IEOR E4707: Financial Engineering: Continuous-Time Models Fall 2010

c 2010 by Martin Haugh

Introduction to Stochastic Calculus
These notes provide an introduction to stochastic calculus, the branch of mathematics that is most identied
with nancial engineering and mathematical nance. We will ignore most of the technical details and take an
engineering approach to the subject. We will cover more material than is strictly necessary for this course.
Any material that is not required, however, should be of value for other courses such as Term Structure Models.
We make the following assumptions throughout.
There is a probability triple (, F, P) where
P is the true or physical probability measure
is the universe of possible outcomes. We use to represent a generic outcome, typically a
sample path(s) of a stochastic process(es).
the set
1
F represents the set of possible events where an event is a subset of .
There is also a ltration, {F
t
}
t0
, that models the evolution of information through time. So for example,
if it is known by time t whether or not an event, E, has occurred, then we have E F
t
. If we are working
with a nite horizon, [0, T], then we can take F = F
T
.
We also say that a stochastic process, X
t
, is F
t
-adapted if the value of X
t
is known at time t when the
information represented by F
t
is known. All the processes we consider will be F
t
-adapted so we will not
bother to state this in the sequel.
In the continuous-time models that we will study, it will be understood that the ltration {F
t
}
t0
will be
the ltration generated by the stochastic processes (usually a Brownian motion, W
t
) that are specied in
the model description.
Since these notes regularly refer to numeraires and equivalent martingale measures (EMMs), readers
should be familiar with these concepts in advance. My Martingale Pricing Theory lecture note, for
example, covers these topics in a discrete-time, discrete-space framework. In Section 12 of these notes we
will discuss martingale pricing theory in the continuous-time setting and state without proof
2
the two
Fundamental Theorems of Asset pricing.
1 Martingales and Brownian Motion
Denition 1 A stochastic process, {W
t
: 0 t }, is a standard Brownian motion if
1. W
0
= 0
2. It has continuous sample paths
3. It has independent, normally-distributed increments.
1
Technically, F is a -algebra.
2
These Theorems, however, are stated and proven in the discrete-time, discrete-space setting of the Martingale Pricing
Theory lecture notes.
Introduction to Stochastic Calculus 2
Denition 2 An n-dimensional process, W
t
= (W
(1)
t
, . . . , W
(n)
t
), is a standard n-dimensional Brownian
motion if each W
(i)
t
is a standard Brownian motion and the W
(i)
t
s are independent of each other.
Denition 3 A stochastic process, {X
t
: 0 t }, is a martingale with respect to the ltration, F
t
, and
probability measure, P, if
1. E
P
[|X
t
|] < for all t 0
2. E
P
[X
t+s
|F
t
] = X
t
for all t, s 0.
Example 1 (Brownian martingales)
Let W
t
be a Brownian motion. Then W
t
, W
2
t
t and exp
_
W
t

2
t/2
_
are all martingales.
The latter martingale is an example of an exponential martingale. Exponential martingales are of particular
signicance since they are positive and may be used to dene new probability measures.
Exercise 1 (Conditional expectations as martingales) Let Z be a random variable and set
X
t
:= E[Z|F
t
]. Show that X
t
is a martingale.
2 Quadratic Variation
Consider a partition of the time interval, [0, T] given by
0 = t
0
< t
1
< t
2
< . . . < t
n
= T.
Let X
t
be a Brownian motion and consider the sum of squared changes
Q
n
(T) :=
n

i=1
[X
t
i
]
2
(1)
where X
t
i
:= X
t
i
X
t
i1
.
Denition 4 (Quadratic Variation) The quadratic variation of a stochastic process, X
t
, is equal to the
limit of Q
n
(T) as t := max
i
(t
i
t
i1
) 0.
Theorem 1 The quadratic variation of a Brownian motion is equal to T with probability 1.
The functions with which you are normally familiar, e.g. continuous dierentiable functions, have quadratic
variation equal to zero. Note that any continuous stochastic process or function
3
that has non-zero quadratic
variation must have innite total variation where the total variation of a process, X
t
, on [0, T] is dened as
Total Variation := lim
t0
n

i=1
|X
t
k
X
t
k1
|.
This follows by observing that
n

i=1
(X
t
k
X
t
k1
)
2

_
n

i=1
|X
t
k
X
t
k1
|
_
max
1kn
|X
t
k
X
t
k1
|. (2)
If we now let n in (2) then the continuity of X
t
implies the impossibility of the process having nite total
variation and non-zero quadratic variation. Theorem 1 therefore implies that the total variation of a Brownian
motion is innite. We have the following important result which will prove very useful when we price options
when there are multiple underlying Brownian motions, as is the case with quanto options for example.
3
A sample path of a stochastic process can be viewed as a function.
Introduction to Stochastic Calculus 3
Theorem 2 (Levys Theorem) A continuous martingale is a Brownian motion if and only if its quadratic
variation over each interval [0, t] is equal to t.
Another interesting result is the following:
Theorem 3 Any non-constant continuous martingale must have innite total variation.
We know from our discrete-time models that any arbitrage-free model must have an equivalent martingale
measure. The same is true in continuous-time models. Theorem 3 then implies that if we work in
continuous-time with continuous (deated) price processes, then these processes must have innite total
variation.
3 Stochastic Integrals
We now discuss the concept of a stochastic integral, ignoring the various technical conditions that are required
to make our denitions rigorous. In this section, we write X
t
() instead of the usual X
t
to emphasize that the
quantities in question are stochastic.
Denition 5 A stopping time of the ltration F
t
is a random time, , such that the event { t} F
t
for all
t > 0.
In non-mathematical terms, we see that a stopping time is a random time whose value is part of the information
accumulated by that time.
Denition 6 We say a process, h
t
(), is elementary if it is piece-wise constant so that there exists a sequence
of stopping times 0 = t
0
< t
1
< . . . < t
n
= T and a set of F
t
i
-measurable
4
functions, e
i
(), such that
h
t
() =

