CM2-09: Brownian motion and martingales Page 1

Brownian motion and

martingales
Syllabus objectives
4.4 Stochastic models for security prices

4.4.2 Explain the definition and basic properties of standard Brownian motion (or
Wiener process).

The Actuarial Education Company © IFE: 2019 Examinations

 Page 2 CM2-09: Brownian motion and martingales

0 Introduction
Essentially, a stochastic process is a sequence of values of some quantity where the future values
cannot be predicted with certainty. This and the following chapter are concerned with
continuous-time stochastic processes that have applications in financial economics. These
chapters are of a very mathematical nature and you may find some of it hard-going. It is more
important that you gain a higher-level understanding than that you master the pure maths that
underlies it. For example, you may find it useful to learn Ito’s Lemma as a procedure rather than
trying to understand the pure mathematical concepts.

The most important process studied here is the Wiener process, also known as Brownian motion,
which is the subject of Section 1. These two terms will be used interchangeably. We define this
as a process with continuous sample paths and independent and normally distributed increments.
A Brownian motion is the continuous-time version of a random walk, as we will see. The graph in
Section 1.2 shows a typical sample path.

If security prices can be modelled in some way in terms of Brownian motion, this will be useful for
pricing certain types of options. This is discussed further in Parts 3 and 4 of the course.

Section 2 of this chapter introduces martingales. A martingale is a process whose current value is
the best estimate of its future values. We will see later that martingale theory has important
applications in relation to financial derivatives.

The notation used in financial economics generally is not standardised and similar notation can
refer to different quantities: readers should check the definitions provided in each section. In
particular, the value of a random stochastic process can be equivalently written as Xt or X (t) .
Furthermore, standard Brownian motion can be denoted by Bt (as in the Tables), or Wt or Zt as
found throughout the Core Reading.

The Core Reading in this chapter is adapted from course notes written by Timothy Johnson.

© IFE: 2019 Examinations The Actuarial Education Company

 CM2-09: Brownian motion and martingales Page 3

1 Introduction to Brownian motion

1.1 Introduction
In 1895, Louis Bachelier embarked on a doctorate on the ‘Theory of Speculation’.
Bachelier’s approach was fairly conventional at the time; he would model an asset price as
a random walk. At the start of his thesis he argues that:

At a given instant the market believes neither in a rise nor in a fall of the true price.

Which means that:

The mathematical expectation of the speculator is zero.

His innovation was to consider the walk to be continuous, rather than a discrete-time
random walk. This is analogous to moving from the binomial to the normal distribution.

Bachelier was unable to mathematically define the paths he discussed. A little later, in
1903–1904, Einstein used a similar model to represent the motion of atoms/molecules in a
liquid. Einstein also failed to define the path he was working with. However, since his
paper was used as evidence that atoms existed it became important in physics that the
paths were rigorously defined. This was done by Norbert Wiener in 1921. Today physicists
will refer to the paths used by Bachelier as ‘Brownian motion’, a physical process, while
mathematicians refer to them as a ‘Wiener process’, which is a mathematical object.

The phenomenon of ‘Brownian motion’ is named after the nineteenth century botanist Robert
Brown who observed the random movement of pollen particles in water. The path of a
two-dimensional Brownian motion process bears a resemblance to the track of such pollen
particles.

1.2 Definition of the Wiener process (standard Brownian motion)

A stochastic process Wt , t 0 is a Wiener process if:

(i) W0 0

(ii) Wt has continuous sample paths.

This means that the graph of Wt as a function of t doesn’t have any breaks in it.

(iii) For any 0 s t the increment Wt Ws is normally distributed,

Wt Ws N (0, t s ).

This property shows that the increments are stationary in that their statistical properties
rely on the size of the interval t s . The concept of stationarity is discussed further in
Subject CS2.

(iv) Wt has independent increments, that is for any sequence of times

0 t1 t2 t n we have that the increments Wt n Wt n 1 , ,Wt 3 Wt 2 ,Wt 2 Wt1 are
independent random variables.

Alternatively, Wt Ws is independent of FsW (Wu s) , the natural filtration of Ws .

The Actuarial Education Company © IFE: 2019 Examinations

 Page 4 CM2-09: Brownian motion and martingales

The natural filtration FtW represents the history of the process up to and including time t.
This concept is covered in more detail in the martingale section of the chapter. FtW may
be written as (Wu t ) to denote that it is the filtration generated by the process Wt .

The fact that a Wiener Process has independent increments implies it is Markovian
(in fact it is ‘strong Markovian’).

Intuitively, a Markov process is one where, if we know the latest value of the process, we
have all the information required to determine the probabilities for the future values.
Knowing the historical values of the process as well would not make any difference.
Markov processes are also discussed in Subject CS2.

Since Wt N (0, t ) it should be clear that P (Wt ,t ) 1.

Property (iii) combined with property (i) gives us Wt Wt W0 N(0,t) , which results in
E Wt 0.

