Lecture Notes

Introduction to Stochastic Calculus for Economists
V. Filipe Martins-da-Rocha
August 8, 2023
ii
Contents
1 Basic Probability Theory 1

1.1 Time and Uncertainty . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3 Random Variable and Stochastic Process . . . . . . . . . . . . . . . . . . . . 5
1.4 Independence . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.5 Expectation and Conditional Expectation . . . . . . . . . . . . . . . . . . . 8
1.6 Martingale . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2 Discrete Time Dynamic Trading 11

2.1 Assets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.2 Portfolios . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
2.3 Self-Financing Trading Strategies . . . . . . . . . . . . . . . . . . . . . . . . 16
2.4 Hedging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
2.5 Arbitrage Opportunities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
2.6 Stochastic Discount Factor . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
2.7 Equivalent Martingale Measure . . . . . . . . . . . . . . . . . . . . . . . . . 21
2.8 Discounted Value of Self-Financing Trading Strategies . . . . . . . . . . . . 23
2.9 The Cox-Ross-Rubinstein Binomial Market Model . . . . . . . . . . . . . . 23
2.9.1 Equivalent Martingale Measure . . . . . . . . . . . . . . . . . . . . . 24
2.9.2 Asset Pricing: European Call . . . . . . . . . . . . . . . . . . . . . . 25
2.9.3 Replicating Strategy . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
3 Continuous Time 27
3.1 Scaled Random Walk . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
3.1.1 Symmetric Random Walk . . . . . . . . . . . . . . . . . . . . . . . . 28
3.1.2 Increasing the Frequency of Tosses . . . . . . . . . . . . . . . . . . . 31
3.2 Brownian Motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
3.2.1 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.3 Itô’s Integral . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
3.4 Itô-Doeblin Formula . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
3.5 Itô Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.6 Itô Processes: Examples . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
3.6.1 Linear constant coefficient . . . . . . . . . . . . . . . . . . . . . . . . 52
3.6.2 Geometric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
3.6.3 Square-root . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
iii
iv CONTENTS
3.6.4 Mean-Reverting Process . . . . . . . . . . . . . . . . . . . . . . . . . 56

3.6.5 Ornstein-Uhlenbeck Process . . . . . . . . . . . . . . . . . . . . . . . 59
3.6.6 Stochastic Volatility Process . . . . . . . . . . . . . . . . . . . . . . 61
3.7 Risk-Neutral Pricing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.7.1 Discounted Prices and Values . . . . . . . . . . . . . . . . . . . . . . 63
3.7.2 Martingale Equivalent Measure . . . . . . . . . . . . . . . . . . . . . 63
3.7.3 Risk-Neutral Measure . . . . . . . . . . . . . . . . . . . . . . . . . . 64
3.7.4 Hedging . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.8 Black-Scholes-Merton Formula . . . . . . . . . . . . . . . . . . . . . . . . . 65
3.9 Martingale Representation Theorem . . . . . . . . . . . . . . . . . . . . . . 70
3.10 Complete Markets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
Chapter 1
Basic Probability Theory
We set off on an expedition through the vast subject of probability theory. No doubt,
the reader has some familiarity with this place and will recognize some of the early land-
marks of the journey. Based on the binomial no-arbitrage pricing model presented in
Shreve (2004a), we recall the standard concepts of probability theory (σ-algebra, proba-
bility measure, random variable, independence, expectation, conditional expectation, and
martingale) in the simple framework of a coin toss space.
1.1 Time and Uncertainty

Time is discrete and runs from 0 to infinite, that is, t ∈ {0, 1, . . .}. At every period t, we
can observe ωt the outcome of a random device with k possible results
ωt ∈ Z := {z1 , . . . , zk }.
The likelihood of each outcome is described by a probability p on Z. Recall that p is a

