Treball final de grau

GRAU DE MATEMÀTIQUES

Facultat de Matemàtiques i Informàtica

Universitat de Barcelona

AN INTRODUCTION TO
STOCHASTIC VOLATILITY
MODELS

Autor: Matias Puig

Director: Dr. Jose Manuel Corcuera

Realitzat a: Departament
de Matemàtiques i Informàtica

Barcelona, 29 de juny de 2017

Contents

Introduction i

1 Historical Background on Stock Price Models 3

2 Stochastic Integration 7
2.1 Martingales and Brownian motion . . . . . . . . . . . . . . . . . 7
2.2 Integral of mean square integrable processes . . . . . . . . . . . 9
2.3 Extension of the integral . . . . . . . . . . . . . . . . . . . . . . . 12

3 Fundamental Theorems 17
3.1 Itô’s formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
3.2 Stochastic differential equations . . . . . . . . . . . . . . . . . . . 18
3.3 Girsanov’s theorem . . . . . . . . . . . . . . . . . . . . . . . . . . 19
3.4 Martingale representation theorem . . . . . . . . . . . . . . . . . 21

4 The Black-Scholes Model 25

4.1 Risk-neutral measure . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.2 Arbitrage and admissibility . . . . . . . . . . . . . . . . . . . . . 28
4.3 Completeness . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
4.4 Pricing and hedging . . . . . . . . . . . . . . . . . . . . . . . . . 32
4.4.1 Pricing a put option . . . . . . . . . . . . . . . . . . . . . 33
4.4.2 Pricing a call option . . . . . . . . . . . . . . . . . . . . . 34
4.4.3 Hedging . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

5 Stochastic Volatility 39
5.1 Empirical motivations . . . . . . . . . . . . . . . . . . . . . . . . 39
5.2 A general approach for pricing . . . . . . . . . . . . . . . . . . . 39
5.3 Heston’s model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

6 Conclusions 47

7 Bibliography 49
Bibliography: Articles . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Bibliography: Books . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
Bibliography: Online . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50

i
Introduction 1

The main goal of this work is to introduce the stochastic volatility models in
mathematical finance and to develop a closed-form solution to option pricing
in Heston’s stochastic volatiltiy model, following the arguments in Heston
1993.

No background in mathematical finance will be assumed, so another main

goal of this work is to develop the theory of stochastic integration and to
introduce the Black-Scholes market model, the benchmark model in mathe-
matical finance. Standard topics in the framework of market models, such as
trading strategies, completeness and replication, and the notion of arbitrage,
will also be reviewed.
2 Introduction
Chapter 1

Historical Background on Stock

Price Models

Louis Bachelier, in his thesis "Théorie de la Spéculation", made the first contri-
bution of advanced mathematics to the study of finance in 1900. This thesis
was well received by academics, including his supervisor Henry Poincaré,
and was published in the prestigious journal Annales Scientifiques de l’École
Normale Supérieure. In this pioneering work, the Brownian motion is used for
the modelling of movements in stock prices. In the words of Louis Bachelier,
in Bachelier 1900:

La détermination des mouvements de la Bourse se subordonne à

un nombre infini de facteurs: il est dès lors impossible d’en espérer
la prévision mathématique, [...] et la dynamique de la Bourse ne
sera jamais une science exacte.
Mais il est possible d’étudier mathématiquement l’état statique du
marché à un instant donné, c’est-à-dire d’établir la loi de proba-
bilité des variations de cours qu’admet à cet instant le marché. Si
le marché, en effet, ne prévoit pas les mouvements, il les consid-
ère comme étant plus ou moins probables, et cette probabilité peut
s’évaluer mathématiquement.

Bachelier argued that, over a short time period, fluctuations in price are inde-
pendent of the current price and the past values of the price, and that these
fluctuations follow a zero mean normal distribution with variance propor-
tional to the time difference. He also assumed that the prices are continuous,
therefore modelled as a Brownian motion (see Bachelier 2011).

Many years later, in the famous article by Black and Scholes, Black and Sc-
holes 1973, prices are modeled as a geometric Brownian motion, whose fluctua-
tions have a lognormal distribution. This model is based on the assumption
that the log returns of a stock price are independent and normally distributed,
with variance proportional to the time difference. The log returns are defined
as log( pi ) log( p j ), where pi and p j denote the prices at times i and j, respec-
tively, with i > j.

The returns of a stock price are defined as the increments

pi pj
ri =
pj
Hence, the approxamation

3
4 Historical Background on Stock Price Models

Figure 1.1: Sample path of a standard Brownian motion, used by Bachelier to

model fluctuations in stock prices

log(1 + r ) ⇡ r, when r ⌧ 1 (1.1)

Gives the approximation

pi pj
= ri ⇡ log(1 + ri ) = log( pi /p j ) = log( pi ) log( p j ) (1.2)
pj
So that the lognormal distribution for the stock price increments proposed by
Black and Scholes obeys to the intuitive idea that the price returns are inde-
pendent and normaly distributed.

Explicitely, as stated in Black and Scholes 1973, the model proposed by Black
and Scholes relies on the following assumptions of an "ideal" market:

(a) The interest rate is known and constant through time

(b) The distribution of stock prices at the end of any finite interval is lognor-
mal

(c) The stock pays no dividends

(d) The variance rate of the return on the stock is constant

(e) The stock price is continuous over time

Empirical observations of stock price distributions have motivated numerous

extensions of the Black-Scholes model in which one or more of the previ-
ous assumptions are relaxed. One of the main criticisms of the Black-Scholes
model is that the normal distribution of stock price returns does not explain
the significant presence of outliers in the distribution of returns (see, for ex-
ample, Mandelbrot 1963). Conveniently, relaxing some of the previous as-
sumptions results in distributions with a higher presence of outliers, or fatter
 5

Figure 1.2: Sample path of a Geometric Brownian motion (with µ = s = 1),

used by Black and Scholes to model stock prices

tails, which better adjust to the reality of observed prices.

Examples of such extensions include models with dividend payments, stock

prices with jumps (non-continuous over time), and models in which the distri-
bution of returns is non-Gaussian, among others. These extensions are widely
used by practitioners and are described in most text books on mathematical
finance (see, for example, Musiela and Rutkowski 2006).

In this project we will concentrate on the stochastic volatility models, in which

the assumption that the variance rate of the return is constant is relaxed.
6 Historical Background on Stock Price Models
Chapter 2

Stochastic Integration

Let (W, F , P ) be a probability space, and T an index set. Consider a func-

tion X (t, w ) having two arguments in T and W. By fixing t, X (t, ·) t2T is
a family of random variables defined in W. On the other hand, by fixing w,
X (·, w ) w2W can be seen as a family of ’random maps’. We will define a
stochastic process following the first point of view and show the equivalence
between the two.

Definition 2.1 (Stochastic Process). Let E be a metric space with the borel s-field.
A stochastic process is a family of E-valued random variables ( Xt )t2T defined in a
probability space (W, F , P )

In order to define a random map as a function-valued random variable, we

must specify the function space and a s-field in this space. The most natural
function space to consider is the space E T of all maps from T to E. The s-field
G will be the product s-field, that is the one generated by cylinder sets of
the form pt1 1 ( B1 ) \ . . . \ ptn 1 ( Bn ), with B1 , . . . , Bn 2 B( E) and pt being the
natural projection defined by the rule f 7! f (t).

Let ( Xt )t2T be a stochastic process and define Y (w)(t) = Xt (w). Is the

map Y : (W, F , P ) ! ( E T , G) G -measurable? Indeed, for any cylinder set
C = pt1 1 ( B1 ) \ . . . \ ptn 1 ( Bn ) 2 G , Y 1 (C ) = Xt1 1 ( B1 ) \ . . . \ Xtn 1 ( Bn ) 2 F .
Since G is generated by the cylinder sets, we conclude that Y is G -measurable.
On the other hand, if Y : (W, F , P ) ! ( E T , G) is G -measurable, for any
t 2 T and w 2 W define Xt (w ) := Y (w )(t). Then, for any B 2 B( E),
Xt 1 ( B) = {w | Xt (w ) 2 B} = {w |Y (w )(t) 2 B} = Y 1 (pt 1 ( B)) 2 F .

Given the equivalence between both definitions, we will consider E-valued

stochastic processes on T as any of the definitions above. A sample path of X is
a function x (t) = X (t, w). We say that a E-valued stochastic process on T has
paths in U ⇢ E T if its sample paths are included in U. A stochastic process X
is said to be continuous if its paths are included in C( T ), the set of continuous
functions. The process X is continuous at t0 2 T if its sample paths x (t) are
continuous at t0 almost surely. Two stochastic processes X and Y are said to
be versions of each other if P ( Xt = Yt ) = 1 for all t 2 T. The finite dimensional
distributions of a stochastic process X are the joint distributions of the random
vectors ( Xt1 , . . . , Xtn ), t1 , . . . , tn 2 T.

2.1 Martingales and Brownian motion

We will now consider stochastic processes with the index set T = R + , and
with values on R.

7
8 Stochastic Integration

Definition 2.2 (Filtration, adapted process). A filtration (Ft )t 0 is an increasing

family of s-algebras included in the s-algebra F . A process is said to be adapted to
the filtration (Ft )t 0 if for each t, Xt is Ft measurable.

Given a process ( Xt )t 0 , we can define the natural filtration Ft = s ({ Xs , s 

t}), for which the process X is adapted to. Moreover, it is frequent to consider
the completion of the filtration (Ft )t 0 - that is, the filtration in which all the
F -negligible sets are F0 -measurable. We will refer to this filtration as the one
generated by the process X, without making explicit that it is the completion.

Definition 2.3 (Martingale). A process ( Xt )t 0 adapted to a filtration (Ft )t 0 is a

martingale if E [| Xt |] < • 8t 0 and E [ Xt |Fs ] = Xs a.s for all s < t.

The following result will be useful when defining the stochastic integral in
the next section:

Theorem 2.4 (Doob inequality). If ( Mt )t 0 is a martingale with continuous sam-

ple paths, then:
⇥ ⇤
E sup | Mt |2  4E | MT |2
t T

Proof. Refer to Lamberton and Lapeyre 2007 (page 58, exercise 13).

We will need some further concepts to define a Brownian motion. A stochastic

process is said to have independent increments if for all 0  t1 < . . . < tn , Xt1
Xt0 , Xt2 Xt1 , . . . , Xtn Xtn 1 are independent. It is said to have stationary
increments if for t1 < t2 , Xt2 Xt1 ⇠ Xt2 t1 X0 .

Definition 2.5 (Brownian motion). A Brownian motion (also known as Wiener

process) is a stochastic process (Wt )t 0 with continuous sample paths and indepen-
dent and stationary increments.

The following result shows further properties of the Brownian motion which
are not explicit in the definition. For a proof of this result, refer to Corcuera
n.d.

Proposition 2.6. Let Wt be a Brownian motion. Then Wt W0 ⇠ N (rt, s2 t), for

some r 2 R.

A standard Brownian motion is a Brownian motion such that W0 = 0 a.e and

Wt ⇠ N (0, t). From now on, we will usually refer to a standard Brownian
motion as a Brownian motion, without specifying that it is standard.

Proposition 2.7 (Martingale property of Brownian motion). A standard Brow-

nian motion is a martingale.

Proof. Let X be a Brownian motion and s < t.

E [ Xt |Fs ] Xs = E [ Xt Xs |Fs ] = E [ Xt Xs ] = 0

The second equality is due to the fact that Xt Xs is independent to Fs due

to the independent increments property.