i
e
i
()I
[t
i
,t
i+1
)
(t)
where I
[t
i
,t
i+1
)
(t) = 1 if t [t
i
, t
i+1
) and 0 otherwise.
Denition 7 The stochastic integral of an elementary function, h
t
(), with respect to a Brownian motion,
W
t
, is dened as
_
T
0
h
t
() dW
t
() :=
n1

i=0
e
i
()
_
W
t
i+1
() W
t
i
()
_
. (3)
Note that our denition of an elementary function assumes that the function, h
t
(), is evaluated at the
left-hand point of the interval in which t falls. This is a key component in the denition of the stochastic
integral: without it the results below would no longer hold. Moreover, dening the stochastic integral in this
way makes the resulting theory suitable for nancial applications. In particular, if we interpret h
t
() as a trading
strategy and the stochastic integral as the gains or losses from this trading strategy, then evaluating h
t
() at
the left-hand point is equivalent to imposing the non-anticipativity of the trading strategy, a property that we
always wish to impose.
For a more general process, X
t
(), we have
_
T
0
X
t
() dW
t
() := lim
n
_
T
0
X
(n)
t
() dW
t
()
where X
(n)
t
is a sequence of elementary processes that converges (in an appropriate manner) to X
t
.
4
A function f() is F
t
measurable if its value is known by time t.
Introduction to Stochastic Calculus 4
Example 2 We want to compute
_
T
0
W
t
dW
t
. Towards this end, let 0 = t
n
0
< t
n
1
< t
n
2
< . . . < t
n
n
= T be a
partition of [0, T] and dene
X
n
t
:=
n1

i=0
W
t
n
i
I
[t
n
i
,t
n
i+1
)
(t)
where I
[t
n
i
,t
n
i+1
)
(t) = 1 if t [t
n
i
, t
n
i+1
) and is 0 otherwise. Then X
n
t
is an adapted elementary process and, by
continuity of Brownian motion, satises lim
n
X
n
t
= W
t
almost surely as max
i
|t
n
i+1
t
n
i
| 0. The Ito
integral of X
n
t
is given by
_
T
0
X
n
t
dW
t
=
n1

i=0
W
t
n
i
(W
t
n
i+1
W
t
n
i
)
=
1
2
n1

i=0
_
W
2
t
n
i+1
W
2
t
n
i
(W
t
n
i+1
W
t
n
i
)
2
_
=
1
2
W
2
T

1
2
W
2
0

1
2
n1

i=0
(W
t
n
i+1
W
t
n
i
)
2
. (4)
By the denition of quadratic variation the sum on the right-hand-side of (4) converges in probability to T. And
since W
0
= 0 we obtain
_
T
0
W
t
dW
t
= lim
n
_
T
0
X
n
t
dW
t
=
1
2
W
2
T

1
2
T.
Note that we will generally evaluate stochastic integrals using Itos Lemma (to be discussed later) without
having to take limits of elementary processes as we did in Example 2.
Denition 8 We dene the space L
2
[0, T] to be the space of processes, X
t
(), such that
E
_
_
T
0
X
t
()
2
dt
_
< .
Theorem 4 (Itos Isometry) For any X
t
() L
2
[0, T] we have
E
_
_
_
_
T
0
X
t
() dW
t
()
_
2
_
_
= E
_
_
T
0
X
t
()
2
dt
_
.
Proof: (For the case where X
t
is an elementary process)
Let X
t
=

i
e
i
()I
[t
i
,t
i+1
)
(t) be an elementary process where the e
i
()s and t
i
s are as dened in Denition 6.
We therefore have
_
T
0
X
t
() dW
t
() :=

n1
i=0
e
i
()
_
W
t
i+1
() W
t
i
()
_
We then have
E
_
_
_
_
T
0
X
t
() dW
t
()
_
2
_
_
= E
_
_
_
n1

i=0
e
i
()
_
W
t
i+1
() W
t
i
()
_
_
2
_
_
=
n1

i=0
E
_
e
2
i
()
_
W
t
i+1
() W
t
i
()
_
2
_
+ 2
n1

0i<jn1
E
_
e
i
e
j
()
_
W
t
i+1
() W
t
i
()
_ _
W
t
j+1
() W
t
j
()
_
Introduction to Stochastic Calculus 5
=
n1

i=0
E
_

_
e
2
i
() E
t
i
_
_
W
t
i+1
() W
t
i
()
_
2
_
. .
= t
i+1
t
i
_

_
+ 2
n1

0i<jn1
E
_

_e
i
e
j
()
_
W
t
i+1
() W
t
i
()
_
E
t
j
__
W
t
j+1
() W
t
j
()
_
. .
=0
_

_
= E
_
n1

i=0
e
2
i
()(t
i+1
t
i
)
_
= E
_
_
T
0
X
t
()
2
dt
_
which is what we had to show.
Theorem 5 (Martingale Property of Stochastic Integrals) The stochastic integral,
Y
t
:=
_
t
0
X
s
() dW
s
(), is a martingale for any X
t
() L
2
[0, T].
Exercise 2 Check that
_
t
0
X
s
() dW
t
() is indeed a martingale when X
t
is an elementary process. (Hint:
Follow the steps we took in our proof of Theorem 4.)
4 Stochastic Dierential Equations
Denition 9 An n-dimensional Ito process, X
t
, is a process that can be represented as
X
t
= X
0
+
_
t
0
a
s
ds +
_
t
0
b
s
dW
s
(5)
where W is an m-dimensional standard Brownian motion, and a and b are n-dimensional and n m-dimensional
F
t
-adapted
5
processes, respectively
6
.
We often use the notation
dX
t
= a
t
dt +b
t
dW
t
as shorthand for (5). An n-dimensional stochastic dierential equation (SDE) has the form
dX
t
= a(X
t
, t) dt +b(X
t
, t) dW
t
; X
0
= x (6)
where as before, W
t
is an m-dimensional standard Brownian motion, and a and b are n-dimensional and
n m-dimensional adapted processes, respectively. Once again, (6) is shorthand for
X
t
= x +
_
t
0
a(X
s
, s) dt +
_
t
0
b(X
s
, t) dW
s
. (7)
While we do not discuss the issue here, various conditions exist to guarantee existence and uniqueness of
solutions to (7). A useful tool for solving SDEs is Itos Lemma which we now discuss.
5
a
t
and b
t
are F
t
-adapted if a
t
and b
t
are F
t
-measurable for all t. We always assume that our processes are F
t
-adapted.
6
Additional technical conditions on a
t
and b
t
are also necessary.
Introduction to Stochastic Calculus 6
5 Itos Lemma
Itos Lemma is the most important result in stochastic calculus, the sine qua non of the eld. We rst state
and give an outline proof of a basic form of the result.
Theorem 6 (Itos Lemma for 1-dimensional Brownian Motion)
Let W
t
be a Brownian motion on [0, T] and suppose f(x) is a twice continuously dierentiable function on R.
Then for any t T we have
f(W
t
) = f(0) +
1
2
_
t
0
f

