Continuous Time Finance II: Lecture Notes: Prof. R Udiger Frey, Ruediger - Frey@wu - Ac.at

Continuous Time Finance II: Lecture Notes
Prof. R
udiger Frey,
ruediger.frey@wu.ac.at
Version from May 8, 2012, Comments welcome
Contents
1 Stochastic Processes in Continuous Time
1.1
1.2
1.3
Stochastic Processes, Stopping Times and Martingales . . . . . . . . . . .
1.1.1
Basic Notions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.1.2
Classes of Processes . . . . . . . . . . . . . . . . . . . . . . . . . .
Stopping Times and Martingales . . . . . . . . . . . . . . . . . . . . . . .
1.2.1
Stopping Times . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.2.2
The optional sampling theorem . . . . . . . . . . . . . . . . . . . .
Brownian Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
1.3.1
Definition and Construction . . . . . . . . . . . . . . . . . . . . . .
1.3.2
Some stochastic properties of Brownian motion: . . . . . . . . . . .
1.3.3
Quadratic Variation . . . . . . . . . . . . . . . . . . . . . . . . . .
2 Pathwise It
o-Calculus
12
2.1
It
os formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
12
2.2
Properties of the It
o-Integral . . . . . . . . . . . . . . . . . . . . . . . . .
15
2.2.1
Quadratic Variation . . . . . . . . . . . . . . . . . . . . . . . . . .
15
2.2.2
Martingale-property of the Ito-integral . . . . . . . . . . . . . . . .
15
2.3
Covariation and d-dimensional Ito-formula
. . . . . . . . . . . . . . . . .
18
2.3.1
Covariation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
18
2.3.2
The d-dimensional Ito-formula . . . . . . . . . . . . . . . . . . . .
19
3 The Black-Scholes Model: a PDE-Approach
21
3.1
Asset Price Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
21
3.2
Pricing and Hedging of Terminal Value Claims . . . . . . . . . . . . . . .
22
3.2.1
Basic Notions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
3.2.2
The pricing-equation for terminal-value-claims . . . . . . . . . . .
24
The Black-Scholes formula . . . . . . . . . . . . . . . . . . . . . . . . . . .
24
3.3.1
24
3.3
The formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
3.3.2
Application of the formula: volatility estimation . . . . . . . . . .
4 Further Tools from Stochastic Calculus

4.1
Stochastic Integration for Continuous Martingales . . . . . . . . . . . . .

M2
and
M2,c
25
27
27
4.1.1
The Spaces
. . . . . . . . . . . . . . . . . . . . . . .
27
4.1.2
Stochastic Integrals for elementary processes . . . . . . . . . . . .
30
4.1.3
Extension to General Integrands . . . . . . . . . . . . . . . . . . .
31
4.1.4
Kunita-Watanabe characterization . . . . . . . . . . . . . . . . . .
32
4.2
It
o Processes and the Feynman-Kac formula . . . . . . . . . . . . . . . .
34
4.3
Change of Measure and Girsanov Theorem for Brownian motion . . . . .
37
4.3.1
Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
37
4.3.2
Density martingales . . . . . . . . . . . . . . . . . . . . . . . . . .
38
4.3.3
The Girsanov Theorem . . . . . . . . . . . . . . . . . . . . . . . .
39
5 Financial Mathematics in Continuous-Time

5.1
42
Basic Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
42
5.1.1
The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
42
5.1.2
Martingale Measures and Arbitrage Opportunities . . . . . . . . .
44
5.1.3
Hedging and risk-neutral pricing of contingent claims . . . . . . . .
45
5.1.4
Change of numeraire . . . . . . . . . . . . . . . . . . . . . . . . . .
46
5.2
The Black-Scholes Model Revisited . . . . . . . . . . . . . . . . . . . . . .
47
5.3
Fundamental Theorems of Asset Pricing in a Generalized Black-Scholes Model 49

5.3.1
A multidimensional Black-Scholes model . . . . . . . . . . . . . . .
49
5.3.2
Existence of an equivalent martingale measure . . . . . . . . . . .
51
5.3.3
Market completeness in the generalized Black-Scholes model . . . .
52
6 Term-Structure Modelling
54
6.1
Bonds and Interest-Rates . . . . . . . . . . . . . . . . . . . . . . . . . . .
54
6.2
Short-Rate Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
56
6.2.1
Martingale Modelling and Term-Structure Equation . . . . . . . .
56
6.2.2
Affine Term Structure . . . . . . . . . . . . . . . . . . . . . . . . .
57
6.2.3
Calibration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
58
HJM-Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
59
6.3.1
63
6.3
6.4
The Musiela parametrization. . . . . . . . . . . . . . . . . . . . . .
Pricing and Hedging of Multi-Currency Derivatives with Interest Rate Risk 63

6.4.1
The Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
64
6.4.2
Options on Lognormal Claims: Theory . . . . . . . . . . . . . . . .
66
ii
6.4.3
Options on Lognormal Claims: Examples . . . . . . . . . . . . . .
A Mathematical Background
69
71
A.1 Conditional Expectation . . . . . . . . . . . . . . . . . . . . . . . . . . . .
71
A.1.1 The elementary case . . . . . . . . . . . . . . . . . . . . . . . . . .
71
A.1.2 Conditional Expectation - General Case . . . . . . . . . . . . . . .
72
iii
Chapter 1
Stochastic Processes in
Continuous Time
1.1
Stochastic Processes, Stopping Times and Martingales
1.1.1
Basic Notions
We work on a probability space (, F, P ) with filtration {Ft }. Recall that a filtration is

a family of -fields {Ft , t 0} such that Ft Fs for s > t. As usual Ft is interpreted
as the set of events which are observable at time t such that the filtration represents
information-flow over time.
A stochastic process X = (Xt )t0 is a family of random variables on (, F, P ). We
introduce the following notions:
The process X is called adapted, if for all t > 0 the random variable (rv) Xt is
Ft -measurable.
The marginal distribution of the process at a given t 0 is the distribution (t) of
the rv Xt .
Consider a finite set of time points (t1 , ..., tn ) in [0, ). Then (Xt1 , ..., Xtn ) is a random vector with values in Rn and distribution (t1 , ..., tn ), say. The class of all
such distributions is called the set of finite-dimensional distributions of X. The
finite-dimensional distributions satisfy a set of obvious consistency requirements,
moreover, they determine the stochastic properties of a stochastic process; see for
instance Chapter 2 of Karatzas & Shreve (1988).
Fix some . The mapping
X () : [0, ) R,
t Xt ()
is called trajectory or sample path of X; a stochastic process can be viewed as

random draw of a sample-path. We are only interested in processes whose sample
paths have certain regularity properties. Of particular interest will be processes with
continuous sample paths like Brownian motion or right continuous with left limits
(RCLL) such as the Poisson processes (see below).
1
Equality of stochastic processes.

processes.
There are two nations of equality for stochastic
Definition 1.1. Given two stochastic processes X and Y . Then X is called modification
of Y if for all t 0 we have

P { : Xt () = Yt ()} = 1.
The processes are called indistinguishable if

P { : Xt () = Yt () t > 0} = 1.
We obviously have that if X and Y are indistinguishable then X is a modification of Y .
For the converse implication extra regular assumptions on the trajectories are needed.
Lemma 1.2. Suppose that X and Y have right-continuous trajectories and that X is
a modification of Y . Then X and Y are indistinguishable.
S
Proof. Put Nt := { : Xt ( 6= Yt (} and let N := qQ[0,) Nq . Since Q is
countable and since X is a modification of Y we have 0 = P (Nq ) = P (N ). We want to
show that for \N we have Xt () = Yt () t 0. This is clear for t Q. For t R\Q
there is a sequence qn Q with qn t. By definition of N we have Xqn () = Yqn () for
all \N . Since X and Y have right-continuous trajectories we moreover get
Xt () = lim Xqn ()) = lim Yqn () = Yt (),
n
which proves the claim.
1.1.2
Classes of Processes
1. Martingales: An adapted stochastic process X with E(|Xt |) < for all t > 0 is
a submartingale if t, s with t > s we have E(Xt |Fs ) Xs .
a supermartingale if t, s with t > s we have E(Xt |Fs ) Xs .
a martingale if X is both a sub- and a supermartingale, i.e. if E(Xt |Fs ) = Xs for
all t, s.
Important examples for martingales are the Brownian motion and the compensated Poisson process. Both processes will be introduced below.
2. Semimartingales In financial modelling we often encounter processes which are the
sum of a completely unpredictable part modelled by a martingale and a systematic
predictable component such as the long-term growth rate of an asset. If the systematic
component of such a process satisfies certain regularity properties these processes are
called semimartingales. A formal definition of semimartingales is given in Definition 2.12
below.
3. Markov-Processes: An adapted stochastic process X is called Markov process, if
for all t, s > 0 and all bounded functions f : R R
E(f (Xt+s ) | Ft ) = E(f (Xt+s ) | (Xt )).
2
(1.1)
Here (Xt ) denotes the -field generated by the rv Xt , a notation which we will use
throughout these notes. Intuitively speaking a process is Markovian if the conditional
distribution of future values Xt+s , s 0, of the process is completely determined by the
present value Xt of the process; in particular given the value of Xt , past values Xu , u < t of
the process do not contain any additional information which is useful for predicting Xt+s .
There is an obvious link to the efficient market hypothesis: if X models the price fluctuations of some asset the assertion that X is Markovian is equivalent to saying that the
market for the asset satisfies the weak form of the efficient market hypothesis; see for
instance the survey Malkiel (1987).
Remark 1.3. A Markov process X is called a strong Markov-process, if (1.1) holds for
all stopping-times and not only for deterministic times t.1 All Markov-processes we will
encounter are also strong Markov processes, but there are a few pathological exceptions.
4. Diffusions: A diffusion is a strong Markov process with continuous trajectories such
that for all (t, x) the limits:
1
E(X(t + h) X(t)|X(t) = x) and
h0 h
1
2 (t, x) = lim E((X(t + h) X(t))2 |X(t) = x)
h0 h
(t, x) = lim
(1.2)
(1.3)
exist. Then (t, x) is called the drift, 2 (t, x) the diffusion coefficient. The name diffusion stems from applications in physics; the most important mathematical examples are
solutions to stochastic differential equations.
5. Point processes and the Poisson process: Assume that certain relevant events
for instance claims in an insurance context or defaults of counterparties in a financial
context occur at random points in time 0 < 1 < . . .. The corresponding point process Nt is then given by Nt := sup{n, n t}, i.e. Nt measures the number of events which
have occurred up to time t.
The Poisson process is a special point process. To construct it we take a sequence Yn of
independent exponentially distributed random variables with P (Yn x) = 1 ex and
P
define n := nj=1 Yj , such that Yn is the waiting time between event n1 and n . The
process Nt = sup{n : n t} is then a Poisson process with intensity . It has among
others the following properties
k
P (Nt = k) = et (t)
k! , k = 0, 1, . . . , t 0.
Nt+u Nt is independent of Ns for s t and Poisson-distributed with parameter
(u).
The compensated Poisson process Mt := Nt t is a martingale; in particular
E(Nt ) = t.
1.2
Stopping Times and Martingales
Throughout this section we work on a probability space (, F, P ) with filtration {Ft }.

1
As in discrete time a random variable with values in [0, ] will be called a stopping time if for
all t 0 the set { , () t} belongs to the sigma-field Ft ; see Section 1.2 below.
1.2.1
Stopping Times
Definition 1.4. A rv : [0, ] is called stopping time wrt. {Ft } if t 0 holds that
{ t} Ft .
Remark 1.5. can be interpreted as the time of the occurence of an observed event.
{ = } means that the event never occurs.
T
Lemma 1.6. Let {Ft } be right-continuous, i.e. Ft = >0 Ft+ , for all t 0. Then
: [0, ] is a stopping time if and only if { < t} Ft , t 0.
T
Proof. It holds { t} = >0 { < t + }. Let { < t} Ft , t 0. Hence { < t + }
T
Ft+ and { t}
Ft+ = Ft . For the converse statement note that
>0
{ < t} =
[
Q+
{ t } Ft .
|
{z
}
Ft Ft
The most important example for stopping times are first hitting times for Borel sets.
Definition 1.7. Given a stochastic process X and a Borel set A in R. Define A :=
inf{t 0 : Xt A}. Then the rv A is called first hitting time into the set A.
Next we address the question when A is a stopping time.
Lemma 1.8. Let X be {Ft }-adapted and right-continuous and let A Rd be open. If the
filtration {Ft } is right-continuous, then the hitting time A is a stopping time.
Proof. Suppose, that {Ft } is right-continuous. We only have to show {A < t} Ft t > 0.
Since X is right-continuous and A is open, it holds that
[
{A < t} =
{Xq A} .
q[0,t)Q
As {Xq A} Fq Ft and as the union of countable sets from Ft also belongs to Ft ,

the claim follows.
Lemma 1.9. Let X be continuous and A R closed. Then A is a stopping time.
Proof. Define open sets An A by An := {x : d(x, A) < 1/n}. Since X is continuous and
A is closed, it holds that
{ : A () t} = { : s [0, t], Xs () A} = { : n N s [0, t] with Xs () An }.
As all An are open, the last set equals
{ : n Nq Q [0, t], Xq () An } =
nN qQ[0,t]
{ : Xq () An } .
{z
}
|
Fq Ft
The right-hand side consists of countably many operations on sets from Ft , hence it
belongs to Ft .
4
The sigma-field F .
to a stopping time .
Next, we define the -field generated by all observable events up
Definition 1.10. Given a stopping time . Then we call

F := A F : A { t} Ft t 0
(1.4)
the -field of the events which are observable up to .

F is indeed a -field as is easily checked; a more intuitive characterization of F will be
given in Lemma 1.16 below. It is easily seen that the rv () is F -measurable: For
all t0 0 we have
{ t0 } { t} = { (t0 t)} Ft0 t Ft .
Let S and T be stopping times, where S T . Intuitively, we can say, that if event A
is observable up to time S, then event A is also observable up to T , so that one would
expect the inclusion FS FT . This is in true.
Lemma 1.11. Given two stopping times S and T with S T . Then FS FT .
Proof. Since S T , it holds {T t} {S t}. Using this we have for A FS :
A {T t} = A {S t} {T t} ,
|
| {z }
{z
}
Ft , as AFS
Ft , as T
stopping time
and hence, A FT .
Lemma 1.12. Given two stopping times S and T . Then
i) S T := min{S, T } is a stopping time.
ii) S T := max{S, T } is a stopping time.
iii) FST = FS FT .
Proof. We have {S T t} = {S t} {T t}, {S T t} = {S t} {T t}. As
S and T stopping times, that means {S t} Ft and {T t} Ft , claims i) and ii) are
proved.
iii) is proved as follows. From Lemma 1.11 we see, that, FST FS and FST FT ,
hence FST FS Ft . Now, let A FS FT . As S and T stopping times we have
A {S t} Ft and A {T t} Ft . It holds
(A {S t}) (A {T t}) = A ({S t} {T t}) = A {S T t} Ft ,
and hence A FST .
Let (Xt )t0 be a right-continuous stochastic process und let T be a stopping time. We
define the stopped rv XT by
XT () := XT () () 1{T <} (),
(1.5)
Lemma 1.13. Let (Xt )t0 be an {Ft }-adapted and right-continuous stochastic process
and let T be a stopping time. Then the rv XT is FT -measurable.
For a proof we refer to Protter (1992) or to Karatzas & Shreve (1988).
Definition 1.14. Given a right-continuous stochastic process (Xt )t0 and a stopping time
T . Then the in T stopped process X T = (XtT )t0 is defined by
(
XT () (), T () t.
XtT () := XtT () () =
(1.6)
Xt (),
T () > t.
Lemma 1.15. If X is adapted and right-continuous, then the stopped process X T is
adapted.
Proof. We have XtT = Xt 1{t<T } + XT 1{tT } . The first summand is Ft -measurable,
since it is a product consisting of two Ft -measurable rvs. For the second summand we
conclude as follows. XT 1{tT } = XT t 1{tT } . The rv XT t is FT t -measurable and
FT t Ft . The rv 1{tT } is Ft -measurable, as T is a stopping time.
Finally we give a more intuitive interpretation of FT , which legitimates the description
-field of the observable events up to time T .
Lemma 1.16. Let T be a stopping time. If P (T < ) = 1, then FT = (X T , X adapted and cadlag).
Proof. Let X be adapted and cadlag. Then the rv XtT = XT t is FT t -measurable. Since
FT t FT the rv XtT is also FT -measurable, hence FT (X T , X adapted and cadlag).
Now, let A FT and define a stochastic process X = (Xt )t0 by Xt () = 1A ()
1{T t} (). The process X is cadlag. It holds {Xt = 1} = A {T t} Ft , hence
S
X is adapted. Moreover we have A = n {Xn () = 1}, since T is finite. Hence, A
(X T , X adapted and cadlag).
1.2.2
The optional sampling theorem
The following results gives a crucial link between stopping times and martingales.
Theorem 1.17 (Optional sampling theorem). Consider an adapted stochastic process
X = (Xt )t0 with E(|Xt |) < , t 0. Then the following statements are equivalent.
(1) X is a martingale.
(2) For all bounded stopping times ( () C for some C > 0, all ) one has
E(X ) = E(X0 ).
(3) Given two stopping times S and T such that S T C for some C > 0. Then
E(XT | FS ) = XS .
We omit the proof; see for instance Protter (1992), Section I.2.
Corollary 1.18. Let X be a martingale with right-continuous trajectories and let be a
stopping time. Then the stopped process X with Xt = Xt is also a martingale.
6
See again Protter (1992), Section I.2 for a proof.

Corollary 1.19 (Martingale inequality). Let X be a right-continuous martingale such
that Xt > 0 a.s. Then we have for C > 0

1
P sup Xt > C E(XT ).
C
t0
Proof. Put TC := inf{t 0 : Xt > C}. Since Xt > 0, we have for an arbitrary n N

1
1
XTC n = E(X0 ),
P sup Xt > C E
C
C
0tn
where the last equality is due to Theorem 1.17, (2). For n we obtain the result by
monotone convergence.
1.3
Brownian Motion
Brownian motion is the most important building block for continuous-time asset pricing
models. It has a long history in the modelling of random events in science.
Around 1830 R. Brown, a Scottish botanist, discovered that molecules of water in a
suspension perform an erratic movement under the buffeting of other water molecules.
While Browns research had no relation to mathematics this observation gave Brownian
motion its name. In 1900 Bachelier introduced Brownian motion as model for stock-prices;
see Bachelier (1900). In 1905 Einstein proposed Brownian motion as a mathematical
model to describe the movement of particles in suspension. The first rigorous theory of
Brownian motion is due to N. Wiener (1923); therefore Brownian motion is often referred
to as Wiener process.
1.3.1
Definition and Construction
Definition 1.20. A stochastic process X = (Xt )t0 on (, F, P ) is standard one-dimensional

Brownian motion, if
(i) X0 = 0 a.s.
(ii) X has independent increments: for all t, u 0 the increment Xt+u Xt is independent of Xs for all s t.
(iii) X has stationary, normally distributed increments: Xt+u Xt N (0, u).
(iv) X has continuous sample paths.
In honor of Brown and Wiener Brownian motion is often denoted by (Bt )t0 or by (Wt )t0 .
Definition 1.21. Standard Brownian motion in Rd is a d-dimensional process Wt =
(Wt1 , ..., Wtd ) where W 1 , ..., W d are d independent standard Brownian motions in R1 .
Definition 1.20 has some elementary consequences:
(i) Wt = Wt W0 is N (0, t)-distributed.
7
(ii) Let t > s. Then the covariance of Wt and Ws is given by cov(Wt , Ws ) = E[Wt Ws ] =
E[(Wt Ws )Ws ] + E[Ws2 ] = E[Wt Ws ]E[Ws ] + s = s.
(iii) The finite-dimensional distributions of W are multivariate normal distributions with
mean 0 and covariance matrix given in (ii).
Theorem 1.22. A stochastic process with the properties of Definition 1.20 exists.
Remarks: 1) There are various methods to construct Brownian motion and hence to
prove Theorem 1.22, which are also useful for simulating Brownian sample paths; see for
instance Karatzas & Shreve (1988).
2) Theorem 1.22 is more than a mere exercise in mathematical rigour: if we replace
the normal distribution in Definition 1.20 (iii) by a fat-tailed -stable distribution, the
corresponding process referred to as -stable motion necessarily has discontinuous
trajectories; see for instance Section 2.4 of Embrechts, Kl
uppelberg & Mikosch (1997) for
details.
1.3.2
Some stochastic properties of Brownian motion:
Proposition 1.23. Let Wt be standard Brownian motion and define Ft := (Ws , s t).
Then a) (Wt )t0 b) (Wt2 t)t0 and c) exp(Wt 1/2 2 t) are martingales with respect
to the filtration {Ft }.
Proof. We start with claim a). Let t > s; by point (ii) of Definition 1.20 the increment
Wt Ws is independent of Fs . Hence we get
E[Wt |Fs ] = E[Wt Ws + Ws |Fs ] = E[Wt Ws ] + Ws = Ws .
To prove claim b) we first show that E[Wt2 Ws2 |Fs ] = E[(Wt Ws )2 |Fs ]. We have
E[(Wt Ws )2 |Fs ] = E[Wt2 2Wt Ws + Ws2 |Fs ] = E[Wt2 |Fs ] 2Ws E[Wt |Fs ] + Ws2
= E[Wt2 |Fs ] 2Ws2 + Ws2 = E[Wt2 Ws2 |Fs ].
2 W 2 |F ] = (t s). By the first step of
The claim is proved if we can show
s
t
s
that E[W
the proof this is equivalent to E (Wt Ws )2 |Fs = (t s). Now Wt Ws is independent
of Fs ; hence E[(Wt Ws )2 |Fs ] = E[(Wt Ws )2 ] = t s, as Wt Ws N (0, t s).
1
Sketch of c) (Bitte ausarbeiten !!!) Gt = e(Wt Ws 2 (ts)) , then E[Gt ] = Gs E[e(Wt Ws 2 (ts)) ]
1
1
and E[Gt |Fs ] = Gs E[e(Wt Ws 2 (ts)) |Fs ] = Gs E[e(Wt Ws 2 (ts)) ] = Gs (properties of t
the lognormal distribution).
1.3.3
Quadratic Variation
Fix some point in time T , which represents the time-point where our model ends. To
define first and quadratic variation we need the notion of a partition of the interval [0, T ].
Definition 1.24. A partition of [0, T ] is a set of time-points t0 = 0 < t1 < . . . < tn = T .
The mesh of this partition is given by | | := sup1in |ti ti1 |.
Definition 1.25 (First Variation). Consider a function X : [0, T ] R. The first variation
of X on [0, T ] is defined as
)
(
X
(1.7)
Var(X) := sup
|X(ti ) X(ti1 )| , a partition of [0, T ] [0, ] .
ti
If Var(X) < X is said to be of finite variation.

