Mathematical Finance

Electronic copy of this paper is available at: http://ssrn.
com/abstract=976593
Mathematical Finance
Introduction to continuous time
Financial Market models
Dr. Christian-Oliver Ewald
School of Economics and Finance
University of St.Andrews
Electronic copy of this paper is available at: http://ssrn.com/abstract=976593
Abstract
These are my Lecture Notes for a course in Continuous Time Finance
which I taught in the Summer term 2003 at the University of Kaiser-
slautern. I am aware that the notes are not yet free of error and the
manuscrip needs further improvement. I am happy about any com-
ment on the notes. Please send your comments via e-mail to ce16@st-
andrews.ac.uk.
Working Version
March 27, 2007
1
Contents
1 Stochastic Processes in Continuous Time 5
1.1 Filtrations and Stochastic Processes . . . . . . . . . . . . 5
1.2 Special Classes of Stochastic Processes . . . . . . . . . . . 10
1.3 Brownian Motion . . . . . . . . . . . . . . . . . . . . . . . 15
1.4 Black and Scholes Financial Market Model . . . . . . . . 17
2 Financial Market Theory 20
2.1 Financial Markets . . . . . . . . . . . . . . . . . . . . . . . 20
2.2 Arbitrage . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3 Martingale Measures . . . . . . . . . . . . . . . . . . . . . 25
2.4 Options and Contingent Claims . . . . . . . . . . . . . . . 34
2.5 Hedging and Completeness . . . . . . . . . . . . . . . . . 36
2.6 Pricing of Contingent Claims . . . . . . . . . . . . . . . . 38
2.7 The Black-Scholes Formula . . . . . . . . . . . . . . . . . 42
2.8 Why is the Black-Scholes model not good enough ? . . . . 46
3 Stochastic Integration 48
3.1 Semi-martingales . . . . . . . . . . . . . . . . . . . . . . . 48
3.2 The stochastic Integral . . . . . . . . . . . . . . . . . . . . 55
3.3 Quadratic Variation of a Semi-martingale . . . . . . . . . 65
3.4 The Ito Formula . . . . . . . . . . . . . . . . . . . . . . . . 71
3.5 The Girsanov Theorem . . . . . . . . . . . . . . . . . . . . 76
3.6 The Stochastic Integral for predictable Processes . . . . . 81
3.7 The Martingale Representation Theorem . . . . . . . . . 84
1
4 Explicit Financial Market Models 85
4.1 The generalized Black Scholes Model . . . . . . . . . . . . 85
4.2 A simple stochastic Volatility Model . . . . . . . . . . . . 93
4.3 Stochastic Volatility Model . . . . . . . . . . . . . . . . . . 95
4.4 The Poisson Market Model . . . . . . . . . . . . . . . . . . 100
5 Portfolio Optimization 105
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 105
5.2 The Martingale Method . . . . . . . . . . . . . . . . . . . 108
5.3 The stochastic Control Approach . . . . . . . . . . . . . . 119
2
Introduction
Mathematical Finance is the mathematical theory of nancial markets.
It tries to develop theoretical models, that can be used by practition-
ers to evaluate certain data from real nancial markets. A model
cannot be right or wrong, it can only be good or bad ( for practical use
). Even bad models can be good for theoretical insight.
Content of the lecture :
Introduction to continuous time nancial market models.
We will give precise mathematical denitions, what we do understand
under a nancial market, until this let us think of a nancial market as
some place where people can buy or sell nancial derivatives.
During the lecture we will give various examples for nancial deriva-
tives. The following denition has been taken from [Hull] :
A nancial derivative is a nancial contract, whose value at expire is
determined by the prices of the underlying nancial assets ( here we
mean Stocks and Bonds ).
We will treat options, futures, forwards, bonds etc. It is not necessary
to have nancial background.
3
During the course we will work with methods from
Probability theory, Stochastic Analysis and Partial Differ-
ential Equations.
The Stochastic Analysis and Partial Differential Equations methods
are part of the course, the Probability Theory methods should be known
from courses like Probability Theory and Prama Stochastik.
4
Chapter 1
Stochastic Processes in
Continuous Time
Given the present, the price S
t
of a certain stock at some future time t
is not known. We cannot look into the future. Hence we consider this
price as a random variable. In fact we have a whole family of random
variables S
t
, for every future time t. Lets assume, that the random
variables S
t
are dened on a complete probability space (, F, P), now
it is time 0 , 0 t < and the -algebra F contains all possible infor-
mation. Choosing sub -algebras F
t
F containing all the information
up to time t, it is natural to assume that S
t
is F
t
measurable, that is
the stock price S
t
at time t only depends on the past, not on the future.
We say that (S
t
)
t[0,)
is F
t
adapted and that S
t
is a stochastic pro-
cess. Throughout this chapter we assume that (, F, P) is a complete
probability space. If X is a topological space, then we think of X as a
measurable space with its associated Borel -algebra which we denote
as B(X).
1.1 Filtrations and Stochastic Processes
Let us denote with I any subset of R.
Denition 1.1.1. A family (F
t
)
tI
of sub -algebras of F such that F
s

5
F
t
whenever s < t is called a ltration of F.
Denition 1.1.2. A family (X
t
, F
t
)
tI
consisting of F
t
-measurable R
n
-
valued randomvariables X
t
on (, F, P) and a Filtration (F
t
)
tI
is called
an n-dimensional stochastic process.
The case where I = N corresponds to stochastic processes in discrete
time ( see Probability Theory, chapter 19 ). Since this section is devoted
to stochastic processes in continuous time, from now on we think of I
as a connected subinterval of R
0
.
Often we just speak of the stochastic process X
t
, if the reference to the
ltration (F
t
)
tI
is clear. Also we say that (X
t
)
tI
is (F
t
)
tI
adapted. If
no ltration is given, we mean the stochastic process (X
t
, F
t
)
tI
where
F
t
= F
X
t
= (X
s
|s I, 0 s t)
is the -algebra generated by the random variables X
s
up to time t.
Given a stochastic process (X
t
, F
t
)
tI
, we can consider it as function
of two variables
X : I R
n
, (, t) X
t
().
On I we have the product -algebra F B(I) and for
I
t
:= {s I|s t} (1.1)
we have the product -algebras F
t
B(I
t
).
Denition 1.1.3. The stochastic process (X
t
, F
t
)
tI
is called measur-
able if the associated map X : I R
n
from (1.1) is (FB(I))/B(R
n
)
measurable. It is called progressively measurable, if for all t I the
restriction of X to I
t
is (F
t
B(I
t
)/B(R
n
) measurable.
In this course we will only consider measurable processes. So from
now on, if we speak of a stochastic process, we mean a measurable
6
stochastic process.
Working with stochastic processes the following space is of fundamen-
tal importance :
(R
n
)
I
:= Map(I, R
n
) = { : I R
n
} (1.2)
i.e. the maps from I to R
n
. For any t I we have the so called evalua-
tion map
ev
t
: (R
n
)
I
R
n
(t)
Denition 1.1.4. The -algebra on (R
n
)
I
cyl
:= (ev
s
|s I)
generated by the evaluation maps is called the -algebra of Borel
cylinder sets.
F
t
:=
cyl,t
:= (ev
s
|s I, s t)
denes a ltration of
cyl
. Whenever we consider (R
n
)
I
as a measur-
able space, we consider it together with this -algebra and this ltra-
tion.
The space (R
n
)
I
has some important subspaces :
C(I, R
n
) := { : I R
n
| is continuous } (1.3)
C
+
(I, R
n
) := { : I R
n
| is right-continuous } (1.4)
C
(I, R
n
) := { : I R
n
| is left-continuous } (1.5)
7
Always, we consider these spaces as measurable spaces together
with the associated -algebras of Borel cylinder sets and their l-
tration. These are dened as the corresponding restrictions of the -
algebras from Denition 1.1.4. to these spaces.
In addition to (1.1) we can also consider the stochastic process (X
t
, F
t
)
tI
as a map
X : (R
n
)
I
, (t X
t
()). (1.6)
Exercise 1.1.1. Show the map in (1.6) is F/
cyl
measurable.
Denition 1.1.5. The stochastic process (X
t
, F
t
)
tI
has continuous
paths if Im(X) C(I, R
n
), where Im(X) denotes the image of X. In
this case, we often just say X is continuous. We say X has contin-
uous paths almost surely if P{ |X() C(I, R
n
)} = 1. X is
called right-continuous if Im(X) C
+
(I, R
n
) and in the same way as
before has right-continuous paths almost surely if P{ |X()
C
+
(I, R
n
)} = 1
Via the map (1.6) under consideration of Exercise 1.1.1 the proba-
bility measure P on (, F) induces a probability measure P
X
on (R
n
)
I
which is also called distribution of X.
Denition 1.1.6. Let (X
i
t
, F
i
t
)
tI
i = 1, 2 be two stochastic processes
dened on two not necessarily identical probability spaces (
i
, F
i
, P
i
).
Then (X
1
t
)
tI
and (X
2
t
)
tI
are called equivalent if they have the same
distribution, that is P
X
1
1
= P
X
2
2
. Often we write (X
1
t
)
tI
(X
2
t
)
tI
.
Clearly equivalence of stochastic processes is an equivalence rela-
tion. For a stochastic process (X
t
, F
t
)
tI
consider the stochastic pro-
cess (ev
X
t
)
tI
on ((R
n
)
I
,
cyl
, P
X
) given by the evaluation maps. Then
(X
t
)
tI
(ev
X
t
)
tI
and (ev
X
t
)
tI
is called the canonical representation
of (X
t
)
tI
. If (X
t
, F
t
)
tI
has continuous paths then the stochastic pro-
cess denoted by the same symbol (ev
X
t
)
tI
on (C(I, R
N
),
cyl
, P
X
) is called
8
the canonical continuous representation and clearly again (X
t
)
tI

(ev
X
t
)
tI
.
Conclusion : If one is only interested in stochastic pro-
cesses up to equivalence one can always think of the underly-
ing probability space as ((R
n
)
I
,
cyl
, P
X
) or (C(I, R
N
),
cyl
, P
X
) in
the continuous case. What characterizes the stochastic process
is the probability measure P
X
.
In some cases though, equivalence in the sense of Denition 1.1.6
is not strong enough. The following denitions give stricter criteria on
how to differentiate between stochastic processes.
t
, F
t
)
tI
(Y
t
, G
t
)
tI
be two stochastic processes on
the same probability space (, F, P). Then (Y
t
, G
t
)
tI
is called a modi-
cation of (X
t
, F
t
)
tI
if
P{ |X
t
() = Y
t
() } = 1 , t I.
(X
t
, F
t
)
tI
and (Y
t
, G
t
)
tI
are called indistinguishable, if
P{ |X
t
() = Y
t
() t I } = 1.
The following two exercises are good to understand the relation-
ships between equivalence, modication and indistinguishability.
Exercise 1.1.2. Under the assumptions of Denition 1.1.7. Prove that
the following implications hold :
(X
t
, F
t
)
tI
and (Y
t
, G
t
)
tI
indistinguishable
(X
t
, F
t
)
tI
and (Y
t
, G
t
)
tI
are modications of each other
(X
t
, F
t
)
tI
and (Y
t
, G
t
)
tI
are equivalent.
9
Give examples for the fact, that in general the inverse implication
does not hold. But :
Exercise 1.1.3. If in addition to the assumptions of Denition 1.1.7
we assume that (X
t
, F
t
)
tI
and (Y
t
, G
t
)
tI
are continuous and (Y
t
, G
t
)
tI
is a modication of (X
t
, F
t
)
tI
, then (X
t
, F
t
)
tI
and (Y
t
, G
t
)
tI
are indis-
tinguishable. How can the last two conditions be relaxed such that the
implication still holds ?
The last two denitions in this section concern the underlying l-
trations.
Denition 1.1.8. A ltration (F
t
)
tI
is called right-continuous if
F
t
= F
t+
:=
sI,s>t
F
s
. (1.7)
It is called left-continuous if
F
t
= F
t
:=
_
sI,s<t
F
s
. (1.8)
Denition 1.1.9. Let I = [0, T] or I = [0, ]. A ltration (F
t
)
tI
satises
the usual conditions if it is right continuous and F
0
contains all P
null-sets of F.
Exercise 1.1.4. Let I = [0, ). Showthe ltration (
cyl,t
)
tI
of (C(I, R
n
),
cyl
)
is right-continuous as well as left-continuous.
1.2 Special Classes of Stochastic Processes
There are two very important classes of stochastic processes, one is
martingales the other is Markov processes, and there is the most im-
portant ( continuous ) stochastic process Brownian motion which be-
longs to both classes and will be treated in the next section. So far,
let (X
t
, F
t
)
tI
be a stochastic process dened on a complete probability
space (, F, P).
10
Denition 1.2.1. If E(|X
t
|) < t I then (X
t
, F
t
)
tI
is called a
1. martingale if s t we have E(X
t
|F
s
) = X
s
2. supermartingale if s t we have E(X
t
|F
s
) X
s
3. submartingale if s t we have E(X
t
|F
s
) X
s
During the course we will see many examples of martingales as well
as sub- and supermartingales.
Exercise 1.2.1. Let (X
t
, F
t
)
tI
be a stochastic process with independent
increments, that means X
t
X
s
is independent of F
u
u s.Consider
the function : I R, (t) = E(X
t
). Give conditions for that imply
X
t
is a martingale or submartingale or supermartingale.
Exercise 1.2.2. Let Y be a random variable dened on a complete prob-
ability space (, F, P) such that E(|Y |) < and let (F
t
)
tI
be a ltration
of F. Dene
X
t
:= E(Y |F
t
) , t I.
Show (X
t
, F
t
)
tI
is a martingale.
For stochastic integration a class slightly bigger than martingales
will play an important role. This class is called local martingales. To
dene it, we rst need to dene what we mean by a stopping time :
Denition 1.2.2. A stopping time with respect to a ltration (F
t
)
tI
is an F measurable random variable : I {} such that for all
t I we have
1
(I
t
) F
t
. A stopping time is called nite if ()
I. A stopping time is called bounded if there exists T
I such that
P{|() T
} = 1.
The following exercises leads to many examples of stopping times.
Exercise 1.2.3. Let (X
t
, F
t
)
tI
be a continuous stochastic process with
11
values in R
n
and let A R
n
be a closed subset. Then
: R
() := inf{t I|X
t
() A}
is a stopping time with respect to the ltration (F
t
)
tI
.
Exercise 1.2.4. Let
1
resp.
2
be stopping times on (, F, P) with respect
to the ltrations (F
t
)
tI
resp. (G
t
)
tI
. Let F
t
G
t
= (F
t
, G
t
). Then
1

2
: R
(
1

2
)() = min(
1
(),
2
())
is a stopping time with respect to the ltration (F
t
G
t
)
tI
.
Given a stochastic process and a stopping time we can dene a new
stochastic process by stopping the old one. In case the stopping time
is nite, we can dene a new random variable. The denitions are as
follows :
t
, F
t
)
tI
be a stochastic process and a stop-
ping time with respect to (F
t
)
tI
. Then we dene a new stochastic pro-
cess (X
t
)
tI
with respect to the same ltration (F
t
)
tI
as
X
t
() =
_
X
t
() , t ()
X
()
() , t > ()
(1.10)
If is nite, we dene a random variable X
on (, F, P) as
X
() := X
()
(). (1.12)
Also we can dene a new -algebra :
12
Denition 1.2.4. Let be a stopping time with respect to the ltration
(F
t
)
tI
. Then
F
:= {A F|A
1
(I
t
) F
t
t I} (1.14)
is called the -algebra of events up to time .
This is indeed a -algebra. The following is a generalization of The-
orem 19.3 in the Probability Theory lecture.
Theorem 1.2.1. Optional Sampling Theorem Let (X
t
, F
t
)
tI
be a
right-continuous martingale and
1
,
2
be bounded stopping times with
respect to (F
t
)
tI
. Let us assume that
1

2
P-almost sure. Then
X
1
= E(X
2
|F
1
). (1.16)
If (X
t
)
tI
is only a submartingale ( supermartingale ) then (1.14) is
still valid with = replaced by ( ).
Denition 1.2.5. A stochastic process (X
t
, F
t
)
tI
is called a local mar-
tingale if there exists an almost surely nondecreasing sequence of stop-
ping times
n
, n N with respect to (F
t
)
tI
converging to almost sure,
such that (X
n
t
, F
t
)
tI
is a martingale for all n N.
The class of local martingales contains the class of martingales.
This follows from the Optional Sampling theorem. We leave the de-
tails as an exercise.
Exercise 1.2.5. Show that every martingale is a local martingale.
A relation between local martingales and supermartingales is es-
tablished by the following :
Exercise 1.2.6. A local martingale (X
t
, F
t
)
tI
which is bounded below
is a supermartingale. Bounded below means that there exists c R such
that P{|X
t
() c , t I} = 1.
13
In plain words, the martingale property means, that the process,
given the present time s has no tendency in future times t s, that is
the average over all future possible states of X
t
gives just the present
state X
s
. In difference to this, the Markov property, which will follow
in the next denition means that the process has no memory, that is
the average of X
t
knowing the past is the same as the average of X
t
knowing the present. More precise :
Denition 1.2.6. (X
t
, F
t
)
tI
is called a Markov process if
E(X
t
|F
s
) = E(X
t
|(X
s
)) 0 s t < (1.18)
Sometimes the Markov property (1.16) is referred to as the elemen-
tary Markov property, in contrast to the strong Markov property which
will be dened in a later section. So far, if we just say Markov we
mean (1.16). Markov processes will arise naturally as the solutions of
certain stochastic differential equations. Also Exercise 1.2.1 provides
examples for Markov processes.
Besides martingales and Markov processes there is another class of
processes which will occur from time to time in this text. It is the class
of simple processes. This class is not so important for its own, but is
important since its construction is simple and many other processes
can be achieved as limits of processes from this class. Because of the
simple construction, they are called simple processes.
Denition 1.2.7. An n-dimensional stochastic process (X
t
, F
t
)
t[0,T]
is
called simple with respect to the ltration (F
t
)
t[0,T]
if there exist 0 =
t
0
< t
1
< ... < T
m
= T and
i
: R
n
such that
0
is F
0
,
i
is F
t
i1
and
X
t
() =
0
() 1
{0}
+
m
i=1
i
() 1
(t
i1
,t
i
]
, t, . (1.19)
In case that in the denition above m = 0, we call (X
t
, F
t
)
t[0,T]
a
constant stochastic process.
14
1.3 Brownian Motion
Brownian Motion is widely considered as the most important ( continu-
ous ) stochastic process. In this section we will give a short introduction
into Brownian motion but we wont give a proof for its existence. There
are many nice proofs available in the literature, but everyone of them
gets technical at a certain point. So as in most courses about Mathe-
matical Finance, we will keep the proof of existence for a special course
in stochastic analysis.
Denition 1.3.1. Let (W
t
, F
t
)
t[0,)
be an R-valued continuous stochas-
tic process on (, F, P). Then (W
t
, F
t
)
t[0,)
is called a standard Brow-
nian motion if
1. W
0
= 0 a.s.
2. W
t
W
s
N(0, t s)
3. W
t
W
s
independent ofF
s
.
An R
n
valued process W
t
is called an n-dimensional Brownian mo-
tion with initial value x R
n
if
W
t
= x + (W
1
t
, ..., W
n
t
), t [0, )
where W
i
t
are independent standard Brownian motions.
It is not true that given a complete probability space (, F, P), there
exists a standard Brownian motion or even an n-dimensional Brownian
motion on this probability space, sometimes the underlying probability
space (, F, P) is just too small. Nevertheless the following is true :
Proposition 1.3.1. There is a complete probability space (, F, P) such
that there exists a standard Brownian motion (W
t
, F
t
)
t[0,)
on (, F, P).
15
This process is unique up to equivalence of stochastic processes ( see
Denition 1.1.6 )
Brownian motion can be used to build a large variety of martingales.
One more or less simple way to to do this is given by the following
exercise.
Exercise 1.3.1. Let (W
t
, F
t
)
t[0,)
be a standard Brownian motion, and
R a real number. For all t [0, ) dene
X
t
= e
W
t
1
2
2
t
.
Then (X
t
, F
t
)
t[0,)
is a ( continuous ) martingale.
In the following we will take a closer look at the canonical contin-
uous representation (ev
W
t
)
t[0,)
of any standard Brownian motion W (
see page 6 ) dened on (C([0, ), R),
cyl
, P
W
). The measure P
W
is called
the Wiener measure. Sometimes the Brownian motion W is also called
Wiener process, hence the notation W.
For t > 0 consider the density functions ( also called Gaussian kernels )
of the standard normal distributions N(0, t) dened on R as
p(t, x) =
1
2t
e
x
2
2t
The following proposition characterizes the Wiener measure.
Proposition 1.3.2. The Wiener measure P
W
on (C([0, ), R),
cyl
) is the
unique measure which satises m N and choices 0 < t
1
< t
2
< ... <
t
m
< and arbitrary Borel sets A
i
B(R),1 i m
P
W
({ C([0, ), R
n
)| (t
1
) A
1
, ..., (t
m
) A
m
}) =
_
A
1
p(t
1
, x
1
)dx
1
_
A
2
p(t
2
t
1
, x
2
x
1
)dx
2
...
_
A
m
p(t
m
t
m1
, x
m
x
m1
)dx
m
.
16
One proof of the existence of Brownian motion goes like this : Use
the denition in Proposition 1.3.2 to dene a measure on ((R
n
)
[0,)
,
cyl
).
For this one needs the Daniell/Kolmogorov extension theorem( [Karatzas/Shreve],
page 50 ). The result is a measure P
W
and a process on ((R
n
)
[0,)
,
cyl
, P
W
)
which is given by evaluation and satises all the conditions in Deni-
tion 1.3.1 except the continuity. Then one uses the Kolomogorov/Centsov
theorem ( [Karatzas/Shreve], page 53 ) to show that this process has a
continuous modication. This leads to a Brownian motion W
t
.
Exercise 1.3.2. Prove Proposition 1.3.2. Hint : Consider the joint den-
sity of (W
t
0
, W
t
1
W
t
0
, ..., W
t
m
W
t
m1
) of any standard Brownian motion
W and use the density transformation formula ( Theorem 11.4. in the
Probability Theory Lecture) applied on the transformation g(x
1
, ..., x
m
) :=
(x
1
, x
1
+ x
2
, ..., x
1
+ x
2
+ ... + x
m
).
1.4 Black and Scholes Financial Market Model
In this section we will introduce into the standard Black-Scholes model
which describes the motion of a stock price and bond. First consider the
following situation. At time t = 0 you put S
0
0
units of money onto your
bank account and the bank has a constant deterministic interest rate
r > 0. If after time t > 0 you want your money back, the bank pays you
S
0
t
= S
0
0
e
rt
. (1.20)
Let us consider the logarithm of this
ln(S
0
t
) = ln(S
0
0
) + rt. (1.21)
So far there is no random, although in reality the interest rate is far
from being constant in time and also nondeterministic. Now consider
the price S
1
t
of a stock at time t > 0 and let S
1
denote the price at time
t = 0. If we look at equation (1.19) the following approach seems to be
natural
17
ln(S
1
t
) = ln(S
1
0
) +
bt + random. (1.22)
This equation means, that in addition to the linear deterministic
trend in equation (1.19) we have some random uctuation. This ran-
domuctuation depends on the time t, hence we think of it as a stochas-
tic process and denote it in the following with w
t
. Since we know the
stock price S
1
t
at time t = 0 there is no random at time t = 0 hence we
can assume that
w
0
= 0 a.s . (1.23)
Furthermore we assume that w
t
is composed of many similar small
perturbations with no drift resulting from tiny little random events (
events in the world we cannot foresee ) all of which average to zero.
The farther we look into the future, the more of these tiny little ran-
dom events could happen, the bigger the variance of w
t
is. We assume
that the number of these tiny little events that can happen in the time
interval [0, t] is proportional to t. Hence by the central limit theorem
an obvious choice for w
t
is
w
t
N(0,
2
t).
Since the number of tiny little events which can happen between
time s and time t is proportional to t s we assume
w
t
w
s
N(0,
2
(t s)) , s < t. (1.24)
Also we assume once we know the stock price S
1
s
at some time s > 0
the future development S
1
t
for t > s does not depend on the stock prices
S
1
u
for u < s before s. This translates to
w
t
w
s
is independent of w
u
, u < s. (1.25)
By comparison of equations (1.14)-(1.16) with ( 1 - 3 ) in denition
18
1.3.1 we must choose w
t
= W
t
, where (W
t
)
t[0,)
is a Brownian motion.
We get the following equation for the stock price S
1
t
S
1
t
= S
1
0
e
bt+W
t
For applications it is useful to re-scale this equation using b :=

