You are on page 1of 561

Advanced Derivatives Pricing and Applications

Dr Dale Roberts
Australian National University

Semester 2, 2014
FINM3007/MATH3015/MATH6015

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

1 / 561

Week 1

Vectors and Matrices


Functions of Bounded Variation
Stieltjes Integration
Gaussian Random Variables

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

2 / 561

Dale Roberts (c) 2011-2014

Vectors and Matrices

Last Modified: 14/10/2014 08:19

3 / 561

Vectors
A vector (or column) of length n is an n 1 array

x1
x2

x = (x 1 , x 2 , . . . , x n )0 =
... .
xn
The superscript 0 means transpose and converts the row
[x 1 x 2 x n ]

to a column.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

4 / 561

Vector-Vector Product
Let x and y be vectors of length n.
We often write x Rn and y Rn .
The dot product or inner product of x and y is given by
0

x y := x y =

n
X

x i yi = x 1 y1 + x 2 y2 + + x n yn .

The norm of a vector x is given by

Dale Roberts (c) 2011-2014

kxk :=

x0x =

v
uX
n
t

x i2

i=1

Last Modified: 14/10/2014 08:19

5 / 561

Matrices
An m n matrix is an array of objects consisting of m rows and n columns

a11 a12 a1n


a22 a2n
a
A = [ai j ] = 21

am1 am2 amn
The element ai j is the entry in the i th row and j th column.
So a column of length n is simply a n 1 matrix and a row of length n is a
1 n matrix.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

6 / 561

Matrices
A n n diagonal matrix is a matrix of the form

d1 0
0
0
0 d2 0
.
..
.
.
D=
.

.
.

. 0 dn1 0
0
0
dn

wherein all entries not on the main diagonal are zero.


Also denoted
D = diag(di )1n = diag(d1 , d2 , . . . , dn )
A special case is the identity matrix

Dale Roberts (c) 2011-2014

I = diag(1, 1, . . . , 1).
Last Modified: 14/10/2014 08:19

7 / 561

Matrix-Vector Product
The (matrix-vector) product Ax is defined to be the column of length m
whose entry in the i th row is
(Ax)i =

n
X

ai j x j = ai1 x 1 + ai2 x 2 + + ain x n

(1)

j=1

If A = [1 2 n ] is a single row (i.e., m = 1), then (1) becomes



x1
n
x2 X

Ax = [1 2 . . . n ] .. =
j x j.
.

Dale Roberts (c) 2011-2014

j=1

xn

Last Modified: 14/10/2014 08:19

8 / 561

Matrix-Vector Product

If the i th row of A is denoted by


ai = [ai1 ai2 ain ],

then

Dale Roberts (c) 2011-2014

(Ax)1
a1 x
(Ax)2 a2 x

Ax =
... = ...

(Ax)m

Last Modified: 14/10/2014 08:19

am x

9 / 561

Matrix-Vector Product
If A = [ai j ] and B = [bi j ] are two matrices of the same size. Then
A = B ai j = bi j ,

i, j

A B = [ai j bi j ]
A = [ai j ]
The following laws of matrix-vector multplication hold:
(A)x = (Ax) = A(x)

(Associative law)
A(x + y) = Ax + Ay (Distributive law)
(A + B)x = Ax + B x (Distributive law)
where A, B are n m matrices; , real numbers; and x, y vectors of
length n.
Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

10 / 561

Matrix-Matrix Product
Let A be an m n matrix. A j denotes the j th column of A and we write
A = [A1 A2 An ].
(A)i j denotes the element of A at row i and column j . Hence, if A = [ai j ]
then (A)i j = ai j .
Let A = [ai j ] be an m n matrix and B = [bi j ] be a p q matrix with
columns B j . The (matrix-matrix) product AB is defined if and only if n = p
and in this case
AB = [AB1 AB2 ABq ]

is a m q matrix.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

11 / 561

Determinants

The determinant of n n matrix A is denoted by det A = |A|.


When n = 1, det A = |[a]| = a.
When n = 2,

a b
det
= ad bc.
c d
Determinants of matrices of size n > 2 are computed in terms of
determinants of matrices of size n 1.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

12 / 561

Determinants
The minor Mi j of ai j in A = [ai j ] is the determinant of the submatrix of A
obtained by deleting the i th row and j th column.
The cofactor of ai j is Ci j = (1)i+ j Mi j .
For each i, j the following hold:

Dale Roberts (c) 2011-2014

|A| =
|A| =

n
X
k=1
n
X

aik Cik (Expansion along i th row)


ak j Ck j (Expansion along j th row)

k=1

Last Modified: 14/10/2014 08:19

13 / 561

Determinants
Example
Expanding along the first row:


1 2 3








5 6
4 6
4 5
2



det A = 4 5 6 = 1
7 9 + 3 7 8 = 0.
8 9
7 8 9
Expanding along the second column:






1 3
4 6
1 3
+ 5



|A| = 2
7 9 8 4 6 = 0.
7 9

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

14 / 561

Inverses of Square Matrices


A matrix B is an inverse of A if and only if AB = I and BA = I .
In this case, we write B = A1 .
A is said to be invertible or nonsingular when it has an inverse.

Example

cos( ) sin( )
A=
.
sin( ) cos( )

cos( ) sin( )
=
.
sin( ) cos( )

You can verify that AA1 = I and A1 A = I .

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

15 / 561

Dale Roberts (c) 2011-2014

Functions of Bounded Variation

Last Modified: 14/10/2014 08:19

16 / 561

Nondecreasing Functions

If f1 and f2 are nondecreasing functions on a closed interval [a, b] then


f1 + f2 is also a non-decreasing function but f1 f2 need not be
The function f := f1 f2 is in the class of functions of bounded variation

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

17 / 561

Setup

f : [a, b] R
P = {(x i1 , x i ) : i = 1, . . . , k} partition of [a, b]
P = set of all partitions of [a, b]

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

18 / 561

Variation

Definition
Take f : [a, b] R, P = {(x i1 , x i ) : i = 1, . . . , k} a partition of [a, b], and P
is the set of all partitions of [a, b]. The variation Vab ( f ) of f on [a, b] is
Vab ( f

) = sup

k
X

| f (x i ) f (x i1 )| : P P

i=1

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

19 / 561

Functions of Bounded Variation

Definition
If Vab ( f ) is finite then f is of bounded variation on [a, b].
Note: If Vab ( f ) = then we say that f has unbounded variation. For example,
Brownian motion has paths of unbounded variation.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

20 / 561

Example 1

Example
If function f is non-decreasing on [a, b], then for any P P :
k
X
i=1

| f (x i ) f (x i1 )| =

k
X

f (x i ) f (x i1 ) = f (b) f (a)

i=1

Therefore,
Vab ( f ) = f (b) f (a).

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

21 / 561

Example 2
Example
The function

f (x) =

0,

x =0

sin( 1x ),

x > 0.

f
/ BV[0, 2 ].
1.0

0.5

0.1

0.2

0.3

0.4

0.5

0.6

-0.5

-1.0

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

22 / 561

Example 3
Example
The function

f (x) =

x =0

0,
x

sin( 1x ),

x > 0.

f
/ BV[0, 2 ].
0.6

0.4

0.2

0.1

0.2

0.3

0.4

0.5

0.6

-0.2

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

23 / 561

Dale Roberts (c) 2011-2014

Stieltjes Integration

Last Modified: 14/10/2014 08:19

24 / 561

Stieltjes Integral

Useful in probability theory as it allows one to handle continuous and


discrete random variables using the same theory
The Stieltjes generalisation of Riemann (and Lebesgue) integrals involve
two functions
Integrating one function against the increments of another function

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

25 / 561

Stieltjes Sum

Let f : [a, b] R and g : [a, b] R. For any P P , P := {I1 , I2 , . . . , I n } with


I k := (x k1 , x k ) choose points zk [x k1 , x k ] and form the sum
S( f , g; P) :=

n
X

f (zk )(g(x k ) g(x k1 )).

k=1

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

26 / 561

Riemann-Stieljes Integral
Definition
If lim|P|0 S( f , g; P) exists, then this limit is called the Riemann-Stieljes integral
of f with respect to g and is denoted
Z b
Z b
f (t) d g(t)

or

f d g.
a

The function f is called the integrand and the function g is called the
integrator.
Note: A Lebesgue-Stieljes integral can also be constructed.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

27 / 561

Linearity in the Integrand

Lemma
If

Rb
a

f1 d g and

Rb
a

( f1 + f2 ) d g =
a

Dale Roberts (c) 2011-2014

f2 d g exist and , R, then


b

f1 d g +
a

Last Modified: 14/10/2014 08:19

f2 d g.
a

28 / 561

Linearity in the Integrator

Lemma
If

Rb
a

f d g1 and

Rb
a
b

f d(g1 + g2 ) =
a

Dale Roberts (c) 2011-2014

f d g2 exist and , R, then


b

f d g1 +
a

Last Modified: 14/10/2014 08:19

f d g2 .
a

29 / 561

Integration by Parts

Lemma
If f integrable with respect to g , then g is integrable with respect to f and
Z b
b Z b

g d f = f g
f d g.
a

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

30 / 561

Existence Criteria
Theorem
If f C[a, b] and g BV[a, b] then
Z b
f dg

exists.

Theorem
If f BV[a, b] and g C[a, b], then
Z b
f dg

exists.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

31 / 561

Outlook to Stochastic Integration


We want to be able to integrate a function f against the paths of a Brownian
motion (Wt ) t0 .
However, each path t 7 Wt () is of unbounded variation but continuous
We could potentially write
b

f (t) dWt
a

if f BV[a, b] and use Stieltjes integration. This is called path-by-path


integration.
What about
Z b
Z b
1
sin
dWt or
Wt dWt ?
t
a
a
Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

32 / 561

Dale Roberts (c) 2011-2014

Gaussian Random Variables

Last Modified: 14/10/2014 08:19

33 / 561

Gaussian Random Variables


A random variable X on a probability space (, F , P) is normally distributed or
Gaussian with parameters and 2 , written as X N(, 2 ), if
 x 
P{X < x} =
, x R

where is the normal distribution function


Z x

Dale Roberts (c) 2011-2014

(x) :=

2
p1 e u /2 du.
2

Last Modified: 14/10/2014 08:19

34 / 561

Characterisation

The moment generating function characterises the class of Gaussian random


variables.

Lemma
A random variable X N(, 2 ) under a measure P if and only if
1

E[e X ] = e + 2

Dale Roberts (c) 2011-2014

for all real .

Last Modified: 14/10/2014 08:19

35 / 561

Gaussian Shifts

Lemma (Gaussian shift)


Let X N(0, 1), let h be a measurable function, and c R. Then
E[e cX h(X )] = e c

Dale Roberts (c) 2011-2014

/2

E[h(X + c)]

Last Modified: 14/10/2014 08:19

36 / 561

Gaussian Shifts

Example
Take h(x) = 1 then

E[eZ ] = e

/2

Note: Z N(0, 2 )

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

37 / 561

Gaussian Shifts

Example
If Z N(0, 1) then X := + Z N(, 2 ) so
E[e X ] = E[e+Z ]

We identify h(x) = e so h(Z) = e and


E[e X ] = E[e+Z ] = E[eZ h(Z)] = e

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

/2

e = e+

/2

38 / 561

Multivariate Random Variables

Consider vector X whose elements are random variables on the same


probability space (, F , P):
X = (X 1 , X 2 , . . . , X n )0
E[X i ] = i and Cov[X i , X j ] = qi j for i, j = 1, . . . , n.
We call the vector := (1 , . . . , n )0 the mean and the matrix
Q := (qi j )ni, j=1 the covariance matrix of X .

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

39 / 561

Multivariate Gaussians
Definition
Assume det Q > 0, then the (random) vector X has a multivariate normal
distribution if it has the density

1
1
0 1
(x) = p
exp (x ) Q (x ) , x Rn .
n
2
(2) det Q
We write X N(, Q) if this is the case.
For example: x Rn ,
Z

Dale Roberts (c) 2011-2014

P{X > x} =

(x 1 , x 2 , . . . , x n ) d x 1 d x 2 d x n .

x1

x2

xn

Last Modified: 14/10/2014 08:19

40 / 561

Multivariate Gaussians
If Z N(0, Q) and c = (c1 , c2 , . . . , cn )0 Rn then X := c T Z is Gaussian with
X N(0, c 0Qc).
If C Rmn , i.e. a m n matrix, then we have
X = C Z N(0, CQC 0 ),

and CQC 0 is a m m covariance matrix.


Given a Rn , X is alternatively characterized by its moment generating
function

1 0
0X
0
E[e ] = exp Q .
2

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

41 / 561

Multivariate Gaussian Shifts

Lemma
Let X N(0, Q), let h be a measurable function of x Rn , and c Rn . Then
0

1 0

E[e c X h(X )] = e 2 c Qc E[h(X + c)]

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

42 / 561

Week 2

Brownian Motion
Monte Carlo Method
Simulating Stochastic Processes

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

43 / 561

Dale Roberts (c) 2011-2014

Brownian Motion

Last Modified: 14/10/2014 08:19

44 / 561

Probability Space
A probability space (, F , P) consists of three parts:
the sample space (the set of all possible outcomes),
the set of events F (each event is a set containing zero or more outcomes),

Example (Coin flip)


= {H, T }. F contains 4 events: {H}, {T }, {} (neither head nor tails) and
{H, T } (either heads or tails). F = {{}, {H}, {T }, {H, T }}. Then
P({}) = 0,

Dale Roberts (c) 2011-2014

P({H}) = P({T }) =

Last Modified: 14/10/2014 08:19

1
2

P({H, T }) = 1

45 / 561

The Set of Events

The set of events F (the set of all subsets of ) is a -algebra as:


F contains the sample space F
F is closed under complements: if A F then ( \ A) F
F is closed under countable unions: if A1 , A2 , . . . F then i Ai F

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

46 / 561

Stochastic Processes

Definition
A stochastic process X is a family (X t )0tT of random variables defined on the
same probability space (, F , P) taking values in Rn (the state space).
The process X can be viewed in terms of:
a (random) path t 7 X t ()
the probability distribution of X at some fixed time t , i.e., P{X t < x}.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

47 / 561

Finite-dimensional Distributions

Definition
If X := (X t )0tT is a Rn -valued stochastic process, then for every finite
sequence of times t 1 , t 2 , . . . , t k the probability distribution of the random
variable
X := (X t 1 , X t 2 , . . . , X t k )0
is called the finite-dimensional distribution of X .
Note: Since X is Rn valued and we are evaluating at k times t 1 , . . . , t k then X is
Rnk -dimensional.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

48 / 561

Filtrations and Adaptedness


Definition
Given a probability space (, F , P), a filtration is an increasing collection of
events (F t ) t0 and bounded above by F , i.e.,
Fs F t F ,

s<t

We say that a stochastic process X = (X t ) t0 is adapted to the filtration if, for


every time point t , the random variable X t only relies on the information up
to time t , i.e., F t .
Examples: 2X 1 is adapted to F1 but X 1 is not adapted to F1/2 .

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

49 / 561

Brownian Motion
Definition
A stochastic process W := (Wt ) t0 is a called a Brownian motion (under P) if it
has all of the following properties:
1. the paths t 7 Wt () are continuous for P-almost all ;
2. W (0) = 0, P-almost surely;
3. for 0 s < t < , Wt Ws N(0, t s);
4. for k N and 0 t 1 < t 2 < . . . < t k , the increments Wt i+1 Wt i with i < k,
are independent.
We shall denote by F t = (F t ) t0 the natural filtration generated by the
process W : the event set F is generated by all values of Ws up to time s = t .

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

50 / 561

Existence

Theorem (Wiener 1923)


Brownian motion exists.
In homage to Norbert Wiener, a Brownian motion is also known as a Wiener
process.
Trivia: Wiener (1894 - 1964) was rejected for a job at University of Melbourne.
Ended up getting a job at MIT where he stayed the rest of his life.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

51 / 561

Simple Properties

The following properties of a Brownian motion W can easily be proved:


Wt N(0, t);
Cov[Ws , Wt ] = min(s, t);
1

E[e Wt ] = e 2 t ;
2

E[Wt4 ] = 3t 2 .

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

52 / 561

Invariances

Brownian motion (BM) also has a number of useful invariance properties. In


particular, if W is a BM then so are the following processes:
the reflected process X t := Wt ;
the scaled process X t := 1c Wc 2 t for c > 0;
the stationary process X t := Wt+u Wu for u > 0;
the inverted process X t := tW1/t for t > 0 and X 0 = W0 = 0.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

53 / 561

Donskers Theorem
Theorem (Donsker 1951)
Let (n )nN be a sequence of independent and identically distributed random
variables with E[i ] = 0 and 0 < Var[i ] = 2 < . Let
S0 = 0,

Sn =

n
X

i .

i=1

We construct a sequence of X n of stochastic processes by


X n (t, ) =

1
1
p S[nt] () + (nt [nt]) p [nt]+1 ()
n
n

for t [0, 1], n N. Then this sequence converges (weakly) towards the
Brownian motion B := (B t )0t1 , i.e., we have X n B in probability dist.
Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

54 / 561

Dale Roberts (c) 2011-2014

Monte Carlo Method

Last Modified: 14/10/2014 08:19

55 / 561

Main Idea

The main idea of the Monte Carlo method is to approximate an expected value
E[X ] by an arithmetic average of the results of a big number of independent
experiments which all have the same distribution as X .
Trivia: In 1946, physicists needed to perform complicated calculations for the
Manhattan project. They couldnt solve the problem using standard
mathematical methods. The Mathematician Stanislaw Ulam (1909 - 1984)
suggested using random experiments. Being a secret project, this required a
code name. Von Neumann chose the name Monte Carlo.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

56 / 561

Strong Law of Large Numbers

One of the most powerful theorems of probability theory


Has been a central object of research during the history of the subject
It is the basis of the Monte Carlo method

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

57 / 561

Strong Law of Large Numbers


Theorem (SLLN)
If (X n )nN be a sequence of integrable, R-valued random variables that are
independent, identically distributed and defined on the probability space
(, F , P). Let = E[X 1 ]. Then, we have for P-almost all
n

1X
X i ()
n i=1

The arithmetic mean of the (realisations of) X i tends to the theoretical mean
of every X i .

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

58 / 561

Crude Monte Carlo Method

Approximate E[X ] by the arithmetic mean


N
1X
X i ()
N i=1

for some finite N N.


Here, the X i () are the results of N independent experiments that the same
probability distribution as X .

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

59 / 561

Unbiased Estimator
Theorem
Let (X n )nN be a sequence of integrable R-valued random varaibles that are i.i.d.
as X . Then the Monte Carlo estimator
N
1X
X N :=
Xi,
N i=1

N N,

is an unbiased estimator for = E[X ], i.e., we have


E[X N ] = .

Note: This ensures that the MC Estimator is correct in the mean, it does not
help us get a feel for the absolute value of the error.
Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

60 / 561

Increasing Accuracy
We look at the standard deviation of the difference between X N and
As

1 X
2
Var[X
]
=
i
N 2 i=1
N
p
the standard deviation is of order O(1/ N )
Var[X N ] = Var[X N ] =

This means that to increase the accuracy of the crude MC estimate by one
digit (i.e., reduce the std. dev by a factor of 0.1) requires increasing N by a
factor of 100.
To acheive higher accuracy, we need a significant effort.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

61 / 561

Central Limit Theorem


Theorem (CLT i.i.d case)
Let (X n )nN be a sequence of i.i.d. R-valued random variables on (, F , P).
Assume that they all have finite variance 2 = Var[X ]. Then, the normalised and
centralised sum of these random variables converges in distribution to the
standard normal distribution:
N

X
1
D
X i N N(0, 1) as N .
p
N i=1
We can infer from the CLT that for large values of N the crude Monte Carlo
estimator is approximately N(, 2 /N )-distributed.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

62 / 561

Approximate Confidence Interval


An approximate (1 )-confidence interval for the expectation is
N
1 X

N
1X

X i z1/2 p ,
X i + z1/2 p
N i=1
N N i=1
N

where z1/2 is the 1 /2-quantile of the standard Normal distribution.


As 97.5%-quantile is about 1.96, an approximate 95% confidence is
N
1 X

N
1X

Xi 2p ,
Xi + 2p
N i=1
N N i=1
N

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

63 / 561

Approximate Confidence Interval


Typically, the standard deviation is unknown. We need to estimate 2 by the
sample variance
v
v

u
u
N
N
X
t 1 X
t
N
1
N =
(X i X N )2 =
X i2 X 2N
N 1 i=1
N 1 N i=1
Giving our approximation of 95% confidence
N
1 X

N
N 1 X

Xi 2p ,
Xi + 2pN
N i=1
N N i=1
N

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

64 / 561

Calculating
Example ()
The classic example of the Monte Carlo method is calculating the value of .
Consider the part of the unit circle C with center at the origin (0, 0) that
intersects with the positive unit square [0, 1]2 . Our experiment is to randomly
choose points p1 , p2 , . . . , pN of the unit square (i.e., uniformly distributed) and
consider
X i = 1{pi C} .
We have then P(pi C) = /4 as the probability of hitting C is its area r 2 /4
and r = 1. Since
E[1{pi C} ] = P(pi C) = /4
we can estimate by
N
4X
() =
1{pi C} ().
N i=1
Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

65 / 561

Monte Carlo Integration


Example (MC Integration)
Suppose you wanted to calculate the value of the deterministic n-dimensional
integral of the form
Z
g(x) d x.
[0,1]n

Introduce the density f (x) of the n-dimensional Uniform distribution on


[0, 1]n via
f (x) = 1{x[0,1]n } (x), x Rn
then
I=

g(x) d x =

g(x) f (x) d x = E[g(X )]

[0,1]n

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

66 / 561

Monte Carlo Integration


We approximate by
N
1X
I N () =
g(X i ())
N i=1

where X 1 , . . . , X N are all i.i.d. uniformly distributed random variables on [0, 1]n
p
Remark: By the SLLN the rate of convergence is O(1/ N ) but is independent
of the dimension n. As opposed to deterministic quadrature formulas that
have a rate of convergence O(N 2/n ). The MC approximation outperforms
when n > 4.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

67 / 561

Dale Roberts (c) 2011-2014

Simulating Stochastic Processes

Last Modified: 14/10/2014 08:19

68 / 561

Simulating Stochastic Processes

In practice, we often need to calculate expected values by the Monte Carlo


method
If X = (X t )0tT is a stochastic process, we often want to calculate
= E[g(X )] = E[g(X t , 0 t T )]

where g is a function of the whole path of X .

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

69 / 561

Algorithm for Stochastic Processes

Approximate E[g(X )] by the arithmetic mean


N
1X
g(X i ()).
N i=1

where X 1 , X 2 , . . . are independent replications of paths of the stochastic


process X .