Brownian Motion

1
0.8
0.6
Value, x

0.4
0.2
0
-0.2 0 0.2 0.4 0.6 0.8 1
-0.4
Time, t

Figure 9.1: a typical sample path of Brownian motion

Brownian motion can be viewed as the continuous version of a simple symmetric random walk.

1.3 Brownian motion in general

Standard Brownian motion is a special case of the more general form of Brownian motion.

The term Brownian motion refers to a process { Zt , t 0} that satisfies criteria (ii) and (iv) above,
but with the distribution in criteria (iii) being replaced with N (t s), 2 (t s) .

Here is the drift coefficient and is known as the diffusion coefficient (or volatility).

Standard Brownian motion is obtained when 0, 1 and Z0 0.

It turns out that Brownian motion is the only process with stationary independent increments and
continuous sample paths. This is far from obvious and we won’t prove it here.

© IFE: 2019 Examinations The Actuarial Education Company

CM2-09: Brownian motion and martingales Page 5

The relationship between standard Brownian motion and Brownian motion is the same as the
relationship between a standard normal distribution, N(0,1) , and a general N( , 2 ) distribution.
A Brownian motion with given diffusion and drift coefficients can be constructed out of a standard
Brownian motion {Wt ,t 0} by setting:

Zt Z0 Wt t

Question

Let Wt be a standard Brownian motion. Prove that Zt Z0 Wt t is a Brownian motion with

diffusion coefficient and drift .

Solution

The second property of a Brownian motion – that it has continuous sample paths – is met because
Zt is driven by only time t and the continuous process Wt .

The increments of Zt are independent of the past because:

Zt Z s Wt Ws t s

and we know that the increments Wt Ws have this property.

The third property we require is that:

Zt Zs N t s , 2 t s

This follows because Wt Ws N 0, t s and therefore:

Zt Zs Wt Ws t s

N 0,t s t s

N 0, 2 t s t s

N t s , 2 t s

Question

How can a Brownian motion, Zt , that has drift and diffusion parameter and a starting value
of Z0 be converted into a standard Brownian motion?

The Actuarial Education Company © IFE: 2019 Examinations

 Page 6 CM2-09: Brownian motion and martingales

Solution

Just invert the relationship given:

Zt Z0 t
Wt

This is analogous to converting an observed value x from a general normal distribution, N( , 2 ) ,

into a value from the standard normal distribution, N(0,1) , by calculating the standardised value:

x
z

1.4 Properties of Brownian motion

Standard Brownian motion has a number of other properties inherited from the simple symmetric
random walk. A simple symmetric random walk is a discrete-time stochastic process:

1
n 1 with probability
2
Xn Zi where Zi
1
i 1 1 with probability
2

The value of the process increases or decreases randomly by 1 unit (= ‘simple’) with equal
probability (= ‘symmetric’).

If we reduce the step size progressively from 1 unit until it is infinitesimal (and rescale the X
values accordingly), the simple symmetric random walk becomes standard Brownian motion. An
important consequence of this is that a standard Brownian motion returns infinitely often to zero,
or indeed any other level.

Many of the properties of standard Brownian motion can be demonstrated using the following
decomposition. For s t :

Wt Ws (Wt Ws )

a decomposition in which the first term is known at time s and the second is independent of
everything up to and including time s .

In calculations involving Brownian motion, we often need to split up Wt in this way, so that we
can work with independent increments.

Covariance of a Wiener Process

An important characteristic of a process is the covariance between its value at s 0
and t s :

Cov (Ws ,Wt ) E [(Ws E [Ws ])(Wt E [Wt ])]

E [Ws (Ws (Wt Ws )]

© IFE: 2019 Examinations The Actuarial Education Company

CM2-09: Brownian motion and martingales Page 7

This follows from the fact that E[Wt ] E[Ws ] 0 , and then by applying the decomposition
Wt Ws (Wt Ws ) .

By independence of increments:

Cov (Ws ,Wt ) E [Ws2 ] E [Ws ]E [(Wt Ws )]

Var (Ws ) 0

This follows from the fact that E[Ws2 ] Var (Ws ) E 2[Ws ] s 0 .

In general, Cov (Ws ,Wt ) min{ s, t } .

The importance of this result is that, in fact, if a stochastic process has the property that:

Cov ( X s , X t ) min{ s, t }

then the process X t is a Wiener process (this is Lévy’s Theorem).

Scaled Wiener process

Given a positive constant c and a Wiener process Wt define the stochastic process X t by:

Xt cWt / c

The ‘clock’ of the process X t has been scaled by a factor c. For example, the process has
been slowed down and magnified if c 1 (and speeded up and shrunk if c 1 ).

By applying Lévy’s Theorem:

Cov ( X t u , Xt ) Cov cW t u , cW t
c c

c Cov W t u ,W t
c c

c min t u,t
c c

t
c
c
t

assuming u 0 .

Since Cov ( X t u , Xt ) min{t u , t } , X t is a Wiener process.