Remarks on notation: 1) Following standard conventions we denote by Var(f ) the first
variation of a function f , whereas var(Y ) stands for the variance of a random variable Y .
2) Whenever a summation formula such as (1.7) contains the index t1 it is understood
that the corresponding summand is equal to zero.
Definition 1.26 (Quadratic Variation). Consider again a function X : [0, T ] R and a
sequence n of partitions of [0, T ] such that |n | 0 as n . Define for t [0, T ] the
quadratic variation of X along the partition n by
X
(X(ti ) X(ti1 ))2 .
Vt2 (X; n ) :=
ti n ; ti <t
Assume that for all t [0, T ] the limit hXit := limn Vt2 (X; n ) exists and that the
function t hXit is continuous. In that case X is said to admit the continuous quadratic
variation hXit .
In principle hXit might depend on the sequence n . However, we are mainly interested
in the case where X is a sample path of a continuous semimartingale such as Brownian
motion. It can be shown that in this case hXit is independent of the sequence of partitions
used in its definition. Obviously hXit is increasing in t and hence in particular of finite
variation.
We now discuss the relation between first and quadratic variation.
Proposition 1.27. If X : [0, T ] R is continuous and of finite variation its quadratic
variation hXit is zero.
By negating this result we have
Corollary 1.28. If X is continuous and if the function t hXit is strictly increasing,
X is of infinite first variation on every subinterval [a, b] of [0, T ].
Proof. (of Proposition 1.27) Choose a sequence of partitions n of [0, T ] such that limn |n | =
0. Then
X
X
(X(ti ) X(ti1 ))2 sup |X(ti ) X(ti1 )|
|X(ti ) X(ti1 )|
ti n
ti n ; ti t
ti n
sup |X(ti ) X(ti1 )| Var(X).
(1.8)
ti n
Now note that Var(X) < and that supti n |X(ti ) X(ti1 )| 0 for n as X is
continuous and as limn |n | = 0. Hence the right side of (1.8) converges to zero which
proves the proposition.
The following result allows us to conclude that the quadratic variation of the sample paths
of a continuous semimartingale is determined by the quadratic variation of its martingale
part.
Proposition 1.29. Assume that X is continuous with continuous quadratic variation
hXit and consider a continuous function A : [0, T ] R which is of finite first variation.
Then Y (t) = X(t) + A(t) is of continuous quadratic variation and we have hY it = hXit .
Proof. We have
X
(Yti Yti1 )2 =
(Xti Xti1 )2 +
+ 2
(Ati Ati1 )2
ti n ; ti t
ti n ; ti t
ti n ; ti t
(Xti Xti1 )(Ati Ati1 )
ti n ; ti t
P
P
Now ti n ; ti t (Xti Xti1 )2 converges to hXit by assumption and ti n (Ati Ati1 )2
converges to zero as A is continuous and of finite variation. The last term can be estimated
as follows:
X
(Xti Xti1 )(Ati Ati1 ) sup |Xti Xti1 |Var(A),
ti1 n
ti n ; ti t
which converges to zero as X is continuous.

Now we deal with quadratic variation of the sample paths B () of Brownian motion.
Roughly speaking, for (almost) all we have hB ()it = t. The following theorem
makes this relation precise.
Theorem 1.30. Consider a sequence of partitions n of [0, T ] such that limn |n | = 0.
2
Then we have for all t [0, T ] that E Vt2 (B (); n ) t 0 as n .
Proof. For a fixed partition n we have
!2
!2
E
(Bti Bti1 ) t
= E
ti n ,ti <t
(Bti Bti1 ) (ti ti1 )
ti n ,ti <t
E (Bti Bti1 )2 (ti ti1 )
2
ti n ,ti <t
as (Bti Bti1 ) and (Btj Btj1 ) are independent for i 6= j. Now note that
2

E (Bti Bti1 )2 (ti ti1 ) = var (Bti Bti1 )2 .
It is well known that for a N (, 2 )-distributed rv we have var( 2 ) = 2 4 . Hence
var(Bti Bti1 )2 = 2(ti ti1 )2 , and we get
!2
E
(Bti Bti1 )2 t
=2
ti n ,ti <t
(ti ti1 )2 |n |t 0 ,
ti n ,ti <t
which proves the theorem.
10
The type of convergence in Theorem 1.30 is known as L2 -convergence. It implies in

particular that Vt2 (B (); n ) converges to t in probability as n . By exploiting
the fact that every sequence of random variables which converges in probability has a
subsequence which converges almost surely we obtain the following
Corollary 1.31. There exists a sequence n of partitions of [0, T ] with limn |n | = 0
such that almost surely Vt2 (B (); n ) t for every t [0, T ].
This corollary is important as it shows that the pathwise Ito-calculus developed in Section
2 applies to sample paths of Brownian motion.
Combining Theorem 1.30 and Corollary 1.28 yields another surprising property of Brownian sample paths.
Corollary 1.32. Sample paths of Brownian motion are of infinite first variation.
Remark 1.33. The sample paths of Brownian motion have many surprising properties.
We refer the reader to Karatzas & Shreve (1988) and in particular to Revuz & Yor (1994)
for further information.
We have seen that Brownian motion is a martingale with continuous trajectories and
quadratic variation hB ()it = t. The following theorem, which is usually referred to as
Levys characterization of Brownian motion, establishes the converse:
Theorem 1.34. If M is a martingale with continuous trajectories such that M0 = 0 and
hM it = t t then M is Brownian motion.
11
Chapter 2
Pathwise It
o-Calculus
Motivation. Consider a function f : R R which is once continuously differentiable
(abbreviated f is a C 1 -function) with derivative f 0 and a C 1 -function X : R+ R with
derivative X := t
X(t). The fundamental theorem of calculus yields
Z t
Z t
0
f (X(s))X(s)ds =:
f 0 (Xs )dXs .
f (X(t)) f (X(0)) =
(2.1)
0
A similar expression for the difference f (Xt ) f (X0 ) can be given if X is not C 1 but only
continuous and of finite variation:
Proposition 2.1. Consider a continuous function X : [0, T ] R which is of finite
variation and a C 1 -function f : R R with derivative f 0 . Let n denote a sequence of
partitions of [0, T ] with limn |n | = 0. Then we have that
Z t
X
0
lim
f (Xti1 )(Xti Xti1 ) =:
f 0 (Xs )dXs
(2.2)
n
ti n ,ti t
exists. Moreover, we have the change of variable rule

Z t
f (Xt ) = f (X0 ) +
f 0 (Xs )dXs .
(2.3)
Proposition 2.1 is a special case of Itos formula (Theorem 2.2 below), hence we omit the
proof.
2.1
It
os formula
In this section we derive the It

o-formula in financial texts often referred to as Itos
Lemma which extends the chain-rule (2.3) to functions with infinite first but finite
quadratic variation. Our exposition is based on the so-called pathwise Ito-calculus developed by F
ollmer (1981); this approach allows us to give an elementary and relatively simple
derivation of most results from stochastic calculus which are needed for the Black-Scholes
option pricing model without having to develop the full theory of stochastic integration.
Throughout this section we consider a continuous function X : [0, T ] R which admits
a continuous quadratic variation hXit in the sense of Definition 1.26. As shown in Corollary 1.31 this is true for paths of Brownian motion. More generally, it can be shown
12
that the sample paths of every continuous semimartingale admit a continuous quadratic
variation.
Rt
As hXit is increasing in t, the integral 0 g(s)dhXis is defined for every continuous function
g : [0, T ] R in the ordinary Riemann-sense; as hXit is continuous this integral is
moreover a continuous function of the upper bound t. Now we can state
Theorem 2.2 (It
os formula). Given a continuous function X : [0, T ] R with continuous quadratic variation hXit . Let F : R R denote a twice continuously differentiable
function. Then we have for t T
Z t
Z
1 t 00
0
F (Xt ) = F (X0 ) +
F (Xs )dXs +
F (Xs )dhXis
(2.4)
2 0
0
where
F 0 (Xs )dXs := lim
F 0 (Xti1 )(Xti Xti1 ).
(2.5)
ti n ; ti t
Remarks: 1)R The existence of the limit in (2.5) is shown in the proof of the theorem.
t
The integral 0 F 0 (Xs )dXs is called Ito-integral; it is a continuous function of the upper
boundary t as is immediately apparent from (2.4).
2) The classical case of Proposition 2.1, where X is of finite variation
R t is a special case of
Theorem 2.2. If hXit is non-zero the additional correction-term 12 0 F 00 (Xs )dhXis enters
our formula for the differential F (Xt ) F (X0 ). We will see that this term is of crucial
importance for most results in continuous-time finance.
3) Note that the sums used in defining the Ito-integral are non-anticipating, i.e. the
integrand F 0 (Xs ) is evaluated at the left boundary of the interval [ti1 , ti ]; we will see in
Section 3.2 below that this makes the Ito-integral the right tool for the modeling of gains
from trade.
4) Often formula (2.4) is expressed in the following short-hand notation: dF (Xt ) =
F 0 (Xt )dXt + 12 F 00 (Xt )dhXit .
5) It is possible to give extensions of this theorem to the case where X has discontinuous
sample paths; see for instance Chapter II.7 of Protter (1992).
Proof. As a first step we establish the following
Lemma 2.3. For every piecewise continuous function g : [0, T ] R we have
X
lim
g(ti1 )(Xti Xti1 )2 =
g(s)dhXis .
(2.6)
ti n ; ti t
Proof of the Lemma. Recall the definition of Vt2 (X; n ) in Definition 1.26. For indicator
functions of the form g(t) = 1(a,b] (t) the convergence in (2.6) translates as

lim Vb2 (X; n ) Va2 (X; n ) = hXib hXia ,
n
which is satisfied by definition, as X admits the continuous quadratic variation hXit . For a
general piecewise continuous function g the claim of the Lemma follows if we approximate
g by piecewise constant functions.
13
Now we turn to the theorem itself. Consider ti , ti1 n , such that ti t and denote by
(X)i,n the increment Xti Xti1 . We get from a Taylor-expansion of F
1
F (Xti ) F (Xti1 ) = F 0 (Xti1 )(X)i,n + F 00 (Xt)(X)2i,n
2
1
= F 0 (Xti1 )(X)i,n + F 00 (Xti1 )(X)2i,n + Ri,n ,
2
where t is some point in the interval (ti1 , ti ), and where Ri,n := 21 (F 00 (Xt)F 00 (Xti1 ))(X)2i,n .
Define n := max{|Xt Xti1 | , t [ti1 , ti ], ti n }. As X is continuous and as |n | 0
for n we have n 0, n . Moreover,

1
00
00
|Ri,n |
max |F (x) F (y)| (X)2i,n =: n (X)2i,n .
2 |xy|<n
Now n 0, for n as F 00 is uniformly continuous and as n 0. Hence

X

X
X

|Ri,n | n
(X)2i,n 0 as n ,
Ri,n

ti n
ti n
as X admits the continuous quadratic variation hXit . Now

X
F (Xt ) F (X0 ) = lim
F (Xti ) F (Xti1 )
n
= lim
ti n ; ti t
F 0 (Xti1 )(X)i,n +
ti n ; ti t
1
2
F 00 (Xti1 )(X)2i,n +
ti n ; ti t
Ri,n .
ti n ; ti t
We have just shown that the sum over the RRi,n tends to zero. Moreover, by Lemma 2.3
P
t 00
00
2
ti n ; ti t F (Xti1 )(X)i,n converges to 0 F (Xs )dhXis . Hence the limit
Z
F 0 (Xs )dXs := lim
F 0 (Xti1 )(Xti Xti1 )
ti n ; ti t
exists, and we obtain the It

o-formula (2.4).
Some Examples:
1) Take F (x) = xn . Applying the Ito-formula yields
Xtn
X0n
Z
+n
Xsn1 dXs
n(n 1)
+
2
Xsn2 dhXis .
In short notation this can be written as dXtn = nXtn1 dXt + n(n1)

Xtn2 dhXit . In the
2
special case where X is a sample path of a Brownian motion B with B0 = 0 we obtain
Z t
Z t
Z t
Bt2 = 2
Bs dBs +
dhBis = 2
Bs dBs + t
0
2) Take F (x) = ex . We get eXt = eX0 +

deXt = eXt dXt + 12 eXt dhXit .
Rt
0
eXs dXs +
14
R
1 t
2
eXs dhXis , or in short notation
2.2
Properties of the It
o-Integral
2.2.1
Quadratic Variation
Throughout this section we consider a continuous function X(t) with continuous quadratic
variation hXit .
2.4. Let F C 1 (R); then the function t F (Xt ) has quadratic variation
RProposition
t
2
0
0 (F (Xs )) dhXis .
Rt
Corollary 2.5. For f C 1 (R) the It
o-integral It := 0 f (Xs )dXs is well-defined; its
Rt
quadratic variation equals hIit = 0 f 2 (Xs )dhXis .
Proof. Denote by n a sequence of partitions of [0, T ] with |n | 0. Then
X
X
2
2
F (Xti ) F (Xti1 ) =
F 0 (Xti )(X)i,n , ti (ti1 , ti )
ti n ; ti t
ti n ; ti t
X
ti n ; ti t
F 0 (Xti1 )2 (X)2i,n +
(F 0 (Xti )2 (F 0 (Xti1 ))2 (X)2i,n .

{z
}
|
ti n ; ti t
0, n
Rt
The first sum converges to 0 (F 0 (Xs ))2 dhXis by Lemma 2.3; a similar argument as in
the proof of Theorem 2.2 shows that the second sum converges to zero as n .
Rx
To proof the Corollary we define F (x) = 0 f (y)dy, such that F 0 = f . As F is a C 2 Rt
function the existence of the integral It = 0 F 0 (Xs )dXs follows from the previous theorem.
Moreover, we get from It
os formula that
Z t
Z
1 t 0
F (Xt ) = F (X0 ) +
f (Xs )dXs +
f (Xs )dhXis =: F (X0 ) + It + At .
2 0
0
As the function A is ofR finite variation we get hIit = hF (X)it . By Proposition (2.4), we
t
know that hF (X)it = 0 f 2 (Xs )dhXis , which proves the corollary.
Example: RWe compute the quadratic variation
of the square of Brownian Rmotion B. We
Rt
t
t
2
have Bt = 0 2Bs dBs + t. Define It := 0 2Bs dBs . We get hB 2 it = hIit = 0 4Bs2 ds.
2.2.2
Martingale-property of the It
o-integral
Up to now we have only used analytic properties of the function X such as the fact that X
admits a continuous quadratic variation in our analysis of the Ito-integral. If X(t) is the
sample path of a stochastic process
R tsuch as Brownian motion we may study probabilistic
properties of the process It () = 0 f (Xs ())dXs (). In particular we may consider the
case that our integrator is a martingale.
If M is a martingale with trajectories ofR continuous quadratic variation and f a C 1 funct
tion we expect the It
o-integral It := 0 f (Ms )dMs to inherit the martingale property
from M , as It is defined as limit of non-anticipating sums, It = limn Itn with Itn =
P
n
ti n ; ti t f (Mti1 )(Mti Mti1 ) . The martingale property of the It is just a variation
of the you cant gain by betting on a martingale argument used already in our proof that
the discounted gains from trade of an admissible selffinancing strategy are a martingale
15
under an equivalent martingale measure in Chapter ?? (Lemma ??). Unfortunately some

integrability problems arise when we pass from the approximating sums to the limit such
that only a slightly weaker result is true. To state this result we need the notion of a local
martingale.
Definition 2.6. A stochastic process M is called a local martingale, if there are stopping
times T1 . . . Tn . . . such that
(i) limn Tn () = a.s.
(ii) (MTn t )t0 is a martingale for all n.
Obviously every martingale in the sense of Section 1.1.2 (every true martingale) is a local
martingale. The opposite assertion is not true; see for instance Remark 2.10 below.
Theorem 2.7. Consider a local martingale M with continuous trajectories
and continuous
Rt
quadratic variation hM it and a function f C 1 (R). Then It () = 0 f (Ms ())dMs () is
a local martingale.
Partial proof. We restrict ourselves to the case where M is a bounded martingale and
where f is bounded; the general case follows by localization (introduction of an increasing
sequence of stopping time (Tn )nN ). The proff goes in two steps.
a) Let n be a sequence of partitions with |n | 0, and fix n. Then the discrete-time
P
process Ikn := ti n , ik f (Mti1 )(Mti Mti1 ), k n, is a martingale wrt the discrete
filtration {Fkn }k with Fkn := Ftk , as can be seen from the following easy argument.
n
n
E(Ikn Ik1
|Fk1
) =E(f (Mtk1 )(Mtk Mtk1 )|Ftk1 ) = f (Mtk1 )E((Mtk Mtk1 )|Ftk1 ),
and the last term is obviously equal to zero as M is a martingale. Note that here we have
used the fact that the It
o-integral is non-anticipating.
b) Let s < t. We will show that E(It ; A) = E(Is ; A) for all A Fs , as this implies that
E(It |Fs ) = Is . Choose tn , sn n with tn & t, sn & s and tn > sn . By Step a) we have
E(Itn ; A) = E(Isn ; A);
moreover Itn It , Isn Is , as I has continuous paths. Moreover, one can show that
(Itn )n and (Isn )n are uniformly integrable (using the boundedness of M and f ), so that
the claim follows from the theorem of Lebesgue.
Rt
Remark 2.8. If f is defined only on a subset G R the process It = 0 f (Ms )dMs can
be defined up to the stopping-time = inf{t > 0, Mt
/ G} and it can be shown that It
is a local martingale until .
In applications one often needs to decide if a local martingale M is in fact a true martingale. The following Proposition provides a useful criterion for this
Proposition 2.9. Let M be a local martingale with continuous trajectories. Then the
following two assertions are equivalent.
(i) M is a true martingale and E(Mt2 ) < t 0.
(ii) E(hM it ) < t.
16
If either (i) or (ii) holds, we have E(Mt2 ) = E(hM it ).

For a proof and a generalization to discontinuous martingales see Chapter II.6 of Protter
(1992).
Remark 2.10. The following process is an example of a local martingale which is not
a true martingale. Consider a three-dimensional Brownian motion Wt = (Wt1 , Wt2 , Wt3 )
with W0 = (1, 1, 1) and define
Mt =
1
1
=p
.
||Wt ||
(Wt1 )2 + (Wt2 )2 + (Wt3 )2
Then M is a local martingale, as can be checked using the Ito-formula in higher dimensions,
but it is not a full martingale; see again Chapter II.6 of Protter (1992) for details.
The following Proposition shows that interesting martingales with continuous trajectories
are necessarily of infinite variation.
Proposition 2.11. Consider a local martingale M with continuous trajectories of finite
variation Then the paths of M are constant, i.e. Mt = M0 almost surely.
Note that there are martingales with discontinuous non-constant trajectories of finite
variation; an example is provided by the compensated Poisson process; see Section 1.1.2.
Proof. By It
os-formula we get for Mt2
Mt2
M02
Ms dMs + hM it =
+2
M02
Z
+2
Ms dMs ,
0
as hM it = 0 by Proposition 1.27. The martingale-property of the Ito-integral implies that

Mt2 is a local martingale. Assume that both Mt and Mt2 are a real martingales.1 Then
we have

0 E (Mt M0 )2 = E Mt2 2Mt M0 + M02 = M02 2M02 + M02 = 0 ,
which shows that E (Mt M0 )2 = 0 so that Mt = M0 a.s.
We close this section with a formal definition of semimartingales.
Definition 2.12. A stochastic process X with RCLL paths is called a semimartingale if
X has a decomposition of the form Xt = X0 + Mt + At where M is a local martingale
and A is an adapted process with continuous trajectories of finite variation; M is called
the martingale part of X, A the finite variation part.
Note that the decomposition of X into a martingale part and a finite variation part is
unique by Proposition 2.11.
1
For an argument how to deal with the case where Mt2 is only a local martingale we refer the reader
to Protter (1992).
17
2.3
2.3.1
Covariation and d-dimensional It

o-formula
Covariation
Fix a sequence n of partitions of [0, T ] with n 0 and continuous functions X, Y which

admit a continuous quadratic variation hXit andhY it along the sequence n .
Definition 2.13. Assume that for all t [0, T ] the following limit exists:
X
(Xti Xti1 )(Yti Yti1 ) =: hX, Y it .
lim
n
ti n ; ti t
Then hX, Y it is called covariation of X and Y .