(W
s
) ds +
_
t
0
f

(W
s
) dW
s
. (8)
Proof: (Sketch) Let 0 = t
0
< t
1
< t
2
< . . . < t
n
= t be a partition of [0, t]. Clearly
f(W
t
) = f(0) +
n1

i=0
_
f(W
t
i+1
) f(W
t
i
)
_
. (9)
Taylors Theorem implies
f(W
t
i+1
) f(W
t
i
) = f

(W
t
i
)(W
t
i+1
W
t
i
) +
1
2
f

(
i
)(W
t
i+1
W
t
i
)
2
(10)
for some
i
(W
t
i
, W
t
i+1
). Substituting (10) into (9) we obtain
f(W
t
) = f(0) +
n1

i=0
f

(W
t
i
)(W
t
i+1
W
t
i
) +
1
2
n1

i=0
f

(
i
)(W
t
i+1
W
t
i
)
2
. (11)
If we let := max
i
|t
i+1
t
i
| 0 then it can be shown that the terms on the right-hand-side of (11) converge
to the corresponding terms on the right-hand-side of (8) as desired. (This should not be surprising as we know
the quadratic variation of Brownian motion on [0, t] is equal to t.)
A more general version of Itos Lemma can be stated for Ito processes.
Theorem 7 (Itos Lemma for 1-dimensional Ito process)
Let X
t
be a 1-dimensional Ito process satisfying the SDE
dX
t
=
t
dt +
t
dW
t
.
If f(t, x) : [0, ) R R is a C
1,2
function and Z
t
:= f(t, X
t
) then
dZ
t
=
f
t
(t, X
t
) dt +
f
x
(t, X
t
) dX
t
+
1
2

2
f
x
2
(t, X
t
) (dX
t
)
2
=
_
f
t
(t, X
t
) +
f
x
(t, X
t
)
t
+
1
2

2
f
x
2
(t, X
t
)
2
t
_
dt +
f
x
(t, X
t
)
t
dW
t
The Box Calculus
In the statement of Itos Lemma, we implicitly assumed that (dX
t
)
2
=
s
t
dt. The box calculus is a series of
simple rules for calculating such quantities. In particular, we use the rules
dt dt = dt dW
t
= 0 and
dW
t
dW
t
= dt
Introduction to Stochastic Calculus 7
when determining quantities such as (dX
t
)
2
in the statement of Itos Lemma above. Note that these rules are
consistent with Theorem 1. When we have two correlated Brownian motions, W
(1)
t
and W
(2)
t
, with correlation
coecient,
t
, then we easily obtain that dW
(1)
t
dW
(2)
t
=
t
dt. We use the box calculus for computing the
quadratic variation of Ito processes.
Exercise 3 Let W
(1)
t
and W
(2)
t
be two independent Brownian motions. Use Levys Theorem to show that
W
t
:= W
(1)
t
+
_
1
2
W
(2)
t
is also a Brownian motion for a given constant, .
Example 3
Suppose a stock price, S
t
, satises the SDE
dS
t
=
t
S
t
dt +
t
S
t
dW
t
.
Then we can use the substitution, Y
t
= log(S
t
) and Itos Lemma applied to the function
7
f(x) := log(x) to
obtain
S
t
= S
0
exp
__
t
0
(
s

2
s
/2) ds +
_
t
0

s
dW
s
_
. (12)
Note that S
t
does not appear on the right-hand-side of (12) so that we have indeed solved the SDE. When

s
= and
s
are constants we obtain
S
t
= S
0
exp
_
(
2
/2) t + dW
t
_
(13)
so that log(S
t
) N
_
(
2
/2)t,
2
t
_
.
Example 4 (Ornstein-Uhlenbeck Process)
Let S
t
be a security price and suppose X
t
= log(S
t
) satises the SDE
dX
t
= [(X
t
t) +] dt + dW
t
.
Then we can apply Itos Lemma to Y
t
:= exp(t)X
t
to obtain
dY
t
= exp(t) dX
t
+ X
t
d (exp(t))
= exp(t) ([(X
t
t) +] dt + dW
t
) + X
t
exp(t) dt
= exp(t) ([t +] dt + dW
t
)
so that
Y
t
= Y
0
+
_
t
0
e
s
(s + 1) ds +
_
t
0
e
s
dW
s
(14)
or alternatively (after simplifying the Riemann integral in (14))
X
t
= X
0
e
t
+t +e
t
_
t
0
e
s
dW
s
. (15)
Once again, note that X
t
does not appear on the right-hand-side of (15) so that we have indeed solved the
SDE. We also obtain E[X
t
] = X
0
e
t
+ t and
Var(X
t
) = Var
_
e
t
_
t
0
e
s
dW
s
_
=
2
e
2t
E
_
__
t
0
e
s
dW
s
_
2
_
=
2
e
2t
_
t
0
e
2s
ds (by Itos Isometry) (16)
=