b +
1
2
2
t and writing
S
1
t
= S
1
0
e
(b
1
2
2
)t+W
t
The model of a nancial market consisting of one bond and one stock
modeled as in equations (1.18) and (1.24) together with the appropriate
set of trading strategies is called the standard Black-Scholes model.
The valuation formula for Europeans call options in this model is called
the Black-Scholes formula and was nally awarded with the Nobel-
Prize in economics in 1997 for Merton and Scholes ( Black was already
dead at this time ).
Exercise 1.4.1. Let S
t
= S
0
e
(b
1
2
2
)t+W
t
denote the price of a stock in
the standard Black-Scholes model. Compute the expectation E(S
t
) and
variance var(S
t
).
19
Chapter 2
Financial Market Theory
In this section we will introduce into the theory of nancial markets.
The treatment here is as general as possible. At the end of this chap-
ter we will consider the standard Black-Scholes model and derive the
Black-Scholes formula for the valuation of an European call option.
Throughout this chapter (, F, P) denotes a complete probability space
and I = [0, T] for T > 0.
2.1 Financial Markets
In the Introduction we gave a naive denition of what we think of a
nancial market is :
some place, where people can buy or sell nancial
derivatives.
Hence we have to model two things. First the nancial derivatives,
second the actions ( buy and sell ) of the people ( so called traders )
who take part in the nancial market. The actions of the traders are
henceforth called trading strategies. Precisely :
20
Denition 2.1.1. A nancial market is a pair
M
m,n
= ((X
t
, F
t
)
tI
, ) (2.1)
consisting of
1. an R
n+1
valued stochastic process (X
t
, F
t
) dened on (, F, P), such
that (F
t
)
tI
satises the usual conditions and F
0
= (, , P
null-sets)
2. a set which consists of R
m+1
valued stochastic processes (
t
)
tI
adapted to the same ltration (F
t
)
tI
.
The components X
0
t
, ..., X
m
t
of X
t
are called tradeable components,
the components X
m+1
t
, ..., X
n
t
are called nontradeable. We denote the
tradeable part of X
t
with X
tr
t
= (X
0
t
, ..., X
m
t
). The elements of are
called trading strategies.
The interpretation of Denition 2.1.1 is as follows : We think of
the tradeable components as the evolution in time of assets which are
traded at the nancial market ( for example stocks or other nancial
derivatives ) and of the nontradeable components as additional ( non-
tradeable ! ) parameter, describing the market. The set is to be
interpreted as the set of allowed trading strategies. Sometimes is
also called the portfolio process. For a trading strategy the i-th
component
i
t
denotes the amount of units of the i-th nancial deriva-
tive owned by the trader at time t. In some cases we will assume that
carries some algebraic or topological structure, for example vector
space ( cone ), topological vector space, L
2
-process etc. We will spec-
ify this structure, when we really need it. The assumption on the 0-th
ltration F
0
makes sure, that any F
0
measurable random variable on
(, F) is constant almost sure. This describes the situation that at time
t = 0 we know completely whats going on.
Denition 2.1.2. Let M
m,n
= ((X
t
, F
t
)
tI
, ) be a nancial market and
let = (
t
)
tI
be a trading strategy, then we dene the correspond-
ing value process as
21
V
t
() =
t
X
tr
t
=
m
i=0
i
t
X
i
t
. (2.2)
The value process gives us the worth of our portfolio at time t.
In some cases it is helpful to consider the (X
i
t
)
tI
in units of another
stochastic process (N
t
, F
t
)
tI
This leads to the notion of a numeraire.
Denition 2.1.3. Consider a nancial market M
m,n
= ((X
t
, F
t
)
tI
, ).
A stochastic process (N
t
, F
t
)
tI
is called a numeraire if it is strictly
positive almost sure, that is
P{|N
t
() > 0} = 1 , t I. (2.3)
The numeraire is called a market numeraire if there exists a trad-
ing strategy such that (N
t
)
tI
and (V
t
())
tI
are indistinguishable.
Given a numeraire (N
t
)
tI
we denote with
X
t
:=
X
t
N
t
. (2.4)
the discounted price process and for any
V
t
() :=
V
t
()
N
t
. (2.5)
the discounted value process.
As an example of a market numeraire one could think of a nancial
market where (N
t
)
t[0,)
= (X
0
t
)
t[0,)
given by X
0
t
= e
rt
represents a
bank account with deterministic interest rate r > 0. From now on,
we will assume that there exists a market numeraire in the market
M
m,n
= ((X
t
, F
t
)
tI
, ) and that it is given by the component of X
0
t
( the
last thing is not really a restriction, think about it ! ) So in our case the
component

X
0
t
of the discounted price process is always constant equal
to 1.
22
2.2 Arbitrage
In a nancial market a risk free opportunity to make money is called
an arbitrage. In our setting :
Denition 2.2.1. An arbitrage in a nancial market M
m,n
= ((X
t
, F
t
)
tI
, )
is a trading strategy such that
V
0
() = 0 almost sure
V
T
() 0 almost sure
P(V
T
() > 0) > 0
M
m,n
= ((X
t
, F
t
)
tI
, ) is called arbitrage free if there exist no ar-
bitrages in .
Lemma 2.2.1. Let M
m,n
= ((X
t
, F
t
)
tI
, ) be a nancial market such
that contains all constant positive R
m+1
-valued processes and carries
the algebraic structure of a cone. Then M
m,n
is arbitrage free if there is
no trading strategy which satises
V
0
() < 0 P-almost sure
V
T
() 0 P-almost sure
Proof. Assume that M
m,n
is arbitrage free and there is a trading strat-
egy such that V
0
() < 0 and V
T
() 0 P-almost sure. Since
V
0
() is F
0
we know that it is constant almost sure. Let c denote this
constant. Then c < 0 and also c =
c
X
0
0
< 0. Using the assumption on
we can dene a new trading strategy as
=
_
_
_
_
_
_
_
_
0
c
m
_
_
_
_
_
_
_
_
.
23
It follows from the conditions on in Denition 2.1.1 that is again
a trading strategy, i.e. . We have V
0
( ) = V
0
() cX
0
0
= V
0
()
V
0
() = 0 almost sure and
V
T
( ) = V
T
() cX
0
T
.
Since V
T
() 0 , c < 0 and the numeraire X
0
T
> 0 almost sure, we
have V
T
(( ) > 0 almost sure. Hence P{V
T
(( ) > 0} = 1 > 0 and is an
arbitrage in M. But this is a contradiction to the assumption that M
is arbitrage free, hence such a can not exist in .
Ideally a nancial market is arbitrage free, but sometimes this is
not the case. If arbitrages exist in a nancial market, then mostly only
for a short period of time. This is because if so, there are probably peo-
ple who want to exploit the arbitrage and by exploitation of the arbi-
trage the arbitrage possibility vanishes. One can say, that the nancial
market has an arbitrage free equilibrium but sometimes differs from
that equilibrium. The main implication of the no arbitrage condition
in this general setup is given by the following proposition. It is often
called the No Arbitrage Principle
Proposition 2.2.1. No Arbitrage Principle 1 Let , be trading
strategies in an arbitrage free nancial market M
m,n
= ((X
t
, F
t
)
tI
, )
such that V
T
() = V
T
() P-almost sure. If is a vector space and con-
tains all constant positive processes, then V
0
() = V
0
() P-almost sure.
Proof. Since V
0
() and V
0
() are F
0
measurable, they are constant al-
most sure. Let us assume V
0
() = V
0
(), then w.l.o.g. V
0
() < V
0
()
almost sure. Now consider the trading strategy . Then we
have V
t
() = V
t
()V
t
() for all t [0, T]. In particular V
0
() < 0
and V
T
() = 0 almost sure. From Lemma 2.2.1 it follows that M
m,n
is not arbitrage free, which is a contradiction to the assumption that it
is arbitrage free.
24
Making more assumptions on the vector space of trading strate-
gies, one can indeed prove, that the two value processes V
t
() and V
t
()
are indistinguishable.
2.3 Martingale Measures
Whereas the measure P on the underlying measurable space (, F) is
somehow articial and can be thought of as a subjective evaluation of
the state of the nancial market ( for example from the point of view
of one distinguished trader ) martingale measures can be interpreted
as an objective evaluation of the market. Therefore martingale mea-
sures are often used to price certain nancial derivatives in a way, that
no one can take advantage by trading these derivatives. In the right
setup existence of martingale measures implies nonexistence of arbi-
trage. We will use martingale measures in the next section for the
pricing of options and contingent claims.
Denition 2.3.1. A probability measure P
on (, F, P) is called an
equivalent martingale measure for the nancial market M
m,n
=
((X
t
, F
t
)
tI
, ), if
1. P and P
have the same null-sets and

2. for any tradeable component X
i
t
the discounted process (

X
i
t
)
t[0,T]
is a local P
martingale.
We denote the set of equivalent martingale measure for M
m,n
as
P(M
m,n
).
Condition (2.) above is the same as saying (

X
tr
t
) is a local P
-martingale.
The following denition seems to be very technical but in fact is quite
useful when working with many measures at the same time.
Denition 2.3.2. Let g : R be a real valued F measurable func-
tion. Then g is called universally integrable in the nancial mar-
25
ket M
m,n
= ((X
t
, F
t
)
tI
, ) if for all equivalent martingale measures
P
P(M
m,n
) we have
E
P
(|g|) =
_
|g|dP
< . (2.6)
Often we just speak of a universally integrable function, though in
fact we mean universally integrable in the nancial market M
m,n
=
((X
t
, F
t
)
tI
, ), when the context is clear.
To establish the connection between martingale measures and ar-
bitrage, we must restrict ourself to a special class of trading strategies
which is from the real world nancial market point of view very nat-
ural. Let us assume that at time t
0
a trader enters a market ( only
consisting of tradeable assets ) and buys assets according to
t
0
. Then
the worth of his portfolio at time t
0
is
V
t
0
() =
t
0
X
t
0
.
Now he chooses not to change anything with his portfolio until time t
1
.
Then his portfolio at time t
1
still consists of
t
0
and hence has worth
V
t
1
() =
t
0
X
t
1
. (2.7)
At time t
1
though he chooses to rearrange his portfolio and reinvest
all the money from V
t
1
() according to
t
1
. After this rearrangement
the worth of his portfolio calculates as
V
t
1
() =
t
1
X
t
1
. (2.8)
Since he only used the money from V
t
1
() and didnt consume any of
the money the two values from ( 2.7 ) and ( 2.8 ) must coincide. Hence
we have
26
(
t
1

t
0
) X
t
1
= 0, (2.9)
or equivalently after division by the numeraire
(
t
1

t
0
)

X
t
1
= 0. (2.10)
In the general framework of a nancial market M
m,n
trading at any
time is allowed. For this purpose consider for any t [0, T] partitions
Z(t) : 0 = t
0
< t
1
< ... < t
k
= t and dene
Z(t)
(d)

X =
k
i=0
(
t
i+1

t
i
)

X
tr
t
i+1
Z(t)
d

X =
k
i=0
t
i
(

X
tr
t
i+1

X
tr
t
i
).
The concept above can then be generalized as follows : For any t
[0, T] and any sequence of partitions Z
l
(t) such that lim
l
|Z
l
(t)| = 0
we have
lim
l
Z
l
(t)
(d)

X = 0 a.s . (2.11)
or equivalently since
V
t
() =
t

X
t
=

V
0
() +
k1
i=0
(
t
i+1

X
tr
t
i+1

t
i

X
tr
t
i
)
=

V
0
() +
Z(t)
(d)

X +
Z(t)
d

X,
V
0
() + lim
n
Z
n
(t)
d

X =

V
t
() a.s . (2.12)
This motivates the following denition of a self nancing trading
strategy.
27
m,n
= ((X
t
, F
t
)
tI
a trading strategy .
1. is called self nancing if for any t [0, T] and any sequence
of partitions Z
l
(t) such that lim
l
|Z
l
(t)| = 0 we have
V
0
() + lim
l
Z
l
(t)
d

X =

V
t
() a.s . (2.13)
2. is called strictly self nancing if the convergence in ( 2.13 ) is
locally dominated by universally integrable functions. This means
that there are universally integrable functions g
j
, a sequence of
stopping times
j
j N such that
j

j+1
, lim
j
j
= P-
almost sure and
|
Z
l
(t)
j
d

X
j
| g
j
. (2.14)
We denote the set of strictly self nancing trading strategies in with
s
The functions g
j
in ( 2.14 ) may depend on the sequence of partitions
Z
l
(t). From the real world point of view the condition ( 2.14 ) is not so
restrictive, since rst of all, there is only nitely much money in the
world at all, second there are nancial ( and other ) restriction to the
behavior of each individual trader.
Proposition 2.3.1. Let be a strictly self nancing trading strat-
egy in a nancial market M
m,n
= ((X
t
, F
t
)
tI
, ) and P
P(M), then
the discounted value process

V
t
() follows a local P
-martingale.
28
Proof. Since the nontradeable components of (X
t
)
t[0,T]
have no effect
on the value process we can assume that m = n, i.e. there are no
nontradeable components. Also we can assume that

V
0
() = 0. Let
be strictly self nancing and
j
be a sequence of stopping times
as in (2.14). Let
j
be another sequence of stopping times as in Def-
inition 1.2.5 such that corresponding to Denition 2.3.1 the stopped
discounted price processes

X
j
are P
-martingales. Let us dene new

stopping times
j
=
j
. Then

X
j
are still P
- martingales ( use the

optional sampling theorem ) and (2.14) is satised with
j
instead of
j
.
For t [0, T] it follows from the theorem about dominated convergence
and the strictly self nancing condition that
E
P
(|
j
t
()|) E
P
(g
j
) <
where g
j
is as in (2.14). Now let 0 s t T and consider se-
quences of partitions
Z
l
(s) : 0 = t
0
< t
1
< ... < t
k
l
= s
Z
l
(s, t) : s = t
k
l
< t
k
l
+1
< ... < t
k
l
= t
Z
l
(t) : 0 = t
0
< t
1
< ... < t
k
l
= s < t
k
l
+1
... < t
k
l
= t
such that lim
l
|Z
l
(t)| = 0 and hence also lim
l
|Z
l
(s)| = 0 and
lim
l
|Z
l
(s, t)| = 0. From the martingale property of

X
j
it follows that
E
P
(

X
j
t
i+1

X
j
t
i
|F
t
i
) = 0, i.
Now let i k
l
, then t
i
s and we have
E
P
(
j
t
i
(

X
j
t
i+1

X
j
t
i
|F
s
) = E
P
(
j
t
i
E
P
(

X
j
t
i+1

X
j
t
i
|F
t
i
)
. .
=0
|F
s
) = 0.
This implies
29
E
P
(
Z
l
(s,t)
j
d

X
j
|F
s
) = 0, l.
On the other side
Z
l
(s)
j
d

X
j
is by denition F
s
measurable, hence
E
P
(
Z
l
(s)
j
d

X
j
|F
s
) =
Z
l
(s)
j
d

X
j
, l.
Since
Z
l
(t)
j
d

X
j
=
Z
l
(s)
j
d

X
j
+
Z
l
(s,t)
j
d

X
j
we conclude
E
P
(
Z
l
(t)
j
d

X
j
|F
s
) =
Z
l
(s)
j
d

X
j
, l.
Building the limit on both sides for l it follows again from the
dominated convergence theorem and the strictly self nancing condi-
tion that
E
P
(
j
t
()|F
s
) =

V
j
s
()
which shows that

V
j
t
() follows a P
-martingale for all j. By the

properties of the sequence of stopping times
j
it follows that

V
t
() fol-
lows a local P
-martingale.
Though the self nancing and strictly self nancing condition seems
to be very natural, it is however not so easy to determine whether a
trading strategy is self nancing ( strictly self nancing ) or
not. The following Exercise shows that in the case of simple trading
strategies this is much easier.
Exercise 2.3.1. Let M
m,n
= ((X
t
, F
t
)
t[0,T]
, ) be a nancial market
such that only consists of simple processes ( see Denition 1.2.7 ).
Showthat if is given as
t
() =
0
()1
{0}
+
m
i=1
i
()1
(t
i1
,t
i
]
, t,
and a partition 0 = t
0
< t
1
< ... < t
m
= T, then is strictly self nancing
if and only if
30
(
t
i+1

t
i
) X
tr
t
i+1
= 0 i,
To use the strictly self nancing condition effectively, we have to
introduce more notation.
m,n
= ((X
t
, F
t
)
tI
, ) be a nancial market. A
trading strategy is called tame if there exists c R such that
V
t
() c P-almost sure. We denote the set of tame trading strategies in
with
t
.
The following Theorem shows us the rst relation between arbi-
trage and martingale measures.
Theorem 2.3.1. Let M
m,n
= ((X
t
, F
t
)
tI
P(M
m,n
) = .Then the nancial market M
m,n
s,t
= ((X
t
, F
t
)
tI
,
s

t
) is
arbitrage free. Here
s

t
denotes the trading strategies in which
are strictly self nancing and tame.
Proof. Let
s

t
such that V
0
() = 0 and V
T
() 0 almost sure.
Let P
P(M
m,n
) = . Since is strictly self nancing it follows from
Proposition 2.3.1 that

V
t
() follows a local P
martingale. Since is
tame and hence

V
t
() is bounded from below, it follows from Exercise
1.2.6 that

V
t
() follows a supermartingale. Therefore we have
E
P
(
V
T
()) = E
P
(
V
T
()|F
0
)

V
0
() = 0
Hence, since also

V
T
() 0 almost sure we have

V
T
() = 0 almost
sure. This then implies that P{V
T
() > 0} = 0 and M
m,n
s,t
is arbitrage
free.
Example 2.3.1. Martingale Measure for the Black-Scholes Mar-
ket Let M
1,1
= ((X
t
, F
t
)
t[0,T]
, ) be the standard Black-Scholes model
with an arbitrary set of trading strategies modeled on the underlying
probability space (, F, P) such that the numeraire X
0
t
= e
rt
represents
a bank account and X
1
t
= e
(b
1
2
2
)t+W
t
the price of a stock. We dene a
new probability measure P
on (, F) as follows : For any A F let