function from Z to (0, 1) satisfying
X
p(z) = 1.
z∈Z
Example 1.1. If Nature tosses a fair coin every period, then Z = {H, T } with “H”
for Heads and “T ” for Tails. The probability p is uniform, i.e., p(H) = p(T ) = 1/2. See
Figure 1.1. If Nature rolls a dice with six faces, we have Z = {1, 2, 3, 4, 5, 6} and p(z) = 1/6
for every z ∈ Z. The probability does not have to be uniform.1
Example 1.2. To simplify the presentation, we assume that Z is time and history inde-
pendent and p is stationary and history independent. The set Z of possible shocks at date t
could depend on time t and on the history (ω1 , ω2 , . . . , ωt−1 ) of realized shocks preceding
time t. The probability of some z ∈ Z at date t could also depend on the history of previ-
ous shocks. In that case, we use the notation pt (z|ω1 , ω2 , . . . , ωt−1 ) for the probability of
the current shock being z given that the history of past shocks is (ω1 , ω2 , . . . , ωt−1 ). Take
1
There may be two possible outcomes Z = {h, ℓ} at every period (h for high and ℓ for low) with
p(h) > p(ℓ).
1
2 CHAPTER 1. BASIC PROBABILITY THEORY
Figure 1.1: Tossing a coin at every period

1.2. INFORMATION 3
Figure 1.2: Two possible states of nature ω and ω̂
for instance the case where Z = {h, ℓ} and the probability that the shock at date t + 1 is
ωt+1 depends on the current shock ωt and is denoted by pt+1 (ωt |ωt ) where the conditional
probability pt+1 (z ′ |z) is time-independent and is given by
(
′ ′ α, if z ′ ̸= z
pt+1 (z |z) = p(z |z) :=
1 − α, otherwise.
This corresponds to the case where the probability that the shock switches (from h to ℓ,
or from ℓ to h) is α ∈ (0, 1).
Uncertainty is represented by all possible complete histories of shocks. A complete
history of shocks, also called state of nature, is then an infinite sequence
ω = (ω1 , ω2 , . . . , ωt , . . .) = (ωt )t⩾1 , with ωt ∈ Z, for all t ⩾ 1.
The set of states of nature is the set Z N of sequences with values in Z. We use the simpler
notation Ω := Z N .
1.2 Information
Given a state of nature ω = (ωt )t⩾1 ∈ Ω and a date T , we denote by ω T the history of
shocks (or realizations) in ω up to date T , i.e.,
ω T := (ω1 , ω2 , . . . , ωT −1 , ωT ).
Figure 1.3: Date-2 event
The mapping ω 7→ ω T from Z N to Z T is denoted by projT . It represents the available

information at date t. Indeed, at date t, we can observe the outcome ωt ∈ Z of the
random experiment, and we can remember the outcomes ωs for every preceding dates s ∈
{1, 2, . . . , t − 1}.
Fix a partial history of shocks ω̄ T = (ω̄1 , . . . , ω̄T ) and denote by ΩT (ω̄ T ) the set of all
complete histories that are consistent with the partial history ω̄ T in the following sense2
ΩT (ω̄ T ) := {ω ∈ Ω : projT (ω) = ω̄ T } = {ω ∈ Ω : ωt = ω̄t , ∀t ∈ {1, . . . , T }}.
If the set Z has at least two elements, the set ΩT (ω̄ T ) has infinitely many states of nature.
At date T , we cannot identify which states of nature in ΩT (ω̄ T ) will prevail. We can only
say whether or not the history of shocks up to date T is ω̄ T . We say that we can observe
whether the true state of nature belongs or not to the set ΩT (ω̄ T ).
A set of the form Ωt (ω̄ t ) is called a date-t event. The probability of Ωt (ω̄ t ), denoted
by P(Ωt (ω̄ t )) (we can also use the notation Pt (ω̄ t )) is defined by
P(Ωt (ω̄ t )) := p(ω̄1 )p(ω̄2 ) . . . p(ω̄t ).
Any finite or countable unions of date-events (of possible different dates) is called
an event. We can define the probability P(A) of any event by using the rule that
2
See Figure 1.3.
1.3. RANDOM VARIABLE AND STOCHASTIC PROCESS 5
P(A ∪ B) = P(A) + P(B) when A ∩ B = ∅. For instance, if ω̄ T and ω̂ τ are two partial
histories with τ < T and ω̄ τ ̸= ω̂ T , then the two events ΩT (ω̄ T ) and Ωτ (ω̂ τ ) have an empty
intersection and therefore
P ΩT (ω̄ T ) ∪ Ωτ (ω̂ τ ) = p(ω̄1 )p(ω̄2 ) . . . p(ω̄T ) + p(ω̂1 )p(ω̂2 ) . . . p(ω̂τ ).

