Fixed Income Securities and Interest Rate Derivatives:

a Quantitative Approach

ZIQI XU

University of Oxford

April 16, 2021

 1

To all the people who care about me.

Contents

1 Basic Considerations 1
1.1 Notations and Statements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Assumptions and Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1

2 Zero-coupon Bond, Forward Rate and Short Rate 2

2.1 T −bond and Zero-coupon(Zero) Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2.2 ZCB Price, Forward Rate and Short Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

2.3 Linking Everything Together(Part 1): B(t, T ) and f (t, T ) . . . . . . . . . . . . . . . . . . . . . . 3
2.4 Linking Everything Together(Part 2): f (t, T ) and R(t) . . . . . . . . . . . . . . . . . . . . . . . . 7
2.5 Linking Everything Together(Part 3): B(t, T ) and R(t) . . . . . . . . . . . . . . . . . . . . . . . 8

2.6 A Simple Multi-dimensional Discussion of HJM Condition . . . . . . . . . . . . . . . . . . . . . . 12

3 Short Rate Models and Affine Methods for ZCB Pricing 15

3.1 Stochastic Differential Equations of Short Rate . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.2 Affine Solutions of ZCB Price . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

3.2.1 Affine Model Solution for Ho-Lee Short Rate Model . . . . . . . . . . . . . . . . . . . . . 16

3.2.2 Affine Model Solution for Hull-White(Extended Vasicek) Short Rate Model . . . . . . . . 16
3.3 Model Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.3.1 Example 1: Ho-Lee Model Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17

3.3.2 Example 2: Hull-White(Extended Vasicek) Model Calibration . . . . . . . . . . . . . . . . 18

4 LIBOR and Forward LIBOR 21

4.1 Martingale Representation Theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.2 Stochastic Representation of Assets and T −forward Measure . . . . . . . . . . . . . . . . . . . . 23
4.3 LIBOR and Forward LIBOR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

4.4 Black Caplet Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

4.5 Forward LIBOR Volatility . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

5 Bonds and Interest Rate Derivatives Pricing 33

5.1 Fixed Coupon-bearing Bonds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

5.2 Floating Rate Notes(FRN) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

2
CONTENTS 3

5.3 Interest Rate Swaps(IRS) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

5.4 Interest Rate Options . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

5.4.1 Caps and Floors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

5.4.2 Swaptions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Chapter 1

Basic Considerations

1.1 Notations and Statements

(1) ZCB means zero coupon bonds in this text. It is true that the notional amount of a ZCB
can be any positive quantity and the ZCB from 0 to T whose notional amount is 1 unit of
currency is usually called the T −bond. However, when we say ZCB in this text, we mean a
T −bond if we do not give other illustrations.
(2) If the word “risk neutral measure” appears in the text, then it means the risk neutral
measure when we use bank account as our numeraire. The measure is typically denoted by Q.
Under Q, the discounted price process of any tradable asset should be a martingale. However,
if we say “risk neutral measure for asset N (t)”, then we mean the risk neutral measure when
we use N (t) as our numeraire. The measure is typically denoted by QN . Under QN , the price
S(t)
of any tradable asset S(t) denominated by N (t) should be a martingale, i.e. the process N (t)
is
a QN −martingale.
(3) All the notations N (·) in this text mean the cumulative density function of the standard
normal distribution.

1.2 Assumptions and Remarks

All the following contents are based on some assumptions that are listed below.
(1)We never consider the credit risk for the contents in this text.
(2) The market has enough liquidity and no transaction cost or tax fees will be considered.
Under this condition, we are able to buy or sell ZCB with any maturity T , although this is
typically not the case of the real market.
(3) In order to better follow the ideas developed in the following chapters, readers should
understand the Black-Scholes model, risk neutral pricing and basic concepts about the change
of numeraire first.

1
Chapter 2

Zero-coupon Bond, Forward Rate and

Short Rate

2.1 T −bond and Zero-coupon(Zero) Rate

A ZCB that pays 1 unit of currency(we will use $ as the currency next) at maturity T
is called a T -bond. The price of a T -bond at time T is denoted by B(t, T ), 0 ≤ t ≤ T . The
zero-coupon rate for time interval [t, T ] is important for us to find the price of a T -bond:

B(t, T ) = e−R(t,T )(T −t) , (2.1)

i.e. the zero-coupon rate is the discount rate for the currency between t and T . Obviously we
have B(T, T ) = 1, ∀T ≥ 0. If we stand at a fixed time t and let T moves forward, we obtain a
mapping T 7→ R(t, T ) which is the term structure of the zero-coupon rate.

2.2 ZCB Price, Forward Rate and Short Rate

We first consider the forward rate for a future time interval [T, T + ∆T ] where ∆T > 0
is a small time increment and we have 0 ≤ t ≤ T ≤ T + ∆T . The continuously compounded
forward rate set at time t for the investment on the future time interval [T, T + ∆T ] is defined
by
log B(t, T + ∆T ) − log B(t, T )
f (t; T, T + ∆T ) = − . (2.2)
∆T
To understand this, we consider the following trading process. At time t, we short one share
B(t,T )
of ZCB B(t, T ) and use the proceeds to long B(t,T +∆T )
shares of ZCB B(t, T + ∆T ). Then at
time T , because we short B(t, T ) at t, we must pay $1 now. And next at time T + ∆T , because
we long B(t,T )
B(t,T +∆T )
shares of B(t, T + ∆T ) at t, we could receive $ B(t,T
B(t,T )
+∆T )
. This is equivalent
to investing $1 at T and receive $ B(t,T
B(t,T )
+∆T )
at T + ∆T . So the forward rate for [T, T + ∆T ]
should satisfy

B(t, T )
1 ef (t;T,T +∆T )∆T = , (2.3)
B(t, T + ∆T )

2
CHAPTER 2. ZERO-COUPON BOND, FORWARD RATE AND SHORT RATE 3

and (2.3) will lead to (2.2).

One fact that should be carefully understood is that although the forward rate f (t; T, T +
∆T ) is the interest rate for future investment(between T and T + ∆T ), it is actually deter-
mined at current time point t instead of some future time point, i.e. f (t; T, T + ∆T ) is F(t)−
measurable.
In order to compute the forward rate for a future time point T , we only need to take limit
as ∆T → 0:
log B(t, T + ∆T ) − log B(t, T ) ∂ log B(t, T )
f (t, T, T ) = lim − =− . (2.4)
∆T →0 ∆T ∂T
This is the instantaneous forward rate for T that is set at t and we simply write it as
∂ log B(t, T )
f (t, T ) = − . (2.5)
∂T
Let θ > t and take integral for both sides of (2.5) with respect to T on interval [t, θ], we obtain
Z θ Z θ
∂ log B(t, T )
f (t, T )dT = − dT = − log B(t, θ) + log B(t, t) = − log B(t, θ). (2.6)
t t ∂T
Let θ = T in (2.6) and it follows that
Z T  Z T 
f (t, u)du = − log B(t, T ) ⇔ B(t, T ) = exp − f (t, u)du . (2.7)
t t

The mapping T 7→ f (t, T ) is the forward rate term structure.

Moreover, in notation f (t, T ) if we let T = t, then the instantaneous forward rate f (t, T )
becomes f (t, t) := R(t) and this is the interest rate of investment at t that we can lock at t.
R(t) is the short rate that is usually assumed to follow the following SDE

dR(t) = µ(t, R(t))dt + σ(t, R(t))dW M (t), (2.8)

where M ∈ {P, Q}. So bond price, forward rate and short rate are highly related, later we will
try to quantitatively find the relationships by probabilistic models.

2.3 Linking Everything Together(Part 1): B(t, T ) and f (t, T )

We consider the one-dimensional case but all the formulae are correct for multi-dimensional
case(after some small and trivial modifications such as changing x2 to ||x||22 when x switches
from a scalar to a vector). Assume that we work under two probability measures(actual measure
P and risk neutral measure Q) and we work with three elements under each measure: bond
price, forward rate and short rate. So in total we should conduct analysis on six elements. Now
assume that under measure P, we have

dR(t) = a(t, T )dt + b(t, T )dW P (t), (2.9)

CHAPTER 2. ZERO-COUPON BOND, FORWARD RATE AND SHORT RATE 4

dB(t, T )
= m(t, T )dt + v(t, T )dW P (t), (2.10)
B(t, T )
df (t, T ) = α(t, T )dt + σ(t, T )dW P (t), (2.11)

where the coefficients such as a(t, T ) are written as the functions of t and T . However, this
does not mean t and T are the only two factors that have impacts on these coefficients. In fact,
we can write R(t) explicitly and get the form like a(t, T, R(t)). But for convenience, we only
write two parameters t and T here.
We now seek a way to convert everything from P to Q or from Q to P. Recall the Gir-
sanov’s theorem and the first fundamental theorem of asset pricing before going further.

(Girsanov’s Theorem) Let W (t), t ∈ [0, T ] be a Brownian motion on (Ω, F, P) and let F(t)
be a filtration generated by the Brownian motion. Assume λ(t), t ∈ [0, T ] is an adapted process.
Define  Z t
1 t 2
Z 
Z(t) = exp − λ(u)dW (u) − λ (u)du (2.12)
0 2 0
and Z t
Q
W (t) = W (t) + λ(u)dW (u). (2.13)
0
Let random variable Z = Z(T ) and define measure Q by
Z
Q(A) = Z(ω)dP(ω), ∀A ∈ F. (2.14)
A

Then EZ = 1 and W Q (t) is a Q−Brownian motion, provided that

Z T
E λ2 (u)Z 2 (u)du < ∞. (2.15)
0

(First Fundamental Theorem of Asset Pricing) A market model is free of arbitrage if it

has a risk-neutral probability measure(this risk-neutral measure do not have to be unique).

Now we continue the exploration in the fixed income world. Define the discount factor as
 Z t 
D(t) = exp − R(u)du (2.16)
0

Changing the Brownian motion WP (t) in (2.9) to (2.11) into WQ (t) according to Girsanov’s
theorem yields
dR(t) = a(t, T )dt + b(t, T )(dW Q (t) − λ(t)dt), (2.17)
dB(t, T )
= m(t, T )dt + v(t, T )(dW Q (t) − λ(t)dt), (2.18)
B(t, T )
df (t, T ) = α(t, T )dt + σ(t, T )(dW Q (t) − λ(t)dt), (2.19)
CHAPTER 2. ZERO-COUPON BOND, FORWARD RATE AND SHORT RATE 5