31
P
(A) := E
P
(Z
T
1
A
) =
_
Z
T
1
A
(2.15)
where Z
t
:= e
W
t
1
2
2
t
and :=
rb
. is called the Market Price of

Risk. Then P
is an equivalent martingale measure for the nancial

market M
1,1
. We postpone the proof of this until section 2.7. The ambi-
tioned reader though should try to do the proof at this point. It is a very
good exercise.
In fact we would also like the discounted value process to be a ( real
) martingale, at least for some martingale measure P
P(M). This
leads to the following denition :
m,n
= ((X
t
, F
t
)
tI
strictly self nancing and tame trading strategy is called admis-
sible if there exists P
P(M) such that the discounted value process
V
t
() is a martingale. We denote the set of admissible trading strategies
with
a
and with M
m,n
a
= ((X
t
, F
t
)
tI
,
a
) the corresponding nancial
market.
The following corollary follows directly from the denition of admis-
sible and Theorem 2.3.1.
Corollary 2.3.1. Let M
m,n
= ((X
t
, F
t
)
tI
P(M
m,n
) = . Then the corresponding nancial market M
m,n
a
is arbi-
trage free.
The following Theorem sharpens the No Arbitrage Principle in the
presence of martingale measures.
Theorem2.3.2. No Arbitrage Principle 2 Let M
m,n
= ((X
t
, F
t
)
tI
, )
and ,
a
admissible trading strategies such that V
T
() = V
T
().
Then the corresponding value processes V
t
() = V
t
() are indistinguish-
able.
32
Proof. From the assumptions and the denition of admissibility it fol-
low that there exist P
1
, P
2
P(M
m,n
) such that

V
t
() follows a P
1
mar-
tingale and

V
t
() follows a P
2
martingale. Because of the strictly self
nancing condition and tameness

V
t
() also follows a P
1
supermartin-
gale. Hence
V
t
() = E
P
1
(
V
T
()|F
t
) = E
P
1
(
V
T
()|F
t
)

V
t
() a.s.
Interchanging the roles of and in the argumentation above com-
pletes the proof.
We have seen so far that the existence of an equivalent martingale
measure for a nancial market implies that it is arbitrage free. But
does arbitrage freeness also imply the existence of an equivalent mar-
tingale measure ? In general not.
Denition 2.3.6. A nancial market M
m,n
= ((X
t
, F
t
)
tI
, ) satises
the Fundamental Law of Asset Pricing if the following two condi-
tions are equivalent :
1. M
m,n
is arbitrage free.
2. P(M
m.n
) = .
A fundamental theorem of asset pricing in this context is some kind
of theorem which asserts that some class of nancial markets satises
the fundamental law of asset pricing. Most of the results so far have
been established in the context of semi-martingales ( see [Delbaen/Schachermayer]
[Stricker] ). In the context presented here, this is still up to further re-
search.
Conjecture 1. Fundamental Theoremof Asset Pricing Let M
m,n
=
((X
t
, F
t
)
tI
, ) be a nancial market where is an ample cone, then the
corresponding market M
m,n
= ((X
t
, F
t
)
tI
,
a
) satises the fundamental
law of asset pricing.
33
2.4 Options and Contingent Claims
The reason for why mathematical advanced models of nancial mar-
ket have been developed is not only that some mathematician actually
wanted to develop mathematical models, only they could understand,
but more that people ( traders ) were in need for formulas to compute
the right ( fair ) prices for what they called options.
Though options are traded on stock exchanges all over the world it
is actually not so easy to describe mathematical what they are until
one introduces a more general thing that is called a contingent claim.
One could say an option is something which gives you the right ( not
the obligation ) to buy some other thing at some time ( in the future )
for some predetermined price. To get a rst idea consider the following
example :
You want to give a grill party in about three weeks from now and there-
fore you need meat. Since you expect a lot of people you need a lot of
meat. You decide to go to a butcher and ask how much you need and
what is the price. The butcher tells you exactly how much you need,
but he also tells you that obviously you cannot buy the meat today ( be-
cause then it would be rotten in three weeks ) and that he cannot say
how much the meat will cost in three weeks. he tells you there might
be another food scandal on its way and the meat prices could jump up
and then you would have to pay much more than the price today. Nev-
ertheless the butcher offers you the following : You pay him 5 Euro
and then you can buy the meat in three weeks for todays price, even if
the price in three weeks is much higher than today. Should you accept
the offer, is 5 Euro a fair price ? Maybe the butcher tries to tricks you.
What would be a fair price for such an offer ?
What is described above is what often is called an option. Lets think
about it like this. The Butcher offers you to buy the meat in three
34
weeks from now for todays price, lets call this price K. Let S
T
denote
the price of the meat at time T = 3 weeks . Then if S
T
K you save
S
T
K Euro. If S
T
< K then you buy meat at the ( spot ) price in three
weeks and save nothing. Since the price S
T
is not known you consider
this price as a random variable S
T
(). The money you save can also be
considered as a random variable via
g() =
_
S
T
() K, if S
T
K
0, if S
T
< K
(2.16)
So g can be thought of some random payment in the future. Some-
thing like this is mathematically known as a contingent claim.
m,n
= ((X
t
, F
t
)
tI
contingent claim g is an F
T
-measurable random variable, such that
g 0 a.s. , > 1 s.t. E(g
) < .
In the example above the intention to buy the call on meat was
something like an insurance against the risk that the price of meat
increases. This kind of behavior is called Hedging. Maybe some other
people do not want to give a grill party, but are interested in some prot
by trading in the meat market. They could act as follows. They buy the
call on meat and hope that the meat price increases. Then at time T
they buy the meat from the butcher at price K and sell it immediately
back to the butcher at the price S
T
and earn S
T
K. This kind of be-
havior is called Speculation.
Our aim in the next section will be to dene the right prices for such
contingent claim but before let us consider some types of contingent
claims traded at nancial markets with at least two tradeable assets
(X
1
t
) and (X
2
t
). We have :
35
European Call : (X
1
T
K)
+
= max(S
T
K, 0)
European Put : (K X
1
T
)
+
Call on maximum : (max(X
1
T
, X
2
T
) K)
+
Call on average : (
_
T
0
X
1
t
dt K)
+
Down and out : (X
1
T
K
1
)
+
1
min
0tT
X
1
t
>K
2
where K,K
1
,K
2
are constants. K and K
1
are called strike prices, K
2
is a downside barrier and K
1
> K
2
.These options are still of an elemen-
tary structure compared to other options traded at nancial markets.
People there are still inventing more and more complicated options, in-
creasing the need for mathematician to evaluate these options. ( Maybe
the mathematician invent them themselves. )
2.5 Hedging and Completeness
The seller of a contingent claim must somehow make sure that at ex-
piry time T he can fulll his obligations. If he does this by investing in
the nancial market, then we speak of hedging.
m,n
= ((X
t
, F
t
)
tI
g : R be a contingent claim. A trading strategy is called a
1. Hedging strategy for the contingent claim g if V
T
() = g a.s.
2. Super Hedging strategy if V
T
() g a.s.
Often we loosely speak of a hedge respectively super hedge, mean-
ing a hedging strategy respectively super hedging strategy. If there exists
36
a hedging strategy for g then g is called -attainable. If there exists a
super hedging strategy for g then g is called -super attainable.
If there exist a hedging strategy for the contingent claim g then
the seller of g can invest in the market corresponding to the trading
strategy and makes sure, that at time T he can fulll his obligations.
Also, if there are two hedging strategies and for the same contin-
gent claim g, both belonging to
a
, the the value processes (V
t
())
t[0,T]
and (V
t
())
t[0,T]
are indistinguishable by the No-Arbitrage Principle
2. However, the existence of such trading strategies in general is not
guaranteed.
Denition 2.5.2. A nancial market M
m,n
= ((X
t
, F
t
)
tI
, ) is called
complete, if for any contingent claim g : R there exists a hedging
strategy . Otherwise the market is called incomplete.
In the following we will see complete and incomplete markets, more
incomplete markets in fact. We will see that some version of the stan-
dard Black-Scholes nancial market model is complete. Incomplete
markets are more difcult, we will see the reason for this in the next
section. However incomplete markets arise quite naturally and the
Black-Scholes model seems not to give the right picture for what is go-
ing on at real world nancial markets. The algebraic structure of the
set of contingent claims is that of a cone. It is however not clear, that if
g
1
and g
2
are -attainable, g
1
+g
2
is also -attainable. To conclude this
in general, we would need at least that is a cone.
We saw in section 2.3. that arbitrage has something to do with the
existence of martingale measure. Completeness has something to do
with the uniqueness of martingale measure at least if one considers
the following sub class of equivalent martingale measures :
Denition 2.5.3. A probability measure P
P(M
m,n
) is called a strong
equivalent martingale measure for the nancial market M
m,n
= ((X
t
, F
t
)
tI
, )
if for all
a
the discounted value process V
t
() follows a P
-martingale.
We denote the set of strong equivalent martingale measures with P
s
(M
m,n
).
37
The next theorem establishes the connection between completeness
and uniqueness of equivalent martingale measures.
Theorem2.5.1. (Uniqueness of equivalent Martingale Measures)
Let M
m,n
= ((X
t
, F
t
)
tI
,
a
) be a complete nancial market, such that
F
T
= F, P(M
m,n
) = and E
P
(X
0
T
)
< for a > 0. Then |P

s
(M
m,n
)| =
1.
Proof. Let P
1
, P
2
P
s
(M
m,n
) and A F = F
T
. Then
g := 1
A
X
0
T
: R
is a contingent claim. Since M
m,n
is complete there exists a hedge

a
such that V
T
() = 1
A
X
0
T
almost sure. This of course implies
V
T
() = 1
A
. Hence we have
P
i
(A) = E
P
i
(1
A
|F
0
) = E
P
i
(
V
T
()|F
0
) =

V
0
(), i = 1, 2,
and in fact P
1
(A) = P
2
(A).
2.6 Pricing of Contingent Claims
The pricing of contingent claims is one of the major topics of this course.
By pricing we mean how to associate a price process to a contingent
claim in a nancial market, such that if this price process is consid-
ered as a tradeable asset ( means the nancial market is traded at the
market ), the corresponding extended market ( with admissible trad-
ing strategies ) is arbitrage free. For the whole section let M
m,n
=
((X
t
, F
t
)
t[0,T]
, ) a nancial market.
Denition 2.6.1. Let g : R be a contingent claim in M
m,n
. A
price process (g
t
)
t[0,T]
for g is a non negative (F
t
)
t[0,T]
adapted stochastic
process such that g
T
= g almost sure.
38
Before thinking about the question whether the extended nancial
market is arbitrage free or not, we must give a precise mathematical
formulation of how to extend nancial markets at all.
m,n
= ((X
t
, F
t
)
tI
, ) and N
k,l
= ((Y
t
, G
t
)
tI
, )
be nancial markets. Then we dene the product of M
m
,
n
and N
l,k
as
M
m
,
n
N
l,k
:= (MN)
m+k+1,n+l+1
= ((Z
t
, H
t
)
tI
, ),
with
Z
t
= (X
0
t
, ..., X
m
t
, Y
0
t
, ..., Y
k
t
, X
m+1
t
, ..., X
n
t
, Y
k+1
t
, ..., Y
l
t
)
t,
H
t
= (F
t
, G
t
) and the Cartesian product of the two sets of
trading strategies.
Given a contingent claim g and a price process (g
t
)
t[0,T]
for this
claim, we consider the nancial market
M
(g
t
)
= ((g
t
, F
t
)
t[0,T]
,
simple
)
where
simple
denotes the one-dimensional simple processes ( see
Denition 1.2.7 ). We can now dene the extended nancial market,
where discrete trading with the contingent claim is allowed as follows
:
Denition 2.6.3. Let (g
t
) be a price process for a contingent claim g
in a nancial market M
m,n
= ((X
t
, F
t
)
tI
, ). We dene the extended
nancial market as
M
m+1,n+1
(g
t
) = M
m,n
M
(g
t
)
where the product is given by ( 2.16 ).
39
Given a price process for a contingent claim, the contingent claim
will only be traded if the price process gives no ( signicant ) advantage
to neither the seller or the buyer of the contingent claim. We would
consider such a price as a fair price.
Denition 2.6.4. A price process (g
t
) for a contingent claimg in a nan-
cial market M
m,n
is called a fair price if the extended nancial market
M
m+1,n+1
a
(g
t
) with admissible trading strategies is arbitrage free.
Let us assume now that P
P(M
m,n
) is an equivalent martingale
measure and g : R is a contingent claim in this market. Then we
can dene
g
t
:= X
0
t
E
P
((X
0
T
)
1
g|F
t
).
Clearly g
t
is a price process for the contingent claim g. The follow-
ing theorem tells us that equivalent martingale measures compute fair
prices for contingent claims.
Theorem 2.6.1. Let M
m,n
= ((X
t
, F
t
)
tI
g a contingent claim in this market. Let P
P(M
m,n
) then the price
process g
t
:= X
0
t
E
P
((X
0
T
)
1
g|F
t
) is fair.
Proof. We have to showthat M
m+1,n+1
a
(g
t
) is arbitrage free .FromCorol-
lary 2.3.1. it follows that this is indeed the case if P(M
m+1,n+1
(g
t
)) = .
In fact we show that P
P(M
m+1,n+1
(g
t
)). But this is almost clear
since by denition of an equivalent martingale measure,

X
t
is a local
P
-martingale and also g

t
= (X
0
t
)
1
g
t
= E
P
((X
0
T
)
1
g|F
t
) is a local P
-
martingale ( see also Exercise 1.2.2 ).
In this way, martingale measures should be interpreted as pricing
systems. They should be considered as ( linear ) functionals on the
space of contingent claims with values in the space of fair prices. The
interesting thing is, that is some cases if |P(M
m,n
)| > 1 not all of them
give the same prices. It might be, that one equivalent martingale mea-
sure prices a contingent claim more expensive than another one, but
40
still both prices are fair. Nevertheless the question arises, which of
the equivalent martingale measures one should use to price contin-
gent claims. Further research into this direction needs more advanced
methods in functional analysis and topology then presented here.
A different method to price contingent claims is directly related to
hedging strategies.
Denition 2.6.5. let g be a contingent claim in a nancial market
M
m,n
= ((X
t
, F
t
)
tI
, ). We dene the Hedging Price of g as
hedge
= inf{x R| hedge
a
s.t. V
0
() = x}
and the Super Hedging Price as
shedge
= inf{x R| super hedge
a
s.t. V
0
() = x}
The hedging price tells us exactly the minimum investment at time
0 to get the pay out g at time T. This seems to be a quite reasonable
procedure to price contingent claims. The question how these prices
are related to martingale measures and whether they are fair or not
will be answered by the following proposition and corollary.
Proposition 2.6.1. If g is a contingent claim in M
m,n
= ((X
t
, F
t
)
tI
, )
and
a
a hedge for g, then the discounted value process V
t
() is a
fair price process for g. Furthermore we have
V
t
() X
0
t
E
P
((X
0
T
)
1
g), P
P(M
m,n
).
Proof. We have g = V
T
() a.s. Since is admissible, there exists P

P(M
m,n
) such that

V
T
() is a P
martingale. Hence
V
t
() = X
0
t

V
t
() = X
0
t
E
P
(
V
T
()|F
t
)
= X
0
t
E
P
((X
0
T
)
1
V
T
()|F
T
)
= X
0
t
E
P
((X
0
T
)
1
g|F
t
)
41
and V
t
() is a fair price process by Theorem 2.6.1. For any other
equivalent martingale measure

P
P(M
m,n
)

V
t
() follows a

P
-super
martingale ( local martingale and bounded from below ! ) and hence
V
t
() = X
0
t

V
t
() X
0
t
E
V
T
()|F
t
)
= X
0
t
E
((X
0
T
)
1
V
T
()|F
T
)
= X
0
t
E
((X
0
T
)
1
g|F
t
)
Corollary 2.6.1. Let g be a contingent claim in M
m,n
then
hedge
X
0
0
E
P
((X
0
T
)
1
g), P
P(M
m,n
).
2.7 The Black-Scholes Formula
In this section we restrict ourself to one distinguished nancial market
model, the standard Black-Scholes model of section 1.4. and Example
2.3.1, and one distinguished contingent claim, the European call op-
tion from section 2.5. The model is given within a complete probability
space (, F, P) which carries a standard Brownian motion (W
t
, F
t
) by a
two dimensional price process
(X
t
)
t[0,T]
=
_
X
0
t
X
1
t
_
t[0,T]
=
_
e
rt
S
0
e
(b
1
2
2
)t+W
t
1
_
t[0,T]
where X
0
t
represents a bank account with deterministic interest
rate r > 0 and X
1
t
the price of a stock with initial price S
0
> 0. We
do not x the set of trading strategies here, but we think of the set
of simple processes as allowed trading strategies. This is reasonable
at least, since at real world markets, trading only happens in discrete
time steps. We denote this market with M
1,1
BS
. In Example 2.3.1 we
dened a new measure P
on (, F) as
42
P
(A) := E
P
(Z
T
1
A
) =
_
Z
T
1
A
where Z
t
:= e
W
t
1
2
2
t
and :=
rb
. Since Z
T
> 0 it is clear that P and
P
are equivalent measures. Let us now show that the process dened
by
W
t
:= W
t
t, t (2.17)
is a Brownian motion with respect to the measure P
. Clearly W
0
=
W
0
= 0 P- and hence also P
-almost sure. Also W
t
W
s
= W
t
W
s
(t
s) is independent of F
s
. Let us now compute the distribution function
of W
t
W
s
. Notice rst, that if A F
t
, we have
P
(A) = E
P
(1
A
Z
T
) = E
P
(1
A
E
P
(Z
T
|F
t
) = E
P
(1
A
Z
t
),
since (Z
t
)
t[0,T]
is a P-martingale ( see Exercise 1.3.1 ). We have
P
(W
t
W
s
x) = E
P
(1
{W
t
W
s
x}
Z
t
)
= E
P
(E
P
(1
{W
t
W
s
x}
Z
t
Z
1
s
|F
s
)Z
s
)
= E
P
(E
P
(1
{W
t
W
s
x}
e
(W
t
W
s
)
1
2
2
(ts)
. .
independent of F
s
|F
s
)Z
s
)
= E
P
(1
{W
t
W
s
x+(ts)}
e
(W
t
W
s
)
1
2
2
(ts)
) E
P
(Z
s
)
. .
=1
Using the density function of W
t
W
s
N(0, t s) ( with respect to
the measure P )we get that the last expression above is equal to
_
x+(ts)
2(ts)
e
y
1
2
2
(ts)
e
y
2
2(ts)
dy
=
..
z=y(ts)
_
x
2(ts)
e
z+
1
2
2
(ts)
e
(z+(ts))
2
2(ts)
dz
43
In the last integral almost everything cancels, and what is left is
the expression
_
x
1
_
2(t s)
e
z
2
2(ts)
dy.
But this ( as a function of x ) is the distribution function of an
N(0, t s) distributed random variable. Hence we have proven that
W
t
W
s
N(0, ts) under the measure P
. Since obviously the process

(W
t
)
t[0,T]
has continuous paths, we have proven that it is a standard
Brownian motion under P
.
Clearly

X
0
t
1 is a martingale under P
. Let us now consider X

1
t
=
x e
(b
1
2
2
)t+W
t
. By a simple transformation we get
X
1
t
= S
0
e
(b
1
2
2
)t+W
t
= x e
W
t
(rb)t+rt
1
2
2
t
= S
0
e
W
t
(r
1
2
2
)t
.
X
1
t
= e
rt
X
1
t
= x e
W
t

1
2
2
t
.
Since (W
t
)
t[0,T]
is a P
Brownian motion the last expression follows

a P
-martingale ( see again Exercise 1.3.1 ). Hence we have proven that

P
is an equivalent martingale measure for the market M

1,1
BS
and that
(M
1,1
BS
)
a
is arbitrage free. Let us now consider a European call option
with strike price K given by
g : R
g() = (X
1
T
( K)
+
.
Using Theorem 2.6.1 we can compute a fair price for this option as
g
0
= E
P
(e
rT
(X
1
T
K) 1
{X
1
T
K}
)
= E
P
(e
rT
X
1
T
1
{X
1
T
K}
)
. .
=(I)
e
rT
K P
(X
1
T
K)
. .
=(II)
44
Let us compute the expressions (I) and (II) :
(I) = E
P
(S
0
e
W
1
2
2
T
1
{ln(S
0
)+(r
1
2
2
)T+W
T
ln(K)}
= S
0
_

ln(
K
S
0
)(r
1
2
2
)T
2T
e
x
1
2
2
T
e
x
2
2T
dx
= S
0
_

ln(
K
S
0
)(r
1
2
2
)T
2T
e
(xT)
2
2T
dx.
We denote
d
1
:=
ln(
S
0
K
) + (r +
1
2
2
)T
T
. (2.18)
an by substitution of y =
xT
T
we get
(I) = S
0
_

d
1
1
2
e
y
2
2
dy
= S
0
(d
1
),
where denotes the standard normal distribution function. Let us
now compute the second part (II). We have
(II) = e
rT
K P
{
W
ln(
K
S
0
) (r
1
2
2
)T
T
. .
d
1
+
T
}
Since
W
T
N(0, 1) under P
, we get
(II) = e
rT
K (d
1
T).
Hence we have proven the following theorem :
Theorem 2.7.1. Black-Scholes Formula A fair price for a European
call option with strike price K in the standard Black-Scholes model is
given by
45
C(S
0
, K, ) = S
0
(d
1
) e
rT
K (d
1
T)
where d
1
is given by the expression d
1
:=
ln(
S
0
K
)+(r+
1
2
2
)T
T
.
In the theorem above we do only speak of a fair price because it
might be, that there are other fair prices corresponding to other equiv-
alent martingale measures for the standard Black-Scholes model. In
fact, there are no other equivalent martingale measures than the one (
2.21 ), but we cannot ( or dont want to ) prove this at this place.
Exercise 2.7.1. Compute a fair price for a digital call option with strike
price K in the standard Black-Scholes model. The pay-out of a digital
Call option is given as
g() := 1
{X
1
T
K}
.
Exercise 2.7.2. Compute a fair price for the Buttery-Spread option
with strike price K in the standard Black-Scholes model. The pay-out
of a Buttery-Spread option is given as
g() =
_
_
0 if X
1
T
K
X
1
T
K if K X
1
T
2K
2K X
1
T
if 2K X
1
T
3K
0 if 3K X
1
T
(2.19)
Why is this option been called Buttery-Spread ?
2.8 Why is the Black-Scholes model not good
enough ?
The Black-Scholes model certainly is still the most famous and one
of the most applied models. On the other side, it is known that the
reality at nancial markets looks different. In this short section we
46
will demonstrate why this is so. The standard Black-Scholes model
assumes for the stock price the price process
S
t
= S
0
e
(b
1
2
)t+W
t
,
where the coefcients b and are assumed to be constants. is
called the volatility of the stock. Let us consider the fair price of a
European Call option given by Theorem 2.7.1 as
C(S
0
, K, ) = S
0
(d
1
) e
rT
K (d
1
T), (2.20)
Let us consider this price as a function of the volatility. From a
rational point of view, it should be clear that this function is mono-
tonically increasing. Nevertheless the ambitious reader should do the
following exercise
Exercise 2.8.1. Show the fair price of a European call option in the
standard Black-Scholes model given by ( 2.24 ) is monotonically increas-
ing as a function of the volatility .
Hence we can look at the price at which European call options for
this stock are traded on some stock exchange, call this price C
real
(S
0
, K)
and solve
C(S
0
, K, ) = C
real
(S
0
, K) (2.21)
for . Using this volatility for the computation in ( 2.24 ) for exactly
this option then of course gives the right ( real world price ). The fun-
damental problem though is, that when solving ( 2.27 ) in praxis, for
different strike prices K, in general one gets different values for . So
we have = (K). This in fact contradicts the assumption made by the
Black-Scholes model that is considered to be constant. (K) is called
the implied volatility, the graph of (K) is often called volatility smile
for its typical shape. There are various methods to improve the Black-
Scholes model, and we will learn about some of them in the remaining
of this lecture.
47
Chapter 3
Stochastic Integration
In section 2.3 we dened the sums
Z
d