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

70 / 561

Special Cases
If the function g only depends on the value of the stochastic process X at a
particular time, i.e., if we have
g(X ) = h(X T )

for a fixed time T and R-valued function h, then we only have to know the
distribution of X T . Either the distribution of X T is known, or we have to
simulate the whole path to get X T .
If g only depends on X at a finite number of times t 1 , t 2 , . . . , t n , i.e.,
g(X ) = h(X t 1 , . . . , X t n )

for some function h then we are in the random variable case.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

71 / 561

General Case

If g cannot be reduced to one of the special cases then we need to simulate


the true path of X . Unfortunately, this is not possible and we need to find
approximation methods. There is no general approach and methods have to
be tailored to the specific problem.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

72 / 561

Simulating Paths: Basics

The fact that the random variables X t for all t s are related to each other
has the consequence that we cannot simply simulate independent
random variables that have the same distribution as the different X t
We have to take care of the relation among the X t which can be extremely
strong
Since the times t [0, T ] are uncountable, we cannot simulate the true
path of the process X . We need to approximate the path at finite times
0 = t1 < t2 < < t n = T

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

73 / 561

Algorithm
Let 0 = t 0 < t 1 < < t n = T be a partition of [0, T ]. Let Pk denote the
conditional distribution of X t k given X t k1 . We obtain the approximate path
X () by:
1. Set X 0 () = 0
2. For k = 1 to n do:
(a) Simulate a random number Yk () with Yk Pk
(b) Set X t k () = X t k1 () + Yk ()
(c) Between t k+1 and t k obtain X t via linear interpolation

Dale Roberts (c) 2011-2014

X t () = X t k1 () +

t t k1
Yk (),
t k t k1

Last Modified: 14/10/2014 08:19

t (t k1 , t k ).

74 / 561

Brownian Motion
Let 0 = t 0 < t 1 < < t n = T be a partition of [0, T ]. We obtain the
approximate path t 7 B t () by:
1. Set B0 () = 0
2. For k = 1 to n do:
(a) Simulate a random number
p Zk N(0, 1)
(b) Set B t k () = B t k1 () + t k t k1 Zk
(c) Between t k=1 and t k obtain X t via linear interpolation

Dale Roberts (c) 2011-2014

B t () = B t k1 () +

t t k1
(B t () B t k1 ()),
t k t k1 k

Last Modified: 14/10/2014 08:19

t (t k1 , t k ).

75 / 561

Week 3

Stochastic Integration
Stochastic Differential Equations
Simulating Solutions of SDE

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

76 / 561

Dale Roberts (c) 2011-2014

Stochastic Integration

Last Modified: 14/10/2014 08:19

77 / 561

Simple Functions
Let F = (F t ) t0 and h an F-adapted process on [0, T ] taking values in R.

Definition
We call h a simple process if for a partition := {0 = t 0 < t 1 < < t n = T }
and F t i -measurable random variables (hi )ni=1 , the process h(t) := h(, t)
satisfies
h(, t) =

n1
X

hi ()1(t i ,t i+1 ] (t)

for 0 t T and

i=0

Note: In the easy case, h is a simple function given by h(t) =


for 0 t T where hi R for i = 1, . . . , n.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

Pn1
i=0

hi 1(t i ,t i+1 ] (t)

78 / 561

Properties

Simple process and functions allow us to approximate arbitrary processes


and functions
We approximate the area under the curve by rectangles
We build up the concept of stochastic integrals by starting with the case
for simple processes

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

79 / 561

Stochastic Integral for Simple Processes

Definition
The stochastic integral (aka. It integral) of a simple process h is defined as
Z T
n1
X
hi ()(Wt i+1 Wt i )

h(t) dWt =

i=0

whenever E[h2i ] < for all i = 0, . . . , n 1.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

80 / 561

Square-integrable Simple Processes

We required E[h2i ] < for all i = 0, . . . , n 1 for the stochastic integral


to exist.
2
We write H[0,T
for the set of all F-adapted simple processes h that satisfy
]

E[h2 ] d t < .

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

81 / 561

Properties of Stochastic Integral

Theorem
2
For simple processes h H[0,T
we have
]

Z T

h(t) dWt = 0

1. E
0

2. E

Dale Roberts (c) 2011-2014

h(t) dWt

E[h(t)2 ] d t

Last Modified: 14/10/2014 08:19

82 / 561

Proof of 1

Since,
Z
0

h(t) dWt =

n1
X

hi ()(Wt i+1 Wt i )

i=0

taking expectations of both sides and from the fact that


E[Wt i+1 Wt i ] = 0

gives the result.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

83 / 561

Proof of 2
From the definition, we have
Z T

h(t) dWt =

n1
X

hi ()(Wt i+1 Wt i ).

i=0

So
E

h(s) dWs

=E

n1
X

2
hi (Wt i+1 Wt i )

i=0

=E

n1
X

h2i (Wt i+1

Wt i )

i=0

n1

+ 2E
hi h j (Wt i+1 Wt i )(Wt j+1 Wt j )
i, j=0

i6= j

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

84 / 561

Proof of 2
For i < j , by the tower property of conditional expectations
E[hi h j (Wt i+1 Wt i )(Wt j+1 Wt j )] = E[hi h j (Wt i+1 Wt i )E[(Wt j+1 Wt j )|F t j ]] = 0

Moreover,
E[h2i (Wt i+1 Wt i )2 ] = E[h2i E[(Wt i+1 Wt i )2 |F t i ]]

= E[h2i ](t i+1 t i )

Substituting, we obtain
n1
Z T
Z
X

2
2
=
E
h(t) dWt
E[hi ](t i+1 t i ) =
0

Dale Roberts (c) 2011-2014

i=0

Last Modified: 14/10/2014 08:19

E[h(t)2 ] d t

85 / 561

Martingales

Definition
A F-adapted stochastic process X := (X t ) t0 is a martingale if E[|X t |] < for
each t > 0 and E[X t |Fs ] = X s for all s t .
The following processes are examples of martingales:
the Brownian motion W ;
X t := Wt2 t ;
E t := exp(cWt 12 c 2 t), where c R.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

86 / 561

Stochastic Integrals as Martingales

Lemma
2
If h is a simple process in H[0,T
, then the process
]

Z
X t :=

h(s) dWs ,

tT

is a martingale.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

87 / 561

Construction

We have seen how to construct stochastic integrals of simple processes


This construction can now be extended to the full stochastic integral
under some conditions on the integrand h
The construction uses the idea that we can approximate the integrand h
by simple processes and then take the limit
The full mathematical proof is about taking care of all the technical
aspects of convergence to the limiting object and existence of this object

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

88 / 561

Square-integrable Processes

2
Let h := (h t )0tT be a stochastic process. We say that h is in the set of H[0,T
]
processes if, for all t [0, T ], we have

Z t

|h(s)|2 ds < .

E
0

2
Note: Processes in H[0,T
can be approximated arbitrarily close by simple
]
2
processes in H[0,T ] .

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

89 / 561

Construction

Theorem
2
Let W be a BM. For every h H[0,T
the stochastic integral
]

I t (h) :=

h(s) dWs ,

tT

(2)

exists.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

90 / 561

Properties
Let W be a BM. Then the stochastic integral given by (2) has the following
properties:
2
linearity: I t (h + g) = I t (h) + I t (g) for h, g H[0,T
and , R;
]
2
then I is a continuous martingale;
If h H[0,T
]

The It isometry holds:

Z T
2
Z
h(t) dWt =
E
0

Dale Roberts (c) 2011-2014

E[|h(t)|2 ] d t.

Last Modified: 14/10/2014 08:19

91 / 561

It Processes
An It process is the sum of an absolutely continuous drift plus a continuous
local martingale of the form
Z t
Z t
X t = X0 +

(s) ds +

(s) dWs ,

(3)

2
where H[0,T
and is a progressively measurable process satisfying
]
Rt
|(s)| ds < for all t > 0, such that the above integrals are defined.
0
We also write
dX t = (t) d t + (t) dWt .

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

92 / 561

Identifying Martingales from It Processes

If X is an It process with representation dX t = (t) d t + (t)dWt which


RT
satisfies E[( 0 s2 ds)1/2 ] < , then

Dale Roberts (c) 2011-2014

X is a martingale X is driftless ( t 0).

Last Modified: 14/10/2014 08:19

93 / 561

Novikovs Condition

If d X t = t X t dWt for some F -previsible process t , then

Z T
1
2 ds
< X is a martingale.
E exp
2 0 s

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

94 / 561

Quadratic covariation
In addition to (3), let
Yt = Y0 +

(s) ds +

(s) dWs ,

be another It process defined in terms of the same Brownian motion W . The


quadratic covariation of the processes X and Y is given by
Z t
X , Y t =

(s)(s) ds,

and X , X is called the quadratic variation of the process X .

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

95 / 561

Product Rule
Given X t and Yt adapted to the same Brownian motion W ,
dX t = t dWt + t d t,
dYt = t dWt + t d t.

Then
d(X t Yt ) = X t dYt + Yt dX t + t t d t

In the other case, if X t and Yt are adapted to two different and independent
,
Brownian motions W and W
dX t = t dWt + t d t
t + t d t
dYt = t d W

Then
d(X t Yt ) = X t dYt + Yt d X t
Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

96 / 561

Dale Roberts (c) 2011-2014

Stochastic Differential Equations

Last Modified: 14/10/2014 08:19

97 / 561

Stochastic Differential Equations


Let (t, x) and (t, x) be measurable functions of t [a, b] and x R.
Consider the stochastic differential equation (SDE)
d X t = (t, X t ) d t + (t, X t ) dWt ,

X a = R,

(4)

which should be interpreted as the stochastic integral equation


Z t
Z t

Dale Roberts (c) 2011-2014

Xt = +

(s, X s ) ds +

(s, X s ) dWs ,

a t b.

(5)

Last Modified: 14/10/2014 08:19

98 / 561

Solutions
A stochastic process X := (X t ) t0 is a (strong) solution to (4) if it satisfies:
The stochastic process (t, X t ) H 2 so that
Z t
(s, X s ) dWs

is a well-defined It integral for each t [a, b];


Almost all sample paths of (t, X t ) satisfy
Z b

|(s, X s )| ds < ;

For each t [a, b], (5) holds almost surely.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

99 / 561

Existence and Uniqueness

To obtain solutions of (4), one needs to impose conditions on the functions


(t, x) and (t, x). Depending on the conditions imposed one can obtain:
existence of (potentially multiple) solutions;
existence of a unique solution;

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

100 / 561

Lipschitz and Linear Growth


We say that a measurable function f (t, x) on [a, b] R satisfies the Lipschitz
condition in x if there exists a constant K > 0 such that
| f (t, x) f (t, y)| K|x y|,

a t b, x, y R.

We say that a measurable function f (t, x) on [a, b] R satisfies the Linear


growth condition in x if there exists a constant K > 0 such that

Dale Roberts (c) 2011-2014

| f (t, x)| K(1 + |x|),

a t b, x R.

Last Modified: 14/10/2014 08:19

101 / 561

Existence of Unique Solution

Theorem
Suppose that (t, x) and (t, x) satisfy the Lipschitz and linear growth
conditions in x for all t 0. Then for every initial time a R+ and point R,
there exists a unique solution X of the SDE (4).

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

102 / 561

Dale Roberts (c) 2011-2014

Simulating Solutions of SDE

Last Modified: 14/10/2014 08:19

103 / 561

Overview

As with ODE, most SDE do not admit explicit solutions


Therefore, we need to consider numerical methods to solve them
However, there are fundamental differences between ODE and SDE

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

104 / 561

Differences between ODE and SDE

The two main differences between ODE and SDE are:


A solution to an SDE is a (function-valued) random variable, and thus we
obtain different solutions for different .
A (strong) solution to an SDE is not smooth, as the underlying Brownian
motion is not smooth at all.
The second difference means that numerical schemes for ODEs that rely on
this property are not automatically good when adapted to the SDE situation.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

105 / 561

Goal of Approximation

Numerical schemes for SDE depend on our specific goal. Are we interested in
obtaining a path t 7 Xe (t, ) that is as close as possible to the (unknown)
solution path t 7 X (t, ),
or in computing an expectation of a functional E[g(X )] of the SDE?
The first is called strong approximation and the second is weak
approximation.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

106 / 561

Exact Simulation
In some rare cases, we can find an explicit solution to the SDE. For example,
the SDE
dX t = aX t d t + bX t dWt , X 0 = x,
has the explicit solution

1 2
X t = x exp (a b )t + bWt = f (t, Wt )
2

Then calculating E[g(X T )] for some function g : R R is no different from


simulating a random variable.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

107 / 561

Approximating Paths

The task becomes a lot more difficult when we want to calculate the
expectation of the function of the whole path t 7 X t
Even in the case of Brownian motion, we need to find a way to
approximate the path
We follow the same approximation idea: approximate the paths on an
partition of the interval [0, T ] by simulating the Brownian motion on that
partition and then constructing the values of X t on the partition

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

108 / 561

Approximating Paths
Theorem
Assume that the SDE
d X t = a(t, X t ) d t + (t, X t ) dWt ,

X0 = x

has an explicit solution of the form X (t) = f (t, Wt ) with f a continuous,


real-valued function. Let Xen be an approximation of X that is constructed by
Xen (t) = f (t, Wt ),

if t = iT /n for some i = 0, 1, . . . , n

and extended to all t [0, T ] by linear interpolation. Then, for each bounded
and continuous function g of the path of X , we have
E[g(Xen )] E[g(X )],

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

109 / 561

Strong Approximation

Consider the SDE


d X t = a(t, X t ) d t + (t, X t ) dWt ,

X0 = x

and consider a partition = {0 = t 0 < t 1 < < t n = T } of the interval


[0, T ]. We approximate the SDE by X 0 = x and for i = 0, . . . , n 1 by
X t i+1 = X t i + a(t i , X t i )(t i+1 t i ) + (t i , X t i )(Wt i+1 Wt i ).

This is called the Euler-Maruyama scheme.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

110 / 561

Euler-Maruyama Scheme

Let t := T /N for a given N so that t i = it . Then approximate the SDE by:


1. Set Xe (0) = X (0) = x
2. For i = 0 to N 1 do
(a) Simulate a Zi N(0, 1)
(b) Set
p
Xet i+1 = Xet i + a(t i , Xet i )t + (t i , Xet i ) t Zi

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

111 / 561

Week 4

Its Formula
Multidimensional Stochastic Calculus
Stochastic Exponential
Black-Scholes Market

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

112 / 561

Dale Roberts (c) 2011-2014

Its Formula

Last Modified: 14/10/2014 08:19

113 / 561

Chain Rule
Remember the (calculus) chain rule states that
d
f (g(t)) = f 0 (g(t))g 0 (t)
dt
Or for a function f (t, x t ) in terms of total derivatives
df
f d t f d xt
=
+
dt
t dt x dt

and multiplying throught by d t gives

Dale Roberts (c) 2011-2014

d f (t, x t ) =

f
f
(t, x t )d t +
(t, x t )d x t
t
x

Last Modified: 14/10/2014 08:19

114 / 561

Its Formula

Its formula is one of the most useful tools in stochastic calculus


It is a stochastic version of the chain rule
In the stochastic version, we get an extra correction term

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

115 / 561

Its Formula for Brownian Motion


Let (Wt ) t0 be a BM. What are the dynamics of f (t, Wt )?

Theorem
For a function f : [0, T ] R such that the partial derivatives
f
2
we have
exist and are continuous and for which x H[0,T
]
f (t, Wt ) f (0, 0) =

Z
0

f
(s, Ws ) dWs +
x

Z
0

f
1
(s, Ws ) ds+
s
2

f
t

, x and x 2

2f

2f
(s, Ws ) ds
x2

Alternatively written:

Dale Roberts (c) 2011-2014

d f (t, Wt ) =

f
f
12f
(t, Wt ) dWt +
(t, Wt ) d t +
(t, Wt ) d t
x
t
2 x2

Last Modified: 14/10/2014 08:19

116 / 561

Applications

E[Wtn ]
Solving SDEs

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

117 / 561

Its Formula for Processes


Let (X t ) t0 satisfy dX t = (t, X t ) d t + (t, X t ) dWt . What are the dynamics of
f (t, X t )?

Theorem
For a function f (t, x) that is twice differentiable in x and once in t . Then for
Yt = f (t, X t ), we have
d Yt =

f
f
2f
1
(t, X t ) d t +
(t, X t ) dX t + (t, X t )2 2 (t, X t )d t
t
x
2
x

Expanding out the dX t term,





2
f
f
f
1
2 f
(t, X t ) + (t, X t )
(t, X t ) + (t, X t )
(t,
X
)
d
t+(t)
(t, X t ) dWt
t
t
x
2
x2
x

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

118 / 561

Dale Roberts (c) 2011-2014

Multidimensional Stochastic Calculus

Last Modified: 14/10/2014 08:19

119 / 561

Multidimensional Brownian Motion


Definition
A standard n-dimensional Brownian motion W := (W (t)) t0 is defined as

W1 (t)
W2 (t)

W (t) =
...
Wn (t)

and satisfies the following two properties:


1. Each of the processes (W j (t)) t0 with j = 1, . . . , n, is a one-dimensional
Brownian motion.
2. The Brownian motions Wi and W j are independent for i 6= j and index
values i, j = 1, . . . , n.
Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

120 / 561

Multidimensional Brownian Motion


Correlation of Brownian motions can be modeled using weighted sums of
independent Brownian motions or multiplication by a covariance matrix.

Example
Let (W1 (t)) t0 and (W2 (t)) t0 be two independent Brownian motions. For
1 +1, let

W (t) = W1 (t) + 1 2 W2 (t).


Then (W (t)) t0 is a Brownian motion and
E(W1 (t) W (t)) = t.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

121 / 561

Multidimensional Stochastic Integration

We would like to extend the idea of stochastic integration to multiple


dimensions.
To handle things like:
Z T

W1 (t) dW2 (t)

or

sin(W12 (t) + W22 (t)) dW2

We will use vectors and matrices.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

122 / 561

Multidimensional Stochastic Integration


2
be the set of Rn -valued adapted processes
Let H[0,T
]

h(t) = [h1 (t) h2 (t) hn (t)]

such that

Z

kh(t)k d t < .

Recall: kxk =

Dale Roberts (c) 2011-2014

x 12 + x 22 + + x n2 .

Last Modified: 14/10/2014 08:19

123 / 561

Multidimensional Stochastic Integration


Let h be Rn -valued, (W (t)) t0 be a n-dimensional Brownian motion, and
times S < T .
2
For every h H[0,T
one can define the stochastic integral
]
T

h(t) dW (t) =

I T :=
S

n Z
X
i=1

h j (t) dW j (t).

2
If h H[0,T
then the process (I t ) tS is a martingale and
]

2
h(t) dW (t)

=E

kh(t)k2 d t .

I is one-dimensional in this case (i.e., h a vector).


Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

124 / 561

Multidimensional Stochastic Integration


Let H be a m n matrix of progressive processes

h11 (t) h1n (t)


.. .
H(t) = ...
.
hm1 (t) hmn (t)
2
then
If hi j H[0,T
]

I =

Z
S

H(t) dW (t) =

h11 (t)
...

h1n (t)
dW1 (t)
.. ..
.
.

hm1 (t) hmn (t)

dWn (t)

In other words, I is a m 1 matrix (i.e., column vector) whose i th element


is given by
n Z T
X
(I )i =
hi j (t) dW (t).

Dale Roberts (c) 2011-2014

j=1

Last Modified: 14/10/2014 08:19

125 / 561

Multidimensional Diffusion Processes


Since we have defined multidimensional stochastic integration, we can
now consider multidimensional diffusion processes or It processes.
Of the form
dX t = a(t) d t + b(t) dW (t),

X (0) = x R

where a(t) R, b(t) = [b1 (t) b2 (t) bn (t)] with W an n-dim BM.
Or of the form
d X t = a(t) d t + B(t) dW (t),

X (0) = x Rn

where a(t) Rn , B(t) Rmn , and W an m-dim BM.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

126 / 561

Multidimensional Diffusion Processes


A stochastic differential with drift and diffusion coefficient a and b for X in
component form is given by
dX i (t) = ai (t) d t +

m
X

bi j (t)dW j (t),

j=1

for t in [0, T ] and i = 1, . . . , n.


The corresponding integral form for the components is
Z t
m Z t
X
X i (t) = X i (0) +

ai (s)ds +

bi j (s)dWsj ,

j=1

for t in [0, T ] and i = 1, . . . , n.


The stochastic differential is just an abbreviation for the corresponding
stochastic integral equation.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

127 / 561

Multidimensional Diffusion Processes


The SDE in vector/matrix form is given by
d X (t) = a(t)d t + b(t)dW (t),

X (0) = x Rn

The column vectors X (t), a(t) and W (t) are given by

a1 (t)
X 1 (t)
X (t) = ... , a(t) = ... and W (t) = [W1 (t), . . . , Wm (t)] T .
X n (t)

an (t)

The matrix b(t) is given by

Dale Roberts (c) 2011-2014

b11 (t) . . .
..
...

b(t) =
.

bn1 (t) . . .
Last Modified: 14/10/2014 08:19

b1m (t)
.. .
.
bnm (t)
128 / 561

Quadratic Variation
Consider the n-dimensional It process
Z t
X (t) = X (0) +

a(s) ds +

B(s) dW (s).

The quadratic variation of the component process X i (t), for i = 1, . . . , n, is


the random integral given by
Z t
Bi (s)0 Bi (s)ds

X i t =

where Bi (t) is the i th column of B(t).

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

129 / 561

Quadratic Covariation
Consider the n-dimensional It process
Z t

a(s) ds +

X (t) = X (0) +

B(s) dW (s).

The quadratic covariation of the component processes X i and X j , for


i, j = 1, . . . , n, is the random integral given by
Z t
Bi (s)0 B j (s)ds.

X i , X j (t) =

The stochastic differential for X i , X j (t) is denoted by dX i , X j t and

Dale Roberts (c) 2011-2014

dX i , X j t = Bi (t)0 B j (t) d t.

Last Modified: 14/10/2014 08:19

130 / 561

Multidimensional It Formula
The multi-dimensional It formula gives the stochastic differential for the
process Y (t) = f (t, X (t)) for some transformation function
f (t, x) = ( f1 (t, x), . . . , f m (t, x))0
The function f is assumed to map from [0, ) Rn Rm and be of class
f (t,x)
2 f (t,x)
C 1,2 (R, Rn ): i.e., the functions (t, x) t and (t, x) x i x j are
continuous on (0, ) Rn for i, j = 1, . . . , n.
Its formula states that the stochastic differential for
Y (t) = (Y1 (t), Y2 (t), . . . , Ym (t))0 for t in [0, T ] is given by

Dale Roberts (c) 2011-2014

X f (t, X (t))
f k (t, X (t))
k
dt +
d X i (t)
d Yk (t) =
t

x
i
i=1
n
2
X
f k (t, X (t))
1
+
dX i , X j (t).
2 i, j =1 x i x j
Last Modified: 14/10/2014 08:19

131 / 561

Multidimensional It Formula
The multi-dimensional It formula using drift and diffusion coefficients is
given by

n
f (t, X (t))
f (t, X (t)) X
d Y (t) = d f (t, X (t)) =
ai (t)
+
t
xi
i=1
!
n
2

f
(t,
X
(t))
1 X
bi (t) T b j (t)
+
dt
2 i, j =1
xi x j
n

m
X
X
f (t, X (t))
+
bi j (t)
dW j (t).

x
i
j=1 i=1
This
by substitution of the differentials for d X i (t) and

is obtained

d X i , X j (t) using the covariation form of Its formula.
Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

132 / 561

Change of Measure

Girsanovs theorem explains how the dynamics change for an


n-dimensional Brownian motion under a change in measure.
The change in measure from P to some other measure Q is related to what
is called the Radon-Nikodym derivative for the two measures.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

133 / 561

Change of Measure
Let (W (t)) t0 be an m-dimensional Brownian motion under P .
Let ( (t)) t0 be an F-adapted n-dimensional diffusion process that
satisfies the Novikov condition.
Then there is a measure Q with the properties:
1. Q is equivalent to P .
f with
2. The process W
Z t
f(t) = W (t) +
(s) ds
W
0

is an m-dimensional BM under Q.
In differential form we have
f(t) = dW (t) + (t)d t.
dW
Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

134 / 561

Girsanovs Theorem
Theorem (Girsanov)
2
and X satisfy dX t = (t) dWt . If E t (X ) is a uniformly integrable
Let H[0,T
]
martingale with E (X ) > 0 then
n Z T

Z T
X
dQ
1
i (s) dWsi +
= E T (X ) = exp
2i (s) ds
dP
2
0
0
i=1

defines an equivalent probability measure Q P and the process


Z t
f(t) := W (t) +
W
(u) du
0

is a Q-Brownian motion.
Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

135 / 561

Martingale Representation Theorem


Suppose we have two n-dimensional diffusion processes X (t) and Y (t)
given by
d X (t) = H(t) dW (t)

d Y (t) = F (t) dW (t).