2
2
_
1 e
2t
_
. (17)
7
Note that f() is not a function of t.
Introduction to Stochastic Calculus 8
These moments should be compared with the corresponding moments for log(S
t
) in the previous example.
Theorem 8 (Itos Lemma for n-dimensional Ito process) Let X
t
be an n-dimensional Ito process
satisfying the SDE
dX
t
=
t
dt +
t
dW
t
.
where X
t
R
n
,
t
R
n
,
t
R
nm
and W
t
is a standard m-dimensional Brownian motion. If
f(t, x) : [0, ) R
n
R is a C
1,2
function and Z
t
:= f(t, X
t
) then
dZ
t
=
f
t
(t, X
t
) dt +

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

i,j

2
f
x
i
x
j
(t, X
t
) dX
(i)
t
dX
(j)
t
where dW
(i)
t
dW
(j)
t
= dt dW
(i)
t
= 0 for i = j and dW
(i)
t
dW
(i)
t
= dt.
Exercise 4 Let X
t
and Y
t
satisfy
dX
t
=
(1)
t
dt +
(1,1)
t
dW
(1)
t
dY
t
=
(2)
t
dt +
(2,1)
t
dW
(1)
t
+
(2,2)
t
dW
(2)
t
and dene Z
t
:= X
t
Y
t
. Apply the multi-dimensional version of Itos Lemma to nd the SDE satised by Z
t
.
6 The Martingale Representation Theorem
The martingale representation theorem is an important result that is particularly useful for constructing
replicating portfolios in complete nancial models.
Theorem 9 Suppose M
t
is an F
t
-martingale where {F
t
}
t0
is the ltration generated by the n-dimensional
standard Brownian motion, W
t
= (W
(1)
t
, . . . , W
(n)
t
). If E[M
2
t
] < for all t then there exists a unique
8
n-dimensional adapted stochastic process,
t
, such that
M
t
= M
0
+
_
t
0

T
s
dW
t
for all t 0
where
T
s
denotes the transpose of the vector,
s
.
Example 5 Let F = W
3
T
and dene M
t
= E
t
[F]. We will show that
M
t
= 3
_
t
0
(T s +W
2
s
) dW
s
(18)
which is consistent with the Martingale Representation theorem. First we must calculate M
t
. We do this using
the independent increments property of Brownian motion and obtain
M
t
= E
t
[W
3
T
] = E
t
[(W
T
W
t
+W
t
)
3
]
= E
t
[(W
T
W
t
)
3
]
. .
= 0
+ E
t
[W
3
t
] + 3E
t
[W
t
(W
T
W
t
)
2
] + 3 E
t
[W
2
t
(W
T
W
t
)]
. .
= 0
= W
3
t
+ 3W
t
(T t). (19)
8
To be precise, additional integrability conditions are required of
s
in order to claim that it is unique.
Introduction to Stochastic Calculus 9
We can now apply Itos Lemma to (19) to obtain
dM
t
= 3W
2
t
dW
t
+
1
2
6W
t
dt + 3(T t) dW
t
3W
t
dt
= 3(W
2
t
+T t) dW
t
which, noting M
0
= 0, is (18).
7 Gaussian Processes
Denition 10 A process X
t
, t 0, is a Gaussian process if (X
t
1
, . . . , X
t
n
) is jointly normally distributed for
every n and every set of times 0 t
1
t
2
. . . t
n
.
If X
t
is a Gaussian process, then it is determined by its mean function, m(t), and its covariance function,
(s, t), where
m(t) = E[X
t
]
(s, t) = E[(X
s
m(s))(X
t
m(t))] .
In particular, the joint moment generating function (MGF) of (X
t
1
, . . . , X
t
n
) is given by
M
t
1
,...,t
n
(
1
, . . . ,
n
) = exp
_

T
m(t) +
1
2

T

_
(20)
where m(t) = (m(t
1
) . . . m(t
n
))
T
and
i,j
= (t
i
, t
j
).
Example 6 (Brownian motion)
Brownian motion is a Gaussian process with m(t) = 0 and (s, t) = min(s, t) for all s, t 0.
Theorem 10 (Integration of a deterministic function w.r.t. a Brownian motion) Let W
t
be a
Brownian motion and suppose
X
t
=
_
t
0

s
dW
s
where
s
is a deterministic function. Then X
t
is a Gaussian process with m(t) = 0 and (s, t) =
_
min(s,t)
0

2
s
ds.
Proof: (Sketch)
(i) First use Itos Lemma to show that
E
_
e
uX
t

= 1 +
1
2
u
2
_
t
0

2
s
E
_
e
uX
s

ds. (21)
If we set y
t
:= E
_
e
uX
t

then we can dierentiate across (21) to obtain the ODE

dy
dt
=
1
2
u
2

2
t
y.
This is easily solved to obtain the MGF for X
t
,
E
_
e
uX
t

= exp
_
1
2
u
2
_
t
0

2
s
ds
_
(22)
Introduction to Stochastic Calculus 10
which, as expected, is the MGF of a normal random variable with mean 0 and variance
_
t
0

2
s
ds.
(ii) We now use (22) and similar computations to show that the joint MGF of (X
t
1
, . . . , X
t
n
) has the form
given in (20) with m(t) = 0 and (s, t) =
_
min(s,t)
0