Any finite or countable intersection of any finite or countable family of events is also called
an event. Similarly, any finite or countable union of any finite or countable intersection of
any finite or countable union of events is also called an event. We can repeat this process
countably many times.
We denote the set of all events by F. It is the smallest set of subsets of Ω satisfying
the following properties:
1. The sets ∅ and Ω are events;
2. For any date t and partial history ω t , the set Ωt (ω t ) is an event;
3. If A and B are two events, then A ∩ B and A ∪ B are also events;
4. If A is an event, then Ac := Ω \ A is also an event;
5. If (An )n∈N is a sequence of events, then ∪n∈N An and ∩n∈N An are also events.
Any set of sets satisfying properties 1., 3., 4. and 5. is called a σ-algebra.
Fix an arbitrary date t. The set of unions of sets of the form Ωt (ω t ) for all possible
partial histories ω t is denoted by Ft . It represents the events that can be observed at
date t, or the information available at date t. The class (a set of sets is also called a class)
Ft is finite. The elements of the form Ωt (ω t ) are called the atoms of Ft , and they form a
partition of Ω. See Figure 1.4.
Observe that [
Ωt (ω t ) = Ωt+1 ((ω1 , . . . , ωt , z)).
z∈Z
This implies that Ft ⊂ Ft+1 . To simplify notations, we pose F0 := {∅, Ω}. The sequence
(Ft )t⩾0 is called a filtration and represents the evolution of information.
1.3 Random Variable and Stochastic Process

Fix a function X : Ω −→ R. Given a real number x ∈ R, we use the following notation
{X ⩽ x} := {ω ∈ Ω : X(ω) ⩽ x}.
A random variable is a function X : Ω −→ R such that {X ⩽ x} is an event (i.e.,

belongs to F) for any real number x ∈ R. The set {X ⩽ x} is the event: “the random
variable X is below x”. We can define in a similar manner the events {X < x}, {X ⩾ x}
and {X > x}.
Fix a date t. A random variable X : Ω → R is said to be Ft -measurable when for
every level x ∈ R, the event {X ⩽ x} belongs to Ft . Given the information observed
up to date t, we can say whether X ⩽ x or not. We can verify that X is Ft -measurable
if, and only if, if for every complete history ω ∈ Ω, the value X(ω) only depends on the
Figure 1.4: The set of date-2 events forms a partition of Ω

1.4. INDEPENDENCE 7
information ω t observed up to date t, in the sense that X(ω) = X(ω

e t ) for some function
Xe : Ωt → R where Ωt is the finite set of all histories ω t up to date t. Equivalently, X
is Ft -measurable when the outcome of X only depends on the information available at
date t. Sometimes we abuse notations and write X(ω t ) instead of X(ω) (or instead of
e t )) when X is Ft -measurable.
X(ω
A very useful random variable is the indicator function 1A of an event A ∈ F defined
by (
1, if ω ∈ A
1A (ω) :=
0, otherwise.
Observe that if the event A belongs to Ft , then the random variable 1A is Ft -measurable.
Finally, observe that X is F0 -measurable if, and only if, X is constant.
A stochastic process is a sequence (Xt )t⩾0 of random variables such that, for every
t ⩾ 1, the random variable Xt : Ω → R is Ft -measurable.
1.4 Independence
Recall that two events A and B are independent when
P(A ∩ B) = P(A)P(B).
Fix an arbitrary event B such that P(B) > 0. The probability of an event A conditional
to B is defined by
P(A ∩ B)
P(A|B) := .
P(B)
It represents the probability that the true state of nature belongs to A when we know that
it belongs to B. Observe that if two events A and B are independent and P(B) > 0, then
P(A|B) = P(A).
Independence means that knowing that the true state belongs to B does not modify the
likelihood that it belongs to A.
Two random variables X and Y are independent when for every x, y ∈ R, the
events {X ⩽ x} and {Y ⩽ y} are independent.
An event A ∈ F is said to be independent of the information Ft when A and
B are independent for any event B ∈ Ft observable at date t. We can verify that A is
independent of the information Ft if, and only if, for every atom Ωt (ω t ) of Ft , the events
A and Ωt (ω t ) are independent.
A random variable X is independent of the information Ft when for every
x ∈ R, the event {X ⩽ x} is independent of Ft .
Fix an arbitrary date t and some arbitrary shock z ∈ Z. Let Ωt+1 (z) be the event of
all complete histories ω ∈ Ω such that ωt+1 = z. The set Ωt+1 (z) is the event that the
outcome of the random experiment at date t + 1 is z. Observe that for any partial history
ω t , the date-t event Ωt (ω t ) and the event Ωt+1 (z) are independent. This follows from the
following property
P(Ωt (ω t ) ∩ Ωt+1 (z)) = P(Ωt+1 ((ω1 , . . . , ωt , z))) = p(ω1 ) . . . p(ωt ) p(z).
| {z }
P(Ωt (ω t ))
The event Ωt+1 (z) is independent of the information Ft . Actually, any event of the form
Ωτ (zτ ) for τ > t and any zτ ∈ Z is also independent of the information Ft . We can
go further. Take any event A that is the countable union of countable intersections of
countable unions of . . . of events of the form Ωτ (zτ ) for τ > t and any zτ ∈ Z. The event
A is also independent of the date t information Ft . Informally, any event that only depends
of the outcomes of the random experiment occurring strictly after date t is independent
of the information displayed by the σ-algebra Ft .
1.5 Expectation and Conditional Expectation