H and F are n n progressive processes.


The martingale representation theorem states that there exists a unique
matrix process (t) = [i j (t)] such that

Dale Roberts (c) 2011-2014

d Y (t) = (t) dX (t)

or

dYi (t) =

n
X

i j (t)d X j (t).

j=1

Last Modified: 14/10/2014 08:19

136 / 561

Dale Roberts (c) 2011-2014

Stochastic Exponential

Last Modified: 14/10/2014 08:19

137 / 561

Stochastic Exponential
The stochastic exponential of an It process X is defined as
1

E t (X ) = e X t 2 X ,X t

The stochastic exponential E t (X ) has the following properties:


E t := E t (X ) is a positive It process and the unique solution to the SDE
d E t = E t dX t ,

E0 = e X 0 ;

E (X ) is a continuous local martingale if X is a local martingale;


E (0) = 1;
E (X )E (Y ) = E (X + Y )eX ,Y ;
E (X )1 = E (X )eX ,X .

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

138 / 561

Dale Roberts (c) 2011-2014

Black-Scholes Market

Last Modified: 14/10/2014 08:19

139 / 561

Black-Scholes Market
The Black-Scholes market typically consists of a risky asset S whose price
dynamics is given by the solution of the SDE
dS t = S t ( d t + dWt ),

S0 = 1

and a risk-free asset B with value B t satisfying


dB t = B t r d t

B0 = 1.

Writing dS t = S t d X t with
dX t = d t + dWt ,

X 0 = 0,

it follows from the stochastic exponential that


S t = S0 E (X ) = S0 exp(t 12 2 t).
Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

140 / 561

Self-financing Portfolios
A portfolio or trading strategy is a pair of processes (, ) that give the amount
of holding at time t of the asset pair (S, B). The value process of this portfolio is
V (t) = (t)S(t) + (t)B(t)

The portfolio is called self-financing for (S, B) if


Z t
|(u)| dSu <

for all t > 0,

and there is no inflow or outflow of capital during the trading. In other words,
the change in value of the portfolio over any time period is only due to the
changes in the value of B and S :

Dale Roberts (c) 2011-2014

dV (t) = (t)S(t) + (t)B(t).


Last Modified: 14/10/2014 08:19

141 / 561

Arbitrage Portfolio

A arbitrage portfolio is a self-financing portfolio (, ) with value process V


satisfying
V (0) = 0, V (T ) 0, P{V (T ) > 0} > 0,
for some T > 0.
If no arbitrage portfolio exists then we say that the model is arbitrage-free.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

142 / 561

Martingale Measures

We call a measure Q an equivalent martingale measure to P (EMM), denoted


Q P, if the discounted price process
Z(t) :=

S(t)
B(t)

is a Q-martingale.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

143 / 561

Fundamental Theorem of Asset Pricing


Theorem
Let Q be the EMM. Suppose we have a contingent claim X T F T . If
EQ [X T2 ] < then the claim is replicable and the value at time t of any
replicating portfolio is given by
Vt = B t EQ [B T1 X T |F t ].

In particular, under the Black-Scholes model, the fair price at time zero for the
option is
V0 = EQ [er T X T ]
fT ) where W
f is a Q-BM.
where X T = f (S T ) and S T = S0 exp((r 21 2 )T + W

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

144 / 561

Week 5
Black-Scholes Model
Definition of Implied Volatility
Empirical Observations
Arbitrage Bounds
Asymptotics
Approximating Implied Volatility
Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

145 / 561

Dale Roberts (c) 2011-2014

Black-Scholes Model

Last Modified: 14/10/2014 08:19

146 / 561

Black-Scholes Model
Black-Scholes Model
Under real-world measure P, we assume there exists deterministic r , , and
such that
B t = exp(r t)
S t = S exp(Wt + t),
where r is the risk-free interest rate, is the stock volatility and is the stock
drift, and S := S0 is the current stock price.
There are no transaction costs and both instruments are freely and
instantaneously tradable either long or short at the price quoted.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

147 / 561

Black-Scholes Formula
Explicit solution was obtained by Black and Scholes (1973) relying heavily on
the notion of no-arbitrage in Merton (1973).

Black-Scholes-Merton (BSM) Formula


C = S(d1 ) er T K(d2 )

where

log( KS ) + (r 12 2 )T
d2 :=
,
p
T

p
d1 := d2 + T

Jointly awarded Nobel prize in economics in 1997.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

148 / 561

Benchmark Model

Widely acknowledged that assumptions underlying the BSM model are far
from realistic
BSM model enjoys unrivalled popularity in practice due to its role as a
benchmark model
Provides a convenient mapping device from the space of option prices to a
single real number called implied volatility (IV)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

149 / 561

Dale Roberts (c) 2011-2014

Definition of Implied Volatility

Last Modified: 14/10/2014 08:19

150 / 561

Implied Volatility

Unknown parameter in the BSM formula is : the volatility


Backed out of market option prices, IV gives a straight-forward
comparison of the relative expensiveness of options across various strikes,
expiries, and underlying assets.
It is interesting to plot IV against relative strikes and time to expiry

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

151 / 561

Dale Roberts (c) 2011-2014Fig.

Matthias R. Fengler

IV Surface

100 %

80 %

60 %

40 %

20 %
0.5
1
1.5
2

Moneyness [X/S]

Time to maturity [years]

1. IV surface of DAX index


from 28
Oct. 2008, traded at the EUREX.
Lastoptions
Modified: 14/10/2014
08:19

152 / 561

IV Smile and Term Structure

The function resulting for a fixed expiry is frequently called the implied
volatility smile due to the U-shaped pattern
For a fixed strike across several expiries one speaks of the term structure of
implied volatility
Practioners like to think of them as stemming from a smooth and
well-behaved surface

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

153 / 561

Applications

Market makers use the IV surface to quote prices for strike-expiry pairs
which are illiquid or not listed
Pricing engines used to price exotic options are calibrated against the IV
surface
Risk managers use stress scenarios defined on IV surface to visualise and
quantify the risk inherent to option portfolios

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

154 / 561

Formal Definition
Concept of IV was first introduced by Latan and Rendelman (1976).

Definition
Given an observed market price of European option Ce with strike K and time
to maturity T , the implied volatility of this option is
e:

e ) Ce = 0
C (K, T,

e R+ .
By monotonicity of the BSM price in , there exists a unique solution
Put options can also be used (Put-Call Parity).

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

155 / 561

IV Surface

e is just a constant. However, given the same underlying asset, it varies


The IV
across strikes K and expiry T .

Definition
An implied volatility surface is the mapping
e : (t, K, T ) 7
e t (K, T ).

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

156 / 561

Dale Roberts (c) 2011-2014

Empirical Observations

Last Modified: 14/10/2014 08:19

157 / 561

Empirical Observations

IV surfaces can be observed for indices, stocks, etc.


Different asset classes may display different features: smiles may be more
shallow, symmetric or even upward-sloping
Smiles are very pronounced for short expiries and become flattish for
longer dated options
This has not always been the case. Strong asymmetry appeared after the
1987 market turmoil
For equity options (indices and stocks), the smile is negatively skewed

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

158 / 561

Implied Volatility Skew


We call m := log(K/S) the log moneyness.

Definition
We define the implied volatility skew as

e 2

m m=0

The skew often increases during times of crisis (e.g., dot-com crash 2001-2003,
Sept. 11 2001, GFC 2008).

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

159 / 561

Further Empirical Observations

When expiry is small, the fluctuations of the skew are larger


The IV term structure is typically upward sloping (i.e., increasing levels of
IV for longer dated options) in calm times, while in times of crisis it is
downward sloping with short-dated options having higher levels of IV
than longer dated ones
Returns of IV and returns of underlying asset are negatively correlated
IV appears to be mean-reverting (though this is hard to show statistically)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

160 / 561

Dale Roberts (c) 2011-2014

BSM implied volatility

Time Series of ATM-IV


45 %

1Y ATMIV

DAX
8000

40 %
7000
35 %
6000
30 %
5000
25 %
4000
20 %
3000
15 %
2001

2003

2005

Last Modified: 14/10/2014 08:19

2007

2009

161 / 561

Dale Roberts (c) 2011-2014 Fig.

Matthias R. Fengler

Time Series of IV Skew

0.2
50

0.6

1M skew
1Y skew

0.8

IV term structure

40

30

20

1.2

10

IV term structure

Skew

0.4

1.4
0
1.6
2001

2003

2005
Time

2007

2009

3. Time series of 1M and


IV 14/10/2014
skew (left
Last1Y
Modified:
08:19 axis, gray line and black line

162 / 561

Dale Roberts (c) 2011-2014

implied volatility
Empirical Quantiles ofBSMATM-IV

0
0.1
0.2

Skew

0.3
0.4
0.5
95%
75%
Median
25%
5%

0.6
0.7
0.8
0.9

1M 3M

1Y
Time to expiry

2Y

Fig. 4. Empirical quantiles


of the ATM IV skew as a function of time to expiry.
!
Last Modified: 14/10/2014 08:19
2

163 / 561

Dale Roberts (c) 2011-2014

Arbitrage Bounds

Last Modified: 14/10/2014 08:19

164 / 561

Arbitrage Bounds on IV Surface


IV cannot simply take any value due to constraints imposed by the
no-arbitrage principles. For IV, these are very difficult to write down, but easily
stated indirectly in the option price domain.

Arbitrage Bounds
e must be such that the call price is bounded above and below:
The IV
e) S
(S er T K)+ C (K, T,

The call price must be decreasing and convex function in K :


er T

Dale Roberts (c) 2011-2014

C
e) 0
(K, T,
K

and

Last Modified: 14/10/2014 08:19

2C
e) 0
(K, T,
K2

165 / 561

Arbitrage Bounds on IV Surface


The quantity x := K/F T is called forward-moneyness where F T is the price of
the forward with expiry T .

Calendar Arbitrage Bounds


Let T1 < T2 be the expiry dates of two call options whose strike prices K1 and
K2 are related by forward-moneyness, i.e. x 1 = x 2 . Then
e 2 ) C (K1 , T1 ,
e 1)
C (K2 , T2 ,

must hold.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

166 / 561

Dale Roberts (c) 2011-2014

Asymptotics

Last Modified: 14/10/2014 08:19

167 / 561

Asymptotics

Asymptotics are mathematical results obtained when you take one of the
variables to infinity or to zero
These types of results provide interesting theoretical information about
the model
They are often easier to establish and help reduce the complexity

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

168 / 561

Asymptotics for the IV Surface

Many results were conjectures until recently


Understanding the behaviour of IV for far expiries and far strikes is of utter
importance for extrapolation problems often arising in practice
These results are model independent
m = log(K/S) is the (simple) log-moneyness

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

169 / 561

IV Surface Flattens
The first theorem shows that the IV surface flattens for infinitely large expiries.

Theorem (Rogers and Tehranchi 2009)


For any M > 0 we have
lim

sup

T m ,m [M ,M ]
1
2

e (m1 , T )| = 0.
|e
(m2 , T )

Note: sup means supremum which is equivalent to the maximum if there are
finitely many elements in the set.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

170 / 561

Flattening Rate
The rate of flattening of the IV skew can be made more precise by the following
result.

Theorem (Rogers and Tehranchi 2009)


1. For any 0 m1 < m2 , we have
e (m2 , T )2
e (m1 , T )2

m2 m1
T

2. For any m1 < m2 0,


e (m2 , T )2
e (m1 , T )2

m2 m1
T

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

171 / 561

Representation Formula
The level of the IV for far expiry can be written in terms of the expected share
price.

Theorem (Tehranchi 2009)


For any M > 0, we have


v


t
8


e (m, T ) log E[S T 1] = 0
lim
sup
T m[M ,M ]

T

Note: a b := min(a, b).

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

172 / 561

At Short Expiry
Theorem (Roper and Rutkowski 2009)
If C (K, ) = (S K)+ for some > 0 then
e (K, T ) = 0.
lim

T 0+

Otherwise,
p
2C (K, T )

,
p

S T
e (K, T ) =
lim
T 0+
| log(S/K)|

,
p
2T log(C (K, T ) (S K)+ )

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

S=K
S 6= K

173 / 561

At Short Expiry

The asymptotics are different when S = K and S 6= K


The ATM case is similar to Brenner and Subrahmanyam approximation
formula
The limit does not need to exist

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

174 / 561

At Far Strike
Lee has established a one-to-one correspondance between the large-strike tail
and the number of moments of S T and the small-strike tail and the number of
moments of S T1 .

Theorem (Lee 2004)


1+p

Define e
p := sup{p : E[S T ] < } and
R = lim sup
m

Then R [0, 2] and


e
p=

e2

|m|/T

R 1
1
+
.
2R
8
2

Equivalently,

R = 2 4( e
p2 + e
pe
p).
Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

175 / 561

At Far Strike
Theorem (Lee 2004)
q

Define e
q := sup{p : E[S T ] < } and
L = lim sup
m

Then L [0, 2] and


e
q=

e2

|m|/T

1
1
+ L .
2 L
8
2

Equivalently,

L = 2 4( e
q2 + e
qe
q).

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

176 / 561

Dale Roberts (c) 2011-2014

Approximating Implied Volatility

Last Modified: 14/10/2014 08:19

177 / 561

Approximation Formulas

There is no closed-form formula for IV, even for European options


e (e.g., Goal Seek in Excel)
A numerical solver has to be used to find
Approximation formulas can be useful when you dont have a solver
(Spreadsheet)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

178 / 561

BSF Approximation
The simplest approximation to IV is due to Brenner and Subrahmanyam (1988)
and Feinstein (1988).

BSF Approximation
e

v
t 2 C
T S

This formula is only accurate when S = Ker T .

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

179 / 561

Li Approximation
A more accurate formula was proposed by Li (2005) and hold for ITM (S  K )
and OTM (S  K ) options.

Li Approximation

q
2z T2 p1T 8z 2 p62z


1 
4(KS)2
2
p
+ S(S+K)
2 T

where z = cos( 31 arccos( p332 )), :=


:= |Ker T S|SC 2 .

Dale Roberts (c) 2011-2014

p
2
S+K (2C

if 1.4
if > 1.4

+ Ker T S) and

Last Modified: 14/10/2014 08:19

180 / 561

Corrado-Miller Approximation
Other approximation formulas often lack a rigourous mathematical
foundation. The most prominent amongst these are those suggested by
Corrado and Miller (1996) and Bharadia et al. (1996)

Corrado-Miller Approximation

v
p
u
2
1
2
SX t
SX
(S X )2
ep

C
+
C

2
2

T S+X

where X = Ker T .

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

181 / 561

Bharadia et al. Approximation

Bharadia et al. Approximation


v
t 2 C (S X )/2
e
.

T S (S K)/2

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

182 / 561

Accuracy of Approximations

Isengildina-Massa et al. (2007) investigate the accuracy of six


approximation formulas
According to their criteria, Corrado and Miller (1996) is the best, followed
by Li (2005) and Bharadia et al. (1996)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

183 / 561

Week 5b

Bonds

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

184 / 561

Dale Roberts (c) 2011-2014

Bonds

Last Modified: 14/10/2014 08:19

185 / 561

Types of Bonds

A bond is a securitized form of a loan


The buyer of the bond lends the issuer an initial price P in return for a
predetermined sequence of payments
The payments may be fixed in nominal terms or linked to some index (e.g.,
CPI)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

186 / 561

Types of Bonds

In UK, government bonds are called gilt-edged securities (gilts)


Most other countries: treasury bills or treasury notes
In Australia: Treasury Fixed Coupon Bonds and Treasury Capital Indexed
Bonds

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

187 / 561

RBA Treasury Fixed Coupon Bonds

Pays interest on a semi-annual basis at the prescribed coupon rate,


applied to the face value
At maturity, the face value amount is repaid
More details:

http://www.aofm.gov.au/content/_download/Treasury_Bond_
Information_Memo_18_December_2009.pdf

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

188 / 561

Treasury Capital Indexed Bonds

Pays interest on a quarterly basis at the prescribed coupon rate, applied to


the face value
However, the face value is adjusted by indexing the principal to inflation
At maturity, investors receive the adjusted capital value of the security
More details:

http://www.aofm.gov.au/content/_download/Treasury_
Indexed_Bond_Information_Memo_18_December_2009.pdf

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

189 / 561

Australian Bonds (small investors)

Subject to availability, a range of series across different maturity dates will


be on sale at any one time.
Applications may be made for amounts of $1,000 face value and in
multiples of $1,000 up to $250,000 per investor per day (all series
combined).
Indicative prices:

http://www.rba.gov.au/fin-services/bond-facility/xls/
otc-prices.xls

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

190 / 561

Types of Bonds

Bonds are also issued by state government (NSW, QLD, etc), banks, and
companies
Bonds issued by companies are called corporate bonds

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

191 / 561

Types of Bonds

Some corporate bonds trade on the ASX (here)


Some are OTC like BHP bonds (here)
Australias corporate bond market is very small compared to the US or
Europe

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

192 / 561

Types of Bonds

Bonds with identical characteristics but sold by different issuers may have
different prices
Example: 2 bonds of 20Y with 6% coupon, one issued by a company and
one issued by RBA
Bond issued by company will probably trade at a lower price because
market participants take into account the possibility of default on the
payments or the redemption process

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

193 / 561

Types of Bonds

We will not consider or model credit risk in this course


We only model risk-free government bonds
Government bonds are not always riskless: think about Greece and US at
the moment

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

194 / 561

Fixed Coupon Bonds

We will only concentrate on fixed coupon (fixed-interest) government


bonds that have no probability of default
We pay a price P for a bond in return for a stream of payments c1 , c2 , . . . , cn
at times t 1 , t 2 , . . . , t n from now
The amount of the payments are fixed at the time of issue

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

195 / 561

RBA Treasury Fixed Coupon Bonds

Pays interest on a semi-annual basis at the prescribed coupon rate,


applied to the face value
At maturity, the face value amount is repaid
More details:

http://www.aofm.gov.au/content/_download/Treasury_Bond_
Information_Memo_18_December_2009.pdf

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

196 / 561

Notation

g : coupon rate per 100 nominal


n: number of coupon payments
t (or ): fixed time between payments
t 1 : time of first payment ( t 1 t )
t j = t j1 + t for j = 2, . . . , n
t n : time to redemption

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

197 / 561

RBA Treasury Fixed Coupon Bonds

Interest is paid on semi-annual basis: t = = 0.5


We assume Australian Government will not default: P(T, T ) = 1 for all T
We assume that RBA offers all maturities T

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

198 / 561

Fixed Coupon Bonds

first coupon
0
if bond has gone ex-dividend
c j = gt for j = 2, . . . , n 1
cn = 100 + gt
c1 =

Dale Roberts (c) 2011-2014

gt

Last Modified: 14/10/2014 08:19

199 / 561

Clean and Dirty Prices

Bond prices are often quoted in two different forms


The dirty price is the actual amount paid in return for the right to the full
amount of each future coupon payment and the redemption proceeds
The clean price is an artificial price which is typically quoted in the market
place

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

200 / 561

Dirty price

Assume semi-annual (twice a year) coupon payments


If the bond has gone ex-div, then the dirty price will give the buyer the
right to the full coupon payable in just over six months
But not the coupon due in a few days
The dirty price of a bond will drop by an amount approximately equal to a
coupon payment at the time it goes ex-div
The dirty price of a bond typically rises steadily in between ex-div dates

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

201 / 561

RBA Prices

Banks selling price includes coupon interest accrued since the last
coupon payment and up to the date of purchase
Purchasers of Bonds subsequently receive a full coupon payment when it
next falls due, irrespective of when the purchase takes place in the coupon
period.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

202 / 561

RBA Prices

The only exception to this is when Bonds are purchased during the
ex-interest period, which commences seven days prior to a coupon
payment date. Investors who purchase Bonds during the ex-interest
period do not receive the impending coupon payment.
As a result, Bond prices in the ex-interest period do not include accrued
interest and are commensurately lower.
Question: Are RBA prices clean or dirty?

