Martingales

1
Martingales are a key concept in probability and in mathematical …nance. The
term ‘martingale’ may refer to very di¤erent ideas e.g. a stochastic process
that has no drift. Essentially, this is the idea of a fair (random) game. We
encounter Martingales through three distinct, but closely connected ideas:

1. Martingales as a class of stochastic process;

2. Exponential martingales, which are a speci…c and extremely useful example

of a martingale;

3. Equivalent martingale measures, where we look for a probability measure

Q such that a given stochastic process S (t) is a martingale under Q re-
gardless of its nature under P. The correspondence between the measures
P and Q is done through a change of measure.

2
Discrete Time Martingales

A discrete time stochastic process fMt : t = 0; : : : ; T g such that Mt is Ft-

measurable for T = f0; : : : ; T g is a martingale if E jMtj < 1 and

E [Mt+1jFt] = Mt (1)

3
The …rst equation represents a standard integrability condition.

The second equation tells you that the expected value of M at time t + 1
conditional on all the information available up to time t is the value of M at
time t. In short, a Martingale is a driftless process.

If we take expectation on both sides of eqn. 1, then

E [Mt+1] = E [Mt]

This is due to the Tower Property of conditional expectations.

Martingales are a very nice mathematical object. They “get rid of the drift”
and enable us to focus on what probabilists consider is the most interesting
part: the statistical properties of purely random processes.

4
In addition, Doob and Meyer have developed a powerful theory centred around
martingales.

Continuous Time Martingales

Next, we generalize our de…nitions to continuous time: A continuous time

stochastic process
n o
Mt : t 2 R+
such that Mt is Ft-measurable for t 2 R+ is a martingale if

E jMtj < 1
and
E [MtjFs] = Ms; 0 s t:

5
Lévy’s Martingale Characterisation: Let Xt; t > 0 be a stochastic process
and let Ft be the …ltration generated by it. Xt is a Brownian motion i¤ the
following conditions are satis…ed:

1. X0 = 0 a.s.;

2. the sample paths t 7! Xt are continuous a.s.;

3. Xt is a martingale with respect to the …ltration Ft;

4. jXtj2 t is a martingale with respect to the …ltration Ft:

The Lévy characterization can be contrasted with the classical de…nition of a

Brownian motion as a stochastic process Xt satisfying:

6
 1. X0 = 0 a.s.;

2. the sample paths t 7! X (t) are continuous a.s.;

3. independent increments: for t1 < t2 < t3 < t4 the increments Xt4

Xt3 , Xt2 Xt1 are independent;

4. normally distributed increments: Xt Xs N (0; jt sj).

Lévy’s characterization neither mentions independent increments nor normally

distributed increments.

Instead, Lévy introduces two easily veri…able martingale conditions.

7
Itô Integrals and Martingales
Next, we explore the link between Itô integration and martingales.

Consider the stochastic process Y (t) = X 2(t). By Itô, we have

Z T
X 2(T ) = T + 2X (t)dX (t)
0
Taking the expectation, we get
"Z #
T
E[X 2(T )] = T + E 2X (t)dX (t)
0
Now, the quadratic variation property of Brownian motions implies that

E[X 2(T )] = T
and hence
"Z #
T
E 2X (t)dX (t) = 0
0

8
Therefore, the Itô integral
Z T
2X (t)dX (t)
0
is a martingale.

In fact, this property is shared by all Itô integrals.

The Itô integral is a martingale

Let g (t; Xt) be a function on [0; T ] and satisfying the technical condition.
Then the Itô integral
Z T
g (t; Xt)dXt
0
is a martingale.

So, Itô integrals are martingales.

9
But does the converse hold? Can we represent any martingale as an Itô inte-
gral?

The answer is yes!