Theorem 2.14. hX, Y it exists if and only if hX + Y it exists; in that case we have the
following so-called polarization-identity
hX, Y it =
1
(hX + Y it hXit hY it ) .
2
(2.7)
Proof. Recall the notation (X)i,n = Xti Xti1 , for ti , ti1 n . We have
X
hX + Y it = lim
((X)i,n + (Y )i,n )2
n
= lim
ti n ; ti t
(X)2i,n +
n
ti n ; ti t
(Y )2i,n + 2
ti n ; ti t
= hXit + hY it + 2 lim
(X)i,n (Y )i,n
ti n ; ti t
(X)i,n (Y )i,n .
(2.8)
ti n ; ti t
Hence the last limit on the right hand side of (2.8) exists iff hX + Y it exists. Solving for
this limit yields the polarization identity.
Note that hX, Y it is of finite variation as it is the difference of monotone functions. We
now use the polarization identity to compute the covariation for a few important examples.
1) If X is a continuous function with continuous quadratic variation hXit and A a continuous function of finite variation we have hX + Ait = hXit and hence hX, Ait = 0.
2) Consider two independent Brownian motions B 1 , B 2 on our probability space (, F, P ).
Then hB 1 (), B 2
()it = 0. To prove this claim we have to compute hB 1 + B 2 it . Note
that (Bt1 + Bt2 )/ 2 is again a Brownian motion and has therefore quadratic variation
equal to t. Hence
1
1
(hB 1 + B 2 it hB 1 it hB 2 it ) = (2t t t) = 0 .
2
2
1
3) Consider a continuous
quadratic variation, and
R t function X with continuous
Rt
R t C -functions
f and g. Define Yt := 0 f (Xs )dXs and Zt := 0 g(Xs )dXs . Then hY, Zit = 0 f (Xs )g(Xs )dhXis .
This follows from the polarization identity and the following computation:
Z t
Z t
hY + Zit =
(f + g)2 (Xs )dhXis = hY it + hZit + 2
f (Xs )g(Xs )dhXis .
0
Example 3) is a special case of a more general rule for stochastic Ito-integrals.

18
2.3.2
The d-dimensional It
o-formula
Theorem 2.15 (d-dimensional It

o-formula). Given continuous functions X = (X 1 , . . . , X d ) :
[0, T ] R with continuous covariation
(
hX k it , k=l,
k
l
hX , X it =

1/2 hX k + X l it hX k it hX l it , k 6= l
and a twice continuously differentiable function F : Rd R. Then
F (Xt ) = F (X0 ) +
d Z
X
i=1
d Z
1 X t 2
i
F (Xs )dXs +
F (Xs )dhX i , X j is .
xi
2
0 xi xj
i,j=1
Remark on notation: For x

F we often write Fxi , xi xj F is denoted by Fxi ,xj . In
i
short-notation the d-dimensional Ito-formula hence writes as:
dF (Xt ) =
d
X
Fxi (Xt )dXti +
d
1 X
Fxi ,xj (Xt )dhX i , X j it .
2
i,j=1
i=1
Example: Let W = (W 1 , . . . , W d ) be d-dimensional Brownian motion so that that

hWtk , Wtl it = kl t where kl = 1 if k = l and kl = 0 otherwise. Hence we have
F (Wt ) = F (W0 ) +
d Z
X
i=0
Fxi (Ws )dWsi
1X
+
2
i=0
Fxi ,xi (Ws )ds .
(2.9)
Corollary 2.16 (It

os product formula). Given X, Y with continuous quadratic variation
hXit , hY it and covariation hX, Y it . Then
Z
Xt Yt = X0 Y0 +
Ys dXs + hX, Y it .
Xs dYs +
0
Proof. Apply Theorem (2.15) to F (x, y) = xy.

In short notation the product formula can be written as d(XY )t = Xt dYt + Yt dXt +
dhX, Y it .
Corollary 2.17 (It
o-formula for time-dependent functions). Given a continuous function X with continuous quadratic variation hXit and a function F (t, x) which is once
continuously differentiable in t and twice continuously differentiable in x. Then
Z t
Z t
Z
1 t
F (t, Xt ) = F (0, X0 ) +
Ft (s, Xs )ds +
Fx (s, Xs )dXs +
Fxx (s, Xs )dhXis .
2 0
0
0
We now consider several applications of the d-dimensional Ito-formula.
1) Geometric Brownian motion: Given a Brownian motion W , an initial value S0 > 0 and
constants , with > 0 we define geometric Brownian motion S by

St = S0 exp Wt + ( 1/2 2 )t .
19
Geometric Brownian motion will be our main model for the fluctuation of asset prices
in Section 3. Using the It
o-formula we now derive a more intuitive expression for the
dynamics of S. Define Xt := Bt and Yt := ( 1/2 2 )t and note that hXit = 2 t and
hY it = hX, Y it = 0. Let F (x, y) := S0 exp(x + y) such that Fx = Fy = Fxx = F . By
definition St = F (Xt , Yt ), and we get
t
Z
1 t
F (Xs , Ys )dYs +
F (Xs , Ys )dXs +
St = S0 +
F (Xs , Ys )dhXis
2 0
0
0
Z t
Z t
Z
1 2
1 t
= S0 +
F (Xs , Ys )dBs +
F (Xs , Ys )( )ds +
F (Xs , Ys ) 2 ds
2
2
0
0
0
Z t
Z t
Ss ds .
(2.10)
Ss dBs +
= S0 +
Z
In our short-notation the equation solved by S can be Rwritten as dSt = St dt + St dBt .

t
In the special case where = 0 we get that St = S0 + 0 Ss dBs is a local martingale.2
2) Brownian motion and the reverse heat-equation. Consider a function F (t, x) that solves
the reverse heat-equation Ft (t, x) + 1/2Fxx (t, X) = 0 and a Brownian motion B. Then
F (t, Bt ) is a local martingale. The proof is again based on Itos formula. We get
Z
F (t, Bt ) = F (0, B0 ) +
Fx (s, Bs )dBs +
0
Ft (s, Bs )ds +
0
= F (0, B0 ) +
Z
Fx (s, Bs )dBs +
Z
= F (0, B0 ) +
1
2
Fxx (s, Bs )dhBis

0
1
(Ft + Fxx )(s, Bs )ds
2
Fx (s, Bs )dBs .
0
In case that Fx (t, Bt ) is sufficiently integrable F (t, Bt ) is even a real martingale. In that
case one easily obtains a probabilistic representation of the solution of our reverse heat
equation. For more on the interplay between solutions of partial differential equations and
stochastic processes we refer to Chapter 4 and 5 of Karatzas & Shreve (1988).
It can be shown that in that case S is even a full martingale.
20
Chapter 3
The Black-Scholes Model: a

PDE-Approach
We now have all the mathematical tools at hand we need to study pricing and hedging of
derivatives in the classical Black-Scholes model.
3.1
Asset Price Dynamics
As in the classical paper Black & Scholes (1973) we consider a market with two traded
assets, a risky non-dividend-paying stock and a riskless money market account. The price
of the stock at time t is denoted by St1 , the price of the money market account by St0 . For
simplicity we work with a deterministic continuously compounded interest rate r such that
St0 = exp(rt). We now look for appropriate models for the dynamics of the stock-price.
As usual we work on a filtered probability space (, F, P ), {Ft } supporting a standard
Brownian motion Wt representing the uncertainty in our market.
In his now famous PhD-thesis Bachelier (1900) proposed to model asset prices by an
arithmetic Brownian motion, i.e. he suggested the model St1 = S0 +Wt +t for constants
, > 0. While this was a good first approximation to the dynamics of stock prices,
arithmetic Brownian motion has one serious drawback: as St1 is N (S0 +t, 2 t) distributed,
the asset price can become negative with positive probability, which is at odds with the
fact that real-world stock-prices are always nonnegative because of limited liability of the
shareholders.
Samuelson (1965) therefore suggested replacing arithmetic Brownian motion by geometric
Brownian motion

1
St1 = S01 exp Wt + ( 2 t) .
(3.1)
2
We know from (2.10) that this model solves the linear stochastic differential equation
(SDE)
dSt1 = St1 dt + St1 dWt .
Geometric Brownian motion - often referred to as Black-Scholes model - is nowadays widely
used as reference model both in option pricing theory and in the theory of portfoliooptimization; we therefore adopt it as our model for the stock price dynamics in this
21
section. Model (3.1) implies that log-returns

1
1
ln St+h
ln St1 = (Wt+h Wt ) + ( 2 )h
2
are N (( 21 2 )h, 2 h)-distributed; in particular the volatility is the instantaneous
variance of the log-returns. Moreover, under (3.1) log-returns over non-overlapping time
periods are stochastically independent.
There are a number of practical and theoretical considerations which make geometric
Brownian motion attractive as a model for stock price dynamics.
Geometric Brownian motion fits asset prices data reasonably well, even if the fit
is far from perfect. For an overview the empirical deficiencies of the Black-Scholes
model we refer to Chapter I of Cont & Tankov (2003) or Section 4.1 of McNeil, Frey
& Embrechts (2005).
Geometric Brownian motion allows for explicit pricing formulae for a relatively large
class of derivatives. Moreover, there are a number of good discrete-time approximations to the Black-Scholes model such as the CRR model presented in Section ??.
The Black-Scholes model is quite robust as a model for hedging of derivatives: if real
asset-price dynamics are not too different from geometric Brownian motion hedging
strategies computed using the Black-Scholes model perform reasonably well. There
is now a large literature on model risk in derivative pricing; see for instance the
collection of papers in Gibson (2000) or Cont (2006); we present a brief discussion
of the model-risk related to volatility-misspecification in Subsection ?? below.
There are also a number of theoretical considerations in favour of the Black-Scholes model.
The model is in line with the efficient markets hypothesis. Moreover, there are
a number of economic models which show that the Black-Scholes model can be
sustained as a model for economic equilibrium; see for instance He & Leland (1993)
for a rational expectations model and Follmer & Schweizer (1993) for a model based
on the temporary equilibrium concept.
The Black-Scholes model is an arbitrage-free and complete model, making derivative
pricing straightforward from a conceptual point of view.
3.2
Pricing and Hedging of Terminal Value Claims
Consider now a contingent claim with maturity date T and payoff H. As in the discretetime setup of Chapter ?? we want to find a dynamic trading strategy replicating the claim;
such a strategy can be used for pricing and hedging purposes. It can be shown that in the
framework of the Black-Scholes model such a strategy exists for every claim whose payoff
is measurable with respect to the information generated byR the asset price. However, such
t
a result requires the notion of the stochastic Ito-integral 0 s dSs1 for general predictable
processes which we do not have at our disposal. We therefore restrict our analysis to
so-called terminal value claims whose payoff is of the form H = h(ST1 ). For these claims
22
one can find Markov hedging strategies which are functions of time and the current stockprice. This includes most examples which are relevant from a practical viewpoint; as
shown in Section ?? extensions to path-dependent derivatives are also possible. For the
general theory we refer the reader to Bingham & Kiesel (1998) or to the advanced text
Karatzas & Shreve (1998).
3.2.1
Basic Notions
Trading strategy: A Markov trading strategy is given by a pair of smooth functions

((t, S), (t, S)), where (t, St1 ) and (t, St1 ) give the number of stocks respectively of units
of the money market account in the portfolio at time t. The value at time t of this strategy
is given by V (t, St1 ) = St1 (t, St1 ) + (t, St1 )S 0 (t). Note that the strategy can alternatively
be described by specifying the functions (t, S) and V (t, S); the position in the money
market account is then given by the function (t, S) := (V (t, S) S(t, S)) /S 0 (t).
Gains from trade: To motivate the subsequent definitions we introduce piecewise
constant approximations to our trading strategies. Consider a sequence n of partitions
with |n | 0 and define
X
nt () =
(ti1 , St1i1 ()) 1(ti1 ,ti ] (t)
(3.2)
ti n
tn ()
(ti1 , St1i1 ()) 1(ti1 ,ti ] (t)
(3.3)
ti n
and Vtn = nt St1 + tn St0 . A similar argument as the one used in the proof of Lemma ??
now yields that this piecewise constant strategy is selffinancing if and only if we have for
all ti n
Vtni = V0 + Gnt , where Gnt =
i
X

ntj (St1j St1j1 ) + tnj (St0j St0j1 ) .
j=1
Now recall that by definition of the Ito-integral Gnt converges to

Hence the following definition is natural.
Rt
0
Rt
(s, Ss1 )dSs1 + 0 (s, Ss1 )dSs0 .
Definition 3.1. Given a Markov trading strategy ((t, St1 ), (t, St1 )) induced by smooth
functions , : [0, T ] R+ R.
(i) The gains from trade of this strategy are given by
Z
Gt =
(s, Ss1 )dSs1
Z
+
(s, Ss1 )dSs0 .
(ii) The strategy is selffinancing, if V (t, St1 ) = V (0, S0 ) + Gt for all t T .

(iii) Consider a terminal value claim with payoff h(ST1 ). A selffinancing strategy is a
replicating strategy for the claim if V (T, S) = h(S) for all S > 0; in that case
V (t, St1 ) is the fair price of the claim at time t.
23
3.2.2
The pricing-equation for terminal-value-claims
We now derive a partial differential equation (PDE) for the value of the replicating strategy. We have the following
Theorem 3.2. Let V : [0, T ] R+ R be a continuous function which solves the PDE
1
Vt (t, S) + 2 S 2 VSS (t, S) + rSVS (t, S) = rV (t, S) ,
2
(t, S) [0, T ) R+ .
(3.4)
Then the hedging strategy with stock-position (t, S) = VS (t, S) and value V (t, S) is selffinancing. If V satisfies moreover the terminal condition V (T, S) = h(S), the strategy
replicates the terminal value claim with payoff h(ST1 ) and the fair price at time t of the
claim equals V (t, St1 ).
Proof. As a first step we compute the quadratic variation of geometric Brownian motion.
Recall that
Z t
Z t
1
1
Ss1 ds =: Mt + At .
Ss dWs +
St = S0 +
0
As A is of finite variation we get
hS 1 i
= hM it =
Rt
0
2 (Ss1 )2 ds.
Now we turn to the first claim. We get from Itos formula

Z t
Z t
Z
1 t
1
1
1
1
1
V (t, St ) = V (0, S0 ) +
VS (s, Ss )dSs +
Vt (s, Ss )ds +
VSS (s, Ss1 )dhS 1 is
2
0
0
0

Z t
Z t
1
1
1
1
1
2
1 2
1
Vt (s, Ss ) + (Ss ) VSS (s, Ss ) ds .
= V (0, S0 ) +
VS (s, Ss )dSs +
2
0
0
Using the PDE (3.4) and the definition of this equals
=V
(0, S01 )
(s, Ss1 )dSs1
= V (0, S01 ) +
r(V (s, Ss1 ) (s, Ss1 )Ss1 )ds
(s, Ss1 )dSs1 +
(s, Ss1 )dSs0 ,
where (t, St1 ) = (V (t, St1 )(t, St1 )St1 )/S 0 (t) is the position in the money-market account
which corresponds to our strategy. Hence our strategy is selffinancing. The remaining
claims are obvious.
3.3
3.3.1
The Black-Scholes formula

The formula
To price a European call option we have to solve the PDE (3.4) with terminal condition
h(S) = (S K)+ . To solve this problem analytically one usually reduces the PDE (3.4)
to the heat equation by a proper change of variables. This technique is useful also for the
implementation of numerical schemes to solve the pricing PDE; see for instance Wilmott,
Dewynne & Howison (1993). Details are given in the following Lemma.
24
Lemma 3.3. Define (t) = 2 (T t) and z(t, S) = ln S (r + 12 2 )(T t). Denote by

u(t, z) : [0, T ] R R the solution of the heat-equation ut = 12 uzz with initial condition
u(0, z) = (ez K)+ . Then C(t, S) := er(T t) u( (t), z(t, S)) solves the terminal value
problem for the price of a European call.
Proof. We have C(T, S) = u( (T ), z(T, S)) = u(0, ln S) = (S K)+ , so that the function
C has the right value at maturity. Moreover,

C
= er(T t) ru 2 u + (1/2 2 + r)uz
t

2C
C
= er(T t) uz 1/S ,
= er(T t) uzz (1/S)2 uz (1/S)2
2
S
S
Plugging these expressions into the PDE (3.4) yields the result.
It is well-known that the solution u of the heat-equation with initial condition u(0, z) =
u0 (z) equals
Z
1
2
u0 (z + x )ex /2 dx.
u(, z) =
2
From this follows after tedious but straightforward computations (see for instance Sandmann (1999) or Wilmott et al. (1993))
Theorem 3.4. Denote by N () the standard normal distribution function. The noarbitrage price of a European call with strike K and time to maturity T in the BlackScholes model with volatility and interest rate r is given by
CBS (t, S; , r, K, T ) := SN (d1 ) er(T t) KN (d2 ) ,
with
(3.5)
ln S/K + (r + 12 2 )(T t)
d1 =
and d2 = d1 T t .
T t
The corresponding hedge-portfolio consists of S

CBS = N (d1 ) units of the risky asset and
rt
rT
(CBS (t, S) N (d1 )S)/e = e
KN (d2 ) units of the money market account.
A probabilistic derivation of the Black-Scholes formula is given in Section 5.2 below.
3.3.2
Application of the formula: volatility estimation
For an extensive discussion how the Black-Scholes formula can be applied in practice we
refer to Cox & Rubinstein (1985) and Hull (1997). Here we content ourselves with a
few remarks about possible approaches to determine the volatility . As volatility is not
directly observable in contrast to the other parameters in the Black-Scholes formula
finding a good value for is by for the most problematic part in applying the BlackScholes formula. The fact that in real markets volatility is rarely constant but tends to
fluctuate in a rather unpredictable manner makes matters even worse.1 There are two
common approaches to determining .
1
The stochastic nature of volatility has given rise to the development of the stochastic volatility models;
see for instance Frey (1997) for an overview.
25
1) Historical volatility: This approach is based on statistical considerations. Recall

that under (2.10) log-returns over non-overlapping periods of length are independent
and N (( 12 2 ), 2 ) distributed. Given asset price data at times ti , i = 1, . . . , N
with ti ti1 = (e.g. daily returns) define Yi = ln Sti ln Sti1 . The standard estimator
from elementary statistics for , the volatility of the log-returns over the time-period
is given by
!1
N
N
2 2
1 X
1 X
=
Yi Y
, where Y =
Yi .
N 1
N
i=1
i=1
Estimated historical volatility is then given by

hist =
/ .
2) Implied volatility The idea underlying the implied volatility concept is the use of
observed prices of traded derivatives to find the prediction of the market for the volatility
of the stock. To explain the concept we consider the following example:
Assume that a call option with strike K and maturity T is traded at time t and at a
given stock-price (St1 ) for a price of Ct . The implied volatility
impl is then given by the
solution to the equation
CBS (t, (St1 ) ;
impl , K, T ) = Ct .
As CBS is strictly increasing in a unique solution to this equation exists; it is usually
determined by numerical procedures.
In practice traders tend to use a combination of both approaches, implied volatility being
the slightly more popular concept.
26
Chapter 4
Further Tools from Stochastic

Calculus
4.1
Stochastic Integration for Continuous Martingales
Rt
In this section we want to define the stochastic integral 0 Hs dXs for semimartingales of
the form X = X0 + M + A, where M is a continuous martingale and A is a continuous
FV-process, and where H is a suitable limit of integrands of the form
Htn =
n1
X
hti ()1(ti ,ti+1 ] (t),
ti n ,
i=0
where n is a fixed partition of (0, T ], and where each of the hti is Fti -measurable. This
extends the pathwise It
o calculus to a larger class of integrands. This is reasonable from a
financial viewpoint, as we want to work with a larger class of trading strategies
R t than just
Markov strategies of the form (t, St ). The key point is to define the integral 0 Hs dMs , as
M will typically have paths of infinite first variation, so that standard Stieltjes integration
does not apply. To overcome these problems we use special properties of the space M2 of
square integrable martingales.
4.1.1
The Spaces M2 and M2,c
Throughout we work on a given a filtered probability space (, F, P ), {Ft } with fixed

horizon T .
Definition 4.1. The space of all martingales M = (Mt )0tT with M0 = 0 and E[MT2 ] <
is denoted M2 ; M2,c M2 denotes the subspace of all martingales with continuous
trajectories.
Note that the Jensen inequality gives for M M2 , t T that
h
i

E[Mt2 ] = E (E[MT |Ft ])2 E E[MT2 |Ft ] = E[MT2 ],
so that Mt is square integrable for all 0 t T .
Lemma 4.2. M2 is a Hilbert-space with scalar product
hM, N iM2 = E[MT NT ] = hMT , NT iL2 .
27
Proof. hM, N iM2 is a scalar product. In particular, if hM, M iM2 = 0, we have MT = 0

a.s. which yields Mt = E[MT |Ft ] = 0 for all t. The Hilbert-space property is obvious,
as M2 can be identified with L2 (, FT , P ) by identifying a martingale (Mt )t with its
terminal value MT via Mt = E[MT |Ft ].
The following result establishes a link between the pathwise supremum the and the terminal value of a martingale.
Theorem 4.3 (Doob inequality). Let p > 1 and (Mt )0tT be a martingale. Define q by
1
1
p + q = 1. Then we have for t T
k sup |Ms |kLp qkMt kLp .
(4.1)
st
In particular we get with MT := suptT |Mt | that

h
i

E (MT )2 4E MT2 .
(4.2)
We omit the proof.

Lemma 4.4. M2,c is a closed subspace of M2 and hence a Hilbert-space.

Proof. Consider a sequence M n M2,c with M n M , i.e. with limn E (MTn MT )2 =
0. We want to show that the limit (Mt )t0 with Mt = E [MT |Ft ] has continuous sample
paths. By Doobs inequality we get
h
i
h
i
E ((M n M )T )2 4E (MTn MT )2 0,
0
i.e. (M n M )T = suptT |Mtn Mt | 0 in L2 (, FT , P ). It follows that (M n M )T

converges to zero a.s. for a subsequence n0 , which implies the result, as the limit of
continuous functions in the supremum-norm is continuous.
Quadratic variation. Recall the definition
X
Vt2 (M ; n ) :=
(Mti Mti1 )2 =
ti n ; ti <t
(Mi,n )2 ,
(4.3)
ti n ; ti <t
for the quadratic variation of M along a partition n of [0, T ].

Theorem 4.5. Consider a sequence (n )n of partitions of [0, T ] such that |n | 0 and
let M M2,c . Then
1. Vt2 (M, n ) converges in probability to an increasing and continuous process hM it
with hM i0 = 0.