Once we have a probability measure P on events of Ω, we can define the expected value
EP (X) of a random variable X : Ω → R. When X only takes finitely many values, in the
sense that the set
X(Ω) := {x ∈ R : ∃ω ∈ Ω, x = X(ω)} = {X(ω) : ω ∈ Ω}
is finite, then X
E(X) := xP({X = x}).
x∈X(Ω)
For a given date t, we look for a random variable Xt : Ω → R that is Ft -measurable and
close to X in the following sense:
EP (Xt Y ) = EP (XY )
for any Ft -measurable random variable Y : Ω → R. A random variable Xt satisfying the

above property does exist and is denoted by EP [X|Ft ]. It is called the Ft -conditional
expectations of X. Fix a state of nature ω and let ω t = projt (ω). Choosing Y to be the
indicator of the event Ωt (ω t ), we have
1
EP [X|Ft ](ω) = EP X1Ωt (ωt ) .
P(Ωt (ω t ))
The conditional expectation operator satisfies the following important properties:
(i) if X is already Ft -measurable, then EP [X|Ft ] = X;
(ii) positivity: if X ⩾ 0, then EP [X|Ft ] ⩾ 0;
(iii) linearity: for every random variables X, Y and scalars λ, µ ∈ R, we have EP [λX +
µY |Ft ] = λ EP [X|Ft ] + µ EP [Y |Ft ];
(iv) tower property: if s < t, then EP [EP [X|Ft ]|Fs ] = EP [X|Fs ];
(v) taking out what is known: if Z is Ft -measurable, then EP [ZX|Ft ] = Z EP [X|Ft ];
(vi) Jensen: if u : R → R is concave, then EP [u(X)|Ft ]⩽u(EP [X|Ft ]);
(vii) role of independence: if X is independent of Ft , then EP [X|Ft ] = EP [X].

1.5. EXPECTATION AND CONDITIONAL EXPECTATION 9
The interpretation of the last property is the following: if the events of the form
{X ⩽ x} are independent of the information Ft , then knowing the results of the random
experiments up to date t does not give us any relevant information regarding the distri-
bution of the possible values of X. In other words, knowing whether the partial history is
ω t , ω̂ t or ω̄ t does not allow me to infer any particular information about the likelihood of
X taking values below some fixed level x. The conditional expectation EP [X|Ft ] must be
constant.
Given a random variable X : Ω → R defined on a probability space (Ω, F, P), we can
consider the law of X, denoted by PX , that is the probability on R satisfying the following
property
PX (E) = P({ω ∈ Ω : X(ω) ∈ E})
for every (well-behaved) subset of R. The distribution function of X is the function
FX : R → [0, 1] defined by
FX (c) := PX ((−∞, c]) = P({ω ∈ Ω : X(ω) ⩽ c}.
The law of X is said to have a probability density function (pdf ), denoted by f X ,
when Z Z ∞
X X
P (E) = f (x)dx = 1E (x)f X (x)dx
E −∞
where 1E : R → {0, 1} is the indicator function of E defined by 1E (x) = 1 if x ∈ E and
1E (x) = 0 if x ̸∈ E.
We can use the pdf of X to compute expectations as follows. For a (well-behaved)
function g : R → R, we have
Z ∞
E[g(X)] = g(x)f X (x)dx.
−∞
Many of the random variables we analyze in this course have a normal distribution N (µ, σ)
with mean µ and variance σ 2 . The pdf of the normal distribution is the function
(x − µ)2

1
x 7−→ √ exp − .
2πσ 2 2σ 2
If the distribution of X is N (µ, σ), then for every y ∈ R, we can verify that
E[exp(yX)] = exp µy + σ 2 y 2 /2 .

