-Lee model.

6. (a) Suppose that 𝑎 = 0.25, 𝑏 = 0.06, and 𝜎 = 0.022 in Vasicek’s model with the initial
short-term interest rate being 𝑟(0) = 5.8%. Calculate the price of a 2.20 year European
call option on a coupon-bearing bond that will mature in three years’ time. Suppose that
the bond pays a coupon of 6% semi-annually. The par value of the bond is 100 and the
strike price of the option is 99. The strike price is the cash price (not the quoted price)
that will be paid for the bond.
(b) Use the answer to (a) and put-call parity arguments to calculate the price of a put option
that has the same terms as the call option.

60
 (c) Plot the yield curve for these values of 𝛾, 𝑏, 𝜎 and 𝑟 out to 5 years.

7. Perform the derivation of the 𝜃(𝑇 ) function in the Hull-White-Vasicek model.

8. Suppose we are using Vasicek’s model with a volatility of 4%, 𝛾 = 15%, 𝜂 = 2% and short
rate 𝑟 = 10%.

(a) Graph the yield curve.

(b) Find the price of a one year call option on a three year zero coupon bond par 1, struck
at 97c.

61
Chapter 5

The LIBOR market model

The input is a model for the LIBOR forward rates. These forward rates are of course extracted
from the bootstrapped yield curve, while the volatilities are found from the cap/caplet Black model
Black [1976] inputs.
In each Black model for maturity date 𝑇 , the forward rates will be log-normally distributed in what
is called the 𝑇 -forward measure, where 𝑇 is the pay date of the option. The fact that we can ‘legally’
discount the expected payoff under this measure using today’s yield curve is a consequence of some
profound academic work of Geman et al. [1995]. We can proceed to use Black’s model without
knowing any of the theory of the LMM (or of Geman et al. [1995]); however, the Black model
cannot safely be used to value more complicated products where the payoff depends on observations
at multiple dates.
The LIBOR market model is then a collection of Black models. Each of these models will have its
own pricing measure. However, one can move from any one of these measures to any other by using
a Radon-Nikodym change of measure. By changing all of the measures to a single measure (the one
of furthest tenor) enables us to consider a single pricing measure.
In addition to the volatility inputs, the matrix of correlations between each of these forward rates
is required - the correlations are the linkers between the forward rates. These correlations are not
observable in the market. They either need to be determined from market data in the calibration
step, or are given by exogenous considerations and then are inputs to further calibration routines.
In the latter case, the fit to the market will be less satisfactory than in the former, but these models
may subsequently perform in a more realistic fashion than methods where the fit is tighter.
The LIBOR market model is also commonly known as the BGM model after Brace et al. [1997].
The model was also initially developed by Miltersen et al. [1997] and Jamshidian [1997].
The LMM has become a standard in the banking industry, and is commonly used for pricing most
exotic interest rate derivatives. Even in the case where other models would clearly be easier to
use, the industry prefers to extend this model: possibly because it is a natural extension of the
Black-Scholes world, and the amenability of Monte Carlo pricing for almost any product variation.
Nevertheless, the ability of these models to capture rate curve dynamics is more than questionable
Manuel Huyet [2007]: in reality there is clear evidence of jumps, regimes, and skew. The normali-

62
ty/lognormality assumption is of course quite questionable. The LMM is only now being extended
to a regime which is skew aware e.g. Svoboda-Greenwood [2007], Rebonato [2007].