(A Differential Formula) If we denote the instantaneous forward rate by f (t, T ), then we

have the following equations to compute the differential of bond price B(t, T ):
 Z T  Z T
d − f (t, v)dv = f (t, t)dt − df (t, v)dv (2.20)
t t

and  Z T 
dB(t, T ) = d exp − f (t, v)dv . (2.21)
t
Intuition of the Calculation: Just using the definition of derivative and obtain
 R 
T RT RT
d − t f (t, v)dv f (t + , v)dv − f (t, v)dv
= − lim t+ t
dt →0 
RT RT R t+
t
f (t + , v)dv − t f (t, v)dv − t f (t + , v)dv
= − lim
→0 
RT RT R t+
f (t + , v)dv − f (t, v)dv f (t + , v)dv
= − lim t t
+ lim t
→0  →0 
Z T
f (t + , v) − f (t, v) F (t + , t + ) − F (t + , t)
= − lim dv + lim
→0 t  →0 
Z T
df (t, v) dF (t + , v)
=− dv + lim
t dt →0 dv v=t
Z T
df (t, v)
=− dv + lim f (t + , t)
t dt →0
Z T
df (t, v)
=− dv + f (t, t),
t dt
(2.22)
RT
which means that (2.20) holds. Now let process X(t) = − t f (t, v)dv and function g(x) = ex .
By Ito’s formula, the differential of the bond price is

dB(t, T ) = dg(X(t))
 Z T   Z T 2
1
= B(t, T ) f (t, t)dt − df (t, v)dv + B(t, T ) f (t, t)dt − df (t, v)dv
t 2 t
 Z T  Z T 2
1
= B(t, T ) R(t)dt − P
[α(t, v)dt + σ(t, v)dW (t)]dv + B(t, T ) σ(t, v)dv dt (2.23)
t 2 t
1
= R(t)B(t, T )dt − α̃(t, T )B(t, T )dt − σ̃(t, T )B(t, T )dW P (t) + σ̃ 2 (t, T )B(t, T )dt
  2
1 2
= R(t) − α̃(t, T ) + σ̃ (t, T ) B(t, T )dt − σ̃(t, T )B(t, T )dW P (t),
2
where we have defined Z T
α̃(t, T ) = α(t, v)dv (2.24)
t
and Z T
σ̃(t, T ) = σ(t, v)dv. (2.25)
t
By the first fundamental theorem of asset pricing, we know that our model is free of
arbitrage if we could find a risk neutral measure Q. Finding such a measure is tantamount to
CHAPTER 2. ZERO-COUPON BOND, FORWARD RATE AND SHORT RATE 6

finding a process λ(t) mentioned in the Girsanov’s theorem. The discounted price of a T −bond
is D(t)B(t, T ) and this process should be a Q−martingale which means that the SDE of this
process under Q should have no dt terms(or the price process B(t, T ) under Q should have drift
R(t), both are totally identical requirements). In fact we have

d(D(t)B(t, T )) = B(t, T )dD(t) + D(t)dB(t, T ) + dD(t)dB(t, T )

 
1 2
= − R(t)D(t)B(t, T )dt + D(t)B(t, T ) R(t) − α̃(t, T ) + σ̃ (t, T ) dt − D(t)B(t, T )σ̃(t, T )dW P (t)
2
 
1
=D(t)B(t, T ) −α̃(t, T ) + σ̃ 2 (t, T ) dt − D(t)B(t, T )σ̃(t, T )(dW Q (t) − λ(t)dt)
2
 
1 2
=D(t)B(t, T ) σ̃(t, T )λ(t) − α̃(t, T ) + σ̃ (t, T ) dt − D(t)B(t, T )σ̃(t, T )dW Q (t)
2
(2.26)
For the purpose of removing dt terms, we must require
1
σ̃(t, T )λ(t) − α̃(t, T ) + σ̃ 2 (t, T ) = 0. (2.27)
2
We treat (2.27) as an equation with respect to λ(t) and if it has at least a solution, then this
means that we could find risk neutral measures corresponding to these solutions of process
λ(t). If we fix T , then, obviously, in our one-dimensional model(driven by only one Brownian
motion), equation (2.27) has a solution as we will show next. However, if we let the maturity
T change and consider more than one Brownian motions(the multi-dimensional model), it is
highly possible that we will meet with a system of linear equations which may not always have
solutions, so in this complex situation, the risk neutral measure may not exist and thus no-
arbitrage condition is violated. A short discussion of multi-dimensional model can be seen in
section 2.6 of this chapter.
By taking derivative with respect to T at both sides of (2.27), we have

σ(t, T )λ(t) − α(t, T ) + σ̃(t, T )σ(t, T ) = 0, (2.28)

so the solution is
α(t, T )
λ(t) = − σ̃(t, T ), 0 ≤ t ≤ T. (2.29)
σ(t, T )
Equation (2.28) is the well-known Heath-Jarrow-Morton(HJM) no-arbitrage condition, if it has
at least a solution, then the model presented in (2.9) to (2.11) admits no arbitrage opportunity.
The HJM condition for multi-dimensional case is similar, so we only need to replace scalars in
the one-dimensional case by the corresponding vectors.
Moreover, as we have shown in (2.29), the solution of (2.28) is unique. So not only does the
existence of the risk-neutral measure lead to a no arbitrage model, but also the uniqueness of
the risk-neutral measure causes the market to be complete(by the second fundamental theorem
of asset pricing) and thus every interest rate derivatives can be hedged by constructing hedging
CHAPTER 2. ZERO-COUPON BOND, FORWARD RATE AND SHORT RATE 7

portfolios by trading ZCB(like what we did in hedging an European call option by constructing
portfolios by trading the underlying stock and the bank account). Additionally, later we will also
show that the λ(t) here is just the market price of risk(MPR). To recall the second fundamental
theorem of asset pricing, see next.
(Second Fundamental Theorem of Asset Pricing) A market model is complete if every
derivative security can be hedged. If a market model has a risk-neutral probability measure,
then the model is complete ⇔ the risk-neutral probability measure is unique.
Finally, we should understand that (2.28) is the HJM condition under measure P. In
some applications, we want to utilize the HJM condition under measure Q which is not hard
to derive. What we need to do is changing the α(t, T ) in (2.28) into the drift of forward rate
f (t, T ) under measure Q(sometimes written as αQ (t, T )). (2.11) and (2.19) show that αQ (t, T ) =
α(t, T ) − λ(t)σ(t, T ). So we have σ(t, T )λ(t) − (αQ (t, T ) + λ(t)σ(t, T )) + σ̃(t, T )σ(t, T ) = 0,
which means that αQ (t, T ) = σ̃(t, T )σ(t, T ). This gives us a way to determine the drift of
forward rate under Q to ensure that the model does not violate HJM condition. One possible
advantage of applying HJM condition under Q is that the formula of αQ (t, T ) does not contain
MPR λ(t) but under P the drift α(t, T ) is affected by λ(t).

2.4 Linking Everything Together(Part 2): f (t, T ) and R(t)

Now we delve into the relationship between the short rate and the forward rate. From
(2.11) we can get
Z t Z t
f (t, T ) = f (0, T ) + α(l, T )dl + σ(l, T )dW P (l). (2.30)
0 0

Recall that we have R(t) = f (t, t), so by replacing T in (2.30) by t, we solve the short rate:
Z t Z t
R(t) = f (t, t) = f (0, t) + α(l, t)dv + σ(l, t)dW P (l). (2.31)
0 0

Then we shall find the differential of R(t), so we first rewrite (2.31) as

Z t Z t
R(t) = f (0, t) + α(l, t)dl + σ(l, t)dW P (l)
Z0 t 0
Z t Z t 
∂f (0, s) ∂α(l, s)
= f (0, 0) + ds + α(l, l) + ds dl
0 ∂T 0 l ∂T
Z t Z t 
∂σ(l, s)
+ σ(l, l) + ds dW P (l)
0 l ∂T
Z t Z t Z t Z s  (2.32)
∂f (0, s) ∂α(l, s)
= R(0) + ds + α(l, l)dl + dl ds
0 ∂T 0 0 0 ∂T
Z t Z t Z s 
P ∂σ(l, s) P
+ σ(l, l)dW (l) + dW (l) ds
0 0 0 ∂T
Z t  Z s Z s 
∂ P
= R(0) + f (0, s) + α(l, s)dl + σ(l, s)dW (l) ds
0 ∂T 0 0
CHAPTER 2. ZERO-COUPON BOND, FORWARD RATE AND SHORT RATE 8
Z t Z t
+ α(l, l)dl + σ(l, l)dW P (l)
0 0
Z t Z t Z t
∂
= R(0) + f (s, s)ds + α(l, l)dl + σ(l, l)dW P (l).
0 ∂T 0 0
Then we take differential at both sides of (2.32), the result is
 
∂
dR(t) = f (t, t) + α(t, t) dt + σ(t, t)dW P (t). (2.33)
∂T

In fact, equation (2.33) and (2.10) should be identical, so their dt terms and dW P (t) terms
should be the same respectively, i.e.
∂
a(t, T ) = f (t, t) + α(t, t) (2.34)
∂T
b(t, T ) = σ(t, t). (2.35)

From (2.34) and (2.35) we know that the value of T does not have impacts on a(t, T ) or b(t, T ),
so we can also write them as a(t) and b(t) for convenience.
∂f (0,s)
Remark: We clarify the notations here again to avoid any confusion. For example, ∂T
means the partial derivative of function f (t, T ) with respect to T evaluated at point (0, s).

2.5 Linking Everything Together(Part 3): B(t, T ) and R(t)

We trade stock and bank account to build a self-financing portfolio when we tried to find
the price of European options. Similarly, we want to seek a method to repeat the same routine.
So one may easily come to the idea: we trade interest rate and bank account to replicate a ZCB.
However, problem is not that straightforward since the interest rate is not a financial product
that is publicly traded like equity products(stocks). So we can only trade bank account. Only
trading bank account makes it impossible for us to replicate a ZCB. A solution to this problem
is that we can introduce another ZCB into the replicating portfolio.
Suppose the replicating portfolio has value X(t) at time t. It contains ∆(t) shares of ZCB
B(t, T ) and the remaining value of the portfolio is balanced by bank account G(t). Our target
is replicating ZCB B(t, S) using X(t). The dynamic for X(t) is

dX(t) = ∆(t)dB(t, T ) + R(t)(X(t) − ∆(t)B(t, T ))dt. (2.36)

Insert (2.10) into (2.36) and obtain

dX(t) = [∆(t)B(t, T )m(t, T )+R(t)(X(t)−∆(t)B(t, T ))]dt+∆(t)v(t, T )B(t, T )dW P (t). (2.37)

We also have dynamic for ZCB B(t, S) from (2.10):

dB(t, S) = m(t, S)B(t, S)dt + v(t, S)B(t, S)dW P (t). (2.38)

CHAPTER 2. ZERO-COUPON BOND, FORWARD RATE AND SHORT RATE 9

To replicate B(t, S) by X(t), we must require X(t) = B(t, S) and thus dX(t) = dB(t, S), thus
we have
∆(t)B(t, T )m(t, T ) + R(t)(X(t) − ∆(t)B(t, T )) = m(t, S)B(t, S) (2.39)

and
∆(t)v(t, T )B(t, T ) = v(t, S)B(t, S). (2.40)

Equation (2.40) yield that

v(t, S)B(t, S)
∆(t) = (2.41)
v(t, T )B(t, T )
which is the hedge ratio like the Greek letter ∆ of options. Next, inserting (2.41) and X(t) =
B(t, S) into (2.39) yields that
m(t, T ) − R(t) m(t, S) − R(t)
= . (2.42)
v(t, T ) v(t, S)
The left-hand side of (2.42) is a function of (t, T ) while the right-hand side is a function of
(t, S), but they are equal. The only possibility is that both sides are functions of t and do not
depend on T or S. Let
m(t, T ) − R(t) m(t, S) − R(t)
= = θ(t). (2.43)
v(t, T ) v(t, S)
The process θ(t) is called the market price of risk(MPR). In the following analysis we will see
the process θ(t) is just the process λ(t) mentioned in section 2.3 that we used to change the
measure from P to Q.
Since the interest rate model (2.9) has Markovian solutions, the ZCB price must be a
function of t and r(t), so there exist a Borel measurable function p(t, x) such that B(t, T ) =
p(t, R(t)). Then by Ito’s formula we obtain
∂p ∂p 1 ∂ 2p
dB(t, T ) = dt + dR(t) + 2
dR(t)dR(t)
∂t ∂x 2 ∂x
(2.44)
∂p 1 2 ∂ 2 p
 
∂p ∂p
= + a(t) + b (t) 2 dt + b(t) dW P (t).
∂t ∂x 2 ∂x ∂x
Under the risk neutral measure Q, the discounted process of B(t, T ) should be a Q−martingale.
Let  Z t 
D(t) = exp − R(u)du . (2.45)
0
be the discount factor and thus
∂p 1 2 ∂ 2 p
 