The following is another important property of the Brownian motion:

2.2 Integral of mean square integrable processes 9

Lemma 2.8. Let (Wt )t 0 be a Brownian motion under the probability P. For any
fixed M > 0,

P (max Wt (w )  M ) = 0
t 0

P (min Wt (w ) M) = 0
t 0

Proof. This is a consequence of the distribution of the first passage time of a

Brownian motion. We define the first passage time TM to a level M 2 R as:

TM (w ) = inf{t 0, Wt (w ) = M }
It can be proved that the first passage time distribution has the following
property (see Karatzas and Shreve 2012, p.80, equation 6.2):

P ( TM < t) = 2P (Wt > M) !1

t!•
Since Wt is normally distributed. This proves the result, as any level M will
be reached almost surely.

2.2 Integral of mean square integrable processes

The fact that a Brownian motion is almost surely nowhere differentiable - a
proof of this result can be found in Karatzas and Shreve 2012-, means that it is
not possible to define the integral of a function f over a Brownian motion as
RT 0
0 f ( s )|Ws |ds, for any fixed T > 0. We will construct the stochastic integral
by defining it among a class of simple processes and then extend it to a larger
class (as in the definition of the Lebesgue Integral). From now on, we will
consider processes with an index set [0, T ].

Definition 2.9 (Mean Square Integrable Process). An adapted process ( Xt )t 0 is

RT
said to be mean square integrable if E ( 0 Xt2 dt) < •. We will denote this class of
processes as S2 .

We aim to define the integral for mean square integrable processes. By iden-
tifying processes that are versions of one another and defining the natural
RT
norm k X k2S2 = E ( 0 Xt2 dt), it can be shown that S2 is a Hilbert space, with
RT
the scalar product < X, Y >= E ( 0 Xt Yt dt).

For constructing the integral, it is necessary to identify the so-called elemen-

tary processes for which the integral can be defined trivially, and so that it
can be extended to the whole S2 space.

Definition 2.10 (Simple Process). A process ( Xt )tT 2 S2 is a simple process if

there exists a partition 0 = t0 < . . . < tn = T of [0, T ] for which Xt = Xti for
t 2 [ t i , t i +1 [ .

Definition 2.11. The integral of a simple process ( Xt )0tT from 0 to T is defined

as:
Z T
I (X )T =
0
Xt dWt := Â Xti (Wti+1 Wti ) (2.1)
i

The integral process is defined as the process ( I ( X )t )tT , where I ( X )t = I ( Xs 1{st} ).

Note that I ( X )t is given by:
10 Stochastic Integration

n 1
I ( X )t = Â Xti (Wti+1 ^t Wti ^t )
i =0

So the integral process is pathwise continuous.

Let E ⇢ S2 be the set of simple processes. Note that the integral of a simple
process is a random variable. In addition, it is in L2 , and the integral is an
isometry between these two spaces:

Lemma 2.12 (Isometry Property for Simple Processes). The integral defines an
isometry between E and a subspace of L2 (W, F T , P ).

Proof. Let I be the integral operator. We have to prove that:

h Z T i h Z T i
2
k I ( X )k2L2 =E Xt dW =E Xt2 dt = k X k2S2
0 0
Now,

n 1
I ( X )2 = Â Xt2i (Wti+1 Wti )2 + 2 Â Xti Xt j (Wti+1 Wti )(Wt j+1 Wt j )
i =0 0 i < j < n

Where t0 = 0 and tn = T. Note that Wt j+1 Wt j is independent of Ft j and has

0 expectation.

E Xti Xt j (Wti+1 Wti )(Wt j+1 Wt j ) = E Xti Xt j (Wti+1 Wti ) E Wt j+1 Wt j = 0

Now, applying a similar reasoning,

E [ Xt2i (Wti+1 Wti )2 ] = E [ Xt2i ]E [(Wti+1 Wti )2 ] = E [ Xt2i ](ti+1 ti )

Because Wti+1 Wti ⇠ N (0, ti+1 ti ). We obtain:

n 1
E [ I ( X )2 ] = Â E[Xt2i ](ti+1 ti )
i =0

Since X is constant on every interval [ti , ti+1 [, the expression above is equal
to:

n 1 hZ i +1 i hZ T i
ÂE Xt2 dt = E Xt2 dt
i =0 i 0

The following result will allow us to extend the integral to the S2 space:

Lemma 2.13. The simple processes are dense in S2

The proof of this result requires several steps and will be omitted - it can be
found in Karatzas and Shreve 2012 (page 134, problem 2.5). However, to give
an idea on how it is accomplished, note that any continuous process ( Xt )0tT
can be approximated in S2 ([0, T ]) by simple processes Xtn defined as:

T2n 1
Xtn = Â X (i2 n
)1]i/2n ,(i+1)/2n [ (t)
i =0
2.2 Integral of mean square integrable processes 11

However, it is not trivial to obtain a similar result for more general adapted
processes.

To extend the integral, we apply the general result that given two metric
spaces X and Y and a dense subset A of X, any uniformly continuous map
f : A ! Y can be extended in a unique way to X by defining fˆ( x ) = lim f ( an )
n
where an is any sequence in A converging to x, and this extension is well
defined and continuous. In addition, if Y is a vector space, the extension
preserves the supremum norm. In our case, we can extend the integral
I : E ! L2 (W, F T , P ) to Î : S2 ! L2 (W, F T , P ) and it remains an isome-
try.
RT
Given a process ( Xt )0tT , we define the process I ( X ) t
by I ( X )t = 0 Xs 1[0,t] (s) dWs .
This process is a continuous martingale:

Lemma 2.14. [Martingale Property of the Integral] The process I ( X ) t

is a mar-
tingale.

Proof. Consider a first case in which X is a simple process, and let s < t. Then,
we can assume tn = t and s 2 [tk , tk+1 [.
Z t n

0
Xs dW = Â Xti (Wti+1 Wti )
i =0

k 1 n
= Â Xti (Wti+1 Wti ) + Xtk (Wtk+1 Wtk ) + Â Xti (Wti+1 Wti ) (2.2)
i =0 i = k +1

Now, the first part of the last equation is an Fs measurable random variable,
so its conditional expectation to Fs is the same variable.

For the second part, using the martingale property of the Brownian motion,

E[ Xtk (Wtk+1 Wtk )|Fs ] = Xtk E[Wtk+1 Wtk |Fs ] = Xtk (Ws Wtk )
Finally, for j > k, and by the tower property of the conditional expectation,

E[ Xt j (Wt j+1 Wt j )|Fs ] = E[ E[ Xt j (Wt j+1 Wt j )|Ft j ]|Fs ] = E[ Xt j (Wt j Wt j )|Fs ] = 0

So the last term in the equation is 0. Adding the results, we obtain:

k 1
E[ I ( X )t |Fs ] = Â Xti (Wti+1 Wti ) + Xtk (Ws Wtk ) = I ( X )s
i =0
We have proven that the integral of a simple process is a martingale. To
extend this result to the S2 space, it is sufficient to show that the martingale
property is preserved by limits in L2 . This is a consequence of the fact that
the conditional expectation is continuous in L2 . To prove this, recall that for
X, Y in L2 :
" # " #
⇣ ⌘2 ⇣ ⌘2
E E [ X |Ft ] E [Y |Ft ] =E E[X Y |Ft ]
h i h i
 E E [( X Y )2 |Ft ] = E ( X Y )2
Where the inequality is a consequence of Jensen’s inequality.
12 Stochastic Integration

Lemma 2.15 (Continuity of the integral). The process I ( X ) t

has almost surely
continuous sample paths.

The proof is based on the proof in Lamberton and Lapeyre 2007 (page 38,
Proposition 3.4.4).

Proof. Let X n be a sequence of simple processes converging to X in S2 . Then,

by the Doob inequality,
" Z #
h i ⇣ T ⌘2
2 n+ p
E sup I ( X n+ p )t I ( X n )t  4E Xs Xsn dWs
t T 0

hZ T i
n+ p 2
= 4E Xs Xsn ds !0
0 n!•

Thus, sup I ( X n+ p )t I ( X n )t converges to 0 in L2 (W, F , P ), so there exists a

t T
subsequence f(n) such that:

sup I ( X f(n+1) )t I ( X f(n) )t ! 0, a.e

t T n!•

Hence, taking a subsequence if necessary, the simple (and continuous) pro-

cesses I ( X f(n) ) converge uniformly to I ( X ) almost surely, so X is almost
surely continuous.

2.3 Extension of the integral

We would like to extend the integral to the bigger space:
Z T
S = { X adapted process, P | Xs |2 ds < • = 1}
0
hR i RT
T
If E 0 | Xs | ds < •, then the set {w, 0 | Xs (w )|2 ds = •} must have mea-
2
RT
sure zero, or otherwise the expectation of 0 | Xs |2 ds would be infinity. So
there is an inclusion S2 ✓ S .

The extension can be achieved by using a technique called localization. For

this, we will define the concept of a stopping time: a map t : W ! T such that
{t  s} 2 Fs for all s 2 T . Moreover, given a process X 2 S , we say that a
sequence of stopping times tn is localising for X in S2 if:

1. The sequence (tn )n2N is increasing

2. 8n, the stopped process Xtn = Xt 1[ttn ] 2 S2

3. lim tn = T, a.e.
n!•

Let X 2 S , and consider the increasing sequence of stopping times defined

by:
n Z t o
2
tn = in f t 2 [0, T ] | Xs | ds n
0
Defined with the convention in f ∆ = T. These are indeed stopping times. To
Rt
show this, we begin by observing that 0 | Xs |2 ds is measurable:

Rt
Lemma 2.16. Given a process X 2 S , 0 Xs2 ds is Ft -measurable.
2.3 Extension of the integral 13

Proof. Considering the integral as a Riemann integral, it can be expressed as a

Rt
limit of partial sums. Then it is clear that 0 Xs2 ds is measurable, being a limit
of measurable functions almost surely.

Now, for any t 2 T,

Z s Z t
2
{tn > t} = {w |8s  t, | Xu (w )| du < n} = {w | | Xu (w )|2 du < n}
0 0
Rs
Because 0 | Xu (w )|2 du is an increasing function of s. So tn is a stopping time.

The sequence tn is localising for X in S2 . Indeed, since X 2 S ,

Z T
a.e w 2 W, 9 Nw 2 N, | Xs (w )|2 ds < Nw
0
So that lim tn = T, a.e. Clearly, the process Xtn = Xt 1{ttn } is mean square
n!•
integrable, so we can define its integral as usual. Now, we would like to define
the integral process of X as the following limit:
Z t Z t
I ( X )t = Xs dWs := lim Xsn dWs , 0tT (2.3)
0 n!• 0

In order to prove that the integral is well defined, it is necessary to show that
the limit exists, almost surely. We state this result in the following theorem:

Theorem 2.17. For any process X 2 S , the integral process of X as defined in 2.3
exists almost surely.