X
tr
and
Z
d

X
tr
for certain
partitions and assumed some kind of convergence of these sums when
|Z| 0. We now want to make this concept more precise. The concept
is called stochastic Integration.
3.1 Semi-martingales
Let us rst consider a generalization of the notion of a simple processes.
Denition 3.1.1. An n-dimensional stochastic process (X
t
, F
t
)
t[0,T]
is
called simple predictable if there exist a nite sequence of stopping
times
0 = T
0
T
1
... T
m
= T
with respect to the ltration (F
t
)
t[0,T]
and
i
: R
n
such that
|
i
| < a.s.,
i
is F
T
i
measurable and
X
t
() =
0
() 1
{0}
+
m
i=1
i
() 1
(T
i1
,T
i
]
, t, . (3.1)
We denote the class of simple predictable processes with E. We say a
sequence of processes X
k
E converges to X E if
48
X
k
E
//
X
: sup{|X
k
t
() X
t
()||(, t) I} 0 as k
Denition 3.1.2. A process Y = (Y
t
)
tI
is called rcll ( right continuous
with left limits ) if
Y () C
+
(I, R
n
) a.s.
Y
t
() := lim
st
Y
s
() exists for a.a. and for all t I.
In this case we dene a new stochastic process as Y
:= (Y
t
)
tI
. A
process X = (X
t
)
tI
is called lcrl ( left continuous with right limits ) if
X() C
(I, R
n
) a.s.
X
t+
() := lim
st
X
s
() exists for a.a. and for all t I.
and in this case as before we dene X
+
:= (X
t+
)
tI
.We dene the
corresponding jump processes as
Y := Y Y
, X := X
+
X.
We can now dene, what a semi-martingale is.
Denition 3.1.3. A process Y is called a total semi-martingale if Y
is rcll, (F
t
)
tI
adapted and the map
_
()dY : E Map((, F), (R, B(R))
X
0
Y
0
+
n
i=1
i
(Y
T
i+1
Y
T
i
)
where X is given by (3.1), is continuous in the following sense :
X
k
E
//
X

_
X
k
dY
P
//
_
XdY .
49
Y is called a semi-martingale if t I the stopped process Y
t
is a
total semi-martingale. We denote the class of semi-martingales with S.
In case Y S and X E we write
_
t
0
XdY :=
_
XdY
t
.
In the denition above Y was assumed to be R-valued. Most de-
nitions in this chapter correspond to the one-dimensional case but can
be generalized to the n-dimensional case without taking serious effort.
For example an n-dimensional semi-martingale is a R
n
vector valued
stochastic process whose components are semi-martingales.
Remark 3.1.1. 1. S is a vector space.
2.
_
XdY is bilinear as a function of the integrand X and the inte-
grator Y .
3.
_
t
0
XdY is F
t
measurable and (
_
t
0
XdY )
tI
is an (F
t
) adapted rcll
stochastic process.
4. S = S(P) actually depends on P, but if Q << P then
Z
k
P
//
Z

Z
k
Q
//
Z
hence every semi-martingale with respect to P is also a
semi-martingale with respect to Q. If in particular we have P Q
then S(P) = S(Q).
5. S = S((F
t
)) also depends on the ltration, but if (G
t
) is a sub-
ltration of (F
t
) and Y is also (G
t
)-adapted, then Y is also a semi-
martingale with respect to (G
t
), since every elementary process with
respect to (G
t
) is also one with respect to (F
t
).
The following elementary properties of the stochastic integral ( for
elementary processes with respect to a semi-martingale ) are left as an
exercise.
Exercise 3.1.1. Let Y = (Y
t
)
tI
be a semi-martingale dened on a
complete probability space (, F, P) with given ltration (F
t
)
tI
and let
X, X
1
, X
2
be elementary processes. Show :
50
1.
_
t
0
cdY = cY
t
for any c R considered as a constant deterministic
process.
2. |Y
s
| < a.s. for all s I.
3. Y 1
{Y 0}
, Y 1
{Y <0}
and |Y | are semi-martingales.
4. If Y 0 a.s. and X
1
s
X
2
s
a.s. s [0, t] then
_
t
0
X
1
dY
_
t
0
X
2
dY
5. If X is bounded and Y is a martingale, then (
_
t
0
XdY )
tI
is also a
martingale.
The next propositions will give us various examples of semi-martingales.
Proposition 3.1.1. Let Y = (Y
t
)
tI
be an rcll (F
t
)
tI
adapted process
s.t. t and a.s. we have
_
t
0
|dY
s
|() := sup{
Z(t)
|Y
t
i+1
Y
t
i
| : Z(t) : 0 = t
0
< t
1
... < t
m
= t} <
( we say Y has nite variation on compacts ). Then Y S.
Proof. Let X E be given as in (3.1). We have
|
_
XdY
t
| = |
0
Y
t
0
+
n
i=1
i
(Y
t
T
i+1
Y
t
T
i
)|
|
0
||Y
t
0
| +
n
i=1
|
i
||Y
t
T
i+1
Y
t
T
i
|
sup
(,t)I
|X
t
()| (|Y
0
| +
_
t
0
|dY
s
|)
If
X
k
E
//
X
then
X
k
X
E
//
0 and we follow from our Assump-
tions,
51
that for any > 0 we have lim
n
P(|
_
(X
k
X)dY
t
| > )
lim
n
P( sup
(,t)I
|X
k
t
() X
t
()|
. .
0 as k0
)| (|Y
0
| +
_
t
0
|dY
t
|)
. .
< a.s.
> ) = 0
The most interesting stochastic processes nevertheless do not have
nite variation on compacts. So we have to go on. In the context of
martingales we dened local martingales as something which becomes
after appropriate stopping a martingale. The same procedure works in
the semi martingale context but :
t
)
tI
be an rcll (F
t
)
tI
adapted process s.t.
there exists an increasing sequence of stopping times T
n
s.t. limT
n
=
a.s. and n we have Y
T
n
S, then Y S ( in words : every local semi-
martingale is a semi-martingale ).
Proof. Let X E be given as in (3.1). Dene stopping times
S
n
= T
n
1
{T
n
t}
+ 1
{T
n
>t}
.
We have
P(|
_
XdY
t
| ) P(|
_
t
0
XdY | , S
n
= ) +P(|
_
t
0
XdY | , S
n
< )
P(|
_
t
0
XdY
T
n
t
| ) +P(S
n
< ).
Now let
X
k
E
//
X
i.e.
X
k
X
E
//
0 . Then n
lim
k
P(|
_
(X
k
X)dY
t
| ) lim
k
P(|
_
(X
k
X)dY
T
n
t
| )
. .
=0
+P(S
n
< )
52
The statement of the proposition now follows from lim
n
P(S
n
<
) = lim
n
P(T
n
t) = 0.
Denition 3.1.4. Let Y = (Y
t
)
tI
be a martingale. We say Y is a
1. L
2
(P)-martingale if t E
P
(Y
2
t
) =
_
Y
2
t
dP <
2. L
2
(P )-martingale if E
P
(Y ) =
_
I
Y
2
d(P ) <
where denotes the Lebesgue-measure on R. Further more, if Y is arbi-
trary, the we say Y is an
1. L
2
loc
(P)-martingale if there exists an increasing sequence of stop-
ping times T
n
s.t. limT
n
= a.s. and Y
T
n
is a L
2
(P)-martingale
2. L
2
loc
(P)-martingale if there exists an increasing sequence of stop-
ping times T
n
as above s.t. Y
T
n
is a L
2
(P )-martingale
Exercise 3.1.2. Let Y be an L
2
(P)-martingale. Show that for any se-
quence of stopping times 0 = T
0
T
1
... T
m+1
< a.s. one has
E(
m
i=0
(Y
T
i+1
Y
T
i
)
2
) = E(Y
2
T
m+1
).
The following proposition gives at once a large variety of examples
for semi-martingales.
t
)
tI
be an rcll L
2
loc
(P)-martingale. Then
Y S.
Proof. W.l.o.g. we assume Y
0
= 0 and Y is an L
2
(P)-martingale. Let
X E as in (3.1). Then
E((
_
t
0
XdY
t
)
2
) = E((
n
i=0
i
(Y
t
T
i+1
Y
t
T
i
))
2
)
( sup
(,t)I
|X
t
()|)
2
E((Y
t
T
n+1
)
2
)
( sup
(,t)I
|X
t
()|)
2
E(Y
2
t
)
53
where the last inequality follows from the Jensen-inequality ( Prob.
Theory Theorem 19.1 and Lemma 19.1 ). Hence
X
k
E
//
X
and hence
X
k
X
E
//
0 implies
_
X
k
dY
t
L
2
(P)
//
_
XdY .
It follows from Prob. Theory Theorem 15.2c) that L
2
convergence
implies convergence in probability. Hence we have
_
X
k
dY
t
P
//
_
XdY
which proves the proposition.
Let us now consider our rst explicit example of a semi-martingale
:
Corollary 3.1.1. Any Brownian motion (W
t
)
tI
is a semi-martingale.
Proof. Obviously (W
t
) is an L
2
loc
(P) martingale.
t
)
tI
be a continuous local martingale,
then Y S.
Proof. Let T
n
be an increasing sequence of stopping times s.t. lim
n
T
n
=
a.s. and Y
T
n
is a martingale. We dene another increasing sequence
of stopping times
S
n
:= inf{t||Y
t
| n}
Clearly lim
n
S
n
= a.s. and hence also lim
n
T
n
S
n
= a.s..
T
n
S
n
is also an increasing sequence of stopping times s.t. Y
T
n
S
n
=
(Y
T
n
)
S
n
is a martingale. Furthermore we have
E((Y
T
n
S
n
. .
n
)
2
) E(n
2
) = n
2
< .
Hence Y
t
is an L
2
loc
(P)-martingale and by Proposition 3.1.3 this im-
plies that Y is a semi-martingale.
54
Let us now describe what semi-martingales weve got so far. For
this the following denition is useful.
Denition 3.1.5. An rcll process Y = (Y
t
)
tI
is called decomposable,
if
Y
t
= Y
0
+ M
t
+ A
t
t I
where M
0
= A
0
= 0, M is an rcll L
2
loc
(P)-martingale and A rcll and
of nite variation on compacts.
Corollary 3.1.2. Let Y = (Y
t
)
tI
be a decomposable process, then Y is a
semi-martingale.
Proof. Follows from Proposition 3.1.1 and Proposition 3.1.3.
The processes in mathematical nance are often of the form Y
t
=
Y
0
+

M
t
+A
t
where everything is as above with the exception that

M
t
is
assumed to be a local martingale. It follows from Proposition 3.1.4 that
in this case Y is also a semi-martingale. In fact most semi-martingales
are also decomposable but we wont go deeper into this.
3.2 The stochastic Integral
We already dened stochastic integration in case the integrator is a
semi-martingale and the integrand is an elementary process. The class
of integrators is already sufcient, but we want to enlarge the class of
integrands. For this we exploit the continuity property in Denition
3.1.1. In fact we show that this continuity is actually stronger than it
was assumed in Denition 3.1.1.
Denition 3.2.1. Let X
k
be a sequence of stochastic processes. We say
X
k
converges ucp ( uniformly on compacts in probability ) if
55
t I : sup
0st
|X
k
s
X
s
|
P
//
0 :
X
k
ucp
//
X
The following proposition shows that we can get any lcrl process as
an ucp limit of elementary processes.
Proposition 3.2.1. Let X be an lcrl adapted process. Then there exists
a sequence X
k
of elementary processes, such that
X
k
ucp
//
X
.
Proof. As before we dene an increasing sequence of stopping times as
n
:= inf{t : |X
t
| n}.
We have lim
n
n
= and hence
lim
n
P( sup
0st
|X
n
s
X
s
| ) lim
n
P( sup
0s
n
=0
..
|X
n
s
X
s
| )
. .
=0
+ lim
n
P( sup
n
st
|X
n
s
X
s
| )
lim
n
P(
n
t) = 0
Hence
X
n
ucp
//
X
and all X
n
are bounded processes. Hence we can
assume in the following that X is also bounded. Dene the rcll process
Z as Z = X
+
. For > 0 dene a sequence of stopping times
0
:= 0,
n+1
:= inf{t|t >
n
and |Z
t
Z
n
| > } n.
n
is an increasing sequence of stopping times s.t. lim
n
n
=
a.s. ( here we need boundedness ). We dene
Z
:=
n=0
Z
n
1
[
n
,
n+1
)
Then we have that
56
|Z
t
Z
t
| a.s. t I
and hence that Z
Z uniformly and bounded 0. We dene an

lcrl process

Z
as
:= X
0
1
{0}
+
n=0
Z
n
1
(
n
,
n+1
]
.
Then

Z
= X uniformly and bounded as 0. Now let

X
k
:= X
0
1
{0}
+
k
n=0
Z
1/k
n
1
(
1/k
n
k,
1/k
n+1
k]
.
Clearly X
k
E and for k > t we have X
k
s
=

Z
1/k
s
0 s t. Hence
we get
lim
k
P( sup
0st
|X
k
s
X
s
| ) = lim
k
P( sup
0st
|
Z
1/k
s
X
s
|
. .
1/k
) = 0
which shows that
X
k
ucp
//
X
.
t
)
tI
be a semi-martingale and X
k
E a
sequence converging ucp to X E, i.e.
X
k
ucp
//
X
, then
(
_
t
0
X
k
dY )
tI
ucp
//
(
_
t
0
XdY )
tI
.
Proof. W.l.o.g. we assume that Y is a total semi-martingale and X = 0.
Let us rst assume X
k
E and
X
k //
0 uniformly and bounded. let
> 0. We dene stopping times
k
:= inf{t :
_
t
0
X
k
dY | }
Clearly X
k
1
[0,
k
]
E and
57
P( sup
0st
|
_
s
0
X
k
dY | ) = P( sup
0st
|
_

k
s
0
X
k
dY | )
= P( sup
0st
|
_
s
0
X
k
1
[0,
k
]
dY | )
= P( sup
0st
|
_
X
k
1
[0,
k
]
. .
E
//
0
dY
s
| ) 0 as k .
This shows that
X
k
E
//
0 (
_
t
0
X
k
dY )
tI
ucp
//
0 . Nowlet
X
k
ucp
//
0
with X
k
E. Let > 0, > 0, t > 0. It follows from above that there ex-
ists > 0 s.t. whenever sup
(,t)I
|X
t
()| we have P(sup
0st
|
_
s
0
XdY | >
) < /2. Dene stopping times
k
:= inf{s : |X
k
s
| > }.
We dene
X
k
:= X
k
1
[0,
k
]
1
{
k
>0}
.
Then

X
k
E and sup
(,t)I
|
X
k
t
()| . If
k
t then
sup
0st
|
_
s
0
X
k
dY | = sup
0st
|
_
s
0
X
k
dY |.
Hence we have
P( sup
0st
|
_
s
0
X
k
dY | > ) P( sup
0st
|
_
s
0
X
k
dY | > ) +P(
k
< t)
/2 +P( sup
0st
|X
k
s
| > )
. .
0 as k
from which the statement of the proposition follows.
We want to dene the stochastic integral with integrand an arbi-
58
trary adapted lcrl process X as a limit of integrals of elementary pro-
cesses. The following lemma shows that this limit indeed exists.
Lemma 3.2.1. Let X
n
E be a sequence of elementary processes s.t.
X
n
ucp
//
X
where X is lcrl adapted. Then t the sequence
_
t
0
X
n
dY
converges in probability.
Proof. Applying Lemma 15.3 Prob. Theory, we have to showthat
_
t
0
X
n
dY
is a Cauchy-sequence in probability. For this let > 0, > 0. For any
> 0 and m, n N we have
P(|
_
t
0
X
m
dY X
n
dY | > ) P(|
_
t
0
(X
m
X
n
)dY | > , sup
0st
|X
m
s
X
n
s
| )
+ P( sup
0st
|X
m
s
X
n
s
| > )
Since Y S we can nd > 0 such that whenever X E with
sup
(,t)I
|X
s
()| we have P(|
_
t
0
XdY | > ) < /2). For this we
have
P(|
_
t
0
(X
m
X
n
)dY | > , sup
0st
|X
m
s
X
n
s
| ) < /2.
Furthermore, since
X
n
ucp
//
X
sup
0st
|X
n
X|
P
//
0 we can
nd n
0
N s.t.
P( sup
0st
|X
n
s
X
s
| > /2) < /4 n n
0
.
Since
{ sup
0st
|X
n
s
X
m
s
| > } { sup
0st
|X
n
s
X
s
| > /2} { sup
0st
|X
m
s
X
s
| > /2}
we have m, n n
0
59
P( sup
0st
|X
n
s
X
m
s
| > ) P( sup
0st
|X
n
s
X
s
| > /2) +P( sup
0st
|X
m
s
X
s
| > /2)
< /4 +/4 = /2.
Hence m, n n
0
we have
P(|
_
t
0
X
m
dY
_
t
0
X
n
dY | > ) < /2 +/2 = ,
which shows that the sequence
_
t
0
X
n
dY is indeed a Cauchy-sequence
in probability.
We are now able to dene the stochastic integral with integrand an
adapted lcrl stochastic process.
Denition 3.2.2. Let X be adapted and lcrl, Y S. Then we dene
_
t
0
XdY := P lim
k
_
t
0
X
k
dY
where (X
k
) is an arbitrary sequence of elementary processes, s.t.
X
k
ucp
//
X
and Plim
k
denotes the limit in probability. We consider (
_
t
0
XdY )
tI
as
a stochastic process and call it the stochastic integral of X with respect
to Y . We denote this stochastic process as
_
XdY .
The stochastic integral is well dened, since by Proposition 3.2.1
there exists at least one sequence of elementary processes converging
ucp to X and by Proposition 3.2.2, if there are two sequences of elemen-
tary processes converging ucp to X, then the corresponding stochastic
integrals converge ucp to the same stochastic process.From the deni-
tion and Remark 3.1.1 it is also clear that
_
XdY is adapted and rcll.
The stochastic process
_
XdY should not be confused with the ( image
of the ) map in 3.1.4.
As a rst example we want to compute the stochastic integral
_
WdW
60
where W is a Brownian-motion. For this we will make use of the fol-
lowing lemma.
Lemma 3.2.2. Let Z
n
: 0 = t
n
0
< ... < t
n
k
n
= t be a sequence of partitions
such that lim
n
|Z
n
| = 0 then
i
(W
t
n
i+1
W
t
n
i
)
2 P
//
t .
Proof. Dene X
n
:=
i
(W
t
n
i+1
W
t
n
i
)
2
. Then by telescoping
X
n
t =
i
(W
t
n
i+1
W
t
n
i
)
2
(t
n
i+1
t
n
i
)
. .
G
i
The G
i
s are independent random variables. Furthermore, with
W
t
n
i+1
W
t
n
i
_
t
n
i+1
t
n
i
G with G N(0, 1).
we have
G
i
(G
2
1)(t
n
i+1
t
n
i
).
E(G
i
) = 0 for all i and
E((X
n
t)
2
) = E((
i
G
i
)
2
) =
i
E(G
2
i
)
= E((G
2
1)
2
i
(t
n
i+1
t
n
i
)
2
)
E((G
2
1)
2
)
. .
<
|Z
n
|
..
0
t 0 as n
This implies
X
n
L
2
(P)
//
t which then implies
X
n
P
//
t .
Let us return to our example
_
WdW. We know that W S and
also that W is continuous, in particular lcrl. Hence this integral is well
61
dened. Let us take a sequence of partitions Z
n
: 0 = t
n
0
< ... < t
n
k
n
<
s.t. lim
n
|Z
n
| = 0 and lim
n
t
n
k
n
= . We dene
X
n
:=
i
W
t
n
i
1
(t
n
i
,t
n
i+1
]
Exercise 3.2.1. Show
X
n
ucp
//
W
.
We compute
_
t
0
X
n
dW =
i
W
t
n
i
(W
t
t
n
i+1
W
t
t
n
i
)
=
1
2
i
(W
t
t
n
i+1
+ W
t
t
n
i
)(W
t
t
n
i+1
W
t
t
n
i
)
. .
=(W
t
t
n
i+1
)
2
(W
t
t
n
i
)
2
1
2
i
(W
t
t
n
i+1
W
t
t
n
i
)(W
t
t
n
i+1
W
t
t
n
i
)
=
1
2
W
2
t
n
k
n
t
. .
W
2
t
1
2
i
(W
t
n
i+1
t
W
t
n
i
t
)
2
. .
t ( Lemma 3.2.2 )
Building the limit for n we get
_
t
0
WdW =
1
2
W
2
t