∂p
d(D(t)B(t, T )) =D(t) −R(t)B(t, T ) + + (a(t) − λ(t)b(t)) + b (t) 2 dt
∂t ∂x 2 ∂x (2.46)
∂p
+ D(t)b(t) dW Q (t).
∂x
D(t)B(t, T ) is a Q−martingale when the dt terms in (2.46) is 0, i.e.
∂p ∂p 1 2 ∂ 2 p
−R(t)B(t, T ) + + (a(t) − λ(t)b(t)) + b (t) 2 = 0. (2.47)
∂t ∂x 2 ∂x
CHAPTER 2. ZERO-COUPON BOND, FORWARD RATE AND SHORT RATE 10

Since we have B(t, T ) = p(t, R(t)), (2.47) can be converted into

∂p ∂p 1 ∂ 2p
−R(t)p(t, R(t)) + (t, R(t)) + (a(t) − λ(t)b(t)) (t, R(t)) + b2 (t) 2 (t, R(t)) = 0. (2.48)
∂t ∂x 2 ∂x
(2.48) holds for any R(t) so it must hold for any positive real number x ∈ R+ . In other words,
the function p(t, x) is the solution of the following partial differential equation
∂p ∂p 1 ∂ 2p
−xp(t, x) + (t, x) + (a(t) − λ(t)b(t)) (t, x) + b2 (t) 2 (t, x) = 0. (2.49)
∂t ∂x 2 ∂x
In section 2.3 we explained that we do not explicitly write the R(t) in the coefficients. For
example, we write a(t, T, R(t)) as a(t, T ). And in section 2.4 we showed that a and b do not
depend on T , then we write them as a(t) and b(t) respectively. But we must understand that
a(t) and b(t) are actually a(t, R(t)) and b(t, R(t)). So the final version of PDE (2.49) is
∂p ∂p 1 ∂ 2p
−xp(t, x) + (t, x) + (a(t, x) − λ(t)b(t, x)) (t, x) + b2 (t, x) 2 (t, x) = 0. (2.50)
∂t ∂x 2 ∂x
PDE (2.50) is pricing PDE of ZCB and is also called the term structure PDE. Since the value
of a ZCB at the maturity date is 1, we can also add the terminal condition here: p(T, x) =
1( because p(T, R(t)) = B(T, T ) = 1).
Now let us study the relationship between θ(t) and (2.50). Before we go further, recall the
one-dimensional Feynman-Kac’s theorem first.

(One-dimensional Feynman-Kac’s Theorem) Suppose we are in probability space (Ω, F, M),

where M is the probability measure. Let X(t) be an Ito diffusion process given by

dX(t) = µ(t, X(t))dt + σ(t, X(t))dW (t) (2.51)

with generator
∂ 1 ∂2
+ σ 2 (t, x) 2
A = µ(t, x) (2.52)
∂x 2 ∂x
and initial condition X t,x (t) = x. Let function ψ(t, x), (t, x) ∈ [0, +∞) × R and function
u(t, x), (t, x) ∈ [0, +∞) × R which satisfies the PDE

 ∂u (t, X(t)) + Au(t, X(t)) − ψ(t, X(t))u(t, X(t)) = 0

∂t
. (2.53)
 u(T, X(T )) = f (X(T ))


Then we have h RT i
u(t, x) = EM e− t ψ(s,X(s))ds f (X(T ))|Xt = x
h RT i . (2.54)
= EM e− t ψ(s,X(s))ds f (X t,x (T ))
It should be noticed that if we choose X(t) to be the short rate process R(t), u(t, x) to be p(t, x),
ψ(t, x) to be x and f (x) to be 1(but we do not say what is the generator A now, because we
CHAPTER 2. ZERO-COUPON BOND, FORWARD RATE AND SHORT RATE 11

will discuss this later in this section), then by Feynman-Kac’s theorem, we have the following
observations: if function p(t, x) satisfies the PDE

 ∂p (t, R(t)) + Ap(t, R(t)) − R(t)p(t, R(t)) = 0

∂t
. (2.55)
 p(T, R(T )) = 1


Then we have
h RT i
p(t, x) = EM e− t R(s)ds |R(t) = x . (2.56)
or we can equivalently write
h RT i h RT i
− t R(s)ds − t R(s)ds
M
p(t, R(t)) = E e |R(t) = E e
M
|F(t) . (2.57)

We extremely hope that PDE (2.55) is the same as PDE (2.50)(after we change the p(t, x) in
(2.50) into p(t, R(t))), because if they are identical, then the solution of (2.55) is just the price
of ZCB and also, the value of the conditional expectation in (2.57) will also equal to the price
of the ZCB. Notice that this conditional expectation is computed under measure M, but later
we will show what M is. Now we take generator A to be
∂ 1 ∂2
A = (a(t, x) − λ(t)b(t, x)) + b2 (t, x) 2 . (2.58)
∂x 2 ∂x
In this way, (2.55) will be equivalent to (2.50). We see (2.51) and (2.52) to obtain the way in
which we can write down the SDE of X(t) by generator A. Similarly, we can write down the
SDE of R(t) under measure M by (2.58), which is

dR(t) = (a(t, x) − λ(t)b(t, x))dt + b(t, x)dW M (t), (2.59)

where W M (t) is the Brownian motion under measure M. We have selected an excellent generator
A to make p(t, x) represent the ZCB price at time t given the initial short rate R(t) = x. Next,
we focus on (2.57). Using the discount factor in (2.45) and the fact that p(T, R(T )) = 1, we
can rewrite (2.57) as

D(t)p(t, R(t)) = EM [D(T )p(T, R(T ))|F(t)] . (2.60)

(2.60) tells us that the discounted ZCB price under measure M is a martingale. So we know
that measure M is just the risk neutral measure Q. And from (2.59) we obtain the short rate
SDE under Q:
dR(t) = (a(t, x) − λ(t)b(t, x))dt + b(t, x)dW Q (t), (2.61)
Apparently, (2.61) agrees with (2.17), so currently, everything works well.
Finally, we start from (2.10) to see the discounted ZCB price under measure Q:
d(D(t)B(t, T )) = D(t)B(t, T )[−R(t) + m(t, T )]dt + D(t)B(t, T )v(t, T )dW P (t)
= D(t)B(t, T )[−R(t) + m(t, T )]dt + D(t)B(t, T )v(t, T )d[W Q (t) − λ(t)dt]
= D(t)B(t, T )[−R(t) + m(t, T ) − λ(t)v(t, T )]dt + D(t)B(t, T )v(t, T )dW Q (t).
(2.62)
CHAPTER 2. ZERO-COUPON BOND, FORWARD RATE AND SHORT RATE 12

Again, because D(t)B(t, T ) is a Q−martingale, the dt terms in (2.62) should be 0, i.e.

m(t, T ) − R(t)
−R(t) + m(t, T ) − λ(t)v(t, T ) = 0 ⇒ λ(t) = . (2.63)
v(t, T )
Through comparing (2.63) and (2.43), the MPR process θ(t) must be the same as λ(t). So the
MPR process is just the process λ(t) that establishes a bridge between measure P and measure
Q. As a results, the ZCB price process under Q is represented by
dB(t, T )
= R(t)dt + v(t, T )dW Q (t) (2.64)
B(t, T )
which has a drift R(t) as we can expect.

2.6 A Simple Multi-dimensional Discussion of HJM Condition

In section 2.3 we have discussed the HJM no-arbitrage condition(see (2.28)). But this is
just the one-dimensional case. There are two sources of increasing dimensions of this problem.
First, the forward rate could be driven by more than one independent Brownian motions:
d
X
df (t, T ) = α(t, T )dt + σj (t, T )dWjP (t). (2.65)
j=1

In (2.65) we have d(d ≥ 2) Brownian motions in total while previously, in (2.11) we merely
introduced one Brownian motion in our model. Secondly, we never changed T . However, in
financial market we have forward rates with different values of T , for example T1 , T2 , ..., TK . This
means that we could generate K equations that represent HJM condition and each equation
has d independent Brownian motions. We will show these two sources in detail next and write
down the multi-dimensional version of HJM condition.
We use the same approach as what is shown in (2.20) to (2.29). The differential of the
integral of forward rate is
 Z T 
d − f (t, v)dv
t
Z T
=R(t)dt − df (t, v)dv
t
" d
#
Z T X
=R(t)dt − α(t, v)dt + σj (t, v)dWjP (t) dv
t j=1
(2.66)
Z T  d Z
X T 
=R(t)dt − α(t, v)dv dt − σj (t, v)dv dWjP (t)
t j=1 t

d
X
=R(t)dt − α̃(t, T )dt − σ̃j (t, T )dWjP (t),
j=1

where we define Z T
α(t, v)dv = α̃(t, T ) (2.67)
t
CHAPTER 2. ZERO-COUPON BOND, FORWARD RATE AND SHORT RATE 13

and Z T
σ(t, v)dv =σ̃(t, T ) = (σ̃1 (t, T ), σ̃2 (t, T ), ..., σ̃d (t, T ))>
t
Z T Z T Z T > (2.68)
= σ1 (t, v)dv, σ2 (t, v)dv, ..., σd (t, v)dv .
t t t

(2.66) also yields that

2 " d
#2
  Z T X
d − f (t, v)dv = R(t)dt − α̃(t, T )dt − σ̃j (t, T )dWjP (t)
t j=1
" d
#2
X
= σ̃j (t, T )dWjP (t)
j=1 (2.69)
d
X
= σ̃j2 (t, T )dt
j=1

= ||σ̃(t, T )||2 dt.

Then we compute the differential of B(t, T ) under Q by Ito’s formula:

 Z T 
dB(t, T ) =d exp − f (t, v)dv
t
 Z T    Z T 2
1
=B(t, T )d − f (t, v)dv + B(t, T ) d − f (t, v)dv
t 2 t
" d
#
X 1
=B(t, T ) R(t) − α̃(t, T ) − σ̃j (t, T )dWj (t) + B(t, T )||σ̃(t, T )||2 dt
P

j=1
2
  d
1 2
X  
=B(t, T ) R(t) − α̃(t, T ) + ||σ̃(t, T )|| dt − B(t, T ) σ̃j (t, T ) dWjQ (t) − λj (t)dt
2 j=1
" d
#
1 X
=B(t, T ) R(t) − α̃(t, T ) + ||σ̃(t, T )||2 + σ̃j (t, T )λj (t) dt
2 j=1
d
X
− B(t, T ) σ̃j (t, T )dWjQ (t).
j=1
(2.70)
The drift of the above SDE should be R(t) so we require
d
1 X
−α̃(t, T ) + ||σ̃(t, T )||2 + σ̃j (t, T )λj (t) = 0. (2.71)
2 j=1

We then take derivative with respect to T at both sides of (2.71):

d
X d
X
− α(t, T ) + σ̃j (t, T )σj (t, T ) + σj (t, T )λj (t) = 0
j=1 j=1
d d
(2.72)
X X
⇔ σj (t, T )λj (t) = α(t, T ) − σ̃j (t, T )σj (t, T ).
j=1 j=1
CHAPTER 2. ZERO-COUPON BOND, FORWARD RATE AND SHORT RATE 14

This is the HJM condition for a single date T . It is actually a linear equation with respect to
λ(t) = (λ1 (t), λ2 (t), ..., λd (t))> . If we consider different values of date T : T1 , T2 , ..., TK , then we
have a system of linear equations:

Pd Pd
 j=1 σj (t, T1 )λj (t) = α(t, T1 ) − j=1 σ̃j (t, T1 )σj (t, T1 )




 Pd σj (t, T2 )λj (t) = α(t, T2 ) − Pd σ̃j (t, T2 )σj (t, T2 )


j=1 j=1
. (2.73)
...






 Pd

σj (t, TK )λj (t) = α(t, TK ) −
Pd
σ̃j (t, TK )σj (t, TK )
j=1 j=1

This can be further written into a matrix form

    Pd 
σ1 (t, T1 ) σ2 (t, T1 ) ... σd (t, T1 ) λ1 (t) α(t, T1 ) − j=1 σ̃j (t, T1 )σj (t, T1 )
    
 σ (t, T ) σ2 (t, T2 ) ... σd (t, T2 )    α(t, T ) − Pd σ̃ (t, T )σ (t, T ) 
  λ2 (t)

 1 2 2 j=1 j 2 j 2
= .
  