2. Mt2 hM it is a martingale, in particular we have E Mt2 = E [hM it ], t T .
3. hM it is uniquely defined by the requirements
(i) hM it is increasing and continuous, hM i0 = 0.
(ii) Mt2 hM it is a martingale.
28
Remark 4.6. 1. The hard parts of the theorem are 1 and 2; the uniqueness of hM it
follows immediately from the fact that a continuous martingale with trajectories of finite
variation is a.s. constant. 2. The convergence of Vt2 (M, n ) hM it in probability
implies that Vt2 (M, n0 ) hM it almost surely for a subsequence n0 ; hence the pathwise
Ito-calculus of Chapter 1 applies to martingales from M2,c
Recall that for two functions X, Y : (0, T ] R such that hXit and hY it exist, the
quadratic covariation
X
Xi,n Yi,n
hX, Y it = lim
n
ti n ; ti t
exists if and only if hX + Y it exists and that hX, Y it = 21 (hX + Y it hXit hY it ), the
so-called polarization identity. Now for M, N M2,c , M + N M2,c and hM + N it exists
by Theorem 4.5. Hence we get
Corollary 4.7. For M, N M2,c the covariation
X
hM, N it = lim
Mi,n Ni,n
n
ti n ; ti t
exists. hM, N i is characterized by the following properties

(i) hM, N it is a continuous FV-process with hM, N i0 = 0.
(ii) Mt Nt hM, N it is a martingale.
The result follows from Theorem 4.5 and the polarization identity.
Note that h, i is bilinear and symmetric on M2,c , i.e. hM1 + M2 , N i = hM1 , N i +
hM2 , N i. Note that for M, N M2,c , hM, N i gives the conditional covariance of the
increments of M and N , i.e. we have for t s 0
E [(Mt Ms ) (Nt Ns ) |Fs ] = E [hM, N it hM, N is |Fs ] .
(4.4)
The proof of (4.4) is easy. We have

E [(Mt Ms )(Nt Ns )|Fs ] = E [Mt Nt Mt Ns Ms Nt + Ms Ns |Fs ]
= E [Mt Nt |Fs ] 2E [Mt Ns |Fs ] + E [Ms Ns |Fs ]
|
{z
} |
{z
}
=2Ms Ns
=Ms Ns
= E [Mt Nt |Fs ] Ms Ns
= E [hM, N it hM, N is |Fs ] ,
where we have used the fact that Mt Nt hM, N it is a martingale.
Example 4.8 (Constructing correlated Brownian motions.). Let W1 and W2 be independent Brownian
pmotions and recall from Section 2.3.1 that hW1 , W2 it = 0. Put B1 = W1 ,
B2 = W1 + 1 2 W2 . Then B1 and
p B2 are standard one-dimensional Brownian motions, and hB1 , B2 it = hW1 , W1 it + 1 2 hW2 , W1 it = t.
29
4.1.2
Stochastic Integrals for elementary processes
We begin the construction of

defined as follows:
:=
Rt
0
Hs dMs by taking H from the class of elementary processes,
n1
n
o
X
H : Ht () =
hti ()1(ti ,ti+1 ] (t) ,
(4.5)
i=0
for deterministic 0 t0 < t1 < ... < tn < T = tn+1 , n N, and bounded, Fti -measurable
random variables hti , 1 i n. Note that the elements of are bounded and leftcontinuous. From a financial point of view, elements of correspond to simple, piecewise
constant trading strategies.
Definition 4.9. For H and M M2,c we define the stochastic integral (H.M ) by
Z t
X

hti Mti+1 t Mti t , t T,
Hs dMs :=
(H.M )t :=
0
0in
H is called integrand, M integrator.

Lemma 4.10 (Properties of (H.M R) for H ). We have (H.M ) M2,c and the quadratic
t
variation is given by h (H.M ) it = 0 Hs2 dhM is .
Corollary 4.11 (It
o-isometry). We have
Z t

Z t

2
2
2
E (H.M )t = E ( Hs dMs ) = E
Hs dhM is .
0
(4.6)
Proof. (of Lemma 4.10) The continuity of (H.M ) is obvious, as M is continuous. In order
to show that (H.M ) is a martingale, by the optional sampling theorem we have to show
that for all bounded stopping times with T a.s. we have E [(H.M ) ] = (H.M )0 .
Now (H.M )0 is obviously equal to zero. Moreover,
X
E [(H.M ) ] = E
hti (Mti+1 Mti )
0in

E hti E Mti+1 Mti |Fti .
0in

By the optional sampling theorem E Mti+1 Mti |Fti = 0, and the martingale propRt
erty of (H.M ) follows. In order to show that h(H.M )it = 0 Hs2 dhM is , we show that
R
t
(H.M )2t 0 Hs2 dhM is is a martingale, which gives the result by Statement 3 of Theorem
4.5. Again we use the optional sampling theorem. Consider a stopping time T . We
have
n+1
X

E (H.M )2 =
E hti1 htj1 Mi Mj ,
(4.7)
i,j=1
where M
h i = Mti Mti1
i . Suppose that i < j. By conditioning on Ftj1 we see
that E hti1 htj1 Mi Mj = 0. Hence (4.7) equals

n+1
X
i=1
h
i n+1
X h
i
E h2ti1 (Mi )2 =
E h2ti1 hM iti hM iti1 ,
i=1
30
using (4.4). Now we obviously have hM it = hM i t , so that

n+1
X
h2ti1
hM iti hM iti1
i
i=1
n+1
X
h
i
E h2ti1 hM iti hM iti1
i=1
Z
= E

Ht2 dhM it .
hR
i

T
The relation E (H.M )2T = E 0 Hs2 dhM is , which we obtain by putting T , shows

that (H.M )T L2 (, FT , P ), as H is bounded and as E [hM iT ] = E MT2 < since
M M2 .
4.1.3
Extension to General Integrands
We use the It
o-isometry to extend the integral (H.M
hR ) from toi a larger class of integrands.
T
We begin with an abstract interpretation of E 0 Hs2 dhM is for M M2,c fixed.
= [0, T ], F = FT B([0, T ]) and define a measure PM on ,
F by
Put
Z T

PM (A) = E
1A (, t) dhM it , A F.
(4.8)
With this definition we have for H

M H 2 = E
kHkL2 (,
F ,PM ) = E
hR
T
Z
Hs2 dhM is
F)
with
Hs2 dhM is is the L2 -norm of H (regarded as a random variable on ,
respect to the measure PM . By Corollary 4.7 the mapping
i.e. E
I : M2,c ,
H 7 (H.M )
2,c
is therefore an isometry from (, k kL2 (,
F ,PM ) ) to (M , k kM2 ). Denote by the closure
n
of in k kL2 (,
F ,PM ) , and let H . By definition there exists a sequence H with
F,
PM ); in particular the sequence H n is Cauchy. By the Ito-isometry
H n H in L2 (,
n
the sequence I(H ) = (H n .M ) is therefore Cauchy in M2 respectively M2,c . Since M2,c
is a Hilbert space (Lemma 4.2), the limit limn (H n .M ) exists in M2,c , and the following
definition makes sense.
F,
PM ) the stochastic integral (H.M ) is the uniquely
Definition 4.12. For H L2 (,
defined element of M2,c such that

lim k(H n .M ) (H.M )kM2 = 0
for all H n with kH n HkL2 (,

F ,PM ) 0.
Corollary 4.13. Let M M2,c , H1 , H2 and 1 , 2 R. Then we have

Linearity of (H.M ): (1 H1 + 2 H2 . M ) = 1 (H1 . M ) + 2 (H2 . M ) and
hR
i

t
It
o-isometry: E (H.M )2t = E 0 Hs2 dhM is , for t T .
In the next theorem we give a partial characterization of the abstract set .
2,c
Theorem 4.14.
hR Let M Mi . Then contains all adapted, left-continuous processes
T
H such that E 0 Hs2 dhM is < .
We omit the proof.

31
Extensions.
ber of ways.
The definition of the stochastic integral (H.M ) can be extended in a num-
a) Localisation.
R T 2 If H is adapted, leftcontinuous and M is a continuous local martingale
with 0 Hs dhM is < P -a.s. we hmay find an increasing
sequence of stopping times
i
RT 2
2,c
n
n
n % T such that M M , E 0 Hs dhM is < . Then, for t < n we put
(H.M )t = (H.M n )t ,
(4.9)
where the right hand side is defined by Definition 4.12. It is easily shown that
(4.9) gives a consistent definition of (H.M )t , and that (H.M )t is a continuous local
martingale.
b) Semimartingales as integrands. Let X = X0 + M + A, where M is a continuous
local martingale and A a continuous
R T 2 FV-process.
R T Then one defines for H adapted
and left continuous such that 0 Hs dhM is + 0 |Hs | d|A|s < ,
Z
Z
0
Hs dAs ,
Hs dMs +
Hs dXs =
where the integral with respect to M is constructed as before and

ordinary Stieltjes integral.
4.1.4
Rt
0
Hs dAs is an
Kunita-Watanabe characterization
Theorem 4.15 (Kunita-Watanabe characterization). Let H , M M2,c . Then we

have for all N M2,c
Z t
Hs dhM, N is .
(4.10)
h(H.M ), N it =
0
Conversely if L
have L = (H.M ).
M2,c
satisfies hL, N it =
Rt
0
Hs dhM, N is , t T for all N M2,c , we
Partial proof. Establishing (4.10) for H is a technical approximation argument and

will be omitted. Uniqueness of a process L that satisfies (4.10) is easy to show. (4.10)
implies that
Z T

E [LT NT ] = E [hL, N iT ] = E
Hs dhM, N is
for all N M2,c ,
0
i.e. the scalar product hL, N iM2 is determined from (4.10) for all N M2,c , , so that L
is uniquely determined.

Further Properties of (H.M ). Theorem 4.15 allows us to establish a number of very
useful properties of the stochastic integral (H.M ).
Quadratic covariation. If we put N = M we get
Z t
Z t
h(H.M ), M it =
Hs dhM, M is =
Hs dhM is ,
0
32
and hence
Z t
Hs dhH.M, M i
h(H.M )it = h(H.M ), (H.M )it =
0
Z s
Z t

Hu dhM iu
Hs d
=
0
0
Z t
=
Hs2 dhM is ,
(4.11)
by the chain-rule of Stieltjes-calculus, i.e. we have shown that h(H.M )it =

for H (and not just in ).
Rt
0
Hs2 dhM is
Chain rule
Consider two integrands H1 , H2 such that
hR of stochastic integration.
i
T
2
E 0 (H1,s H2,s ) dhM is < . Then we have (H1 . (H2 . M )) = (H1 H2 . M ) or, in
the long version,
Z
Z
t
H1,s d(H2 . M )s =
0
H1,s H2,s dMs .
(4.12)
Relation (4.12) is the chain rule of stochastic integration.

Proof. Let L = H1 . (H2 . M ). By Theorem 4.15 we get
Z t
Z t
Z s
hL, N it =
H1,s dh(H2 . M ), N is =
H1,s d
H2,u dhM, N iu
0
0
0
Z t
=
H1,s H2,s dhM, N is = h(H1 H2 . M ), N it .
0
Hence Lt =
Rt
0
H1,s H2,s dMs as claimed.
The next result is important in the context of incomplete markets.

Theorem 4.16 (Kunita Watanabe decomposition). Consider martingales N, M1 , . . . , Mn
M2,c . Then there is a unique decomposition of the form
n Z t
X
Nt = N0 +
Hs,i dMs,i + Lt , t T,
i=1
with H1 , . . . , Hn and L M2,c strongly orthogonal to M1 , . . . , Mn , that is with

hL, Mi i 0 for all 1 i n.
Note that L is also weakly orthogonal to M1 , . . . , Mn , that is E(LT MT,i ) = L0 M0,i = 0
as LMi is a martingale.
Sketch of the proof. We let n = 1. By the Ito-isometry, the space of stochastic integrals
MH = {(H.M ) : H } is a closed subspace of M2,c . Hence we can project the rv
NT N0 on that space, that is there is a representation
Z T
NT = N0 +
Hs dMs + LT
0
RT
and E(L 0 Hs dMs ) = 0 for all H . Define the martingale L via Lt = E(LT | Ft ).
Then L is the desired martingale. It remains to show that hL, M i 0. Here one uses
MH implies that also the stopped martingale
that MH is stable under stopping, i.e. M
MH for an arbitrary stopping time ; we omit the details.
M
33
4.2
It
o Processes and the Feynman-Kac formula
It
o processes. It
o processes are solutions of stochastic differential equations driven by
Brownian motion; they will be our basic model for asset price dynamics.
Definition 4.17. Given a d-dimensional Brownian motion W = (Wt,1 , ..., Wt,d )t0 , a time
point t0 0, some vector x Rn and functions : R+ Rn Rn and : R+ Rn Rnd .
Then the n-dimensional process X = (Xt,1 , ..., Xt,n )t0 is called an It
o process with initial
value t0 , x, drift and dispersion matrix if X satisfies the SDE
Z
i (s, Xs ) ds +
Xt,i = xi +
t0
d Z
X
j=1
ij (s, Xs ) dWsj ,
t t0 .
(4.13)
t0
In short notation (4.13) is often written in the form dXt = (t, Xt ) dt + (t, Xt ) dWt .
Moreover, one frequently takes t0 = 0.
An intuitive way to understand the SDE (4.13) is the Euler approximation. For a small
time step t one has with tn = nt
Xtn+1 = Xtn + (tn , Xtn )t +
d
X
j (tn , Xtn )n,j
(4.14)
j=1
where n,j are iid N (0, t ). (4.14) can be used to generate (approximations of) the
trajectories of X on a computer.
Define the n n matrix C(t, X) = (t, X) 0 (t, X). Then we have
hXi , Xj it = h
d Z
X
ik dWs,k ,
k=1 0
d Z t
X
k,l=1 0
d Z t
X
k=1
Z t
d Z
X
l=1
jl dWs,l it
ik jl d hWk , Wl is
| {z }
=kl s
ik jk ds
Cij (s, Xs ) ds.

0
Hence C is sometimes called instantaneous covariance matrix of X.

Let f : Rn R be a smooth function and define the differential operator A by
Af (t, x) =
n
X
i=1
n
f
1 X
2f
i (t, x)
(x) +
Cij (t, x)
(t, x).
xi
2
xi xj
(4.15)
i,j=1
A is known as generator of the Ito process X. We have

Lemma 4.18. For f : Rn R smooth
R t f such that f and its first two derivatives are
f
bounded the process Mt = f (Xt ) 0 t (s, Xs ) + Af (s, Xs ) ds is a martingale.
34
Proof. We consider only the time-independent case. Applying Itos formula in n dimensions we get
f (Xt ) = f (X0 ) +
n Z
X
i=1
= f (X0 ) +
+
1
2
n Z
X
i,j=1
i=1 0
n
XZ t
i,j=1 0
= f (X0 ) +
n Z
f
1 X t 2f
(Xs ) dXs,i +
(Xs ) dhXi , Xj is
xi
2
0 xi xj
n
XX
f
(Xs )i (s, Xs ) ds +
xi
Z
0
i=1 j=1
f
(Xs )i (s, Xs ) dWs,j
xi
2f
Cij (s, Xs ) ds
xi xj
n X
d Z
X
i=1 j=1
f
(Xs )i (s, Xs ) dWs,j +
xi
Af (s, Xs )ds.
0
Rt
Hence f (Xt ) 0 Af (s, Xs ) ds can be represented as a sum of stochastic integrals wrt
Brownian assumed motion, and is therefore a local martingale. Since f and its derivatives
are bounded by assumption, this expression is a true martingale and the claim follows.
In the following theorem we give conditions ensuring that a solution to the SDE (4.13)
exists.
Theorem 4.19. Suppose that , satisfy Lipschitz and growth conditions of the form
k(t, X) (t, Y )k KkX Y k,
t 0,
k(t, X) (t, Y )k KkX Y k,
t 0,
k(t, X) + (t, X)k K(1 + kXk),
t 0.
Then for all initial values (t0 , x) [0, )Rn a unique solution to (4.13) exists. Moreover,
this solution is (FtW )-adapted, where FtW = (Ws,i : 1 i d, s t).
The Feynman Kac formula. Lemma 4.18 forms the basis for a close interplay between
stochastic processes and solution of parabolic PDEs, which is extremely fruitful in financial
mathematics. The basic result is the celebrated Feynman Kac formula.
We start with the one dimensional case, which is notationally easier. Consider scalar
functions (t, x), (t, x), and r(t, x) and some function on R. Suppose that F (t, x) is a
solution to the terminal value problem
F
F
1
2F
(t, x) + (t, x)
(t, x) + 2 (t, x) 2 (t, x) =r(t, x)F (t, x),
t
x
2
x
F (T, x) =(x).
(4.16)
(4.17)
We want to represent F in terms of some Ito process X. Consider the SDE

dXt = (t, Xt )dt + (t, Xt )dWt ,
Xt0 = x
and recall that the infinitesimal generator of X equals

Af (t, x) = (t, x)
f
1
2f
+ 2 (t, x) 2 ,
x 2
x
35
f C 1,2 ,
(4.18)
so that (4.16) can be written as F

t + AF = rF . Consider now the processR Zt =
Rt
t
exp( t0 r(s, Xs )ds)F (t, Xt ). Application of Ito s Lemma Yields with Dt = exp( t0 r(s, Xs )ds)
dZt = d(Dt F (t, Xt )) = F (t, Xt )dDt + Dt dF (t, Xt ) + dhD, F (t, X)it .
Now D is an FV-process which solves the ODE dDt = r(t, Xt )Dt dt, so that hD, F (t, X)it
F
0. Moreover, the It
o formula gives dF (t, Xt ) = ( F
t +AF )(t, Xt )dt+ x (t, Xt )dWt . Hence
we get, using the chain rule for stochastic integration
dZt = r(t, Xt )F (t, Xt )Dt dt + {(
+ Dt
f
+ AF )(t, Xt )Dt }dt
t
F
(t, Xt )dWt .
x
Using the PDE (4.16) for F we see that the dt terms vanish, i.e.
Z t
F
Zt = Zt0 +
(s, Xs )Ds (s, Xs )dWs
t0 x
is a (local) martingale and a true martingale given sufficient integrality, which we assume
form now on. Moreover, by definition Dt0 = R1 so that Zt0 = F (t0 , Xt0 ) = f (t0 , x). The
T
martingality of Z now gives, as ZT = exp t0 r(s, Xs )ds (XT ) because of (4.17),

F (t0 , Xt0 ) = Zt0 = Et0 ,x (ZT ) = Et0 ,x exp(

r(s, Xs ))(XT ) .
(4.19)
t0
Formula (4.19) is called Feynman Kac formula. It can be used in two ways:
We can use probabilistic techniques or Monte-Carlo simulation to compute the expectation on the rhs of (4.19) in order to solve numerically the PDE (4.16), (4.17).
We can try to solve the PDE (4.16), (4.17), perhaps numerically, in order to compute
the expectation on the rhs of (4.19).
Both approaches are frequently used in financial mathematics. For a generalization of
(4.19) to multi-dimensional processes and for a precise statement of the necessary integrality conditions we refer to Section 5.7 of Karatzas & Shreve (1988).
Example 4.20 (The Black-Scholes PDE). In Theorem 3.2 we showed that the fair price
Vt = (t, St ) of a terminal-value claim with payoff h(ST ) in the Black-Scholes Model solves
the terminal value problem
1
ut + rSuS + 2 S 2 uSS = ru,
2
u(T, S) = h(S).
In order to give a probabilistic interpretation we consider the SDE

dSt = rSt dt + St dWt
2
with generator A = rS S
+ 12 2 S 2 S
2 , and we obtain from (4.19)

u(t0 , S) = Et0 ,S er(T t0 ) h(ST ) , t0 T.
(4.20)
This can be viewed as risk-neutral pricing formula in continuous time. In the next Chapter
we give a derivation of (4.20) using only probabilistic techniques.
36
4.3
4.3.1
Change of Measure and Girsanov Theorem for Brownian motion

Motivation
In discrete-time models, it was shown that the price in t < T of any attainable claim with
FT -measurable payoff H is given by the risk-neutral pricing formula

H
Q
(4.21)
Ht = St,0 E
|Ft , t T.
ST,0
Here St,0 > 0 represents the numeraire at time t (often St,0 = exp(rt)), and Q satisfies
the following properties:
(i) Q P , i.e. Q(A) > 0 P (A) > 0,
A FT .
(ii) Theh discounted

price processes St,i = St,i /St,0 , 1 i n, are Q-martingales, i.e.
i
E Q St,i |Fs = Ss,i , 0 s t T .
In this section we provide the mathematical tools for extending (4.21) to continuous-time
models driven by Brownian motion.
By the Radon-Nikodym theorem, the equivalence of P and Q implies the existence of an
FT -measurable rv Z with P (Z > 0) = 1 such that for all A FT
Q(A) = E P [Z1A ] = E P [Z; A] .
(4.22)
In particular, we get E P [Z] = Q() = 1. Conversely, any FT -measurable rv Z with

E[Z] = 1 and P (Z > 0) = 1 can be used to define a measure Q on FT via (4.22); P
and Q are obviously equivalent. Z is called the (Radon-Nikodym) density of Q wrt P ,
denoted by Z = dQ
dP . Note that (4.22) implies that for every FT -measurable X 0 we
have E Q [X] = E P [ZX] (by measure theoretic induction).
Example 4.21. For finite = {1 , ..., k } two measures P and Q are equivalent if for
1 i k : P ({i }) > 0 Q({i }) > 0; it is immediate from (4.22) that dQ
dP () =
Q({})/P ({}) in that case.
In the next Example we show how a change-of-measure can be used to alter the mean of
a normally distributed random variable.
2
Example 4.22. Let X
N (0, ) on some (, F, P ). Define a random variable Z by
2
Z = exp 2 X 12 2 . Then we get, using standard rules for the mean of lognormal
random variables

h

1 2
i 1
E[Z] = exp 2 exp E 2 X + var 2 X
2

2
2
1
1
= exp 2 exp
= 1,
2
2 2
so that we can define a measure Q by putting dQ

dP = Z. Note that we have for any bounded
g : R R:

Z
x2
1
1 2
Q
P
E [g(X)] = E [Zg(x)] =
g(x) exp
x
e 22 dx
2
2
2
2 2

Z
2
1
(x )
=
g(x) exp
dx,
2 2
2 2
37
which shows that X N (, 2 ) under Q.