The proof is based on the one by Lamberton and Lapeyre 2007. We will need
the following proposition:

Proposition 2.18. Let H 2 S2 and t be an Ft -stopping time. Then,

Z T
I ( H )t = 1{st } Hs dWs , a.s.
0

Proof. For any H 2 S2 , we define

Z T Z T Z t
Hs dWs := Hs dWs Hs dWs
t 0 0
Let A 2 Ft . The following property holds:
Z T Z T
1 A Hs 1{s>t} dWs = 1 A Hs dWs
0 t
This is clearly true for simple processes, and the property can be extended to
mean-square integrable processes by a density argument.
n
Now, let t = Â ti 1 Ai , where all Ai are disjoint and Fti -measurable. In this
i =1
case, for each i, the process 1 Ai 1{s>ti } is adapted because 1 Ai is Fs measurable
if s > ti , and the process is zero otherwise. Hence,

Z T Z T n n Z T

0
1{s>t } Hs dWs =
0
 1 A i 1{ s > t i } Hs dWs =  1 Ai 1{s>ti } Hs dWs
i =1 i =1 0
14 Stochastic Integration

n Z T Z T
= Â 1 Ai ti
Hs dWs =
t
Hs dWs
i =1

It follows that:
Z t Z T
Hs dWs = 1{st } Hs dWs
0 0
Now, consider an arbitrary stopping time t and define a decreasing sequence
tn by:

2n 1
( k + 1) T
tn = Â 2n
1{ kT t  (k+1)T }
2n 2n
k =0
Rt
Clearly, tn converges to t almost surely. Since the map t 7! 0 Hs dWs is
Rt Rt
almost surely continuous, 0 n Hs dWs converges to 0 Hs dWs almost surely.
Rt RT
By the previous discussion, we know that 0 n Hs dWs = 0 1{stn } Hs dWs for
all n 1. Now,

Z T Z T
! Z T
!
2
E 1{stn } Hs dWs 1{st } Hs dWs =E 1{t <stn } Hs2 ds
0 0 0

The last expression converges to 0 by the dominated convergence theorem.

Consequently,
Z tn Z T Z T
L2
Hs dWs = 1{stn } Hs dWs ! 1{st } Hs dWs
0 0 n!• 0
Rt RT
In particular, a subsequence of 0 n Hs dWs converges almost surely to 0 1{st } Hs dWs .
Rt Rt
This concludes the proof, because we also know that 0 n Hs dWs ! 0 Hs dWs
n!•
almost surely.
Rt
Proof of the Theorem. To see that 0 X n dW converges, note that because tn are
increasing, Xtn = 1{ttn } Xtn+1 . The previous proposition implies that
Z t Z t^tn
Xsn dWs = Xsn+1 dWs
0 0
RT
Hence, on the set {w, 0 Xs (w )2 ds < n},

I ( X n )t = I ( X m )t
for all m n (2.4)
RT S RT
Since {w, 0 Xs (w )2 ds < •} = {w, 0 Xs (w )2 ds < n} and the set has
n 2N
probability 1, the sequence I ( X n )t converges almost surely.

By construction, and by 2.4, the extended integral has almost surely continu-
ous sample paths. Note that the extended integral does not necessarily have
the martingale property. However, we will show that it is a local martingale.

Definition 2.19. An adapted process Xt is a local martingale if there exists a se-

quence (tn )n2N of increasing stopping times such that P (lim tn = T ) = 1 and such
n
that for each n, the stopped process X tn (t) := Xt^tn is a martingale.

Proposition 2.20. Let X 2 S . Then, the integral process I ( X )t is a local martingale.

2.3 Extension of the integral 15

Proof. The proof is based on the one by Capinski, Kopp, and Traple 2012
(p.140, Proposition 4.25).

Consider again the sequence of stopping times

n Z t o
2
tn = in f t 2 [0, T ] | Xs | ds n
0
Rt
Recall that I ( X )t = lim Mn (t), where Mn (t) := 0 Xsn dWs is a martingale and
n!•
Xtn = Xt 1{ttn } . Now,
Z T
tk
X (t) = X (t ^ tk ) = lim Mn (t ^ tk ) = lim 1[0,t^tk ] Xsn dWs
n n 0
Z t Z t
= lim 1[0,tk ] 1[0,tn ] Xs dWs = 1[0,tk ] Xs dWs = Mk (t)
n 0 0
As for n k, tn tk .

The following result will be useful in the next sections:

Proposition 2.21. Let M be a non-negative local martingale. Then, M is a super-

martingale. Moreover, if E [ Mt ] = M0 is constant, M is a martingale.

The proof is based on the one in Musiela and Rutkowski 2006 (p. 591, Propo-
sition A.7.1). We will need the following result (see Musiela and Rutkowski
2006, p.580, Lemma A.1.2 for a proof):

Lemma 2.22 (Conditional Form of Fatou’s Lemma). Let ( Xn )n2N be a sequence

of random variables in a probability space (W, F , P ) and let G be a sub-s-field of F .
Suppose that there exists a random variable Z, such that Xn Z for all n, and such
that E [ Z ] > •. Then,

E [lim inf Xn |G]  lim inf E [ Xn |G]

n n

Proof. (Proposition)
Since M is a local martingale, there exists an increasing sequence of stopping
times (tn )n2N such that P (lim tn = T ) = 1 and such that for each n, the
n
stopped process Mttn is a martingale.
Let 0  s  t  T. Because M is non-negative, and by the conditional form
of Fatou’s lemma,

E [ Mt |Fs ] = E [lim inf Mttn |Fs ]  lim inf E [ Mttn |Fs ]

n n

= lim inf Mstn = Ms

n
So that M is a supermartingale.

Note that in particular, E [ Mt ]  M0 < • for all t.

Finally, in the case of constant expectation, it is clear that if E [ Mt |Fs ]  Ms

a.s and E [E [ Mt |Fs ]] = E [ Mt ] = E [ Ms ], then E [ Mt |Fs ] = Ms , a.s.
16 Stochastic Integration
Chapter 3

Fundamental Theorems

The standard results that follow are basic for the development of the mathe-
matical finance theory.

Let (W, F , F = (Ft ), P ) be a filtered probability space, and Wt an Ft -Brownian

motion. We begin with the following definition:

Definition 3.1 (Itô Process). An Itô process is a process that satisfies the equation:
Z t Z t
X t = X0 + as ds + b s dWs
0 0
Where:

1. X0 is F0 -measurable.

2. at and b t are Ft -adapted and measurable processes.

RT
3. 0 |as |ds < •, a.s.
RT
4. 0 | b s |2 ds < •, a.s.

Equivalently, we say that the process satisfies the equation (in differential notation):

dXt = at dt + b t dWt

Lemma 3.2. The expression of an Itô process is unique.

Proof. Refer to Capinski, Kopp, and Traple 2012 (p. 98, Theorem 3.27).

3.1 Itô’s formula

Theorem 3.3 (Itô’s formula). Let g 2 C1,2 ([0, t] ⇥ R, R ) and ( Xt ) be an Itô pro-
cess, with:

dXt = at dt + b t dWt
Then Yt = g(t, Xt ) is also an Itô process, and satisfies the formula:

1
dYt = gt + gx at + gxx b2t dt + gx b t dWt (3.1)
2
Where all partial derivatives of g are evaluated at (t, Xt ).

For a proof of Itô’s formula, refer to Capinski, Kopp, and Traple 2012 (p. 136,
section 4.7). The following example will be particularly relevant:

17
18 Fundamental Theorems

Example 3.4. Let St be an Itô process such that dSt = µSt dt + sSt dWt , and let
g(t, x ) = log(x). Then, if Yt = g(t, St ),

1 2
dYt = (µ s )dt + sdWt
2
Or, equivalently:

1 2
St = exp (µ s )t + sWt
2
We will also require a two-dimensional version of Itô’s formula, adapted from
Musiela and Rutkowski 2006.

Theorem 3.5 (Itô’s formula, two dimensions). Let (Wt1 ), (Wt2 ) be Brownian
motions, g 2 C1,2 ([0, t] ⇥ R2 , R ) and ( Xt ), (Yt ) be Itô processes, with:

dXt = at dt + bt dWt1

dYt = at dt + b t dWt2
Then Zt = g(t, Xt , Yt ) is also an Itô process, and satisfies the formula:

1 1
dZt = gt + gx at + gy at + gxx bt2 + gyy b2t + gxy bt b t dt (3.2)
2 2

+ gx bt dWt1 + gy b t dWt2
Where all partial derivatives of g are evaluated at (t, Xt , Yt ).

3.2 Stochastic differential equations

In the next chapter we will frequently consider stochastic differential equa-
tions of the form:

dXt = a(t, Xt )dt + b(t, Xt )dWt (3.3)

X0 = x 0
In this situation, it is important to know if a solution to the equation exists,
and in that case, if it is unique. The following theorem, adapted from Capin-
ski, Kopp, and Traple 2012 (p.160, Theorem 5.8), provides sufficient conditions
for this to happen.

Theorem 3.6. [Existence and uniqueness of stochastic differential equations] Con-

sider the stochastic differential equation 3.3, and assume that the coefficient functions
a(t, x ), b(t, x ) are Lipschitz with respect to x and uniformly continuous with respect
to t. Moreover, assume that they have linear growth. This means that there exists a
constant C > 0 such that:

| a(t, x )| + |b(t, x )|  C (1 + | x |), 8 x 2 R, t 2 [0, T ] (3.4)

Then, 3.3 has a unique solution with continuous paths and that is mean square inte-
grable.

Proof. Refer to Capinski, Kopp, and Traple 2012 (p.160, Theorem 5.8).
3.3 Girsanov’s theorem 19

The natural extension of this result to multiple dimensions is also valid. That
is, consider a multi-dimensional version of 3.3,

dXt = a(t, Xt )dt + b(t, Xt )dWt (3.5)

X0 = x 0
Where ( Xt ) is a d-dimensional stochastic process (understood as d compo-
nents of one dimensional stochastic processes) and Wt is a d-dimensional
Brownian motion (understood as d components of one dimensional Brow-
nian motions), and x0 2 R d . Then, if the component functions a(t, x ), b(t, x )
are Lipschitz with respect to x and uniformly continuous with respect to t,
a unique regular solution to the stochastic differential equation exists (see
Musiela and Rutkowski 2006, p.639, Theorem A.3.1).

An important property that we would like the solutions of stochastic differ-

ential equations of the form 3.3 to satisfy is the Markov property.

Definition 3.7 (Markov Property). A stochastic process ( Xt ) in a filtered probabil-

ity space (W, F , F, P ) satisfies the Markov property if, for any bounded measurable
function f : R ! R and any s  t,

E [ f ( Xt )|Fs ] = E [ f ( Xt )|F Xs ] (3.6)

Where (F Xt ) is the filtration generated by ( Xt ).

In our case, (Ft ) is the filtration generated by the Brownian motion (Wt ).
The following result guarantees that under certain conditions on the coeffi-
cient functions, this property is satisfied by solutions of stochastic differential
equations. It has been adapted from Capinski, Kopp, and Traple 2012 (p.174,
Theorem 5.14).

Theorem 3.8. [Markov Property] Consider a stochastic differential equation of the

form 3.3, with coefficients a(t, x ) and b(t, x ) that are Lipshitz continuous with respect
to x, uniformly continuous with respect to t, and satisfy the linear growth condition
3.4. Then, the solution Xt has the Markov property 3.6. That is,

E [ f ( Xt )|FWs ] = E [ f ( Xt )|F Xs ]

Proof. Refer to Capinski, Kopp, and Traple 2012 (p.174, Theorem 5.14).

3.3 Girsanov’s theorem

Theorem 3.9 (Girsanov, simple version). Let (Wt ) be a standard Brownian Mo-
tion and g 2 R. Then, W̃t := Wt + gt is a standard Brownian motion under the
probability P̃ defined by:

dP̃ 1 2
= exp gWT g T
dP 2
Proof. The proof has been adapted from Capinski and Kopp 2012.