The Actuarial Education Company © IFE: 2019 Examinations

Page 8 CM2-09: Brownian motion and martingales

The scaled process could also be re-parameterised as:

1
Xt Wat
a

1
with a .
c

Time-inverted Wiener process

Given a Wiener process Wt define the stochastic process X t by:

Xt tW1/ t

The time-inverted Wiener process is itself a Wiener process, as can be shown by Lévy’s
Theorem.

Let u 0.

Then we have:

Cov ( X t u , X t ) Cov (t u )W 1 , tW1

t u t

(t u )t Cov W 1 ,W 1
t u t

Since 1/ (t u ) 1/ t and by the covariance of Wiener processes:

1 1
Cov ( X t u , Xt ) (t u )t min ,
t u t
1
(t u )t
t u
t

Therefore Xt is also a Wiener process.

The time-inverted Wiener process is useful in proving limiting properties. For example,
since tW1/ t is a Wiener process, then:

Wt
lim lim W1/ t W0 0
t t t

Correlated Wiener processes

The process defined by:

Zt Wta 1 2
Wtb

where Wta and Wtb are independent Wiener processes and 1 1 defines the
correlation between Zt and Wta .

© IFE: 2019 Examinations The Actuarial Education Company

CM2-09: Brownian motion and martingales Page 9

It should be obvious that E [ Zt ] 0.

Question

Show that E[ Zt ] 0 .

Solution

The process Zt is only a weighted sum of two Wiener processes, both of which have zero
expectation. Therefore we have:

E [ Zt ] E Wta 1 2
Wtb

E Wta E 1 2
Wtb

E Wta 1 2
E Wtb

2
0 1 0 0

The variance of the process is then given by:

Var (Zt ) Var Wta 1 2

Wtb

Since Wta and 1 2

Wtb are independent:

2
Var (Zt ) Var Wta 1 2
Var Wtb

2 2
t 1 t

Similarly, the variance of the increment Zt u Zt is u.

Question

Show that Var (Zt u Zt ) u for u 0 .

The Actuarial Education Company © IFE: 2019 Examinations

Page 10 CM2-09: Brownian motion and martingales

Solution

Var (Zt u Zt ) Var Wta u 1 2

Wtb u Wta 1 2
Wtb

Var Wta u Wta 1 2

Wtb u Wtb

Var Wta u Wta Var 1 2

Wtb u Wtb

2 2
(t u t) (1 ) (t u t)
u

Furthermore, the covariance of Zt and Wta is given by:

Cov (Zt ,Wta ) t

The correlation between Zt and Wta is defined as:

Cov (Zt ,Wta )

Var (Zt )Var (Wta )

Non-differentiability of sample paths

There are various ways of proving the fact that the Wiener process is non-differentiable.
dWt
Consider the following proof by contradiction. If the derivative existed, then we can
dt
say that:

Wt Ws dWt Ws Wt dWt
lim or lim
t s 0 t s dt s t 0 s t dt

The first inequality assumes that t s , and the second covers the case when s t . These
statements come from the definition of a derivative as the convergence of a function’s gradient:

f (x h) f (x)
f (x) lim
h 0 h

If such a derivative existed, then we could find an arbitrarily small to measure the difference
between the derivative and the gradient.

However, (in the case t s ):

Wt Ws dWt dWt 1
N ,
t s dt dt t s

which will have a positive probability of being greater than , and so the statement is
uncertain. In fact, since the variance increases as t s 0 , it never holds, almost surely.

© IFE: 2019 Examinations The Actuarial Education Company

 CM2-09: Brownian motion and martingales Page 11

That is to say, it’s impossible to bound a normal distribution.

The fundamental theorem of calculus is that given a derivative, f ( x ) , then the integral f ( x ) ,
is understood as:

b b b
df ( x )
f (b) f (a ) f ( x )dx f (a ) dx f (a ) df ( x )
dx
a a a

dWt
However, since the idea of is meaningless, the stochastic integral,
dt

b
dWt
a

cannot be handled using classical calculus and there appears to be no way of

understanding how a Wiener process behaves between two times.

In other words, because a Wiener process isn’t differentiable anywhere then it’s not clear how
integrals should be handled when the variable of integration is Wt . Stochastic integrals will be
dealt with in the next chapter.

1.5 Geometric Brownian motion

As mentioned at the start of this chapter, Brownian motion was used by Bachelier to model the
movements of the Paris stock exchange index.

However successful the Brownian motion model may be for describing the movement of market
indices in the short run, it is useless in the long run, if only for the reason that a standard
Brownian motion is certain to become negative eventually. It could also be pointed out that the
Brownian motion model predicts that daily movements of size 100 or more would occur just as
frequently when the process is at level 100 as when it is at level 10,000.

A more useful model is:

St e Zt

where Zt is the Brownian motion process Zt Z0 Wt t . Thus St , which is called geometric

2
Brownian motion, is lognormally distributed with parameters Z0 t and t . So the values of
2
log St are normally distributed with mean Z0 t and variance t.