5.1 The model for a single forward rate

Let 𝐿𝑖 (𝑡) := 𝐿(𝑡; 𝑡𝑖−1 , 𝑡𝑖 ) be the FRA-rate for the period from 𝑡𝑖−1 to 𝑡𝑖 . Clearly

𝐶(𝑡, 𝑡𝑖 ) = 𝐶(0, 𝑡𝑖−1 ) [1 + 𝐿𝑖 (𝑡)𝛼𝑖 ] (5.1)

where 𝑡 is today, 𝐶(𝑡, ⋅) are the capitalisation factors extracted from the yield curve, 𝛼𝑖 is the time
from 𝑡𝑖−1 to 𝑡𝑖 in years, observing the relevant day count convention. It follows that
[ ]
1 𝑍(𝑡, 𝑡𝑖−1 )
𝐿𝑖 (𝑡) = −1 (5.2)
𝛼𝑖 𝑍(𝑡, 𝑡𝑖 )
1
𝑍(𝑡, 𝑡𝑖 )𝐿𝑖 (𝑡) = [𝑍(𝑡, 𝑡𝑖−1 ) − 𝑍(𝑡, 𝑡𝑖 )] (5.3)
𝛼𝑖
Of course this is just §1.7 again. In particular, 𝑍(𝑡, 𝑡𝑖 )𝐿𝑖 (𝑡) is a tradeable asset. If we consider the
measure for which 𝑍(𝑡, 𝑡𝑖 ) is the numeraire - this is known as the 𝑡𝑖 -forward measure - then 𝐿𝑖 (𝑡) is
a martingale under this measure. It now follows from the Martingale Representation Theorem1

𝑑𝐿𝑖 (𝑡) = 𝐿𝑖 (𝑡)𝑣𝑖 (𝑡)′ 𝑑𝑊 𝑖 (𝑡) (5.4)

⎡ ⎤
𝑊1𝑖 (𝑡)
⎢ 𝑊2𝑖 (𝑡) ⎥
⎢ ⎥
𝑖
Here 𝑊 (𝑡) = ⎢ ⎢ .. ⎥ are correlated Brownian motions under the 𝑡𝑖 -forward measure, and
⎥
⎣ . ⎦
𝑊𝑛𝑖 (𝑡)
⎡ ⎤
𝑣1𝑖 (𝑡)
⎢ 𝑣2𝑖 (𝑡) ⎥
⎢ ⎥
𝑣𝑖 (𝑡) = ⎢
⎢ .. ⎥ are deterministic functions, and they are each 0 as soon as we reach 𝑡𝑖−1 . In fact
⎥
⎣ . ⎦
𝑣𝑛𝑖 (𝑡)
we can (and do) assume that all except the 𝑖𝑡ℎ entry of 𝑣𝑖 (𝑡) are 0.

5.2 The pricing of caplets and caps

Since 𝑣𝑖 (𝑡) is deterministic, we have
∫ 𝑡𝑖−1
Σ𝑡𝑖 [ln 𝐿(𝑡𝑖−1 )∣ℱ𝑡 ] = ∥𝑣𝑖 (𝑠)∥2 𝑑𝑠
𝑡
∫ 𝑡𝑖−1
𝑡𝑖
𝔼 [ln 𝐿(𝑡𝑖−1 )∣ℱ𝑡 ] = ln 𝐿(𝑡) − 21 ∥𝑣𝑖 (𝑠)∥2 𝑑𝑠
𝑡
1 This theorem tells you that that 𝑑𝐿𝑖 (𝑡) = 𝑣𝑖 (𝑡)′ 𝑑𝑊 𝑖 (𝑡). Since 𝐿𝑖 (𝑡) is never 0 this can be rewritten in the given
form.

63
Thus the value of a caplet with strike 𝐾 is given by

𝑉 (𝑡) = 𝑍(𝑡, 𝑡𝑖 )𝛼𝑖 𝔼𝑡𝑖 (𝐿(𝑡𝑖−1 ) − 𝐾)+ ∣ℱ𝑡

[ ]