1
2
t. (3.2)
This result is under some aspect very interesting and shows very
good how stochastic integration is different from ordinary integration,
where we have
_
xdx =
1
2
x
2
. The additional term
1
2
t comes from
Lemma 3.2.2 and may surprise those, who prefer deterministic think-
ing. We will later see that from the probabilistic point of view, its exis-
tence is very natural. Let us now state ( without proof ) some properties
of the stochastic integral.
Proposition 3.2.3. Let X be an adapted lcrl process and Y S. Then
1. For any stopping time we have
_
0
XdY =
_
0
X 1
[0,]
dY =
_
0
XdY
where
_
0
means integration over the whole parameter
set I.
62
2. (
_
XdY )
s
= X
s
(Y )
s
, hence if Y is a continuous semi-martingale,
then
_
XdY is a continuous process.
3.
_
XdY depends on the measure P but if Q << P then Q
_
XdY
and P
_
XdY are Q indistinguishable.
4. Let

X be another adapted lcrl process and

Y be another semi-
martingale. Let
A = {|X() =

X(), Y () =

Y ()}.
Then
_
XdY and
_

Xd
Y coincide on A.
5. Associativity :
_
XdY is itself a semi-martingale and with

X as
above we have
_

Xd
_
XdY =
_

XXdY.
The following proposition is a generalization of statement (5) in Ex-
ercise 3.1.1. in case Y is an L
2
loc
(P)-martingale.
Proposition 3.2.4. Let Y be an L
2
loc
(P)-martingale and X adapted and
lcrl. Then
_
XdY is also an L
2
loc
(P)-martingale.
Proof. W.l.o.g. we assume that Y is in fact an L
2
(P)-martingale with
Y
0
= 0 and X is bounded ( the general case then follows via appropriate
stopping as in the proofs before ). Let X
k
E be a sequence such that
X
k
ucp
//
X
. W.l.o.g. this convergence is bounded by a constant c, i.e.
|X
k
| c for all k. Then ( Exercise 3.1.1 (5) )
_
X
k
dY is a martingale.
Furthermore we have
E((
_
t
0
X
k
dY )
2
) = E(
i
X
k
k
i
(Y
t
k
i+1
Y
t
k
i
))
2
c
2
E((Y
t
k
n
k
+1
)
2
) c
2
E(
2
t
)
63
where the last two inequalities are implied by Exercise 3.1.2 and
the Jensen inequality. Hence the sequence
_
t
0
X
k
dY is bounded in L
2
(P)
and so is uniformly integrable. Hence by Exercise 1 Sheet 7 we have
E(
_
t
0
XdY |F
s
) = lim
k
E(
_
t
0
X
k
dY |F
s
)
= lim
k
_
s
0
X
k
dY
=
_
s
0
XdY
Corollary 3.2.1. If Y is a continuous local martingale and X is adapted
and lcrl, then
_
XdY is a local martingale.
Proof. In the proof of Proposition 3.1.4 we showed that any continuous
local martingale is an L
2
loc
(P)-martingale.
Denition 3.2.3. Let Z
n
: 0 =
n
0

n
1
...
n
k
n
< be a sequence of
random partitions. We say Z
n
tends to the identity if
1. lim
n
sup
k

n
k
= a.s.
2. ||Z
n
|| = sup
k
|
n
k+1
n
k
| 0 a.s. as n .
We say that Z
n
is rening, if Z
n
Z
n+1
.
The following Theorem ( without proof ) says that one can compute
the stochastic integral by using random partitions tending to the iden-
tity.
Theorem3.2.1. Let Y S and X adapted and lcrl or rcll. Furthermore
let Z
n
be a sequence of random partitions as in Denition 3.2.3 tending
to the identity. Then
i
X
n
i
(Y
n
i+1
Y
n
i
)
ucp
//
_
X
dY .
64
3.3 Quadratic Variation of a Semi-martingale
Denition 3.3.1. Let X, Y S. The quadratic covariation of X and
Y is dened as
[X, Y ] := XY
_
X
dY
_
Y
dX. (3.3)
The quadratic variation of X is dened as
[X] := [X, X] = X
2
2
_
X
dX. (3.4)
Clearly [X, Y ] depends bilinearly and symmetrically on X and Y .
Furthermore we have the Polarization identity :
[X, Y ] =
1
2
([X + Y ] [X] [Y ]). (3.5)
The following proposition justies the name quadratic covariation.
Proposition 3.3.1. Let X, Y S. Then the process [X, Y ] is adapted
and of nite variation on compacts. If X = Y then [X] is increasing a.s.
Furthermore for any rening sequence of partitions Z
n
: 0 =
n
0
...
n
k
n
< tending to the identity we have
X
2
0
+
i
(X
n
i+1
X
n
i
)(Y
n
i+1
Y
n
i
)
ucp
//
[X, Y ] . (3.7)
Proof. W.l.o.g. we assume X
0
= 0 and X = Y ( the general case follows
from the polarization identity ). We have
(X
2
)
n
k
n
=

k
n
1
i=0
(X
2
)
n
i+1
(X
2
)
n
i
ucp
//
X
2
. (3.8)
We know from Theorem 3.2.1 that
65
i
X
n
i
(X
n
i+1
X
n
i
)
ucp
//
_
X
dY (3.9)
Substratcting 2-times (3.9) from (3.8) under consideration of
(X
2
)
n
i+1
(X
2
)
n
i
= (X
n
i+1
+X
n
i
)(X
n
i+1
X
n
i
)
we get
i
(X
n
i+1
X
n
i
)
2
ucp
//
X
2
2
_
X
dY = [X] .
In the sum above, any summand is positive and the greater the
index t in [X]
t
the more summands are involved in the sum. Hence it
is clear that the process [X] is increasing a.s. and in particular of nite
variation on compacts. That it is adapted follows immediately from the
denition and the corresponding properties of the stochastic integral.
Proposition 3.3.2. Under the assumptions of Proposition 3.3.1 we have
1. [X, Y ] = X
0
Y
0
.
2. [X, Y ] = (X)(Y ).
3. For any stopping time we have [X
, Y ] = [X, Y
] = [X
, Y
] =
[X, Y ]
.
Proof. (1.) and (3.) follow directly from the denition and the corre-
sponding properties of the stochastic integral. To show (2.) we can
assume w.l.o.g. that X = Y . From Proposition 3.2.3 (2) we have
_
X
dX = X
X.
Hence
66
(X)
2
= (X X
)
2
= X
2
2XX
+ X
2
= X
2
X
2
+ 2X
(X
X)
= (X
2
) 2X
X
and
[X] = (X
2
) 2
_
X
dX = (X)
2
.
As an immediate consequence of Denition 3.3.1 we get the follow-
ing proposition :
Proposition 3.3.3. Integration by Parts Formula Let X, Y S then
X Y S and
XY =
_
X
dY +
_
Y
dX + [X, Y ]. (3.10)
We see that the classical Integration by Parts Formula from deter-
ministic analysis is complemented by the quadratic covariation term of
the two semi-martingales.
Denition 3.3.2. For any rcll process X the continuous process X
c
given as
X
c
t
= X
t
0st
(X)
s
(3.11)
is called the continuous part of X.
67
Exercise 3.3.1. Under the assumptions of above, show X
c
is a well
dened continuous stochastic process.
For any X S we have
[X]
c
t
= [X]
t
0st
([X])
s
= [X]
t
0st
(X)
2
s
= [X]
t
X
2
0

0<st
(X)
2
s
.
Denition 3.3.3. X S is called pure quadratic jump, if [X]
c
= 0.
In this case [X]
t
= X
2
0
+
0<st
(X)
s
.
Exercise 3.3.2. Show that any rcll process of nite variation on com-
pacts is pure quadratic jump.
Corollary 3.3.1. Let (W
t
)
t[0,)
be a Brownian motion. Then (W
t
) is
of innite variation on compacts. More generally any continuous semi-
martingale s.t. [X] is not constant a.s. is of innite variation on com-
pacts.
Proof. If X would be of nite variations on compacts, then by Exercise
3.3.2 and continuity we would have [X] = X
2
0
= const. which would
contradict the assumption.
We have seen before, that [W]
t
= t and W
2
t
[W]
t
=
_
t
0
WdW follows
a continuous local martingale.
Exercise 3.3.3. Show by direct computation, that W
2
t
[W]
t
= W
2
t
t
follows a martingale.
This in fact is no coincidence, as the following proposition shows.
Lemma 3.3.1. Let X be a continuous local martingale, then X
2
[X] is
also a continuous local martingale. If [X] 0 then X X
0
is constant.
68
Proof. By denition X
2
[X] = 2
_
X
dX and from Corollary 3.2.1 it

follows that this is a local martingale. We follow from Proposition 3.2.3
that
_
X
dX = X
X 0.
Hence the local martingale X
2
[X] is also continuous. If [X] 0
then X
2
is a non-negative local martingale and hence a super martin-
gale. Let us rst assume that X
0
= 0. Then
E(X
2
t
) E(X
2
0
) = 0
which shows that X 0. If X
0
= 0 then consider

X = X X
0
and
repeat the argument.
Proposition 3.3.4. Let X be a continuous local martingale and
be stopping times. If [X] is constant on [, ] [0, ), then X is too.
Proof. Consider the continuous local martingale

X := X
. Using
Proposition 3.3.2 we have
[

X] = [X
, X
]
= [X
, X
] 2 [X
, X
]
. .
=[X,X]
=[X
]
+[X
, X
]
= [X]
[X]
= 0,
which together with Lemma 3.3.1 implies that

X 0 and hence X
constant on [, ] [0, ).
Proposition 3.3.5. Let X, Y be L
2
loc
(P) martingales. Then [X, Y ] is the
unique adapted rcll process A of nite variation on compacts, s.t.
69
1. XY A is a local martingale.
2. A = XY and A
0
= X
0
Y
0
.
Proof. It follows from Proposition 3.3.1 that [X, Y ] is of nite variation
on compacts and adapted. Let us now show that [X, Y ] indeed satises
conditions (1.) and (2.) above. By denition
[X, Y ] = XY
_
X
dY
_
Y
dX,
and it follows from Proposition 3.2.4 that XY [X, Y ] is a local mar-
tingale. Hence condition (1.) is satised. It follows from Proposition
3.3.2 that (2) is also satised. Now choose A = [X, Y ] and suppose B is
another process satisfying the conditions from above. Then the process
A B = (XY B) (XY A) is a local martingale and
(A B) = A B = XY XY = 0.
Hence A B is a continuous local martingale with (A B)
0
= A
0
B
0
= 0. Since A B is of nite variation on compacts it follows from
Exercise 3.3.2 that [A B] = 0.It then follows from Lemma 3.3.1 that
A B 0, hence A = B, which shows the uniqueness part in the
proposition.
The proposition above can be quite useful in determining the pro-
cess [X, Y ].
Exercise 3.3.4. Let X be an L
2
(P)-martingale, then E(X
2
t
) = E([X]
t
).
The next two propositions show how to compute the quadratic vari-
ation between two stochastic Integrals. We will omit the proofs. They
can be found in the book [Protter].
Proposition 3.3.6. Let Y
1
, Y
2
S and X
1
, X
2
be lcrl adapted processes.
Then
70
[
_
X
1
dY
1
,
_
X
2
dY
2
] =
_
X
1
X
2
d[Y
1
, Y
2
]
Proof. [Protter], page 68.
Proposition 3.3.7. Let Z be rcll adapted and X, Y S. Let Z
n
: 0
n
0
...
n
k
n
be a sequence of random partitions tending to the identity,
then
i
Z
n
i
(X
n
i+1
X
n
i
)(Y
n
i+1
Y
n
i
)
ucp
//
_
Z
d[X, Y ] .
Proof. [Protter], page 69.
3.4 The Ito Formula
Let f be a sufciently smooth function, and Y S be a semi-martingale.
Is f(Y ) still a semi-martingale and if, how can we compute f(Y ) as in
ordinary calculus, where we have the formula f(y) =
_
y
0
f
(x)dx. An el-
ementary example where this question arises is for example the stan-
dard Black-Scholes model, where we have
S
t
= e
(b
1
2
2
)t+W
t
.
Is this a semi-martingale ? The Ito formula will completely answer
all those questions. We will state it in a very general form, where it is
also known as the Ito-Wentzel Formula, but only proof the case where
the semi-martingale Y is continuous. This corresponds to the classical
Ito Formula.
Theorem 3.4.1. Ito-Wentzel Formula Let Y S and f C
2
(R) then
f(Y ) S and
71
f(Y
t
) = f(Y
0
) +
_
t
0
f
(Y
)dY +
1
2
_
t
0
f
(Y
)d[Y ]
c
+
0<st
(f(Y
s
) f(Y
s
) f
(Y
s
)(Y )
s
).
Proof. We assume Y is continuous, where we have to show that
f(Y
t
) = f(Y
0
) +
_
t
0
f
(Y
)
dY +
1
2
_
t
0
f
(Y
)d[Y ].
let us consider the Taylor-expansion of f at the point x R. We
have
f(y) = f(x) +f
(x)(y x) +
1
2
f
(x)(y x)
2
+ R(x, y),
where R(x, y) r(|y x|) (y x)
2
for an increasing function r :
R
+
R
+
with lim
u0
r(u) = 0. W.l.o.g. Y
0
= 0 ( compose f with a simple
translation ). We x t > 0 and denote with Z
n
a rening sequence of
random partitions tending to the identity
Z
n
: 0
n
0
...
n
k
n
= t ...
n
l
n
< .
We have
f(Y
t
) f(Y
0
) =
k
n
1
i=0
(f(Y
n
i+1
) f(Y
n
i
)
=
k
n
1
i=0
f
(Y
n
i
)(Y
n
i+1
Y
n
i
)
+
k
n
1
i=0
f
(Y
n
i
)(Y
n
i+1
Y
n
i
)
2
+
k
n
1
i=0
R(Y
n
i+1
, Y
n
i
).
Using Theorem 3.2.1 and Proposition 3.3.7 we nd that
72
k
n
1
i=0
f
(Y
n
i
)(Y
n
i+1
Y
n
i
)
P
//
_
t
0
f
(Y
)dY
k
n
1
i=0
f
(Y
n
i
)(Y
n
i+1
Y
n
i
)
2
P
//
1
2
_
t
0
f
(Y
)d[Y ]
Furthermore we have
|
k
n
1
i=0
R(Y
n
i+1
, Y
n
i
)| sup
i
r(|Y
n
i+1
Y
n
i
|)
k
n
1
i=0
(Y
n
i+1
Y
n
i
)
2
,
where the second expression on the right side converges in prob-
ability to [Y ]
t
. For each xed the path s Y
s
() is continuous
and hence uniformly continuous on the compact interval [0, t]. Since
lim
n
sup
i
|
n
i+1

n
i
| = 0 this implies lim
n
sup
i
|Y
n
i+1
Y
n
i
| = 0 and
hence using the properties of the function r
lim
n
sup
i
r(|Y
n
i+1
Y
n
i
|) = 0.
which implies that |
k
n
i=0
R(Y
n
i+1
, Y
n
i
)|
P
//
0 .
There is also a multidimensional version of the Ito-formula. The
proof is just a multidimensional modication of the previous proof, so
we will omit it.
Theorem 3.4.2. Let Y = (Y
1
, ..., Y
n
) be an n-tuple of semi-martingales
and f C
2
(R
n
, R), then f(Y ) S and
f(Y
t
) = f(Y
0
) +
_
t
0
n
i=1
x
i
f(Y
)dY
i
+
1
2
_
t
0
n
i,j=1
2
x
i
x
j
f(Y
)d[Y
i
, Y
j
]
c
+
0<st
(f(Y
s
) f(Y
s
)
n
i=1
x
i
f(Y
s
)(Y )
s
).
73
As one can see, the transformation rules can be quite complicated
since in general second order terms are involved. To circumvent this
on can introduce the so called Fisk-Stratonovich-integral.
Denition 3.4.1. Let X, Y S. We dene the Fisk-Stratonovich-
integral of X with respect to Y as
_
X dY :=
_
X
dY +
1
2
[X, Y ]
c
.
The Fisk-Stratonovich integral transformsimpler than the Ito-integral
as the following proposition shows.
Proposition 3.4.1. Let Y S and f C
3
(R), then
f(Y
t
) = f(Y
0
) +
_
t
0
f
(Y
) dY +
0st
(f(Y
s
) f(Y
s
)).
Exercise 3.4.1. Prove Proposition 3.4.1.
Of course there is also a multidimensional version of Proposition
3.4.1. Let us study two important examples on how the Ito-formula
can be applied. The rst one gives a characterization of Brownian mo-
tion, the second one is about the stochastic exponential, which in the
framework of stochastic analysis is what the exponential function is in
standard calculus.
Proposition 3.4.2. L evys characterization of Brownian motion
: Let X = (X
t
, F
t
)
t[0,T]
be a stochastic process. Then the following two
statements are equivalent :
1. X is a standard Brownian motion ( on the interval [0, T] )
2. X is a continuous local martingale s.t. [X]
t
= t t [0, T].
Proof.
74
Denition 3.4.2. Let X S. The stochastic exponential of X is dened
as the stochastic process
E(X)
t
:= exp(X
t
1
2
[X, X]
c
t
)
0<st
(1 + X
s
)exp(X
s
).
Proposition 3.4.3. Let Z S such that Z > 0 a.s. Then X :=
_
1
Z
dZ
is well dened and the unique semi-martingale which satises Z = Z
0
E(X).
Proof. We only give a proof for the case when Z is continuous. We apply
the Ito formula to the function
f : R
2
R
(x, y) e
x
1
2
y
which obviously satises E(X)
t
= f(X
t
, [X]
t
). Hence we get
dE(X) = E(X)dX
1
2
E(X)d[X] +
1
2
E(X)d[X]
= E(X)dX.
We also have dX =
1
[
Z]dZ which is equivalent to dZ = ZdX. The
Integration by parts formula yields :
d(
Z
E(X)
) = Zd(
1
E(X)
) +
1
E(X)
dZ + d[Z,
1
E(X)
].
Applying the Ito-formula to
1
E(X)
= f(X, [X]) yields :
75
d(
1
E(X)
) = (E(X))
1
dX +
1
2
E(X)
1
d[X] +
1
2
E(X)
1
d[X]
= E(X)
1
(dX + d[X]).
d(
Z
E(X)
) = E(X)
1
(ZdX + Zd[X] + dZ
..
=ZdX
d[Z, X]
. .
=Zd[X]
= 0.
hence Z/E(X) const. and since E(X)
0
= 1 this constant must be
Z
0
. If Y S. So we have proven Z = Z
0
E(X). If Y S would
be another semi-martingale which satises Z = Z
0
E(Y ) then by the
same argumentation as for X we have
dY =
1
E(Y )
dE(Y ) =
1
Z
dZ = dX
which shows X = Y . Hence X is uniquely determined by this prop-
erty.
Remark 3.4.1. An immediate consequence of Proposition 3.4.3 is that
for any X S we have dE(X) = E(X)
dX.
3.5 The Girsanov Theorem
Assume you have been given two equivalent measures P Q and a
semi-martingale X = M + A decomposed into a local martingale ( un-
der P ) part and a process A which has nite variation on compacts. In
general M is no local martingale under Q. Nevertheless the Girsanov
theorem says, there is a decomposition of X = N + C into a local mar-
tingale ( under Q ) part and a process C of nite variation on compacts
and it tells you precisely how to compute N and C. We will see that this
theorem is the main tool in the computation of equivalent martingale
measures for nancial markets.
76
Theorem 3.5.1. ( Girsanov ) Let P Q equivalent probability mea-
sures and let Z
s
= E
P
(
dQ
dP
|F
s
) denote the density process. Furthermore
let X = M + A where M is a local martingale under P and A is of nite
variations on compacts. Then the process N dened as
N
t
= M
t
_
t
0
1
Z
s
d[Z, M]
s
0st
(
1
Z
)
s
(Z)
s
)(M)
s
is a local martingale under Q, C := X N is of nite variation on
compacts and X = N + C.
Before we can proof this, wee need two additional lemmas.
Lemma 3.5.1. If Y S and X is rcll of nite variation on compacts.
Then
[X, Y ] = X
0
Y
0
+
0<st
(X)
s
(Y )
s
.
Proof. If Y is continuous for any sequence Z
n
: 0 =
n
0
...
n
k
n
<
of random partitions tending to the identity, we have
|
i
(X
n
i+1
X
n
i
)(Y
n
i+1
Y
n
i
) sup
i
|Y
n
i+1
Y
n
i
|
. .
0 continuity !
_
|dX|
. .
<
0
Hence it follows from Proposition 3.3.1 that [X, Y ]
c
= [X, Y ]X
0
Y
0
=
0. In case Y is not continuous we show that [X, Y ]
c
= [X, Y
c
]. Then we
can conclude from above that [X, Y ]
c
= [X, Y
c
]
c
= 0. We have
i
(X
n
i+1
X
n
i
)(Y
c,
n
i+1
Y
c,
n
i
) [X, Y
c
]. (3.12)
77
We can assume that between
n
i
and
n
i+1
X and Y jump at most one
time. Then
i
(X
n
i+1
X
n
i
)(Y
c,
n
i+1
Y
c,
n
i
) =
i
(X
n
i+1
X
n
i
)(Y
n
i+1
Y
n
i
n
i
<s
n
i+1
(Y )
s
)
=
i
(X
n
i+1
X
n
i
)(Y
n
i+1
Y
n
i
i
(X
n
i+1
X
n
i
)
. .
n
i
<s
n
i+1
(X)
s
)
n
i
<s
n
i+1
(Y )
s
)
. .
0<st
(X)
s
(Y )
s
which shows that the expression on the left side in (3.12) also con-
verges to [X, Y ]
c
and hence [X, Y
c
] = [X, Y ]
c
.
Lemma 3.5.2. Let Q P be two equivalent measures and let Z
t
= E
P
(
dQ
dP
be the corresponding density process. Then the following two statements
are equivalent :
1. M is a local martingale under Q.
2. M Z is a local martingale under P.
Proof. For a localizing sequence
n
of stopping times, we have lim
n
n
=
0 Pa.s. lim
n
n
= 0 Qa.s. Hence after stopping we can restrict our-
selves to martingales. Let 0 s t and A F
s
. Then r s
_
A
M
r
dQ =
_
A
M
r
dQ
dP
dQ =
_
A
M
r
Z
r
dP.
Hence we have
_
A
M
t
dQ =
_
A
M
s
dQ
_
A
M
t
Z
t
dP =
_
A
M
s
Z
s
dP
78
which implies
E
Q
(M
t
|F
s
) = M
s
E
P
(M
t
Z
t
|F
s
) = M
s
Z
s
.
Proof. ( of Theorem 3.5.1 ) Z,M and by Proposition 3.3.3 also
Z M [Z, M] =
_
Z
dM +
_
M
dZ (3.13)
are local martingales under P. It follows from
1
Z
s
= E
Q
(
P
Q
|F
s
)
that
1
Z
s
is a local martingale under Q. It follows from Lemma 3.5.2
and (3.13) that
M
1
Z
[Z, M] =
1
Z
(
_
Z
dM +
_
M
dZ) (3.14)
is a local martingale under Q. Integration by parts ( under Q) yields
1
Z
[Z, M] =
_
1
Z
d[Z, M] +
_
[Z, M]d(
1
Z
) + [[Z, M],
1
Z
].
let us dene
N :=
_
[Z, M]d(
1
Z
). (3.15)
We have
1
Z
t
[Z, M]
t
=
_
1
Z
s
d[Z, M]
s
+