 

 ... ... ... ...   ...

 
  ... 

Pd
σ1 (t, TK ) σ2 (t, TK ) ... σd (t, TK ) λd (t) α(t, TK ) − j=1 σ̃j (t, TK )σj (t, TK )
(2.74)
As a result, the risk neutral measure Q

• does not exist if (2.74) has no solution;

• exists and is unique if (2.74) has a unique solution;

• exists and is not unique if (2.74) has more than one solutions.
Chapter 3

Short Rate Models and Affine Methods

for ZCB Pricing

3.1 Stochastic Differential Equations of Short Rate

In (2.9) and (2.17) we have shown the general form of short rate models under measure P
and Q respectively without giving the formulae of a(t, T ) and b(t, T ). Now we shall try to give
these formulae and see the bond price that we could obtain in these models in the next section.
All the stochastic differential equations of short rate R(t) are give under risk neutral measure
Q in this section and Girsanov’s theorem can convert this to the version under measure P if
necessary. Now we list some of the well-known short rate models below.

• Vasicek model: dR(t) = (a − bR(t))dt + σdW Q (t);

• Hull-White(extended Vasicek model): dR(t) = (a(t) − b(t)R(t))dt + σ(t)dW Q (t);

• Cox-Ingersoll-Ross(CIR) model: dR(t) = (a − bR(t))dt + σ R(t)dW Q (t);

p

• Hull-White(extended CIR model): dR(t) = (a(t) − b(t)R(t))dt + σ(t) R(t)dW Q (t);

p

• Ho-Lee model: dR(t) = a(t)dt + σdW Q (t).

Remark: In all the stochastic differential equations above, a, b, σ are constants but a(t), b(t), σ(t)
are nonrandom functions.

3.2 Affine Solutions of ZCB Price

Under the risk neutral measure Q, the bond price B(t, T ) = p(t, R(t)) is a martingale, so

D(t)B(t, T ) = EQ [D(T )B(T, T )|F(t)]

(3.1)
= EQ [D(T )|F(t)].

This leads to the risk neutral pricing formula of zero-coupon bond:

  Z T  
1 Q
B(t) = E [D(T )|F(t)] = E exp −
Q
R(s)ds F(t) . (3.2)
D(t) t

15
CHAPTER 3. SHORT RATE MODELS AND AFFINE METHODS FOR ZCB PRICING 16

Suppose we use the short rate model

dR(t) = ψ(t, R(t))dt + σ(t, R(t))dW Q (t). (3.3)

By the same arguments in chapter 2, with the support of Feynman-Kac formula, we obtain the
partial differential equation of bond price p(t, x):
∂p ∂p 1 ∂ 2p
−xp(t, x) + (t, x) + ψ(t, x) (t, x) + σ 2 (t, x) 2 (t, x) = 0, (3.4)
∂t ∂x 2 ∂x
with the terminal condition p(T, x) = 1 for any x ≥ 0.
For some short rate models, we could derive a closed-form solution by trying a possible
ansatz. If we try an ansatz of the form

p(t, x) = exp (−A(t, T ) − C(t, T )x) , (3.5)

then we say that we are solving an affine model. We insert (3.5) into (3.4) and it follows that
d d 1
−x − A(t, T ) − x C(t, T ) − ψ(t, x)C(t, T ) + σ(t, T )2 C 2 (t, T ) = 0, (3.6)
dt dt 2
with terminal condition A(T, T ) = C(T, T ) = 0.

3.2.1 Affine Model Solution for Ho-Lee Short Rate Model

If we assume that the short rate behaves like Ho-Lee model, then (3.6) becomes
d d 1
−x − A(t, T ) − x C(t, T ) − a(t)C(t, T ) + σ 2 C 2 (t, T ) = 0. (3.7)
dt dt 2
Because (3.7) holds for all possible values of x, we must have
d
C(t, T ) + 1 = 0, C(T, T ) = 0 (3.8)
dt
and
d 1
− A(t, T ) − a(t)C(t, T ) + σ 2 C 2 (t, T ) = 0, A(T, T ) = 0. (3.9)
dt 2
Solving ordinary differential equations (3.8) and (3.9) yields

C(t, T ) = T − t (3.10)

and Z T
1
A(t, T ) = − σ 2 (T − t)3 + a(s)(T − s)ds, (3.11)
6 t
for any 0 ≤ t ≤ T .

3.2.2 Affine Model Solution for Hull-White(Extended Vasicek) Short Rate Model

The Ho-Lee model is relatively simple so the derivation of (3.10) and (3.11) is not so hard.
Now we assume that the short rate follows Hull-White(extended Vasicek) model and show the
affine solution of bond price for this short rate assumption. In this case (3.6) is equivalent to
d d 1
−x − A(t, T ) − x C(t, T ) − [a(t) − b(t)x]C(t, T ) + σ 2 (t)C 2 (t, T ) = 0. (3.12)
dt dt 2
CHAPTER 3. SHORT RATE MODELS AND AFFINE METHODS FOR ZCB PRICING 17

Then we write two ordinary differential equations

d
−1 − C(t, T ) + b(t)C(t, T ) = 0 (3.13)
dt
and
d 1
− A(t, T ) − a(t)C(t, T ) + σ 2 (t)C 2 (t, T ) = 0. (3.14)
dt 2
Let  Z s 
Db (s) = exp − b(v)dv . (3.15)
0
Consider the differential
d d
[Db (s)C(s, T )] = − b(s)Db (s)C(s, T ) + Db (s) C(s, T )
ds  dt 
d (3.16)
= − Db (s) b(s)C(s, T ) − C(s, T )
dt
= − Db (s).
Integrating (3.16) from s = t to s = T gives
Z T Z T  Z s 
−Db (t)C(t, T ) = − Db (s)ds ⇒ C(t, T ) = exp − b(v)dv ds. (3.17)
t t t

Finally, by integrating both sides of (3.14) from u = t to u = T , we obtain

Z T 
1 2 2
A(t, T ) = a(u)C(u, T ) − σ (u)C (u, T ) du. (3.18)
t 2
In summary, it is not difficult to understand how to derive these two ordinary differential
equations. The problem is how to solve them: for complex short rate models, these ordinary
differential equations are exceedingly intractable. Therefore, numerical methods should be
applied to rescue the calculation process. A closed-form solution could also be derived when
the short rate follows the Cox-Ingersoll-Ross model, but this derivation is more difficult than
the two examples we discussed above. We skip it in this material.

3.3 Model Calibration

In this section we briefly discuss the calibration of short rate model parameters. We show
this problem by two examples.

3.3.1 Example 1: Ho-Lee Model Calibration

Consider the Ho-Lee model for short rate:

dR(t) = a(t)dt + σdW Q (t), (3.19)

where σ is a given constant and a(t) is an unknown function. The target is finding the function
a(t). The ZCB price B(t, T ) given by this model is

B(t, T ) = exp (−A(t, T ) − C(t, T )R(t)) , (3.20)

CHAPTER 3. SHORT RATE MODELS AND AFFINE METHODS FOR ZCB PRICING 18

where Z T
1
A(t, T ) = a(u)(T − u)du − σ 2 (T − t)3 , (3.21)
t 6
and
C(t, T ) = T − t. (3.22)

So the forward rate is

∂ ∂
f (t, T ) = − log B(t, T ) = [A(t, T ) + C(t, T )R(t)]
∂T ∂T
Z T (3.23)
1
= a(u)du − σ 2 (T − t)2 + R(t).
t 2
Next, we observe the forward rate curve f ∗ (0, T ) from the real financial market. We should
notice that here T is not fixed because f ∗ (0, T ) is a forward curve and it should contain many
values of forward rates that have different dates T . Our goal is making the theoretical forward
curve f (0, T ) matches with the observed forward rate curve f ∗ (0, T ), i.e.
Z T
∗ 1
f (0, T ) = a(u)du − σ 2 T 2 + R(0). (3.24)
0 2
Taking derivative with respect to T on both sides of (3.24), we have
∂ ∗
f (0, T ) = a(T ) − σ 2 T. (3.25)
∂T
For any T ≥ 0, we can find a(T ) by (3.25) and thus we complete the calibration of a(t):
∂ ∗
a(t) = f (0, t) + σ 2 t, t ≥ 0. (3.26)
∂T

3.3.2 Example 2: Hull-White(Extended Vasicek) Model Calibration

Consider the Hull-White(extended Vasicek) model for short rate:

dR(t) = (θ(t) − ar(t))dt + ηdW Q (t), (3.27)

where a and η are given constants and θ(t) is an unknown function which we plan to obtain.
Assume that the ZCB price has the following form

B(t, T ) = exp (−A(t, T ) − C(t, T )R(t)) . (3.28)

By the same approach in section 3.2 we have

1
C(t, T ) = [1 − exp (−a(T − t))] (3.29)
a
and Z T Z T
1
A(t, T ) = θ(u)C(u, T )du − η 2 C 2 (u, T )du. (3.30)
t 2 t

According to the relationship f (t, T ) = − ∂ log∂T

B(t,T )
and after a tedious calculation, we obtain
the SDE of f (t, T ):
η2
df (t, T ) = [exp (−a(T − t)) − exp (−2a(T − t))] dt + η exp (−a(T − t)) dW Q (t). (3.31)
a
CHAPTER 3. SHORT RATE MODELS AND AFFINE METHODS FOR ZCB PRICING 19

We can prove that (3.31) satisfy the HJM condition. Indeed, we have
Z T
η2
[exp (−a(T − t)) − exp (−2a(T − t))] = η exp (−a(T − t)) η exp (−a(u − t)) du. (3.32)
a t

This is just the HJM condition equation under measure Q. Taking integral at both sides of
(3.31) from u = 0 to u = t yields
Z t 2
η
f (t, T ) =f (0, T ) + [exp (−a(T − u)) − exp (−2a(T − u))] du+
0 a
Z t
η exp (−a(T − u)) dW Q (u)
0
η2
 
1 1
=f (0, T ) + 2 exp(−a(T − t)) − exp(−aT ) − exp(−2a(T − t)) + exp(−2aT ) +
a 2 2
Z t
η exp (−a(T − u)) dW Q (u).
0
(3.33)
Let T = t in (3.33) and obtain
η2 1
 
1
R(t) =f (t, t) = f (0, t) + 2 − exp(−at) + exp(−2at) +
a 2 2
Z t
η exp (−a(t − u)) dW Q (u)
0
Z t (3.34)
η2 2
=f (0, t) + 2 [1 − exp(−at)] + η exp (−a(t − u)) dW Q (u)
2a 0
2 Z t
η
=f (0, t) + C 2 (0, t) + η exp (−a(t − u)) dW Q (u).
2 0

Suppose f ∗ (0, T ), T ≥ 0 be the forward rate curve observed in the real financial market. So
we should choose an optimal function θ(t) to make the theoretical forward rate f (0, t) matches
with the real forward rate f ∗ (0, t):
t
η2
Z
∗
R(t) = f (0, t) + C 2 (0, t) + η exp (−a(t − u)) dW Q (u). (3.35)
2 0

(3.35) can be treated as an equation with respect to θ(t) because R(t) is a function of θ(t). The
solution of (3.35) is the calibration result of θ(t). To get the solution, we can take differential
for (3.35):
Z t
∂f ∗ (0, t)
 
2 ∂C(0, t)
dR(t) = dt + η C(0, t) dt + η −a exp(−at) Q Q
exp(au)dW (u)dt + dW (t)
∂T ∂T 0
Z t
∂f ∗ (0, t)
 