It is clear that W̃ has continuous sample paths. We will prove directly that
W̃t has independent and normally distributed increments, by computing the
20 Fundamental Theorems

probability P̃ ( A) where A = \in=1 Ai and Ai = {W̃ti W̃ti 1  ai }, for any

given partition t0 , . . . , tn of T , and every a1 , . . . , an 2 R. By definition,

⇥ ⇤ h 1 2 i
P̃ ( A) = EP̃ 1 A = EP exp gWT g T 1A
2
Now, WT can be expressed as Âin=1 Wti Wti 1 . Similarly, T = Âin=1 (ti ti 1)
n
and 1 A = ’ 1 Ai . We obtain:
i =1
" #
n
1 2
P̃ ( A) = EP ’ exp g(Wti Wti 1 )
2
g ( ti ti 1) 1 Ai
i =1

The increments Wti Wti 1

are independent, and the indicators sets Ai can be
expressed as:

Ai = {Wti Wti 1
+ g ( ti ti 1)  ai }
So, because of the independence property,

n h 1 2 i n
P̃ ( A) = ’ P exp
E g(Wti Wti 1 )
2
g ( ti ti 1 ) 1 Ai = ’ P̃( Ai ) (3.7)
i =1 i =1

Now, because the increments Wti Wti 1

are distributed as N (0, ti ti 1 ),

h 1 2 i
P ( Ai ) = EP exp g(Wti Wti 1 ) g ( ti ti 1 ) 1 Ai
2
Z
1 2 1 x2
= exp gx g ( ti ti 1 ) p exp dx
{ x + g ( ti ti 1 ) ai } 2 2p (ti ti 1) 2( t i t i 1)

Z ⇣ 2⌘
1 x g ( ti ti 1 )
= p exp dx
{ x + g ( ti ti 1 ) ai } 2p (ti ti 1) 2( t i t i 1 )
Z ⇣ ⌘
1 z2
= p exp dz (3.8)
{ z  ai } 2p (ti ti 1) 2( t i t i 1)

Equation 3.8 proves that W̃ti W̃ti 1 are distributed as N (0, ti ti 1 ), and
equation 3.7 proves that they are independent.

A generalized version of Girsanov’s theorem will be required in the Stochastic

Volatility chapter. The theorem has been adapted from Musiela and Rutkowski
2006 (p.648, Theorem A.15.1).

Theorem 3.10 (Girsanov’s theorem, a generalized version). Let (Wt ) be a d-

dimensional Brownian motion in the filtered probability space (W, F , F, P ), and g be
an R d -valued F-adapted stochastic process, such that:
" Z T Z T
#
⇣ 1 ⌘
E exp gs dWs gs2 ds =1 (3.9)
0 2 0

Rt
Define W̃t := Wt + 0 gs ds, and define the probability P̃ by:
⇣ Z T Z T ⌘
dP̃ 1
= exp gs dWs gs2 ds (3.10)
dP 0 2 0

Then, (W̃t ) is a Brownian motion under the probability P̃.

3.4 Martingale representation theorem 21

3.4 Martingale representation theorem

In this section, we will prove the martingale representation theorem. These re-
sults have been adapted from Corcuera n.d., whose notations and arguments
will be followed closely. We will need some preliminary results.

Theorem 3.11 (Martingale L p -convergence). Let ( Mn )n2N be a martingale in a

filtered probability space (W, F , F = (Fn ), P ). Assume that, for some p > 1, Mn 2
L p = L p (W, F , P ) 8n 2 N, and sup k Mn k L p < •. Then, for some M• 2 L p ,
n 2N
Lp
Mn ! M•
n!•

Proof. Refer to Kallenberg 2006 (p.109, Corollary 6.22).

Lemma 3.12. Let (Gn ) be a filtration in a probability space (W, F , P ) and let X 2
L2 = L2 (W, F , P ). Then,

L2
E [ X |Gn ] ! E [ X |G• ]
n!•
Where G• = s(Gn , n 2 N ).

Proof. Define Xn := E [ X |Gn ]. Clearly, Xn is a martingale with respect to the

filtration (Gn ), and by the properties of the conditional expectation,

sup k Xn k L2  k X k L2 < •
n 2N
By theorem 3.11,

L2
Xn !Y
n!•

For some Y 2 L2 . It remains to show that Y = E [ X |G• ] =: X• . By the

continuity of the conditional expectation in L2 (this was proved in 2.14),

E [Y |Gn ] = E [ lim Xm |Gn ] = lim E [ Xm |Gn ] = Xn

m!• m!•
By the tower property of the conditional expectation, we also have that
h i
E [ X• |Gn ] = E E [ X |G• ] Gn = Xn
So that, for all n 2 N,

E [Y |Gn ] = E [ X• |Gn ]
Hence, for every G 2 [n2N Gn ,

E [Y1G ] = E [ X• 1G ] (3.11)
Define the collection

C = { G 2 G• | E [Y1G ] = E [ X• 1G ]} (3.12)
Clearly, by 3.11,

[n2N Gn ⇢ C ⇢ G• (3.13)
We wish to show that C = G• .
22 Fundamental Theorems

Now, if we prove that C is a s-algebra, we will have, by equation 3.13 and the
fact that [n Gn generates G• , that C = G• .

Condition (1) is clear since X 2 [n2N Gn ⇢ C . For condition (2), if B ⇢ A 2 C ,

E [Y1 A\ B ] = E [Y1 A ] E [Y1B ] = E [ X• 1 A ] E [Y1B ] = E [ X• 1 A\ B ]

So that A \ B 2 C . For condition (3) note that, if ( An )n2N is a sequence in C ,
by the dominated convergence theorem,

E [Y1[n An ] = E [lim Y1[nm=1 Am ] = lim E [Y1[nm=1 Am ]

n n
= lim E [ X• 1[nm=1 Am ] = E [ X• 1[n An ]
n
Hence, [n An 2 C , and C = G• .

Now, Bn := { X• Y > n1 } 2 G• , so

E [ X• 1Bn ] = E [Y1Bn ]
This implies that P ( Bn ) = 0 8n 2 N, so that P ([n Bn ) = 0. Hence X•  Y a.s,
and a similar argument shows that X• Y a.s.

Lemma 3.13. Let (Wt ) be a Brownian motion in the filtered probability space (W, F , F, P ),
where F is generated by (Wt ). Consider the set J of stepwise functions f : [0, T ] !
R of the form:
n
f = Â l i 1] t i 1 ,ti ]
i =0
With li 2 R and 0 = t0 < . . . < tn = T. For each f 2 J , define
nZ T 1
Z T o
f
ET = exp f (s)dWs f 2 (s)ds
0 2 0
f
Let Y 2 L2 (F T , P ), and assume that Y is orthogonal to E T for all f 2 J . Then,
Y = 0.
f
Proof. Let f 2 J and Y 2 L2 (F T , P ) orthogonal to E T . Define Gn := s(Wt0 , . . . , Wtn ).
By assumption,
!
n n o
E exp  li Wti Wti 1
Y =0
i =1

Taking the conditional by Gn we obtain:

!
n n o
E exp  li Wti Wti 1
E [Y |Gn ] =0
i =1

Let X : W ! Rn be defined by:

X = (Wt1 , Wt2 Wt1 , . . . , Wtn Wtn 1 )

Decomposing Y as Y = Y+ Y , we get:

! !
n n o n n o
E exp  li Wti Wti 1
E [Y+ |Gn ] = E exp  li Wti Wti 1
E [Y |Gn ]
i =1 i =1
(3.14)
3.4 Martingale representation theorem 23

Applying the pushforward

h measure theorem to 3.14, i and because of the fact
that E [Y |Gn ] = E Y W1 = a1 , . . . , Wn Wn 1 = an :
a1 =W1 ,...,an =Wn Wn 1
Z n n o
exp  li xi E [Y+ |Gn ]( x1 , . . . , xn )dP X ( x1 , . . . , xn ) (3.15)
Rn i =1
Z n n o
=
Rn
exp  i i E[Y |Gn ](x1, . . . , xn )dP X (x1, . . . , xn )
l x (3.16)
i =1
Note that 3.15 and 3.16 are, respectively, the Laplace transforms of E [Y+ |Gn ]( x1 , . . . , xn )
and E [Y |Gn ]( x1 , . . . , xn ) with respect to the measure P X . We admit the result
on the uniqueness of the Laplace transform, which in this case implies that:

E [Y+ |Gn ]( x1 , . . . , xn ) = E [Y |Gn ]( x1 , . . . , xn ), P X a.s (3.17)

So that E [Y+ |Gn ]( x1 , . . . , xn ) = E [Y |Gn ]( x1 , . . . , xn ) for all ( x1 , . . . , xn ) 2 R n \
A for some set A ⇢ R n with P X ( A) = 0. This means that E [Y+ |Gn ](w ) =
E [Y+ |Gn ](w ), 8w 2 X 1 ( A). Finally,

E [Y+ |Gn ] = E [Y+ |Gn ] P a.s. (3.18)

Since this is true for any Gn as defined above, by 3.12 we have that

E [Y± |Gn ] ! E [Y± |s(Gn ; n 2 N )] = Y±

n!•
Hence, Y = 0.

Proposition 3.14. Let (W, F , F = (Ft ), P ) be a filtered probability space with F =

FT , let (Wt ) be a Brownian motion and let F 2 L2 (W, FT , P ). Then, there exists an
adapted, mean square integrable process (Yt ) such that
Z T
F = E[ F] + Yt dWt
0

Proof. Consider the Hilbert space H of centered random variables in L2 (W, F T , P ),

RT
and its subspace I consisting of the random variables of the form 0 Yt dWt ,
for some adapted, mean square integrable process (Yt ). Note that proving the
proposition is equivalent to proving that I = H.

If, on the contrary, I ( H, then there would exist a centered, non-trivial ran-
dom variable Z 2 H orthogonal to I . We will prove that this is not possible.

f f
Indeed, suppose that such Z exists. Take Yt := Et , with Et as defined in
Lemma 3.13. Then,
h Z T i
f
E Z· Et dWt =0
0
And also
h Z T i
f
E Z· 1+ Et dWt =0 (3.19)
0
f
Now, a similar argument as the one in 3.4 shows that Et is the solution to the
following stochastic differential equation:

f f
dEt = f (t)Et dWt
24 Fundamental Theorems

In particular,
Z T
f f
ET = 1+ f (t)Et dWt (3.20)
0
From 3.19 and 3.20 we conclude:

f
E[ZET ] = 0 (3.21)
Since this is true for any f defined as in 3.13, by 3.13 we conclude that Z =
0.

Theorem 3.15 (Martingale Representation Theorem). Let ( Mt ) be a square inte-

grable martingale with respect to the filtration (Ft ), and (Wt ) a Brownian motion.
Then, there exists an adapted, mean square integrable process ( Xt ) such that, in
differential notation:

dMt = Xt dWt , a.s

Proof. Applying the previous proposition applied to MT , we obtain the exis-

tence of a mean square integrable process (Yt ) such that:
Z T Z T
MT = E [ MT ] + Yt dWt = M0 + Yt dWt
0 0
Since ( Mt ) is a martingale, for any t 2 [0, T ] we have:
h Z T i Z t
Mt = E [ MT |Ft ] = E M0 + Yt dWt Ft = M0 + Yt dWt
0 0
Rt
Where we have used the fact that 0 Yt dWt is a martingale (see 2.14).
Chapter 4

The Black-Scholes Model

The Black-Scholes model consists of two stocks S and b, in a time frame [0, T ],
where b is a bank account such that b(t) = ert , and the risky stock S satisfies
the differential equation:

dSt = µSt dt + sSt dWt (4.1)

Where Wt is a Brownian motion in the filtered probability space (W, F , F, P ),

and P represents the so-called empirical or physical probability. The dis-
counted stock price is defined as St⇤ = bt 1 St .