N
t
+ [
FV
..
[Z, M],
1
Z
]
t
=
..
Lemma 3.5.1
_
1
Z
s
d[Z, M]
s
+

N
t
+
0st
(
1
Z
)
s
[Z, M]
s
. .
(Z)
s
)(M)
s
.
Since the sum of two local martingales is a gain a local martingale
79
we get by adding (3.15) and (3.14) a local martingale under Q
N
t
+M
t
_
t
0
1
Z
s
d[Z, M]
s

N
t
0st
(
1
Z
)
s
(Z)
s
)(M)
s
= M
t
_
t
0
1
Z
s
d[Z, M]
s
0st
(
1
Z
)
s
(Z)
s
)(M)
s
but this is exactly the process N dened in the Theorem. Further-
more
C
t
= X
t
N
t
= A
t
+M
t
N
t
= A+
_
1
Z
s
d[Z, M]
s
+
0st
(
1
Z
)
s
(Z)
s
)(M)
s
is of nite variation on compacts. This nishes the proof.
Remark 3.5.1. If in the setting of the Girsanov Theorem either the local
martingale part M under P or the density process Z is continuous, then
N = M [

Z, N]
where

Z =
_
1
Z
dZ satises E(

Z) = Z.
The Girsanov Theorem tells you, how you can decompose a semi-
martingale in its local martingale part and its nite variation part after
a change of measure. The following proposition shows how to apply the
Girsanov Theorem effectively in the Brownian motion setting. In fact
this proposition is also sometimes called the Girsanov Theorem.
Proposition 3.5.1. Let W be a d-dimensional Brownian motion on
(, F, P) with respect to the ltration (F
t
). Let H be a lcrl adapted d-
dimensional process such that the local martingale E(
_
HdW) is in
fact a martingale, then the process dened by
80
X
t
:= W
t
+
_
t
0
H
s
ds
is a Brownian motion on [0, T] with respect to the probability mea-
sure Q dened by Q(A) = E
P
(1
A
Z
T
) where Z
T
:= E(
_
HdW)
T
.
Proof. W.l.o.g. we assume d = 1. It follows from Remark 3.4.1 that
dZ
t
= Z
t
d(
_
t
0
H
s
dW
s
) = Z
t
H
t
dW
t
. Clearly
dQ
dP
= Z
T
and since Z
T
is a martingale, we also have E
P
(
dQ
dP
|F
t
) = Z
t
. Hence we are in the
framework of the Girsanov Theorem and it follows that
N
t
:= W
t
_
t
0
1
Z
s
d[Z, W]
s
is a continuous local martingale under Q. Furthermore by Proposi-
tion 3.3.6 we have
[Z, W]
t
=
_
t
0
Z
s
H
s
d[W, W]
s
=
_
t
0
Z
s
H
s
ds.
This implies that X
t
= W
t
+
_
t
0
H
s
ds = N
t
is a continuous local mar-
tingale under Q. Since
_
t
0
H
s
ds is of nite variation on compacts we
have [X]
t
= [W]
t
= t and the proposition follows from the L evy charac-
terization of Brownian motion.
3.6 The Stochastic Integral for predictable
Processes
In this section we will enlarge the set of integrands. So far we are
able to integrate lcrl processes with respects to semi-martingales. The
space of lcrl processes seem to be large but is still not large enough. In
the economic setting we think of the space of integrands as the space
of trading strategies. We wish that our nancial market is complete (
at least the Black-Scholes Versions ). To achieve this, lcrl processes are
unfortunately not enough. In this section we sketch how to extend the
81
stochastic integral to predictable processes. We will make use of the
following Theorem which translates the term semi-martingale into the
more familiar terms of local martingales and nite variation processes.
Theorem 3.6.1. Characterization of Semi-martingales Let X be
an adapted rcll process. Then the following things are are equivalent :
1. X S.
2. X is decomposable ( see Denition 3.1.5 ).
3. Given > 0 there exists a local martingale M such that |M|
a.s. and M
0
= 0, a process A which is of nite variation on
compacts s.t. A
0
= 0 and X
t
= X
0
+ M
t
A
t
.
The decomposition in (3) is in fact unique, if one assumes one extra
condition on the process A which is called naturality.
Denition 3.6.1. For a semi-martingale Y decomposed as in (3) of The-
orem 3.7.1 in Y
t
= X
0
+ M
t
+ A
t
we dene
||Y ||
H
2 = (E([M]
))
1/2
+ (E(
_

0
|dA
s
|
2
))
1/2
.
where [M]
= lim
t
[M]
t
and |dA
s
| is the total variation of A as
dened in Proposition 3.1.1. We dene the space
H
2
:= {Y S : ||Y ||
H
2 < }.
Denition 3.6.2. The -algebra
P := (X|X : I R lcrl adapted )
is called the predictable -algebra on I. A process X is called
predictable if it is P/B(R) measurable.
82
Clearly any lcrl process is predictable. In the following Denition
we dene a metric on the set of all predictable processes. This metric
depends on the semi-martingale Y , which serves as the integrator.
Denition 3.6.3. Let Y H
2
and X,

X predictable processes. We dene
d
Y
(X,

X) := (E(
_

0
(X
s

X
s
)d[M]
s
))
1/2
+ (E(
_

0
|X
s

X
s
||dA
s
|
2
))
1/2
.
The integrals in this denition are path-wise Riemann-Stieltjes inte-
grals.
With these denitions we are now able to generalize the stochastic
integral to the case of predictable integrands.
Theorem 3.6.2. Let Y H
2
and X a predictable process with paths
that are bounded on compact intervals. Then there exists a sequence X
n
of lcrl processes such that
X
n
d
Y
//
X
and lim
n
X
n
dY exists in H
2
.
We dene
_
XdY as this limit.
The following is called the Ito-isometry.
Proposition 3.6.1. Let Y H
2
a local martingale and X a predictable
process with paths that are bounded on compact intervals. Then
E((
_

0
XdY )
2
) = ||
_

0
XdY ||
H
2 = E(
_

0
X
2
t
d[Y ]
t
).
Proof. After stopping we can assume that
_
XdY is a martingale. It
then follows from Exercise .. that
E((
_

0
XdY )
2
) = E(lim
t
[
_
XdY ]
t
)
= E(lim
t
_
t
0
X
s
d[Y ]
s
)
= E(
_

0
X
s
d[Y ]
s
)
83
Taking Y = W a Brownian motion we get the following Corollary :
Corollary 3.6.1. E((
_
T
0
X
s
dW
s
)
2
) = E(
_
t
0
X
2
s
ds).
3.7 The Martingale Representation Theo-
rem
In this section we present and prove the martingale representation the-
orem, which has many application, but for us its relation to the ques-
tion of completeness of nancial markets is most important. We will
study this relation in the next chapter.
Theorem3.7.1. Let Wbe a n-dimensional Brownian motion and (F
t
) =
(F
B
t+
) be the augmented, right continuous Brownian Filtration. Further-
more let Y be a m-dimensional rcll local martingale with respect to the
ltration (F
t
). Then there exists a predictable (m n)-matrix valued
process X such that
Y =
_
XdW.
84
Chapter 4
Explicit Financial Market
Models
In this chapter we give explicit models for nancial markets. Of course
we can only present some of the large variety of nancial models, but
we try to give at least some example for any class of market model, such
there is generalized Black Scholes, diffusion type stochastic volatility
and markets with jump components. Throughout the whole chapter
the triple (, F, P) stands for a complete probability space.
4.1 The generalized Black Scholes Model
This model is a generalization of the model presented in Chapter 1 in
the sense that it does not assume that the trend and the volatility are
given constants, but stochastic processes which are adapted to the un-
derlying ltration. For the underlying ltration we take F
t
= F
W
t+
the
augmented and right-continuous Brownian ltration which is gener-
ated by a m-dimensional Brownian motion W on (, F, P). The price
process X = (X
0
, X
1
, ..., X
n
) which contains only tradeable assets and
is modeled over the nite interval [0, T] is given as
85
X
0
t
= exp(
_
t
0
r
s
ds)
X
i
t
= X
i
0
exp(
_
t
0
(b
i
s
1
2
m
j=1
2
ij,s
)ds +
m
j=1
2
ij,s
dW
j
s
) , i {1, .., n}
Here (r
t
), (b
t
) and (
t
) are scalar resp. vector resp. matrix valued
predictable ( with respect to the ltration (F)
t
) stochastic processes
which are pathwise bounded. We do further assume that there exists a
constant K > 0 such that ||
t
x|| K||x|| for all x R
m
. This property
is often called uniform coercitivity. using the Ito-formula we can write
the price-processes in differential form :
dX
0
t
= X
0
t
r
t
dt
dX
i
t
= X
i
0
(b
i
t
dt +
m
j=1
2
ij,t
dW
j
t
) , i {1, .., n}.
Application of the Integration by parts formula on

X
i
t
= X
i
t
(X
0
t
)
1
gives
d

X
0
t
= 0
d

X
i
t
=

X
i
0
((b
i
t
r
t
) +
m
j=1
2
ij,t
dW
j
t
) , i {1, .., n}.
For the set of trading strategies we assume the following : =
(
0
,
1
, ...,
n
) if it is a predictable R
n+1
-valued stochastic process
process such that
1.
_
T
0
|
0
t
|dt < and
_
T
0
(
i
t
X
i
t
)
2
dt < a.s.
2. V
t
() =
t
X
t
=
_
t
0
n
i=1
i
t
dX
i
t
d(

X) = d

X self-nancing
3. The price-process V
t
() is bounded from below. tameness
86
The equivalence in two follows from application of the Integration
by parts formula. We denote the corresponding nancial market as
M
genBS
= ((X
t
, F
t
)
t[0,T]
, )
and call it the generalized Black-Scholes nancial market. Let
us now compute the equivalent martingale measures for this nancial
market. Assume that P
P(M
genBS
) and let Z
t
:= E
P
(
dP
dP
|F
t
) denote
the corresponding density process. Clearly Z is a strictly positive mar-
tingale with respect to the Brownian ltration, hence by Theorem 3.7.1
continuous.

X
0
1 is a martingale in any case but for

X
i
for 1 i n
to be a local martingale under P
, it follows from the Girsanov theorem

( Theorem 3.5.1 and Remark 3.5.1 ) that we must have
X
i

X
0
= M
i
[

Z, M
i
] (4.1)
where

X
i
=

X
i
0
+ M
i
+ A
i
is the decomposition of X into initial
value, local martingale part and nite variation part under the mea-
sure P ( see Theorem 3.6.1 (3) ) and

Z =
_
1
Z
dZ. Since

Z is also a local
martingale with respect to the Brownian ltration, by Theorem 3.7.1 (
Martingale Representation Theorem ) there must be an m-dimensional
predictable process such that
d
Z = dW=
m
j=1
j
dW
j
. (4.2)
If we use this form of

Z to compute the covariation process in (4.1)
we get
d[

Z, M
i
]
t
=

X
i
t
m
j=1
ij,t
j
t
dt (4.3)
which we can use to compute d

X
i
as
87
d

X
i
t
=

X
i
t
(
m
j=1
ij,t
j
t
dW
j
t

m
j=1
ij,t
j
t
dt). (4.4)
Since the decomposition into initial value, local martingale part and
nite variation part is unique by comparison with the expression for
d

X
i
on the previous page we must have b
i
t
r
t
=
m
j=1
sigma
ij,t
j
t
for
1 i n. Using vector notation we can write this as
t

t
= r
t
b
t
a.s t [0, T]
where r
t
= (r
t
, ..., r
t
)
T
. .
m
is the vector which has r
t
in any component.
Such a process is called a state price process or market price of
risk . For any such process we dene the corresponding measure P
via P
(A) = E
P
(1
A
Z
T
) where Z
T
= E(
_
dW) ( here we need that
t
is pathwise bounded, which follows from the uniform coercitivity of
t
) and the discussion above shows that P
P(M
genBS
). Hence we have
proven the following theorem :
Theorem 4.1.1. P(M
genBS
) = {P
| is a state price process }.

If
t
does not have full rank, then in general there is no such
at all. In this case no equivalent martingale measure exists. It can
be shown that in this case the market is not arbitrage free. If the
t
has full rank, then more then one state processes may exist. In this
case the market is arbitrage free, but their is no unique fair price for a
contingent claim on this market ( of course however if traders x one
of these prices everything works ne and no one can take advantage ).
The best situation is, if
t
is invertible a.s. for all t [0, T]. In this case
t
must be
1
t
(r
t
b
t
) and we have :
Theorem 4.1.2. If det(
t
) = 0 a.s. for all t [0, T] then we have
P(M
genBS
) = {P
|
t
=
1
t
(r
t
b
t
)}
and |P(M
genBS
)| = 1.
88
In the last case we have a unique equivalent martingale measures
and hence fair prices can only be computed by one formula. For a one
dimensional model where the processes r
t
, b
t
and
t
are constants and
a European call option this is the price computed in Theorem 2.7.1. Let
us now consider the question, whether arbitrage freeness implies the
existence of martingale measures. In fact we can proof the following
theorem :
Theorem 4.1.3. If M
genBS
is arbitrage free, then P(M
genBS
) = . Hence
M
genBS
satises the Fundamental Law of Asset Pricing ( see def. 2.3.6
).
Proof. Let be a trading-strategy. Assume there exists A [0, T]
s.t. ( P)(A) > 0 and
T
= 0
T
(b r) = 0
on A, where r denotes the vector with all entries equal to r. Let us
dene a new process
i
as
i
:=
1
X
i
t
sgn(
T
(b r))
i
1
A
for i = 1, .., n
and
0
s.t. = (
0
,
1
, ...,
n
) is self-nancing and V
0
() = 0 ( this is
always possible ). Here sgn denotes the sign function, which is 1, if the
argument is greater or equal than zero and 1 else. We have
89
d(
V
t
()) =
n
i=1
i
t
d

X
i
t
=
n
i=1
i
t
X
i
t
((b
i
t
r
t
)dt +
m
j=1
ij,t
dW
j
t
)
= 1
A
i=1
sgn(
T
(b r))
t

i
1
X
i
t
X
i
t
((b
i
t
r
t
)dt +
m
j=1
ij,t
dW
j
t
)
= 1
A
|
T
t
(b
t
r
t
)|
d
t + 1
A
sgn(
T
t
(b
t
r
t
))
n
j=1
(
n
i=1
i
t

ij,t
)
. .
=(
T
)
j
=0 onA
dW
j
t
= 1
A
|
T
t
(b
t
r
t
)|dt.
Clearly

V
t
() =
_
t
0
d
V
s
() =
_
t
0
1
A
|
T
(b r)|dt 0 for all 0 t T
and since (
P)(A) > 0 and |

T
(b r)|dt > 0 on A we have
E(
V
T
()) = E(
_
T
0
1
A
|
T
t
(b
t
r
t
)|dt
=
_
[0,T]
1
A
|
T
t
(b
t
r
t
)|d( P) > 0.
This implies P(V
t
() > 0) = P(
V
t
() > 0) > 0 and is an arbitrage
strategy. By our assumption, the market is arbitrage free, such a trad-
ing strategy cannot exist. Hence any A as chosen on the beginning of
this proof must satisfy (
P)(A) = 0 and
T
= 0 (b r) a.s.
Therefore we have (b r) (ker(
))
= im() and there exists a

process s.t.
= b r
( the argumentation here is not really precise, one has to show that
90
this can be chosen as a adapted ( in fact predictable ) process, this is
very technical ). As is Theorem 4.1.1. we have P
P(M
genBS
).
Let us now consider the question of completeness. We assume that
det(
t
) = 0 a.s. for all t [0, T], hence n = m and there exists a unique
equivalent martingale measure P
= P
and by Proposition 3.5.1

W
t
= W
t
_
t
0
s
ds
is a Brownian motion under P
. We also have
d

X
i
t
=

X
i
t
n
j=1
ij,t
dW
j
t
.
Let us consider a contingent claim g. We assume that
E
P
(e
_
T
0
r
t
dt
g) < . (4.5)
The last condition is equivalent to E
P
(e
_
T
0
r
t
dt
g) < by the explicit
form of P
. We consider the martingale ( with respect to the Brownian

ltration )
g
t
:= E
P
(e
_
T
0
r
t
dt
g|F
t
). (4.6)
It follows from Theorem 3.7.1 that there exists a predictable R
n
val-
ued stochastic process H such that
g =
_
H dW. (4.7)
We dene the matrix valued process as

ij,s
:= X
i
s
ij,s
, 1 i, j n. (4.8)
Clearly det(
s
) = (
n
i=1
X
i
s
) det(
sigma
s
) = 0 a.s. for all t [0, T]. We dene
s
= (
1
s
, ...,
n
s
) by
91

s
:= H
s

1
s

i
s
=
n
k=1
H
k
s

1
ki,s
, 1 i n. (4.9)
Then we have
n
i=1
i
s
d

X
i
s
=
n
i=1
i
s
n
j=1
X
i
s
ij,s
. .

ij,s
dW
j
s
=
n
j=1
n
k=1
H
k
s
(
n
i=1

1
ki,s

ij,s
)
. .
=
kj
dW
j
s
=
n
j=1
H
j
s
dW
j
s
= d g
s
.
we dene
0
s
:= g
s
n
j=1
j
s
X
j
s
. We show that = (
0
,
1
, ...,
n
) is a
self-nancing trading-strategy. We have
V
t
() =
t

X
t
= ( g
t
j=1
j
t

X
j
t
) 1 +
n
j=1
j
t

X
j
t
= g
t
which then implies
d
V
t
() = d g
t
=
n
i=1
i
t
d

X
i
t
=
..
d

X
0
t
=0
t
d

X
t
.
Hence is self-nancing. Last but not least we have
V
T
() =

V
T
() X
0
T
= g
T
e
_
T
0
r
t
dt
= E
P
(g e
_
T
0
r
t
dt
|F
T
) e
_
T
0
r
t
dt
= g.
Since also satises conditions (1) and (3) on page 74 we have a
found a hedge for the contingent-claim g. hence we have proven
the following theorem :
92
Theorem 4.1.4. If in the generalized Black-Scholes nancial market
model det(
t
) = 0 a.s. for all t [0, T], then this model is complete in the
sense that any contingent claim g which satises E
P
(e
_
T
0
r
t
dt
g) <
can be hedged by a trading strategy .
Corollary 4.1.1. If |P(M
genBS
)| = 1 then M
genBS
is complete in the
sense above.
The inverse implication is left as an exercise.
Exercise 4.1.1. Show if M
genBS
is complete, then |P(M
genBS
)| = 1.
The following theorem summarizes the results :
Theorem 4.1.5. The following two statements are equivalent :
1. |P(M
genBS
)| = 1
2. M
genBS
is complete.
4.2 A simple stochastic Volatility Model
Let us assume that on (, F, P) there exists a 1-dimensional Brownian
motion (W
t
) and an independent process (
t
) given by
t
:=
_
_
2 0 t with probability 1
1 < t T with probability 1/2
3 < t T with probability 1/2
where : [0, T] is a stopping time and F
t
:= (W
s
,
s
|0 s t)
+
.
Because of the independence of W and (W
t
) is also a Brownian motion
with respect to F
t
( check Denition 1.3.1 ). Dene
X
0
t
:= 1
X
1
t
:= X
1
0
exp(
_
t
0
s
dW
s
1
2
_
t
0
2
s
ds) = E(
_
dW)
t
.
93
we consider these as tradeable assets ( Bond and Stock ) and the
volatility
t
as nontradeable asset. The trading-strategies are dened
via the same conditions as in section 4.1. We can think of this market
as a standard Black-Scholes market where at random time the mar-
ket conditions change ( either the market gets more volatile or it calms
down ). However, since the numeraire is chosen to be 1 we have