The main result of this section, Girsanovs theorem for Brownian motion, can be thought
of as the infinite-dimensional analogue of this example: by a proper change of measure it
is possible to alter the drift of an Ito process. In particular, by choosing Q P properly,
it is usually possible to turn an asset price into a martingale.
4.3.2
Density martingales
Now we return to the change of measure for stochastic processes. Suppose that we have
a filtered probability space (, F, P ), (Ft )0tT , and a strictly positive, FT -measurable
random variable Z with E[Z] = 1, and define the measure Q by E Q [X] = E P [XZ], X
bounded, FT -measurable. Define the associated density martingale by
Zt = E P [Z|Ft ] ,
0 t T.
(4.23)
Lemma 4.23. (Zt )0tT is a martingale, and for every Ft -measurable random variable
Y we have E Q [Y ] = E P [Y Zt ], t T .
Proof. The martingale property of (Zt )t0 is obvious. For the second claim note that by
iterated conditioning

E Q [Y ] = E P [Y Z] = E P E P [Y Z|Ft ] = E P Y E P [Z|Ft ] = E P [Y Zt ] .
The next lemma extends this result to conditional expectations.

Lemma 4.24 (abstract Bayes formula). Given 0 s t T and let Y be some Ft measurable integrable random variable. Then
1 P
E Q [Y |Fs ] =
E [Y Zt |Fs ]
(4.24)
Zs
Proof. We have to check that the right hand side of (4.24) satisfies the characterizing
equation of conditional expectations, i.e.

1 P
Q
E
E [Y Zt |Fs ] 1A = E Q [Y 1A ] , A Fs .
Zs
Using Lemma 4.23, the left hand side equals, as A Fs ,

1 P
P
E Zs E [Y Zt |Fs ] 1A = E P E P [Y Zt 1A |Fs ] = E P [Y Zt 1A ] ,
Zs
which is equal to E Q [Y 1A ] by Lemma 4.23.
Lemma 4.25. Let Q P with dQ
dP = Z. An adapted process (Mt )0tT is a Q-martingale
if and only if the process (Mt Zt )0tT is a P -martingale.
Proof. By the Bayes formula (4.24) we have for t T with ZT := Z
1 P
E Q [MT |Ft ] =
E [MT ZT |Ft ] .
Zt
(4.25)
If (Mt Zt )0tT is a P -martingale the right hand side equals Z1t Mt Zt = Mt and M is
a Q-martingale. Conversely, if M is a Q-martingale the left hand side equals Mt and
multiplying (4.25) with Zt gives Mt Zt = E P [MT ZT |Ft ].
38
4.3.3
The Girsanov Theorem
We begin with the one-dimensional version. Let (Wt )0tT be a one-dimensional Brownian
motion on the filtered probability space (, F, P ), (Ft )0tT . Let (t )0tT be an adapted
process, and define

Z t
Z
1 t 2
ds , 0 s T.
(4.26)
Zt = exp
s dWs
2 0 s
0
We show
R t that Zt 1 can
R t 2be written as a stochastic integral with respect to W . Define
Xt = 0 s dWs 2 0 s ds. Then
Z t
Z .
s2 ds,
hXit = h s dWs it =
0
and since Zt = exp (Xt ), It

os formula gives with f (x) = ex :
1
dZt = f 0 (Xt ) dXt + f 00 (Xt ) dhXit
2
1 Xt 2
Xt
= e dXt + e t dt
2
1 2
= Zt dXt + t Zt dt
2
1
1
= t Zt dWt t2 Zt dt + t2 Zt dt,
2
2
Rt
so that Zt = Z0 + 0 s Zs dWs . It follows that Z is a nonnegative local martingale and
hence a supermartingale. Z is a true martingale if and only if E[ZT ] = Z0 = 1. In that
case Z can be used as a Radon-Nikodym density.
Theorem 4.26 (Girsanov). Suppose that (Zt )0tT defined in (4.26) is a martingale.
Define a probability measure Q on FT by putting dQ

dP = ZT . Then the process W defined
by
Z t
Wt = Wt
s ds, 0 t T
(4.27)
0
is a Brownian motion under Q.

R
t + t s ds is under Q a Brownian motion with drift
Remark 4.27. The process Wt = W
0
(t )0tT ; in that sense Theorem 4.26 generalises Example 4.22.
Proof. The proof is based on the Levy-characterization of Brownian motion:
Theorem 4.28 (Levy). Given a filtered probability space (, F, P ), (Ft )t0 . An adapted
process M = (Mt )t0 with continuous trajectories is a one-dimensional Brownian motion
if and only if the following conditions hold:
1. M is a local martingale.
2. hM it = t, where hM i is the pathwise quadratic variation defined for instance in
(4.3).
39
Rt
it , and since W is
Condition 2 is easily verified: Since 0 s ds is FV, we have hW it = hW
a P -Brownian motion, we have hW it = t P -a.s. and hence also Q-a.s. (as P Q). The
is a Q-local martingale. By Lemma 4.25 this
theorem is proven if we can show that W
is equivalent to showing that

R t (Wt Zt )0tT ia a P -local martingale. Recall that Z has the
representation Zt = Z0 + 0 s Zs dWs . Hence we get, using Itos product formula
t Zt = W
t dZt + Zt dW
t + dhW
, Zit .
dW
Now we have
, Zit
hW
(i)
(ii)
Z .
hW, Zit = hW,
s Zs dWs it
0
Z t
s Zs dhW, W is
0
Z t
s Zs ds,
(4.28)
differ only by a continuous FV process and (ii) the Kunitausing (i) that W and W
Watanabe characterization (Theorem 4.15). Hence we get, using (4.28) and the chain-rule
for stochastic integrals (4.13),
t Zt = W
t t Zt dWt + Zt dWt Zt t dt + Zt t dt
dW
t t Zt + Zt ) dWt .
= (W
t Zt has a representation as an integral with respect to
This shows that the product W
the P -Brownian motion (Wt )0tT and is therefore a P -local martingale, proving the
result.
Remark 4.29. The following conditions are sufficient to ensure that (Zt )0tT
tingale, so that Theorem 4.26 applies.
i
hR
T
(a) E 0 s2 Zs2 ds < (integrable quadratic variation).
dis a mar- dimensional

version
(b) E [ZT ] = 1. (This is necessary and sufficient.)

h
R
i
T
(c) Novikov-criterion: E exp 21 0 |s |2 ds < .
We discuss two mathematical applications of Theorem 4.26.
1. Solution of SDEs via change of measure. Given some function b(s, X) : R+ R,
bounded (but not Lipschitz). We want to construct a solution X of the SDE
dXt = b(t, Xt ) dt + dWt ,
0 t T,
(4.29)
with W some Brownian motion. Start with some probability space (, F, P ), (Ft )0tT ,
supporting a Brownian motion (Xt )0tT . Define a measure Q via

Z T

Z
dQ
1 T 2
= ZT := exp
(s, Xs ) dXs
(s, Xs ) ds .
dP FT
2 0
0
Rt
By Theorem 4.26, Wt := Xt 0 (s, Xs ) ds is a Q-Brownian motion; hence
Z t
Xt = Wt +
(s, Xs ) ds
0
solves equation (4.29) for a Q-Brownian motion W .

40
2. Maximum-likelihood estimation for a Brownian motion with drift. We

observe a process Xt = Wt + t, R a constant and W a Brownian motion and
want to estimate the unknown parameter . A possible approach is maximum-likelihood
estimation (MLE):
F)
whose distribution P
Consider some random variable X with values in the space (,
F)
depends on an unknown parameter A . Suppose moreover, that there is a
on (,
F)
such that P P and
common reference measure P on (,
dP
= L(X; ),
dP
A .
of X, the (abstract) ML-estimator

Given some observation X
for is given by
),
= arg max L(X,

A
(4.30)
most likely. ML-estimators typically

i.e.
is the value of that makes the observation X
have desirable properties such as asymptotic normality and consistency.
Let us apply the approach to the Brownian motion with drift X. In that case we choose
= C([0, T ], R), P equal to the Wiener measure, and we let Xt (
) =
(t) (the so-called
coordinate process on C([0, T ], R)). Moreover, we define P by

dP
1 2
= exp XT T , R,
(4.31)
dP
2
By definition of the Wiener measure, X is a P Brownian motion. Moreover, as in the
previous application, Xt t is a P -Brownian motion, so that under P Xt = (Xt
t )0tT of
t) + t is a Brownian motion with drift . Given an observed trajectory (X
X, the ML-estimator is hence given by

1 2
= arg max exp XT T

R
2
1
T 2 T.
= arg max X
(4.32)
R
2
T
Differentiating (4.32) with respect to gives X
T = 0 and hence the ML-estimator
XT
= T .
41
Chapter 5
Financial Mathematics in
Continuous-Time
5.1
5.1.1
Basic Concepts
The Model
We start with a general model for a frictionless security market with continuous trading.
Fix some horizon date T and consider some probability space (, F, P ) with a filtration
{Ft }0tT . There are d + 1 traded assets with price process S = (S 0 , . . . , S d ). We
assume throughout that S 0 , . . . , S d follow semimartingales with continuous trajectories.
We assume
that S 0 (t) > 0 P a.s. for all t and use S 0 as numeraire; often S 0 (t) =
Rt
exp( 0 rs ds) for some adapted process r with rs 0. In that case S 0 represents the
so-called savings account and r is the short rate of interest.
d+1
Selffinancing trading
adapted and left-continuous pro strategies: A R -valued
0
d
i
cess (t) = t , . . . , t is called a trading strategy; t represents the number of units of
security i in the portfolio
time t. We assume that i is sufficiently integrable so that
R t at
i
the stochastic integral 0 s dSsi is well defined.
P
Definition 5.1. (i) The value of the portfolio at time t equals Vt = Vt = di=0 it Sti ;
V = (Vt )0tT is called value process of the strategy.
(ii) The gains process of the strategy is given by

Gt =
u dSu :=
0
d Z
X
i=0
iu dSui
(iii) The strategy is called selffinancing, if for all t [0, T ] Vt = V0 + Gt .

The economic justification of (ii) and (iii) is the same as in the case of the Markov strategies
discussed in Chapter 3.
Numeraire invariance: We choose S 0 as our numeraire and introduce the discounted
price process
St = (1, St1 /St0 , . . . , Std /St0 ),
(5.1)
42
Pd
i i
and the discounted value process Vt = Vt /St0 =
i=0 t St . Intuitively, one would
expect that a change of numeraire has no substantial economic implications. The next
result confirms this fact.
Proposition 5.2 (Numeraire Invariance). Selffinancing strategies remain selffinancing
after a change of numeraire.
Proof. Let Xt = St0
1
. We have, using Ito s product formula, that
dStk = dStk Xt = Stk dXt + Xt dStk + dhS k , Xit ,
(5.2)
P
and similarly dVt = Vt dXt + Xt dVt + dhX, V it . Now we have Vt = dk=0 kt Stk by
P
definition and, as is selffinancing, dVt = dk=0 kt dStk . This gives the following dynamics
of the discounted wealth process V :
dVt =
d
X
kt Stk dXt +
k=0
d
X
k=0
d
X
d
X
Xt kt dStk +
k=0
d
X
kt dhS k , Xit
k=0
kt {Stk dXt + Xt dStk + dhS k , Xit }
(5.3)
kt dStk ,
k=0
P
as follows by comparing (5.3) and (5.1). Hence we have shown that dVt = dk=0 kt dStk
so that the discounted value process can be represented as sum of the discounted initial
value and of the gains process wrt the discounted securities prices.
Corollary 5.3.
(i) A trading strategy is selffinancing if and only if
V = V0 + G
t
with
=
G
t
d
X
kt dStk .
(5.4)
k=1
(ii) A selffinancing strategy is is completely determined by the initial investment V0 and

the position 1 , . . . , d in the assets S 1 , . . . , S d .
0
Proof. Statement
R t 0 0 (i) follows immediately from the numeraire invariance (note that St 1,
so that 0 t dSt 0).
For the second statement note that V0 and 1 , . . . , d uniquely determine the discounted
wealth process Vt by (5.4) and hence the undiscounted valued process Vt ; the position
0t is then given by
d
X
0t = (St0 )1 {Vt
kt Stk }.
k=1
43
5.1.2
Martingale Measures and Arbitrage Opportunities
Definition 5.4 (Arbitrage). A selffinancing portfolio such that P (Vt 0

t T ) = 1, V0 = 0 and P (VT > 0) > 0 is an arbitrage opportunity.
for all
Arbitrage opportunities represent a chance to create risk free profits and should not exist
in a well-functioning market.
Definition 5.5 (Martingale measure). A probability measure Q is called an equivalent
martingale measure if
(i) Q P on FT , i.e. for every A FT , Q(A) = 0 P (A) = 0.
(ii) The discounted price processes Sk , 1 k d are Q-martingales.
Note that a martingale measure is always related to a given Rnumeraire (due to the dist
counting in (ii)). If the numeraire is of the form St0 = exp( 0 rs ds), r the spot rate of
interest, Q is also called spot martingale measure.
Lemma 5.6. Q is a spot martingale measure, if and only if every price process S i has
Q-dynamics of the form
dSti = rt Sti dt + dMti ,
0 i d,
(5.5)
where M 0 , . . . , M d are Q-martingales.

Rt
Proof. We get, as (St0 )1 = exp( 0 rs ds), that d(St0 )1 = rt (St0 )1 dt. Ito s product
formula gives, as (St0 )1 is of finite variation, that
dSti = rt Sti dt + (St0 )1 dSti ,
which is a martingale if and only if S i has dynamics (5.5).
The next result shows that the existence of an equivalent martingale measure excludes
arbitrage opportunities.
Theorem 5.7. If the model S 0 , . . . , S d admits an equivalent martingale measure Q, there
are no arbitrage opportunities.
Proof. Step 1: For any selffinancing strategy such that Vt R 0, the discounted wealth
P
t
process Vt is a Q-supermartingale. We have Vt = V0 + dk=1 0 ks dSsk . The discounted
price processes are Q-martingales by assumption, hence Vt is a local Q-martingale, as
it is a sum of stochastic integral wrt martingales. Since Vt 0, it follows that V is a
Q-supermartingale, using that every nonnegative local martingale is a supermartingale by
the Fatou-Lemma for conditional expectations.
Step 2: Suppose now that P (VT > 0) > 0. It follows that Q(VT > 0) > 0, as Q P .
Since V is a Q-supermartingale we get V E Q (V ), and this is strictly positive as
0
Q(VT > 0) > 0. Hence also V0 > 0, so that an arbitrage opportunity cannot exist.
Remark 5.8. The converse statement (the fact that absence of arbitrage implies the
existence of equivalent martingale measures) is in principle correct as well. This is the
famous first fundamental theorem of arbitrage, established in full generality by Delbaen
& Schachermayer (1994), see for instance [ ]. A good discussion of the more technical
aspects surrounding this result is also given in Chapter 10 of Bjork (2004).
44
5.1.3
Hedging and risk-neutral pricing of contingent claims
A contingent claim X is an FT -measurable random variable, to be interpreted as payoff

of some financial claims. In this section we introduce basic concepts related to hedging
and pricing of contingent claim. We assume throughout that the model admits some
martingale measure related to the numeraire S 0 , and denote a fixed such measure by Q.
Definition 5.9. A selffinancing trading strategy is called Q-admissible, if the discounted
is a Q-martingale; (Q) denotes the linear space of all Q-admissible
gains process G
strategies.
From now on we restrict our attention to contingent claims X such that E Q (|X|/ST0 ) <
:= X/S 0 L1 (, FT , Q)).
(i.e. X
T
L1 (, FT , Q) is called atDefinition 5.10. (i) A contingent claim X such that X

tainable, if there is at least one admissible strategy (Q) such that VT = X, Q
a.s.; any such strategy is called replicating strategy for X.
0
(ii) The financial market model is called complete, if any contingent claim X with XS
T
L1 (, FT , Q) is attainable.
The financial motivation for the definition is of course the fact that by initially investing
V0 and then following the trading strategy the claim can be replicated without any
further cost or risk. Note that in reality the performance of the trading strategy may be
sub-optimal due to market frictions or model errors, and perfect replication of the claim
may be impossible.
The next result links attainability to martingale representation.
= X/S 0 L1 (, FT , Q).
Lemma 5.11. Consider a contingent claim X such that X
T
t ) admits a
Then X is attainable if and only if the Q-martingale M with Mt = E Q (X|F
representation as stochastic integral of the form
Mt = x +
d Z
X
i=1
iu dSui ,
0 t T,
(5.6)
for some constant x and some Q-admissible strategy 1 , . . . , d .

Proof. Suppose that (5.6) holds. We define the replicating strategy by putting V0 =
xS 0 (0), and by using 1t , . . . , dt as position in S 1 , . . . , S d . The discounted wealth process
of the strategy satisfies (by Corollary 5.3 (ii))
VT = x +
d Z
X
i=1
is dSsi =
X
,
ST0
where the last equality follows from (5.6). Again by Corollary 5.3, we thus have X =
V0 + GT , and the if part follows. The converse statement is true by Corollary 5.3.
Market completeness can be characterized by the uniqueness of an equivalent martingale
measure.
Theorem 5.12 (Second fundamental theorem of asset pricing). Consider a market that
admits (at least) one martingale measure Q. Then the market is complete if and only if
the measure Q is unique.
45
The if part will be proven below; for the more difficult only if part we refer to the
literature.
Now we turn to the risk-neutral pricing of contingent claims. Consider an arbitrage-free
market and a contingent claim X. Denote by (t; X), 0 t T a candidate price process
for X.
Proposition 5.13.
1. If the extended model with price processes S 0 , S 1 , . . . S d , (, X)
admits an equivalent martingale measure and is thus arbitrage-free, we must have
X/ST0 L1 (, FT , Q) and
(t, X) = St0 E Q
X
0 |Ft
ST
(5.7)
for some martingale measure Q for the original market (S 0 , . . . , S d ).

2. Two different martingale measure Q1 , Q2 will in general lead to two different price
processes 1 (t, X), 2 (t, X) which are both consistent with absence of arbitrage.
3. If X is attainable, we have for any martingale measure Q with X/ST0 L1 (, FT , Q)
the relation
X

St0 E Q 0 |Ft = Vt ,
ST
where is a Q-admissible replicating strategy for X. In particular, the fair arbitragefree price of an attainable claim is uniquely defined.
Proof. 1) (5.7) is necessary to ensure that there is some martingale measure for the extended market.
2) We have seen examples for this in the discrete-time setup.
for some Q-admissible
3) If X is attainable we have the representation X/ST0 = x + G
T
is a martingale
strategy . Taking conditional expectation gives, as G
E(
X
= V ,
|Ft ) = x + G
t
t
ST0
(5.8)
where the last equality follows form Corollary 5.3. Multiplying both side of (5.8) with St0
proves the claim.
Remark 5.14. If the market is complete, we must have for any two martingales Q1 and
Q2 and any bounded random variable X that
E Q1 (X) = V (0) = E Q2 (X),
= XS 0 (which exists by market completeness).
where is a replicating strategy for X
T
Hence we must have Q1 = Q2 , which is the if part of Theorem 5.12.
5.1.4
Change of numeraire
We now discuss the technique of change of numeraire, which will be very useful for the
valuation of claims under stochastic interest rates in the next section. We consider a
46
model with d + 1 assets S 0 , S 1 , . . . , S d and assume that there is a martingale measure Q

such that the discounted asset prices Sti = Sti /St0 , 1 i d are Q-martingales. Consider
now a new asset Xt > 0. The following theorem shows which martingale measure we have
to use in order to work with X as new numeraire.
Theorem 5.15 (Change of numeraire). Consider a non-dividend-paying numeraire Xt
such that Xt /St0 is a true Q-martingale. Define a measure QX by putting
dQX
Xt S 0
|Ft = t := 0 0 .
dQ
St X0
(5.9)
Then the processes Sti /Xt , 1 i d (the basic security process discounted using X) are
QX -martingales. Moreover, we have for every contingent claim H such that H/ST0
L1 (, FT , Q) the change-of-numeraire formula
X
St0 E Q (H/ST0 |Ft ) = Xt E Q (H/XT |Ft ).
(5.10)
We mention that the theorem remains true for a general change of numeraire where one
starts with an arbitrary numeraire X and goes to a new numeraire X.