Consider a sub σ-algebra G ⊆ F. Intuitively, the σ-algebra G contains less information

than F. Fix two random variables X, Y : Ω → R such that X is G-measurable and Y is
independent of G. Intuitively, this means that if I can verify that the events in G occurred
or not, then I can see the value taken by X. On the contrary, knowing all the events in G
does not give me more information about the value of Y .
Let h : R × R → R be a well-behaved function. We can compute E[h(X, Y )|G] using
the following result.
Lemma 1.1 (Lemma of Independence). Let g : R → R be the function defined by
g(x) := E[f (x, Y )], for all x ∈ R.
We have E[h(X, Y )|G] = g(X).
The intuition is as follows. If I can observe the events in G, I can know the value
x of the random variable X. Therefore, I take this value “as a constant” and compute
E[h(x, Y )|G]. Since Y is independent of G, the transformation h(x, Y ) is also independent
of G. Therefore, we have E[h(x, Y )|G] = E[h(x, Y )]. To recover the random variable
E[h(X, Y )|G], we replace x with the random variable X.
1.6 Martingale
A stochastic process (Xt )t⩾0 is said to be a martingale when EP [Xt+1 |Ft ] = Xt for every
t ⩾ 0. This implies that each Xt is Ft -measurable. Using the tower property, we have that
(Xt )t⩾0 is a martingale if, and only if, for every τ > t, we have EP [Xτ |Ft ] = Xt . Observe
that if X is a random variable, then the process (EP [X|Ft ])t⩾0 is a martingale.
Consider the case where Z = {H, T } is the set of possible outcomes when tossing a
coin. Assume that p(H) = 1/2 (fair coin). Let Xt be the random variable defined by
(
+1, if ωt = H
Xt (ω) :=
−1, if ωt = T.
The random variable Xt measures the gain when a gambler wins 1 if the coin comes up
Heads and loses 1 if it comes up Tails. Observe that Xt only depends on the outcome of
the random experiment at date t. In particular, it is independent of the information Ft−1 .
Denote by Mt the gambler’s accumulated gain at date t and defined by
Mt = X1 + X2 + . . . + Xt .
For simplicity, we pose M0 := X0 := 0. The process (Mt )t⩾0 is also known as the sym-
metric random walk. We can verify that the process (Mt )t⩾0 is a martingale. Indeed,
fix two dates ti+1 > ti , then
Mti+1 − Mti = Xti +1 + . . . + Xti+1 .
For every t > ti , the random variable Xt is independent of the information Ft . This
implies that
EP Mti+1 − Mti |Fti = EP [Xti +1 ] + . . . + EP [Xti+1 ].
Since the coin is fair, we have EP [Xt ] = 0 for every t, and we get the desired result:

EP Mti+1 |Fti = Mti .
In other words, the gambler’s conditional expected fortune after the next trial, given the
history, is equal to the present fortune.
Chapter 2
Discrete Time Dynamic Trading
This chapter represents a pivotal milestone in our journey through the fascinating world
of Stochastic Calculus applied to finance. As we delve into this chapter, we will explore
the fundamental concepts that lay the groundwork for comprehending continuous-time
dynamic trading.
In finance, a central objective is to optimize our decisions in the presence of uncer-
tainty and risk. The discrete-time setting provides a stepping stone to understanding the
intricacies of financial markets and how participants navigate within them. By examining
the standard general equilibrium model of financial markets, we can gain valuable insights
into the principles governing asset pricing and investment decisions.
Key Notions to be Discussed:
1. Assets: We will begin by exploring the notion of assets, which represent financial
instruments or securities traded in the market. Understanding the characteristics
and valuation of assets is essential to building a solid foundation for dynamic trading
strategies.
2. Portfolios: As investors, assembling portfolios is critical to our decision-making