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

203 / 561

Zero-Coupon Bonds

A ZCB is a bond with a coupon rate of zero and a nominal value of 1


We denote the price at time t of a ZCB that matures at time T by P(t, T )
Sometimes we call this a T -ZCB or T -bond
Note that the value of $1 due immediately is P(t, t) = 1 for all t
Arbitrage arguments give P(t, T ) 1 for all T

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

204 / 561

Dale Roberts (c) 2011-2014

ZCB Cash Flow

P(t, T )
t

1
T

P(t, T )
t

1
T

Last Modified: 14/10/2014 08:19

205 / 561

Modelling Assumptions

There exists a frictionless market for T -bonds for every T > 0


P(T, T ) = 1 for all T
P(t, T ) is differentiable in T

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

206 / 561

Dale Roberts (c) 2011-2014

Term structure
Price P(t, T )
1

0
Years

T 7 P(t, T )
Last Modified: 14/10/2014 08:19

207 / 561

Dale Roberts (c) 2011-2014

T -Bond Price Process


1

0
Years

t 7 P(t, T )
Last Modified: 14/10/2014 08:19

208 / 561

Modelling Assumptions

In reality these assumptions are not always satisfied


Zero-coupon bonds are not traded for all maturities
It might happen that P(T, T ) 1 if the issuer defaults
The condition that P(t, T ) is differentiable in T is purely technical: it
ensures that T 7 P(t, T ) is a smooth curve

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

209 / 561

Simple vs. Continuous Compounding


Compounding Rate
1
R
2
R/2
m
R/m
..
..
.
.

eR = 1 + R + o(R)

t =0
1
1
1
..
.

t =1
1+R
(1 + R/2)2
(1 + R/m)m
..
.
eR

for small R

e0.04 = 1.04081 but e0.04 1.04081 = 8.1bp

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

210 / 561

Week 6
Interest Rates
Arbitrage-Free Pricing
Fundamental Theorem of Asset Pricing
The Interest Rate Setting
Blacks model

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

211 / 561

Dale Roberts (c) 2011-2014

Interest Rates

Last Modified: 14/10/2014 08:19

212 / 561

Spot Rates
The spot rate at time t for maturity at time T is defined as the yield to
maturity of the T -bond:
R(t, T ) =
In other words,

log P(t, T )
Tt


P(t, T ) = exp (T t)R(t, T )

If we invest $1 at time t in the T -bond for T t years, then this will


accumulate at an average rate of R(t, T ) over the whole period

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

213 / 561

Forward Rates

The forward rate at time t which applies between times T and S


(t T < S) is defined as
F (t, T, S) =

1
P(t, T )
log
ST
P(t, S)

The forward rate arises from a contract: agree at time t that we will invest
$1 at time T in return for e(ST )F (t,T,S) at time S

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

214 / 561

No-arbitrage argument

e (ST )F (t,T ,S)

By definition, contract has value zero at time t provided F (t, T, S) is the fair
forward rate between T and S
We argue that:

Dale Roberts (c) 2011-2014

F (t, T, S) =

P(t, T )
1
log
ST
P(t, S)

Last Modified: 14/10/2014 08:19

215 / 561

No-arbitrage argument
Suppose this is not true and that
F (t, T, S) >

1
P(t, T )
log
ST
P(t, S)

Then setup a portfolio at time t :

Units
Value
Forward
+1
0
T -Bond
+1
P(t, T )
S -Bond P(t, T )/P(t, S) P(t, T )
0

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

216 / 561

No-arbitrage argument

Hold the contracts until maturity of the Forward


At time T the portfolio is worth 0
At time S , the portfolio is worth
e(ST )F (t,T,S)

P(t, T )
>0
P(t, S)

We have an arbitrage: we started with a portfolio with value 0 at time t


and have a sure profit at S

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

217 / 561

No-arbitrage argument
Hence, we cannot have
F (t, T, S) >

1
P(t, T )
log
ST
P(t, S)

By reversing the argument, we cannot have


F (t, T, S) <

1
P(t, T )
log
ST
P(t, S)

F (t, T, S) =

P(t, T )
1
log
ST
P(t, S)

We conclude that

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

218 / 561

Forward-rate curve

The instantaneous forward rate curve (or forward-rate curve) at time t is


for t < T given by
f (t, T ) = lim F (t, T, S)
ST

..
.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

219 / 561

Forward-rate curve

lim F (t, T, S) =

ST

log P(t, T )
T

T P(t, T )
P(t, T )

Therefore,

Dale Roberts (c) 2011-2014

P(t, T ) = exp

f (t, u) du

Last Modified: 14/10/2014 08:19

220 / 561

Dale Roberts (c) 2011-2014

Forward-rate curve

f (t, u)

A=

RT
t

f (t, u) du

Last Modified: 14/10/2014 08:19

221 / 561

Dale Roberts (c) 2011-2014

Forward-rate curve

F (t, u, u + )

u u+

f (t, u) = lim F (t, u, u + )


0

Last Modified: 14/10/2014 08:19

222 / 561

Forward Rates

It is a lot easier to work with the instantaneous forward-rate f (t, T ) than


the forward F (t, T, S) between T and S
The forward curve T 7 f (t, T ) is positive for all T t
This implies that P(t, T ) is a decreasing function of T

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

223 / 561

Short Rate
R(t, T ) is the risk-free rate of interest over the fixed period from t to T
When we talk about the risk-free rate of interest we mean the
instantaneous risk-free rate:
r(t) = lim R(t, T ) = R(t, t) = f (t, t)
T t

Think of r(t) as the rate of interest on a bank account: this can change on
a daily basis
r(t) is often called the short rate

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

224 / 561

Short Rate
R(t, T )

r (t)
t

Dale Roberts (c) 2011-2014

tT

r (t) = limT t R(t, T )

Last Modified: 14/10/2014 08:19

225 / 561

Proxies for Short Rate

The short rate r(t) is fundamental to no-arbitrage pricing


However, it can not be directly observed
The overnight interest rate is not usually considered a good proxy for r(t)
Some people use a slightly longer rate as a proxy: the seven-day
Eurodollar rate
Ideally, you want to use a liquid security as a proxy

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

226 / 561

Par Yields

The par-yield curve (t, T ) specifies the coupon rates,


100(t, T ),

at which new bonds should be priced if they are to be issued at par


Price of 100 per 100 nominal

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

227 / 561

Par Yields
Par yield for a maturity T (annual coupons, = 1), solve following
equation for coupon rates (t, T ):
100 = 100(t, T )

T
X

P(t, s) + 100P(t, T )

s=t+1

Solution given by

Dale Roberts (c) 2011-2014

1 P(t, T )
(t, T ) = P T
s=t+1 P(t, s)

Last Modified: 14/10/2014 08:19

228 / 561

Yield to Maturity

Yield to maturity usually applies to coupon bonds


It is the internal rate of interest (IRR) of the bond
Given the price of a bond P , coupons c1 , c2 , . . ., then

find :

P(t, T ) =

n
X

c j e(t j t)

j=1

is the YTM

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

229 / 561

Relationships

For a given t , each of the curves P(t, T ), f (t, T ) and R(t, T ) uniquely
determines the other two
Recall that we have

Dale Roberts (c) 2011-2014

P(t, T ) = exp (R(t, T )(T t))


Z T

= exp

f (t, u) du

Last Modified: 14/10/2014 08:19

230 / 561

Dale Roberts (c) 2011-2014

Relationships

f (t, u)
R(t, T )

A = R(t, T )(T t) =

RT
t

f (t, u) du

P(t, T ) = exp(A)

Last Modified: 14/10/2014 08:19

231 / 561

Dale Roberts (c) 2011-2014

Example

T
F (0, T, T + 1)

0
1
2
3
0.0420 0.0500 0.0550 0.0560

P(0, T ) = exp

T 1
X

F (0, t, t + 1)

t=0

T
P(0, T )

0
1
2
3
0.95887 0.91211 0.86239 0.81628

Last Modified: 14/10/2014 08:19

232 / 561

Dale Roberts (c) 2011-2014

Example

R(0, T ) =
T
F (0, T, T + 1)
P(0, T )
R(0, T )

log P(0, T )
T +1

0
1
2
3
0.0420 0.0500 0.0550 0.0560
0.95887 0.91211 0.86239 0.81628
0.0420 0.0460 0.0490 0.05075

Last Modified: 14/10/2014 08:19

233 / 561

Dale Roberts (c) 2011-2014

Arbitrage-Free Pricing

Last Modified: 14/10/2014 08:19

234 / 561

Arbitrage

We are able to construct at time 0 some portfolio which has a NPV of zero
At some fixed time T > 0, the portfolio gives us a sure profit
This is sometimes called a free lunch

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

235 / 561

Arbitrage

The concept of arbitrage becomes clear when we consider static portfolios


Also known as buy-and-hold strategies

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

236 / 561

Arbitrage with Static Portfolios

Suppose we can invest in n assets


Asset i has price Pi (t) at time t
Asset i pays no dividend or coupons
We have x i units of asset i in our portfolio
Portfolio value is

Dale Roberts (c) 2011-2014

V (t) =

n
X

x i Pi (t)

i=1

Last Modified: 14/10/2014 08:19

237 / 561

Arbitrage with Static Portfolios

At time t = 0:
V (0) =

n
X

x i Pi (0) = 0

i=1

At time T > 0:

Dale Roberts (c) 2011-2014

P(V (T ) 0) = 1
P(V (T ) > 0) > 0

Last Modified: 14/10/2014 08:19

238 / 561

Principle of No Arbitrage

The Principle of No Arbitrage states simply that such arbitrage


opportunities do not exist
If they did exist, smart investors could make infinite amounts of money

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

239 / 561

Equivalent Forms for PNA

We cannot construct a riskless portfolio which returns more than the


risk-free rate of return
If two portfolios A and B give rise to identical (but possibly random) future
cashflows with certainty, then A and B must have the same value at the
present time (aka. law of one price)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

240 / 561

Example: Parallel Yield Curve Shifts

It is easy to find a model which admits arbitrage


Suppose that

P(0, T ) = exp

f (0, u) du

for some initial forward-rate curve f (0, T )

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

241 / 561

Example: Parallel Yield Curve Shifts

Probability space (, F , P)
Our model is defined so that at time 1 the forward rate curve will be
f (1, u) = f (0, u) + ,

for u > 1

: R is a random variable

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

242 / 561

Dale Roberts (c) 2011-2014

Example: Parallel Yield Curve Shifts

u 7 f (0, u)

Last Modified: 14/10/2014 08:19

243 / 561

Dale Roberts (c) 2011-2014

Example: Parallel Yield Curve Shifts

u 7 f (1, u)
{
u 7 f (0, u)

Last Modified: 14/10/2014 08:19

244 / 561

Dale Roberts (c) 2011-2014

Example: Parallel Yield Curve Shifts

u 7 f (1, u)
u 7 f (0, u)

Last Modified: 14/10/2014 08:19

245 / 561

Dale Roberts (c) 2011-2014

Example: Parallel Yield Curve Shifts

{

Last Modified: 14/10/2014 08:19

246 / 561

Example: Parallel Yield Curve Shifts


Suppose exists: T1 -Bond, T2 -Bond, T3 -Bond with 1 < T1 < T2 < T3
$1

T1

$1

T2

$1

T3

Let x i be the units held at time 0 of the Ti -Bond

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

247 / 561

Example: Parallel Yield Curve Shifts


At time t = 1 we have

Dale Roberts (c) 2011-2014

P(1, T ) = exp

f (1, u) du

( f (0, u) + ) du

= exp

= exp

f (0, u) du e(T 1)

P(0, T ) (T 1)
e
P(0, 1)

Last Modified: 14/10/2014 08:19

248 / 561

Example: Parallel Yield Curve Shifts

Remember this for later:

Dale Roberts (c) 2011-2014

P(1, T ) =

P(0, T ) (T 1)
e
P(0, 1)

Last Modified: 14/10/2014 08:19

(6)

249 / 561

Example: Parallel Yield Curve Shifts


For an arbitrage, we need:

Dale Roberts (c) 2011-2014

3
X

x i P(0, Ti ) = 0

i=1
3
X

x i P(1, Ti ) 0 with prob. 1

i=1
3
X

x i P(1, Ti ) > 0 with positive prob.

i=1

Last Modified: 14/10/2014 08:19

250 / 561

Example: Parallel Yield Curve Shifts


The value of the portfolio at time 1 is

Dale Roberts (c) 2011-2014

V1 () =
=
=

3
X
i=1
3
X

x i P(1, Ti )
xi

P(0, Ti ) (Ti 1)
e
P(0, 1)

by (6)

i=1
(T2 1)

e
g()
P(0, 1)

Last Modified: 14/10/2014 08:19

251 / 561

Example: Parallel Yield Curve Shifts

Where
g() =

3
X

x i P(0, Ti )e(Ti T2 )

i=1

The trick is to choose time T2 , add/substract T2 , to get

Dale Roberts (c) 2011-2014

e(Ti 1) = e(Ti T2 ) e(T2 1)

Last Modified: 14/10/2014 08:19

252 / 561

Example: Parallel Yield Curve Shifts


If value of initial portfolio is zero,
3
X

x i P(0, Ti ) = 0

i=1

then g(0) = 0
Also, at time 1 and for a shift ,

Dale Roberts (c) 2011-2014

V1 () < 0 g() < 0

Last Modified: 14/10/2014 08:19

253 / 561

Example: Parallel Yield Curve Shifts

As
g() =

3
X

x i P(0, Ti )e(Ti T2 )

i=1

g() is continuous and twice differentiable in


We are looking for the situation where V1 () > 0 for all 6= 0
As g(0) = 0, we must also have g 0 (0) = 0

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

254 / 561

Example: Parallel Yield Curve Shifts


The condition g 0 (0) = 0 implies
3
X

x i (T2 Ti )P(0, Ti ) = 0

i=1

Which, in turn, implies


3
X

x i Ti P(0, Ti ) = 0

i=1

since T2

3
X

x i P(0, Ti ) = 0

i=1

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

255 / 561

Example: Parallel Yield Curve Shifts

Finally, it is sufficient that g 00 () > 0 for all to ensure that g() > 0 for all
6= 0
As
00

g () =

3
X

x i (T2 Ti )2 P(0, Ti )e(Ti T2 )

i=1

g () > 0 for all if and only if x 1 and x 3 are both greater than or equal to
0
00

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

256 / 561

Example: Parallel Yield Curve Shifts


Take x 2 = 1
Eq (6) and x 2 = 1 imply that at least one of x 1 and x 3 must be greater
than zero
But

3
X

x i Ti P(0, Ti ) = 0

i=1

implies that x 1 and x 3 must be both positive or both negative


It follows that g 00 () > 0 for all

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

257 / 561

Example: Parallel Yield Curve Shifts

Hence, g() > 0 for all 6= 0


We can conclude that
V1 () =

3
X

x i P(1, Ti ) > 0

i=1

for all 6= 0
Arbitrage!

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

258 / 561

Example 2

Suppose that P(0, t) = e0.08t for all t > 0


And

P(1, t + 1) =

e0.1t
e

0.06t

if = 1
if = 0

for the random variable : {0, 1}


Spot and forward rate curves will randomly shift up or down by 2%

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

259 / 561

Example 2

Suppose we hold x 1 , x 2 , x 3 units of T1 -Bond, T2 -Bond, and T3 -Bond such


that
x 2 P(0, 2) = 1

Dale Roberts (c) 2011-2014

x 1 P(0, 1) + x 2 P(0, 2) + x 3 P(0, 3) = 0


x 1 P(0, 1) + 2x 2 P(0, 2) + 3x 3 P(0, 3) = 0

Last Modified: 14/10/2014 08:19

260 / 561

Example 2
Solving the system of equations:
1
= 1.173511
P(0, 2)
1
x3 =
= 0.635624
2P(0, 3)
1
x1 =
= 0.541644
2P(0, 1)
x2 =

At time 1, the value of this portfolio is 0.00021 if = 1 or 0.00022 if = 0

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

261 / 561

Example 2

The model is not arbitrage-free


It violates the law of one price

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

262 / 561

Conclusion

Parallel shifts of the yield curve cannot occur at any time in the future

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

263 / 561

Dale Roberts (c) 2011-2014

Fundamental Theorem of Asset Pricing

Last Modified: 14/10/2014 08:19

264 / 561

Fundamental Theorem of Asset Pricing


Suppose that the risk-free rate r(t) is stochastic
Randomness is underpinned by the probability space (, F , P)
P is the real-world probability measure
The cash account is given by
B(t) = B(0)ex p

Z


r(s) ds

And satisfies d B(t) = r(t)B(t) d t

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

265 / 561

Fundamental Theorem of Asset Pricing

Theorem (FTAP)
(i) Bond prices evolve in arbitrage-free manner if and only if exists measure Q
)
equivalent to P under which P(t,T
B(t) is a martingale for 0 < t < T
(ii) If (i) holds, then the market is complete if and only if Q is the unique
)
measure under which P(t,T
B(t) are martingales

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

266 / 561

Fundamental Theorem of Asset Pricing

The measure Q is often called the equivalent martingale measure


It is also called risk-neutral measure

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

267 / 561

Fundamental Theorem of Asset Pricing

Corollary
Hence

P(t, T ) = EQ exp



r(s) ds F t

where F t is the -algebra generated by the price histories up to time t , and EQ


is the expectation with respect to Q

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

268 / 561

Fundamental Theorem of Asset Pricing

X is some F T -measurable claim payable at T


V (t) is fair value of X at time t
Then V (t)/B(t) is also a martingale under Q
Hence,

Dale Roberts (c) 2011-2014

V (t) = EQ exp



r(u) du X F t

Last Modified: 14/10/2014 08:19

269 / 561

Example: Forward Pricing

Forward contract: K will be paid at time T in return for repayment of 1 at


time S where T < S
Equivalently, K is paid at T in return for delivery of an S -bond with value
P(T, S)
How much is this contract worth at time t < T ?
X = P(T, S) K at time T

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

270 / 561

Dale Roberts (c) 2011-2014

Example: Forward Pricing

V (t) = EQ exp

= EQ e

RT

r(u) du



r(u) du X F t
h

RS

i i

F T F t
i

r(u) du

EQ e

h RT
t r(u) du
KEQ e
F t
i
i
h RS
h RT


= EQ e t r(u) du F t KEQ e t r(u) du F t
t

= P(t, S) K P(t, T )

Last Modified: 14/10/2014 08:19

271 / 561

Example: Forward Pricing

We now choose K so that V (t) = 0


This gives
K=

P(t, S)
P(t, T )

The contract can be hedged at no cost at time t by buying one unit of


P(t, S) and selling P(t, S)/P(t, T ) units of P(t, T )

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

272 / 561

Dale Roberts (c) 2011-2014

The Interest Rate Setting

Last Modified: 14/10/2014 08:19

273 / 561

The Interest Rate Setting

What is special about the interest rate setting?


There are a few new things to think about...

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

274 / 561

Factors

A one-factor model is one under which there is only one source of


randomness affecting the bond prices
In a one-factor model all price changes are perfectly (but non-linearly)
correlated
If we know one change in quantity (e.g. risk-free rate r(t)) we know the
change in all asset prices

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

275 / 561

Factors

A multifactor model has multiple sources of randomness


Price changes are not perfectly correlated
Example: m bonds, each bond driven by a different Brownian motion

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

276 / 561

A Bond is a Derivative

A bond is a derivative itself


In a one-factor model, the price of any bond is derived from our
knowledge of the short-rate r(t)
The short-rate is the underlying

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

277 / 561

Types Of Model

There are two main types of models: Equilibrium and No-arbitrage


Equilibrium models are built on assumptions about how the economy
works
Take into account the varying risk preferences of different investors
Relationship between the economy and the term-structure of interest
rates

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

278 / 561

Types Of Model

This means constructing a simple stochastic model for the evolution of


the risk-free rate
Invoke the FTAP to derive theoretical bond prices
theoretical prices evolve in a way that is arbitrage-free
It may happen that theoretical prices is different from the observed prices
in the market, giving rise to arbitrage opportunities (cheap/dear analysis)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

279 / 561

Types Of Model

Often short-rate models are seen as equilibrium models


This is not always true
It is difficult to prove that a short-rate model has an equilibrium derivation

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

280 / 561

Types Of Model

No-Arbitrage Models use the observed term structure at the current time
as the starting point
Future prices evolve in a way which is consistent with this initial price
structure and which is arbitrage free
These models are used to price short-term derivatives

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

281 / 561

Types Of Model

Actual vs. theoretical price differences in equilibrium models get


magnified when pricing derivatives
A 1% error could lead to a 10% error in the price of an option
On the other hand, no-arbitrage models have strange dynamics for
quantities like r(t) in the long-term which are hard to justify

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

282 / 561

Dale Roberts (c) 2011-2014

Blacks model

Last Modified: 14/10/2014 08:19

283 / 561

Types Of Model

Black, F. (1976). The pricing of commodity contracts. Journal of Financial


Economics.
Assumed that futures prices have the same lognormal property assumed
for stock prices
Gives us Black-Scholes style formula for valuing futures options

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

284 / 561

Blacks formula is given by


C = er t (F0 (d1 ) K(d2 ))
P = er t (K(d2 ) F0 (d1 ))

where
log(F0 /K) + 2 T /2
d1 =
,
p
T

Dale Roberts (c) 2011-2014

p
log(F0 /K) 2 T /2
d2 =
= d1 T .
p
T

Last Modified: 14/10/2014 08:19

285 / 561

Example
European put future option on crude oil. TTM is 4 months, current futures
price is $20, exercise price is $20, risk-free rate is 5% per annum, and the
volatility of the futures price is 25% per annum.
F0 = 20,

K = 20, r = 0.05, T = 4/12, = 0.25.


p
p
T
T
log(F0 /K) = 0, d1 =
= 0.07216, d2 =
= 0.07216.
2
2
(d1 ) = 0.4712, (d2 ) = 0.5288
V = e0.054/12 (20 0.5288 20 0.4712) = 1.12

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

286 / 561

Blacks model for a Bond option

Traders became comfortable with the lognormal assumption and


volatility to describe uncertainty
This lead to the formula being applied to interest rate derivatives (Bond
options, interest rate caps, and swap options)
During 1980s and 1990s, the volume of trading in interest rate derivatives
in both over-the-counter and exchange-traded markets increased very
quickly

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

287 / 561

Blacks model for a Bond option

Consider a European call option on a variable whose value is V . Define:


T:
F:
F0 :
K:
P(t, T ):
VT :
:

Dale Roberts (c) 2011-2014

Maturity of the option


Forward price of V for a contract with maturity T
Value of F at time zero
Strike price of the option
Price at time t of a zero-coupon bond paying $1 at time T
Value of V at time T
Volatility of F

Last Modified: 14/10/2014 08:19

288 / 561

Blacks model for a Bond option

Blacks model assumes:


1. VT has a lognormal
distribution with standard deviation of log VT
p
equal to T
2. The expected value of VT is F0
Discounts the expected payoff by multiplying by P(0, T )
European call payoff is max(VT K, 0) at time T

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

289 / 561

Blacks model for a Bond option


Lognormal assumption implies that the expected payoff is
E(VT )(d1 ) K(d2 )

where E(VT ) is the expected value of VT and


log(E(VT )/K) + 2 T /2
d1 =
,
p
T

p
d2 = d1 T .

This gives
C = P(0, T ) (F0 (d1 ) K(d2 )) ,

Dale Roberts (c) 2011-2014

P = P(0, T ) (K(d2 ) F0 (d1 ))

Last Modified: 14/10/2014 08:19

290 / 561

Blacks model for a Bond option

Blacks model can be extended to allow for the situation where the payoff
is calculated at time T but the payoff is actually made at some time S > T
This basically changes the discounting:
C = P(0, S) (F0 (d1 ) K(d2 )) ,

Dale Roberts (c) 2011-2014

P = P(0, S) (K(d2 ) F0 (d1 ))

Last Modified: 14/10/2014 08:19

291 / 561

Blacks model for a Bond option

Blacks model does not assume that the evolution of V or F is driven by a


(geometric) Brownian motion
All that is required is that VT is lognormally distributed at time T
is referred to as the volatility of F or the forward volatility of V
doesnt say anything about the standard deviation of log V at times
other than T

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

292 / 561

Validity of Blacks model

Blacks model is appropriate when interest rates are constant or


deterministic
In this situation, the forward price of V equal its futures price

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

293 / 561

Validity of Blacks model

When interest rates are stochastic Blacks model involves two approximations:
E(VT ) = F0 . In a risk-neutral world, E(VT ) is equal to its futures price. The
forward price and futures price are not the same when interest rates are
stochastic
The stochastic behaviour of interest rates is not taken into account for the
discounting

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

294 / 561

Validity of Blacks model

It is inappropriate to price interest rate derivatives by just modelling the


dynamics of underlying security
Consistent pricing must be based on the evolution of the entire term
structure of interest rates
Using Blacks model navely can lead to arbitrage opportunities

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

295 / 561

Week 7

Continuous-Time Interest Rate Models


The Martingale Approach
The PDE Approach
Additional Comments

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

296 / 561

Dale Roberts (c) 2011-2014

Continuous-Time Interest Rate Models

Last Modified: 14/10/2014 08:19

297 / 561

One-Factor Models for the Risk-Free Rate

Continuous-time setting
One-factor models for the term structure of interest rates
Bond pricing given a one-factor diffusion model for the risk-free rate r(t)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

298 / 561

One-Factor Models for the Risk-Free Rate

We assume that r(t) is modelled by


d r(t) = a(t)d t + b(t)dW (t)
(W (t)) t0 is a standard Brownian motion under the real-world probability
measure P
a(t) and b(t) are previsible processes
F t = ({W (s) : s t}) is the -algebra generated by the history of W (s)
up to time t

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

299 / 561

One-Factor Models for the Risk-Free Rate

a(t) = a(r(t)) and b(t) = b(r(t))


Ensures that r(t) is Markov and time homogeneous

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

300 / 561

Dale Roberts (c) 2011-2014

One-Factor Models for the Risk-Free Rate

Model
a(r)
Merton (1973)

Dothan (1978)
r
Vasicek (1977)
( r)
CIR (1985)
( r)
Pearson-Sun (1994)
( r)
Brenann-Schwartz (1979)
( r)
Black-Karasinski (1991)
r r log r

Last Modified: 14/10/2014 08:19

b(r)

p r
r
r
r

301 / 561

Dale Roberts (c) 2011-2014

One-Factor Models for the Risk-Free Rate

Model r(t) 0?
AR?
Simple Formula?
M
N
N
Y
D
Y
N
N
V
N
Y
Y
CIR
Y
Y
Y
PS
Y if > 0
Y
N
BS
Y
Y
N
BK
Y
Y ( > 0)
N

Last Modified: 14/10/2014 08:19

302 / 561

One-Factor Models for the Risk-Free Rate

All models are approximation to reality but some are better than others
One must consider:
Desired characteristics
Ease of implementation
How well does model approximate reality
Fast vs. slow calculation

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

303 / 561

One-Factor Models for the Risk-Free Rate

Sometimes simple formula doesnt mean fast calculation


Ex: analytical formulae exist for CIR however involve using the law of the
non-central chi-squared distribution
In R: dchisq,pchisq,rchisq
Ex: Vasicek and other models only require Normal distribution

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

304 / 561

Other Characteristics

Are bond and derivative prices straightforward and simple to calculate


numerically?
Is the model flexible enough to cope with new and more complex
derivative products?
Does the model produce dynamics which are realistic?