Martingale Representation Theorem: If Mt is a martingale, then there

exists a function g (t; Xt) satisfying the technical condition such that
Z T
MT = M0 + g (t; Xt)dXt
0

Example Using only Itô and the fact that Itô integrals are martingales, we will
show that
h i
E X 2(T ) = T:

10
Consider the function F (t; Xt) = XT2 , then by Itô’s lemma,
Z Z
2 2 1 T T
XT = X0 + 2dt + 2XtdXt
2 0 0
Z T Z T
= dt + 2 XtdXt
0 0
since X0 = 0
Taking the expectation,
"Z # "Z #
h i T T
E XT2 = E dt + 2E XtdXt
0 0

Now,
Z T
XtdXt
0
hR i
is an Itô integral and as a result E T X dX =0
0 t t

11
Moreover,
"Z #
T
E dt = E [T ] = T
0
We can conclude that
h i
E X 2(T ) =T
As an aside, we can usually exchange the order of integration between the time
integral and the expectation so that
"Z # Z T
T
E f (Xt)dt = E [f (Xt)] dt
0 0
This is due to an analysis result known as Fubini’s Theorem.

12
Itô Integral
Recall the usual Riemann de…nition of a de…nite integral
Z b
f (x) dx
a

y = f (x )
h

a = x0 x1 x i −1 xi x i +1 b = xN
x

which represents the area under the curve between x = a and x = b; where
the curve is the graph of f (x) plotted against x:

13
Assuming f is a "well behaved" function on [a; b] ; there are many di¤erent
ways (which all lead to the same value for the de…nite integral).

Start by partitioning [a; b] into N intervals with end points x0 = a < x1 <
x2 < :::::: < xN 1 < xN = b; where the length of an interval dx =
xi xi+1 tends to zero as N ! 1: So there are N intervals and N + 1
points xi:

Discretising x gives
xi = a + idx
Now consider the de…nite integral
Z T
f (t) dt:
0
With Riemann integration there are a number of ways we can approximate this:

14
1. left hand rectangle rule;
Z T NX1
f (t) dt = lim f (ti) (ti+1 ti)
0 N !1
i=0

2. right hand rectangle rule;

Z T NX1
f (t) dt = lim f (ti+1) (ti+1 ti)
0 N !1
i=0

3. trapezium rule;
Z T NX1
f (t) dt = lim 1 (f (t ) + f (t
2 i i+1 )) (ti+1 ti)
0 N !1
i=0

15
 4. midpoint rule
Z T NX1
f (t) dt = lim f 12 (ti + ti+1) (ti+1 ti)
0 N !1
i=0

In the limit N ! 1; f (t) we get the same value for each de…nition of the
de…nite integral, provided the function is integrable.

Now consider the stochastic integral of the form

Z T Z T
f (t; X ) dX = f (t; X (t)) dX (t)
0 0
where X (t) is a Brownian motion. We can de…ne this integral as
NX1
lim f (ti; Xi) (Xi+1 Xi) ;
N !1
i=0

16
where Xi = X (ti) ; or as
NX1
lim f (ti+1; Xi+1) (Xi+1 Xi) ;
N !1
i=0
or as
NX1
lim f ti+ 1 ; Xi+ 1 (Xi+1 Xi) ;
N !1 2 2
i=0

where ti+ 1 = 12 (ti + ti+1) and Xi+ 1 = X ti+ 1 or in many other ways.
2 2 2
So clearly drawing parallels with the above Riemann form.

Very Important: In the case of a stochastic variable dX (t) the value of the
stochastic integral does depend on which de…nition we choose.

In the case of a stochastic integral, the de…nition

NX1
I = lim f (ti; Xi) (Xi+1 Xi) ;
N !1
i=0

17
is special. This de…nition results in the Itô Integral.

It is special because it is non-anticipatory; given that we are at time ti we

know Xi = X (ti) and therefore we know f (ti; Xi). The only uncertainty is
in the Xi+1 Xi term.