2 ∂C(0, t)
= dt + η C(0, t) dt + η −a exp(−at) exp(au)dW (u)dt + ηdW Q (t)
Q
∂T ∂T 0
∂f ∗ (0, t) η2 2
 
2 ∂C(0, t) ∗
= dt + η C(0, t) dt − a R(t) − f (0, t) − C (0, t) dt + ηdW Q (t)
∂T ∂T 2
 ∗
aη 2 2

∂f (0, t) 2 ∂C(0, t) ∗
= + η C(0, t) + af (0, t) + C (0, t) − aR(t) dt + ηdW Q (t).
∂T ∂T 2
(3.36)
CHAPTER 3. SHORT RATE MODELS AND AFFINE METHODS FOR ZCB PRICING 20

The drift of SDE (3.36) should be θ(t) − aR(t) so the calibration of θ(t) is
∂f ∗ (0, t) ∂C(0, t) aη 2 2
θ(t) = + η 2 C(0, t) + af ∗ (0, t) + C (0, t), t ≥ 0. (3.37)
∂T ∂T 2
There are many other methods to conduct model calibration such as methods from opti-
mization, data science and machine learning. These approaches are useful especially for models
that have a large number of parameters and are hard to tackle by humans. These contents are
not consistent with the main purpose of this material. Readers may see other textbooks or
research papers for more about model calibration.
Chapter 4

LIBOR and Forward LIBOR

4.1 Martingale Representation Theorem

In the option pricing model, we tried to build a portfolio X(t) to replicate the option price
and thus evaluate the price of the option. But why does this replication exist? Although we can
give the Black-Scholes formula without explicitly show this existence, we need to understand
this problem in depth. The martingale representation theorem(MPT) ensures the finding of
replication of an option to be successful so we spend some time to look at it. Moreover, in
the analysis of LIBOR rate, MPT also plays an important role and can generate splendid results.

(Martingale Representation Theorem) Assume that W (t) is a Brownian motion and F(t)
is the filtration generated by W (t), where 0 ≤ t ≤ T . Let M (t), 0 ≤ t ≤ T be a martingale
that satisfies

• M (t) is F(t)−measurable;

• E[M (t)|F(s)] = M (s), ∀0 ≤ s ≤ t ≤ T .

Then there exists an adapted process Γ(t), 0 ≤ t ≤ T such that

Z t
M (t) = M (0) + Γ(u)dW (u), 0 ≤ t ≤ T (4.1)
0

or we can also informally write

dM (t) = Γ(t)dW (t). (4.2)

The theorem shows that the collection of all the Ito integrals with respect to a Brownian mo-
tion path W (t) is very “dense”: every martingale with respect to the filtration generated by
W (t) can be represented by an Ito integral with respect to W (t)(plus an initial value). In
other words, given a Brownian motion W (t) and its filtration F(t), we can construct all the
martingales that are F(t)−measurable.

21
CHAPTER 4. LIBOR AND FORWARD LIBOR 22

(Martingale Representation Theorem with Brownian motion under Q) Assume that

W (t) is a Brownian motion and F(t) is the filtration generated by W (t), where 0 ≤ t ≤ T . We
could define the risk neutral measure Q as we did in discussing the Girsanov’s theorem. Next,
let M̃ (t), 0 ≤ t ≤ T be a Q−martingale. Then there exists an adapted process Γ̃(t), 0 ≤ t ≤ T
such that Z t
M̃ (t) = M̃ (0) + Γ̃(u)dW Q (u), 0 ≤ t ≤ T (4.3)
0
or we can also informally write
dM̃ (t) = Γ̃(t)dW Q (t). (4.4)

Notice that the process in this theorem is F(t)−adapted and F(t) is the filtration generated
by W (t) instead of W Q (t).
Now let us consider replicate an European option value V (t) by trading stock S(t) and
bank account G(t). The replication portfolio value process X(t) is self-financing. Assume
that we hold ∆(t) shares of stock at time t in the portfolio X(t) with the remaining value
of the portfolio balanced by the bank account. Then we write down the following SDEs(let
D(t) = e−rt be the discount factor):

dS(t) = rS(t)dt + σS(t)dW Q (t), (4.5)

dG(t) = rG(t)dt, (4.6)

and
d(D(t)X(t)) = −rD(t)X(t)dt + D(t)[∆(t)dS(t) + r(X(t) − ∆(t)S(t))dt]
(4.7)
= σ∆(t)D(t)S(t)dW Q (t).
Because X(t) replicates V (t), we have

d(D(t)V (t)) = d(D(t)X(t)). (4.8)

In addition, the discounted value of the option should be a Q−martingale, so by the MPT, we
know that there exists some adapted process Γ̃(t), such that

d(D(t)V (t)) = Γ̃(t)dW Q (t). (4.9)

We combine (4.5) to (4.7) and obtain

Γ̃(t)
∆(t) = . (4.10)
σD(t)S(t)
Now that we explicitly acquire the formula of ∆(t), we are confident that the replicating port-
folio exists. This is guaranteed by the existence of process Γ̃(t).
CHAPTER 4. LIBOR AND FORWARD LIBOR 23

4.2 Stochastic Representation of Assets and T −forward Measure

In this section we will discuss the stochastic representation of asset price processes under
risk neutral measure Q with the help of MRT introduced in 4.1. Then we change our study
from measure Q to a new probability measure QN which is the risk neutral measure when we
use asset N (t) as the numeraire.
If we choose a T −bond as the numeraire, i.e. let N (t) = B(t, T ), the corresponding risk
neutral measure is frequently called the T −forward measure(denoted by QT ). This measure is
extremely important not only in the pricing of some complicated derivatives but also in building
suitable models for forward LIBOR. So together with section 4.1, section 4.2 shows the basic
preliminaries that are necessary to understand models for forward LIBOR and Black caplet
formula.
Define the discount factor by
 Z t 
D(t) = exp − R(u)du . (4.11)
0

We consider asset N (t) whose price SDE is driven by d Brownian motions under risk neu-


tral measure Q: W Q (t) = W1Q (t), W2Q (t), ..., WdQ (t) . Then we know that its discounted
process D(t)N (t) is a Q−martingale. By the MRT(multi-dimensional MRT which we do
not write in 4.1, but the form is similar), there exists a multi-dimensional process v ∗ (t) =
(v1 (t), v2 (t), ..., vd (t)) such that
d
X
∗
d(D(t)N (t)) = v (t) · dW (t) =
Q
vi∗ (t)dWiQ (t), (4.12)
i=1

where the notation x · y is the inner product of vector x and y. Obviously in financial market
the price of asset and the discount factor is positive, so D(t)N (t) 6= 0. Then we can let
v ∗ (t)
v(t) = , (4.13)
D(t)N (t)
and thus write the price dynamic as

d(D(t)N (t)) = D(t)N (t)v(t) · dW Q (t). (4.14)

This is the so-called stochastic representation of asset N (t) under measure Q. Or by Ito’s
formula we can equivalently convert (4.14) into other formats. For example, we have

dN (t) = R(t)N (t)dt + N (t)v(t) · dW Q (t), (4.15)

and Z t   Z t 
1 2
N (t) = N (0) exp R(s) − ||v(s)||2 ds + v(s) · dW (s) .
Q
(4.16)
0 2 0
To intuitively understand (4.16), we can use a trick to reduce the number of Brownian motions
in (4.15) from d to 1. First we show a formula that can merge more than one Brownian motions
into a single Brownian motion.
CHAPTER 4. LIBOR AND FORWARD LIBOR 24

(Merge Formula of Brownian Motions) Let real numbers e1 , e2 , ..., ed ∈ [0, 1] and
d
X
e2i = 1. (4.17)
i=1

Assume W1 (t), W2 (t), ..., Wd (t) are d independent Brownian motions, and define
d
X
W (t) = ei Wi (t). (4.18)
i=1
Pd
Fake Proof: Obviously, we know that the initial value of W (t) is W (0) = i=1 ei Wi (0) = 0.
W (t) also has continuous path and it is a martingale(a linear combination of Brownian motions
is a martingale). Finally, we have
d
! d
! d
!
X X X
dW (t)dW (t) = d ei Wi (t) d ej Wj (t) = e2i dt = dt. (4.19)
i=1 j=1 i=1

By Levy’s theorem, we can prove that W (t) is an one-dimensional Brownian motion.

Next, let
v
q u d
uX
σ(t) = ||v(t)||2 = v12 (t) + v22 (t) + ... + vd2 (t) = t vi2 (t). (4.20)
i=1

Then (4.15) is equivalent to

dN (t) = R(t)N (t)dt + N (t)[v1 (t)dW1Q (t) + v2 (t)dW2Q (t) + ... + vd (t)dWdQ (t)]
v1 (t)dW1Q (t) + v2 (t)dW2Q (t) + ... + vd (t)dWdQ (t)
= R(t)N (t)dt + N (t)σ(t)
σ(t)
! (4.21)
v1 (t)W1Q (t) + v2 (t)W2Q (t) + ... + vd (t)WdQ (t)
= R(t)N (t)dt + N (t)σ(t)d
σ(t)

= R(t)N (t)dt + N (t)σ(t)dH Q (t),

where we let
v1 (t)W1Q (t) + v2 (t)W2Q (t) + ... + vd (t)WdQ (t)
Q
H (t) = , (4.22)
σ(t)
and H Q (t) is a Q−Brownian motion by the merge formula of Brownian motions(notice that
Pd  vi (t) 2
i=1 σ(t) = 1). (4.21) is the so-called generalized geometric Brownian motion, so its
solution can be written explicitly as
Z t   Z t 
1 2
N (t) = N (0) exp R(s) − σ (s) ds + Q
σ(s)dH (s)
0 2 0
Z t   Z t 
1 2
 Q Q Q

= N (0) exp R(s) − ||v(s)||2 ds + v1 (s)W1 (s) + v2 (s)W2 (s) + ... + vd (s)Wd (s) ds
0 2 0
Z t   Z t 
1 2
= N (0) exp R(s) − ||v(s)||2 ds + v(s) · dW (s)
Q
0 2 0
(4.23)
CHAPTER 4. LIBOR AND FORWARD LIBOR 25

which is just (4.16). This fake proof is just assisting readers in understanding (4.16) from a
view of decreasing the number of Brownian motions. Some of the derivations in the fake proof
is problematic. The real proof can be obtained by Ito’s formula which we skip here since it is
straightforward.
Next we consider the change of measure from Q to QN (like what we did in the past from P
to Q). We start from the probability space (Ω, F, Q). In Girsanov’s theorem(multi-dimensional
case, similar to the one-dimensional case in Chapter 2, the only difference is changing the scalars
to vectors), let λ(t) = −v(t), 0 ≤ t ≤ T , then we obtain process
Z t
1 t
Z 
2
Z(t) = exp v(s) · dW (s) −
Q
||v(s)||2 ds , (4.24)
0 2 0
and the QN −Brownian motion
Z t
QN
W (t) = W (t) −
Q
v(s)ds. (4.25)
0

Set Z = Z(T ) as the Radon-Nikodym derivative, we have

Z
N
Q (A) = Z(ω)dQ(ω), ∀A ∈ F. (4.26)
A

Under the assistance of properties of conditional expectation, we can quickly show the following
two results.

• Result 1: For any F(t)−measurable random variable Y , we have EQ [Y ] = EQ [Y Z(t)], 0 ≤

N

t ≤ T.