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

305 / 561

Other Characteristics

Does the model fit historical data well? (e.g., mean and variance)
If the model keeps rates positive, does it allow forward rates, spot rates
and par yields to take values close to zero?
Does the model have equilibrium derivation?

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

306 / 561

Other Characteristics

One factor models generally fail many of these criteria


This is due to their dependence on a single-factor (e.g., dynamics of r(t))
This is why we will consider multi-factor models in the second half of
semester

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

307 / 561

Other Characteristics

Over the next couple of weeks we will consider one-factor models


We will consider a number of concepts that will carry over to the
multi-factor setting
We will consider a general approach to pricing whereby we obtain a
coherent set of price dynamics that do not admit arbitrage

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

308 / 561

Dale Roberts (c) 2011-2014

The Martingale Approach

Last Modified: 14/10/2014 08:19

309 / 561

The Martingale Approach

We will now discuss the Martingale approach to pricing

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

310 / 561

The Martingale Approach


We consider an abstract model with only one source of randomness
Risk-free rate r(t)
d r(t) = a(t)d t + b(t)dW (t)
P(t, T ) is price of T -bond at time t
d P(t, T ) = P(t, T )[m(t, T )d t + S(t, T )dW (t)]
a(t), b(t), m(t, T ) and S(t, T ) are previsible functions (possibly
stochastic)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

311 / 561

The Martingale Approach

We also have the risk-free cash account B(t)


B(t) satisfies the SDE

dB(t) = r(t)B(t)d t

The solution is

Dale Roberts (c) 2011-2014

B(t) = B(0) exp

Z


r(u) du

Last Modified: 14/10/2014 08:19

312 / 561

The Martingale Approach

The risk premium at time t on the risky bond P(t, T ) is defined as the
excess expected rate of return on the bond m(t, T ) over the risk-free rate
of interest r(t)
The risk premium represents the extra reward we get for investing in the
risky asset rather than the risk-free cash account

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

313 / 561

The Martingale Approach

The market price of risk is the previsible process


(t) =

m(t, T ) r(t)
S(t, T )

The market price of risk represents the excess expected return per unit of
volatility

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

314 / 561

The Martingale Approach

Consider a claim X S at some time S < T


What is the no-arbitrage price, V (t), at time t < S for this claim?

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

315 / 561

Fundamental Theorem of Asset Pricing

Theorem
There exists a measure Q P with

V (t) = EQ exp



r(u) du X F t

f(t) and W
f(t) is a BM under Q
where d r(t) = (a(t) (t)b(t))d t + b(t)d W

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

316 / 561

Fundamental Theorem of Asset Pricing

In Baxter and Rennie, we had a 3 step approach


Cairns expands this to 5 steps

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

317 / 561

Fundamental Theorem of Asset Pricing


Proof

Define the discounted price process



 Z t
P(t, T )
Z(t, T ) =
r(u) du
= P(t, T ) exp
B(t)
0

where B(t) is the risk-free cash account B(0) = 1

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

318 / 561

Fundamental Theorem of Asset Pricing


Proof

1.
2.
3.
4.
5.

Establish the probability measure Q equiv. to P s.t. Z(t, T ) is a martingale


For t < S < T define D(t) = EQ [B(S)1 X |F t ]
Apply MRT to get holdings (t)
Set (t) = D(t) (t)Z(t, T ), show that portfolio is self-financing
Show that portfolio replicates claim X

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

319 / 561

Fundamental Theorem of Asset Pricing


Proof: Step 1

Apply product rule to


Z(t, T ) =

P(t, T )
= B(t)1 P(t, T )
B(t)

to get

Dale Roberts (c) 2011-2014

d Z(t, T ) = B(t)1 d P(t, T ) + P(t, T )d(B(t)1 ) + dB 1 , P(t)

Last Modified: 14/10/2014 08:19

320 / 561

Fundamental Theorem of Asset Pricing


Proof: Step 1

Apply Its formula:


1
1 2
dB(t)
+
dB(t)
B(t)2
2 B(t)3
r(t)
=
dt
B(t)

d(B(t)1 ) =

as d B(t) = r(t)B(t)d t and dB(t) = 0d t

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

321 / 561

Fundamental Theorem of Asset Pricing


Proof: Step 1

Substituting d P(t, T ) = P(t, T )[m(t, T )d t + S(t, T )dW (t)] and d(B(t)1 )


into product formula gives
r(t)P(t, T )d t
P(t, T )
(m(t, T )d t + S(t, T )dW (t))
+ 0d t
B(t)
B(t)
= Z(t, T )[(m(t, T ) r(t))d t + S(t, T )dW (t)]

d Z(t, T ) =

as

Dale Roberts (c) 2011-2014

dB 1 , P(t) = 0 d t

Last Modified: 14/10/2014 08:19

322 / 561

Fundamental Theorem of Asset Pricing


Proof: Step 1
Recall that the market price of risk is
(t) =

m(t, T ) r(t)
S(t, T )

so defining
f(t) = W (t) +
W

(u) du

f(t) = dW (t) + (t) d t


dW

d Z(t, T ) = Z(t, T )[(m(t, T ) r(t))d t + S(t, T )dW (t)]

Dale Roberts (c) 2011-2014

= Z(t, T )[(m(t, T )r(t)(t)S(t, T ))d t +S(t, T )(dW (t)+(t)d t)]


f(t)]
= Z(t, T )[(m(t, T ) r(t) (t)S(t, T ))d t + S(t, T )d W
f(t)
= Z(t, T )S(t, T )d W
Last Modified: 14/10/2014 08:19

323 / 561

Fundamental Theorem of Asset Pricing


Proof: Step 1

Notice that Z(t, T ) is an exponential process


We need the Novikov condition to hold, i.e.,

Z T

1
EP exp
(u)2 du <
2 0

so we can apply Girsanovs theorem

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

324 / 561

Fundamental Theorem of Asset Pricing


Proof: Step 1
Girsanovs theorem allows us to state that there exists a measure Q
equivalent to P
With Radon-Nikodm derivative given by

Z T
Z T
1
dQ
(u)dW (u)
(u)2 du
= exp
dP
2 0
0
Under which
f(t) = W (t) +
W

(u) du

is a standard Brownian motion

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

325 / 561

Fundamental Theorem of Asset Pricing


Proof: Step 1

Note that under the same change of measure from P to Q we have that
d P(t, T ) = P(t, T )[m(t, T ) d t + S(t, T ) dW (t)],

under P

becomes
f(t)],
d P(t, T ) = P(t, T )[r(t) d t + S(t, T ) d W

under Q

Under Q, all T -bond prices have drift equal to the risk-free rate r(t)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

326 / 561

Fundamental Theorem of Asset Pricing


Proof: Step 1

The process Z(t, T ) under Q satisfies


f(t)
d Z(t, T ) = Z(t, T )S(t, T ) d W
The SDE has no drift
The SDE is a martingale under Q if

Z T

1
2
EQ exp
S(t, T ) d t
<
2 0

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

327 / 561

Fundamental Theorem of Asset Pricing


Proof: Step 2

We have the claim X S = X that is a random variable


We need to make it into a process
For t < S < T define
D(t) = EQ [B(S)1 X S |F t ]
This is a martingale under Q by the tower property for conditional
expectations

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

328 / 561

Fundamental Theorem of Asset Pricing


Proof: Step 3

Z(t, T ) and D(t) are both Q-Martingales


We can apply the Martingale representation theorem (MRT)
By MRT, there exists a previsible process (t) such that
Z t
D(t) = D(0) +

(u)d Z(u, T )

This requires S(t, T ) to be non-zero for all t < S almost surely

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

329 / 561

Fundamental Theorem of Asset Pricing


Proof: Step 4

We setup a (P, B) portfolio with value


V (t) = (t)P(t, T ) + (t)B(t)
We take (t) = D(t) (t)Z(t, T )
We need to show that the portfolio is self-financing

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

330 / 561

Fundamental Theorem of Asset Pricing


Proof: Step 4
Consider a short time interval [t, t + d t)
Over the period, hold (t) of P(s, t) and (t) of B(s)
The change of value over [t, t + d t) is
(t)d P(t, T ) + (t)d B(t)
We will now compute the change of value of V (t) but
V (t) = (t)P(t, T ) + (t)B(t) = B(t)[(t)Z(t, T ) + (t)] = B(t)D(t)

as (t) = D(t) (t)Z(t, T )

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

331 / 561

Fundamental Theorem of Asset Pricing


Proof: Step 4

d V (t) = d[B(t)D(t)]
= B(t)d D(t) + D(t)dB(t) + dB(t)d D(t)
= B(t)(t)d Z(t, T ) + D(t)r(t)B(t)d t + 0d t
f(t) + ((t)Z(t, T ) + (t))r(t)B(t)d t
= (t)B(t)S(t, T )Z(t, T )d W
f(t)) + (t)r(t)B(t) d t
= (t)P(t, T )(r(t)d t + S(t, T )d W
= (t)d P(t, T ) + (t)dB(t)
The portfolio process is self-financing

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

332 / 561

Fundamental Theorem of Asset Pricing


Proof: Step 5

We also have

V (S) = B(S)EQ [B(S)1 X S |FS ] = X S

So the portfolio replicates the claim X S


It follows that for t < S , V (t) is the unique no-arbitrage price at time t for
X S payable at S

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

333 / 561

Dale Roberts (c) 2011-2014

Fundamental Theorem of Asset Pricing


Proof: Step 5

V (t) = B(t)D(t)


B(t)
= EQ
X S F t
B(S)

ZS
= EQ exp



r(u) du X S F t

Last Modified: 14/10/2014 08:19

334 / 561

Fundamental Theorem of Asset Pricing

Uniqueness of the price V (t) is due to:


We are considering a one-factor model
We specified the dynamics under P for P(t, T ) which allows us to establish
a market price of risk
Randomness is driven by a Brownian motion W (t). Other processes (e.g.,
jumps) may mean that V (t) is not uniquely determined

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

335 / 561

Fundamental Theorem of Asset Pricing


Corollary
For all S such that 0 < S < T ,

P(t, S) = EQ exp



r(u) du F t

Proof.
Take X S = 1 in previous theorem.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

336 / 561

Fundamental Theorem of Asset Pricing

Under Q the prices of all tradable assets have the risk-free rate of interest
as the expected growth rate
Under P,
d V (t) = V (t)[(r(t) + (t)V (t)) d t + V (t)dW (t)]
(t)V (t) is called the market risk premium or risk premium

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

337 / 561

Fundamental Theorem of Asset Pricing

Notice that (t)V (t) depends on the market price of risk (t)
This implies that the risk premiums on different assets are closely linked
They can differ only through the volatility in the tradable asset (V (t) or
S(t, T ) for a ZCB)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

338 / 561

Fundamental Theorem of Asset Pricing

We anticipate that ZCB will have a positive risk premium ((t)S(t, T ) > 0
for all T > t ) to reward for the extra risk
Derivatives V (t) for which V (t) has the same sign as S(t, T ) also have a
positive risk premium (e.g., call option on P(t, T ))
Opposite sign for V (t) leads to a negative risk premium (e.g., Put option
on P(t, T ))

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

339 / 561

Dale Roberts (c) 2011-2014

The PDE Approach

Last Modified: 14/10/2014 08:19

340 / 561

The PDE Approach

We shall now discuss the Partial Differential Equation (PDE) approach


given in Vasicek (1977)
Previously, we have used the martingale approach to pricing
The PDE approach is also useful especially for numerical methods
Black and Scholes used a PDE approach in their classic paper

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

341 / 561

The PDE Approach

r(t) is Markov, i.e., the future dynamics of r(t) given its current value is
independent of past behaviour
The market is efficient: no transactions costs and investors are rational

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

342 / 561

The PDE Approach

Our abstract model for r(t) is of the form


d r(t) = a(t, r(t))d t + b(t, r(t))dW (t)
Prices changes for all bonds with different maturities are perfectly (but
nonlinearly) correlated

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

343 / 561

The PDE Approach


By assumption P(t, T ) is only a function of the current time t and r(t), i.e.,
P(t, T ) = f (t, r(t))

for some unknown f (t, x)


By Its formula,

Dale Roberts (c) 2011-2014

f
f
1 2f
dt +
d r(t) + b2 2 d t
t
x
2 x
f
f
1 2f
=
dt +
(ad t + bdW (t)) + b2 2 d t
x
2 x
 t

2
f
f
f
1 f
=
+a
+ b2 2 d t + b
dW (t)
t
x 2 x
x

d f (t, r(t)) =

Last Modified: 14/10/2014 08:19

344 / 561

The PDE Approach


As f (t, r(t)) = P(t, T ), this implies that the price dynamics are given by
an exponential SDE
d P(t, T ) = P(t, T )[m(t, T )d t + S(t, T )dW (t)]
With average price returns


1 P
P 1 2 2P
m(t, T ) =
+a
+ b
P t
r 2 r2
And price returns volatility

Dale Roberts (c) 2011-2014

S(t, T ) =

1 P
b
P r

Last Modified: 14/10/2014 08:19

345 / 561

The PDE Approach

We can now consider the market price of risk and use it to derive an
equation for price of a T -bond
Construct a portfolio of:
V1 (t) of a T1 -bond (i.e., short)
V2 (t) of a T2 -bond (i.e., long) with T2 > T1
Total portfolio worth is V (t) = V2 (t) V1 (t)
We now vary V1 (t) and V2 (t) so that V (t) remains risk-free

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

346 / 561

The PDE Approach


The instantaneous investment gain from t to t + d t is
d P(t, T1 )
d P(t, T2 )
+ V2 (t)
P(t, T1 )
P(t, T2 )
= V1 (t)(m1 d t + S1 dW ) + V2 (t)(m2 d t + S2 dW )

V1 (t)

= (V2 m2 V1 m1 )d t + (V2 S2 V1 S1 )dW


Where for i = 1, 2:
mi = m(t, Ti ),

Si = S(t, Ti )

We now simplify that expression under some assumptions

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

347 / 561

The PDE Approach


If we assume that

Then we get

V1 (t) S(t, T2 ) S2
=
=
V2 (t) S(t, T1 ) S1
V1 (t)S1 V2 (t)S2 = 0

and
V2 m2 V1 m1 =

S1 V
S V
m2 2
m1
S1 S2
S1 S2

So the investment gain over [t, t + d t] is

m2 S1 m1 S2
V (t)
dt
S1 S2

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

348 / 561

The PDE Approach

Therefore, through our choice of portfolio strategy we have obtained a


risk-free investment strategy
Further, by varying V1 (t) appropriately we can ensure that this portfolio is
also self-financing:

m2 S1 m1 S2
dV (t) = V (t)
dt
S1 S2

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

349 / 561

The PDE Approach


Since the portfolio is risk-free, the principle of no-arbitrage states that the
growth rate must equal r(t):

m2 S1 m1 S2
m1 r
m r
= r(t) or
= 2
S1 S2
S1
S2
This must be true for all maturities. Thus for all T > t
m(t, T ) r(t)
= (t)
S(t, T )
(t) is the market price of risk and cannot depend on the maturity date T

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

350 / 561

The PDE Approach


On one hand, we have
m(t, T ) = r(t) + (t)S(t, T )
On the other hand,

Dale Roberts (c) 2011-2014



P 1 2 2P
1 P
+a
+ b
m(t, T ) =
P t
r 2 r2
1 P
S(t, T ) = b
P r

Last Modified: 14/10/2014 08:19

351 / 561

The PDE Approach

We equate the two expressions for m(t, T ) to find

Dale Roberts (c) 2011-2014

P
P 1 2 2P
rP = 0
+ (a b)
+ b
t
r 2 r2

Last Modified: 14/10/2014 08:19

352 / 561

The PDE Approach


Theorem (Feynman-Kac Formula)
Let T > 0 and be a continuous function. Suppose that a function F (t, r)
satisfies

1 2

2
t F (t, r) + f (t, r) r F (t, r) + (t, r) r F (t, r) r F (t, r) = 0
2
F (T, r) = (r)
Then1 M (t) = exp(

Rt
0

r(u) du)F (t, r(t)) is a martingale and

F (t, r(t)) = EQ exp



r (u) du (r(T )) F t

f
where r (t) = r(t) and under Q satisfies dr (t) = f (t, r )d t + (t, r )d W
1

under some technical conditions, see appendix

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

353 / 561

The PDE Approach

We now apply the Feynmann-Kac formula with F (t, r) = P(t, T )


f (t, r) = a(t, r) (t)b(t, r),

(t, r) = b(t, r)

And terminal condition P(T, T ) = (r(t)) = 1 as we are pricing a T -bond


Hence,

Dale Roberts (c) 2011-2014

P(t, T ) = EQ exp

Last Modified: 14/10/2014 08:19



r (u) du F t

354 / 561

The PDE Approach

Under P we have

d r(t) = a(t)d t + b(t)dW (t)

Under Q we have
f(t)
dr (t) = f (t, r (t))d t + (t, r (t))d W
We need to satisfy ourselves that these measures P and Q are equivalent

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

355 / 561

The PDE Approach

Assume that (t) satisfies Novikov condition


R
f(t) = W (t) + t (s) ds
Define W
0

f is a
Then by Girsanovs theorem, there exists Q equiv. to P under which W
BM with Radon-Nikodym derivative
Z T

Z T
dQ
1
= exp
(t)dW (t)
(t)2 d t
dP
2 0
0

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

356 / 561

The PDE Approach


We then have
d P(t, T ) = P(t, T )(m(t, T )d t + S(t, T )dW )
f (t)d t)]
= P(t, T )[m(t, T )d t + S(t, T )(d W
f]
= P(t, T )[(m(t, T ) (t)S(t, T ))d t + S(t, T )d W
f]
= P(t, T )[r(t)d t + S(t, T )d W
Thus under Q, the expected return on any bond is the risk-free rate
This implies that Q is the EMM

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

357 / 561

The PDE Approach

This Feymann-Kac approach can be applied to other derivative payoffs


Simply choose the function appropriately
This is the basis of finite difference methods for pricing options as we
determine P(t, T ) by solving a PDE with a specific terminal condition

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

358 / 561

Dale Roberts (c) 2011-2014

Additional Comments

Last Modified: 14/10/2014 08:19

359 / 561

Market Price of Risk

Note that we started by specifying P dynamics and then moved to Q


dynamics using the market price of risk (t)
The market price of risk links the dynamics of the cash bond B(t) and the
T -bond prices
We cant replicate a payoff X only using the cash bond B(t) therefore one
needs to specify
A short-rate model is not fully determined under P unless one specifies the
market price of risk

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

360 / 561

Start under Risk-Neutral Measure


It is customary from a modelling point-of-view to simply specify the Q
dynamics of r directly
This implies the Q dynamics of all bond prices by the formula

Z T
P(t, T ) = EQ exp

r(u) du
t

Some people back out the market price of risk by statistical methods
from historical observations of price movements
This is actually wrong... statistical observations are under P

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

361 / 561

Week 8

Particular Models
Affine Short-Rate Models

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

362 / 561

Dale Roberts (c) 2011-2014

Particular Models

Last Modified: 14/10/2014 08:19

363 / 561

The Vasicek Model

The Vasicek model is specified under Q by


f(t)
d r(t) = ( r(t)) d t + d W
f(t) is BM under EMM Q and , , are all positive constants
where W
is the risk-neutral long-term mean risk-free rate
is the rate at which r(t) reverts back to
is the local volatility of the short-term interest rates

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

364 / 561

The Vasicek Model


For s > 0, r(t + s) given r(t) is Normally distributed under Q with mean
+ (r(t) )es

and variance

1 e2s
2
This implies that the long-term standard deviation of r(t) is

Dale Roberts (c) 2011-2014

p
2

Last Modified: 14/10/2014 08:19

365 / 561

Bond Prices in Vasicek Model


Theorem
Prices for T -bonds in the Vasicek model are given by
P(t, T ) = exp[A(t, T ) B(t, T )r(t)]

where
B(t, T ) =

1 e(T t)


2
2
A(t, T ) = (B(t, T ) (T t)) 2
B(t, T )2
2
4

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

366 / 561

European Options in Vasicek Model


Theorem
The price of a European call option on a S -bond with strike price K and exercise
date T ( T < S ) is
V (t) = P(t, S)(d1 ) K P(t, T )(d2 )

where

p
1
P(t, S)
+
, d2 = d1 p
log
p
K P(t, T )
2
v
t 1 e2(T t)

p = (1 e(ST ) )

d1 =

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

367 / 561

The Cox-Ingersoll-Ross Model

The Vasicek model has a major problem: the risk-free rate can become
negative
Further, empirical evidence shows that the volatility of r(t) is not constant
but is an increasing function of r(t)
The first tractable model that fixed these problems was the CIR model:

f(t)
d r(t) = ( r(t)) d t + r(t)d W
f is BM under EMM Q
where , , > 0 and W

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

368 / 561

Bond Prices in CIR Model


Theorem
Prices for T -bonds in the CIR model are given by
P(t, T ) = exp[A(t, T ) B(t, T )r(t)]

where
2(et 1)
B(t, T ) =
( + )(et 1) + 2


2
2e(+)t/2
A(t, T ) = 2 log

( + )(et 1) + 2
p
= 2 + 22

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

369 / 561

European Option Prices in CIR Model

Prices of European options are given in terms of the non-central


chi-squared distribution under Q
They involve a pretty nasty formula
Exercise: Implement the pricing formula in R

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

370 / 561

Calibration of Vasicek and CIR Model

In Vasicek and CIR, T -bond prices are given by


P(t, T ) = exp(A(t, T ) B(t, T )r(t))

where A and B depend on the model parameter vector


For Vasicek model: = (, , )
For CIR model: = (, , )

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

371 / 561

Calibration of Vasicek and CIR Model

(0, Ti ) for the Ti -bond with


We have the current market YTMs denoted R
i = 1, . . . , n
We can plot the yield curve (i.e., the term structure)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

372 / 561

2.8%
2.4%
2.0%

Yield

3.2%

Calibration of Vasicek and CIR Model

Dale Roberts (c) 2011-2014

3M

6M

1Y

2Y

5Y

10Y

20Y

Time To Maturity (Log Scale)

Last Modified: 14/10/2014 08:19

373 / 561

Calibration of Vasicek and CIR Model


Recall that the yield to maturity R(t, T ) satisfies
P(t, T ) = exp((T t)R(t, T ))
We also have

P(t, T ) = exp(A(t, T ) B(t, T )r(t))

Equating the two terms and setting t = 0 we get

Dale Roberts (c) 2011-2014

1
R(0, T ) = (A(0, T ) B(0, T )r(0))
T

Last Modified: 14/10/2014 08:19

374 / 561

Calibration of Vasicek and CIR Model

We now choose the parameter vector to minimize the squared distance


(0, Ti ) and the theoretical yields R(0, Ti ):
between the market yields R
n
X


R
(0, Ti ) R(0, Ti ) 2
min

i=1

We are fitting the model under Q


We do not fit to historical data! Historical data gives the past dynamics
under P

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

375 / 561

2.8%
2.4%
2.0%

Yield

3.2%

Calibration of Vasicek and CIR Model

Dale Roberts (c) 2011-2014

3M

6M

1Y

2Y

5Y

10Y

20Y

Time To Maturity (Log Scale)

Last Modified: 14/10/2014 08:19

376 / 561

Calibration of Vasicek and CIR Model

The Vasicek and CIR models are equilibrium models


They are nice and simple: we can obtain formulas for prices
The disadvantage of equilibrium models is that they do no automatically
fit todays term structure
Can you trust the price of a bond option when the model does not price
the underlying bond correctly?
A 1% error in the price of the underlying bond could lead to a 25% error in
an option price

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

377 / 561

Calibration of Vasicek and CIR Model

A hedge fund might view the difference between the model prices and the
market prices as the market is mispricing certain assets and build a
trading strategy to capture the mispricings

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

378 / 561

Calibration of Vasicek and CIR Model

There exist models that fit the current market perfectly: they are called
no-arbitrage models (e.g., Ho & Lee, Hull & White)
In an equilibrium model, todays term structure of interest rates is an
output
In a no-arbitrage model, todays term structure of interest is an input

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

379 / 561

Dale Roberts (c) 2011-2014

Affine Short-Rate Models

Last Modified: 14/10/2014 08:19

380 / 561

Affine Models

We have seen that both the Vasicek and CIR models have T -bond prices
that have the affine form
P(t, T ) = exp(A(t, T ) B(t, T )r(t))
An affine transformation in mathematics is a linear transformation
followed by a translation (i.e., x 7 B x + a).
The form of A and B depend on the model
Q: Are there any other models that have an affine form for P(t, T )?