= 𝑍(𝑡, 𝑡𝑖 )𝛼𝑖 [𝐿(𝑡)𝑁 (𝑑1 ) − 𝐾𝑁 (𝑑2 )]

ln 𝐿(𝑡)
𝐾 ± 2Σ
1
𝑑± = √
Σ
∫ 𝑡𝑖−1
Σ= ∥𝑣𝑖 (𝑠)∥2 𝑑𝑠
𝑡

A cap is a collection of caplets. Thus the cap has value

𝑀
∑
𝑉 (𝑡) = 𝑍(𝑡, 𝑡𝑖 )𝛼𝑖 [𝐿(𝑡; 𝑡𝑖−1 , 𝑡𝑖 )𝑁 (𝑑𝑖1 ) − 𝐾𝑁 (𝑑𝑖2 )]
𝑖=1

ln 𝐿(𝑡;𝑡𝐾
𝑖−1 ,𝑡𝑖 )
± 21 Σ𝑖
𝑑𝑖± = √
Σ𝑖
∫ 𝑡𝑖−1
Σ𝑖 = ∥𝑣𝑖 (𝑠)∥2 𝑑𝑠
𝑡

This shows that the LMM is in every way consistent with the Black model, and this model can
recover all of the Black model inputs that are used for calibration.

5.3 A common measure

In §5.2 we could have used a single Brownian motion and a one dimensional 𝑣(𝑡). However, we want
to price more complex derivatives based on many forward rates, possibly with many payoff times.
So, we use a vector of Brownian motions.
We know numeraires for the measure 𝑡𝑗−1 and for the ‘next’ measure 𝑡𝑗 , so we can explicitly calculate
the likelihood process:

𝑑ℙ𝑗−1 𝑍(𝑡, 𝑡𝑗−1 ) 𝑍(𝑡, 𝑡𝑗 )

𝜂(𝑡) := (𝑡) = /
𝑑ℙ𝑗 𝑍(0, 𝑡𝑗−1 ) 𝑍(0, 𝑡𝑗 )
𝑍(0, 𝑡𝑗 ) 𝑍(𝑡, 𝑡𝑗−1 )
=
𝑍(0, 𝑡𝑗−1 ) 𝑍(𝑡, 𝑡𝑗 )
𝑍(0, 𝑡𝑗 )
= (1 + 𝛼𝑗 𝐿𝑗 (𝑡))
𝑍(0, 𝑡𝑗−1 )

Now
𝑍(0, 𝑡𝑗 )
𝑑𝜂(𝑡) = 𝛼𝑗 𝑑𝐿𝑗 (𝑡)
𝑍(0, 𝑡𝑗−1 )
𝑍(0, 𝑡𝑗 )
= 𝛼𝑗 𝐿𝑗 (𝑡)𝑣𝑗 (𝑡)′ 𝑑𝑊 𝑗 (𝑡)
𝑍(0, 𝑡𝑗−1 )
𝛼𝑗 𝐿𝑗 (𝑡)
= 𝜂(𝑡) 𝑣𝑗 (𝑡)′ 𝑑𝑊 𝑗 (𝑡)
1 + 𝛼𝑗 𝐿𝑗 (𝑡)

64
and so by Girsanov’s theorem

𝛼𝑗 𝐿𝑗 (𝑡)
𝑑𝑊 𝑗−1 (𝑡) = 𝑑𝑊 𝑗 (𝑡) − 𝑣𝑗 (𝑡) 𝑑𝑡
1 + 𝛼𝑗 𝐿𝑗 (𝑡)

By induction,
𝑛
∑ 𝛼𝑗 𝐿𝑗 (𝑡)
𝑑𝑊 𝑖 (𝑡) = 𝑑𝑊 𝑛 (𝑡) − 𝑣𝑗 (𝑡) 𝑑𝑡
𝑗=𝑖+1
1 + 𝛼𝑗 𝐿𝑗 (𝑡)

Note that the last equation is one of 𝑛 × 1 column vectors. It follows that
⎡ ⎤
𝑛
∑ 𝛼 𝐿
𝑗 𝑗 (𝑡)
𝑑𝐿𝑖 (𝑡) = 𝐿𝑖 (𝑡)𝑣𝑖 (𝑡)′ ⎣𝑑𝑊 𝑛 (𝑡) − 𝑣𝑗 (𝑡) 𝑑𝑡⎦ (5.5)
𝑗=𝑖+1
1 + 𝛼𝑗 𝐿𝑗 (𝑡)

5.4 Pricing exotic instruments under LMM

In order to price derivatives, we now do the following

• Use Monte Carlo to evolve the forward rates to the first cash flow at 𝑡𝑖 in the derivative.