2
s
ds. (See Shreves Stochastic Calculus for Finance II for
further details.)
The next theorem again concerns Gaussian processes and is often of interest
9
when studying short-rate term
structure models.
Theorem 11 Let W
t
be a Brownian motion and suppose
t
and
t
are deterministic functions. If
X
t
:=
_
t
0

u
dW
u
and Y
t
:=
_
t
0

u
X
u
du
then Y
t
is a Gaussian process with m(t) = 0 and
(s, t) =
_
min(s,t)
0

2
v
__
s
v

y
dy
___
t
v

y
dy
_
dv.
Proof: The proof is tedious but straightforward. (Again, see Shreves Stochastic Calculus for Finance II for
further details.)
Note for example that Brownian motion with drift and the Ornstein-Uhlenbeck process are both Gaussian
processes. Nonetheless, we saw in Example 4 that these two processes behave very dierently: the
Ornstein-Uhlenbeck process mean-reverts and its variance tends to a nite limit as T . This is not true of
Brownain motion with or without drift.
8 The Feynman-Kac Formula
Suppose X
t
is a stochastic process satisfying the SDE dX
t
= (t, X
t
) dt +(t, X
t
) dW
t
. Now consider the
function, f(x, t), given by
f(t, x) = E
x
t
_
_
T
t

(t)
s
h(X
s
, s) ds +
(t)
T
g(X
T
)
_
where

(t)
s
= exp
_

_
s
t
r(X
u
, u) du
_
and the notation E
x
t
[] implies that the expectation should be taken conditional on time t information with
X
t
= x. Note that f(x, t) may be interpreted as the time t price of a security that pays dividends at a
continuous rate, h(X
s
, s) for s t, and with a terminal payo g(X
T
) at time T. Of course E[] should then be
interpreted as an expectation under an equivalent martingale measure with the cash account as the
corresponding numeraire and r(, ) as the instantaneous risk-free rate.
The Feynman-Kac Theorem
10
states that f(, ) satises the following PDE
f
t
(t, X
t
) +
f
x
(t, X
t
)
t
(t, X
t
) +
1
2

2
f
x
2
(t, X
t
)
2
t
(t, X
t
) r(x, t)f(x, t) +h(x, t) = 0, (x, t) R[0, T)
f(x, T) = g(x), x R
9
See, for example, Hull and Whites one-factor model.
10
Additional technical conditions on , , r, h, g and f are required.
Introduction to Stochastic Calculus 11
Proof: (Sketch proof )
We will use the martingale property of conditional expectations which states that if M
t
:= E
t
[F] where F is a
given xed random variable then M
t
is a martingale. First let
f(t, X
t
) := E
t
_
_
T
t

(t)
s
h(X
s
, s) ds +
(t)
T
g(X
T
)
_
(23)
and note that the random variable inside the expectation on the right-hand-side of (23) is a function of t. This
means we cannot yet apply the martingale property of conditional expectations. However by adding
_
t
0

(t)
s
h(X
s
, s) ds to both sides of (23) and then multiplying both sides by
(0)
t
we obtain

(0)
t
_
f(t, X
t
) +
_
t
0

(t)
s
h(X
s
, s) ds
_
= E
t
_
_
T
0

(0)
s
h(X
s
, s) ds +
(0)
T
g(X
T
)
_
(24)
= E
t
[Z] , say.
Note that Z does not depend on t and so we can now apply the martingale property of conditional expectations
to conclude that the left-hand-side of (24) is a martingale. We therefore apply Itos Lemma to the left-hand-side
and set the coecient of dt to zero. This results in the Feynman-Kac PDE given above.
Remark 1 The Feynman-Kac result generalizes easily to the case where X
t
is an n-dimensional Ito process
driven by an m-dimensional standard Brownian motion.
The Feynman-Kac Theorem plays an important role in nancial engineering as it enables us to compute the
prices of derivative securities (which can be expressed as expectations according to martingale pricing theory) by
solving a PDE instead. It is worth mentioning, however, that in the development of derivatives pricing theory,
the PDE approach preceded the martingale approach. Indeed by constructing a replicating
11
strategy for a given
derivative security, it can be shown directly that the price of the derivative security satises the Feynman-Kac
PDE.
9 Change of Probability Measure
The majority of nancial engineering models price securities using the EMM, Q, that corresponds to taking the
cash account, B
t
, as numeraire. Sometimes, however, it is particularly useful to work with another numeraire,
N
t
, and its corresponding EMM, P
N
say. We now describe how to create new probability measures and how to
switch back and forth between these measures.
Let Q be a given probability measure and M
t
a strictly positive Q-martingale such that E
Q
0
[M
t
] = 1 for all
t [0, T]. We may then dene a new equivalent probability measure, P
M
, by dening
P
M
(A) = E
Q
[M
T
1
A
] .
Note that
(i) P
M
() = 1
(ii) P
M
(A) 0 for every event A and
(iii) It can easily be shown that
P
M
_
_
i
A
i
_
=

i
P
M
(A
i
)
whenever the A
i
s form a sequence of disjoint sets.
11
We are implicitly assuming that the security can be replicated.
Introduction to Stochastic Calculus 12
Points (i), (ii) and (iii) above imply that P
M
is indeed a probability measure. Because the null-sets of Q and
P
M
coincide we can conclude that P
M
is indeed an equivalent probability measure (to Q). Expectations with
respect to P
M
then satisfy
E
P
M
0
[X] = E
Q
0
[M
T
X] . (25)
Exercise 5 Verify (25) in the case where X() =