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

381 / 561

Affine Models
Consider the abstract SDE for r(t)
f(t)
d r(t) = m(t, r(t))d t + s(t, r(t))d W
f is a BM under EMM Q
where W
Suppose that
P(t, T ) = exp(A(t, T ) B(t, T )r(t))
Apply Its formula with f (t, x) = exp(A(t, T ) B(t, T )x) to get

A B
1 2
f(t)
d P(t, T ) = P(t, T )

r(t) Bm + Bs d t Bsd W
t
t
2

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

382 / 561

Affine Models
But under Q, we know that
f(t)]
d P(t, T ) = P(t, T )[r(t)d t + S(t, T )d W

where S(t, T ) is the volatility of P(t, T )


This follows as all tradable assets must have expected returns given by the
risk-free rate r(t) under Q
Set
1
A B
G(t, r) :=

r(t) Bm(t, r) + Bs(t, r)2


t
t
2
We need G(t, r) = r(t) for all t and r . Equivalently,

Dale Roberts (c) 2011-2014

g(t, r) := G(t, r) r(t) = 0,

Last Modified: 14/10/2014 08:19

t, r

383 / 561

Affine Models
Now, to ensure this holds we need
2
2
2g
2 m(t, r) 1
2 (s(t, r) )
= B(t, T )
+ B(t, T )
=0
r2
r2
2
r2

Clearly we need B(t, T ) 6= 0 for all t, T which implies


2 m(t, r) 1
2 (s(t, r)2 )

+ B(t, T )
=0
r2
2
r2
Hence we conclude
2 (s(t, r)2 )
=0
r2

and

2 m(t, r)
=0
r2

Linear functions f (x) = a + bx satisfy f 00 (x) = 0 for all x

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

384 / 561

Affine Models
Theorem
If the drift and volatility of r(t) take the form
m(t, r(t)) = a(t) + b(t)r(t)

and

s(t, r(t)) =

(t)r(t) + (r)

this is sufficient for P(t, T ) to have the affine form


P(t, T ) = exp(A(t, T ) B(t, T )r(t))

where a(t), b(t), (t) and (t) are deterministic functions

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

385 / 561

Affine Models

Example (Vasicek Model)


= 0, = 2 , b = and a = which gives
f(t).
d r(t) = ( r(t))d t + d W

We already know it has the affine form.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

386 / 561

Affine Models

Example (CIR Model)


= 0, = 2 , b = , and a = which gives

f(t).
d r(t) = ( r(t))d t + r(t)d W

Again, we already know it has the affine form.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

387 / 561

Affine Models
Example (Merton Model)
= 0, = 2 , b = 0 which gives
f(t)
d r(t) = ad t + d W

and we can determine


B(t, T ) = T t
1
1
A(t, T ) = 2 (T t)3 a(T t)2
6
2

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

388 / 561

Affine Models
Example (Pearson and Sun Model)
= 2 , = /2 , b = and a = ( + ) which gives

f(t).
d r(t) = ( r(t))d t + r(t) d W

Note:
This model is similar to CIR
Now, the minimum value for r(t) is
One can derive A(t, T ) and B(t, T ) in terms of the A and B from the CIR

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

389 / 561

Week 9

No-Arbitrage Models
The Heath-Jarrow-Morton (HJM) Framework

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

390 / 561

Dale Roberts (c) 2011-2014

No-Arbitrage Models

Last Modified: 14/10/2014 08:19

391 / 561

Introduction

We shall now consider the class of no-arbitrage models


We saw that with the Vasicek and CIR models, our theoretical prices do not
match all the bond prices we observe in the market
By design, no-arbitrage models take the observed prices in the market as
an input into the model
This means that the theoretical prices P(t, T ) match the observed prices
Pobs (t, T ) at the time of calibration t

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

392 / 561

Introduction
Market practitioners prefer no-arbitrage models over equilibrium models
Market makers need to both buy and sell at quoted prices. If their pricing
model gives different prices than the rest of the market then this will give
rise to arbitrage opportunities against the market maker
No-arbitrage models allow market makers to check market scenarios (e.g.,
what happens to other prices if the short-term rate increases?)
Market makers are typically considered with the pricing and hedging of
short-term claims. No-arbitrage models work well in this situation
Banks typically want to price OTC derivatives in a way which is consistent
with the closest-matching traded derivatives

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

393 / 561

Markov Models

We shall consider models for which:


the current bond prices in the market are an input
the probability law of P(s, T ) given F t where t s < T is the same as
P(s, T ) given the current value of the short-rate r(t)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

394 / 561

Affine Term Structures

Recall that a model is called affine if the T -bond prices are given by
P(t, T ) = exp (A(t, T ) B(t, T )r(t))

for functions A and B


Some no-arbitrage models are also affine

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

395 / 561

Affine Term Structures


Theorem
Given an affine short-rate model
d r(t) = (b(t) + (t)r(t))d t +

f(t)
a(t) + (t)r(t)d W

then the functions A and B satisfy the system of ODE:


1
t A(t, T ) = a(t)B 2 (t, T ) + b(t)B(t, T ),
A(T, T ) = 0,
2
1
t B(t, T ) = (t)B 2 (t, T ) (t)B(t, T ) 1, B(T, T ) = 0.
2

and
P(t, T ) = exp (A(t, T ) B(t, T )r(t))

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

396 / 561

Ho and Lee Model


In 1986, Ho and Lee considered the model
f(t)
d r(t) = (t)d t + d W
f(t) is a BM under the EMM Q
where W
This is a more general version of the Merton model where the drift (t) is
now a function of time
We also have the current term structure as an input:

Dale Roberts (c) 2011-2014

P(0, T ) for all T > 0

Last Modified: 14/10/2014 08:19

397 / 561

Ho and Lee Model


Ho and Lee model is
f(t)
d r(t) = (t)d t + d W
So in the affine term structure theorem:
b(t) = (t),

(t) = 0,

a(t) = 2 ,

(t) = 0

To find A and B , we need to solve the ODE:

Dale Roberts (c) 2011-2014

1
t A(t, T ) = 2 B 2 (t, T ) + (t)B(t, T ), A(T, T ) = 0,
2
t B(t, T ) = 1,
B(T, T ) = 0.

Last Modified: 14/10/2014 08:19

398 / 561

Ho and Lee Model


Set g(s) := B(s, T ) and integrate g 0 (s) = s B(s, T ) from t to T :
Z T
Z T
g 0 (s) ds =

1 ds
t

Use the boundary condition B(T, T ) = 0 to get

Dale Roberts (c) 2011-2014

g(T ) g(t) = (T t)

B(t, T ) = T t

=0

Last Modified: 14/10/2014 08:19

399 / 561

Ho and Lee Model


By integrating s A(s, T ) from t to T :
Z T
Z T
Z T
1 2 2
s A(s, T ) ds =
B (s, T ) ds +
(s)B(s, T ) ds
2
t
t
t
Substituting B(t, T ) = T t and using the condition B(T, T ) = 0 we get
Z T
Z T
1 2
A(t, T ) =
(T s)2 ds
(s)(T s) ds
2
t
t
Z T
2
=
(T t)3 +
(s)(T s) ds
6
t

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

400 / 561

Ho and Lee Model

Hence, T -bond prices in the Ho and Lee model are given by


P(t, T ) = exp (A(t, T ) B(t, T )r(t))

with

Dale Roberts (c) 2011-2014

B(t, T ) = T t
2
A(t, T ) =
(T t)3
6

Last Modified: 14/10/2014 08:19

(s)(T s) ds
t

401 / 561

Ho and Lee Model


In an affine model, T -bond prices are given by
P(t, T ) = exp (A(t, T ) B(t, T )r(t))
The forward curve is given by

Dale Roberts (c) 2011-2014

log P(t, T )
T

=
[A(t, T ) B(t, T )r(t)]
T

f (t, T ) =

Last Modified: 14/10/2014 08:19

402 / 561

Ho and Lee Model


For the Ho and Lee model, we have
T B(t, T )r(t) = r(t)
2
T A(t, T ) =
(T t)2 +
2
Hence,

Dale Roberts (c) 2011-2014

2
f (t, T ) = (T t)2 +
2

(s) ds
t

Last Modified: 14/10/2014 08:19

(s) ds + r(t)
t

403 / 561

Ho and Lee Model


Suppose we have all the market prices Pobs (0, T ) for T > 0
Then we have the market forward curve
fobs (0, T ) =

log Pobs (0, T )


T

And our theoretical forward curve


2
f (0, T ) = T 2 +
2

(s) ds + r(0)

Q: How do we calibrate our model so that f (0, T ) = fobs (0, T )?

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

404 / 561

Ho and Lee Model


We now choose
(s) :=
Then

(s) ds =

Z
0

fobs (0, s) + 2 s
s

fobs (0, s) ds +
s

2 s ds

= fobs (0, T ) fobs (0, 0) + 2 T 2


= fobs (0, T ) r(0) + 2 T 2
We now have that

f (0, T ) = fobs (0, T )

Our theoretical forward curve matches the market forward curve exactly!

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

405 / 561

Ho and Lee Model


Given a forward curve f (t, T ) at time t , our T -bond prices are given by
Z T

P(t, T ) = exp

f (t, u) du

In the Ho and Lee model, by integrating we get


T

P(t, T ) = exp

Dale Roberts (c) 2011-2014

fobs (0, u) du + fobs (0, u)(T t)


t


2
2
t(T t) (T t)r(t)
2

Last Modified: 14/10/2014 08:19

406 / 561

Ho and Lee Model


The solution of the short-rate SDE
f(t)
d r(t) = (t)d t + d W

is easily found to be
r(t) = r(0) +

f(t)
(s) ds + W

With our choice of (s) :=

Dale Roberts (c) 2011-2014

2
s f obs (0, s) + s

we get

2 t 2
f(t)
r(t) = fobs (0, t) +
+ W
2

Last Modified: 14/10/2014 08:19

407 / 561

Ho and Lee Model

This means that r(t) fluctuates along a modified initial forward curve

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

408 / 561

Hull and White Model

In 1990, Hull and White proposed a generalisation of the Vasicek model


given by
f(t),

d r(t) = ((t)
r(t)) d t + d W
f(t) is a BM under Q and (t) is a deterministic function of time
where W
In the Vasicek model, (t) = is constant

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

409 / 561

Hull and White Model


This is an affine model with
b(t) = (t),

(t) = ,

a(t) = 2 ,

(t) = 0

We get the same B(t, T ) as the Vasicek model:


B(t, T ) =


1
1 e(T t)

Similar to the Ho and Lee model, by integrating:


Z T
Z T
2
2
A(t, T ) =
B (s, T ) ds +
(s)B(s, T ) ds
2 t
t

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

410 / 561

Hull and White Model


The theoretical initial forward curve is
f (0, T ) = T A(0, T ) + T B(0, T )r(0)
Notice that

1
1 e(T t) = e(T t)

t B(t, T ) = e(T t)

T B(t, T ) = T

Hence,

Dale Roberts (c) 2011-2014

T B(s, T ) = s B(s, T )

Last Modified: 14/10/2014 08:19

411 / 561

Hull and White Model


We now get
f (0, T ) = T A(0, T ) + T B(0, T )r(0)
Z T
Z T
2
2
=
s B (s, T )ds + (s) T B(s, T )ds + T B(0, T )r(0)
2 0
0
Z T
2

e(T s) (s)ds + eT r(0)


= B 2 (0, T ) +
2
0
Z T
2
= 2 (1 eT )2 +
e(T s) (s)ds + eT r(0)
2
0

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

412 / 561

Hull and White Model


We now want to fit this model to the forward curve observed in the
market: fobs (0, T )
If we choose
(s) =

1
2
s fobs (0, s) + fobs (0, t) + 2 (1 et )2

then we get

Dale Roberts (c) 2011-2014

f (0, T ) = fobs (0, T )

Last Modified: 14/10/2014 08:19

413 / 561

Hull and White Model


From the properties of the Ornstein-Uhlenbeck process, the short-rate r(t)
evolves like
Z t
Z t
t
(ts)
f(s)
r(t) = e r(0) +
e
(s) ds +
e(ts) d W
0

As

e(ts) (s) ds = fobs (0, t) et r(0) +

2
(1 et )2
22

we get

Dale Roberts (c) 2011-2014

2
r(t) = fobs (0, t) + 2 (1 et )2 +
2
Last Modified: 14/10/2014 08:19

f(s)
e(ts) d W

414 / 561

Hull and White Model


As the distributional properties of the Hull and White model are similar to
the Vasicek model, pricing European options is straight-forward
The price at time t of a European Call option on an S -bond with maturity
date T and strike price K is
C(t) = P(t, S)(d1 ) K P(t, T )(d2 )

where is the CDF of a standard Normal and

Dale Roberts (c) 2011-2014

d1 =

p
1
P(t, S)
+
,
log
p
K P(t, T )
2

d2 = d1 p

v
t 1 e2(T t)

p := (1 e(ST ) )

2
Last Modified: 14/10/2014 08:19

415 / 561

Dothan Model
In 1978, Dothan proposed the model
d r(t) = r(t) dW (t)

where W (t) is a P-BM


The market price of risk is assumed constant, which gives
f(t)
d r(t) = r(t) d t + r(t) d W
f(t) is a Q-BM
under Q where W
Notice that this is a lognormal model like the Black-Scholes model for a
share price

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

416 / 561

Dothan Model
The solution to the SDE is

r(t) = r(s) exp



2
f(t) W
f(s)) ,

(t s) + (W
2

st

So the Fs -conditional distribution of r(t) is lognormal


EQ [r(t)|Fs ] = r(s)e(ts)
2

VarQ [r(t)|Fs ] = r 2 (s)e2(ts) e (ts) 1


In lognormal models, r(t) cannot go negative which is desirable

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

417 / 561

Dothan Model
No closed-form solution for T -bond prices exists
Recall that the cash-bond is given by
Z t

B(t) = exp

r(u) du
0

Take > 0 small, then

Dale Roberts (c) 2011-2014

EQ [B()] = EQ exp

Z


r(u) ds

r(0) + r()
EQ exp

2
X
EQ e e , X Normal
=
Last Modified: 14/10/2014 08:19

418 / 561

Dothan Model

This means that the Dothan model cannot be used to price Eurodollar
futures
This is why Market Models have been introduced

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

419 / 561

Black-Karasinski Model
In 1991, Black and Karasinski introduced the following model.
Start with Y (t) = log(r(t)), whereby
f(t)
d Y (t) = (t)(log (t) Y (t)) d t + (t) d W
f(t) is a Q-BM
where W
Applying Its formula gives



(t)2
f(t)
log r(t) d t + (t)r(t) d W
d r(t) = (t)r(t) log (t) +
2(t)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

420 / 561

Black-Karasinski Model
Under this model, r(t) has a local mean-reversion level of
 2 
(t)
(t) exp
2(t)
Like the Dothan model, there is no closed-form solution for T -bond prices
It is a lognormal model
Similar to the Dothan model, we get

Dale Roberts (c) 2011-2014

EQ [B()]

Last Modified: 14/10/2014 08:19

as 0

421 / 561

The Heath-Jarrow-Morton (HJM) Framework

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

422 / 561

The HJM Framework

The HJM approach gives a very general framework that we can work in to
develop a no-arbitrage model
It describes how a general model for the term-structure should evolve in a
way that is arbitrage-free and which the initial forward-rate curve is part of
the input

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

423 / 561

The HJM Framework


We start with infinitely many processes f (t, T ) for 0 t < T : one process
for each T R+
We will consider the one-factor version: all f (t, T ) are driven by the same
Brownian motion W (t)
We have that f (t, T ) satisfies
d f (t, T ) = (t, T )d t + (t, T )dW (t)

for any fixed maturity T

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

424 / 561

The HJM Framework


Or equivalently,
f (t, T ) = f (0, T ) +

(s, T ) ds +

(s, T ) dW (s)

The coefficients and may depend upon f (t, T ) or the whole forward
curve at time t
Or even more generally, upon F t = ({W (s) : s t})
As all f (t, T ) depend on the same BM, the changes over the whole
forward rate curve are perfectly related to each other (in a nonlinear way)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

425 / 561

The HJM Framework


Technical Conditions

For all T , (t, T ) and (t, T ) are previsible and depend only upon the
history of W (s) up to time t
Z TZ T
|(s, t)| dsd t < for all T

sup k(s, t)k < for all T (for each -path)


s,tT

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

426 / 561

The HJM Framework

We have the forward curve dynamics under P given by


Z t
Z t
f (t, T ) = f (0, T ) +

(s, T ) ds +

(s, T ) dW (s)

The key insight of HJM (1980) is: if we assume that the dynamics of f (t, T )
lead to an arbitrage-free model then there must be a drift condition
Our aim is now to determine what this drift condition is

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

427 / 561

The HJM Framework

Define the risk-free asset


Define the tradable assets
Change measure from P to EMM Q
Determine the replicating strategy for a claim X
Workout the drift restriction for this all to work

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

428 / 561

The Risk-Free Asset


We now need to define the risk-free asset in the HJM framework
The risk-free asset is simply the cash account B(t) which satisfies
dB(t) = r(t)B(t) d t
Recall that it is called risk-free as there is no BM term: there is no price
risk, in the worst case you get back your initial deposit B(0)
Solving the SDE, we get

Dale Roberts (c) 2011-2014

B(t) = B(0) exp

Z


r(u) du

Last Modified: 14/10/2014 08:19

429 / 561

The Risk-Free Asset


Since
f (t, T ) = f (0, T ) +

(s, T )ds +

(s, T )dW (s)

we get
r(T ) = lim f (t, T ) = f (0, T ) +
t%T

(s, T ) dW (s) +

(s, T ) ds

r(T ) may or may not be Markov depending on the form of (s, T )

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

430 / 561

The Risk-Free Asset


We now plug the HJM formula for r(t) into B(t) to get
B(t) = B(0) exp

Z
0

f (0, u) du +

Z tZ
0

(s, u) duds

Z t Z
0

Dale Roberts (c) 2011-2014

(s, u)dW (s) du =

(s, u)du dW (s)

Noticed that we changed the order of integration


Z t Z u

Z t Z t
0

(s, u)du dW (s)

Last Modified: 14/10/2014 08:19

431 / 561

Tradable Assets
Our tradable assets will be all the zero-coupon bonds with prices P(t, T )
for T > t
Our T -bond prices are given by
T

P(t, T ) = exp

f (t, u) du

Plugging in our forward curve f (t, T ) we get


T

Z t Z

P(t, T ) = exp

Dale Roberts (c) 2011-2014

(s, u) du dW (s)

f (0, u) du
t

Z tZ

Last Modified: 14/10/2014 08:19

(s, u) duds

432 / 561

Tradable Assets
Define the discounted asset price
Z(t, T ) :=

P(t, T )
B(t)

Combining integrals and taking B(0) = 1, we get


Z(t) = exp

S(s, T )dW (s)

f (0, u) du

with

Dale Roberts (c) 2011-2014

S(s, T ) :=

Z tZ

(s, u) duds

(s, u) du

Last Modified: 14/10/2014 08:19

433 / 561

Tradable Assets

S(s, T ) is interpreted as the volatility of P(t, T )


Note that S(s, T ) is negative which makes sense as bond prices go down
when interest rates go up
What are the dynamics of Z(t, T )?