Compare this to a de…nition such as

NX1
lim f (ti+1; Xi+1) (Xi+1 Xi) ;
N !1
i=0
which is anticipatory; given that at time ti we know Xi but are uncertain
about the future value of Xi+1: Thus we are uncertain about both the value
of
f (ti+1; Xi+1)
and the value of (Xi+1 Xi) there exists uncertainty in both of these
quantities. That is, evaluation of this integral requires us to anticipate the
future value of Xi+1 so that we may evaluate f (ti+1; Xi+1) :

18
The main thing to note about Itô integrals is that I is a random variable (unlike
the deterministic case). Additionally, since I is essentially the limit of a sum of
normal random variables, then by the CLT I is also normally distributed, and
can be characterized by its mean and variance.

Example: Show that Itô’s lemma implies that

Z T Z T
3 X 2dX = X (T )3 X (0)3 3 X (t) dt:
0 0
Show that the result also can be found by writing the integral
Z T NX1
3 X 2dX = lim 3 Xi 2 (Xi+1 Xi)
0 N !1
i=0
Hint: use 3b2 (a b) = a3 b3 3b (a b)2 (a b)3.

The Itô integral here is de…ned as

Z T NX1
3X 2 (t) dX (t) = lim 3Xi2 (Xi+1 Xi)
0 N !1
i=0

19
Now note the hint:

3b2 (a b) = a3 b3 3b (a b)2 (a b)3

hence

3Xi2 (Xi+1 Xi)

3
= Xi+1 Xi3 3Xi (Xi+1 Xi)2 (Xi+1 Xi)3 ;
so that
NX1
3Xi2 (Xi+1 Xi) =
i=0

NX1 NX1 NX1

3
Xi+1 Xi3 3Xi (Xi+1 Xi)2
i=0 i=0 i=0
NX1
(Xi+1 Xi)3
i=0

20
Now the …rst two expressions above give
NX1 NX1
3
Xi+1 Xi3 = XN
3 X03
i=0 i=0
= X (T )3 X (0)3 :
In the limit N ! 1; i.e. dt ! 0; (Xi+1 Xi)2 ! dt; so
NX1 Z T
lim 3Xi (Xi+1 Xi)2 = 3X (t) dt
N !1 0
i=0
Finally (Xi+1 Xi)3 = (Xi+1 Xi)2 (Xi+1 Xi) which when N ! 1
behaves like dX 2dX O dt3=2 ! 0:

Hence putting together gives

Z T
X (T )3 X (0)3 3X (t) dt
0
which is consistent with Itô’s lemma.

21
The other important property that the Itô integral has is that it is a martingale.
We know that
Xi+1 Xi
is a martingale; i.e. in the context

E [Xi+1 Xi] = 0:
Since
" #
NP 1
E f (ti; Xi) (Xi+1 Xi) =
i=0
NP 1
f (ti; Xi) E [Xi+1 Xi] = 0
i=0
Thus
"Z #
T
E f (t; X (t)) dX (t) = 0:
0
This is, essentially a consequence of the Itô integral being non-anticipatory, as
discussed earlier. No other stochastic integral has this property.

22
Exercise We know from Itô’s lemma that
Z T Z T
4 X 3 (t) dX (t) = X 4 (T ) X 4 (0) 6 X 2 (t) dt
0 0
Show from the de…nition of the Itô integral that the result can also be found
by initially writing the integral
Z T NX1
4 X 3dX = lim 4 Xi3 (Xi+1 Xi)
0 N !1
i=0
Hint: use 4b3 (a b) = a4 b4 4b (a b)3 6b2 (a b)2 (a b)4.

23
Proving that a Continuous Time Stochastic Process is
a Martingale
Consider a stochastic process Y (t) solving the following SDE:

dY (t) = f (Yt; t)dt + g (Yt; t)dX (t); Y (0) = Y0

How can we tell whether Y (t) is a martingale?

The answer has to do with the fact that Itô integrals are martingales.