Define F = (Ft ) to be the filtration generated by the Brownian motion (Wt ).

We assume that F0 = {∆, W} and F T = F . The coefficients of the stochastic
differential equation 4.1, according to the notations in 3.3, are

a(t, x ) = µx (4.2)

b(t, x ) = sx

These functions clearly satisfy the regularity conditions in 3.6, so there exists
a unique mean-square integrable solution St with continuous sample paths.
By theorem 3.8, this solution also satisfies the Markov property.

An explicit solution for this stochastic differential equation was found in 3.4:

1 2
St = S0 exp{ (µ s )t + sWt } (4.3)
2
The following are some implicit assumptions of this model, as argued in Cor-
cuera n.d.:

• it has continuous trajectories

St Su
• its returns Su are independent of s(Ss , 0  s  u). Indeed,

St Su St 1 2
= 1 = exp{ (µ s )(t u) + s(Wt Wu )} 1
Su Su 2

Which is independent of s(Ss , 0  s  u).

• its returns are stationary:

St Su St u S0
⇠
Su S0

25
26 The Black-Scholes Model

A trading strategy or strategy is a random vector f = (f1 , f2 ) with values in R2 ,

adapted to the filtration (Ft )0tT . The value of the strategy is the process
Vf (t) := ft · (b(t), St ).
A strategy is self-financing if:

dVf (t) = ft1 db(t) + ft2 dSt (4.4)

This is a natural extension of the discrete time expression Dn Vf = fn1 Dn b +

fn2 Dn S. The discounted value process is defined as Vf⇤ (t) = ft · (1, St⇤ ), where
St⇤ = St /b(t) is the discounted stock price.
For 4.4 to make sense, we need to require that:
RT
1. 0 |ft1 |dt < •, a.s
RT
2. 0 (ft2 )2 dt < •, a.s

We will denote be the set of self-financing strategies as F. We will discuss

further restrictions on the set of strategies considered in the model later on.

The following characterization of self-financed strategies will be useful:

Lemma 4.1. Let f be a strategy satisfying integrability conditions (1) and (2). Then,
f 2 F iff

dṼt (f) = ft2 dS̃t

Proof. The proof has been adapted from Lamberton and Lapeyre 2007 (p. 65,
Proposition 4.1.2).

Suppose that f is self-financing. The Itô formula gives

rt
dṼf (t) = rṼf (t)dt + e dVf (t)

Imposing the self-financing condition, we obtain:

dṼf (t) = re rt
(ft1 ert + ft2 St )dt + e rt
(ft1 db(t) + ft2 dSt )

= ft2 ( re rt
St dt + e rt
dSt ) = ft2 dS̃t

The converse can be proved similarly.

4.1 Risk-neutral measure

A risk-neutral measure is a measure P̃ equivalent to P such that the discounted
stock price is a martingale under P̃. The following results guarantee the exis-
tence and uniqueness of a risk-neutral measure. We begin with the following
lemma:

1 2
Lemma 4.2. Given a Brownian motion (Wt ), the process Mt := S0⇤ exp{ 2s t +
sWt } is a martingale.
4.1 Risk-neutral measure 27

Proof. For any s < t,

h i h 1 2 i
E Mt /Ms |Fs = E exp{ s (t s) + s (W̃t W̃s )}|Fs
2
h 1 i
= E exp{ s2 (t s) + s (W̃t W̃s )} (because W̃t W̃s ?Fs )
2
Now,

h 1 2 i
E exp{ s (t s) + s(W̃t W̃s )}
2
Z n
1 2 1 x2 o
= exp s (t s) + sx p exp dx
R 2 2p (t s) 2( t s )
Z n
1 (x s (t s))2 o
= p exp dx = 1
R 2p (t s) 2( t s )

Proposition 4.3. A risk neutral measure exists, defined by:

⇣ dQ ⌘ ⇣ µ r 1 (r µ )2 ⌘
= exp WT⇤ T (4.5)
dP t s 2 s2
µ r
Where the process Wt⇤ := Wt + s t is a Brownian motion under P.

Proof. In view of Girsanov’s theorem, we seek a value g such that by defin-

ing Wt⇤ := Wt + gt, the stock price is a martingale under the probability Q,
defined as in Girsanov’s theorem. The stock price evolves according to:

dSt⇤ = (µ r + gs )St⇤ dt + sSt⇤ dWt⇤

It is a martingale under Q iff the drift term µ r + gs = 0.

Indeed, let r = µ r + gs. Then,

1 2 1 2
St⇤ = S0⇤ exp{(r s )t + sWt⇤ } = exp{rt}S0⇤ exp{ s t + sWt⇤ }
2 2
1 2
The previous lemma implies that exp{ 2s t + sWt⇤ } is a martingale under
Q, which proves the result.

Remark. The stock price St evolves according to the equation:

dSt = rSt dt + sSt dWt⇤ (4.6)

Theorem 4.4. A unique neutral measure Q exists.

Proof. Refer to Capinski and Kopp 2012 (p. 51, Theorem 3.12)
28 The Black-Scholes Model

4.2 Arbitrage and admissibility

An arbitrage opportunity is an opportunity to have positive returns in an in-
vestment with no risk. More specifically, an arbitrage strategy is defined as
following:

Definition 4.5. An arbitrage strategy is a self-financing strategy f = (f1 , f2 ) such

that Vf (0) = 0, Vf ( T ) 0, a.s and P (Vf ( T ) > 0) > 0.

The following example shows that the market model described above, with
the set of self-financing strategies, has arbitrage opportunities. In order to
eliminate these opportunities, it will be necessary to restrict the set of so-
called admissible strategies, introduced below.

Theorem 4.6. Arbitrage opportunities exist within the class of self-financing strate-
gies.

This example is based on the suicide strategy described in Capinski and Kopp
2012 (p.24-27).

We begin by considering the strategy f given by:

1
ft2 = p
sS̃t T t
The risk free component is determined by the self-financing condition and
an (arbitrary) initial value. Now, the strategy is not almost surely square
integrable, but it will be modified later on. Firstly, we will study its properties.

Lemma 4.7. For each M 0,

P (min{t : Ṽf (t) Ṽf (0) M}  T ) = 1

P (min{t : Ṽf (t) Ṽf (0)  M}  T ) = 1

Proof.
1
dṼf (t) = ft2 dS̃t = ft2 sS̃t dWt⇤ = p dWt⇤
T t
So that:
Z t
1
Ṽf (t) Ṽf (0) = p dWu⇤
0 T u
Now, let
Z t
1
g(t) = p du
0 T u
Then, Ṽf (t) Ṽf (0) and W ⇤ (g(t)) have the same distribution. A proof of this
result can be found at Capinski and Kopp 2012 - here we will focus on its
consequences. Recall that - see Lemmma 1.8, for any fixed M > 0:

Q (max Wt⇤  M) = 0
t 0

Q (min Wt⇤ M) = 0
t 0

And since P and Q are equivalent,

4.2 Arbitrage and admissibility 29

P (max Wt⇤  M) = 0
t 0

P (min Wt⇤ M) = 0
t 0

Clearly, g(t) ! •, so that

t! T

P (max Ṽf (t)  M) = P (max W ⇤ (g(t))  M) = P (max W ⇤ (t)  M) = 0

t T t T t 0

P (min Ṽf (t) M ) = P (min W ⇤ (g(t)) M) = P (min W ⇤ (t) M) = 0

t T t T t 0

Proof. (Theorem)

Let Ṽf (0) = 1 and take M = 1, so that:

P (min{t : Ṽf (t)  0}  T ) = 1

Let

t = min{t : Ṽf (t) = 0}  T, a.s

Define the self-financing strategy q as:

qt2 = ft2 1{tt }

And such that Vq (0) = 1. Since t  T almost everywhere, it is clear that
Z T
(qt2 )2 dt < •, a.e
0
So that q 2 F. This suicide strategy begins with a positive wealth and ends
almost surely in bankrupcy. As we will see, this cannot be admissible. In-
deed, we can construct a new strategy from q that is an arbitrage opportunity.
Define:

c1t = qt1 + 1
c2t = qt2

Then Vc (0) = Vq (0) + 1 = 0 and Vc ( T ) = Vq ( T ) + erT = erT . Clearly c 2

F, because, just like q, it is self-financing and satisfies integrability conditions.
The previous properties of c show that it is an arbitrage opportunity.

We will now define which strategies are considered admissible in the model.
Although several alternatives have been proposed in the literature, here we
follow the definition from:

Definition 4.8. A self-financing strategy f is admissible if it is bounded by below.

We will denote the set of admissible strategies as F0 ⇢ F.

The next result guarantees that no arbitrage opportunities exist within the
admissible strategies:
30 The Black-Scholes Model

Proposition 4.9. No admissible strategy is an arbitrage opportunity.

Proof. Suppose that f 2 F0 is an admissible strategy and an arbitrage oppor-

tunity - we will argue by contradiction. In that case, Vf (t) is a local martingale
with respect to Q which is bounded by below, so there exists a constant L such
that Vf (t) + L is non-negative. Hence, Proposition 1.21 implies that Vf (t) is
a super martingale under Q. In particular, EQ [Vf ( T )]  Vf (0) = 0, which
combined with the fact that P (Vf ( T ) > 0) > 0 and that P and Q are equiv-
alent, implies that P (Vf ( T ) < 0) > 0, which condradicts the assumption
Vf ( T ) 0.

Definition 4.10 (European Option). A European option with maturity T is defined

by a non-negative, F T -measurable random variable H = h(ST ), that expresses its
payoff.

Definition 4.11. A strategy f 2 F0 replicates the derivative with payoff H if H =

Vf ( T ). The market model is complete if every European option can be replicated.

We are now able to state a fundamental result in the Black-Scholes model,

relating the price process of a derivative to the value process of the replicat-
ing strategy. The proof of the result will be outlined - however, substantial
technical details that have been omitted can be found at Capinski and Kopp
2012.

Theorem 4.12. Let H be the payoff of a derivative which is replicated by the strategy
f. Assuming that the option price is an Itô process, the No Arbitrage Principle
implies that the price of the option at time t, Vt , is equal to Vf (t) for all t 2 T.

Proof. This proof is based on the one in Capinski and Kopp 2012 (p.21, Theo-
rem 2.16). Assume that this is not the case, and let t0 be any time in which a
difference between Vf (t) and Vt appears with positive probability. Consider
a strategy y with zero initial value and that buys the cheaper of the two and
sells the most expensive short at time t0 and invests the remaining money in
the bank account. The value Vy ( T ) of this strategy is positive with a positive
probability. Indeed, assuming without a loss of generality that Vf (t0 ) > Vt0 ,

Vy ( T ) = Vf (t0 ) Vt0 er(T t0 )

Vf ( T ) + VT (4.7)
= Vf (t0 ) Vt0 er(T t0 )
(because VT = H = Vf ( T )) (4.8)
Which is greater than zero with positive probability. Now, if we show that y
is an admissible strategy, this would violate the No Arbitrage Principle.

The strategy is self-financing by construction, and it remains to show that it

is bounded by below. The reader can refer to Capinski and Kopp 2012 (p.30,
Theorem 2.16) for a proof of this detail.