Proof. First we note that t is in fact a martingale with mean one: we have
X

S0
S00 Xt
T
E Q (T |Ft ) = 0 E Q
|F
=
= t ;
t
X0
X0 St0
ST0
moreover, E Q (T ) = 0 = 1. Hence the measure QX is well-defined. Consider now an
t := Ht /S 0 is a Q-martingale; Ht could
arbitrary stochastic process (Ht )0tT such that H
t
for instance be the price process of one of the traded securities. We get from the Bayes
formula (Lemma 4.24) that
H
1
H
X S0
H S0X

X
i
0
T
T
T 0 T
EQ
|Ft = E Q
T |Ft = 0 t E Q
|F
t
XT
t
XT
S0 Xt
XT X0 ST0
H

S0
T
= t EQ
|F
t
Xt
ST0
ii Ht
= ,
Xt
t is a Q-martingale.
where we have used i the definition of in (5.9) and ii the fact that H
X
The above computation shows that Ht /Xt is a Q -martingale and thus proves the first
part of the theorem. Moreover, we have
H

H

Ht
Ht
X
T
= Xt E Q
|Ft ,
St0 E Q 0 |Ft = St0 0 = Ht = Xt
Xt
XT
ST
St
which proves (5.10).
5.2
The Black-Scholes Model Revisited
Setup. As in the previous section we consider a model with two traded assets, stock
and money market account, with dynamics given by
dSt0 = rSt0 dt,
S00 = 1
dSt1
St1 dWt ,
St1 dt
47
(5.11)
S01
= S0 > 0.
(5.12)
Note that (5.11) implies that St0 = exp(rt). We define the discount factor by Dt =
(St0 )1 = exp(rt). The discounted stock price is given by St := Dt St1 ; the discounted
price of S 0 is by definition equal to 1. In economic terms discounting corresponds to taking
the money market account S 0 as new numeraire. We now use Itos-formula to compute the
dynamics of the discounted stock price. Since dDt = r(Dt )dt, the Ito-product-formula
yields
dSt1 = St1 dDt + Dt dSt1 + dhSt1 , Dt it = rSt1 dt + St1 dt + St1 dWt
= ( r)St1 dt + St1 dWt ,
where we have used that hS 1 , Dit 0 as (Dt ) is of finite variation.
Equivalent Martingale Measures in the Black-Scholes Model. Recall from
Lemma 5.6 that the discounted price Rprocess S1 is a Q-martingale if and only if S 1 itself
t
has dynamics of the form St1 = S0 + 0 rSs1 ds + MtQ , where M Q is a Q-local martingale.
Comparison with the Black-Scholes SDE shows that we need to define Q in such a way
that by going from P to Q the drift is changed from to r .
Remark. If {Ft } is generated by the Brownian motion W and if Q P it can be shown
that the density Q
P is necessarily of the form
Z T

Z T
dQ
2
= exp
s dWs 1/2
s ds
dP
0
0
for some adapted process . This means that in a model based on Brownian motion like
the Black-Scholes model changing the measure is equivalent to changing the drift.
Now we apply Girsanovs theorem to determine the equivalent martingale measure in the
Black-Scholes model. We have
Z t
Z t
1
1
St = S0 +
Ss ds +
Ss1 dWs
0
0

Z t
Z t
Z t
r
1
1
rSs ds +
= S0 +
Ss ds +
Ss1 dWs
0
0
0
Z t
Z t
s
= S0 +
rS 1 ds +
S 1 dW
s
R
t = Wt + t r ds. Define := r ; is often referred to as market price of risk
where W
0
Rt
t we need to turn
of the stock. In order for S to be of the form St1 = S0 + 0 rSs1 ds + M
Rt
t = Wt + ds into a martingale by a change of measure. Girsanovs theorem tells us
W
0
that under Q with
dQ
= exp(WT 1/22 T )
dP
R
is a Q-Brownian motion, hence the Ito-integral M
t = t Ss1 dW
s is a
the process W
0
Q-local martingale. Summing up we have
Proposition 5.16. Consider a Black-Scholes model with stock-price dynamics of the form
dSt1 = St1 dWt + St1 dt for constants , . Then the equivalent martingale measure Q
r
2
is given by the density dQ
dP = GT := exp(WT 1/2 T ), where = is the market
price of risk of the stock. Under Q the stock-price process solves the SDE
t
dSt1 = rSt1 dt + St1 dW
48
(5.13)
Black-Scholes via risk-neutral valuation. We now derive the Black-Scholes formula

from the risk-neutral valuation principle; this is the easiest approach from a computational
perspective. We know that St1 solves equation (5.13) and is thus of the form St1 =
t + (r 1/2 2 )t) . From the risk-neutral pricing-rule which applies as we
S0 exp( W
already know from Chapter 3 that the call can be replicated we get for C0 , the call price
at time zero
C0 = E Q (erT (ST1 K)+ ) = E Q (erT ST1 1(S 1 >K) ) erT K Q(ST1 > K) .
|
{z
}
|
{z T
}
(II)
(I)
We start with the second term. One has

Q(ST1
> K) =
Q(ln ST1
ln K ln S0 (r 1/2 2 )T
W
T >
T
T
> ln K) = Q
!
.
T N (0, 1). By the symmetry of centered normal distributions

Now note that Z = W
T
we have for M R that Q(Z > M ) = Q(Z < M ). Hence

ln S0 /K + (r 1/2 2 )T
1
= N (d2 ) .
Q(ST > K) = Q Z <
T
To deal with term (I) we need to apply Girsanovs theorem once more. Note that erT ST1 =
T 1/2 2 T ) has the form of a Girsanov-density. Define a new
S0 YT where YT := exp( W
measure QS by dQS /dQ = YT . Then we have

S
E Q erT ST1 1{S 1 >K} = S0 E Q YT 1{S 1 >K} = S0 E Q (1{S 1 >K} ) = S0 QS (ln ST1 > ln K) .
T
t t is a QS -Brownian motion. Hence W

T =
Girsanovs theorem implies that WtS := W
2
S
WT + T and
QS (ln St1 > ln K) = QS (ln S0 + (r 1/2 2 )T + WTS + 2 T > ln K) .
The right hand side is now evaluated in exactly the same way as before, yielding

E Q erT ST1 1(S 1 >K) = S0 N (d1 ) .
T
5.3
5.3.1
Fundamental Theorems of Asset Pricing in a Generalized Black-Scholes Model

A multidimensional Black-Scholes model
The traded assets consist of one money-market account S 0 and d stocks S 1 , . . . , S d . Given
a filtrated probability space (, F, P ), {Ft }, 0 t T , we assume that the stock-price is
an Ito-process of the form
dSti
ti Sti dt
Sti
n
X
ij (t)dWtj ,
j=1
49
1 i d,
(5.14)
for an n-dimensional standard Brownian motion Wt = (Wt1 , . . . , Wtn ) and adapted process
ti , 1 i d and ij (t), 1 i d, 1 j n. The initial values are given by
S0 = (S01 , . . . , S0d ). The money-market account satisfies
St0
Z t
= exp( rs ds)
0
for some adapted process (rs )0st modelling the (random) short-rate of the interest. We
now give a more intuitive description of the model (5.14). Define
i (t) =
n
X
ij (t)2
1
(5.15)
j=1
and assume that i (t) > 0 a.s.. Define processes Bti , 1 i d, via
Bti =
n Z
X
j=1
ij (s)
dWsj .
i (s)
(5.16)
It is easy seen that hB i it = t (exercise!), so that B 1 , . . . B d are one-dimensional Brownian

motions by the Levy-characterization of Brownian motion. Moreover, one has for i 6= j
that
Pn
Z t
il (t)jl (t)
i
j
hB , B it =
ij (s)ds with ij (s) = l=1
.
i (t)j (t)
0
Note that ij (t) [1, 1]; it is termed instantaneous correlation of B i and B j . In terms
of the B i the asset price dynamics can be written in the form
dSti = ti Sti dt + i (t)Sti dBti ,
1 i d,
(5.17)
so that i (t) is the volatility of S i at time t. Moreover, we get from Itos formula
1
d ln Sti = (ti i2 (t))dt + i (t)dBti ,
2
so that
i
hln S , ln S it =
t
i
i (s)j (s)dhB , B is =
0
i (s)j (s)ij (s)ds.

0
The quantity ij therefore models the instantaneous correlation of the logarithmic return
processes: for h small we have

j
i
E (ln St+h
ln Sti )(ln St+h
ln Stj )|Ft (i j ij )h.
As in the case of the one-dimensional Black-Scholes model the dynamics of the discounted
stock prices Sti = Sti /St0 are given by
d
X
i
i
i
i
i
i
i
i
dSt =(t rt )St dt + i (t)St dBt = (t rt )St dt + St d(

ij (t)dWtj ).
j=1
50
(5.18)
5.3.2
Existence of an equivalent martingale measure
The following Proposition gives sufficient conditions for the existence of an equivalent
martingale measure, and hence for the absence of arbitrage in the model (5.14).
Proposition 5.17. Suppose that for all 0 t T there is a solution t = (t1 , . . . , tn ) of
the so-called market-price-of-risk equations
ti rt =
n
X
ij (t)tj ,
1 i d,
(5.19)
j=1
and that E exp( 12

tingale measure.
RT
0

ks k2 ds) < . Then the model (5.14) admits an equivalent mar-
Proof. Using the process t from (5.19) we can write the SDE (5.18) for the discounted
stock price S1 , . . . Sd in the form in the form
Z t
n
n

X
X
j
j
i
i
i
i
i
dSt =(t rt )St dt + St

ij (t)dWt = St
sj ds
(5.20)
ij (t)d Wt +
j=1
j=1
R
into
j = W j + t sj ds, 1 j n. We now use the Girsanov theorem to turn W
Define W
t
t
0
a new Brownian motion. Define
Z T
n Z T
X
ZT := exp(
sj dWsj
ks k2 ds).
(5.21)
j=1
The integrability condition on ensures E(ZT ) = 1. Define a new measure Q by dQ

dP |FT =
with is
ZT . The multidimensional Girsanov-theorem shows that under Q the process W
Q-Brownian motion, and the result follow from Lemma 5.6.
Denote by (t) the matrix (ij (t))1id,1jn and by 1 the vector (1, . . . , 1)0 Rd . Suppose that the market price of risk equation does not admit a solution. This implies that
the vector (t rt 1) does not belong to Im(t) (the range of the matrix (t)) and hence
we get for ker 0 (t) = {x Rd : 0 (t)x = 0}

ker 0 (t) = Im(t) 6= {0}.
Hence we can find t Rd such that
d
X
ti (ti rt ) 6= 0,
0 (t)t = 0,
and
i=1
otherwise (t rt 1) ker 0 (t) = Im(t), which contradicts the fact that the marketprice-of-risk equation has no solution. Define a trading-strategy by it = ti /Sti . The
corresponding discounted wealth process V satisfies
dVt =
d
X
ti (ti rt )dt +
i=1
n X
d
d
X
X
(
ij (t)ti )dWtj =
ti (ti rt )dt,
j=1 i=1
Pd
i=1
i
0
0
where we have used that
i=1 ij (t)t = (t)t j = 0 as t ker (t). Hence we
have constructed a locally riskless selffinancing trading strategy whose discounted wealthprocess is non-constant, and it is easy to construct arbitrage-opportunities from this.
Hence we have shown the following result:
51
Theorem 5.18. Modulo integrability conditions, the model (5.14) is arbitrage-free if and
only if the market-price of risk equation (5.19) admits a solution.
5.3.3
Market completeness in the generalized Black-Scholes model
In order to discuss market completeness we need the following martingale representation

theorem for Brownian motion.
Theorem 5.19 (Ito-representation theorem for Brownian motion). Consider a d-dimensional
Brownian motion W = (Wt,1 , . . . , Wt,d )0 on some filtered probability space (, F, P ), {Ft },
0 t T and denote by {FtW } the filtration generated by W . Then every {FtW }-adapted
square integrable martingale M can be written as a stochastic integral with respect to
Brownian
motion,
that is there is a d-dimensional adapted process ht = (h1t , . . . , hdt ) with
R

T
E 0 (his )2 ds < for all i such that
Mt = M0 +
d Z
X
i=1
his dWs,i ,
0 t T.
As before, we consider a model with d + 1 traded assets and dynamics of the form dSt0 =
rt St0 dt, and
dSti = ti Sti dt +
n
X
ij (t)Sti dWtj ,
, 1 i d,
(5.22)
j=1
for an n-dimensional Brownian motion Wt = (Wt1 , . . . , Wtn ). Moreover, we assume that

our filtration is generated by the Brownian motion, that is Ft = (Wsi : s t, 1 i n);
this assumption restricts the contingent claims we may consider and is thus crucial for
establishing completeness of the market.
Proposition 5.20. Suppose that the model (5.22) is free of arbitrage, and that the underlying filtration is generated by the Brownian motion W . Then the market is complete if and only if P .a.s. Im 0 (t) = Rn , 0 t T , where (t) represents the matrix
ij (t) 1id,1jn .
Since (t) is a d n-matrix a necessary condition for completeness is thus the inequality
d n.
Proof. Denote by Q some equivalent martingale measure, which exists as the market is
assumed to be arbitrage-free. As shown before, under Q the discounted state prices have
dynamics
dSti = Sti
n
X
j,
ij (t)dW
t
(5.23)
j=1
j is a Q-Brownian motion. By Lemma 5.11, the statement of the proposition in

for W
equivalent to the following statement:
d Z t
X
Every Q-martingale M has a representationMt = M0 +
is dSsi , if and only if Im 0 (t) = Rd for all t.
i=1
(5.24)
52
We now turn to (5.24). Using It

os representation theorem (Theorem 5.19) and the assumption {Ft } = {FtW } it can be shown that M has a unique representation of the
form
n Z t
X
sj .
Mt = M0 +
hjs dW
(5.25)
j=1
On the other hand, writing (5.24) in short-hand notation and using (5.23) we get that the
martingale representation of M in (5.24) is equivalent to
dMt =
d
X
Sti it
i=1
n
X
j=
ij (t)dW
t
j=1
n X
d
X
j.
Sti ij (t)it dW
t
(5.26)
j=1 i=1
Comparing (5.25) and (5.26) we see that we must have

hjt =
d
X
ij (t)Sti it =
i=1
d
X
ij (t)it ,
(5.27)
i=1
where it = it Sti . This equation is solvable for all h if and only if the range of 0 (t) is the
entire space Rn .
53
Chapter 6
Term-Structure Modelling
In this chapter we give an introduction to the modelling of interest-rate risk. Our presentation is largely based on Bj
ork (2004); further information can be found in the recent
textbook Filipovic (2009).
6.1
Bonds and Interest-Rates
We begin by describing the market for interest-rate related products at a given point in
time t.
Zero coupon bond. A zero coupon bond with maturity T > t, also called T -bond
guarantees the holder e1 to be paid at the maturity date T . The price in t < T is
denoted by p(t, T ). Since we assume that bonds do not default it holds that p(T, T ) = 1.
The family p(t, T ), T t describes the term-structure of interest rates at time t. We
assume that
There is a frictionless market for T -bonds for all T > t
For fixed t the mapping T 7 p(t, T ) is continuously differentiable; we write pT (t, T ) =
T p(t, T ).
Note that the mapping t 7 p(t, T ) (the price trajectory of a bond with fixed maturity
date) is usually quite irregular with sample path properties similar to those of Brownian
motion.
Forward price of bonds. Consider time points t < S < T . The price, contracted in
t, to be paid in S for e1 at time T is the forward price of the T -bond with maturity S,
denoted by F (t, S, T ). It is well-known that F (t, S, T ) = p(t, T )/p(t, S). Alternatively, if
we make a contract at t to invest e1 over the time period [S, T ] we obtain an amount of
1/F (t, S, T ) = p(t, S)/p(t, T ) at T .
From this relation we can define various interest rates.
Definition 6.1. (Interest rates, continuous compounding) Let t < S < T .
54
1. The (continuously compounded) forward rate over (S, T ], contracted at t, R(t; S, T )

is defined by the equation
exp((T S)R(t; S, T )) =
p(t, S)
,
p(t, T )
i.e. as continuously compounded return on the forward investment. One has

R(t; S, T ) =
ln p(t, T ) ln p(t, S)
.
T S
2. The (continuously compounded) spot rate or yield with maturity T , contracted at

t is defined to be
ln p(t, T )
R(t, T ) := R(t; t, T ) =
.
T t
3. The instantaneous forward rate with maturity T , contracted at t, is defined by
f (t, T ) := lim R(t; S, T ) = T ln p(t, T ).
ST
4. The (instantaneous) short-rate of interest is r(t) = f (t, t), i.e. r(t) = T ln p(t, T )|T =t .
Rt
As usual we define the money market account account process by Bt = exp( 0 rs ds), so
that B has dynamics dBt = rt Bt dt with initial value B0 = 1.
Relation between f (t, ) and p(t, ). We get from the fundamental theorem of calculus
Z T
Z T
ln p(t, T ) = ln p(t, S) +
T ln p(t, u)du = ln p(t, S)
f (t, u)du.
S

R
T
and therefore p(t, T ) = p(t, S) exp S f (t, u)du . In particular
Z
p(t, T ) = exp(
f (t, u)du),
(6.1)
as p(t, t) = 1. This shows that there is a one-to-one relation between the family p(t, ) of
bond prices and the family f (t, ) of instantaneous forward rates.
Dynamic Modelling. Term-structure modelling is the task of constructing a probabilistic model for the evolution of the family of zero coupon bond price processes p(, T ).
There are a number of requirements for such a model.
p(, T ) needs to satisfy the final condition p(T, T ) = 1.
Absence of arbitrage between the security prices {p(, T ), T t} (a challenge as we
are dealing with many price processes simultaneously).
Calibration. At the initial date t0 the price p(t0 , T ) has to be consistent with the
bond prices or interest rates observed in the market at time t0 ).
The short rate r(t) is a nominal interest rate and should be nonnegative (but this
requirement is sometimes relaxed).
In these lecture notes we consider two approaches for term-structure modelling, so-called
short-rate models and forward-rate- or HJM models.
55
6.2
6.2.1
Short-Rate Models
Martingale Modelling and Term-Structure Equation
Consider a filtered probability space (, F, Q), (Ft ) supporting an adapted process (rt )t0
which models the instantaneous short rate of interest. We view Q as risk-neutral measure
and define the price of a T -bond at t T by the risk-neutral pricing rule, that is we put

RT
1
t rs ds
(6.2)
=
E
e
p(t, T ) = Bt E
F
Ft .
t
BT
The approach of modelling the dynamics of traded securities directly under some martingale measure (and not under the historical measure P ) is termed martingale modelling; it
is motivated by the observation that at least in a complete market only the risk-neutral
measure Q matters for the pricing of derivative securities. Martingale modelling has
become very popular. The approach has the following advantages:
The approach ensures that the model is arbitrage-free.
Bond prices are automatically consistent with the terminal condition p(T, T ) = 1.
Disadvantages/ problems:
Bond prices defined via (6.2) are not automatically consistent with the prices observed in the market at time t; for this one needs to adjust parameters in the
dynamics of the short rate process (calibration), which can be difficult from a computational viewpoint.
The approach has conceptual difficulties in incomplete markets (but that is less
relevant in the context of term-structure models).
Assumption 6.2. We assume that under the fixed martingale measure Q the short-rate
has dynamics of the form drt = (t, rt )dt + (t, rt )dWt for a Q-Brownian motion W .
Under Assumption 6.2 the short-rate is a Markov process, and we have for p(t, T ) as
defined in (6.2)

Z
p(t, T ) = E exp
t

Z

rs ds Ft = Et,rt exp

rs ds
,
(6.3)
and the right side is obviously a function of t and rt which we denote by F T (t, rt ).
Proposition 6.3 (Term-structure equation.). The function F T solves the parabolic PDE
1
t F + r F + 2 rr F = rF, t < T
2
(6.4)
with terminal condition F T (T, r) = 1.

Proof. The result follows immediately from the Feynman-Kac formula, see equation (4.19).
56
Examples for short-rate dynamics. The following short-rate models are frequently
used in the literature.
Vasicek-model drt = (b art )dt + dWt , a, b > 0.
CIR-model drt = (b art )dt + rt dWt .

Hull-White drt = ((t) art )dt + dWt .
Extended CIR: drt = ((t) art )dt + rt dWt .

Black-Derman-Toy drt = a(t)rt dt + rt dWt .
6.2.2
Affine Term Structure
Next we look for conditions such that the term structure equation is easy to solve. The
main result in this direction is the existence of an affine term structure.
Definition 6.4. A short-rate model has an affine term structure (ATS), if the bond-prices
are of the form
p(t, T ) = F T (t, rt ) = exp(A(t, T ) B(t, T )rt )
(6.5)
for deterministic functions A(t, T ) and B(t, T ), 0 t T .

Note that in a model with an ATS the continuously compounded yield R(t, T ) = T 1t ln p(t, T )
is an affine function of rt ; this explains the name. It is easy to give sufficient conditions
for the existence of an ATS.
Proposition 6.5. Assume that and 2 are of the form
(t, r) = (t)r + (t), 2 (t, r) = (t)r + (t)
(6.6)
Then the model has an ATS (6.5) where A and B satisfy the ODE-system
1
Bt (t, T ) + (t)B(t, T ) (t)B 2 (t, T ) = 1
2
1
At (t, T ) (t)B(t, T ) + (t)B 2 (t, T ) = 0,
2
(6.7)
with terminal conditions B(T, T ) = A(T, T ) = 0.

Proof. Assume that F T is of the form (6.5). This gives, ignoring the arguments of a and
b,
t F T = (At Bt r)F T ,
r F T = BF T ,
rr F T = B 2 F T .
Plugging these terms into the term-structure equation gives after division by F T
1
At Bt r B + 2 B 2 = r.
2
57
(6.8)
If and 2 are affine in r, we get by substituting (6.7) into (6.8)

1
(At B + B 2 ) + r(Bt + 1 + B 2 ) = 0.
2
This equation holds for all values of r if and only if both brackets are zero, which gives
(6.7). The terminal condition on A and B ensures that F T (T, T ) = 1.
Example 6.6. (Term-structure equation in the Vasicek-model.) Here the ODE-system
(6.7) becomes, as = a, = b, = 0, = 2 ,
1
Bt aB = 1, At = bB 2 B 2 .
2
Since the ODE for B is linear, one has, as B(T, T ) = 0,
1
(1 exp (a(T t))),
a
RT
2 RT
and, by integration, A(t, T ) = t bB(s, T )ds + 2 t B 2 (s, T )ds .
B(t, T ) =
6.2.3
Calibration
All short-rate models considered so-far have dynamics which depend on a number of
unknown parameters. In order to apply the model these parameters need to be determined.
Here two different approaches can be distinguished.
a) Statistical estimation from a time series (rs )st of the past values of the short-rate.
Here we face the following conceptual problem: in order to compute the bond prices we
use the Q-dynamics of the short-rate, whereas the statistical estimation gives a estimate
of the parameters under P , and the drift of the short-rate may change in the transition
from P to Q.
b) Calibration to market prices. Schematically, the approach is as follows: Denote current
time by t0 .
Fix a concrete short-rate model, say, the CIR model with parameter vector =
(b, a, ).
Solve the term-structure equation for all maturities T , denote the solution by F T (t0 , rt0 ; ).
Go to the market and obtain observed prices {p (t0 , T1 ), . . . , p (t0 , Tm )} for maturities T1 , . . . , Tm and obtain moreover an estimate of the current short-rate rt0 .
Determine as solution of the minimization problem
min
m
X
wi p (t0 , Ti ) F Ti (t0 , rt0 ; )
2
(6.9)
i=1
where w1 , . . . , wn is a vector of weights.