n
i=1
c
i
I
{A
i
}
, M
T
is constant on each A
i
, and where
A
1
, . . . , A
n
form a partition
12
of .
When we dene a measure change this way, we use the notation dP
M
/dQ to refer to M
T
so that we often write
E
P
M
0
[X] = E
Q
0
_
dP
M
dQ
X
_
.
The following result explains how to switch between Q and P
M
when we are taking conditional expectations. In
particular, we have
E
P
M
t
[X] =
E
Q
t
_
dP
M
dQ
X
_
E
Q
t
_
dP
M
dQ
_ (26)
=
E
Q
t
_
dP
M
dQ
X
_
M
t
. (27)
Since M
t
is a Q-martingale (27) follows easily from (26). We shall not prove (26), however, as it is more
dicult and requires the measure-theoretic denition of a conditional expectation.
Exercise 6 Show that if X is F
t
-measurable, i.e. X is known by time t, then E
P
M
0
[X] = E
Q
0
[M
t
X].
Remark 2 Since M
T
is strictly positive we can set X = I
A
/M
T
in (25) where I
A
is the indicator function of
the event A. We then obtain E
P
M
0
[I
A
/M
T
] = E
Q
0
[I
A
] = Q(A). In particular, we see that dQ/dP
M
is given by
1/M
T
.
Remark 3 In the context of security pricing, we can take M
t
to be the deated time t price of a security with
strictly positive payo, normalized so that its expectation under Q is equal to 1. For example, let Z
T
t
be the
time t price of a zero-coupon bond maturing at time T, and let
13
B
t
denote the time t value of the cash
account. We could then set
14
M
T
:= 1/(B
T
Z
T
0
) so that
M
t
= E
Q
t
_
1
B
T
Z
T
0
_
is indeed a Q-martingale. The resulting measure, denoted by P
T
, is sometimes called the T-forward measure.
Note that we have implicitly assumed (why?!) that in this context, Q refers to the EMM when we take the cash
account as numeraire. We discuss P
T
in further detail in Section 11.
12
The A
i
s form a partition of if A
i
A
j
= whenever i = j and

i
A
i
= .
13
We assume the zero-coupon bond has face value $1 and B
0
= $1.
14
Note that it is not the case that M
t
= 1/(B
t
Z
t
0
) which is not a Q-martingale.
Introduction to Stochastic Calculus 13
10 Girsanovs Theorem
Girsanovs Theorem is one of the most important results for nancial engineering applications. When working
with models driven by Brownian motions
15
it enables us to (i) identify the equivalent martingale measure(s)
corresponding to a given numeraire and (ii) to move back and forth between dierent EMM-numeraire pairs.
Consider then the process
L
t
:= exp
_

_
t
0

s
dW
s

1
2
_
t
0

2
s
ds
_
(28)
where
s
is an adapted process and W
s
is a P-Brownian motion. Using Itos Lemma we can check that
dL
t
= L
t

t
dW
t
so L
t
is a positive martingale
16
with E
P
0
[L
t
] = 1 for all t.
Theorem 12 (Girsanovs Theorem) Dene an equivalent probability measure, Q

, by setting
Q

(A) := E
P
0
[L
T
1
A
]. (29)
Then

W
t
:= W
t
+
_
t
0

s
ds is a standard Q

-Brownian motion. Moreover,

W
t
has the martingale representation
property under Q

.
Remark 4 Suppose
s
= , a constant. Then since

W
t
:= W
t
+t is a standard Q

-Brownian motion, it
implies that W
t
=

W
t
t is a Q-Brownian motion with drift .
Example 7 Let dX
t
=
t
dt +
t
dW
t
and suppose we wish to nd a process,
s
, such that X
t
is a
Q

-martingale. This is easily achieved as follows:

dX
t
=
t
dt +
t
dW
t
=
t
dt +
t
(d

W
t

t
) dt
=
t
d

W
t
if we set
t
=
t
/
t
in which case X
t
is a Q

-martingale. We can obtain some intuition for this result: suppose

for example that
t
and
t
are both positive so that
t
is also positive. Then we can see from the denition of
L
t
in (28) that L
t
places less weight on paths where W
t
drifts upwards than it does on paths where it drifts
downwards. This relative weighting adjusts X
t
for the positive drift induced by
t
with the result that under
Q

, X
t
is a martingale.
Remark 5 Note that Girsanovs Theorem enables us to compute Q

-expectations directly without having to

switch back to the original measure, P.
We can get some additional intuition for the Girsanov Theorem by considering a random walk,
X = {X
0
= 0, X
1
, . . . , X
n
} with the interpretation that X
i
is the value of the walk at time iT/n. In particular,
X
n
corresponds to the value of the random walk at time T. We assume that X
i
:= X
i
X
i1
N(0, T/n)
under P and is independent of X
0
, . . . , X
i1
for i = 1, . . . n.
Suppose now that we want to compute := E
Q
0
[h(X)] where Q denotes the probability measure under which
X
i
N(, T/n), again independently of X
0
, . . . , X
i1
. In particular, if we set := T/n then X
t
approximates standard Brownian motion under P and X
t
approximates Brownian motion with drift under Q.
15
Versions of Girsanovs Theorem are also available for jump-diusion and other processes.
16
In fact we need
s
to have some additional properties before we can claim L
t
is a martingale. A sucient condition is
Novikovs Condition which requires E
P
0
_
exp
_
1
2
_
T
0

2
s
ds
__
< .
Introduction to Stochastic Calculus 14
If we let f() and g() denote the PDFs of N(, T/n) and N(0, T/n) random variables, respectively, then we
obtain
= E
Q
0
[h(X)] =
_
R
n
h(x
1
, . . . , x
n
)
_
n

i=1
f(x
i
)
_
dx
1
. . . dx
n
=
_
R
n
h(x
1
, . . . , x
n
)

i
_
f(x
i
)
g(x
i
)
g(x
i
)
_
dx
1
. . . dx
n
=
_
R
n
h(x
1
, . . . , x
n
)

i
_
f(x
i
)
g(x
i
)
_
_

i
g(x
i
)
_
dx
1
. . . dx
n
= E
P
0
_
h(X
1
, . . . , X
n
)

i
_
f(X
i
)
g(X
i
)
_
_
= E
P
0
_
h(X
1
, . . . , X
n
) exp
_

i
X
i

2
T
2
__
= E
P
0
_
h(X
1
, . . . , X
n
) exp
_
X
n

2
T
2
__
which is consistent with our statement of Girsanovs Theorem in (28) and (29) above. (See Remark 4.)
Remark 6 (i) As in the statement of the Girsanov Theorem itself, we could have chosen (and therefore )
to be adapted, i.e. to depend on prior events, in the random walk.
(ii) Note that Girsanovs Theorem allows the drift, but not the volatility of the Brownian motion, to change
under the new measure, Q