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

434 / 561

Tradable Assets

Using Its formula,

d Z(t, T ) = Z(t, T )

Dale Roberts (c) 2011-2014

Z T
1 2
(t, u)du d t + S(t, T )dW (t)
S (t, T )
2
t

Last Modified: 14/10/2014 08:19

435 / 561

Change of Measure

We now want to make the discounted asset price into a martingale


We achieve this by changing measure
The required change of measure drift (market price of risk) for the T -bond
is
Z T
1
1
(t, u) du
(t) = S(t, T )
2
S(t, T ) t
We assume that (t) satisfies the Novikov condition

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

436 / 561

Change of Measure
By Girsanovs theorem, there exists a measure Q equivalent to P such that
Z t
f(t) = W (t) +
W
(s) ds
0

is a Q-BM
Under Q we have
f(t)
d Z(t, T ) = Z(t, T )S(t, T )d W
Therefore, Z(t, T ) is a martingale under Q (assuming S(t, T ) satisfies the
technical condition)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

437 / 561

Change of Measure

It follows that the dynamics of the T -bond prices under Q are given by


f(t)
d P(t, T ) = P(t, T ) r(t) d t + S(t, T )d W
So all T -bonds have expected return equal to the risk-free rate r(t) under
Q
We now need to determine the relationship between (t, T ), (t, T ), and
(t) so that the model is arbitrage-free

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

438 / 561

Replicating Strategies

Suppose that we have a claim X at time S with S < T on a T -bond


We now determine a hedging strategy for X in terms of the cash holding
with value B(t) and the T -bond with value P(t, T )

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

439 / 561

Replicating Strategies
Find the equivalent measure Q under which Z(t, T ) is a martingale
Define the Q-martingale D(t) = EQ [B(S)1 X |F t ]
Find the previsible process (t) such that
Z t
D(t) = D(0) +

(s) d Z(s, T )

Define (t) = D(t) (t)Z(t, T )


The trading strategy ((t), (t)) represents the number of units of B(t)
and P(t, T ) respectively is self-financing, replicating strategy for the claim
X at time S

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

440 / 561

Replicating Strategies

We have already found the EMM Q under which Z(t, T ) is a martingale


We define D(t) = EQ [B(S)1 X |F t ]
D(t) is automatically a Q-martingale by the tower property for conditional
expectations

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

441 / 561

Replicating Strategies

By the Martingale Representation Theorem, there exists a previsible


process (t) such that
Z t
D(t) = D(0) +

(s) d Z(s, T )

Alternatively written d D(t) = (t)d Z(t, T )

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

442 / 561

Replicating Strategies

Suppose we have a trading strategy which holds (t) units of the T -bond
with price P(t, T )
Set (t) = D(t) (t)Z(t, T ) to be the amount of cash B(t) that we hold
at time t
The value of the portfolio at time t is given by

Dale Roberts (c) 2011-2014

V (t) = B(t)D(t) = B(t)EQ [B(S)1 X |F t ]

Last Modified: 14/10/2014 08:19

443 / 561

Replicating Strategies

The instantaneous change in the portfolio is


dV (t) = d(B(t)D(t))
= D(t)dB(t) + B(t)d D(t)
= r(t)B(t)D(t)d t + B(t)(t)d Z(t, T )

as d B(t) = r(t)B(t)d t and d D(t) = (t)d Z(t, T )

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

444 / 561

Replicating Strategies
We have
d P(t, T ) = d(B(t)Z(t, T )) = r(t)B(t)Z(t, T )d t + B(t)d Z(t, T )
The instantaneous investment gain is

Dale Roberts (c) 2011-2014

(t)dB(t) + (t)d P(t, T )


= [D(t) (t)Z(t, T )]B(t)r(t)d t
+ (t)[r(t)B(t)Z(t, T )d t + B(t)d Z(t, T )]
= r(t)B(t)D(t)d t + (t)B(t)d Z(t, T )
= d V (t)

Last Modified: 14/10/2014 08:19

445 / 561

Replicating Strategies

The investment is self-financing as


(t)dB(t) + (t)d P(t, T ) = d V (t)
Therefore, V (t) is the value at time t of the claim X with maturity S

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

446 / 561

The Arbitrage-Free Market

Suppose that X = 1 at maturity S then X is an S -bond payoff


We have determined that the arbitrage-free price is then

ZS



r(u) du F t
P(t, S) = B(t)EQ B(S)1 |F t = EQ exp

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

447 / 561

The Arbitrage-Free Market

The discounted S -bond is


Z(t, S) =



P(t, S)
= EQ B(S)1 |F t
B(t)

so that Z(t, S) is a Q-martingale


This must be true for any maturity S that we choose
This implies that all bonds must be turned into Q-martingales by the same
change of measure: they all have the same market price of risk (t)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

448 / 561

The Arbitrage-Free Market


Thus for all maturities T ,
1
1
S(t, T )
2
S(t, T )

(t, u) du = (t)
t

Or equivalently,

Dale Roberts (c) 2011-2014

Z
t

1
(t, u) du = S(t, T )2 (t)S(t, T )
2

Last Modified: 14/10/2014 08:19

449 / 561

The Arbitrage-Free Market


Recall that
S(t, T ) =

(s, u) du

so that

And differentiating

Dale Roberts (c) 2011-2014

S(t, T ) = (t, T )
T
RT
t

(t, u) du with respect to T we get

(t, T ) = (t, T ) [(t) S(t, T )]

Last Modified: 14/10/2014 08:19

450 / 561

The Arbitrage-Free Market


Now recall that our original model for the forward dynamics under P was
d f (t, T ) = (t, T )d t + (t, T )dW (t)
Substituting in our arbitrage-free drift (t, T ) and changing to Q gives


f(t) (t)d t
d f (t, T ) = (t, T )d t + (t, T ) W
f(t) (t, T )(t)d t
= (t, T )[(t) S(t, T )]d t + (t, T )d W

Dale Roberts (c) 2011-2014

f(t)
= (t, T )S(t, T )d t + (t, T )d W

Last Modified: 14/10/2014 08:19

451 / 561

The Arbitrage-Free Market


Therefore, under Q we have
f(t)
d f (t, T ) = (t, T )S(t, T )d t + (t, T )d W
Notice that the dynamics are solely determined from the volatility (t, T )
Consequently,

Dale Roberts (c) 2011-2014

r(t) = f (0, t)

(s, T )S(s, t) ds +

Last Modified: 14/10/2014 08:19

f(s)
(s, t)d W

452 / 561

Week 10

Multifactor Models

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

453 / 561

Dale Roberts (c) 2011-2014

Multifactor Models

Last Modified: 14/10/2014 08:19

454 / 561

Introduction

We shall now look at models which include more than one source of
randomness
When you look at historical interest rate data to see that changes in
interest rates with different maturities are not perfectly correlated as
predicted by one-factor model
We can look at some data from

http://www.rba.gov.au/statistics/tables/

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

455 / 561

Yield (%)
5
6

Cash Yield Vs. Bond Yield

Dale Roberts (c) 2011-2014

2000

2005
Year

Last Modified: 14/10/2014 08:19

2010

456 / 561

Long Yield (%)


5
6

Cash Yield Vs. Bond Yield

Dale Roberts (c) 2011-2014

5
6
Short Yield (%)

Last Modified: 14/10/2014 08:19

457 / 561

Introduction

Data analysis shows that multi-factor models may be more appropriate


Clearly, a one-factor time-homogeneous model is inappropriate

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

458 / 561

Introduction

One of the main reasons is to deal with more-complex interest rate


options which refer to two or more stochastic underlying quantities
For example, consider the option with the following payoff at time T :
X = max [R(T, T + 5) R(T, T + 10), 0]
A one-factor model would probably overprice this contract due to the
perfect (non-linear) correlation between the rates in such a model

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

459 / 561

Multi-factor Affine Models

We have already seen one-factor affine models: Vasicek (1977),


Cox-Ingersoll-Ross (1985), Ho-Lee (1986), and Hull-White (1990)
We now look at multi-factor models which have the affine form

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

460 / 561

Multi-factor Affine Models


Consider an n-dimensional diffusion model

X 1 (t)
X 2 (t)

X (t) =
...
X n (t)
The model is called affine if the T -bond prices are of the form

n
X
P(t, T ) = exp A(t, T ) +
B j (t, T )X j (t)

Dale Roberts (c) 2011-2014

j=1

Last Modified: 14/10/2014 08:19

461 / 561

Multi-factor Affine Models


Or written in vector notation as


P(t, T ) = exp A(t, T ) + B(t, T ) T X (t)

where

B1 (t, T )
B2 (t, T )

B(t, T ) =
..

Bn (t, T )
B(t, T ) T is the transpose of B(t, T )

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

462 / 561

Multi-factor Affine Models

The model is time homogeneous if X is time homogeneous and A(t, T )


and B(t, T ) are functions of T t only
We shall now restrict ourselves to the time homogeneous models for
simplicity and use A(T t) and B(T t)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

463 / 561

Multi-factor Affine Models

Recall that for one-dimensional affine models of the short-rate r(t) we


must have

f(t)
d r(t) = (a + br(t))d t + r(t) + d W
Q: What form must the SDE of X (t) have in the n-dimensional case if X (t)
is affine?

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

464 / 561

Multi-factor Affine Models


Theorem (Duffie and Kan 1996)
Suppose P(t, t + ) = exp[A() + B() T X (t)]. Then X (t) has the SDE
f(t),
d X (t) = ( + BX (t))d t + S D(X (t))d W

where = (1 , . . . , n ) T is a constant vector, B = (bi j ) and S = (i j ) are a


constant matrices and D(X (t)) is the diagonal matrix

q
1T X (t) + 1
0

0

0
2T X (t) + 2 0
0

..
.
.
..
..
D(X (t)) =

..
..
...

.
.

0
nT X (t) + n
where each i R and each i = (i1 , , in ) T is a constant vector.
Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

465 / 561

Multi-factor Affine Models

Within this framework, we need to ensure that each of the volatility


processes
iT X (t) + i ,

for i = 1, . . . , n remain positive (preferably strictly positive)


Duffie and Kan (1996) provide conditions on the parameters for the
volatility processes to remain positive

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

466 / 561

Multi-factor Affine Models


Consider the spot rates in this framework
R(t, t + j ) =

1
[A( j ) + B() T X (t)]
j

for j = 1, . . . , n for terms to maturity 1 < 2 < < n


In vector notation, R(t) = (R(t, t + 1 ), . . . , R(t, t + n )) T , giving
R(t) = aR + BR X (t)

for a constant vector aR Rn and constant matrix BR Rnn

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

467 / 561

Multi-factor Affine Models


If the matrix BR is invertible (i.e., exists BR1 such that BR1 BR = I ) then
X (t) = BR1 (R(t) aR )
Hence,



P(t, t + ) = exp A() + B() T BR1 (R(t) aR )



()R(t)
= exp A()
+B

for suitable functions A()


and B ()
Therefore, it the model is affine in X (t) then it can be reformulated in a
way which is affine in R(t)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

468 / 561

Multi-factor Affine Models

These transformations make the n-factor model simple to calibrate if we


are given at time t the market spot rates
R(t, t + 1 ), . . . , R(t, t + n )

Choosing a specific model gives A()


and B () then the bond prices are



()R(t)
P(t, t + ) = exp A()
+B

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

469 / 561

Multi-factor Affine Models


Models that are time-homogeneous under Q do not need to be time
homogeneous under the real-world measure P
Assume that the vector of market prices of risk
(t) = (1 (t), . . . , n (t)) T

is previsible and satisfies the Novikov condition


f(t) be a n-dimensional Q-BM. Then
Let W
Z t
f(t) +
W (t) := W
(s) ds
0

is a (n-dimensional) P-BM

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

470 / 561

Multi-factor Affine Models

The SDE for X (t) under P then becomes

Dale Roberts (c) 2011-2014

d X (t) = [ + BX (t)]d t + SD(X (t))[dW (t) + (t)d t]


= [ + BX (t) + SD(X (t))(t)]d t + S D(X (t))dW (t)

Last Modified: 14/10/2014 08:19

471 / 561

Multi-factor Affine Models

We shall now consider three special cases in the affine framework:


Gaussian multi-factor models
Generalized CIR models
The Longstaff and Schwartz model (1992)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

472 / 561

Gaussian Multi-factor Models

Gaussian models allow interest rates to become negative


The first multifactor extension of the Vasicek (1977) model was developed
by Langetieg (1980)
Later works include Beaglehole-Tenney (1991) for the general theory and
Babbs-Nowman (1999) for model calibration
We shall now consider the theory of time-homogeneous Gaussian models

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

473 / 561

Gaussian Multi-factor Models


Let X (t) be an n-dimensional diffusion process with SDE
f(t)
dX (t) = BX (t)d t + K W
f(t) is a
where B and K are real-valued constant n n matrices and W
n-dimensional Q-BM
The risk-free rate of interest is
r(t) = + T X (t)

where = (1 , . . . , n ) T Rn
If all i 6= 0 we can scale X (t) and assume i = 1

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

474 / 561

Gaussian Multi-factor Models

The matrix B has a spectral decomposition B = BR B L


= diag(1 , . . . , n ) is the diagonal matrix of eigenvalues of B (some
these eigenvalues may be complex)
B L and BR are the matrices of left and right eigenvectors of B . Column i of
BR is the right eigenvector of B corresponding to i
The columns of BR are scaled in a way which ensures BR B L = I

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

475 / 561

Gaussian Multi-factor Models

This decomposition is not unique


For X (t) to be stationary we require that the real parts of all eigenvalues
1 , . . . , n are negative
We have B k = BR k B L which is useful for defining the exponential of a
matrix

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

476 / 561

Gaussian Multi-factor Models


Recall that e x = 1 + x +

x2
2!

x3
3!

+ . . . so we have

exp(B) = BR exp()B L

where, after defining 0 = I ,

Dale Roberts (c) 2011-2014

exp() = I + +
=

2 3
+
+ ...
2!
3!

X
k
k=0

k!

Last Modified: 14/10/2014 08:19

477 / 561

Gaussian Multi-factor Models


Define

Y (t) := et B L X (t)

Applying the n-dimensional Its formula,


f(t)
dY (t) = et B L K d W
Hence, we have

Dale Roberts (c) 2011-2014

Y (t) = Y (0) +

f(u)
eu B L K d W

Last Modified: 14/10/2014 08:19

478 / 561

Gaussian Multi-factor Models


But as Y (t) = et X (t) and BR B L = I we have
X (t) = BR e t Y (t)
Using the fact that BR e t B L = e t B , it follows that
Z t

f(u)
e(tu)B K d W

X (t) = e t B X (0) +

This is an n-dimensional Orstein-Uhlenbeck process (i.e., solution to


n-dim Vasicek model that mean-reverts to zero)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

479 / 561

Gaussian Multi-factor Models


The requirement that the real parts of the eigenvalues 1 , . . . , n are
negative ensures that
exp(t) 0

and

exp(t B) 0

as t
Think of the 1-dimensional case where a R, then

Dale Roberts (c) 2011-2014

et a 0

as

Last Modified: 14/10/2014 08:19

480 / 561

Gaussian Multi-factor Models


Now let
R(T ) =

r(t) d t
0

= T +

T X (t) d t

This is Normally distributed with

Dale Roberts (c) 2011-2014

EQ [R(T )] = T + T BR 1 (e T I)B L X (0)


Z T

T BR 1 (e T I)B L K K T B LT (e T I)1 BRT d t

VarQ [R(T )] =

Last Modified: 14/10/2014 08:19

481 / 561

Gaussian Multi-factor Models


Example
Simplest 2-dimensional case is

1 0
B = BR B L
B=
0 1
then

BR = B L = I, = I

EQ [R(T )] = T + T 1 (e T I)X (0)


= T T 1 (I e t )X (0)

and compare to the solution to the 1-dimensional Vasicek model

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

482 / 561

Gaussian Multi-factor Models


We then have that T -bond prices at t = 0 are given by
Z t


P(0, T ) = EQ
r(t) d t F0
0

= EQ [ex p(R(T ))|X (0)]


= exp EQ [R(T )|X (0)] +


1
VarQ [R(T )|X (0)]
2

Recall identifying Normals: If Z N (, 2 ) then

 Z
1 2 2
E e
= exp +
2

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

483 / 561

Gaussian Multi-factor Models


Example
(Beaglehole and Tenney 1991) We take
r(t) = (X 1 (t) + 1 ) + (X 2 (t) + 2 )
{z
}
|
{z
} |
inst. price inflation

where

inst. real interest

f1 (t),
d X 1 (t) = 1 X 2 (t)d t + 11 d W
f1 (t) + 22 d W
f2 (t).
d X 2 (t) = 2 X 1 (t)d t + 21 d W

Q: What are the matrices B and K ?

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

484 / 561

Gaussian Multi-factor Models


Example
We take
r(t) = + X 1 (t),

where

f1 (t),
d X 1 (t) = 1 (X 2 (t) X 1 (t))d t + 11 d W
f1 (t) + 22 d W
f2 (t).
d X 2 (t) = 2 X 2 (t)d t + 21 d W

In this model, + X 2 (t) behaves like a stochastic (local) mean reversion level
for r(t).

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

485 / 561

Generalized CIR Models


A number of multi-factor CIR models have been proposed
In terms of the general framework
f(t),
dX (t) = ( + BX (t))d t + S D(X (t))d W
q

where D(X (t)) = diag( 1T X (t), . . . , nT X (t)) and


P(t, t + ) = exp[A() + B() T X (t)]
The risk-free rate of interest is then defined as

Dale Roberts (c) 2011-2014

r(t) =

n
X

X i (t)

i=1

Last Modified: 14/10/2014 08:19

486 / 561

Generalized CIR Models


As D(X (t)) is diagonal, the X i (t) are independent
We get that

Dale Roberts (c) 2011-2014



r(u) du F t

n
X

P(t, T ) = EQ exp

= EQ exp

n
Y

EQ exp
t

i=1

= exp

i=1
T

n
X



X i (u) du F t


X i (u) du F t

Ai (T t)

i=1

Last Modified: 14/10/2014 08:19

n
X

Bi (T t)X i (t)

i=1

487 / 561

Generalized CIR Models


P(t, T ) is the product of n one-dimensional CIR
The coefficients are
Ai () =
Bi () =

2i i
2i

2i e(i +i )/2
log
(i + i )(ei 1) + 2i


2(ei 1)
,
(i + i )(ei 1) + 2i

2i + 22i

If we have 2i i > 2i then the probability that X i (t) hits zero is zero

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

488 / 561

The Longstaff and Schwartz Model


A variation of the two-factor CIR model was proposed by Longstaff and
Schwartz in 1992
Take for i = 1, 2 the SDEs
dYi (t) = i (i Yi (t)) d t +

fi (t)
Yi (t)d W

f1 and W
f2 are independent Q-Brownian motions
where W
For constants c1 and c2 , define

Dale Roberts (c) 2011-2014

r(t) := c1 Y1 (t) + c2 Y2 (t),

V (t) := c12 Y1 (t) + c22 Y2 (t)

Last Modified: 14/10/2014 08:19

489 / 561

The Longstaff and Schwartz Model

d r(t) = c1 d Y1 (t) + c2 Y2 (t)

f1 (t)
= c1 1 (1 Y1 (t))d t + Y1 (t)d W

f2 (t)
+ c2 2 (2 Y2 (t))d t + Y2 (t)d W

(c2 1 c1 2 )r(t) + (2 1 )V (t)


= c1 1 1 + c2 2 2
dt
c2 c1
v
v
u
u
t c1 (c2 r(t) V (t))
t c2 (V (t) c1 r(t))
f1 (t) +
f2 (t)
+
dW
dW
c2 c1
c2 c1

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

490 / 561

The Longstaff and Schwartz Model

It can be shown that instantaneous variance of r(t) is V (t)d t which is the


reason why the model was formulated this way
Also, by using r(t) and V (t) we focus on the variables that are believed to
influence the pricing of derivatives, instead of using Y1 (t) and Y2 (t)
r(t) is the level and V (t) is the volatility
If c1 < c2 then the form of r(t) and V (t) means that V (t) is limited to the
region (c1 r(t), c2 (t))
The model has richer dynamics when c1 and c2 are far apart

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

491 / 561

Multifactor Heath-Jarrow-Morton Models


We can easily extend the one-factor HJM framework to the multi-factor
setting
We assume the existence of a given model under the risk-neutral measure
Q
We model the T -bond dynamics
f(t)]
d P(t, T ) = P(t, T )[r(t) d t + S(t, T ) T d W

where the volatility function S(t, T ) = (S1 (t, T ), . . . , Sn (t, T )) T is


f(t) is a standard n-dim BM under Q
previsible and W

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

492 / 561

Multifactor Heath-Jarrow-Morton Models

The forward dynamics are given by


f(t)
d f (t, T ) = (t, T ) T S(t, T )d t + (t, T ) T d W

where

Dale Roberts (c) 2011-2014

(t, T ) = (1 (t, T ), . . . , n (t, T )) T

i (t, T ) =
Si (t, T )
T

Last Modified: 14/10/2014 08:19

493 / 561

Multifactor Heath-Jarrow-Morton Models


Example
(Rebonato 1996)
Suppose that primary sources of randomness in forward-rate curve are:
slope of curve and twist in curve
In a Gaussian setting, this can be modelled by
1 (t, T ) = 11 e1 (T t)
2 (t, T ) = 21 e1 (T t) + 22 e2 (T t)
Example: 0 < 2 < 1 then can we achieve a twist by adding the
constraints 22 > 0 and 21 < 22

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

494 / 561

Week 11

The Forward-Measure Approach

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

495 / 561

Dale Roberts (c) 2011-2014

The Forward-Measure Approach

Last Modified: 14/10/2014 08:19

496 / 561

The Forward-Measure Approach

In pricing problems that we have seen so far we have expressed prices in


terms of expectations under the risk-neutral equivalent martingale
measures Q
Sometimes the calculation can be made easier by making a second
change of measure from Q to the forward measure
In recent years, this idea has been critical in the development of totally
new approaches to interest rate modelling

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

497 / 561

A New Numeraire
For a T -bond, we have the dynamics
f(t)]
d P(t, T ) = P(t, T )[r(t) d t + S(t, T )d W
f(t) is a Q-BM
where W
Previously, we considered the discounted price process
Z(t, U) =

where B(t) = exp

R t
0

P(t, U)
B(t)

r(s) ds is the cash account at time t

The discounted price was a martingale under the same measure Q for all
maturity dates U

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

498 / 561

A New Numeraire
Lets change the numeraire from B(t) to P(t, T ). Thus we consider
Y (t, U) =

P(t, U)
P(t, T )

Then we get

1
1
1
d Y (t, U) =
d P(t, U)+P(t, U)d
+d P(t, U)d
P(t, T )
P(t, T )
P(t, T )

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

499 / 561

A New Numeraire
Then
f(t)
d Y (t, U) = Y (t, U) (S(t, U) S(t, T )) d W
+ Y (t, U)S 2 (t, T ) d t Y (t, U)S(t, U)S(t, T ) d t
c(t) with
Now suppose we define a new process W
c(0) = 0
W

and

c(t) = d W
f(t) S(t, T ) d t
dW

Then

Dale Roberts (c) 2011-2014

c(t) + S(t, T )d t]
d Y (t, U) = Y (t, U) (S(t, U) S(t, T )) [d W
Y (t, U)S(t, T )[S(t, U) S(t, T )]d t
c(t)
= Y (t, U)[S(t, U) S(t, T )]d W
Last Modified: 14/10/2014 08:19

500 / 561

A New Numeraire

So we have
c(t)
dY (t, U) = Y (t, U)[S(t, U) S(t, T )]d W
It appears that we can find a suitable measure under which Y (t, U) for all
U > T are martingales

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

501 / 561

Change of Measure
Define (t) = S(t, T )
By Girsanovs theorem, there exists a measure P T which is equivalent to Q
such that
Z
t

c(t) = W
f(t) +
W

(s) ds

is a Brownian motion under P T


Under the assumption that Novikovs condition holds:

ZU
1
EQ exp(
(t)2 d t) <
2
0
P T is called a forward measure

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

502 / 561

Change of Measure

The forward measure depends on the forward maturity date T but not
upon the maturity date U of the bond under consideration
The Y (t, U) for U > T are all martingales under the measure P T

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

503 / 561

Derivative Payments
Let X be a claim contingent upon F T payable at time T
Define
D(t) = EPT


X
F t = EPT [X |F t ]
P(T, T )

D(t) is martingale under P T


Y (t, U) is also a martingale under P T
By MRT, there exists a previsible process (t) such that

Dale Roberts (c) 2011-2014

d D(t) = (t)d Y (t, U)

Last Modified: 14/10/2014 08:19

504 / 561

A Replicating Strategy
Let (t) = D(t) (t)Y (t, U)
Consider the investment strategy which holds at time t , (t) units of the
numeraire P(t, T ) and (t) units of U -bond
The value of this portfolio at time t is
V (t) = P(t, T )D(t)
Now

Dale Roberts (c) 2011-2014

f(t)]
d P(t, T ) = P(t, T )[r(t)d t + S(t, T )d W
c(t) + S(t, T )d t)]
= P(t, T )[r(t)d t + S(t, T )(d W
c(t))
= P(t, T )([r(t) + S 2 (t, T )]d t + S(t, T )d W

Last Modified: 14/10/2014 08:19

505 / 561

A Replicating Strategy

And we have

Dale Roberts (c) 2011-2014

d D(t) = (t)dY (t, U)


c(t)
= (t)Y (t, U)(S(t, U) S(t, T ))d W

Last Modified: 14/10/2014 08:19

506 / 561

A Replicating Strategy
Thus
d V (t) = D(t)d P(t, T ) + P(t, T )d D(t) + d P(t, T )d D(t)
c(t))
= P(t, T )D(t)((r(t) + S 2 (t, T ))d t + S(t, T )d W

Dale Roberts (c) 2011-2014

c(t)
+ (t)P(t, T )Y (t, U)(S(t, U) S(t, T ))d W
c(t)]
+ (t)P(t, T )Y (t, U)[(r(t) + S 2 (t, T ))d t + S(t, T )d W
c(t)]
[(S(t, U) S(t, T ))d W
c(t)]
= V (t)[(r(t) + S 2 (t, T ))d t + S(t, T )d W
c(t)
+ (t)P(t, U)(S(t, U) S(t, T ))d W
+ (t)P(t, U)S(t, T )(S(t, U) S(t, T )) d t

Last Modified: 14/10/2014 08:19

507 / 561

A Replicating Strategy
The instantaneous investment gain is thus
(t)d P(t, T ) + (t)d P(t, U)
c(t))
= (D(t) (t)Y (t, U))P(t, T )((r(t) + S 2 (t, T ))d t + S(t, T )d W
c(t) + S(t, T )d t))
+ (t)P(t, U)(r(t)d t + S(t, U)(d W
= d V (t)
Hence ((t), (t)) is a self-financing strategy
V (T ) = X so it is also a replicating strategy

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

508 / 561

A Replicating Strategy

Theorem
The price at time t for the claim X payable at time T is
V (t) = P(t, T )EPT [X |F t ]
This price is independent of the choice of the hedging asset P(t, U)
But the hedging strategy depends upon the choice of U

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

509 / 561

Evaluation of a Derivative Price

Suppose that the claim X at T is a function of P(T, U)


Under the martingale measure Q we have

Z T
V (t) = EQ exp



r(s) ds X F t

Evaluation of this expectation requires the knowledge of the joint


RT
distribution of P(T, U) and exp( t r(s) ds) under Q

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

510 / 561

Evaluation of a Derivative Price

On the other hand, under the forward measure approach


V (t) = P(t, T )EPT [X |F t ]
We only need to establish the distribution of X under P T
This is a lot easier to calculate

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

511 / 561

Evaluation of a Derivative Price


Lemma
The forward price at time t for delivery of P(T, U) at time T is given by
EPT [P(T, U)|F t ].