• Result 2: Let 0 ≤ s ≤ t ≤ T and Y is an F(t)−measurable random variable. Then we

have
N 1
EQ [Y |F(s)] = EQ [Y Z(t)|F(s)]. (4.27)
Z(s)
Proof for the above two results are straightforward and so are skipped here.
When we want to find the price at time t of a derivative security V (t), 0 ≤ t ≤ T , if we
only do this work under the risk neutral measure Q, then we must compute
1 Q
V (t) = E [D(T )V (T )|F(t)]. (4.28)
D(t)
This is easy in the past when we were deriving the Black-Scholes formula for European options
because we assumed that the short rate is a constant, i.e. R(t) = r. Therefore (4.28) becomes

V (t) = e−r(T −t) EQ [V (T )|F(t)]. (4.29)

But for advanced interest rate derivatives, we have to deal with stochastic interest rate models,
so in (4.28) both the terms D(T ) and V (T ) are random. Here comes the problem: typically
D(T ) and V (T ) are correlated and because of this fact, the D(T ) part cannot be moved out of
the expectation sign EQ , which leads to more difficulties for the calculation in (4.28).
CHAPTER 4. LIBOR AND FORWARD LIBOR 26

But if we change the measure from Q to QN , this problem can be solved. We know that
V (T ) is the price of a security at time T so it must be F(T )−measurable. Then we can change
the random variable Y in (4.27) into D(T ) and it follows that(also changing t and s in (4.27)
into T and t respectively)
N 1 Q
EQ [V (T )|F(t)] = E [V (T )Z(T )|F(t)]. (4.30)
Z(t)
From (4.23) and (4.24) we could derive the relationship between N (t) and Z(t):
N (t)
Z(t) = D(t) . (4.31)
N (0)
Inserting (4.31) into (4.30), the result is
N 1 Q
N (t)EQ [V (T )|F(t)] = E [V (T )D(T )N (T )|F(t)]. (4.32)
D(t)
Because we want both sides of (4.32) can represent the price of a security V (t) whose payoff
at the maturity T is V (T ), we must require N (T ) = 1, which means that we have to use a
numeraire asset N (t) whose value at T is 1. So which asset should we select to serve as N (t)?
Naturally one can choose the T −bond B(t, T ). In this way, we let N (t) = B(t, T ) and we
rewrite QN as QT when N (t) = B(t, T ). The new measure, QT , is the so-called T −forward
measure. Consequently, the pricing formula (4.32) becomes
1 Q T
V (t) = E [V (T )D(T )|F(t)] = B(t, T )EQ [V (T )|F(t)]. (4.33)
D(t)
So we have successfully separated D(T ) and V (T ) so that under measure QT we only need
to compute the expectation of V (T ) instead of D(T )V (T ). Moreover, the term B(t, T ) is the
price of ZCB, so its appearance should not add extra troubles for the calculation.
1 Q T
V (t) = E [V (T )D(T )|F(t)] = B(t, T )EQ [V (T )|F(t)]. (4.34)
D(t)
In the future sections, we will use (4.34) for the pricing of securities related to forward LIBOR.
The final question for this section is: why should we believe that the probability measure
QT is really the risk neutral measure of the T −bond? To show this, we need to prove that
any asset price denominated by B(t, T ) is a QT −martingale. But instead of directly proving
this, we will work on a more general case: we shall show that the probability measure QN is
just the risk neutral measure of asset N (t) and thus we can prove the result for QT by letting
N (t) = B(t, T ).
Assume now we choose asset N (t) as the numeraire and the stochastic representation of
N (t) under the risk neutral measure(for bank account) Q is

d(D(t)N (t)) = D(t)N (t)v(t) · dW Q (t). (4.35)

CHAPTER 4. LIBOR AND FORWARD LIBOR 27

Now we use the process −v(t) to change the measure from Q to QN as (4.24) to (4.26) show.
S(t)
Now for any asset S(t), we shall show that N (t)
is a QN −martingale. Assume the stochastic
representation(under Q) of S(t) is

d(D(t)S(t)) = D(t)S(t)σ(t) · dW Q (t), (4.36)

then simply by Ito’s formula we can observe that

 
1 1 1
d =− 2 2
d(D(t)N (t)) + 3 d(D(t)N (t))d(D(t)N (t))
D(t)N (t) D (t)N (t) D (t)N 3 (t)
d
!
1 X
= −v(t) · dW Q (t) + vi2 (t)dt
D(t)N (t) i=1
d
!
1 QN
X
2
= −v(t) · (dW (t) + v(t)dt) + vi (t)dt (see equation (4.25))
D(t)N (t) i=1
d
!
1 N
X
=− v(t) · dW Q (t) use the fact that v(t) · v(t) = vi2 (t) .
D(t)N (t) i=1
(4.37)
Moreover, we also have
N
d(D(t)S(t)) = D(t)S(t)σ(t) · dW Q (t) = D(t)S(t)σ(t) · (dW Q (t) + v(t)dt)
N
(4.38)
= D(t)S(t)σ(t) · dW Q (t) + D(t)S(t)σ(t) · v(t)dt.

As a result, the differential of NS(t)

(t)
shall be
   
S(t) D(t)S(t)
d =d
N (t) D(t)N (t)
     
1 1 1
=D(t)S(t)d + d(D(t)S(t)) + d d(D(t)S(t))
D(t)N (t) D(t)N (t) D(t)N (t)
1 N 1 N
= − D(t)S(t) v(t) · dW Q (t) + D(t)S(t)σ(t) · (dW Q (t) + v(t)dt)
D(t)N (t) D(t)N (t)
1 N N
− v(t) · dW Q (t)D(t)S(t)σ(t) · (dW Q (t) + v(t)dt)
D(t)N (t)
S(t) N S(t) N S(t) N N
=− v(t) · dW Q (t) + σ(t) · (dW Q (t) + v(t)dt) − v(t) · dW Q (t)σ(t) · dW Q (t)
N (t) N (t) N (t)
d
!
S(t) N S(t) X
= [σ(t) − v(t)] · dW Q (t) + σ(t) · v(t) − σi (t)vi (t) dt
N (t) N (t) i=1
d
!
S(t) QN
X
= [σ(t) − v(t)] · dW (t) use the fact that σ(t) · v(t) = σi (t)vi (t) .
N (t) i=1
(4.39)
S(t)
Since the dt term disappears in (4.39), we know that N (t)
must be a QN −martingale. Further-
more, because the asset S(t) is arbitrarily selected, we know that QN must be the risk neutral
measure of numeraire N (t). If we choose the T −bond as the numeraire, then the probability
measure QT generated by the Girsanov’s theorem must also be the risk neutral measure of the
T −bond B(t, T ).
CHAPTER 4. LIBOR AND FORWARD LIBOR 28

Additionally, we can see the original volatility vector of S(t) under Q is σ(t), while the
S(t)
new volatility vector of N (t)
under QN is σ(t) − v(t) which is the difference of the volatility
vectors of the two assets.
The risk neutral measure Q is actually the risk neutral measure of bank account G(t)
which has a differential without any Brownian motion terms:

dG(t) = R(t)G(t)dt. (4.40)

This means that the volatility vector of G(t) is a zero vector. We have already seen the stock
price SDEs in the Black-Scholes model:

dS(t) = µS(t)dt + σS(t)dW P (t), under measure P (4.41)

and
dS(t) = rS(t)dt + σS(t)dW Q (t), under measure Q. (4.42)

The fact is that the stock prices processes under both measures share the same volatility
σ. This is caused by choosing the bank account G(t) as the numeraire. Since the volatility
vector for G(t) in this case is 0(the zero vector will become real number 0 if we only consider
models with one Brownian motion like the Black-Schole model), the volatility of stock price
will not experience any change when we change measures from P to Q, i.e. σ P − σ (numeraire) =
σ Q and σ (numeraire) = 0 ⇒ σ P = σ Q .

4.3 LIBOR and Forward LIBOR

Now we introduce some basic concepts in LIBOR market. We use similar way to construct
the forward LIBOR first. Suppose t ≤ T ≤ T + δ. At t we short 1 share of B(t, T ) and long
B(t,T )
B(t,T +δ)
shares of B(t, T + δ). Then at time T we have to pay $1 and at time T + δ, we can
receive $ B(t,T
B(t,T )
+δ)
. In the forward rate model we assume that on interval [T, T + δ] the money
rolls over by a continuously compounding interest rate, which leads to the forward rate. But
this time we assume a simple interest rate on interval [T, T + δ], and this will bring us to the
forward LIBOR, i.e.
 
B(t, T ) 1 B(t, T )
1 · (1 + L(t, T )δ) = ⇒ L(t, T ) = −1 , (4.43)
B(t, T + δ) δ B(t, T + δ)
where the notation L(t, T ) means the forward LIBOR rate which is set at time t and locks the
simple interest rate between T and T + δ. δ is called the tenor of the LIBOR, and typically it
is 90 days or 180 days.
If we have t = T , then the rate L(T, T ) does not lock the simple interest rate for a future
interval. Instead it locks the simple interest rate for an interval that starts from current time
point T , so we usually call it the spot LIBOR.
CHAPTER 4. LIBOR AND FORWARD LIBOR 29

Another important fact is that process L(t, T ), 0 ≤ t ≤ T is actually a QT +δ −martingale.

QT +δ is the risk neutral measure when we choose ZCB B(t, T + δ) as the numeraire. This is
also the T −forward measure discussed in section 4.2 but we change T into T + δ. The reason
for this change is also clear: in (4.43), B(t, T + δ) is at the position of the denominator so
it must be the numeraire. In other words, if we want L(t, T ) to be a martingale under some
probability measure, the most quickest way to do this is choosing B(t, T + δ) as the numeraire
B(t,T )
and thus under its risk neutral measure QT +δ , B(t,T +δ)
is a martingale which causes L(t, T ) to
be a QT +δ −martingale. In the following contents related to the pricing of some widely used
interest rate products, we frequently prefer the measure QT +δ . Finally, we must remember that
L(t, T ) is F(t)−measurable.

4.4 Black Caplet Formula

In many cases, the cash flow structure of interest rate products such as interest rate swaps
is determined by LIBOR. Although for one product, it may have more than one dates when we
can see cash flows are generated, this complex structure can be simplified and in this section
we only consider the simplified model. But this is enough because the cash flows of complex
products can be acquired by repeating the cash flow structure in the simplified model again
and again, and each cash flow can be analyzed by totally the same method that we will show
in the simplified model.
Before introducing the Black caplet formula, let us consider an easier case. What is the
value of a contract at time t that pays L(T, T ) at time T + δ? We complete the calculation by
risk neutral pricing formula and change the measure from Q to QT +δ . We use U (t) to denote
the value at t, then by risk neutral pricing formula, we have
1 Q
U (t) = E [D(T + δ)L(T, T )|F(t)]. (4.44)
D(t)
• Case 1: t ∈ [T, T + δ]. In this case L(T, T ) is F(t)−measurable, so
1 Q
U (t) = E [D(T + δ)L(T, T )|F(t)]
D(t)
 
1 Q (4.45)
= L(T, T ) E [D(T + δ)|F(t)]
D(t)
= L(T, T )B(t, T + δ).

• Case 2: t ∈ [0, T ]. In this case L(T, T ) is not F(t)−measurable, so it will be difficult

to compute U (t) under measure Q. With the rescue of (4.34) and use measure QT +δ , we
obtain
1 Q
U (t) = E [D(T + δ)L(T, T )|F(t)]
D(t) (4.46)
QT +δ
= B(t, T + δ)E [L(T, T )|F(t)].
CHAPTER 4. LIBOR AND FORWARD LIBOR 30

And because we have demonstrated that L(t, T ) is a QT +δ −martingale, (4.46) can be

further written as
U (t) = B(t, T + δ)L(t, T ). (4.47)

In summary, the value of this contract should be a piecewise function:


 B(t, T + δ)L(t, T ), t ∈ [0, T ]

U (t) = . (4.48)
 B(t, T + δ)L(T, T ), t ∈ [T, T + δ]


We will use approach with this style to tackle a more difficult problem and derive the famous
Black caplet formula in the end. Now we consider a contract which pays (L(T, T ) − K)+ =
max(L(T, T ) − K, 0) at time T + δ, where K ≥ 0 is a constant. We will try to find the value
of this contract at time t < T + δ, which is denoted by Y (t). By risk neutral pricing formula,
we have
1 Q
Y (t) = E [D(T + δ)(L(T, T ) − K)+ |F(t)]. (4.49)
D(t)
• Case 1: t ∈ [T, T + δ]. In this case (L(T, T ) − K)+ is F(t)−measurable, so
1 Q
Y (t) = E [D(T + δ)(L(T, T ) − K)+ |F(t)]
D(t)
1 Q (4.50)
= (L(T, T ) − K)+ E [D(T + δ)|F(t)]
D(t)
= (L(T, T ) − K)+ B(t, T + δ).