4.3 Completeness
We aim to prove that any square integrable European option can be replicated
by an admissible strategy. The proof of the following theorem is based on the
proof in Capinski and Kopp 2012:

Theorem 4.13 (A Completeness Theorem). For any square integrable European

option H with respect to Q, there exists an admissible strategy f that replicates H.
4.3 Completeness 31

Proof. We will see that, in fact, such a f exists that satisfies the additional
property that Vf⇤ (t) is a martingale under Q, and
h i
Vf (t) = b(t)Vf⇤ (t) = b(t)EQ H ⇤ (t)|Ft (4.9)

Where H ⇤ (t) = H/b(t). Hence we seek an admissible strategy f with value

given by the previous expression. The martingale representation theorem
guarantees the existence of an adapted, mean square integrable process X (t)
such that:
⇣ h i⌘
⇤
d EQ H (t)|Ft = X (t)dWt⇤ (4.10)

Returning to 4.9 and differentiating on both sides, we have:

dVf (t) = ft1 db(t) + ft2 dSt = ft1 rb(t)dt + ft2 (rSt dt + sSt dWt⇤ ) (Self-financing condition)

⇣ h i⌘ ⇣ h i⌘ h i
d b(t)EQ H ⇤ (t)|Ft = b(t)d EQ H ⇤ (t)|Ft + EQ H ⇤ (t)|Ft db(t) (Itô Formula)

h i
= b(t) X (t)dWt⇤ + rb(t)EQ H ⇤ (t)|Ft dt

Equating both sides and rearranging, we obtain:

" # " #
h i
0 = rb(t)ft1 + rft2 St rb(t)EQ H ⇤ (t)|Ft dt + ft2 sSt b(t) X (t) dWt⇤

By the uniqueness of the expression of an Itô process,

h i
rb(t)ft1 + rft2 St ⇤
rb(t)EQ H (t)|Ft = 0

ft2 sSt b(t) X (t) = 0 (4.11)

Isolating ft2 in the second equality gives:

b(t) X (t)
ft2 = (4.12)
sS(t)
And substituting in the first equality gives:
h i
ft1 = EQ H ⇤ (t)|Ft X (t)/s

We have obtained a unique strategy that may attain H - now we need to verify
all the conditions in the theorem. Firstly, the following calculation shows that
f attains H:

h i h i
Vf (t) = ft ⇤ (b(t), St ) = b(t))EQ H ⇤ (t)|Ft b(t) X (t)/s + b(t) X (t)/s = b(t)EQ H ⇤ (t)|Ft

In particular, Vf ( T ) = H.
h i
The strategy is admissible because Vf⇤ (t) = EQ H ⇤ |Ft is clearly a martingale.

For the self-financing condition,

32 The Black-Scholes Model

dVf⇤ (t) = X (t)dWt⇤

Note that, from 4.12, we have:

X (t) = sft2 St⇤

So that:

dVf⇤ (t) = sft2 St⇤ dWt⇤ = ft2 dSt⇤

4.4 Pricing and hedging

We have shown that the Black Scholes model has no arbitrage opportunities
and is complete in the sense that every european option that is square in-
tegrable with respect to Q is replicable. We also know that the price of a
replicable european option at any time is determined by the conditional ex-
pectation of the value of the replicating strategy at the given time, under the
risk neutral probability. In this section, we will develop pricing formulas for
european options and to obtain the replicating strategy.

Let H be the payoff of a European option with maturity T. We will assume

that H = h(ST ) for some function h. For a call option h( x ) = ( x K )+ and
for a put option h( x ) = (K x )+ . Recall that if f is an admissible strategy
that replicates H, then the value of the option at time t < T is equal to the
value of the strategy f at time t, Vf (t) (see Theorem 4.12). In particular, if Vt
denotes the value of the option:

Vt = ert Ṽf (t) = ert EQ [Ṽf ( T )|Ft ] = ert EQ [e rT

H |Ft ] H |Ft ] = EQ [e r ( T t)

(4.13)
Note that H should be square-integrable with respect to Q, due to the condi-
tions in Theorem 4.13. In that case, the theorem guarantees the existence of a
replicating strategy f, validating the previous argument. Now, if H = h(ST ),
with h : R ! R being a bounded, measurable function, we can express Vt
as a function of St and t - the condition that h is bounded will be needed to
apply the Markov property. Indeed, following the arguments in Lamberton
and Lapeyre 2007 (p. 69, Remark 4.3.3),

" #
⇣ ⌘
r ( T t) r ( T t) s2 /2)( T t)+s(WT⇤ Wt⇤ )
Vt = EQ [e H |Ft ] = EQ e h St e (r Ft

Note that St is Ft -measurable, and WT⇤ Wt⇤ is independent of Ft under Q.

Hence,
" #
⇣ ⌘
r ( T t) (r s2 /2)( T t)+s(WT⇤ Wt⇤ )
Vt = EQ e h St e Ft

" #
⇣ ⌘
r ( T t) s2 /2)( T t)+s(WT⇤ Wt⇤ )
= EQ e h St e (r St (because of the Markov property)
4.4 Pricing and hedging 33

" #
⇣ ⌘
r ( T t) s2 /2)( T t)+s(WT⇤ Wt⇤ )
= EQ e h xe(r St = x
x = St
" #
⇣ ⌘
r ( T t) (r s2 /2)( T t)+s(WT⇤ Wt⇤ )
= EQ e h xe
x = St
Where the last equality is a consequence of the fact that Wt⇤ is inde- if WT⇤
pendent of Ft , then it is independent of St . Now, we can express Vt as:

Vt = P(t, St ) (4.14)

Where
" #
⇣ ⌘
r ( T t) s2 /2)( T t)+s(WT⇤ Wt⇤ )
P(t, x ) = EQ e h xe(r (4.15)

And, since WT⇤ Wt⇤ is distributed as N (0, T t) under Q, the previous ex-
pectation is expressed by the following integral:
Z • ⇣ p ⌘ y2 /2
r ( T t) (r s2 /2)( T t)+sy T t ep
P(t, x ) = e h xe dy (4.16)
• 2p
In the case of calls and puts, P(t, x ) can be calculated explicitely, giving rise
to the Black-Scholes pricing formulas. Firstly, we will need to prove that call
and put options are square integrable with respect to Q. The case of a put
option with payoff H = h(ST ) = (K ST )+ , this is clear since the payoff is
bounded by K. We admit the following result, which shows that this is also
the case for a call option, with payoff H = h(ST ) = (ST K )+ .

Lemma 4.14. The payoff H = h(ST ) = (St K )+ of a call option is square inte-
grable with respect to Q.

Proof. Refer to Capinski and Kopp 2012 (p.55-56, Call options).

4.4.1 Pricing a put option

We can apply the results obtained in 4.15 to the case of a put option with
h( x ) = (K x )+ , since the function h is bounded. In this case,
" #
⇣ ⌘+
r ( T t) (r s2 /2)( T t)+s(WT⇤ Wt⇤ )
P(t, x ) = EQ e K xe

WT⇤ Wt⇤
Let q = T t and z = p , a standard normal random variable under Q.
q
Then, " #
⇣ p ⌘+
rq s2 q/2
P(t, x ) = E Ke xes qz

Define:

log(K/x ) (r s2 /2)q
d= p (4.17)
s q
p
Note that Ke rq xes qz s2 q/2 0 () z  d. We obtain:
" #
⇣ p ⌘
rq s2 q/2
F (t, x ) = E Ke xes qz
1{ z  d }
34 The Black-Scholes Model

Z d ⇣ p ⌘ y2 /2
rq s qy s2 q/2 e
= Ke xe p dy
• 2p
Z d y2 /2 Z d p y2 /2
rq e s qy s2 q/2 e
= Ke p dy xe p dy
• 2p • 2p
The value of the first integral is Ke rq times the cumulative distribution func-
tion of a standard normal random variable evaluated p at d: N (d). For the
second integral,
p the change of variable t = y + s q shows that its value is
xN (d s q ). Finally,

r ( T t)
p
P(t, x ) = Ke N (d) xN (d s T t) (4.18)

Figure 4.1: Black-Scholes pricing of a put option at time t = 0, with K = 100,

T = 2, r = 0.04 and s = 0.1.

4.4.2 Pricing a call option

A call option has a payoff H = h(ST ), where h( x ) = ( x K )+ . Because h is
not bounded, we can’t use the results in 4.15 to obtain a price for the option,
as in the case of put options. However, the price will be obtained by means of
a relationship between the price of a call option and the price of a put option
for a given strike price K, known as the call-put parity. The following theorem
has been adapted from Capinski and Kopp 2012 (p.56, Theorem 3.16):

Theorem 4.15 (Call-put parity). Let Ct , Pt be the price of a call option and a put
option, respectively, at time t, both with strike price K and maturity T. Then,

r ( T t)
Ct Pt = St Ke (4.19)
4.4 Pricing and hedging 35

Proof. We follow closely the arguments and notations in Capinski and Kopp
2012 (p.56, Theorem 3.16).

Note that

ST (ST K )+ + (K ST )+ = K (4.20)
This equation can be deduced by separately considering the cases ST K 0
and ST K  0. Hence,

ST CT + PT = K (4.21)
Multiplying both sides by e rT ,

rT rT rT
ST⇤ e CT + e PT = Ke (4.22)
Now, (St⇤ ), (e rt Ct ), and (e rt Pt ) are Q-martingales (this assertion relies on
the fact that call and put options are replicable), so

rT rT rT rT rt rt
e = E [e |Ft ] = E [ST⇤ e CT + e PT |Ft ] = St⇤ e Ct + e Pt
(4.23)
The result follows by multiplying both sides by ert .

The price of a call option follows from 4.18 and the call-put parity. Indeed,

r ( T t)
Ct = Pt + St Ke (by 4.19)

r ( T t)
p r ( T t)
= Ke N (d) St N ( d s T t ) + St Ke (by 4.18)

p r ( T t)
= St (1 N (d s T t)) Ke (1 N (d))

p r ( T t)
= St N ( d + s T t) Ke N ( d) (4.24)
In particular, the price of a call option is a function of t and St :

Ct = C (t, St ) (4.25)
Where
p r ( T t)
C (t, x ) = xN ( d + s T t) Ke N ( d) (4.26)

4.4.3 Hedging
Now that the pricing formulas have been justified, we aim to obtain explicit
formulas for replicating (or hedging) strategies in the same setting of a Eu-
ropean option with an F T -measurable payoff H = h(ST ) which is square
integrable with respect to Q. We follow the arguments in Fouque, Papanico-
laou, and Sircar 2000, which begin by recalling that the value of the option is
equal to:

Vt = P(t, St ) (4.27)
The fact that Vt is a function of t and St was proven in 4.15 assuming that h
was a bounded function. However, we saw in 4.25 that this is also true for
36 The Black-Scholes Model

Figure 4.2: Black-Scholes pricing of a call option at time t = 0, with K = 100,

T = 2, r = 0.04 and s = 0.1.

call options. Hence, for the following, we assume that either h is bounded or
h( x ) = ( x K )+ .