X = X
and since by Remark 3.4.1 P P(M). On the other side, assume we
have a measure P
such that
P
F
W
T
= P
F
W
T
P
(
T
= 1) = p
P
(
T
= 3) = 1 p
for any 0 < p < 1 then P
P and P
P(M) ( E(
_
dW) remains a
martingale under P
. It is not hard to show, that the value p determines

the martingale measure uniquely and hence that
P(M)
= (0, 1)
where the right hand side denotes the open interval. Since martin-
gale measures exist, the market above is arbitrage free. Is it complete
? For this, let us consider the contingent claim g =
T
. Since (
t
) is also
a martingale with respect to P ( easy exercise ! ) we have
t
= E(g|F
t
).
Assume now that g could be hedged, i.e. there would exist
such that
T
X
T
=
T
. Since must be self-nancing we have
t
X
t
=
_
t
0

s
dX
s
and this is a martingale under P. Hence
t
X
t
= E(
T
X
T
|F
t
) = E(g|F
t
) =
t
.
Since (
_
dW
s
)
s
=
s
(W)
s
0 we have that (X
t
) is continuous.
94
Hence (
t
X
t
) = (
_

s
dX
s
)
s
=
s
(X)
s
0 and
t
X
t
=
t
is contin-
uous with respect to t. This is a contradiction, since
t
jumps at time
. Hence such a trading strategy can not exist and M is not com-
plete. From a rational point of view, the non-completeness is caused by
the extra randomness induced by the volatility and the fact, that this
randomness cannot be traded. The same situation occurs in the more
elaborate stochastic volatility models in the next section.
Exercise 4.2.1. Study how the different values of p determine different
values for the fair price of a European call option.
4.3 Stochastic Volatility Model
let us assume that on the complete probability space (, F, P) there is
given a 2-dimensional Brownian motion W
t
= (W
1
t
, W
2
t
) and that F
t
=
F
W
t
is the Brownian ltration. Consider a market model consisting of
Bond ( Bank account ) : B
t
Stock : S
t
Volatility :
t
where the Bond and the Stock is tradeable, but the volatility is not.
For the trading strategies we assume the same conditions as in section
4.1. To make this model more precise, we assume
B
t
= exp(rt)
S
t
= S
0
exp(
_
t
0
(b
1
2
2
t
)dt + a
_
t
0
t
dW
1
t
+ a a
_
t
0
t
dW
2
t
)
and
t
is predictable and satises
df(
t
) = (
t
)dt + (
t
)dW
2
t
,
0
= const. (4.10)
95
where f : R R is a strictly increasing, two times differentiable
function and , are at least continuous functions ( we will later also
make assumptions on and ). If (4.10) has a solution ( a process
t
which satises (4.10) ) then we get
t
simply by
t
= f
1
(f(
t
)). No-
tice that we can also write equation (4.10) in terms of d by application
of the Ito-formula. The thing is, that sometimes using such an f the
equation looks much nicer. In most cases f will be either f(x) = x or
f(x) = x
2
. The existence of such a
t
in general depends on the func-
tions , and f. The following theorem gives us sufcient conditions
for existence.
Theorem4.3.1. Under the assumptions above, there exists a predictable
process
t
which satises (4.10) if
1. f(x) = x and , are Lipschitz continuous and satisfy the growth
condition
(x)
2
+(x)
2
K (1 +x
2
)
for a constant K > 0.
2. f(x) = x
2
, (x) = ( x
2
) and (x) = x for constants , and .
In both cases
t
is unique up to indistinguishability.
The following is a list of the most famous stochastic volatility mod-
els :
1.) d
t
= (
t
)dt +dW
2
t
Ornstein-Uhlenbeck model
2.) d
t
=
t
dt +
t
dW
2
t
geometric BM model
3.) d
t
=
1
t
(
2
t
)dt +dW
2
t
simple Heston model
4.) d
2
t
= (
2
t
)dt +
t
dW
2
t
Heston model .
Clearly the coefcients of 1.) and 2.) satisfy the assumptions in The-
orem4.3.1 (1) and (4) corresponds exactly to the second case of Theorem
4.3.1. The theorem cannot be applied to equation 3.) however, since the
96
function x
1
x
is not Lipschitz continuous. Nevertheless equation 3.)
can be solved by using the Girsanov theorem. Using the same method
one can compute an explicit solution for 1.). We leave this as an Ex-
ercise. Let us now consider the question of arbitrage freeness of this
market and as always in this case start the quest for equivalent mar-
tingale measures. As in section 4.1 one has to nd a 2-dimensional
process (
s
) s.t. for

Z =
_

s
dW
s
=
_

1
s
dW
1
s
+
_

2
s
dW
2
s
we have
S = M [M,

Z]
where M denotes the martingale part of

S
t
under P. We have
d
S
t
= S
t
(b r)dt + S
t
a
t
dW
1
t
+S
t
a
t
dW
2
t
and hence dM
t
= S
t
a
t
dW
1
t
+ S
t
a
t
dW
2
t
. This implies
d
S
t
= dM
t
d[M,
_

s
dW
s
]
t
= S
t
a
t
dW
1
t
+ S
t
a
t
dW
2
t
(S
t
a
1
t
+ S
t
a
2
t
2
t
)dt.
Uniqueness of the FV part shows
S
t
(b r) = (S
t
a
1
t
+ S
t
a
2
t
2
t
)
which implies that
r b
t
= a
1
s
+ a
2
s
.
Hence we get ( changing from to )
Theorem 4.3.2. In the stochastic volatility models from above, we have
P(M) = {P
|
dP
dP
= E(
_
dW) s.t.
r b
t
= a
1
s
+a
2
s
}.
Under the assumption that we have chosen an equivalent martin-
gale measure for our stochastic volatility model, let us now discuss,
97
how to compute fair prices, when there is no closed form solution for
the volatility process. The keywords are approximation and simula-
tion. To solve the stochastic differential equation we rst approximate
its solution in the same way as in the case of ordinary differential equa-
tion ( Euler method ) and then simulate the approximated solution and
use the Monte Carlo method to compute an approximation of the fair
price of a contingent claim. The approximation procedure we discuss
here is called the stochastic Euler scheme. This is the easiest approxi-
mation scheme. For more elaborate schemes see [?]. Assume now that
X is given as the solution of
dX
t
= (X
t
)dt + X
t
dW
t
, X
0
= s, t [0, T].
Let us assume that N >> 1 so that :=
T
N
<< 1. Then
_
t+
t
(X
s
)ds (X
t
)
_
t+
t
(X
s
)dW
s
(X
t
) (W
t+
W
t
)
. .
N(0,)
where the second factor on the right side of the second equation is
a standard normally distributed with expectation zero and variance
distributed random variable. Hence we get
X
(n+1)
X
n
+ (X
n
) + (X
n
)(W
(n+1)
W
n
).
We use this relationship to implement a recursive scheme. Dene
X
N
0
:= x
X
N
(n+1)
:=

X
N
n
+ (

X
N
n
) + (

X
N
n
)Z
n
Z
n
i.i.d N(0, )
n = 1, .., N.
98
We could now dene the process

X
N
piecewise constant. This leads
to a non continuous process. If we want our approximation to be con-
tinuous we can do this with linear interpolation : For t [i, (i + 1))
we dene
X
N
t
:=

X
N
i
+
t i
(

X
N
(i+1)

X
N
i
).
In any way, this procedure gives us a sequence of stochastic pro-
cesses

X
N
that can easily be simulated on the computer. We would
wish, that if N we have

X
N
X in some sense ( a.s., ucp, in
probability for the nal value...). The following Theorem tells us more.
Theorem 4.3.3. Under the assumptions of Theorem 4.3.1 we have
E(|X
T

X
N
T
|) K
_
1
N
for some constant K > 0. Therefore

X
N
T
L
1
//
X
T
. We speak of strong
convergence of order 1/2. Furthermore for each function g : R R of
polynomial growth rate and only nitely many discontinuities, we have
E(g(X
T
)) E(g(

X
N
T
)) K
1
N
for some other constant K > 0. In this case we speak of weak con-
vergence of order 1.
We do not give a proof of this theorem. Clearly we think of g from
above as a contingent claim. Given a stochastic volatility model on
begins by simulation of the process (
t
) which results in a process (
t
).
This process is then used in a second simulation for the stock price
process :
S
N
i+1
:=

S
N
i
+

S
N
i
r +

S
N
i

i
Z
i
where Z
i
i.i.d N(0, ) and independent from the rst Z
n
used for
99
the simulation of (
t
) ( this corresponds to the case a = 0 ). Then as
above we have :
E(g(S
T
)) E(g(
S
N
T
)) K
1
N
for some constant K > 0 and nally we have
Theorem 4.3.4. Under the assumptions from above :
lim
N
e
rT
E(g(
S
N
T
)) fair price of g.
The last step of the simulation is the computation of the expectation
E(g(
S
N
T
)). This is done by the so called Monte Carlo method, which
relies on the strong law of large numbers. One simulates

S
N
T
n-times
for n >> 1 and gets realizations

S
N,i
T
, i = 1, .., n. By the strong law of
large numbers
1
n

i=1
ng(
S
N,i
T
)
is a good approximation for the fair price of g. This method can also
be used to price options in the standard Black Scholes model ( which by
the way is a stochastic volatility model with d
t
= 0,
t
to compute
fair prices for options where there is no closed form solution available.
A good practical exercise is :
Exercise 4.3.1. Check the quality of the Monte Carlo method by apply-
ing it to the standard Black Scholes model and the European call option
and compare the result to the explicit solution of section 2.7.
In general the quality of this method depends very much on the
underlying model and on the smoothness of the pay-out function g.
4.4 The Poisson Market Model
Let (, F, P) still denote a complete probability space together with a
ltration (F
t
) which satises the usual conditions.
100
Denition 4.4.1. An Z-valued rcll process (N
t
) is called a Poisson
process with intensity if :
1. N
0
= 0
2. N
t
N
s
independent of F
s
3. N
t
N
s
P((ts)) ( Poisson distributed with parameter (ts).
One should compare this denition with the denition of Brown-
ian motion.Besides that the Poisson-process is not continuous, only the
third item has been changed, using the Poisson distribution instead of
the normal distribution. We mentioned in section 1.3. that the Brow-
nian motion can be thought of some kind of random-walk with very
small step-sizes and models the movement of a small particle in some
uid. Let us also give a motivation for the Poisson process. For this let
T
i
be a sequence of exponentially with parameter distributed random
variables ( i.e. P(T
i
t) = 1 e
t
for t 0 and 0 else ). We think of T
i
as the time passing between two similar random events ( for example
the arrival of a customer at some shop ). The S
n
:=
n
i=1
T
i
is the time
when the n-th event happens. We dene
N
t
:= max{n|S
n
t}
as the number of events that happen before time t. One can show
that (N
t
) as dened above is a Poisson process with intensity . From
this description one can also see, that a Poisson process possesses only
jumps of size 1. The following is an easy but useful exercise :
Exercise 4.4.1. Let (N
t
) be a Poisson process of intensity . Show (N
t
t) is a martingale.
It is clear from the denition, that any Poisson process (N
t
) is in-
creasing and hence of nite variation on compacts and also pure quadratic
jump ([N]
c
= 0). Let us consider the following market model :
101
X
t
=
_
B
t
S
t
_
=
_
Bond
Stock
_
.
B
t
1
S
t
= S
0
exp(bN
t
t)
where b, > 0 and (N
t
) is a Poisson process with intensity . For the
set of trading strategies we suppose the same conditions as in section
4.1. We call this market model M
Poisson
. As for any market model,
we are interested whether it is arbitrage free or not. To answer this
question we dene a process Y as
Y
t
:= (exp(b)
1
)N
t
t).
This process is also rcll, has nite variation on compacts and there-
fore satises [Y ]
c
= 0. We have ( see Def. 3.4.2 )
E(Y )
t
= e
Y
t
0<st
(1 + (Y )
s
) exp((Y )
s
)
= e
(exp(b)1)N
t
e
t
0<st
exp(b) (exp((exp(b) 1)))
= e
(exp(b)1)N
t
e
t
exp(b N
t
) ee
(exp(b)1)N
t
= exp(bN
t
t).
Here we used N
t
=
0<st
(N)
s
as well as (Y )
s
= (exp(b)
1)(N)
s
and hence (N)
s
= 0 (Y )
s
= 0. This shows that
S
t
= S
0
E(Y ).
If we can now nd a measure P
P under which Y is a local mar-

tingale, then also S is a local martingale under P
. This follows from

proposition 3.4.3. To compute such a P
we need a version of the Gir-

102
sanov Theorem for Poisson processes.
Theorem 4.4.1. Let (N
t
) be a Poisson process with intensity under
the measure P and T > 0. Dene a new measure P
via
dP
dP
=

N
T
e
(
)
T.
Then (N
t
)
t[0,T]
is a Poisson process on [0, T] under P
with intensity
.
Proof. [?], page 246.
Let us now dene
:= /(exp(b) 1) and dene P
with respect
to this
as in Theorem 4.4.1. Then under P
(N
t
) is a Poisson pro-
cess with intensity
. Hence it follows from Exercise 4.4.1. that

(N
t

t)
t[0,T]
is a P
martingale. Since Y
t
= (exp(b) 1)N
t
t =
(exp(b) 1)(N
t

t) it follows that also Y is a martingale under P
and as mentioned before that

X = X is a martingale under P
. Hence
P
P(M
Poisson
) and M
Poisson
is arbitrage free. One can also show
that M
Poisson
is complete. For this one needs some kind of martin-
gale representation theorem for Poisson processes ( we leave this out ).
Also P(M
Poisson
) = {P
}. Let us now consider a European call option

g = (S
T
K)
+
in M
Poisson
where K denotes the strike price. For the fair
price we have
= E
P
(g)
= E
P
((S
0
exp(bN
T
T) K)
+
)
=
n=0
1
n!
(
T)
n
(S
0
exp(bn T) K)
+
.
The following exercise is very interesting. It should be solved by
experimenting on the computer.
Exercise 4.4.2. In the standard Black/Scholes model take parameters
S
0
= 10, = 0.4,r = 0.025 and T = 5. Compute the fair price of a
103
European call with strike price K = 10 with the formula in section 2.7.
What parameters , b and should one take to get approximately the
same price for the same option in the Poisson market model. Then look
what happens if you slightly change K.
The pure Poisson market model is not of practical use but good to
demonstrate the general concept. Nevertheless it is sometimes used in
combination with the standard Black-Scholes model.
104
Chapter 5
Portfolio Optimization
5.1 Introduction
Consider a nancial market M. Up to now we were mainly concerned
in questions like how to characterize arbitrage freeness, completeness
and how to compute fair prices for contingent claims. A different ques-
tion is how to choose a trading strategy that in some sense performs
in an optimal way. Optimality will be dened via a so called utility
function U which may depend on the terminal wealth of the portfolio
but also on possible consumption. It is clearly that a greater terminal
wealth should have a greater utility, but the utility function should also
pay respect to the fact that trading in nancial markets bears risks.
Also different traders may use different utility functions correspond-
ing to their individual attitude towards risk. However there are some
quite general rules :
1. more money is always better ( has a greater utility ) U is mono-
tonically increasing
2. if your wealth is 1000 Euro your individual increase in utility from
gaining another 1000 Euro is much bigger than if your wealth
would be 1000000 Euro ( change in utility can be understood as
what effect has the gain on your life ) U
(x) is monotonically
decreasing
105
3. if your wealth is innite ( means you already have everything )
you utility can no longer increase lim
x
U
(x) = 0 and vice

versa, if your wealth is zero ( you have nothing ) then your life is
changed dramatically, by any gain lim
x
U
(x) = 0
In this section we do only consider the generalized Black-Scholes
model from section 4.1 in the case where det(
s
) = 0. This is the situ-
ation where we have a complete nancial market and a unique equiv-
alent martingale measure and do not have to worry which equivalent
martingale measure we use. In contrast to what we have done before
though, we now also allow the trader to consume some oft he money he
has in the market. Mathematically precise :
Denition 5.1.1. A consumption strategy is a non-negative predictable,
real valued stochastic process (c
t
)
t[0,T]
s.t.
_
T
0
c
t
dt < P a.s..
Remark 5.1.1. (c
t
) models the consumption-rate,
_
T
0
c
t
dt the money the
trader takes from the market for consumption over the time interval
[0, t].
We do now consider pairs (, c) consisting of a trading strategy
which satisfy the conditions 1.) and 2.) in section 4.1 ( page 74 ) and a
consumption strategy c where condition 3.) in section 4.1. is replaced
by the following self-nancing condition :
dV
t
() =
t
dX
t
c
t
dt.
Since consumption takes place over the whole time interval [0, T] we
must allow a utility function also to depend on a time parameter.
Denition 5.1.2. 1. U : (0, ) R is called a utility function if it
is strictly concave, C
1
and U
(0) := lim
x0
U
(x) = as well as
U
() := lim
x
U
(x) = 0.
106
2. If U : [0, T] R also depends continuously on time and t
U(t, ) is a utility function in the sense of 1.) then U is also called a
utility function.
It is easy to check that the following functions are utility functions.
1. U(x) = ln(x)
2. U(x) =
_
(x)
3. U(x) = x
for 0 < 1
4. U(t, x) = e
t
U
1
(x) for > 0 and U
1
as in Denition 5.1.2.
Denition 5.1.3. For a self-nancing pair (, c) consisting of a trading
strategy and a consumption strategy s.t V
0
() = x as well as utility
functions U
1
: [0, T] (0, ) R and U
2
: (0, ) R we dene the
expected utility as
J(x, , c) = E(
_
T
0
U
1
(t, c
t
)dt + U
2
(V
T
())).
We allow the expression in Denition 5.1.3. to be innite ( there is
nothing to complain about innite utility ) but for technical reasons we
assume that the negative utility has nite expectation, i.e.
E(
_
T
0
U
1
(t, c
t
)
dt + U
2
(V
T
()
)) < .
The aim of Portfolio Optimization is then to compute
max
(,c)A
(x)
J(x, , c)
A
(x) = {(, c) self-nancing pair V

0
() = x,
E(
_
T
0
U
1
(t, c
t
)
dt + U
2
(V
T
()
)) < }
(
, c
) such that the maximum is attained .

107
This problem is also called the continuous portfolio problem. In
the following two sections we will discuss two methods to solve this
problem. The rst one is called the martingale method the second
one the stochastic control approach. The rst method only works
in the the case the market is complete ( which is the case here ), the
second one is from the mathematical point of view more sophisticated
but has the advantage that it also works in the non-complete case. This
method can easily be adapted to the case where the underlying nan-
cial market is not a generalized Black-Scholes market.
5.2 The Martingale Method
The main idea behind this method is to decompose the continuous port-
folio problem into two subproblems, where the rst one is a static op-
timization problem, and the second one a representation problem. The
existence of a solution of the second problem is guaranteed by the com-
pleteness oft he generalized Black-Scholes model which itself is implied
by the martingale representation theorem. From this the method got
its name. Let us rst assume that c 0 and U
1
0.
Problem 1 :G(x) := {g 0 F
T
mb |E
P
(e
_
T
0
r
t
dt
g) = x,
E
P
((U
2
(g)
) < }
compute G
:= max
gG(x)
E(U
2
(g)).
Problem 2 : compute (, 0) A
(x)s.t.V
T
() = G
.
Before we discuss the solution of Problem 1 let us take a look at the
well known Lagrange multiplier method, which helps us, to com-
pute extremal points of a function under some subsidiary condition.
Assume f : R
n
R is strictly concave and g : R
n
R
k
is convex, then
108
x
= max{f(x)|x R
n
, g(x) = 0}

= (
1
, ...,
k
) R
k
s.t.

x
i
f(x
)
k
j=1
x
i
g
j
(x) = 0 i = 1, ..., n
g = (g
1
, ..., g
k
)
and g(x
) = 0
(x
) is a zero of L where
L(x, ) = f(x)
g(x) is the Lagrange-function

For the application of the Lagrange multiplier method in Problem
1, let us dene
H
t
:= e
_
t
0
r
s
ds
E
P
(
dP
dP
|F
t
)
and the Lagrange-function L as
L(g, ) := E(U
2
(g) (H
T
g x)), > 0.
Setting the partial derivatives equal to zero we get :
g
L(g, ) = E(U
2
(g) H
T
) = 0
L(g, ) = E((H
T
g x)) = x E(H
T
g)
= x E
P
(e
_
T
0
r
t
dt
g) = 0.
The second equation says exactly, that and hedging strategy ofg
satises V
0
() = x. Since U
2
: (0, ) (0, ) is bijective ( strictly
decreasing and surjective ), we can dene
109
g := (U
2
)
1
(H
T
).
Obviously g solves the rst equation from above. Substituting this
in the second equation we get
x E(H
T
(U
2
)
1
(H
T
))
. .
=:()
= 0. (5.1)
The function : (0, ) (0, ) has the following properties :
Lemma 5.2.1. is continuous and strictly decreasing. Furthermore
(0) := lim
0
() =
(0) := lim
() = 0.
Proof. That is continuous follows from the continuity of (U
2
)
1
and
the theorem about dominated convergence. Since H
T
is strictly positive
and (U
2
)
1
is strictly decreasing, () is also strictly decreasing. Since
(U
2
)
1
shares the behavior for 0 respectively with (U
2
) so
does ().
Hence () is invertible and we can dene
Y (u) :=
1
(u). (5.2)
Using this notation we see that (5.1) is equivalent to
() = x =
1
(x) = Y (x)
as well as
g = (U
2
)
1
(Y (x)H
T
).
Then the pair ( g, Y (x)) is a zero of L and g is a solution of problem
1. Is it ? Even though the argument above seems to be convincing,
110
g actually stands for a random variable and

g
is not dened. Also
it is not clear at all,if the Lagrange multiplier methods works in this
context. The computations above are just formal computations which
lead us to the right solution. We will later see and proof, that g is indeed
the solution of problem 1. From now on we also consider the case c = 0.
We dene
I
2
(y) := (U
2
)
1
(y)
I
1
(y) := (U
1
)
1
(t, y)
(y) := E(
_
T
0
H
t
I
1
(t, yH
t
)dt +H
T
I
2
(yH
T
)).
A slight modication of the proof of lemma 1 shows that above has
the same properties as in lemma 1. We will later need the following
lemma.
Lemma 5.2.2. Let U be a utility function and I := (U
)
1
. Then
U(I(y)) U(x) +y(I(y) x), 0 < y < x < .
Proof. Since U is concave we have
U(I(y)) U(x) +U
(I(y))(I(y) x)) = U(x) +y(I(y) x).