• Case 2: t ∈ [0, T ]. In this case (L(T, T ) − K)+ is not F(t)−measurable, so it will be

difficult to compute Y (t) under measure Q. With the rescue of (4.34) and use measure
QT +δ , we obtain
1 Q
Y (t) = E [D(T + δ)(L(T, T ) − K)+ |F(t)]
D(t) (4.51)
QT +δ +
= B(t, T + δ)E [(L(T, T ) − K) |F(t)].
T +δ T +δ
The expectation part EQ [(L(T, T ) − K)+ |F(t)] = e−0(T −t) EQ [(L(T, T ) − K)+ |F(t)]
is like the price of an European call option at t with payoff (L(T, T ) − K)+ at maturity
T , strike K, constant interest rate 0. Before we use the Black-Scholes formula to compute
this expectation, we must understand the dynamic of L(t, T ) under meausre QT +δ . Once
again, we use the fact that L(t, T ) is a QT +δ −martingale. According to the MRT, there
exists a process γ(t, T ) such that
T +δ
dL(t, T ) = γ(t, T )L(t, T )dW Q (t). (4.52)

Although the drift is 0 which is a constant, this process is not a geometric Brownian motion
because the volatility γ(t, T ) is not a constant. Actually this is a so-called generalized
geometric Brownian motion. However, the generalized geometric Brownian motion can
CHAPTER 4. LIBOR AND FORWARD LIBOR 31

still be used in the Black-Schole model so this should not be a problem that can stop us
from deriving the results. Using all the information that we have obtaind, we compute the
results of Black-Scholes model under measure QT +δ :
     
1 T 2
RT
log L(t,T ) L(t,T )
+ 21 t γ 2 (s, T )ds
R
K
+ 0 + 2 t
γ (s, T )ds log K
d1 = qR = qR , (4.53)
T 2 T 2
t
γ (s, T )ds t
γ (s, T )ds
     
L(t,T ) 1 T 2 L(t,T ) 1 T 2
R R
log K
+ 0 − 2 t
γ (s, T )ds log K
− 2 t
γ (s, T )ds
d2 = qR = qR , (4.54)
T 2 T 2
t
γ (s, T )ds t
γ (s, T )ds
and
T +δ
EQ [(L(T, T ) − K)+ |F(t)] = L(t, T )N (d1 ) − Ke−0(T −t) N (d2 )
(4.55)
= L(t, T )N (d1 ) − KN (d2 ).
So the contract value for case 2 is

Y (t) = B(t, T + δ) [L(t, T )N (d1 ) − KN (d2 )] . (4.56)

Therefore, the value of this contract is also a piecewise function


 B(t, T + δ)(L(T, T ) − K)+ , t ∈ [T, T + δ]

Y (t) = . (4.57)
 B(t, T + δ) [L(t, T )N (d1 ) − KN (d2 )] , t ∈ [0, T ]


The first piece is easy to derive while the second piece is hard. The formula in the second piece
is the Black caplet formula. Here we only deal with one cash flow but for interest rate products
with more than one cash flows related to LIBOR, we can analyze them one by one and each
one of them can be priced in the way that we have shown above.

4.5 Forward LIBOR Volatility

We utilized MRT to derive (4.52) and γ(t, T ) is the volatility of L(t, T ) under measure
QT +δ . But we only know the existence of γ(t, T ). The formula of γ(t, T ) has not been derived
so this section concentrates on this problem. It turns out that γ(t, T ) is highly related to the
ZCB volatility.
We first derive the SDE of B(t, T + δ) through (2.26):

dB(t, T + δ) = R(t)B(t, T + δ)dt − σ̃(t, T + δ)B(t, T + δ)dW Q (t). (4.58)

(4.58) is an SDE for generalized geometric Brownian motion so we have

  
1 2
B(t, T + δ) = B(0, T + δ) exp R(t) − σ̃ (t, T + δ) t − σ̃(t, T + δ)W (t) .
Q
(4.59)
2
CHAPTER 4. LIBOR AND FORWARD LIBOR 32

Changing the T + δ into T gives the solution of B(t, T ):

  
1 2
B(t, T ) = B(0, T ) exp R(t) − σ̃ (t, T ) t − σ̃(t, T )W (t) .
Q
(4.60)
2
Then by Girsanov’s theorem shown in (4.24) to (4.26) we can find the relationship between the
T +δ
Brownian motion under measure QT +δ (denoted by W Q (t)):
Z t
QT +δ T +δ
W Q
(t) = W (t) + σ̃(s, T + δ)ds ⇒ dW Q (t) = dW Q (t) + σ̃(t, T + δ)dt. (4.61)
0

To understand (4.61), just see (4.25): if we choose asset N (t)(with volatility vector v(t)) as the
numeraire, then we shall set λ(t) = −v(t) to travel between Q and QN (the risk neutral measure
of N (t)). Now if we choose B(t, T + δ)(with volatility vector −σ̃(t, T + δ)) as the numeraire,
then, by the same idea, we set λ(t) = −(−σ̃(t, T + δ)) = σ̃(t, T + δ) to change the measure
from Q to QT +δ (the risk neutral measure of B(t, T + δ)).
Finally we compute the differential of L(t, T ) by (4.43):
 
1 B(t, T )
dL(t, T ) = d
δ B(t, T + δ)
!
B(0, T ) exp R(t) − 12 σ̃ 2 (t, T ) t − σ̃(t, T )W Q (t)


1
= d
B(0, T + δ) exp R(t) − 21 σ̃ 2 (t, T + δ) t − σ̃(t, T + δ)W Q (t)


δ
  
1 B(0, T ) 1 2 1 2
= d exp σ̃ (t, T + δ) − σ̃ (t, T ) t + (σ̃(t, T + δ) − σ̃(t, T )) W (t)
Q
δ B(0, T + δ) 2 2
   
1 2 1 2
let X(t) = σ̃ (t, T + δ) − σ̃ (t, T ) t + (σ̃(t, T + δ) − σ̃(t, T )) W (t)Q
2 2
 
1 B(0, T ) X(t) 1 B(0, T ) X(t) 1
= de = e dX(t) + dX(t)dX(t)
δ B(0, T + δ) δ B(0, T + δ) 2
 
1 B(0, T ) X(t) 1
= e dX(t) + dX(t)dX(t)
δ B(0, T + δ) 2
1 B(t, T )  
= σ̃(t, T + δ) (σ̃(t, T + δ) − σ̃(t, T )) dt + (σ̃(t, T + δ) − σ̃(t, T )) dW Q (t)
δ B(t, T + δ)
1 B(t, T ) T +δ
= (σ̃(t, T + δ) − σ̃(t, T )) dW Q (t)
δ B(t, T + δ)
 
1 T +δ
= + L(t, T ) (σ̃(t, T + δ) − σ̃(t, T )) dW Q (t)
δ
.
(4.62)
We compare (4.62) with (4.52) and then derive the formula of γ(t, T ):
 
1 + δL(t, T )
γ(t, T ) = (σ̃(t, T + δ) − σ̃(t, T )) . (4.63)
δL(t, T )
Chapter 5

Bonds and Interest Rate Derivatives

Pricing

In the previous chapters, we discussed methods to price ZCB securities and a single cash
flow. But interest rate derivatives in the actual financial market have cash flow structures that
are harder to tackle. This chapter will use basic methods mentioned previously to analyze these
actual interest rate derivatives.

5.1 Fixed Coupon-bearing Bonds

This is nearly the easiest product in this chapter. Suppose 0 ≤ t ≤ T0 < T1 < ... < Tn = T.
A fixed coupon-bearing bond makes coupon payments {Ck }nk=1 at times T1 , T2 , ..., Tn and a final
payment for the notional amount(denoted by N ) at Tn . The price of such a product can be
treated as a linear combination of ZCBs, i.e. if we use B C (t, T ) to denote the price, then it
should be n
X
C
B (t, T ) = N B(t, T ) + B(t, Tk )Ck . (5.1)
k=1

The value of Ck is usually determined by the corresponding coupon rate and the time gap
Tk − Tk−1 . If the notional amount is $1, then the price is
n
X
C
B (t, T ) = B(t, T ) + B(t, Tk )Ck . (5.2)
k=1

The assumption that notional amount is $1 holds in the following sections.

5.2 Floating Rate Notes(FRN)

A floating rate note is a coupon-bearing bonding whose coupon payments are determined
by the prevailing LIBOR instead of being fixed like the fixed coupon-bearing bond. This
product is also called floating rate bond, floating coupon bond and LIBOR bond.
Before we price the FRN, we introduce a new notation for LIBOR. In chapter 4 we used
L(t, T ) to represent the forward LIBOR set at time t for future interval [T, T + δ]. But we

33
CHAPTER 5. BONDS AND INTEREST RATE DERIVATIVES PRICING 34

never explicitly wrote the tenor δ in the notation L(t, T ). But in this chapter, we may meet
with forward LIBOR rates with different values of tenor, so if we do not write the tenor, we
may suffer from confusions. As a result, we will make a small modification for the notation,
i.e. we write the tenor for the forward LIBOR: L(t, T ; δ). The meaning of L(t, T ; δ) does not
change: it is still the forward LIBOR set at time t for future interval [T, T + δ], and as shown
in section 4.3, the process L(t, T ; δ), 0 ≤ t ≤ T is a QT +δ −martingale.
Suppose 0 ≤ t ≤ T0 < T1 < ... < Tn = T and define δk = Tk − Tk−1 . The coupon payment
date is T1 , T2 , ..., Tn . Assume the notional amount is $1, then the amount of FRN coupon
payment at time Tk is
Ck = L(Tk−1 , Tk−1 ; δk )δk . (5.3)

So the most important thing that needs to be carefully remembered is that Ck is determined
at time Tk−1 (L(Tk−1 , Tk−1 ; δk ) is F(Tk−1 )−measurable) instead of time Tk . By changing the
measure from Q to QTk and using the martingale property of L(t, Tk−1 ; δk ), we obtain the value
of this coupon payment at time t is(denoted by ck (t))
1 Q
ck (t) = E [D(Tk )Ck |F(t)]
D(t)
1 Q
= E [D(Tk )L(Tk−1 , Tk−1 ; δk )δk |F(t)]
D(t) (5.4)
QTk
=B(t, Tk )E [L(Tk−1 , Tk−1 ; δk )δk |F(t)]
=B(t, Tk )δk L(t, Tk−1 ; δk ).