Now, let f be a replicating strategy for H. Such a strategy exists by theorem

4.13. By Theorem 4.12, P(t, St ) is equal to the value of the strategy Vf (t):

ert ft1 + ft2 St = P(t, St ) (4.28)

Differentiating on both sides, and applying the self-financing condition on the

left side, we obtain:

1
(rert ft1 + ft2 µSt )dt + ft2 sSt dWt = ( Pt + µSt Px + s2 St2 Pxx )dt + sSt Px dWt
2
(4.29)
Where all the partial derivatives of P are evaluated at (t, St ) - we will use
this abbreviation later on without further mention. By the uniqueness of the
expression of an Itô process, we conclude that:

ft2 = Px (t, St ) (4.30)

And from 4.28 we deduce that:

ft1 = e rt
( P(t, St ) Px (t, St )St ) (4.31)

The previous argument also shows the relationship between the pricing func-
tion P(t, x ) and a certain PDE (the Black-Scholes PDE). Indeed, substituting
the value of the hedging strategy in 4.29, we obtain the formula:

r(P Px St + Px µSt )dt + Px sSt dWt (4.32)

4.4 Pricing and hedging 37

1
= Pt + Px µSt + s2 St2 Pxx dt + sSt Px dWt
2
Equating the drift terms in both expressions, we obtain the equation:

1
Pt rP + rPx St + s2 St2 Pxx = 0
2
In particular, the drift terms in equation 4.32 will be equal if P satisfies the
following PDE:

1
Pt rP + rxPx + s2 x2 Pxx = 0, 8 t, x 2 R + (4.33)
2

P(t, x ) = h( x ), 8 t, x 2 R +
Where all the partial derivatives of P are evaluated at (t, x ).
38 The Black-Scholes Model
Chapter 5

Stochastic Volatility

5.1 Empirical motivations

In the Black-Scholes model,

St 1 2
= exp{ (µ s )(t u) + s(Wt Wu )}
Su 2
So that

St Su St 1 2
⇡ log( ) = (µ s )(t u) + s(Wt Wu )
Su Su 2
⇣ ⌘
Which is distributed as N µ 12 s2 )(t u), s (t u) . Hence, the variance
rate of the returns is approximately given by the volatility .
In a Stochastic Volatility framework, the stock price is modelled as:

dSt = µSt dt + st St dWt (5.1)

Where st is now a process instead of a constant. Since the variance rate of re-
turns in the Black-Scholes model is approximately the volatility s, the stochas-
tic volatility model is indeed a generalization of the Black-Scholes model in
which the variance rate of the returns is no longer assumed to be constant, as
argued in Chapter 1.

Moreover, as argued in Gatheral 2011, mixing distributions with different

variances produces distributions with higher peaks and fatter tails, which
means that the modeling of random volatility produces distributions in re-
turns that are better adjusted to the observed stock price distributions (Man-
delbrot 1963). The randomness of the volatility parameter also allows us to
incorporate more information into the model, such as correlation between the
volatility and the stock price, or diffusion properties of the volatility such as
mean reversion, which will be introduced in the Heston model. Further argu-
ments and empirical motivations for the Stochastic Volatility models, such as
the observation of implied volatilities, are out of the scope of this project, but
can be found in detail in Gatheral 2011 and Fouque, Papanicolaou, and Sircar
2000.

5.2 A general approach for pricing

In this section, we follow closely the arguments and notations in Fouque, Pa-
panicolaou, and Sircar 2000 (p. 46, section 2.5). Consider a stochastic volatility
model of the form:

39
40 Stochastic Volatility

dSt = µSt dt + st St dWt (5.2)

Where (St ) represents the stock price, st = f (Yt ) for some positive function
f , and Yt is an Ornstein-Uhlenbeck process, defined by:

dYt = a(m Yt )dt + b t dẐt (5.3)

Where a and m are constants with am > 0, and ( b t ) is an adapted process.

Intuitively, this process pulls toward or reverts to m, known as the long-run
mean level of (Yt ), with a velocity a, known as the rate of mean reversion. We
assume that the Brownian motion ( Ẑt ) satisfies:
q
Ẑt = rWt + 1 r2 Zt (5.4)

Where ( Zt ) is a standard Brownian motion independent of (Wt ), and r 2

[ 1, 1]. This implies that the correlation between ( Ẑt ) and (Wt ) is r. Indeed,

h i
E Ẑt E [ Ẑt ] Wt E [Wt ] E [ Ẑt Wt ] rt
cor ( Ẑt , Wt ) = = pp = =r (5.5)
sd( Ẑt )sd(Wt ) t t t

Thus, this model allows us not only to incorporate randomness into the
volatility, but also to specify the correlation between the volatility and the
stock price, or skewness, as argued in the previous section.

As usual, the filtered probability space is (W, F , F, P ), in this case being

F = (Ft ) the filtration generated by the two Brownian motions. More ex-
plicitely, Ft is generated by the sets {w 2 W|Ws < a, Fs < b} for s  t and the
P-null sets. In particular, ( b t ) is required to be F-adapted.

In the Black-Scholes model, we were able to obtain a pricing formula for

European options with a payoff H = h(ST ) which is square integrable with
respect to Q after proving the existence of the unique risk neutral measure
Q. In this case, we will first assume the existence of a certain risk neutral
measure Q, under which the discounted stock price St⇤ is a martingale. As in
the Black-Scholes model, the No Arbitrage Principle implies that the option
price Vt is given by:

r ( T t)
Vt = EQ [e H |Ft ]

So, for each risk neutral measure Q, we are able to find a reasonable option
price Vt . We will now find a family of equivalent risk neutral measures, by
means of the multidimensional Girsanov theorem.

µ r
Indeed, let qt := st , and define:
Z t
Wt⇤ = Wt + qs ds
0

And for an arbitrary adapted, square integrable process (gt ), define:

Z t
Zt⇤ = Zt + gs ds
0

Let Q g be defined by:

5.2 A general approach for pricing 41

Z T Z T Z T
!
dQ g 1
= exp qs dWs gs dZs (qs2 + gs2 )ds (5.6)
dP 0 0 2 0

Assuming certain regularity conditions for f - for example, that it is bounded

away from 0 - and according to Girsanov’s theorem in two dimensions, each
measure Q g is a risk neutral probability, under which Wt⇤ and Zt⇤ are inde-
pendent Brownian motions. The process (gt ) is called the risk premium factor
or the market price of volatility risk, and parametrizes the space of risk neutral
measures. The Itô formula shows that the stochastic processes (St ), (Yt ), and
( Ẑt⇤ ) are driven by the following dynamics under the risk neutral measure
Qg :

dSt = rSt dt + f (Yt )St dWt⇤ , (5.7)

h q i
dYt = a(m Yt ) b t (rqt + gt 1 r2 dt + b t dẐt⇤ , (5.8)

⇣q ⌘
Ẑt⇤ = rWt⇤ + 1 r2 Zt⇤ (5.9)

Now, assume that gt = g(t, St , Yt ), b t = b(t, St , Yt ). Then, the option price

Vt = EQg [e r(T t) H |Ft ] is also a function of t, St and Yt , because of the
Markov property. Indeed,
" #
⇣ ⌘
r ( T t)
Vt = EQg e h ST Ft

" #
⇣ ⌘
r ( T t)
= EQg e h ST FSt ,Yt (Because of the Markov property)

" #
⇣ ⌘
r ( T t)
= EQg e h ST St = s, Yt = y
s=St ,y=Yt

Under these assumptions, we can proceed as in the previous section to show

that the function P defined by Vt = P(t, St , Yt ) follows a certain PDE. Indeed,

rt rt rt
d( e P(t, St , Yt )) = re P(t, St , Yt )) + e dP(t, St , Yt )
And the two dimensional Itô formula gives:

rt
d e P(t, St , Yt ) = (5.10)
" q
h ⇣ ⌘i
rt
=e Pt rP + rSt Ps + a(m Yt ) b t rqt + gt 1 r2 Py

#
1 2 2 1 2
+ f (Yt ) St Pss + b t Pyy + r f (Yt )St b t Psy dt
2 2
⇣ ⌘ ⇣ ⌘
+e rt f (Yt )St Ps dWt⇤ + e rt ny b t Py dẐt⇤

Note: we frequently abbreviate P(t, s, y) as P. We will not abbreviate in cases where

confusion may arise or to clarify the dependence on t, s, y.

In particular, e rt P(t, St , Yt ) will be a local Q g -martingale if P(t, s, y) is a solu-

tion of the following PDE:
42 Stochastic Volatility

h ⇣ µ r q ⌘i
Pt rP + rsPs + a(m y) b(t, s, y) r + g(t, s, y) 1 r2 Py (5.11)
f (y)

1 1
+ f (y)2 s2 Pss + b(t, s, y)2 Pyy + rb(t, s, y)s f (y) Psy = 0
2 2

5.3 Heston’s model

Heston’s model, as described in Heston 1993, is a particular case of the general
p p
model described above, in which f (y) = y and b(t, s, y) = s y. Hence, the
dynamics of the processes (St ) and (Yt ) are give by:
p
dSt = µSt dt + Yt St dWt (5.12)

p
dYt = a(m Yt )dt + s Yt dẐt (5.13)
Where, as before, Ẑt is defined as:
q
Ẑt = rWt + 1 r2 Zt (5.14)
Note that 5.12 and 5.13 can be expressed in the form 3.5, with:

a(t, s, y) = (µs, a(m y)) (5.15)

p p
b(t, s, y) = ( ys, s y)
In this case, the coefficients are not Lipschitz. However, 5.13 can be rewritten
in the form (see Heston 1993):
p p
d Yt = b Yt dt + ddẐt (5.16)
Indeed, applying Itô’s formula one can obtain an expression or Yt similar to
5.13. In this form, the coefficients of the stochastic differential equation satisfy
the regularity conditions in 3.5, so that a unique solution exists and satisfies
the Markov property.

In this section, we wish to obtain a pricing function P(t, St , Yt ) for a European

call option with maturity T and payoff H = h(St ) = (St K )+ , where St is
modelled according to Heston’s model. Steven L. Heston, in his article Heston
1993, derives a closed-form solution for the call option price C (t, s, y) in this
model. We will follow closely its arguments to achieve the desired formula.

Inspired by the results obtained in the previous section, we require that

C (t, s, y) satisfies the following PDE:

Ct rC + rsCs + a(m y) l(t, s, y) Cy (5.17)

1 1
+ ys2 Css + s2 yCyy + rsysCsy = 0
2 2
Where
⇣ q ⌘
p
l(t, s, y) = s r(µ r) + yg(t, s, y) 1 r2 (5.18)

And with boundary condition C ( T, s, y) = (s K )+ .

5.3 Heston’s model 43

The price C (t, St , Yt ) can also be expressed as follows:

h i
r ( T t)
C (t, St , Yt ) = EQg e ST K +
Ft (5.19)

h i h i
r ( T t) r ( T t)
= EQg e S T 1{ S T K} Ft e KEQg 1{ST K} Ft (5.20)

he r ( T t) S i h i
T r ( T t)
= St E Q g 1{ S T K} Ft e KEQg 1{ST K} Ft (5.21)
St

= St P1 KP(t, T ) P2 (5.22)
h i
e r ( T t) ST
Where P(t, T ) = e r(T t) , P1 = EQg St 1{ S T K } Ft ,
h i
and P2 = EQg 1{ST K } Ft .

Now, P(t, T ) P2 is the price at time t of an option with payoff H = 1{ST K} .

Indeed,
h i
r ( T t)
P(t, T ) P2 = EQg e 1{ S T K} Ft (5.23)

Thus, the equation for P(t, T ) P2 is exactly the same as 5.17. Consider the
change of variables x = ln(s), so that:

∂ ∂ 1
= ,
∂s ∂x s

∂2 ∂2 1 1 ∂
=
∂s2 ∂x2 s2 s2 ∂x
Which gives the following PDE for P2 :

∂P2 1 ∂P2 ∂P2 1 ∂2 P2

+ r y + a(m y) ly + y 2 (5.24)
∂t 2 ∂x ∂y 2 ∂x

1 ∂2 P ∂2 P2
+ s2 y 22 + rsy =0
2 ∂y ∂x∂y
With boundary condition P2 ( T, x, y) = 1{ x log(k)} .