Y (t) is a martingale if and only if it satis…es the martingale condition

E [YtjFs] = Ys; 0 s t
Let’s start by integrating the SDE between s and t to get an exact form for
Y (t):
Z t Z t
Y (t) = Y (s) + f (Yu; u)du + g (Yu; u)dX (u)
s s

24
Taking the expectation conditional on the …ltration at time s, we get
Z t Z t
E [YtjFs] = E Y (s) + f (Yu; u)du + g (Yu; u)dX (u)jFs
s s
Z t
= Y (s) + E f (Yu; u)dujFs
s
where the last line follows from the fact that a Itô integral is a martingale, )
Z t Z s
E g (Yu; u)dX (u)jFs = g (Yu; u)dX (u) = 0:
s s
So, Y (t) is a martingale i¤
Z t
E f (u)dujFs = 0
s
This condition is satis…ed only if f (Yt; t) = 0 for all t. Returning to our SDE,
we conclude that Y (t) is a martingale i¤ it is of the form

dY (t) = g (Yt; t)dX (t); Y (0) = Y0

25
Exponential Martingales

Let’s start with a motivating example.

Consider the stochastic process Y (t) satisfying the SDE

dY (t) = f (t)dt + g (t)dX (t); Y (0) = Y0 (2)

where f (t) and g (t) are two time-dependent functions and X (t) is a standard
Brownian motion.

De…ne a new process Z (t) = eY (t).

How should we choose f (t) if we want the process Z (t) to be a martingale?

26
Consider the process Z (t) = eY (t). Applying Itô to the function we obtain:
dZ 1 d2Z 2 (t)
dZ (t) = dY (t) + dY
dY 2 dY 2
dZ 1 d2Z 2
= (f (t)dt + g (t)dX (t)) + 2
g (t)dt
dY 2 dY
(t) 1 2
= eY f (t) + g (t) dt + eY (t)g (t)dX (t)
2
1 2
= Z (t) f (t) + g (t) dt + g (t)dX (t)
2
Z (t) is a martingale if and only if it is a driftless process.

Therefore for Z (t) to be a martingale we must have

1
f (t) + g 2(t) = 0
2
This is only possible if
1 2
f (t) = g (t)
2

27
Going back to the process Y (t), we must have
1 2
dY (t) = g (t)dt + g (t)dX (t); Y (0) = Y0
2
implying that
Z Z
1 T 2 T
Y (T ) = Y0 g (t)dt + g (t)dX (t)
2 0 0
Hence, in terms of Z (t):
dZ (t) = Z (t)g (t)dX (t):
Using the earlier relationship, we can write Z (T ) = eY (T ):

Let’s simplify this Z (T ) =

( Z Z T )
1 T
exp Y0 g 2(t)dt + g (t)dX (t)
2 0 0
to give
( Z Z T )
1 T
Z (T ) = Z0 exp g 2(t)dt + g (t)dX (t)
2 0 0

28
Because the stochastic process Z (t) is the exponential of another process
(namely Y (t)) and because it is a martingale, we call Z (t) an exponential
martingale.

We have actually just stumbled upon a much more general and very important
result.

29
Key Condition (Novikov Condition)

A trading strategy ( ; ) = f( t; t) ; t 2 [0; :::; T ]g is a previsible process

in that t 2 Ft :

A stochastic process Yt satis…es the Novikov condition if

" Z !#
1 T 2 ds
E exp s <1
2 0
where is a t previsible process.

30
Key Fact

Given a process t satisfying the Novikov condition, then the process Mt

de…ned as we can de…ne the probability measure Q on ( ; F ) equivalent to P
through the Radon Nikodým derivative
Z t Z
1 t 2
Mt = exp s dXs ds ; t 2 [0; T ]
0 2 0 s
is a martingale.

In our earlier example t = g (t); Mt = Z (t).

31
Key Fact (Girsanov’s Theorem)

Given a process t satisfying the Novikov condition, we can de…ne the prob-
ability measure Q on ( ; F ) equivalent to P through the Radon Nikodým
derivative
Z t Z
dQ 1 t 2
= exp s dXs ds ; t 2 [0; T ]
dP 0 2 0 s
In this case, the process XtQ de…ned as
Z t
XtQ = XtP + s dXs
0
as is a standard Brownian motion on ( ; F ; Q):

32