The Actuarial Education Company © IFE: 2019 Examinations

Page 12 CM2-09: Brownian motion and martingales

Brownian Motion

5
4

3
Value, x 2
1

-1 0 0.2 0.4 0.6 0.8 1

Time, t

Figure 9.2: A Brownian motion, Wt , and a geometric Brownian motion, St e Zt

The most important property of St is:

St 0 for all t

From the properties of the lognormal distribution we also have:

E St exp (Z0 t) ½ 2t and Var (St ) E 2 St exp( 2t) 1

Geometric Brownian motion features heavily in this course. For example, Black and Scholes’
Nobel prize-winning formula for pricing European options assumes that the price of the
underlying asset is a geometric Brownian motion.

The properties of St are less helpful than those of Brownian motion. For example, St has neither
independent increments nor stationary increments.

The increments of St are of the form St Ss e Zt e Zs .

But this is not so important because Zt does possess these desirable properties. Analysis of path
properties of St should involve first taking the logarithm of the observations, and then
performing the analysis using techniques appropriate to Brownian motion.

S S e Zt
The log-return log t from time s to time t is given by log t log Z Zt Z s .
Ss Ss e s

It follows by the independent increments property of Brownian motion that the log-returns, and
hence the returns themselves, are independent over disjoint time periods.

© IFE: 2019 Examinations The Actuarial Education Company

 CM2-09: Brownian motion and martingales Page 13

2 Martingales

2.1 Introduction
In simple terms, a martingale is a stochastic process for which its current value is the best
estimate of its future value. So, the expected future value is the current value. Other ways of
thinking of a martingale are that the expected change in the process is zero or that the process
has ‘no drift’.

Note
Throughout this course we will be using the word ‘expected’ in its statistical sense, rather than in
the everyday sense.

Consider a person standing on a ‘never-ending’ ladder. Every minute they move up or down the
ladder one step, depending on whether a tossed coin comes up heads or tails.

In the everyday sense of the word, after the next toss of the coin, we ‘expect’ them to move up or
down (but we don’t know which way). However, in the statistical sense, because
1 1 1 1 0 , we ‘expect’ them to stay exactly where they are, even though there’s no
2 2
way that that can happen!

The idea of martingales is consistent with the original equestrian term ‘martingale’, meaning a
holster used to keep a horse ‘pointing straight ahead’.

Their importance for modern financial theory cannot be overstated. In fact, the whole theory of
pricing and hedging of financial derivatives is formulated in terms of martingales.

For this reason, it may be best to think of a martingale as being a random process that has ‘no
drift’ because the idea of drift is more consistent with the way we think about real financial
assets. We have already seen that it is possible to model the log of a share price, log St , using
Brownian motion with a drift . You can think of as being the rate of the long-term drift of
the log of the share price. It is the underlying non-random trend. It should not come as a great
surprise that when we remove this underlying non-random trend (or drift) and look at log St t,
we obtain a martingale.

Conditional expectation
The features of martingales rely on the application of conditional expectations.

The filtration Ft represents everything that can be known up to and including time t.

Some random variables will be known by time t . We say that Xt is Ft -measurable if the value of
the process is known at time t, ie it belongs to Ft .

A stochastic process Xt , t 0 is said to be adapted to the filtration Ft if Xt is Ft -measurable for

all t .

The Actuarial Education Company © IFE: 2019 Examinations

 Page 14 CM2-09: Brownian motion and martingales

If Ft is the filtration generated by Xt (as opposed to any other process), then it is known as the
natural filtration of Xt and is denoted here by FtX . The following results will be used extensively.

(i) E E X |FtX E X

(ii) if X is Ft -measurable, then E X |Ft X

(iii) if X is independent of Ft , then E X |Ft E X

(iv) any function of Xt is adapted to FtX .

2.2 Wiener processes are martingales

We have the following definitions of continuous-time martingales.

Given a filtered probability space ( , F , Ft , P ) , a stochastic process X t is called a

martingale with respect to the filtration, Ft , if:

X t is adapted to Ft

E Xt for all t

E X t | Fs X s for all s t

The first condition is just a technicality to ensure that the process value can be known with
certainty at time t, and the second is to guarantee that Xt is integrable. In most questions we are
only concerned with the last condition and we’ll assume the first two hold.

Question

In words, what does the E[ Xt |Fs ] in the last condition mean?

Solution

E[ Xt |Fs ] means the expected value of the process at time t, given that we are at time s and we
know the history of the process up to and including time s.

Of all the properties of martingales, the most useful is also the simplest: a martingale has constant
mean, ie E[ Xn ] E[ X0 ] X0 for all n .

© IFE: 2019 Examinations The Actuarial Education Company

CM2-09: Brownian motion and martingales Page 15

Given s t , a supermartingale is such that:

E X t | Fs Xs

while a submartingale is such that:

E X t | Fs Xs

A supermartingale has either negative or zero drift, whereas a submartingale has either positive
or zero drift. So a process which is both a supermartingale and a submartingale must therefore
be a martingale.