Now, substituting C (t, s, y) in 5.17 and using the previous result, we obtain a
PDE for P1 :

∂P1 ⇣ ∂P ⌘ ⇣ ⌘ ∂P
1 1
s rsP1 + rs + P1 + s a(m y) ly
∂t ∂x ∂y

1 ⇣ ∂2 P1 ∂P1 ⌘ 1 ⇣ ∂P1 ⌘
ys +2 + P1 ys + P1
2 ∂x2 ∂x 2 ∂x

1 2 ∂2 P1 ⇣ ∂2 P ∂P ⌘
1
s ys 2 + rsys + 1 =0
2 ∂y ∂x∂y ∂y
Reordering this expression, we obtain the following PDE for P1 :

∂P1 1 ∂P1 ∂P1

+ r+ y + a(m y) ly + rsy (5.25)
∂t 2 ∂x ∂y

1 ∂2 P 1 ∂2 P ∂2 P1
+ y 21 + s2 y 21 + rsy =0
2 ∂x 2 ∂y ∂x∂y
With the boundary condition:
44 Stochastic Volatility

P1 ( T, x, y) = 1{ x log(K )}

We can summarize 5.24 and 5.25 in the expression:

∂Pj ∂Pj ∂Pj 1 ∂2 Pj 1 2 ∂2 Pj ∂2 Pj

+ r + uj y + aj bj y + y 2 + s y 2 + rsy = 0,
∂t ∂x ∂y 2 ∂x 2 ∂y ∂x∂y
(5.26)

Pj ( T, x, y) = 1{ x log(K )}

Where

1
u1 =
2
1
u2 =
2
a j = am, j = 1, 2
b1 = a + l rs
b2 = a + l

Now, continuing with the same notations, assume that x j , j = 1, 2, is a

stochastic process that follows the stochastic differential equation:
q
dx j (t) = r + u j y j (t) dt + y j (t)dWt (5.27)

q
dy j (t) = a j b j y j (t) dt + s y j (t)dẐt

As argued in Heston 1993, Pj is the conditional probability that the option

expires in-the-money:
⇣ ⌘
Pj (t, x, y) = P x j ( T ) log(K ) x j (t) = x, y j (t) = y (5.28)

To prove this, define

h i
f j (t, x, y) = E 1{ x j (T ) log(K )} x j (t) = x, y j (t) = y

Assume that f j (t) = f j (t, x j (t), y j (t)) is sufficiently regular to apply the Itô
formula. In this case,
⇣∂f ∂ fj ∂ fj
j
d f j (t) = + r + u j y j (t) + aj b j y j (t) (5.29)
∂t ∂x ∂y
1 ∂2 f j 1 2 ∂2 f j ∂2 f j ⌘
+ y j (t) 2 + s y j (t) 2 + rsy j (t) dt
2 ∂x 2 ∂y ∂x∂y
⇣ ∂ fj ⌘ ⇣ ∂ fj ⌘
+ r + u j y j (t) dWt + a j b j y j (t) dẐt
∂x ∂y
Where all the partial derivatives of f are evaluated at (t, x j (t), y j (t)).

Now, f j is a martingale, because due to the tower property of conditional

expectations and the Markov property,

⇥ h ⇥ ⇤ i
E f j (t) Fs ]= E E 1{ x j (T ) log(K )} x j (t) = x, y j (t) = y x = x j (t),y=y j (t)
Fs
5.3 Heston’s model 45

h ⇥ ⇤ i
E E 1{ x j ( T ) log(K )} Ft Fs = f j (s)
Thus, the drift term in 5.29 must vanish. We obtain the PDE:

∂ fj ∂ fj ∂ fj
+ r + u j y j (t) + aj b j y j (t) (5.30)
∂t ∂x ∂y

1 ∂2 f j 1 2 ∂2 f j ∂2 f j
+ y j (t) 2 + s y j (t) 2 + rsy j (t) =0
2 ∂x 2 ∂y ∂x∂y
Clearly, f j must satisfy the boundary condition

f j ( T, x, y) = 1{ x log(K )}

Hence, assuming regularity conditions on f j ,

h i
Pj (t, x, y) = E 1{ x j (T ) log(K )} x j (t) = x, y j (t) = y (5.31)

⇣ ⌘
= P xj (T ) log(K ) x j (t) = x, y j (t) = y

Now, applying the Itô formula to Pj (t, x, y) we obtain that it is a martingale

because the drift term vanishes. Since Pj and f j satisfy the same boundary
conditions and are both martingales, Pj = f j and the regularity assumption
on f j is validated.

The probabilities 5.31 are not easily obtained in a closed form. However,
consider the characteristic function g j defined as:
h i
ifx j ( T )
g j (t, x, y; f) = E e x j (t) = x, y j (t) = y

We admit (see Heston 1993, Appendix p.341) that g j can be expressed as:

g j (t, x, y; f) = eC(T t:f)+ D ( T t:f)y+ifx

(5.32)
Where

a⇣ 1 c j edt ⌘
Cj (t; f) = rfit + 2 (b j rsfi + d j )t 2log (5.33)
s 1 cj
bj rsfi + d j 1 ed j t
D j (t; f) = (5.34)
s2 1 ced j t
and

bj rsfi + d
cj = (5.35)
bj rsfi d
q
dj = (rsfi b j )2 s2 (2u j fi f2 ) (5.36)
To obtain the desired result, the characteristic functions can be inverted in the
following way:
Z • he iflog(K ) g i
1 1 j ( t, x, y; f )
Pj (t, x, y) = + Re df (5.37)
2 p 0 if
The following are illustrative plots of the call option price in Heston’s model,
according to the obtained closed-form solution. The underlying code has been
obtained from Roberts n.d.
46 Stochastic Volatility

Figure 5.1: Heston pricing of a call option according to the closed-form solu-
tion (the dashed line) at time t = 0, with K = 100, T = 2, r = 0.04, a = 0.5,
r = 0.5, Y0 = 0.1, m = 0.1, s = 0.1 compared to the Black-Scholes call price
(the regular line) at time t = 0, with K = 100, T = 2, r = 0.04 and s = 0.1,
and the payoff function ( x K )+ (the thick line).

Figure 5.2: Heston pricing of a call option according to the closed-form solu-
tion (the dashed line) at time t = 0, with X0 = 80, K = 100, T = 2, r = 0.04,
a = 1, r = 0.5, m = 0.1, s = 0.1 compared to the Black-Scholes call price (the
regular line) at time t = 0, with K = 100, T = 2, r = 0.04 and S0 = 80.
Chapter 6

Conclusions

The Black-Scholes model, which is the benchmark model in continuous-time

market models, relies on a series of assumptions that were introduced in
Chapter 1: the interest rate is known and constant through time, the dis-
tribution of stock prices at the end of any finite interval is lognormal, the
stock pays no dividends, the variance rate of the return on the stock is con-
stant and the stock price is continuous over time. In this project we have
mathematically formulated the Black-Scholes model, introducing the theory
of stochastic integration and the fundamental results on stochastic processes
as a necessary background, and taking particular care in examining the un-
derlying assumptions of the model, as well as covering the topics of trading
strategies, completeness, and the notion of arbitrage.

The model has been extended to incorporate stochastic volatility, particularly

in the setting of Heston’s stochastic volatility model. Stochastic volatility
models are one of the many extensions of the Black-Scholes model that are
frequently employed by practitioners in financial markets. Other models that
incorporate jumps, or non-gaussian distributions, among others, are not in
the scope of this project but are equally relevant, and can be combined with
the stochastic volatility framework adding more precision. We have shown
that the stochastic volatility framework not only reduces the assumptions and
limitations of the Black-Scholes model, but also allows to capture additional
information in the model, such as correlation between the stock price and the
volatility, and a mean-reverting structure for the volatility. Moreover, this so-
lution captures some aspects of the empirical distribution of stock prices, such
as fat tails and higher peaks, more accurately than the Black-Scholes model.
This precision of the stochastic volatility models in capturing stock price dis-
tributions results in option valuations that are more reliable than the ones
obtained from the Black-Scholes model, as argued in Gatheral 2011. Finally,
despite the fact that the incorporation of stochastic volatility adds complex-
ity to the model, we obtained a closed-form solution for pricing options in
Heston’s model. This closed-form solution reduces the complexity of pricing
options within Heston’s model and is the cause of the high popularity of this
model.

47
48 Conclusions
Chapter 7

Bibliography

Bibliography: Articles
Mandelbrot, Benoit (1963). ?The variation of certain speculative prices? In: The
journal of business 36.4, pp. 394–419.
Black, Fischer and Myron Scholes (1973). ?The pricing of options and corpo-
rate liabilities? In: Journal of political economy 81.3, pp. 637–654.
Heston, Steven L (1993). ?A closed-form solution for options with stochastic
volatility with applications to bond and currency options? In: Review of
financial studies 6.2, pp. 327–343.
Courtault, Jean-Michel et al. (2000). ?Louis Bachelier on the centenary of The-
orie de la speculation? In: Mathematical Finance 10.3, pp. 339–353.

Bibliography: Books
Bachelier, Louis (1900). Theorie de la speculation. Gauthier-Villars.
Fouque, Jean-Pierre, George Papanicolaou, and K Ronnie Sircar (2000). Deriva-
tives in financial markets with stochastic volatility. Cambridge University Press.
Rogers, L Chris G and David Williams (2000). Diffusions, Markov processes and
martingales: Volume 2, Ito calculus. Vol. 2. Cambridge university press.
Jacod, Jean and Philip Protter (2004). Probability essentials. Springer Science &
Business Media.
Kallenberg, Olav (2006). Foundations of modern probability. Springer Science &
Business Media.
Musiela, Marek and Marek Rutkowski (2006). Martingale methods in financial
modelling. Vol. 36. Springer Science & Business Media.
Lamberton, Damien and Bernard Lapeyre (2007). Introduction to stochastic cal-
culus applied to finance. CRC press.
Bachelier, Louis (2011). Louis Bachelier’s theory of speculation: the origins of mod-
ern finance. Princeton University Press.
Gatheral, Jim (2011). The volatility surface: a practitioner’s guide. Vol. 357. John
Wiley & Sons.
Capinski, Marek and Ekkehard Kopp (2012). The Black–Scholes Model. Cam-
bridge University Press.
Capinski, Marek, Ekkehard Kopp, and Janusz Traple (2012). Stochastic calculus
for finance. Cambridge University Press.
Karatzas, Ioannis and Steven Shreve (2012). Brownian motion and stochastic cal-
culus. Vol. 113. Springer Science & Business Media.
Cohn, Donald L (2013). Measure theory. Springer.

49
50 Bibliography

Bibliography: Online
Peter Morters, Yuval Peres (2008). Brownian Motion. url: https://www.stat.
berkeley.edu/~peres/bmbook.pdf.
Corcuera, Jose Manuel. Introduction to Quantitative Finance. url: http://www.
ub.edu/plie/personal_PLiE/corcuera_HTML/ifqfe3.pdf.
Roberts, Dale. Implementations of the Heston stochastic volatility model. url: https:
//github.com/daleroberts/heston.