This approach is widely adapted in practice, as it is less subjective and often easier than
statistical estimation. Moreover the approach (approximately) aligns model and market
prices which is important for the use of market-consistent valuation methods. Nonetheless
there are a number of problems as-well.
58
On the practical side, the model may not be flexible enough so that model- and
market prices diverge widely even for optimal .
With very little price observation there may be many solutions to (6.9), i.e. can
not be determined from observable prices.
The parameter vector may fluctuate a lot over time, whereas in pricing it is
assumed that this parameter is constant.
Time series properties of the short rate (or of observed bond prices) are ignored
completely.
6.3
HJM-Models
Basic approach. In the HJM-approach to term structure modelling one models simultaneously the dynamics of all forward-rates f (t, T ), T t and computes bond-price
dynamics from the relation
Z T

f (t, u)du
p(t, T ) = exp
(6.10)
t
(recall that f (t, T ) = T ln p(t, T )). In this way the calibration to the observed bond
prices is ensured by design, as the current forward-rate curve f (0, T ) is taken as initial
value of the forward-rate dynamics; moreover (6.10) ensures that p(T, T ) = exp(0) = 1.
On the other hand, it is a priori not clear if the model generated in this way is actually
arbitrage-free, as one models infinitely many securities (all bond prices), but has typically
only finitely many sources of randomness in the model; this issue is taken up below.
Assumption 6.7. (Forward-rate dynamics) Given a probability space (, F, P ), (Ft ), P
the historical measure, let W be a d-dimensional Brownian motion on this space. Then
for each fixed T the forward rate f (t, T ) has a stochastic differential of the form
df (t, T ) = (t, T )dt + (t, T )dWt
(6.11)
where
(, T ) and (, T ) are adaptedR process with values in R respective Rd , and where
RT
Pd
t
i=1 0 i (t, T )dWt,i . Moreover and are continuously
0 (t, T )dWt is shorthand for
differentiable in the T -variable.
In order to give conditions on (t, T ) and (t, T ) in (6.11) ensuring that the model is free
of arbitrage we need to compute the dynamics of the bond prices using (6.10).
Proposition 6.8 (Short-rate dynamics and bond price dynamics in HJM). If the family
of forward rates satisfies Assumption 6.7, the short rate r(t) = f (t, t) has the differential
drt = a(t)dt + b(t)dWt , with a(t) = T f (t, t) + (t, t) and b(t) = (t, t).
R

T
Moreover, p(t, T ) = exp t f (t, u)du satisfies
1
dp(t, T ) = p(t, T ){r(t) + A(t, T ) + kS(t, T )k2 }dt + p(t, T )S(t, T )dWt , with
2
Z T
Z T
A(t, T ) =
(t, u)du, and S(t, T ) =
(t, u)du
t
59
(6.12)
(6.13)
(6.14)
Proof. As r(t) = f (t, t) we get from (6.11)

Z t
Z t
(s, t)dWs .
(s, t)ds +
r(t) = f (0, t) +
(6.15)
Rt
Now write (s, t) = (s, s) + s T (s, u)du, and similarly for . Substitution into (6.15)
gives
Z t
Z t
r(t) =
(s, s)ds +
(s, s)dWs
0
0
Z tZ t
Z tZ t
T (s, u)dudWs .
(6.16)
T (s, u)duds +
+ f (0, t) +
0
Changing the order of integration we get

Z tZ
Z tZ t
Z tZ t
1{u>s} T (s, u)duds =
T (s, u)duds =
0
T (s, u)dsdu,
(6.17)
RtRt
RtRu
and similarly 0 s T (s, u)dudWs = 0 0 T (s, u)dWs du. Moreover, f (0, t) = r(0) +
Rt
0 T f (0, u)du. On the other hand one has
Z u
Z u
T f (u, u) = T f (0, u) +
T (s, u)ds +
T (s, u)dWs .
0
Hence the sum in (6.16) equals r(0) +

proposition.
Rt
0
T f (u, u)du, which proves he first part of the
In order to identify
R T the bond price dynamics (6.13) and (6.14) one argues as follows: Define
Y (t, T ) := t f (t, s)ds so that p(t, T ) = exp(Y (t, T )). Recall that under Assumption
6.7,
Z t
Z t
(u, s)dWu .
(u, s)du +
f (t, s) = f (0, s) +
0
This gives, after changing the order of integration,

Z T
Z tZ T
Z tZ
Y (t, T ) :=
f (0, s)ds
(u, s)dsdu
t
(u, s)dsdWu .
Now we can decompose these integrals into

Z T
Z tZ T
Z tZ T
f (0, s)ds
(u, s)dsdu
(u, s)dsdWu
0
0
u
0
u
Z tZ t
Z t
Z tZ t
+
f (0, s)ds +
(u, s)dsdu +
(u, s)dsdWu
0
0
u
0
u
Z t
Z t
= Y (0, T ) +
A(u, T )du +
S(u, T )dWu
0
0
Z t
Z tZ s
Z tZ s
+
f (0, s)ds +
(u, s)duds +
(u, s)dWu ds,
0
(6.18)
(6.19)
where we have used the definition ofRA and S in (6.14)

R t and a similar argument as in (6.17).
t
Now, (6.19) is obviously equal to 0 f (s, s)ds = 0 r(s)ds, and we get that dY (t, T ) =
{r(t) + A(t, T )}dt + S(t, T )dWt . The proof of (6.13) is completed by applying the Itoformula to p(t, T ) = exp(Y (t, T )).
60
Absence of arbitrage Recall that we consider a market with infinitely many assets
but only finitely many sources of uncertainty, namely W1 , . . . , Wd . Hence we need special
conditions on the forward rate dynamics in order to ensure that the model is free of
arbitrage.
Proposition 6.9. Suppose that there is a process t = (t,1 , . . . , t,d )0 such that for all
0tT
Z T
(t, T ) = (t, T )
(t, s)0 ds + (t, T )(t)
(6.20)
t
and such that Zt = exp

is free of arbitrage.
R
t
0 s dWs

2
k
k
ds
is a true martingale. Then the model
s
0
R
1 t
2
Note that Condition (6.20) is a restriction on the drift of the forward rates, as this equation
has to hold simultaneously for all maturity dates T .
Proof. According to Propostion 6.8, bond prices are of the form

1
dp(t, T ) = p(t, T )r(t)dt + p(t, T ) A(t, T ) + kS(t, T )k2 dt + p(t, T )S(t, T )dWt .
2
Define a new measure Q by dQ
dP |FT = ZT ; Q is well-defined as Z is a true martingale.
Rt
f
Moreover, Wt = Wt + 0 s ds is a Q-Brownian Motion by the Girsanov theorem. Now
note that under Q the finite-variation part of p(, T ) equals

Z t
1
2
p(s, T ) r(s) + A(s, T ) + kS(s, T )k S(s, T )s ds.
(6.21)
2
0
We get from (6.20) and the definition of S(s, T ) in (6.14) that
Z
S(s, T )s =
s
Z
=
(s, u)s du (by (6.14))

Z u
0
(s, u) (s, u)
(s, ) d du (by (6.20))
1
= A(s, T ) + kS(s, T )k2 (by (6.14)).
2
Rt
Hence the finite variation part of p(, T ) in (6.21) equals 0 p(s, T )r(s)ds, and discounted
bond prices are Q-local martingales.
Obviously the measure Q constructed in the proof of Proposition 6.9 is a risk-neutral
measure. It is interesting to study the forward-rate and price dynamics under Q
Corollary 6.10. Under Q the forward-rate dynamics are

Z T
ft
df (t, T ) = (t, T )
(t, s)0 ds dt + (t, T )dW
0
ft ,
dp(t, T ) = r(t)p(t, T )dt + p(t, T )S(t, T )dW
f a Q-Brownian motion.
W
61
This follows immediately from the proof of Proposition 6.9. The fact that the risk-neutral
drift of f (t, T ) is of the form
t
(t, T ) = (t, T )
(t, s)0 ds
is also known as HJM drift condition.

Martingale modelling in HJM The previous corollary has shown that the Q-dynamics
of bonds and forward-rates are fully determined from the forward-rate volatilies (t, T ).
For practical purposes the HJM model is therefore used as follows:
1.) Specify the volatilities (t, T ) (the modelling part)
2.) The forward-rate drift Q (t, T ) is specified via the HJM-drift condition.
3.) Go to the market and observe current forward rates f (0, T ).
4.) The forward-rates are given by
f (t, T ) = f (0, T ) +
Q (s, T )ds +
ft .
(t, T )dW
0
5.) Bond prices are given via p(t, T ) = exp{
RT
t
f (t, u)du} or as solutions of
ft ,
dp(t, T ) = p(t, T )r(t)dt + p(t, T )S(t, T )dW
for S(t, T ) =
RT
t
(t, u)du.
Example 6.11 (The Ho-Lee model).

R T Assume that (t, T ) = . Then the drift of the
forward-rates equals Q (t, T ) = t ds = 2 (T t) and the forward rates are given by
2 (T s)ds + Wt
0

t
2
= f (0, T ) + t T
+ Wt .
2
f (t, T ) = f (0, T ) +
In particular, the short-rate satisfies

r(t) = f (t, t) = f (0, t) + 2
t2
+ Wt ,
2

or, in differential form, drt = T f (0, t) + 2 t dt + dWt . Short-rate dynamics of this
form are known as Ho-Lee model (the Vasicek or Hull-White model for a = 0). Note that
the HJM-approach automatically ensures that the model is calibrated to the initial yield
curve.
Example 6.12 (The Vasicek-model). Here we take (t, T ) = ea(T t) for some a > 0.
Hence we get
Q
(t, T ) = (t, T )
(t, s)ds =
t
62

2 a(T t)
e
1 ea(T t) .
a
In order to identify the corresponding short-rate dynamics we use Proposition 6.8. The
volatility of the short-rate dynamics is given by (t, t) = . Identifying the drift (t, r)
is more involved. According to Proposition 6.8, we have (t, rt ) = Q (t, t) + T f (t, t) =
T f (t, t), as Q (t, t) = 0. Now we have, using the forward rate dynamics
Z
T f (t, t) = T f (0, T ) +
Z
T (s, t)ds +
T (s, t)dWs .
0
As (s, t) = ea(ts) , we get T (s, t) = a(s, t). Moreover, T Q (t, s) = aQ (t, s) +

2 e2a(ts) . This gives
Z t
Z t
Z t
T f (t, t) = T f (0, t) a
(s, t)ds a
(s, t)dWs +
2 e2a(ts)ds
0
= af (t, t) + g(t)
with g(t) = T f (0, t)af (0, t)+
short-rate dynamics
Rt
0
2 e2a(ts) ds. Summarizing we thus have the following
dr(t) = (g(t) ar(t))dt + dWt ,

so that the above form of leads to the Vasicek or the Hull-White model.
6.3.1
The Musiela parametrization.
still to do
6.4
Pricing and Hedging of Multi-Currency Derivatives with

Interest Rate Risk
This section, which is based on Frey & Sommer (1996), discusses the valuation and hedging of non path-dependent European options on one or several underlyings in a model
of an international economy which allows for both, interest rate risk and exchange rate
risk. We study options on stocks, bonds, future contracts, interest rates and exchange
rates; their payoff may be in any currency and a relatively complex function of one or
several underlyings. We use an international economy model similar to the one introduced
by Amin & Jarrow (1991) as framework of our analysis. Their model combines a fully
developed stochastic theory of the term structure of interest rates in the sense of Heath,
Jarrow & Morton (1992) with models for the valuation of exchange rate and stock options. The main tools of our analysis are stochastic methods and in particular the change
of numeraire technique introduced in Section 5.1.4. Since pricing formulas are of limited
practical use without knowledge of the corresponding hedge portfolio we present a systematic approach to computing hedging strategies. In order to illustrate the flexibility of
our method we derive explicit formulas for prices and hedge portfolios for a wide range of
examples containing among others currency options or guaranteed-exchange-rate options
or options on interest rates.
63
6.4.1
The Model
Here we introduce an arbitrage-free model of an international economy that incorporates

stochastic interest rates and exchange rates. This model will serve as our framework
for the valuation of derivatives. We consider N countries indexed by n {0, . . . , N }.
Country 0 will be the domestic country. The exchange rate between country 0 and country
n {1, . . . , N } will be denoted by en , that is en units of the domestic currency can be
exchanged for one unit of the foreign currency. When working with only two countries
we simply talk about the domestic and the foreign country and index them with d and
f . The choice of the domestic country is arbitrary and depends on the particular pricing
and hedging problem under consideration. We call an asset a domestic asset if its payoffs
are denominated in the domestic currency. Notice that every asset whose payoffs are
not originally denominated in this currency can be transformed into a domestic asset by
translating its payoffs into the domestic currency using the corresponding exchange rate.
We assume that in all countries zero coupon bonds of all maturities T [0, TF ] are traded.
The zero coupon bond in country n with maturity date T shall be denoted by pn (t, T ) for
t [0, T ]. By assumption pn (T, T ) 1 T, n. The short rate in country n, rn , is given by

rtn =
ln pn (t, T ) .
(6.22)

T T = t
R
n := exp( T r n ds) we denote the savings-account of country n. Apart from zero
By t,T
s
t
coupon bonds we consider other primitive assets such as dividend free stocks. They are
denoted by S n,j , 0 n N, 0 j in where S n,j is the price of asset j in country n.
We now introduce our model of asset price dynamics. When modelling asset price processes one usually starts from assumptions on their dynamics under the so-called historical
probabilities which govern the actual evolution of asset prices. Since we are only interested in the pricing of derivatives by no-arbitrage arguments it is legitimate to model the
asset price dynamics directly under a domestic risk-neutral measure Q. Under such a
measure all non-dividend paying domestic assets are martingales after discounting with
the domestic savings account. This implies that their drift is equal to r0 .
Assumption 6.13. Let there be given a filtered probability space (, F, Q), (Ft )t[0,TF ]
supporting a d-dimensional Brownian Motion W = (Wt )0tTf . We work with the following assumptions on asset price dynamics: We put for the domestic assets
dp0 (t, T ) = rt0 p0 (t, T )dt + 0 (t, T )p0 (t, T )dWt
dSt0,j
= rt0 St0,j dt + 0,j (t)St0,j dWt ,
(6.23)
and for the foreign assets

n
dpn (t, T ) = (rtn n (t, T ) e (t))pn (t, T )dt + n (t, T )pn (t, T )dWt
dStn,j
= (rtn n,j (t) e (t))Stn,j dt + n,j (t)Stn,j dWt .
(6.24)
Finally the dynamics of the exchange rates are given by

n
den (t) = (rt0 rtn )en (t)dt + e (t)en (t)dWt .

n
(6.25)
Here n (t, T ), n,j (t), e (t) : [0, TF ] Rd are deterministic square integrable functions
64
of time. For the bonds we require moreover that n (t, T ) = 0 t T and that n (t, T ) is
smooth in the second argument.
Note that the assumption of deterministic dispersion coefficients is essential as we want to
obtain explicit pricing formulas. An explicit construction of the bond price model is easily
done using the HJM approach explained in the previous section. As shown by Amin &
Jarrow (1991) the dynamics of asset prices and exchange rates given in Assumption 6.13
actually specify an arbitrage-free model of an international economy with Q representing
a domestic risk-neutral measure. The drift terms of the exchange rate and the foreign
assets are determined by absence of arbitrage considerations. As an example we derive
n
the drift of en . Consider the domestic asset Y := en (0,)
. By absence of arbitrage its
0
drift must equal r . Using It
os Lemma to compute the dynamics of Y it is immediate
that the drift of Y equals r0 if and only if the drift of the exchange rate equals the interest
rate differential.
The volatility of asset S n,j is given by n,j (t) := | n,j (t)|. The instantaneous correlations
between the assets in our economy are given by
(S n1 ,j1 , S n2 ,j2 ) :=
n1 ,j1 n2 ,j2 1
.
n1 ,j1 n2 ,j2
Only volatilities and instantaneous correlations matter for the pricing of derivatives, since
they determine the law of the asset prices under the domestic risk neutral measure. In our
analysis this is reflected by the fact that only inner products of the dispersion coefficients
and hence instantaneous covariances enter the pricing formulas. Nonetheless we start
with independent Brownian motions and model correlations by means of the dispersion
coefficients because this facilitates the use of stochastic calculus. To compute these
coefficients from the estimated instantaneous covariance matrix of the processes one may
use the Cholesky decomposition of this matrix.
Finally we note that the price process of a discounted foreign asset is not a martingale
under the domestic risk-neutral measure as can be seen from (6.24); hence this measure
must not be used for the valuation of derivatives paying off in foreign currencies.
For our pricing theory we need to assume that the markets in our economy are complete.
Assumption 6.14. There are d traded domestic assets such that for all t [0, TF ] the
instantaneous covariance matrix of these assets is strictly positive definite.
As shown in Proposition 5.20, his assumption guarantees that every contingent claim
adapted to the filtration generated by the asset prices can be replicated by a dynamic
trading strategy in the d assets and the domestic savings-account. Hence the domestic
risk neutral measure is unique and the price at time t of every domestic contingent claim
H with single FT measurable and integrable payoff HT at time T is given by
i
h
1

0
Ht := E Q t,T
HT Ft .
(6.26)
We now introduce the class of admissible underlyings for the derivative contracts considered in this section. A typical example of the kind of options we want to analyze is the
guaranteed-exchange-rate call. This contract is defined by its terminal payoff [
eSTf K]+ ,
where STf is some primitive foreign asset and e is a guaranteed exchange rate which will
1
Of course similar formulas hold for bonds and exchange rates.
65
be applied at time T to convert the price of the foreign asset into domestic currency. Now,
eSTf is not the time T value of a traded domestic asset. However, it defines a domestic
contingent claim X whose price Xt = E Q [(t,T )1 eSTf |Ft ] is given by
Z
Xt = X0 exp
0
sX dWs
|sX |2 ds
Z
+
rsd ds
h
i
with X0 = E Q (0,T )1 XT
and sX a deterministic Rd -valued function of time. We will see in section 6.4.3 below that
this structure is found in many ostensibly complex option contracts. This motivates the
following definition.
Definition 6.15. A domestic contingent-claim X with a single payoff XT at a certain
date T is called a lognormal claim2 if its price process (Xt )0tT given by

1

Q
d

XT Ft
Xt := E
t,T
admits a representation of the form
Z t

Z t
Z
1 t X2
X
d
Xt = X0 exp
s dWs
rs ds
| | ds +
2 0 s
0
0
(6.27)
with some constant X0 and with deterministic dispersion coefficients X : [0, T ] Rd .

Remark 6.16. The main restriction made in the definition of a lognormal claim is the
assumption of X being deterministic. In fact, whenever XT is strictly positive,

1

Q
d

Xt := E
t,T
XT Ft
is always of the form (6.27) with possibly stochastic volatility X , as can easily be
shown by means of the martingale representation theorem (Theorem 5.19). Note that the
solution of the SDE dXt = rt0 Xt dt + tX Xt dWt is given by
Z t

Z t
Z t
X
X 2
0
Xt := X0 exp
s dWs 1/2
|s | ds +
rs ds .
0
Hence under our assumption on asset price dynamics every primitive domestic asset,
interpreted as contingent claim with payoff equal to the assets price at time T , is a
lognormal claim. However, the class of contingent claims that satisfy Definition 6.15 is
much larger. For instance products and quotients of lognormal claims remain lognormal
claims.
6.4.2
Exchange Options on Lognormal Claims
Next we give a rather general theorem which leads to a unified treatment of the pricing of
European options on various underlyings such as foreign and domestic zero coupon bonds,
foreign or domestic stocks or forward and future contracts on foreign and domestic assets.
2
This name is motivated by the fact that XT is lognormally distributed. This is immediate if one writes
XT = XT /p0 (T, T ) and then expresses the right hand side using (6.27) and the corresponding expression
for p0 (, T ).
66
Theorem 6.17. Let X, Y be lognormal claims. Consider an option to exchange X for Y

at the maturity date T , i.e. a European option with payoff [XT YT ]+ .
1. The price process C = (Ct )0tT of this option is given by
Ct = C(t, Xt , Yt ) := Xt N (d1t ) Yt N (d2t )
where N denotes the one-dimensional standard normal distribution function, and where
d1t and d2t are given by
s
R
Z T
1 T
X Y |2 ds
ln
(X
/Y
)
+
|
t
t
s
s
1
2
1
2 t
qR
dt =
|sX sY |2 ds .
, dt = dt
T
t
X
Y 2
t |s s | ds
2. The hedge portfolio for this option in terms of the lognormal claims X and Y consists
of
C
X
(t) := N (d1t ) units of X and YC (t) := N (d2t ) units of Y .
Proof. The main tool in the proof is the change of numeraire technique as introduced in
Section 5.1.4. We now recall a few facts from this theory. Define for a lognormal claim X
a new equivalent probability measure QX on FT by
dQX
dQ

1
d
XT 0,T
X0
Then for every domestic asset Z whose discounted price process is a martingale under
Q that is for every asset that pays no dividends in [0, T ) the process Z/X is a
martingale under QX , i.e. QX is the martingale measure corresponding to the numeraire
X. Moreover we have the transition formula

1
X
Q
d
E
t,T
XT ZT |Ft = Xt E Q [ZT |Ft ]
(6.28)
In our setup it is easy to determine the law of the asset price processes under QX by
means of the Girsanov
theorem. Applying this theorem to dQX /dQ immediately yields
R
t
that WtX := Wt 0 sX ds is a Brownian Motion under QX .
Now it is easy to proof the first part of the theorem. According to (6.26) the price of the
option is given by

1

d
Ct = E Q t,T
[XT YT ]+ Ft

1
1

Q
d
Q
d
= E
t,T
XT 1{YT /XT <1} Ft E
t,T
YT 1{XT /YT >1} Ft
i
i
h
h
X
Y

= Xt E Q 1{YT /XT <1} Ft Yt E Q 1{XT /YT >1} Ft
The last line follows from (6.28) if we take once X and once Y as numeraire. Now we get
under QX for YT /XT
YT
Yt
=
exp
XT
Xt
Z
(sY
sX )dWsX
67
Z
t
|sY
sX |2 ds
Hence
QX

YT

< 1 Ft = QX ln YT ln XT < 0 Ft
XT
R
R
T Y
1 T
X
X
Y
X
2
( s )dWs
ln Xt ln Yt + 2 t |s s | ds
qR
= QX qt R s
<
T
T
Y
X
2
Y
X
2
t |s s | ds
t |s s | ds
qR
RT
Y
t (s
sX )dWsX /
X 2
Y
Since
and
are deterministic,
t |s s | ds is a standard
normally distributed random variable so that

X YT
Q
< 1 Ft = N (d1t ) .
XT

Analogously we get QY XT /YT > 1 Ft = N (d2t ) , and the first part of the theorem
follows.
Now we turn to the hedging part. Let (Z)M denote the (uniquely determined) martingale
part of a continuous semimartingale Z. To prove the second claim we note that the
proposed selffinancing hedge portfolio duplicates the option if it holds that
2
M
d (C)M = N (d1t )d(X)M
t N (dt )d(Y )t ,
(6.29)
and if moreover the value of the hedge portfolio equals the options price for all 0 t T .
We now check these two conditions. (i) As Ct is a function only of Xt and Yt we get from
Itos Lemma
d (C)M
t =
C
C
(t, Xt , Yt )d (X)M
(t, Xt , Yt )d (Y )M
t +
t .
x
y
Now following El Karoui, Myneni & Viswanathan (1992) we may compute the derivatives
of the option price:

1
C
Q
+
d
(t, Xt , Yt ) = Et
t,T
[XT YT ]
x
Xt

1
XT
d
= EtQ t,T
1{XT YT }
Xt

1 Q d 1
=
E
t,T
1{XT YT } XT
Xt t
As shown in the first part of the proof this expression equals N (d1t ). Similarly we get
C/y (t, Xt , Yt ) = N (d2t ), and hence (6.29).
(ii) By Eulers Theorem we get from the linear homogeneity of C in Xt and Yt
Ct =
C
C
(t, Xt , Yt ) Xt +
(t, Xt , Yt ) Yt = N (d1t )Xt N (d2t )Yt ,
x
y
which shows that also the second condition is satisfied.