Consider:

E Wt | FsW E Ws (Wt Ws ) | FsW

E Ws | FsW E (Wt Ws ) | FsW

Since increments are independent and Wt Ws N (0, t s) :

E Wt | FsW Ws

The Wiener process is a martingale with respect to its natural filtration, noting that
E Wt since Wt almost surely.

Question

An asset’s value at time t (in pence and measured in years) is denoted by At and fluctuates in
value from day to day. Within these random fluctuations, there appears to be an underlying
long-term trend, in that the asset’s value is increasing by 2 pence on average each week.

(i) Assuming that there are exactly 52 weeks in a year, suggest a process based on At that
you think might be a martingale.

(ii) Suppose that the price increments have a continuous uniform distribution such that
At As U[ 52(t s),260(t s)] . Construct a martingale out of At .

Solution

(i) The value of the asset is increasing on average by 2 pence a week. Assuming that there
are exactly 52 weeks in a year, this means the asset is ‘drifting’ by 104 pence a year.
A martingale is a process without drift and so a good suggestion would be to remove this
drift and consider the process:

At 104t

The Actuarial Education Company © IFE: 2019 Examinations

 Page 16 CM2-09: Brownian motion and martingales

(ii) Using the formula for the expected value of a uniform distribution from page 13 of the
Tables, we have:

260(t s) 52(t s)
E At Fs E As ( At As ) Fs As As 104(t s)
2

So for every increase of t s in the time, the process is ‘drifting’ by 104(t s) .

One way to construct a martingale is to subtract the drift. Mathematically, we can

subtract 104t from both sides of the last equation to get:

E At 104t Fs As 104 s

This now fits in with the definition given for a martingale in continuous time.

So the process At 104t is again a martingale.

2.3 Library of martingales

Now consider the stochastic process defined by Wt2 t.

2
E Wt2 t | FsW E Ws (Wt Ws ) | FsW t

2
E Ws2 | FsW 2E Ws (Wt Ws ) | FsW E Wt Ws | FsW t

Ws2 s

so it is a martingale with respect to its natural filtration.

Note how the three expectations above were evaluated:

E Ws2 |FsW Ws2 since the value of Ws is known with certainty at time s

E Ws (Wt Ws )|FsW WsE (Wt Ws )|FsW 0 because the increments of a Wiener

process have zero mean
2
E Wt Ws |FsW Var Wt Ws E 2 Wt Ws t s 0 due to the statistical

properties of Wiener process increments.

Question

Let Wt be a Wiener process. Determine whether the process Wt2 is a supermartingale or a

submartingale.

© IFE: 2019 Examinations The Actuarial Education Company

CM2-09: Brownian motion and martingales Page 17

Solution

Given that Wt2 t is a martingale with zero drift, then Wt2 must have positive drift. This means
that Wt2 is a submartingale.

1 2
Finally, consider the stochastic process defined by exp Wt 2
t .

The same approach is taken here as before; decompose the Wiener process, but also decompose
time t into s (t s) .

2
E exp Wt 1
2
t | FsW E exp (Ws (Wt Ws )) 1
2
2
(s (t s ) | FsW

2 2
exp Ws 1
2
s E exp (Wt Ws ) 1
2
(t s ) | FsW

To evaluate this expectation, consider a random variable Z where Z N , 2 , then its moment

t 1 2t 2
generating function is defined as: MGFZ (t) E[exp(Z t)] e 2 . Therefore when t 1 we
1 2
have: E[exp(Z )] e 2 .

From the MGF of a normal:

1 2 1 Var (Z )
E [exp(Z )] exp 2
exp E [ Z ] 2

1 2
If exp Yt , where Yt (Wt Ws ) 2
(t s ) is a normally distributed random variable, we
have that:

2
E exp (Wt Ws ) 1
2
(t s ) | FsW E [exp(Yt )]

exp E [Yt ] 1 Var (Y )

2 t

1 2 1 2
exp 2
(t s) 2
(t s)

So:

2
E exp Wt 1
2
t | FsW exp Ws 1
2
2
s

meaning that the process exp Wt 1

2
2
t is an FtW -martingale.

The Actuarial Education Company © IFE: 2019 Examinations

Page 18 CM2-09: Brownian motion and martingales

This result is a special case of the fact that:

T T
2
exp f (t )dWt 1
2
(f (t ))2 dt
0 0

is a martingale.

© IFE: 2019 Examinations The Actuarial Education Company

CM2-09: Brownian motion and martingales Page 19

Chapter 9 Summary
Wiener process (standard Brownian motion)
A Wiener process is a stochastic process with the defining properties:
W0 0

it has independent increments

it has stationary increments
it has normally distributed increments, ie Wt Ws N(0,t s)

it has continuous sample paths.

A Wiener process is the continuous-time analogue of a random walk.