. It is interesting to see that we are not so constrained in the case of the random
walk. Have you any intuition for why this is so?
Multidimensional Girsanovs Theorem
The multidimensional version of Girsanovs Theorem is a straightforward generalization of the one-dimensional
version. In particular let W
t
be an n-dimensional standard P-Brownian motion and dene
L
t
:= exp
_

_
t
0

T
s
dW
s

1
2
_
t
0

T
s

s
ds
_
for t [0, T]. Then
17

W
t
:= W
t
+
_
t
0

s
ds is a standard Q

-Brownian motion where dQ

/dP = L
T
.
17
Again it is necessary to make some further assumptions in order to guarantee that L
t
is a martingale. Novikovs condition
is sucient.
Introduction to Stochastic Calculus 15
11 A Useful EMM: The Forward Measure
We now discuss the -forward measure, P

, which is the EMM that corresponds to taking the zero-coupon

bond with maturity as the numeraire. We therefore let Z

t
denote the time t price of a zero-coupon bond
maturing at time t and with face value $1. As usual, we let Q denote the EMM corresponding to taking the
cash account, B
t
, as numeraire. We assume without loss of generality that B
0
= $1 and now use Z

t
to dene
our new measure. To do this, set
dP

dQ
=
1
B

0
. (30)
Exercise 7 Check that (30) does indeed dene an equivalent probability measure.
Now let C
t
denote the time t price of a contingent claim that expires at time . We then have
C
t
= B
t
E
Q
t
_
C

_
(by martingale pricing with EMM Q) (31)
=
B
t
E
P

t
_
C

0
_
E
P

t
[B

0
]
(going from Q to P

using (26)) (32)

=
B
t
Z

0
E
P

t
[C

]
E
Q
t
[1] /E
Q
t
[1/(B

0
)]
(going from P

to Qin the denominator of (32) using (26))

= Z

t
E
P

t
[C

] (by martingale pricing with EMM Q). (33)

We can now nd C
t
, either through equation (33) or through equation (31) where we use the cash account as
numeraire. Computing C
t
through (31) is our usual method and is often very convenient. When pricing
equity derivatives, for example, we usually take interest rates, and hence the cash account, to be deterministic.
This means that the factor 1/B

in (31) can be taken outside the expectation so only the Q-distribution of C

is needed to compute C
t
.
When interest rates are stochastic we cannot take the factor 1/B

outside the expectation in (31) and we

therefore need to nd the joint Q-distribution of (B

, C

) in order to compute C
t
. On the other hand, if we
use equation (33) to compute C
t
, then we only need the P

-distribution of C

, regardless of whether or not

interest rates are stochastic. Working with a univariate-distribution is generally much easier than working with a
bivariate-distribution so if we can easily nd the P

-distribution of C

, then it can often be very advantageous

to work with this distribution. The forward measure is therefore particularly useful when studying term-structure
models.
Switching to a dierent numeraire can also be advantageous in other circumstances, even when interest rates
are deterministic. For example, we will nd it convenient to do so when pricing quanto-options and options on
multiple underlying securities. Moreover, we will see how we can use Girsanovs Theorem to move back and
forth between Q and other EMMs corresponding to other numeraires.
Introduction to Stochastic Calculus 16
12 Applying Stochastic Calculus: Martingale Pricing Theory
We now briey outline how the martingale pricing theory that we have seen in a discrete-time, discrete-space
framework translates to a continuous-time setting. We use S
t
to denote the time t price of a risky asset and B
t
to denote the time t value of the cash account. For ease of exposition we assume the risky asset does not
18
pay
dividends. Let
(s)
t
and
(b)
t
denote the number of units of the security and cash account, respectively, that is
held in a portfolio at time t. Then the value of the portfolio at time t is given by V
t
=
(s)
t
S
t
+
(b)
t
B
t
.
Denition 11 We say
t
:= (
(s)
t
,
(b)
t
) is self-nancing if
dV
t
=
(s)
t
dS
t
+
(b)
t
dB
t
.
Note that this denition is consistent with our denition for discrete-time models. It is also worth emphasizing
the mathematical content of this denition. Note that Itos Lemma implies
dV
t
=
(s)
t
dS
t
+
(b)
t
dB
t
+ S
t
d
(s)
t
+ B
t
d
(b)
t
+ dS
t
d
(s)
t
+ dB
t
d
(b)
t
. .
A
so that the self-nancing assumption amounts to assuming that the sum of the terms in A are identically zero.
Our denitions of arbitrage, numeraire securities, equivalent martingale measures and complete markets is
unchanged from the discrete-time setup. We now state
19
without proof the two fundamental theorems of asset
pricing. These results mirror those from the discrete-time theory.
Theorem 13 (The First Fundamental Theorem of Asset Pricing) There is no arbitrage if and only if
there exists an EMM, Q.
A consequence of Theorem 13 is that in the absence of arbitrage, the deated value process, V
t
/N
t
, of any
self-nancing trading strategy is a Q-martingale. This implies that the deated price of any attainable security
can be computed as the Q-expectation of the terminal deated value of the security.
Theorem 14 (The Second Fundamental Theorem of Asset Pricing) Assume there exists a security
with strictly positive price process and that there are no arbitrage opportunities. Then the market is complete if
and only if there exists exactly one risk-neutral martingale measure, Q.
Beyond their theoretical signicance, these theorems are also important in practice. If our model is complete
then we can price securities with the unique EMM that accompanies it, assuming of course the absence of
arbitrage. This is the case for the Black-Scholes model as well as local volatility models. If our model is
incomplete then we generally work directly under an EMM, Q, which is then calibrated to market data. In
particular, the true data-generating probability measure, P, is often completely ignored when working with
incomplete models. In some sense the issue of completeness then only arises when we discuss replicating or
hedging strategies.
Determining Replicating Strategies
The following example is particularly useful in many nancial engineering applications. While we focus on the
case of a single risky security the result generalizes easily to multiple risky securities. We assume the market is
complete of course, since only in that case can we guarantee that a replicating strategy actually exists.
18
We can easily adapt our denition of a self-nancing trading strategy to accommodate securities that pay dividends.
19
Additional technical conditions are generally required to actually prove these results.
Introduction to Stochastic Calculus 17
Example 8 (Wealth Dynamics and Hedging)
We know the value of the cash-account, B
t
, satises dB
t
= r
t
B
t
dt and suppose in addition that S
t
satises
dS
t
=
t
S
t
dt +
t
S
t
dW
t
. (34)
Then for a portfolio (
(s)
t
,
(b)
t
), the portfolio value, V
t
, at time t satises V
t
:=
(s)
t
S
t
+
(b)
t
B
t
. If