• Calculate the cash flows in that experiment.

• Future value these cash flows forward to time 𝑡𝑛 by using the capitalisation factors that have
arisen from the experiment, NOT using today’s forward capitalisation factor 𝐶(𝑡; 𝑡𝑖 , 𝑡𝑛 ).

• Present value back to today by using today’s discount factor 𝑍(𝑡, 𝑡𝑛 ).

• Continue the Monte Carlo evolution to 𝑡𝑖+1 and repeat.

5.4.1 A simple example

Suppose now is time 𝑡0 . In three months time (time 𝑡1 ) we observe the LIBOR rate. If it is greater
than 10.5%, a caplet on the then three month rate (to be observed at time 𝑡2 ) is knocked in with a
strike of 10%. The payoff occurs in advance, 6 months from now. 9 months from now is time 𝑡3 .
Let 𝐿0 be the current LIBOR rate, 𝐿1 be the rate for the 3x6 period, and 𝐿2 the rate for the 6x9
period. Let 𝑣1 and 𝑣2 be the volatilities and 𝜌 the correlation.

𝑑𝐿1 (𝑡) = 𝐿1 (𝑡)𝑣1 (𝑡)′ 𝑑𝑊 1 (𝑡)

𝑑𝐿2 (𝑡) = 𝐿2 (𝑡)𝑣2 (𝑡)′ 𝑑𝑊 2 (𝑡)

Also
𝛼2 𝐿2 (𝑡)
𝑑𝑊 1 (𝑡) = 𝑑𝑊 2 (𝑡) − 𝑣2 (𝑡) 𝑑𝑡
1 + 𝛼2 𝐿2 (𝑡)
[ ]
′ 2 𝛼2 𝐿2 (𝑡)
⇒ 𝑑𝐿1 (𝑡) = 𝐿1 (𝑡)𝑣1 (𝑡) 𝑑𝑊 (𝑡) − 𝑣2 (𝑡) 𝑑𝑡
1 + 𝛼2 𝐿2 (𝑡)

65
Thus
𝛼2 𝐿2 (𝑡)
𝑑𝐿1 (𝑡) = 𝐿1 (𝑡)𝑣11 (𝑡)𝑑𝑊12 (𝑡) − 𝐿1 (𝑡) 𝜌𝑣11 (𝑡)𝑣22 (𝑡) 𝑑𝑡
1 + 𝛼2 𝐿2 (𝑡)
𝑑𝐿2 (𝑡) = 𝐿2 (𝑡)𝑣22 (𝑡)𝑑𝑊22 (𝑡)
𝜌 𝑑𝑡 = 𝑑𝑊12 (𝑡)𝑑𝑊22 (𝑡)

To find a solution of this scheme, we can proceed as follows: we find 𝑑𝑊12 and 𝑑𝑊22 using the
Cholesky decomposition.2 We then calculate 𝑑𝐿1 and 𝑑𝐿2 and the new values of 𝐿1 and 𝐿2 (we
are now at time 𝑡1 ). If 𝐿1 < 10.5% we can exit - the option expires worthless. If not, we generate
another 𝑑𝑊22 and another 𝑑𝐿2 and the new value of 𝐿2 (we are now at time 𝑡2 ). We then determine
the payoff of the caplet with the strike of 10%. The payoff is in advance, so we capitalise the payoff
with the observed value of 𝐿2 to time 𝑡3 .
We take averages in our Monte Carlo experiment, and discount to today using the current 9m
discount factor.
We have used the most naı̈ve discretisation of the scheme. In reality, smaller time steps need to be
taken. For details see [Brigo and Mercurio, 2006, §6.10].