Other properties of Wiener processes include:

{Wt ,t 0} is a Markov process

{Wt ,t 0} is a martingale, ie E(Wt |Fs ) Ws

{Wt ,t 0} returns infinitely often to 0, or indeed to any other level

Cov(Ws ,Wt ) min s,t

Xt cWt /c ,t 0 is also a Wiener process (scaling property)

Xt tW1/t ,t 0 is also a Wiener process (time inversion property)

the sample path is not differentiable anywhere.

Brownian motion with drift

Brownian motion with drift is related to standard Brownian motion by the equation:

Zt Z0 Wt t

where is the volatility or diffusion coefficient and is the drift.

The Actuarial Education Company © IFE: 2019 Examinations

Page 20 CM2-09: Brownian motion and martingales

Geometric Brownian motion (lognormal model)

For modelling purposes a Brownian motion may have to be transformed, for example by
taking logarithms. A useful model for security prices is geometric Brownian motion:

St e Zt

where Zt is the Brownian process Zt Z0 Wt t . Thus St is lognormally distributed

2
with parameters Z0 t and t.

Martingales
A martingale is a stochastic process such that:
Xt is adapted to Ft

E Xt for all t

E[ Xt |Fs ] X s for all s t

Martingales are processes with no drift. In fact, it can be shown that a martingale has
constant mean, ie:

E[ Xn ] E[ X0 ] for all n

Martingales constructed from Wiener processes

Various martingales can be constructed from Wiener processes, for example, Wt , Wt2 t
Wt 1 2t
and e 2 .

© IFE: 2019 Examinations The Actuarial Education Company

 CM2-09: Brownian motion and martingales Page 21

Chapter 9 Practice Questions

9.1 Assume that the spot rate of interest at time t , S(t) , can be modelled by S(t) = e 2 W (t) , where
Exam style W (t) is a Brownian motion with drift coefficient and volatility coefficient 1 such that W(0) = 0.

(i) Write down an expression for W (t) in terms of a standard Brownian motion, B(t) . [1]

(ii) Show that { S(t) t 0} is a continuous-time martingale. [4]

[Total 5]

9.2 ’Brownian motion is the only process with stationary independent increments and continuous
sample paths.’

(i) Give mathematical definitions of each of the three underlined terms.

(ii) State the distribution of the increments for a standard Brownian motion.

9.3 (i) What is meant by saying that the process {Yt } is a martingale with respect to another
process { Xt } ?

Let Bt ( t 0 ) be a standard Brownian motion.

(ii) Show that Bt and Bt2 kt are both martingales with respect to Bt , for a suitably chosen
value of the constant k , which you should specify.

(iii) Show that there is a value of the constant c , which you should specify, such that
(a bBt )2 ct is a martingale with respect to Bt , where a and b are constants.

9.4 Let Bt ( t 0 ) be a standard Brownian motion process starting with B0 0.

(i) What is the probability that B2 takes a positive value?

(ii) What is the probability that B2 takes a value in the interval ( 1,1) ?

(iii) Show that the probability that B1 and B2 both take positive values is 38 .

(iv) What is the probability that Bt takes a negative value at some time between t 0 and
t 2?

9.5 Consider the statement: ‘If you want to find the variance of X B(s) B(t) , where s t , for a
standard Brownian motion process, you can use the fact that B(s) and B(t) are independent to
get Var ( X ) s t .’

(i) Explain why the statement is not correct, and find a correct expression for Var(X ) .

(ii) Hence show that the general formula for Var B(t1 ) B(t2 ) when t1 ,t2 0 can be
expressed as t1 t2 2min(t1 ,t2 ) .

The Actuarial Education Company © IFE: 2019 Examinations

 Page 22 CM2-09: Brownian motion and martingales

9.6 Let Bt ( t 0 ) be a standard Brownian motion process starting with B0 0.

Exam style
(i) Show that, when s t , E(BsBt ) s . [3]

(ii) Hence find a general formula for the correlation coefficient (Bt1 ,Bt2 ) . [2]
[Total 5]

9.7 (i) Write down a formula for E eaX where X N , 2 and, by differentiating, or
Exam style
otherwise, derive an expression for E XeaX . [2]

(ii) Show that:

aBt 1 a 2t
Xt Bt at e 2

is a martingale, where Bt is a standard Brownian motion, and a is an arbitrary constant.

You may assume that E Xt . [5]
[Total 7]

© IFE: 2019 Examinations The Actuarial Education Company

 CM2-09: Brownian motion and martingales Page 23

Chapter 9 Solutions
9.1 (i) General Brownian motion

A general Brownian motion can be defined as:

W (t) W (0) t B(t)

where B(t) is standard Brownian motion, is the drift, is the volatility coefficient and W (0) is
the value of general Brownian motion at time 0.