Let us now present the main theorem of this section.
Theorem 5.2.1. Let x > 0 and (y) < for all y > 0. Then
g
:= I
2
(Y (x) H
T
) is the optimal wealth
c
t
:= I
1
(t, Y (x) H
t
) is the optimal consumption
111
and there exists a trading-strategy
s.t. (
, c
) A
(x) and V
T
(
) =
g
.
Proof. For simplicity we assume U
1
0 and hence c
0. By denition
we have
E(H
T
g
) = (Y (x)) = x.
It follows from the completeness of the generalized Black-Scholes
model that there exists
s.t. V
T
(
) = g
and V
0
() = x. Further-
more it follows from Lemma 5.2.2 that
U
2
(g
) U
2
(1) +Y (x) H
T
(g
1)
and since a b a
|b| we have
E(U
2
(g
) E(|U
2
(1)| + Y (x) H
T
(g
+ 1))
= |U
2
(1)| +Y (x)(E(H
T
g
)
. .
=x
+E(H
T
1))
= |U
2
(1)| +Y (x)(x +E
P
(e
_
T
0
r
s
ds
)
. .
<
) <
which shows that (
, 0) A
(x). That the expectation under the

risk neutral measure P
is nite follows from the uniform boundedness

of the interest-rate process r
t
, which was an assumption in the gener-
alized Black-Scholes model. Let us now show the optimality of (
, 0).
We assume that (, 0) A
(x) is arbitrary. Then from Lemma 5.2.2 we

deduce that
U
2
(g
) U
2
(V
T
()) + Y (x)H
T
(g
V
T
()).
Building expectations on both sides we get
112
E(U
2
(g
)) j(x, , 0) +Y (x) E(H

T
(g
V
T
())
= J(x, , 0) +Y (x) (x E
P
(
V
T
())
. .
=x
)
= J(x, , 0)
where we have used that under the risk neutral measure P
the
discounted value process

V
T
()) is a martingale. Hence J(x,
, 0)
J(x, , 0) and the theorem is proven.
For computations it is sometimes advantageous to consider the so
called portfolio process.
Denition 5.2.1. Let (, c) be a self-nancing pair consisting of a trading-
strategy and a consumption process c s.t. V
t
() > 0 a.s. for all t [0, T].
Dene
t
:= (
1
t
, ...,
n
t
)
t [0, T]
i
t
:=

i
t
X
i
t
V
t
()
.
The portfolio-process describes howmuch relative to the total wealth
the trader invests into each asset. Clearly we have
1
t
1 = 1
n
i=1
i
t
X
i
t
V
t
()
=
V
t
()
n
i=1
i
t
X
i
t
V
t
()
=

0
t
X
0
t
V
t
()
which shows that
t
also carries information about the 0th asset.
Assume now that (, c) is a self-nancing pair. Then
113
dV
t
() =
n
i=1
i
t
dX
i
t
c
t
dt
=
0
t
X
0
t
r
t
dt +
n
i=1
i
t
X
i
t
(B
i
t
dt +
n
j=1
ij,t
dW
j
t
) c
t
dt
= (1
t
1)V
t
()r
t
dt +
n
i=1
V
t
() : t
i
(b
i
t
dt +
n
j=1
ij,t
dW
j
t
) c
t
dt
= (1
t
1)V
t
()r
t
dt +V
t
()
t
b
t
dt + V
t
()
T
t

t
dW
t
c
t
dt.
Reordering terms, we get the so called Wealth equation ( short
notation WE )
dV
t
= [r
t
V
t
c
t
]dt +V
t
t
((b
t
r
t
)dt +
t
dW
t
)
V
0
= x
It is clear how to compute the portfolio process, given the trading-
and consumption strategy. It is not so so clear how one gets the trading
strategy back given the portfolio process. The following proposition
says that the two concepts trading strategies and portfolio processes
are mainly equivalent.
Proposition 5.2.1. Given any predictable process s.t.
_
T
0
||
t
||
2
dt <
P a.s. and consumption process (c
t
) then WE has a solution V which
itself determines a trading strategy s.t. V
t
() = V and (, c) is self-
nancing. On the other side, given any self-nancing pair, the wealth
process V = V () is a solution of WE for the associated portfolio process
and x = V
0
().
Proof. The second part is clear, the rst part also, if we know that (WE)
has a solution at all. That this is indeed the case under the assump-
tions made, follows from some theorem about the existence of solutions
of linear stochastic differential equations ( [?], page 62 ).
114
Denition 5.2.2. Given a process and a consumption process c as in
proposition 5.2.1, the solution V of (WE) is denoted by V
t
(, c) and is
called the wealth process corresponding to the portfolio-process and
the consumption process c.
To the pair (, c) corresponds according to Proposition 5.2.1 a trad-
ing strategy . Clearly we have V
t
() = V
t
(, c). Note here that given
the trading strategy, it is not necessary for the computation of the
wealth process also to know the consumption process. In the forth
coming we will switch back and forth between either of the two point
of views. The continuous portfolio problem is equivalent to
max
(,c)A
(x)
J(x, , c)
by V
T
() V
T
(, c) and hence can be formulated completely without
using trading strategies. Let us now consider the following example :
As utility functions we take U
1
(t, x) 0 ( hence c 0 ) and U
2
(x) =
lm(x). We have
U
2
(x) =
1
x
I
2
(x) = (U
2
)
1
(x) =
1
x
(y) = E(H
T
1
y H
T
) =
1
y
Y (x) =
1
x
.
Hence by Theorem 5.2.1 the optimal terminal wealth is
g
= I
2
(Y (x)H
T
) =
x
H
T
V
T
(
)H
T
= x.
It follows from the integration by parts formula ( Proposition 3.3.3 )
that
115
dH
t
= d(e
_
t
0
r
s
ds
E(
_
dW)
t
)
= H
t
t
dW
t
H
t
r
t
dt
where
t
=
1
t
(b
t
R
t
) denotes the market price of risk. Using the
wealth equation and Proposition 3.3.6 we can compute
d[V (
), H]
t
= H
t
V
t
(
t

t
t
dt
= H
t
V
t
(
t
(b
t
r
t
)dt.
Another application of the Integration by parts formula leads to the
following computation :
d(H
t
V
t
()) = H
t
dV
t
() + V
t
()dH
t
+ d[V
t
(, H]
t
= H
t
(r
t
V
t
(
dt + V
t
t
((b
t
r
t
)dt +
t
dW
t
)
V
t
(
)(H
t
t
t
opdW
t
H
t
r
t
dt)
= d[V (
), H]
t
= H
t
V
t
(
)(
t

t
)dW
t
.
Using the assumption in section 4.1 on , r and one can easily
show, that the last expression is not only a local martingale but a mar-
tingale under P.Therefore we have that
H
t
V
t
(
) = E(H
T
V
T
(
)|F
t
)
= E(x|F
t
)
= x.
In particular the process (H
t
V
t
(
) is deterministic which implies

that
116
t

t
= 0.
The last equation is equivalent to
t
= (
t
)
1
t
which has to be valid for all t [0, t]. This strategy is the portfolio
process of the optimal trading strategy. For the standard Black-Scholes
model where b, r and are constant one gets
t
=
b r
2
which is constant in time. Note that since the prices of the assets
change over time, this doesnt mean that the optimal trading strategy
is also constant. In fact it is highly non constant and the trader has to
rearrange his portfolio at any time. Unfortunately there is no general
method to nd the optimal portfolio. Under some assumptions there
are methods from Malliavin Calculus ( Clark-Ocone formula ) which
may be used to compute the optimal portfolio process. These methods
wont be considered in this course. We will discuss a method, which is
more elementary, but needs more assumptions.
Proposition 5.2.2. Let x > 0, (y) < for all y > 0 and g
, c
be
the optimal terminal wealth resp. consumption as in Theorem 5.2.1.
Assume there exists f C
1,2
([0, T] R
d
) s.t. f(0) = x and
1
H
t
E(
_
T
t
H
s
c
s
ds + H
T
g
|F
t
) = f(t, W
1
t
, ..., W
d
t
)
then the optimal portfolio process has the following form
t
=
1
V
t
(
, c
)
(
)
1
x
f(t, W
1
t
, ..., W
n
t
).
Proof. Applying the Ito-formula gives :
117
1
H
t
E(
_
T
t
H
s
c
s
ds + H
T
g
|F
t
) = f(0) +
_
t
0
f
s
(s, W
1
s
, ..., W
d
s
)ds
= +
n
i=1
2
f
x
2
i
(s, W
1
s
, ..., W
d
s
)ds
= +
n
i=1
f
x
i
(s, W
1
s
, ..., W
d
s
)dW
s
.
On the other side it follows from the self-nancing condition that (
using the processes H
t
instead of switching to the equivalent martin-
gale measure, compare Proposition 2.6.1 ) that
1
H
t
E(
_
T
t
H
s
c
s
ds + H
T
g
|F
t
) = V
t
()
=
..
(WE)
x +
_
t
0
((r
s
+
s
(b
s
r
s
))V
s
(
, c
) c
s
ds)
+
_
t
0
V
s
(
, c
s

s
dW
s
.
It then follows from the uniqueness of the martingale part ( Theo-
rem 3.6.1 ) that
x
f(t, W
1
t
, ..., W
n
t
)
= V
s
(
, c
s

s
which implies that
s
=
1
V
s
(
, c
)
(
s
)
1
x
f(t, W
1
t
, ..., W
n
t
).
118
5.3 The stochastic Control Approach
We consider a complete probability space (, F, P) together with a l-
tration (F
t
)
t[t
0
,t
1
]
as well as a stochastic differential equation
dX
t
= (t, X
t
, u
t
)dt + (t, X
t
, u
t
)dW
t
. (5.3)
where W
t
is an m-dimensional Brownian motion and u
t
a d-dimensional
predictable stochastic process ( called the control ). One should think
of, that one can choose the process u on which the solution X of (5.3)
depends and hereby control X. The main problem of stochastic con-
trol is the following :
Given so called cost functions L and nd u
such that
J(t
0
, x, u
) = min
uA(t
0
,x)
J(t
0
, x, u) where
J(t
0
, x, u) = E(
_

t
0
L(s, X
t
0
,x
s
, u
s
)ds + (, X
t
0
,x
))
A(t
0
, x) = {u predictable stochastic process with values in
U R
d
|x(5.3) has a unique solution X
t
0
,x
t
s.t. X
t
0
,x
t
0
= x}
where denotes a stopping time. We will later make more precise
assumptions. Before we discuss how to solve this problem which is
also called the stochastic optimal control problem we study how
it is related to the portfolio optimization problem. Let us consider the
Wealth-equation (WE) as a controlled SDE where the control is given
by u = (u
1
, u
2
) = (, c) and
dV
u
t
= (t, V
u
t
, u
t
)dt + (t, V
u
t
, u
t
)dW
t
(5.4)
where
119
(t, x, u) := (r + u
1
(b r))x u
2
(t, x, u) := xu
1

and the upper index u in V
u
t
indicates the dependence of the solution
on the control u. The portfolio optimization problem corresponds to the
stochastic optimal control problem via
J(t, x, u) := E(
_
T
t
U
1
(t, u
2
t
)dt +U
2
(V
T
(, c))
where U
1
and U
2
are utility functions and minimum is replaced by
maximum. Hence methods for solving the stochastic optimal control
problem can be used to solve the portfolio optimization problem. The
theory on stochastic optimal control is highly developed, though some-
times a bit technical in its application. To make the mathematics pre-
cise we need to make some assumptions and denitions :
Q
0
:= [t
0
, t
1
) R
n
Q
0
:= [t
0
, t
1
] R
n
U R
d
closed
: Q
0
U R
n
: Q
0
U R
nm
(, , u), (, , u) C
1
(Q
0
) u
||
t
|| +||
x
|| C
||
t
|| +||
x
|| C
||(t, x, u)|| +||(t, x, u)|| C(1 +||x|| +||u||)
Furthermore we need the following denitions :
120
O :=
_
R
n
or
n-dim manifold with C
3
boundary, embedded in R
n
Q := [t
0
, t
1
) O
Q := [t
0
, t
1
] O
:= inf{t t
0
|(t, X
t
) Q}
|L(t, x, u)| C(1 +||x||
k
+||u||
k
) and
|(t, x)| C(1 +||x||
k
) on QU for a k N.
The value function corresponding to the stochastic optimal con-
trol problem is dened as follows :
V (t, x) := inf
uA(t,x)
J(t, x, u) (t, x) Q.
We denote with a the matrix valued function a :=
and dene for

each u U an operator A
u
dened on C
1,2
(Q) as follows :
(A
u
G)(t, x) := G
t
(t, x) +
1
2
i,j=1
a
i,j
(t, x, u)

2
G
x
i
x
j
+
n
i=1
i
(t, x, u)
G
x
i
(t, x).
Then the following theorem holds :
Theorem 5.3.1. ( Hamilton-Jacobi-Bellmann ) : Let G C
1,2
(Q)
C(

Q) s.t. |G(t, x)| K(1 + ||x||
k
) for a constant K > 0 and k N and
assume G is a solution of
() : inf
uU
(A
u
G(t, x) +L(t, x, u)) = 0 (t, x) Q
G(t, x) = (t, x)(t, x)
Q
where
Q = ([t
0
, t
1
) O) ({t
1
}

O). Then
121
1. G(t, x) J(t, x, (u
t
))(t, x) Q and u A(t, x)
2. If (t, x) Q there exists (u
t
) A(t, x) s.t.
u
s
argmin
uU
(A
u
G(s, X
u
s
) + L(s, X
u
s
, u))s [t, ]
then
G(t, x) = V (t, x) = J(t, x, (u
t
))
Proof. see [?] page 267.
The following algorithm shows how to apply this theorem to solve
the stochastic optimal control problem.
1. solve the minimization problem () in dependence of the unknown
function G and its derivatives
2. let u
s
:= u
(s, x, G(s, x),

s
G(s, x),

x
G(s, x),

2
x
2
G(s, x))
be the outcome of step 1.), then solve
(A
u
t
G(t, x) + L(t, x, u
t
)) = 0 (t, x) Q
G(t, x) = (t, x)
3. check if all assumptions in Theorem 5.3.1 are satised
Let us use this algorithm to solve the following portfolio optimiza-
tion problem in the Black-Scholes market model ( for simplicity we as-
sume constant deterministic coefcients ) with utility functions
U
1
(t, c) :=
1
e
t
c
U
2
(x) =
1
where > 0 and (0, 1) are constants. The HJB-equation for

V (t, x) := sup
uA(t,x)
J(t, x, u) has the form
122
0 = max
u
1
[
1
,
2
]
d
,u
2
[0,)
{
1
2
u
1
u
1
x
2

2
x
2
V (t, x)
+ ((r + u
1
(b r))x u
2
)

x
V (t, x) + U
1
(t, u
2
) +

t
V (t, x)}
V (T, x) = U
2
(x)
for given 0 <
1
<
2
< . In the rst step we compute the mini-
mum in () with L = U
1
, = U
2
( for the transition of minima to
maxima ). For this we build the partial derivatives with respect to u
1
and u
2
and setting them equal to zero gives :
u
1
t
= (
)
1
(b r)
x
V (t, x)
x

2
x
2
V (t, x)
(5.5)
u
2
t
= (e
t

x
V (t, x)
1
1
. (5.6)
In the second step we have to solve the partial differential equation
for V (t, x) :
0 =
1
2
(b r)
)
1
(b r)
(

x
V (t, x))
2
2
x
2
V (t, x)
+r x

x
V (t, x)
+

x
V (t, x)
1
e
t
1
+

t
V (t, x).
To solve the last equation we make the following separation ap-
proach :
V (t, x) = f(t)
1
f(T) = 1 (terminal condition) .

Substituting this in the previous equation we get
123
0 = [
1
2
(b r)
)
1
(b r) 1 1 +r] f(t)
+
1
e
t
1
(f(t))
1
+ f
(t).
Let us dene
a
1
:=
1
2
(b r)
)
1
(b r)
1
1
+ r
a
2
(t) :=
1
e
t
1
.
Using this, we get the following ordinary differential equation for f
:
f
(t) = a
1
f(t) a
2
f(t)
1
f(T) = 1.
The ODE above is still nonlinear. With the substitution g(t) =
(f(t))
1
1
however we get the following linear ODE
g
(t) =
a
1
1
g(t)
a
2
(t)
1
g(T) = 1.
This ODE can be easily solve. The solution is
g(t) = e
a
1
1
(Tt)
+
1
(a
1
)
(e
a
1
1
T
e
a
1
1
t
) e
a
1
1
(Tt)
.
We get the function f as f(t) = (g(t))
1
. Substituting this in (5.5)
respectively in (5.6) we get for the optimal portfolio respectively the
optimal consumption process
124
t
=
1
1
(
)
1
(b r)
c
t
= (e
t
f(t))
1
1
V
t
(
, c
).
The term V
t
(
, c
) is the value process corresponding to the pair

(
, c
). The denition is not recursive as it looks like, since substitut-

ing (
, c
) as above in (WE) determines a wealth process uniquely. The

third step is now to check, that all the assumption for a application of
the HJB theorem are satised. This is a very good exercise.
125
Bibliography
[Bingham/Kiesel] Risk neutral Valuation
[Delbaen/Schachermayer] A general version of the fundamental theo-
rem of asset pricing.
[Karatzas/Shreve] Brownian Motion and Stochastic Calculus
[Hackenbroch/Thalmaier] Stochastische Analysis
[Hull] Options,Futures and other Derivatives
[Korn1] Optionsbewertung und Portfoliooptimierung
[Korn2] Optimal Portfolios
[Korn3] Option Pricing and Portfolio Optimization
[Protter] Stochastic Integration and Stochastic Differential Equations
[Stricker] Arbitrage et loi de martingales
126

Mathematical Finance

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Mathematical Finance

Uploaded by

Copyright:

Available Formats

Electronic copy of this paper is available at: http://ssrn.

have the same null-sets and

-martingales. Let us dene new

- martingales ( use the

-martingale for all j. By the

on (, F) as follows : For any A F let

. is called the Market Price of

is an equivalent martingale measure for the nancial

P(M) such that the discounted value process

< for a > 0. Then |P

-martingale and also g

-almost sure. Also W

. Since obviously the process

. Let us now consider X

Brownian motion the last expression follows

-martingale ( see again Exercise 1.3.1 ). Hence we have proven that

is an equivalent martingale measure for the market M

Z uniformly and bounded 0. We dene an

= X uniformly and bounded as 0. Now let

dX and from Corollary 3.2.1 it

, it follows from the Girsanov theorem

| is a state price process }.

P)(A) > 0 and |

= im() and there exists a

and by Proposition 3.5.1

. We consider the martingale ( with respect to the Brownian

. It is not hard to show, that the value p determines

P under which Y is a local mar-

. This follows from

we need a version of the Gir-

:= /(exp(b) 1) and dene P

as in Theorem 4.4.1. Then under P

. Hence it follows from Exercise 4.4.1. that

t) it follows that also Y is a martingale under P

and as mentioned before that

}. Let us now consider a European call option

(x) = 0 and vice

(x) = {(, c) self-nancing pair V

) such that the maximum is attained .

g(x) is the Lagrange-function

(I(y))(I(y) x)) = U(x) +y(I(y) x).

(x). That the expectation under the

is nite follows from the uniform boundedness

(x) is arbitrary. Then from Lemma 5.2.2 we

)) j(x, , 0) +Y (x) E(H

) is deterministic which implies

and dene for

(s, x, G(s, x),

where > 0 and (0, 1) are constants. The HJB-equation for

f(T) = 1 (terminal condition) .

) is the value process corresponding to the pair

). The denition is not recursive as it looks like, since substitut-

) as above in (WE) determines a wealth process uniquely. The

You might also like