5.4.2 Calibration of the parameters

• A decent yield curve bootstrap algorithm needs to be used. Such issues have already been
discussed.

• Volatilities need to be determined from information in the market. This is the same cap to
caplet problem seen in §3.4.

• Correlations need to be modelled.

• For the Monte Carlo we need low discrepancy sequences in high dimensions. Sobol’ sequences
are most suitable here, see [Jäckel, 2002, Chapter 8].

2 Suppose there are two underlyings. Using excel/vba, we first extract pairs of uniformly distributed random
numbers 𝑈1 , 𝑈2 , then transform them into pairs of independent normally distributed random numbers 𝑍1 , 𝑍2 by
using the inverse of the cumulative normal distribution. We then apply the Cholesky decomposition:
√
𝑊1 = 𝑍1 , 𝑊2 = 𝜌𝑍1 + 1 − 𝜌2 𝑍2 (5.6)

Now 𝑊1 and 𝑊2 are normally distributed random numbers with correlation 𝜌.

66
Bibliography

F. Black, E. Derman, and W. Toy. A one-factor model of interest rates and its application to
treasury bond options. Financial Analysts Journal, pages 33–39, January/February 1990. 22, 52

Fischer Black. The pricing of commodity contracts. Journal of Financial Economics, 3:167–179,
1976. 28, 62

Fischer Black and Piotr Karasinski. Bond and option pricing when short rates are lognormal.
Financial Analysts Journal, 47(4):52–59, 1991. 53

A. Brace, D. Ga̧tarek, and M. Musiela. The market model of interest rate dynamics. Journal of
Mathematical Finance, 7(2):127–155, 1997. 62

M. J. Brennan and E. S. Schwartz. A continuous time approach to the pricing of bonds. Journal of
Banking and Finance, 31:33–155, 1979. 51

M. J. Brennan and E. S. Schwartz. An equilibrium model of bond pricing and a test of market
efficiency. Journal of Financial and Quantitative Analysis, 17:301–329, 1982. 51

Damiano Brigo and Fabio Mercurio. Interest Rate Models - Theory and Practice: With Smile,
Inflation and Credit. Springer Finance, second edition, 2006. 35, 66

Andrew J. G. Cairns. Interest Rate Models: An Introduction. Princeton University Press, 2004. 53

J. C. Cox, J. E. Ingersoll, and S. A. Ross. A theory of the term structure of interest rates. Econo-
metrica, 53:385–407, 1985. 50

Mark Davis and Jan Ob̷lój. Market completion using options. 2007. URL http://arxiv.org/PS_
cache/arxiv/pdf/0710/0710.2792v4.pdf. 44

M Dothan. On the term structure of interest rates. Journal of Financial Economics, 6:59–69, 1978.
45

H. Geman, N. El Karoui, and J. Rochet. Changes of numeraire, changes of probability measure and
option pricing. Journal of Applied Probability, 32, 1995. 39, 62

Patrick S. Hagan and Graeme West. Interpolation methods for curve construction. Applied Mathe-
matical Finance, 13(2):89–129, 2006. 5, 9

67
Patrick S. Hagan and Graeme West. Methods for constructing a yield curve. WILMOTT Magazine,
May:70–81, 2008. URL http://www.finmod.co.za/interpreview.pdf. 5, 9

Patrick S. Hagan, Deep Kumar, Andrew S. Lesniewski, and Diana E. Woodward. Managing smile
risk. WILMOTT Magazine, September:84–108, 2002. URL http://www.wilmott.com/pdfs/
021118_smile.pdf. 44

T. S. Y. Ho and S.-B. Lee. Term structure movements and pricing interest rate contingent claims.
Journal of Finance, 41, 1986. 18

J. Hull and A. White. Pricing interest rate derivative securities. Review of Financial Studies, 3,
1990. 55

J. Hull and A. White. Branching out. Risk, 7:34–37, 1994. 55

John Hull. Convexity adjustments to eurodollar futures. URL http://www.rotman.utoronto.ca/