Since 1 and W (0) =0:

W (t) t B(t) [1]

(ii) S(t ) is a martingale

We write S(t) as St and B(t) as Bt for neatness. Let t s , then:

E St Fs E e 2 ( t Bt ) Fs

2
E e 2 t 2 Bt Fs

2
e 2 t E e 2 Bt Fs

2 2 Bt Bs Bs
e 2 tE e Fs

2 2 Bt Bs
e 2 t e 2 Bs E e Fs

2 2 Bt Bs
e 2 t e 2 Bs E e [2]

The filtration Fs can be left out because of the independent increments property. Now
Bt Bs N 0,t s and so the expectation is of the form E[eaX ] where X N 0,t s and
a 2 . Therefore, we can use the MGF of a normal distribution calculated at point 2 to
determine the expectation.

So, using the MGF formula, from page 11 of the Tables, we get:

2 0( 2 ) 12 (t s)( 2 )2
E St Fs e 2 t e 2 Bs e
2
e 2 s 2 Bs [1]

Ss

The Actuarial Education Company © IFE: 2019 Examinations

 Page 24 CM2-09: Brownian motion and martingales

Finally we check that E St for all values of t .

2
E[ St ] E e 2 t 2 Bt

2
e 2 t E e 2 Bt

2
e 2 t MBt ( 2 )

2 0( 2 ) 1 t ( 2 )2
e 2 te 2 1

where: MBt ( 2 ) is the MGF of Bt at 2 . [1]

[Total 4]

9.2 (i) Mathematical definitions

‘Stationary increments’ means that the distribution of Bt Bs ( t s ) depends only on t s .

‘Independent increments’ means that Bt Bs is independent of the filtration Fr whenever

r s t.

‘Continuous sample paths’ means that the function t Bt ( ) for each particular realisation is
a continuous function of t .

(ii) Distribution of the increments

For a standard Brownian motion, Bt Bs (with t s ) has a N(0,t s) distribution.

9.3 (i) What is a martingale?

Strictly, we should say that the process {Yt } is a martingale with respect to the filtration {Ft } of
the process { Xt } , which means that:

E[Yt |Fs ] Ys for all s t

and E[ Yt ]

and Yt is adapted to Ft

(ii) Show that these processes are martingales

If we use an s subscript to denote the expected value with respect to the filtration at time s ,
then we can write:

Es [Bt ] Es [(Bt Bs ) Bs ] Es [Bt Bs ] Es [Bs ]

Since Bt Bs N(0,t s) and the value of Bs is known at time s , this gives:

Es [Bt ] 0 Bs Bs

© IFE: 2019 Examinations The Actuarial Education Company

CM2-09: Brownian motion and martingales Page 25

We have shown that the expected future value of Bt is equal to its current value (at time s ). We
also need to show that E Bt .

One way to do this is to note that x 1 x 2 for all values of x .

2
So: E Bt E 1 Bt2 1 var(Bt ) E (Bt ) 1 t 02

Therefore, Bt is a martingale (with respect to Bt ).

Similarly:

Es [Bt2 ] E s [{(Bt Bs ) Bs }2 ]

E s [(Bt Bs )2 ] 2E s [(Bt Bs )Bs ] E s [Bs2 ]

vars [Bt Bs ] [E s (Bt Bs )]2 2Bs E s [Bt Bs ] Bs2

(t s) 02 0 Bs2

t s Bs2

So: E s [Bt2 t ] Bs2 s

We can show that E Bt2 t for any value of t, by first noting that:

x2 k x 2 k (k 0) for all values of x .

2
So: E Bt2 t E Bt2 t var(Bt ) E (Bt ) t t 02 t 2t

Since the expected future value of Bt2 t is equal to its current value (at time s ) and the expected
value of its modulus is finite, Bt2 t is a martingale with respect to Bt , and the required constant
is k 1.

(iii) Value of c to make the process a martingale

We know that Bt and Bt2 t are both martingales with respect to Bt .

So, if we use an s subscript to denote the expected value with respect to the filtration at time s ,
then:

Es [Bt ] Bs

and E s [Bt 2 t ] Bs2 s E s [Bt 2 ] Bs2 t s

The Actuarial Education Company © IFE: 2019 Examinations

 Page 26 CM2-09: Brownian motion and martingales

Using these results, we then find that:

E s [(a bBt )2 ] E s [a2 2abBt b2Bt2 ]

a2 2abE s [Bt ] b2E s [Bt2 ]

a2 2abBs b2 (Bs2 t s)

(a bBs )2 b2 (t s)

So: E s [(a bBt )2 b2t ] (a bBs )2 b2 s

ie (a bBt )2 b2t is a martingale with respect to Bt . So the required value of the constant is
c b2 .

2
The technical condition E a bBt ct will hold as in (ii).

9.4 (i) Probability that B2 takes a positive value

B2 B2 B0 N(0,2) . Therefore, B2 is equally likely to be positive or negative (and has zero

probability of being exactly zero).

So: P(B2 0) 12

(ii) Probability that B2 takes a value in the interval ( 1,1)

B2 B2 B0 N(0,2) .