Whenever the lognormal claims X and Y are assets for which liquid markets exist, Theorem 6.17 is sufficient for the construction of a hedge portfolio. Otherwise we must go on
68
and duplicate X and Y by a dynamic trading strategy. The existence of such a strategy
is guaranteed by Assumption 6.14; it can be computed as in the proof of Theorem 6.17.
The following observation then shows how to construct hedging strategies for C from the
hedge portfolios for X and Y . Suppose that the hedge portfolios for X and Y in terms
of domestic assets HiX and HiY for which we assume the existence of liquid markets are
given by
LX
LY
X
X
X
X
X
Y
Pt =
i (t)Hi and Pt =
iY (t)HiY .
i=1
i=1
Then the hedge portfolio for the exchange option on X and Y in terms of HiX and HiY is
given by
LX
LY
X
X
1
X
X
Pt =
N (dt ) i (t)Hi
N (d2t ) iY (t)HiY .
i=1
i=1
The application of this principle is illustrated in certain examples presented below.
6.4.3
Options on Lognormal Claims: Examples
Now we want to use Theorem 6.17 in order to price a number of practically relevant
contracts.
Currency Options: The payoff of a plain vanilla currency option equals [eT K]+ .
Define the domestic assets X := e pf (, T ) and Y := Kpd (, T ); the parameters of their
price processes can be read off from the asset price dynamics and are given by X0 =
e0 pf (0, T ), X (t) = e (t) + f (t, T ) and Y0 = Kpd (0, T ), Y (t) = d (t, T ), respectively.
Since pd (T, T ) = pf (T, T ) = 1 the options payoff equals [XT YT ]+ , and its price can be
computed by means of Theorem 6.17. We obtain
Ct = et pf (t, T )N (d1 ) Kpd (t, T )N (d2 ) with
RT
ln(et pf (t, T )/pd (t, T )) ln K + t | e (s) + f (s, T ) d (s, T )|2 ds
qR
d1 =
T
e
f
d
2
t | (s) + (s, T ) (s, T )| ds
qR
T
f
d
e
f
d
2
and d2 = d1
t | (s) + (s, T ) (s, T )| ds. Since we assume p and p to be
traded assets we can use directly Theorem 6.17 to compute a feasible hedge portfolio.
Currency Converted Options: There are two types of currency converted options.
The payoff of a Foreign Asset/ Domestic Strike Option equals [eT STf K]+ . To deal with
this claim we set X := eS f and notice that this is a lognormal claim with Xt = et Stf and
f
X = e + S . Next set Y := K pd (, T ). Theorem 6.17 can now be directly applied to
give the price and the hedging strategy of this contract. Similarly for a Domestic Asset/
Foreign Strike Option with payoff [STd eT K]+ , where K is in foreign currency we use the
lognormal claims X := S d and Y = Kepf (, T ).
Guaranteed-Exchange-Rate Options:
The payoff of this derivative equals
f
+
f
[
eST eK] , where S is a foreign asset and e some predetermined exchange rate. This
contract can be interpreted as an option to exchange the lognormal claims X and Y with
payoff XT = eSTf and YT := eK. Whereas YT equals the time T value of K e units of
69
pf (, T ), there is no traded asset whose value at T is equal to XT . To price and hedge the
option we therefore have to compute the parameters of X. We have
(Z
T
d (0, T )
p
f
X0 = eS0f f
exp
| f (s, T )|2 + S (s) d (s, T ) + e (s) f (s, T ) (6.30)
p (0, T )
0
)
f
S (s) f (s, T ) f (s, T ) d (s, T ) S (s) e (s)ds

sX
= S (s) + d (s, T ) f (s, T )
The price of the option can now be computed by plugging these parameters into the pricing
formula of Theorem 6.17. Next we want to determine the hedge portfolio for the option.
As XT is not the terminal value of a traded asset we have to go through the procedure
outlined after the proof of Theorem 6.17. To replicate XT by a dynamic trading strategy
of the domestic assets eS f , pd (t, T )
we first note that by (6.30) Xt is given by a function X
f
f
f = X/(e
t pf (t, T )) and
t S f ), X/ep
and e p (t, T ) with derivatives X/eS
= X/(e
t
d = X/p
d (t, T ). As X
is linear homogenous in the prices of these assets, an
X/p
argument similar to the proof of the second part of Theorem 6.17 shows that the hedge
portfolio for X equals
X
eS
f (t) =
Xt
et
Stf
X
ep
f (t) =
Xt
,
et pf (t, T )
pXd (t) =
Xt
d
p (t, T )
t)
Options on Forwards: Assume that (X
0tT is the price process of a lognormal claim.
Consider two points in time T1 and T2 with 0 < T1 T2 . The payoff in T1 of an option
with maturity date T2 is given by
with
maturity dateiT1 on a forward contract on X
h
+
T pd (T1 , T2 )K . For t < T1 price and hedge portfolio of this contract immediately
X
1
t and Kpd (t, T2 ).

follow from Theorem 6.17, if we use the lognormal claims X
Options on Interest Rates: We are mainly interested in contracts where one of the
underlying assets is a foreign or domestic LIBOR rate. For a fixed > 0 (in practice
usually = 0.25 or = 0.5) the LIBOR rate Ln (t, ) prevailing in country n over the
period [t, t + ] is defined by the equation
(1 + Ln (t, ))pn (t, t + ) = 1 ,
that is Ln (t, ) = 1 (1/pn (t, t + ) 1).
Caps: Perhaps the most important LIBOR derivatives are caps and floors. A cap is a
portfolio of caplets. The payoff of a caplet with face value V , underlying interest rate
process Ld (t, ), level K and maturity date T + equals

+
1
d
+
V [L (T, ) K] = V d
(K + 1)
p (T, T + )
As the payoff of this caplet is known already at T we may compute its present value at
T which equals V [1 (K + 1) pd (T, T + )]+ . From this we see that the price and
the hedge portfolio for caplets can be inferred directly from Theorem 6.17 if we use the
lognormal claims X = pd (, T ) and Y = (K + 1) pd (, T + ). Of course this choice of
X and Y reflects the well-known fact that caplets can be considered as options on zero
coupon bonds.
70
Appendix A
Mathematical Background
A.1
Conditional Expectation
Given a probability space (, F, P ) and a random variable X. A priori, the best prediction
for X is E(X). If we have additional information about the outcome of the experiment
modelled by (, F, P ), we can give a better prediction of X. The best-possible prediction
in an L2 -sense is the conditional expectation. We first study this idea in an elementary
setting where information is modelled by a (finite) partition of , which leads us to an
explicit formula for the conditional expectation. In a second step we will use the properties
of this elementary conditional expectation to extend the notion to general probability
spaces.
A.1.1
The elementary case
Definition A.1. A set A = {A1 , . . . , An } of measurable subsets of with P (Ai ) > 0 for
all i is called a partition of if Ai Aj = for i 6= j and if moreover = A1 An .
Now consider a partition A of . Suppose that we have the additional information that
the result of our random experiment belongs to a particular subset Ai0 A. Our best
prediction for the rv X is now
E(X|Ai0 ) :=
1
E(X1Ai0 )
P (Ai0 )
The prediction-mechanism which gives the prediction of X if we are given the additional
information which set from the partition A actually occurs is therefore given by the
random variable
n
n
X
X
E(X1Ai0 )
1Ai ()E(X|Ai ) =
1Ai ()
.
P (Ai )
i=1
i=1
Example: Consider a 2-period binomial model with P (up) = P (down) = 12 . Then

E(S2 ) = S0 ( 14 u2 + 21 ud + 41 d2 ) = S0 ( 12 u + 12 d)2 . Define the partition A = {A1 , A2 } with
A1 = {S1 = uS0 } and A2 = {S1 = dS0 }, i.e. the partition is formed by the value of the
stock-price at t = 1. Then the forecast of S2 given the information contained in A is given
by

1 2 1
1
1 2
1A1 ()E(S2 |A1 ) + 1A2 ()E(S2 |A2 ) = 1A1 ()S0
u + ud + 1A2 ()S0
ud + d
.
2
2
2
2
71
Obviously, this forecast differs depending on whether A1 or A2 actually occurs in t = 1.

Formally, the additional information is described by the -field F A generated by the
partition A.
Definition A.2. Given a partition A = {A1 , . . . , An } of . The -field F A generated by
S
A is the set of all unifications kj=1 Aj , Aj A, k N.
Remark: Usually the -field F A is defined as the smallest -field containing all the sets
A1 , . . . , An . It is easily seen that the two definitions are equivalent.
Definition A.3. Given a partition A = {A1 , ..., An } of with sigma-field F A and a
random variable X. The conditional expectation of X given F A is the random variable
A
E(X|F )() =
n
X
1Ai ()E(X|Ai ) =
i=1
n
X
1Ai ()
i=1
E(X1Ai )
.
P (Ai )
Proposition A.4. Given a partition A of and a rv X. The conditional expectation

E(X|F A ) has the following properties
(i) E(X|F A ) is F A -measurable.
(ii) For every random variable Y which is F A -measurable (i.e. Y is constant on the sets
Ai , i = 1, ..., n) we have E(XY ) = E(E(X|F A )Y ).
Proof. The property (i) is clear, as E(X|F A )is constant on each Aj .
P
measurable it is of the form Y = ni=1 cj 1Aj for constants cj . Hence
E(XY ) =
n
X
cj E(X1Aj ) =
i=1
= E
cj P (Aj )
i=1
n
X
i=1
A.1.2
n
X
E(X1Aj )
cj
1Aj
P (Aj )
As Y is F A -
E(X1Aj )
P (Aj )
!
= E(Y E(X|F A )) .
Conditional Expectation - General Case
The explicit definition of the conditional expectation works only if P (Ai ) 0 for all sets
in our partition. However, in continuous models this is usually not the case. We therefore
use the properties of the conditional expectation obtained in Proposition A.4 to define
the conditional expectation in more general situations.
Definition A.5. Given an integrable rv X on (, F, P ) and a sigma-field G F. A
random variable Z is called conditional expectation of X given G, Z = E(X|G), if
(i) Z is G-measurable.
(ii) E(Y X) = E(Y Z) for all rvs Y which are G-measurable.
Theorem A.6. There is exactly one random variable Z which satisfies (i), (ii).
The proof can be found in any standard textbook on probability theory.
Examples:
72
(1) If G = {, } we have E(X|G) = E(X).

(2) If X is G-measurable we have E(X|G) = X
(3) If X1 , X2 are independent and G := (X2 ) we get for any bounded measurable
function f that E(f (X1 )|G) = E(f (X1 )).
Proposition A.7. The conditional expectation has the following properties:
(1) Linearity: E(c1 X1 + c2 X2 |G) = c1 E(X1 |G) + c2 E(X2 |G).
(2) If Y is G-measurable we have E(Y X|G) = Y E(X|G).
(3) Projectivity of the conditional expectation: Consider sigma-fields G0 G1 F.
Then we have E(X|G0 ) = E(E(X|G1 )|G0 ), in particular E(X) = E(X|G) for every
sub--field G.
Property (3) is often referred to as law of iterated expectations.
Proof. ad 2) We have to check the property (ii) of the definition of the conditional expectation. Let Z be G-measurable. Then
E(Y XZ) = E((Y Z)X) = E(Y ZE(X|G)) = E(Y (ZE(X|G)) ,
as the product (Y Z) is G-measurable.
ad 3) E(X|G0 ) is obviously G0 -measurable. Consider a G0 -measurable random variable
Y . As Y is also G1 -measurable, we have E(Y (E(X|G1 )) = E(XY ). On the other hand
we get from the definition of E(X|G0 ) that E(XY ) = E(Y E(X|G0 )). This shows that
E((X|G1 )Y ) = E(Y E(X|G0 )) so that Definition (A.5)(ii) holds.
73
Bibliography
Amin, K. & Jarrow, R. (1991), Pricing foreign currency options under stochastic interest
rates, Journal of International Money and Finance 10, 310330.
Bachelier, L. (1900), Theorie de la speculation, in P. Cootner, ed., The Random Character
of Stock Market Prices, MIT Press, Cambridge, Mass. (1964), pp. 1778.
Bingham, N. & Kiesel, R. (1998), Risk-neutral Valuation, Springer.
Bjork, T. (2004), Arbitrage Theory in Continuous Time, 2nd edn, Oxford University Press,
Oxford.
Black, F. & Scholes, M. (1973), The pricing of options and corporate liabilities, J. Polit.
Economy 81(3), 637654.
Cont, R. (2006), Model uncertainty and its impact on the pricing of derivative instruments, Math. Finance 16, 519542.
Cont, R. & Tankov, P. (2003), Financial Modelling with Jump Processes, Chapman &
Hall / CRC.
Cox, J. & Rubinstein, M. (1985), Options markets, Prentice Hall, Englewood Cliffs.
Delbaen, F. & Schachermayer, W. (1994), A general version of the fundamental theorem
of asset pricing, Math. Ann. 300, 463520.
El Karoui, N., Myneni, R. & Viswanathan, R. (1992), Arbitrage pricing and hedging
of interest rate claims with state variables: 2, applications, preprint, University of
Paris VI.
Embrechts, P., Kl
uppelberg, C. & Mikosch, T. (1997), Modelling Extremal Events for
Insurance and Finance, Springer, Berlin.
Filipovic, D. (2009), Term-Structure Models A Graduate Course, Springer Finance,
Springer, Berlin Heidelberg.
Follmer, H. (1981), Calcul dIt
o sans probabilites., Vol. Seminaire Proba. XV, Lecture
Notes in Mathematics vol. 850, Springer, Berlin, pp. 144150.
Follmer, H. & Schweizer, M. (1993), A microeconomic approach to diffusion models for
stock prices, Math. Finance 3(1), 123.
Frey, R. (1997), Derivative asset analysis in models with level-dependent and stochastic
volatility, CWI Quarterly 10, 134.
74
Frey, R. & Sommer, D. (1996), A systematic approach to pricing and hedging of international derivatives with interest rate risk, Appl. Math. Finance 3, 295317.
Gibson, R., ed. (2000), Model Risk, Concepts, Calibration and Pricing, Risk Publications,
London.
He, H. & Leland, H. (1993), On equilibrium asset price processes, Rev. Finan. Stud.
6(3), 593617.
Heath, D., Jarrow, R. & Morton, A. (1992), Bond pricing and the term structure of
interest rates: A new methodology for contingent claim valuation, Econometrica
60(1), 77105.
Hull, J. (1997), Options, futures and other derivatives, 3rd ed., Prentice Hall Int., Englewood Cliffs.
Karatzas, I. & Shreve, S. (1988), Brownian Motion and Stochastic Calculus, Springer,
Berlin.
Karatzas, I. & Shreve, S. (1998), Methods of Mathematical Finance, Springer, Berlin.
Malkiel, B. (1987), Efficient market hypothesis, in The New Palgrave: Dictionary of
Economics, Macmillan, London, pp. 127134.
McNeil, A., Frey, R. & Embrechts, P. (2005), Quantitative Risk Management: Concepts,
Techniques and Tools, Princeton University Press, Princeton, New Jersey.
Protter, P. (1992), Stochastic Integration and Differential Equations: A New Approach,
Applications of Mathematics, Springer, Berlin.
Revuz, D. & Yor, M. (1994), Continuous Martingales and Brownian Motion, Springer,
Berlin.
Samuelson, P. (1965), Rational theory of warrant pricing, Industrial Management Review
6, 1331.
Sandmann, K. (1999), Einf
uhrung in die Stochastik der Finanzm
arkte, Springer, Berlin.
Wilmott, P., Dewynne, J. & Howison, S. (1993), Option Pricing, Mathematical Models
and Computation, Oxford Financial Press, Oxford.
75

Continuous Time Finance II: Lecture Notes: Prof. R Udiger Frey, Ruediger - Frey@wu - Ac.at

Uploaded by

Document Information

Original Title

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

Continuous Time Finance II: Lecture Notes: Prof. R Udiger Frey, Ruediger - Frey@wu - Ac.at

Uploaded by

Copyright:

Available Formats

Continuous Time Finance II: Lecture Notes

Stochastic Processes, Stopping Times and Martingales . . . . . . . . . . .

Stopping Times and Martingales . . . . . . . . . . . . . . . . . . . . . . .

The optional sampling theorem . . . . . . . . . . . . . . . . . . . .

Definition and Construction . . . . . . . . . . . . . . . . . . . . . .

Some stochastic properties of Brownian motion: . . . . . . . . . . .

Martingale-property of the Ito-integral . . . . . . . . . . . . . . . .

Covariation and d-dimensional Ito-formula

The d-dimensional Ito-formula . . . . . . . . . . . . . . . . . . . .

3 The Black-Scholes Model: a PDE-Approach

Asset Price Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

Pricing and Hedging of Terminal Value Claims . . . . . . . . . . . . . . .

The pricing-equation for terminal-value-claims . . . . . . . . . . .

The Black-Scholes formula . . . . . . . . . . . . . . . . . . . . . . . . . . .

Application of the formula: volatility estimation . . . . . . . . . .

4 Further Tools from Stochastic Calculus

Stochastic Integration for Continuous Martingales . . . . . . . . . . . . .

Stochastic Integrals for elementary processes . . . . . . . . . . . .

Extension to General Integrands . . . . . . . . . . . . . . . . . . .

Change of Measure and Girsanov Theorem for Brownian motion . . . . .

The Girsanov Theorem . . . . . . . . . . . . . . . . . . . . . . . .

5 Financial Mathematics in Continuous-Time

Martingale Measures and Arbitrage Opportunities . . . . . . . . .

Hedging and risk-neutral pricing of contingent claims . . . . . . . .

The Black-Scholes Model Revisited . . . . . . . . . . . . . . . . . . . . . .

Fundamental Theorems of Asset Pricing in a Generalized Black-Scholes Model 49

A multidimensional Black-Scholes model . . . . . . . . . . . . . . .

Existence of an equivalent martingale measure . . . . . . . . . . .

Market completeness in the generalized Black-Scholes model . . . .

Bonds and Interest-Rates . . . . . . . . . . . . . . . . . . . . . . . . . . .

Martingale Modelling and Term-Structure Equation . . . . . . . .

Affine Term Structure . . . . . . . . . . . . . . . . . . . . . . . . .

The Musiela parametrization. . . . . . . . . . . . . . . . . . . . . .

Pricing and Hedging of Multi-Currency Derivatives with Interest Rate Risk 63

Options on Lognormal Claims: Theory . . . . . . . . . . . . . . . .

Options on Lognormal Claims: Examples . . . . . . . . . . . . . .

A.1 Conditional Expectation . . . . . . . . . . . . . . . . . . . . . . . . . . . .

A.1.1 The elementary case . . . . . . . . . . . . . . . . . . . . . . . . . .

A.1.2 Conditional Expectation - General Case . . . . . . . . . . . . . . .

Stochastic Processes, Stopping Times and Martingales

We work on a probability space (, F, P ) with filtration {Ft }. Recall that a filtration is

is called trajectory or sample path of X; a stochastic process can be viewed as

Equality of stochastic processes.

There are two nations of equality for stochastic

which proves the claim.

Stopping Times and Martingales

Throughout this section we work on a probability space (, F, P ) with filtration {Ft }.

As {Xq A} Fq Ft and as the union of countable sets from Ft also belongs to Ft ,

Next, we define the -field generated by all observable events up

Definition 1.10. Given a stopping time . Then we call

the -field of the events which are observable up to .

The optional sampling theorem

See again Protter (1992), Section I.2 for a proof.

Definition and Construction

Definition 1.20. A stochastic process X = (Xt )t0 on (, F, P ) is standard one-dimensional

Some stochastic properties of Brownian motion:

If Var(X) < X is said to be of finite variation.

sup |X(ti ) X(ti1 )| Var(X).

(Xti Xti1 )(Ati Ati1 )

which converges to zero as X is continuous.

(Bti Bti1 ) (ti ti1 )

E (Bti Bti1 )2 (ti ti1 )

which proves the theorem.