˜hull/Technical%20Notes/TechnicalNote1.pdf. Technical Note No. 1, Options, Futures, and
Other Derivatives. 58

John Hull. Options, Futures, and Other Derivatives. Prentice Hall, sixth edition, 2005. 29, 30, 38

International Swaps and Derivatives Association. EMU and market conventions: Recent develop-
ments, 25 November 1998. URL http://www.isda.org/c_and_a/pdf/mktc1198.pdf. 6

Peter Jäckel. Monte Carlo Methods in Finance. Wiley, 2002. 35, 66

Farshid Jamshidian. An exact bond option pricing formula. Journal of Finance, 44(1):205–209,
1989. 48

Farshid Jamshidian. Libor and swap market models and measures. Finance and Stochastics, 1(4):
293–330, 1997. 62

F.A Longstaff and E.S. Schwartz. Interest rate volatility and the term structure: A two-factor
general equilibrium model. Journal of Finance, 47, 1992a. 51

F.A Longstaff and E.S. Schwartz. A two-factor interest rate model and contingent claims valuation.
Journal of Fixed Income, 1992b. 51

Société Générale Manuel Huyet. Libor market models: the reasons behind the success. A focus on
calibration., 2007. URL http://www-rocq.inria.fr/mathfi/mathfiinria.html. 62

Kristian R. Miltersen, Klaus Sandmann, and Dieter Sondermann. Closed form solutions for term
structure derivatives with log-normal interest rates. Journal of Finance, 52(1):409–430, 1997. 62

National Institute of Standards and Technology. Digital library of mathematical functions, 2010.
URL http://dlmf.nist.gov/. 55

Russell Poskitt. In search of the convexity adjustment: Evidence from the sterling futures and imm
fra markets. Journal of Futures Markets, 28(7):617–633, 2008. 58

68
Ricardo Rebonato. Interest-Rate Option Models. John Wiley and Sons Ltd, second edition, 1998.
52, 54

Ricardo Rebonato. Modern pricing of interest-Rate derivatives: the LIBOR market model and
beyond. Princeton University Press, 2002. 35

Ricardo Rebonato. Volatility and Correlation: The Perfect Hedger and the Fox. Wiley Finance,
2004. 35, 36

Ricardo Rebonato. A time-homogeous, SABR-consistent extension of the LMM: calibration and nu-
merical results, 2007. URL http://www.riccardorebonato.co.uk/papers/LMMSABRText.pdf.
63

R. Rendelman and B. Bartter. The pricing of options on debt securities. Journal of Financial and
Quantitative Analysis, 15, 1980. 45

L.E. Svensson. Estimating and interpreting forward interest rates: Sweden 1992-1994, 1994. URL
http://www.nber.org/papers/W4871. IMF working paper. 55

Simona Svoboda. An investigation of various interest rate models and their calibration
in the South African market. Master’s thesis, University of the Witwatersrand, 2002.
URL http://web.wits.ac.za/NR/rdonlyres/EC8B7EF6-E72C-4EA1-954D-A9D8180564CD/0/
simonamsccomplete.zip. 9, 42

Simona Svoboda. Interest Rate Modelling. Finance and Capital Markets Series. Palgrave Macmillan,
2003. 42, 54

Simona Svoboda-Greenwood. Volatility Specifications in the LIBOR Market Model. PhD the-
sis, Magdalen College, University of Oxford, 2007. URL http://people.maths.ox.ac.uk/
˜svoboda/other/FinalThesis02012008.pdf. 63

O. A. Vasicek. An equilibrium characterisation of the term structure. Journal of Financial Eco-

nomics, 15, 1977. 47

Graeme West. The Mathematics of South African Financial Markets and Instruments: Lecture
notes, 2009. URL http://www.finmod.co.za/safm.pdf. 4, 5

Paul Wilmott. Paul Wilmott on Quantitative Finance. John Wiley and Sons Ltd., 2000. 42

69