Proof: Since Y (t, U) is a martingale under P T we have


EPT [Y (T, U)|F t ] = Y (t, U).

Therefore,
EPT


P(T, U)
P(t, U)
,
F t = EPT [P(T, U)|F t ] =
P(T, T )
P(t, T )

which is the forward price.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

512 / 561

Evaluation of a Derivative Price


Lemma
Suppose that the T -bond volatility S(v, V ) is deterministic. Then under P T ,
P(t, U) is lognormal with
Z T
(S(v, U) S(v, T ))2 d v

VarPT [log P(T, U)|F t ] = b2 =

and
EPT [log P(T, U)|F t ] = a = log

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

P(t, U) 1 2
b
P(t, T ) 2

513 / 561

Vasicek Model
Example
Under the Vasicek model,
S(v, U) =

(1 e(Uv) )

Which gives
2
2
b = 2

Dale Roberts (c) 2011-2014

(e(T v) e(Uv) )2 d v
t

(1 e(UT ) )2 (1 e2(T t) )
3
2
2

Last Modified: 14/10/2014 08:19

514 / 561

Evaluation of a Derivative Price


Lemma
Suppose that a random variable Y has a lognormal distribution with
parameters a and b2 ; that is, log Y N (a, b2 ). Then for any constant K > 0:
E[(Y K)+ ] = E[Y ](h) K(h b),

where

log(E[Y ]/K) + 12 b2
a + b2 log K
h=
=
b
b
and is the CDF of a standard Normal distribution.

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

515 / 561

Evaluation of a Derivative Price


We can apply this lemma to price a call option with maturity T and strike
price K on a U -bond with price P(t, U)
We get
EPT [(P(T, U) K)+ |F t ] = EPT [P(T, U)|F t ](h) K(h b)

where

1 2
1
P(t, U)
log K + b
h=
log
b
P(t, T )
2
1
P(t, U)
1
= log
+ b
b
K P(t, T ) 2

Recall

Dale Roberts (c) 2011-2014

b =

(S(v, U) S(v, T ))2 d v


t

Last Modified: 14/10/2014 08:19

516 / 561

Evaluation of a Derivative Price

Hence, the price at time t of the call option is


V (t) = P(t, U)EPT [(P(T, U) K)+ |F t ]
= P(t, U)(h) K P(t, T )(h b)
Notice the similarity to the Black-Scholes formula
This formula only holds when bond prices are lognormally distributed

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

517 / 561

Equity Options with Stochastic Interest


The forward measure approach can also be used to price equity options
when the cash rate r(t) is stochastic
Suppose we have the usual model
f1 (t)]
d P(t, T ) = P(t, T )[r(t)d t + S(t, T )d W
f(t) is a BM under the martingale measure Q
where W
Consider an equity index R(t) which includes full reinvestment of
dividends
Under Q suppose that R(t) has the SDE
f1 (t) + 2 d W
f2 (t)]
dR(t) = R(t)[r(t)d t + 1 d W
f2 (t) is second BM under Q independent of W
f1 (t)
where W
Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

518 / 561

Equity Options with Stochastic Interest


Let D(t) = EPT [ f (R(T ))|F t ]. This is a martingale under P T
The processes
b (t, U) =
P

P(t, U)
,
P(t, T )

b(t) =
R

R(t)
P(t, T )

are both martingales under P T


By the multidimensional MRT, there exist previsible processes P (t) and
R (t) such that

Dale Roberts (c) 2011-2014

b (t, U) + R (t)d R
b(t)
d D(t) = P (t)d P

Last Modified: 14/10/2014 08:19

519 / 561

Equity Options with Stochastic Interest


Define

b (t, U) R (t)R
b(t)
(t) = D(t) P (t) P

Consider the portfolio process which holds at time t


P (t)

of P(t, U)
R (t) of R(t)
(t) of P(t, T )

The value at time t is


V (t) = P(t, T )D(t) = P(t, T )EPT [ f (R(T ))|F t ]
This portfolio is self-financing and replicating ( t = T ) and V (t) is the
unique no-arbitrage price at time t

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

520 / 561

Equity Options with Stochastic Interest


Theorem
Consider the European call option f (R(T )) = (R(T ) K)+ . The price for this
option at time t is
V (t) = P(t, T )[F (t)(h) K(h b)]

where
h=
b2 =

log(F (t)/K) + 21 b2
b
T

(1 S(v, T ))2 d v + 22 (T t)
t

F (t) =

Dale Roberts (c) 2011-2014

R(t)
P(t, T )

[forward price for R(T )]

Last Modified: 14/10/2014 08:19

521 / 561

Week 12

Market Models

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

522 / 561

Dale Roberts (c) 2011-2014

Market Models

Last Modified: 14/10/2014 08:19

523 / 561

Introduction

Instantaneous forward rates are an unobservable quantity


This makes the forward curve T 7 f (t, T ) difficult to estimate
This lead to the suggestion that maybe one should develop a model of
directly observable (i.e., market quoted) quantities such as LIBOR

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

524 / 561

Introduction

A key modelling assumption was that relevant market rates are lognormal
This was an assumption already used by traders to price some interest
rate derivatives
The market practice was/is to use Blacks formula

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

525 / 561

Blacks Formula
Definition
A caplet with reset date T and settlement date T + pays the holder the
difference between a simple market rate F (T, T + ) and the strike rate
Its cash flow at time T + is
(F (T, T + ) )+
We write Cpl(t; T, T + ) for the price at time t of the caplet with reset
date T and settlement date T +

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

526 / 561

Blacks Formula
Definition
A Cap is a strip of caplets. It consists of:
A number of future dates T0 < T1 < < Tn with Ti Ti1 =
A cap rate
We write Cp(t) for its price at time t
We have
Cp(t) =

n
X

Cpl(t; Ti1 , Ti )

i=1

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

527 / 561

Blacks Formula
It is market practice to price a cap/floor according to Blacks formula
Let t T0 . Blacks formula for the value of the i th caplet is
Cpl(t; Ti1 , Ti ) = P(t, Ti )[F (t; Ti1 , Ti )(d1 (i; t)) (d2 (i; t))]

where
d1,2 (i; t) =

log

F (t;Ti1 ,Ti )

12 (t)2 (Ti1 t)
p
(t) Ti1 t

Here (t) is the cap implied volatility (it is the same for all caplets
belonging to a cap)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

528 / 561

Blacks Formula

Typically, t = 0, T0 = and = 3 months (US) or = 6 months (Europe)


Caps are quoted in the market in terms of their implied volatilities
So to price a Cap, one takes the implied volatility (t) quoted at time t
and plugs it into the Black model

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

529 / 561

Blacks Formula

ATM vols (in %)


13 14 15 16 17 18 19

ATM Cap Volatilities

0
Dale Roberts (c) 2011-2014

10

15
20
Maturity (in years)
Last Modified: 14/10/2014 08:19

25

30
530 / 561

Blacks Formula

This approach is completely heuristic!


It is based on the assumption that the simple market rates F (t; Ti1 , Ti )
are lognormal
It is a challenge for any market realistic interest rate model to match the
given volatility curve observed in the market

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

531 / 561

Market Models

We are now going to look at the development of Market Models which


attempt to construct a model of observable interest rates such as LIBOR
Sometimes denoted LMM (in job ads): LIBOR Market Model
The first developments all happened very quickly in 1997
Miltersen, Sandmann & Sondermann (1997): formulae for zero-coupon
bond options, caps and floors
Further improvements were obtained by Brace, Gatarek, and Musiela
(1997), Jamshidian (1997), and Musiela and Rutkowski (1997)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

532 / 561

LIBOR

There are two types of market interest rates: LIBOR and forward LIBOR
LIBOR is an annualized, simple rate of interest that will be delivered at the
end of a defined period
The -LIBOR denoted by F (T, T + ) means that an investment of 1 at
time T will grow to 1 + F (T, T + ) at time T +
LIBOR is always quoted as an annual rate of interest even though it
typically applies over shorter periods

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

533 / 561

LIBOR

For a fixed > 0, the forward -period LIBOR for the future date T
prevailing at time t is the simple forward rate

1
P(t, T )
L(t, T ) := F (t; T, T + ) =
1
P(t, T + )
With this notation, L(T, T ) = F (T ; T, T + ) is the LIBOR rate

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

534 / 561

LMM: Heuristic Derivation

Let P T + be the (T + ) forward measure


Recall that

P(t, T )
P(t, T + )

is a martingale under P T +

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

535 / 561

LMM: Heuristic Derivation


We also know that

P(t, T )
P(t, T )
d
=
T,T + (t)dW T + (t)
P(t, T + )
P(t, T + )
R T +
where T,T + (t) = T (t, u) du

Using this with the definition of L(t, T ) gives

1 P(t, T )
1
P(t, T )
d L(t, T ) = d
=
T,T + (t)dW T + (t)

P(t, T + )
P(t, T + )
1
= (L(t, T ) + 1) T,T + (t)dW T + (t)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

536 / 561

LMM: Heuristic Derivation


We have
d L(t, T ) =

1
(L(t, T ) + 1) T,T + (t)dW T + (t)

Assumption: there exists a Rn -valued deterministic function (t, T ) such


that
L(t, T )
(t, T )
T,T + (t) =
L(t, T ) + 1
Plugging this in to the above formula, we get

Dale Roberts (c) 2011-2014

d L(t, T ) = L(t, T )(t, T )dW T + (t)

Last Modified: 14/10/2014 08:19

537 / 561

LMM: Heuristic Derivation


The solution of that SDE is given by
Z t
L(t, T ) = L(s, T ) exp

(u, T )dW

T +

1
(u)
2

t
2

k(u, T )k du

So the P T + distribution of log L(t, T ) conditional on F t is Gaussian with


mean
Z T
1
log L(t, T )
k(s, T )k2 ds
2 t

and variance

Dale Roberts (c) 2011-2014

k(s, T )k2 ds
t

Last Modified: 14/10/2014 08:19

538 / 561

LMM: Heuristic Derivation


The time t price of a caplet with reset date T , settlement date T + and
strike rate is thus
i
h R T +

EQ e t r(s)ds (L(T, T ) )+ F t
= P(t, T + )EPT + [(L(T, T ) )+ |F t ]
= P(t, T + )(L(t, T )(d1 (t, T )) (d2 (t, T )))

where

Dale Roberts (c) 2011-2014

d1,2 (t, T ) :=

log

L(t,T )

21

RT

k(s, T )k2 ds
R T
1/2
2 ds
k(s,
T
)k
t

Last Modified: 14/10/2014 08:19

539 / 561

LMM: Heuristic Derivation


This is just Blacks formula for the caplet price with
v
Z T
u
t 1
k(s, T )k2 ds
(t) =
Tt t
This means that any HJM model satisfying
T,T + (t) =

L(t, T )
(t, T )
L(t, T ) + 1

for some deterministic (t, T ) yields Blacks formula for caplet prices

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

540 / 561

LMM: Heuristic Derivation

The question is: Does such a HJM model exist?


The answer is yes, but the construction and proof are not easy
This is carried out in the paper by Brace, Gatarek and Musiela (1997)

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

541 / 561

LMM: Direct Approach

We can also construct the LMM using a direct approach


We dont use the HJM framework (although it can be implicitly implied)
We work with the forward measure concept

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

542 / 561

LMM: Direct Approach

Fix a time horizon TM = M with M N


Fix our probability space (, F , (F t ) tTM , P TM )
Under P TM we have a d -dimensional Brownian motion W TM
PTM is the TM -forward measure
Define the dates Tm := m for m = 0, 1, . . . , M

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

543 / 561

LMM: Direct Approach

We now construct a model of M forward LIBOR rates with maturities


T0 , T1 , . . . , TM 1
For every m M 1, we define (t, Tm ) as the volatility of L(t, Tm )
We assume there is an initial positive and nonincreasing term structure of
M bond prices
P(0, T0 ), . . . , P(0, TM 1 )

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

544 / 561

LMM: Direct Approach


Definition
For notational convenience, if S in an It process and H is a predictable
process we write
Z
H S =

H(u)dS(u)

Examples: Recall the stochastic exponential of an It process X is given by


E t (X ) := exp(X 12 X t ) then
Z t
((, T ) W T ) t =

(u, T )dW T (u)

E t ((, T ) W ) = exp
T

Z
0

Dale Roberts (c) 2011-2014

1
(u, T )dW (u)
2
T

Last Modified: 14/10/2014 08:19

t
2

|(u, T )| du

0
545 / 561

LMM: Direct Approach


First, lets postulate the dynamics of L(t, TM 1 ) = F (t; TM 1 , TM ) as
d L(t, TM 1 ) = L(t, TM 1 )(t, TM 1 )dW TM (t),

t [0, TM 1 ]

with initial condition


1
L(0, TM 1 ) =

P(0, TM 1 )
1
P(0, TM )

The solution is

Dale Roberts (c) 2011-2014

L(t, TM 1 ) = L(0, TM 1 )E t ((, TM 1 ) W TM )

Last Modified: 14/10/2014 08:19

546 / 561

LMM: Direct Approach


Or plugging in the definition of L(0, TM 1 ) we get

1 P(0, TM 1 )
1 E t ((, TM 1 ) W TM )
L(t, TM 1 ) =
P(0, TM )
Now we assume that

Dale Roberts (c) 2011-2014

TM 1 ,TM (t) =

L(t, TM 1 )
(t, TM 1 ),
L(t, TM 1 ) + 1

Last Modified: 14/10/2014 08:19

t [0, TM 1 ]

547 / 561

LMM: Direct Approach


Recall that we changed from Q to the forward measure P T and obtained
Z T
T
f(t) (t, T ) d t, (t, T ) :=
dW (t) = d W
(t, u) du
t

in the HJM setup

Dale Roberts (c) 2011-2014

d f (t, T ) = (t, T )d t + (t, T )dW (t),

Z T

under P

f(t) under Q
(t, u)du d t + (t, T )d W

d f (t, T ) = (t, T )

Last Modified: 14/10/2014 08:19

548 / 561

LMM: Direct Approach


Change from Q to P TM gives new P TM Brownian motion
f(t) (t, TM )d t
dW TM (t) = d W
Change from Q to P TM 1 gives new P TM 1 Brownian motion
f(t) (t, TM 1 )d t
dW TM 1 (t) = d W
Equate both expression to get transformation from P TM to P TM 1 :

Dale Roberts (c) 2011-2014

dW TM 1 (t) = dW TM (t) ((t, TM 1 ) (t, TM ))d t


= dW TM (t) TM 1 ,TM (t)d t

Last Modified: 14/10/2014 08:19

549 / 561

LMM: Direct Approach


Again, we postulate
d L(t, TM 2 ) = L(t, TM 2 )(t, TM 2 )dW TM 1 (t)

1 P(0, TM 2 )
1
L(0, TM 2 ) =
P(0, TM 1 )
Define
TM 2 ,TM 1 (t) =

L(t, TM 2 )
(t, TM 2 )
L(t, TM 2 ) + 1

And so forth... to obtain a family of M processes

Dale Roberts (c) 2011-2014

L(t, T0 ), . . . , L(t, TM 1 ),

Last Modified: 14/10/2014 08:19

t [0, TM ]

550 / 561

LIBOR Dynamics Under Different Measures


Lemma
Let 0 m, n M 1. Then the dynamics of L(t, Tm ) under P Tn+1 is for m < n
given by

n
X
d L(t, Tm ) = L(t, Tm ) (t, Tm )
Tr ,Tr+1 (t) T d t + (t, Tm )dW Tn+1
r=m+1

For m = n,

d L(t, Tm ) = L(t, Tm )(t, Tm )dW Tn+1 (t)

and for m > n we have

d L(t, Tm ) = L(t, Tm ) (t, Tm )

m
X

Tr ,Tr+1 (t) T d t + (t, Tm )dW Tn+1

r=n+1

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

551 / 561

Bond Price Dynamics

We have constructed a market model for LIBOR rates


However, to price bonds and options we need to determine the bond price
dynamics
So now well determine the bond price dynamics and then derive a caplet
pricing formula

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

552 / 561

Bond Price Dynamics


For m = 1, . . . , M we can define the forward price process
P(t, Tm1 )
= L(t, Tm1 ) + 1,
P(t, Tm )
Since

Dale Roberts (c) 2011-2014

t [0, Tm1 ]

P(t, Tm1 )
d
= d L(t, Tm1 )
P(t, Tm )
= L(t, Tm1 )(t, Tm1 )dW Tm (t)
P(t, Tm1 )
Tm1 ,Tm (t)dW Tm (t)
=
P(t, Tm )

Last Modified: 14/10/2014 08:19

553 / 561

Bond Price Dynamics

Therefore, we get that



P(t, Tm1 )
P(0, Tm1 )
=
E t Tm1 ,Tm W Tm ,
P(t, Tm )
P(0, Tm )

t [0, Tm1 ]

which is a P Tm martingale
This is sometimes call the Tm -forward price process of the Tm1 -bond

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

554 / 561

Bond Price Dynamics


We can extend this concept and define the Tm -forward price processes for
all Tk -bonds
When k < m we define
m1

P(t, Tk ) Y P(t, Tr )
=
P(t, Tm ) r=k P(t, Tr+1 )
=

P(t, Tm1 )
P(t, Tk )

P(t, Tk+1 )
P(t, Tm )

For k > m, we define

Dale Roberts (c) 2011-2014

P(t, Tk )
=
P(t, Tm )

P(t, Tm )
P(t, Tm+1 )

Last Modified: 14/10/2014 08:19

P(t, Tk1 )

P(t, Tk )

555 / 561

Bond Price Dynamics


Lemma
For every 1 k 6= m M , the forward price process satisfies

P(t, Tk )
P(0, Tk )
=
E t Tk ,Tm W Tm ,
P(t, Tm )
P(0, Tm )

for the Rd -valued bounded progressive process


Pm1
r=k Tr ,Tr+1 ,
P
Tk ,Tm =
k1
r=m Tr ,Tr+1 ,
Hence,

Dale Roberts (c) 2011-2014

P(t,Tk )
P(t,Tm )

t [0, Tk Tm ]

k<m
k>m

for t [0, Tk Tm ] is a positive P Tm -martingale.

Last Modified: 14/10/2014 08:19

556 / 561

Bond Price Dynamics

We can now derive the nominal Tn -bond prices


n
Y
P(Tm , Tk )
1
=
P(Tm , Tn ) =
P(Tm , Tk1 ) k=m+1 L(Tm , Tk1 ) + 1
k=m+1
n
Y

at dates t = Tm
We have derived an arbitrage-free market model for the bonds with
maturities T0 , . . . , TM since we have shown that P Tm is a martingale
measure for the Tm -bond as numeraire

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

557 / 561

LMM: Pricing Claims

Any Tm -contingent claim X with EPTm [|X |] < can now be priced at time
t Tm in terms of the Tm -bond numeraire via
V (t) = P(t, Tm )EPTm [X |F t ]
We can also express this price relative to any future Tn -bond

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

558 / 561

LMM: Pricing Claims

Lemma
The Tm -bond discounted Tm -claim price satisfies


P(t, Tn )
V (t)
X

=
EPTn
F t
P(t, Tm )
P(t, Tm )
P(Tm , Tn )
for all m < n M .

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

559 / 561

LMM: Pricing Claims


Corollary
Let m + 1 < n M . The time Tm price of the nth caplet with reset date Tn1 ,
settlement date Tn and strike rate is
Cpl(Tm ;Tn1 ,Tn ) = P(Tm , Tn )[L(Tm , Tn1 )(d1 (n; Tm )) (d2 (n; Tm ))]

where
d1,2 (n; Tm ) =

log

L(Tm ,Tn1 )

rR

Tn1
Tm

Dale Roberts (c) 2011-2014

12

R Tn1
Tm

k(s, Tn1 )k2 ds

k(s, Tn1 )k2 ds

Last Modified: 14/10/2014 08:19

560 / 561

LMM: Pricing Claims

This is Blacks formula for the caplet price with


Z Tn1
1
2
(Tm ) =
k(s, Tn1 )k2 ds
Tn1 Tm T

Dale Roberts (c) 2011-2014

Last Modified: 14/10/2014 08:19

561 / 561

You might also like