t
:= (
(s)
t
,
(b)
t
) is self-nancing, then we have
dV
t
=
(s)
t
dS
t
+
(b)
t
dB
t
=
(s)
t

t
S
t
dt +
(s)
t

t
S
t
dW
t
+
(b)
t
r
t
B
t
dt
= V
t
_

(s)
t
S
t
V
t

t
+

(b)
t
B
t
V
t
r
t
_
dt +

(s)
t
S
t
V
t

t
V
t
dW
t
= V
t
[r
t
+
t
(
t
r
t
)] dt +
t

t
V
t
dW
t
(35)
where
t
and (1
t
) are the fractions of time t wealth, V
t
, invested in the risky asset and cash account,
respectively, at time t.
Suppose now that V
t
is the time t value of some derivative security and that we want to determine the
self-nancing trading strategy, (
(s)
t
,
(b)
t
), that replicates it. By writing the dynamics of V
t
as in (35), we can
immediately determine
t
, the fraction of time t wealth that is invested in the risky asset. In particular,
t
is the
coecient of
t
V
t
dW
t
in the dynamics of V
t
. Note also that we can work under any probability measure we
choose since dierent probability measures will only change the drift of V
t
and not the volatility. Once we have

t
we immediately have 1
t
and then also (
(s)
t
,
(b)
t
).
Remark 7 If a security pays dividends or coupons, then in the above statements we should replace the security
price with the total gain process from holding the security. For example, if a stock pays a continuous dividend
yield of q, then Y
t
:= e
qt
S
t
is the total gain process. It is what your portfolio would be worth at time t if you
purchased one unit of the security at time t = 0 and reinvested all dividends immediately back into the stock.
The portfolio (
(s)
t
,
(b)
t
) is then said to be self-nancing if
dV
t
=
(s)
t
dY
t
+
(b)
t
dB
t
=
(s)
t
d(e
qt
S
t
) +
(b)
t
dB
t
=
(s)
t
_
e
qt
dS
t
+ qY
t
dt
_
+
(b)
t
dB
t
. (36)
We shall make use of (36) later in the course when we price options on dividend-paying securities.
The Role of Girsanovs Theorem
Suppose we again adopt the setting of Example 8 and that the risky security does not pay dividends. If we
deate as usual by the cash account then we can let Z
t
:= S
t
/B
t
denote the deated time t value of the risky
asset. Assume also that W
t
is a P-Brownian motion where P is the true data-generating probability measure.
Then it is easy to check using Itos Lemma that
dZ
t
= Z
t
(
t
r) dt + Z
t

t
dW
t
. (37)
But we know from the rst fundamental theorem that Z
t
must be a Q-martingale. Girsanovs Theorem,
however, implies that Z
t
is a Q-martingale only if

W
t
:= W
t
+
_
t
0

s
ds is a Q-Brownian motion with

s
:= (
s
r)/
s
. Indeed, because there is only one such process,
s
, this Q is unique
20
and the market is
therefore complete by the second fundamental theorem. We then obtain that the Q-dynamics for S
t
satisfy
dS
t
= rS
t
dt +
t
S
t
d

W
t
.
20
If, for example, there were two Brownian motions and only one risky security then there would be innitely many processes,

s
, that we could use to make Z
t
a martingale. In this case we could therefore conclude that the market was incomplete.
Introduction to Stochastic Calculus 18
Exercise 8 Suppose now that the risky asset pays a continuous dividend yield, q. Show that the Q-dynamics
of S
t
now satisfy
dS
t
= (r q)S
t
dt +
t
S
t
d

W
t
.
(Hint: Remember that it is now the deated total gains process that is now a Q-martingale.)
We can use these observations now to derive the Black-Scholes formula. Suppose the stock-price, S
t
, has
P-dynamics
dS
t
=
t
S
t
dt +S
t
dW
t
where W
t
is a P-Brownian motion. Note that we have now assumed a constant volatility, , as Black and
Scholes originally assumed. If the stock pays a continuous dividend yield of q then martingale pricing implies
that the price of a call option on the stock with maturity T and strike K is given by
C
0
= E
Q
0
_
e
rT
(S
T
K)
+

(38)
where log(S
T
) N
_
(r q
2
/2)T, T
_
. We can therefore compute the right-hand-side of (38) analytically
to obtain the Black-Scholes formula.
The following exercise is perhaps too dicult at this stage of the course. Later on, however, when we have had
plenty of practice with Girsanovs Theorem and changing probability measures, we will see how straightforward
such questions are.
Exercise 9 Consider an equity model with two securities, A and B, whose price processes, S
(a)
t
and S
(b)
t
respectively, satisfy the following SDEs
dS
(a)
t
= rS
(a)
t
dt +
1
S
(a)
t
dW
(1)
t
dS
(b)
t
= rS
(b)
t
dt +
2
S
(b)
t
_
dW
(1)
t
+
_
1
2
dW
(2)
t
_
where (W
(1)
t
, W
(2)
t
) is a 2-dimensional Q-standard Brownian motion. We assume the cash account is the
numeraire security corresponding to Q (which is consistent with the Q-dynamics of S
(a)
t
and S
(b)
t
) and that the
continuously compounded interest rate, r, is constant. Use the change of numeraire technique to compute the
time 0 price, C
0
of a European option that expires at time T with payo max (0, S
(a)
T
S
(b)
T
).