Standardising:

1 1
P( 1 B2 1) 0.760 (1 0.760) 0.520
2 2

(iii) Probability that B1 and B2 both take positive values

We can write the required probability as:

p P(B1 0,B2 0) P(B1 B0 0,B2 B1 B1 )

If we now write X B1 B0 and Y B2 B1 , then we know from the properties of Brownian

motion that X and Y are independent, each with a N(0,1) distribution.

So the required probability is:

p P( X 0,Y X)

© IFE: 2019 Examinations The Actuarial Education Company

 CM2-09: Brownian motion and martingales Page 27

Since the range of values of Y depends on the value of X , we must use a double integral to
evaluate this:

p (x) (y) dy dx
x 0 y x

where the joint density function is expressed as the product of the individual density functions by
independence.

So:

p (x) (y) dy dx
x 0 y x

(x) (y) x dx
x 0

( x) 1 ( x) dx
x 0

(x) (x)dx
x 0

Finally:

p (x) (x)dx 1{ (x)}2 1 12 1 2 3

0 2 0 2 2 8

(iv) Probability that Bt takes a negative value at some time between 0 and 2

The probability is 1 because Bt will almost surely take a negative value at some point close to
t 0.

9.5 (i) Explain why the statement is not correct

The statement is not correct because B(s) and B(t) , which represent the value of the process at
two different times, are not independent. In fact, they are positively correlated.

The Actuarial Education Company © IFE: 2019 Examinations

 Page 28 CM2-09: Brownian motion and martingales

It is actually the increments B(s) B(0) and B(t) B(s) that are independent. The correct
calculation can be done by expressing X B(s) B(t) in terms of these increments:

Var ( X ) Var B(s) B(t)

Var 2B(0) 2{B(s) B(0)} {B(t) B(s)}

4Var B(0) 4Var B(s) B(0) Var B(t) B(s)

0 4 s (t s) 3s t

(ii) Show the general formula

This 3s t formula works if s t . If t s , we can swap the letters to get 3t s . If s t , we

have:

Var( X ) Var B(s) B(t) Var 2B(s) 4Var B(s) 4 s (or 4t)

which agrees with either formula.

So a general formula would be s t 2min(s,t) , or if we’re using t1 and t2 to denote the times,
t1 t2 2min(t1 ,t2 ) .

9.6 (i) Show that E (Bs Bt ) s

If we express Bt in terms of the increment Bt Bs , which is independent of the value of Bs , we

get:

E[BsBt ] E[Bs {(Bt Bs ) Bs }]

E[Bs (Bt Bs )] E[Bs2 ]

E (Bs )E (Bt Bs ) Var[Bs ] [E (Bs )]2 0 s 0 s [3]

(ii) Find a general formula for the correlation coefficient

We can then calculate the covariance and correlation between these two values:

Cov(Bs ,Bt ) E BsBt E Bs E Bt s

Cov (Bs ,Bt ) s s

and (Bs ,Bt ) [1]
Var (Bs )Var (Bt ) st t

This formula only applies when s t . We can generalise this to cover any positive times t1 and t2 ,
if we write it in the form:

min(t1 ,t2 )
(Bt1 ,Bt2 ) [1]
max(t1 ,t2 )
[Total 2]

© IFE: 2019 Examinations The Actuarial Education Company

 CM2-09: Brownian motion and martingales Page 29

t1 t2
There are various alternative ways of writing this, eg (Bt1 ,Bt2 ) min , .
t2 t1

9.7 (i) Formulae for expectations

1 2 2
a a
E eaX e 2 [½]

This is the MGF of a normal random variable and can be found in the Tables.

We therefore have:

d d a 1 2 a2 a 1 2 a2
2
E XeaX E eaX e 2 a e 2 . [1½]
da da
[Total 2]

(ii) Show that Xt is a martingale

2
E Xt Fs E Bt at eaBt 0.5a t Fs

a Bs Bt Bs 0.5a2t
E Bs Bt Bs at e Fs

2 a B B 2 a B B
E Bs at eaBs 0.5a t e t s Fs E Bt Bs eaBs 0.5a t e t s Fs [2]

Now we can use the fact that we are conditioning on all the information known at time s. Any
terms involving Bs can be taken outside the expectation, as can terms in s and t (which are
fixed, not random points in time). This gives:

2 aB B 2 aB B
Bs at eaBs 0.5a t E e t s Fs eaBs 0.5a t E Bt Bs e t s Fs [1]

We can then drop the conditions since the increments are independent of the past:

2 aB B 2 aB B
Bs at eaBs 0.5a t E e t s eaBs 0.5a t E Bt Bs e t s [1]

2
Using part (i), and noting that Bt Bs N(0,t s) so that 0 and t s , we therefore get:

2 0.5a2 t s 2 0.5a2 t s
Bs at eaBs 0.5a t e eaBs 0.5a t a t s e
2
Bs as eaBs 0.5a s

Xs [1]

as required. Note that the bounded condition is given in the question.

[Total 5]

The Actuarial Education Company © IFE: 2019 Examinations