From (4.43) we acquire the relationship

B(t, Tk−1 )
1 + δk L(t, Tk−1 ; δk ) = . (5.5)
B(t, Tk )
Inserting (5.5) into (5.4) yields
 
B(t, Tk−1 )
ck (t) =B(t, Tk )δk L(t, Tk−1 ; δk ) = B(t, Tk ) − 1 = B(t, Tk−1 ) − B(t, Tk ). (5.6)
B(t, Tk )
So the value of the all the coupon payments of FRN at time t is
n
X n
X
ck (t) = [B(t, Tk−1 ) − B(t, Tk )] = B(t, T0 ) − B(t, T ). (5.7)
k=1 k=1

The value of the FRN at time t(denoted by B F (t, T )) is the sum of the time t value of all
coupon payments and the time t value of the notional amount(which is B(t, T )):
n
X
F
B (t, T ) = ck (t) + B(t, T ) = B(t, T0 ). (5.8)
k=1

Therefore, at the issue date T0 , the FRN value is B(T0 , T0 ) = 1 which is just the notional
amount of this product.
CHAPTER 5. BONDS AND INTEREST RATE DERIVATIVES PRICING 35

5.3 Interest Rate Swaps(IRS)

In an IRS contract, we have two counterparties which swap the fixed payments(determined
by a notional amount and a fixed rate R) and the floating payments(determined by a notional
amount and a floating rate such as LIBOR). Investing in an interest rate swap contract is
equivalent to long a fixed coupon-bearing bond and short an FRN(for the party receiving fixed
payments) or short a fixed coupon-bearing bond and long an FRN(for the party receiving float
payments).
Suppose 0 ≤ t ≤ T0 < T1 < ... < Tn = T and define δk = Tk − Tk−1 . The time points
T1 , T2 , ..., Tn are the dates when the two parties should swap their payments. In the following
analysis we assume the notional amount is $1. At time Tk , the floating payment is

P (float) (Tk ) = δk L(Tk−1 , Tk−1 ; δk ) (5.9)

while the fixed payment is

P (fixed) (Tk ) = δk R. (5.10)

So for the party receiving fixed payments, its net cash flow at Tk is

P (fixed) (Tk ) − P (float) (Tk ) = δk [R − L(Tk−1 , Tk−1 ; δk )]. (5.11)

The value of this net cash flow at time t is(with the help of (5.6))

p(net) (t, Tk ) = δk RB(t, Tk ) − [B(t, Tk−1 ) − B(t, Tk )]. (5.12)

So the contract value at time t for the party receiving fixed payments is
n
X
(fixed)
IRS (t) = p(net) (t, Tk )
k=1
n
X n
X
=R δk B(t, Tk ) − [B(t, Tk−1 ) − B(t, Tk )] (5.13)
k=1 k=1
Xn
=R δk B(t, Tk ) − B(t, T0 ) + B(t, T ).
k=1

The contract value at time t for the party receiving float payments is
n
X
(float)
IRS (t) = −R δk B(t, Tk ) + B(t, T0 ) − B(t, T ), (5.14)
k=1

since the trading of the IRS contract is a zero-sum game.

When t < T0 , the IRS is called a forward interest rate swap(FIRS) initiated at time t while
a spot interest rate swap(SIRS) is initiated at time t = T0 . The forward swap rate is the value
of R in (5.13) and (5.14) that makes the initial value of the IRS to be 0: IRS(t) = 0, which
leads to
B(t, T0 ) − B(t, T )
R∗ (t; T0 , T ) = Pn . (5.15)
k=1 δk B(t, Tk )
CHAPTER 5. BONDS AND INTEREST RATE DERIVATIVES PRICING 36

Under the case where T0 = t, we obtain the spot swap rate

1 − B(t, T )
R∗ (t, T ) = R∗ (t; t, T ) = Pn . (5.16)
k=1 δk B(t, Tk )
Remember that the forward or spot swap rate can merely ensure that the value of contract is 0
at the initial time t. Later with time going by, the contract will have nonzero values. Finally,
for fixed initial time t, the map T 7→ R∗ (t, T ) is the swap rate term structure(or the swap yield
curve).

5.4 Interest Rate Options

For this section, we also define the time points and time increments as 0 ≤ t ≤ T0 < T1 <
... < Tn = T and δk = Tk − Tk−1 .

5.4.1 Caps and Floors

An interest rate cap is a series of call options on the spot LIBOR starting from Tk−1 and
ending at Tk for any k ∈ {1, 2, ..., n}. At any time Tk , the product pays cash flow H(Tk ) =
δk (L(Tk−1 , Tk−1 ; δk ) − R)+ = δk max{L(Tk−1 , Tk−1 ; δk ) − R, 0}, where R ≥ 0 is a fixed constant
called the cap rate. The value of H(Tk ) at time t is
1 Q T
h(t, Tk ) = E [D(Tk )H(Tk )|F(t)] = B(t, Tk )EQ k [H(Tk )|F(t)]
D(t) (5.17)
QTk +
= B(t, Tk )δk E [(L(Tk−1 , Tk−1 ; δk ) − R) |F(t)].
The result of (5.17) can be given by the Black caplet formula (4.53) to (4.55):
h    i
(1) (2)
h(t, Tk ) = B(t, Tk ) L(t, Tk−1 ; δk )N dk − RN dk , (5.18)

where we have   RT
L(t,Tk−1 ;δk ) 1
log R
+ 2 t
γk2 (s, T )ds
(1)
dk = qR , (5.19)
T
t
γk2 (s, T )ds
and   RT
L(t,Tk−1 ;δk ) 1
log R
− 2 t
γk2 (s, T )ds
(2)
dk = qR . (5.20)
T
t
γk2 (s, T )ds
The γk (t, T ) is the volatility of L(Tk−1 , Tk−1 ; δk ) under T −forward measure QTk (see (4.52)) such
that
T
dL(Tk−1 , Tk−1 ; δk ) = γk (t, T )L(Tk−1 , Tk−1 ; δk )dW Q k (t). (5.21)
Therefore, when we change the value of k from 1, 2, ..., n, we shall obtain sequences {h(t, Tk )}nk=1 ,
(1) (2)
{dk }nk=1 , {dk }nk=1 and {γk (t, T )}nk=1 . Finally, the value of the interest rate cap is the sum of
all the elements in {h(t, Tk )}nk=1 :
n
X n
X h    i
(cap) (1) (2)
V = h(t, Tk ) = B(t, Tk ) L(t, Tk−1 ; δk )N dk − RN dk . (5.22)
k=1 k=1
CHAPTER 5. BONDS AND INTEREST RATE DERIVATIVES PRICING 37

Interest rate floor contracts are similar: they pay cash flow δk (R − L(Tk−1 , Tk−1 ; δk ))+ at any
time Tk for each k ∈ {1, 2, ..., n}, where R ≥ 0 is still a fixed constant called the floor rate. The
pricing of floors can be finished by the same methods presented above.

5.4.2 Swaptions

A (T0 − t) × (T − T0 ) swaption with strike R is a contract which starts from time t and, at
its exercise date T0 , gives the holder the right to enter into an IRS contract with swap payment
dates {Tk }nk=1 and fixed swap rate R. If the right offered by the swaption is enter into the IRS
as the party receiving float payments, then this swaption is a payer’s swaption. Otherwise, the
swaption is a receiver’s swaption. Remember that the existence period of the swaption is time
interval [t, T0 ] so T0 is the maturity date of the swaption, and that time interval [T0 , T ] is the
start date to the end date of the underlying IRS.
Now assume that we want to price a payer’s swaption. The first step for us to do this
is computing the payoff of the swaption at its maturity T0 . Then we can discount the payoff
to t under the risk neutral measure Q to obtain the value of the swaption at time t. Because
this is a payer’s option, its payoff at T0 depends on the swap value for the party receiving float
payments at T0 : IRS(float) (T0 ) defined in (5.14). To be specific, the payoff can be written as

J(T0 ) = max{IRS(float) (T0 ), 0}

( n
)
X
= max −R δk B(T0 , Tk ) + B(T0 , T0 ) − B(T0 , T ), 0
k=1
(5.23)
( n
! )
X
= max 1 − R δk B(T0 , Tk ) + B(T0 , T ) , 0
k=1
Pn
The part R k=1 δk B(T0 , Tk )+B(T0 , T )
in (5.23) is just the price of a fixed coupon-bearing bond
from T0 to T with coupon payments Ck = Rδk . Let B C (T0 , T ) = R nk=1 δk B(T0 , Tk )+B(T0 , T )
P

and transform (5.23) into

J(T0 ) = max 1 − B C (T0 , T ), 0

(5.24)
As a result, the swaption price at time t should be
1 Q
J(t) = E [D(T0 )J(T0 )|F(t)]
D(t)
(5.25)
1 Q
E D(T0 ) max 1 − B C (T0 , T ), 0 |F(t) ,
 
=
D(t)
which is the price of a put option on the fixed coupon-bearing bond B C (t, T ) with maturity T0 ,
strike price $1. But to find J(t), we must find the SDE of the price of the fixed coupon-bearing
bond B C (t, T ), which is troublesome. We can numerically solve this problem by breaking the
bond into several ZCBs and then simulating B C (t, T ) by simulating the price of each ZCB first.
But we have another way to derive J(t). (5.15) shows the forward swap rate, and we can
treat the denominator nk=1 δk B(t, Tk ) as a new asset A(t; T0 , T ) = nk=1 δk B(t, Tk ) because
P P
CHAPTER 5. BONDS AND INTEREST RATE DERIVATIVES PRICING 38

A(t; T0 , T ) is strictly positive for any 0 ≤ t ≤ T0 . We rewrite (5.15) as

B(t, T0 ) B(t, T )
R∗ (t; T0 , T ) = − . (5.26)
A(t; T0 , T ) A(t; T0 , T )
If we choose asset A(t; T0 , T ) as the numeraire, and donote the risk neutral measure of A(t; T0 , T )
B(t,T0 ) B(t,T )
by QA . Then, under measure QA , we know that both A(t;T0 ,T )
and A(t;T0 ,T )
are QA −martingales
and thus the forward swap rate process R∗ (t; T0 , T ) is also a QA −martingale. Recall (4.32), we
let N (t) in (4.32) be A(t; T0 , T ) and then QN will become QA , so the value of the swaption is
" ( n
) #
1 Q X
J(t) = E D(T0 ) max −R δk B(T0 , Tk ) + 1 − B(T0 , T ), 0 F(t)
D(t) k=1
 
1 Q ∗
= E D(T0 ) max {−RA(T0 ; T0 , T ) + R (T0 ; T0 , T )A(T0 ; T0 , T ), 0} F(t)
D(t) (5.27)
 
1 Q
= E D(T0 ) max {A(T0 ; T0 , T )[R∗ (T0 ; T0 , T ) − R], 0} F(t)
D(t)
 
1 Q ∗
= E D(T0 )A(T0 ; T0 , T ) max {R (T0 ; T0 , T ) − R, 0} F(t) .
D(t)
Observing the term max {R∗ (T0 ; T0 , T ) − R, 0} in (5.27), we can treat the swaption as a call
option on the forward swap rate R∗ (t; T0 , T ) with maturity T0 , strike price R. We use (4.32) to
convert the calculation of expectation from Q to QA :
A
J(t) = A(t; T0 , T )EQ [max {R∗ (T0 ; T0 , T ) − R, 0} |F(t)] . (5.28)

The work next is totally the same as the proof of the Black caplet formula. Because A(t; T0 , T )
is a QA −martingale, by martingale representation theorem, there exists a process ψ(t, T ) such
that
A
dA(t; T0 , T ) = ψ(t; T0 , T )A(t; T0 , T )dW Q (t), (5.29)

which is a generalized geometric Brownian motion. The final solution of J(t) is given by
 ∗  RT
log R (t;T
R
0 ,T )
+ 12 t 0 ψ 2 (s; T0 , T )ds
d1 = qR , (5.30)
T0 2
t
ψ (s; T0 , T )ds
 ∗ 
R (t;T0 ,T ) 1 T0 2
R
log R
− 2 t
ψ (s; T0 , T )ds
d2 = qR , (5.31)
T0 2
t
ψ (s; T0 , T )ds
and
J(t) = A(t; T0 , T ) [R∗ (t; T0 , T )N (d1 ) − RN (d2 )] . (5.32)

In real financial market, what we need to do before applying (5.32) is the calibration for
A(t; T0 , T ) = nk=1 δk B(t, Tk ) which can be obtained by the market data of ZCB prices with
P

different maturities.
CHAPTER 5. BONDS AND INTEREST RATE DERIVATIVES PRICING 39

The End.