process. We will discuss the construction and management of portfolios to achieve
specific investment objectives.
3. Self-financing Trading Strategies: A vital concept in dynamic trading is the

idea of self-financing strategies. We will analyze how such strategies enable us to
maintain a balanced portfolio without introducing additional external capital.
4. Hedging: Managing risk is a paramount concern for investors. We will investi-

gate how hedging techniques help mitigate exposure to potential losses arising from
adverse market movements.
5. Arbitrage Opportunities: Arbitrage opportunities are unique instances where

risk-free profits can be earned by exploiting market inefficiencies. We will explore
the conditions under which such opportunities arise and their implications.
6. Stochastic Discount Factor: The stochastic discount factor bridges the present
and future values of assets. We will study its significance in asset pricing and valu-
ation.
11
12 CHAPTER 2. DISCRETE TIME DYNAMIC TRADING
7. Equivalent Martingale Measures: We will introduce equivalent martingale mea-

sures to prepare the transition to continuous time dynamic trading. These are fun-
damental in pricing derivatives and understanding market dynamics.
By mastering these concepts, you will develop a solid understanding of discrete-time

dynamic trading, which will pave the way for more sophisticated analyses in continuous
time settings. As we progress through this chapter, I encourage you to actively engage
with the material, ask questions, and participate in discussions. A deeper grasp of these
concepts will empower you to apply stochastic calculus principles effectively in real-world
financial scenarios.
2.1 Assets
There is a fixed set J of d + 1 assets (or securities),
J = {j0 , j1 , . . . , jd }.
It would make sense to allow the set J to depend on t since some assets can be issued at
some date t > 0 and have maturity at some finite horizon T > t. We assume that the set
J does not depend on time to simplify notations.1
The stochastic process (Stj )t⩾0 represents the time evolution of security j’s price. For
every date t, the price Stj only depends on the information available at date t; that is the
random variable is Ft -measurable. Stj (ω) is the amount (in units of some numéraire or
composite consumption good) needed to purchase one unit of asset j at date t contingent
to the state of nature ω. The vector in RJ of asset prices at date t contingent to history
ω is denoted by
St (ω) := (Stj (ω))j∈J ∈ RJ .
The space RJ is the set of functions from J to R. When J has d + 1 elements, it can
be identified with the space Rd+1 of vectors with d + 1 coordinates. Indeed, the vector
St (ω) ∈ RJ can be identified with a vector in Rd+1 as follows
Stj0 (ω)
 
 S j1 (ω) 
 t
(Stj0 (ω), Stj1 (ω), . . . , Stjd (ω)) or  . .

 .  .
Stjd (ω))
One unit of asset j may give the right to its owner to receive an amount δtj (ω) (in units
of the numéraire) at date t contingent to history ω. For instance, when one unit of asset
j is one stock of some firm, δtj (ω) is the per-stock dividend paid by the firm.
Asset j0 is a specific asset called a “bank account” or “money market account”. It is
a riskless security in the sense that Stj0 (ω) does not depend on ω. Moreover, it does not
pay dividends, i.e., δtj0 (ω) = 0 for every t and every ω.
We abuse notation and write Stj0 instead of Stj0 (ω). The (unit) price of the riskless
asset at t = 0 is normalized to 1. The amount Stj0 corresponds to the proceeds you can
1
If an asset has finite maturity T < ∞, we can set its prices and dividends equal to zero after T .
2.1. ASSETS 13
withdraw from your bank account at date t if you invested one unit of numéraire at the
initial period 0. To purchase one unit of asset j0 at date t, we need Stj0 units of numéraire.
In that case, it is as if we had deposited 1 unit of numéraire at the initial period 0. In
particular, if we liquidate our position on this asset at any date τ > t, we get Sτj0 units of
numéraire. The amount
j0
St+1
Rt+1 := j0
St
is called the gross return at date t + 1. Investing one unit of numéraire at date t, you can
withdraw Rt+1 units of numéraire at date t + 1. The equation defines the interest rate
j0
St+1 − Stj0
1 + rt+1 = Rt+1 , or, equivalently, rt+1 = .
Stj0
In the case of a constant interest rate rt = r for every t, we get that
Stj0 = (1 + r)t .
We have assumed that trade occurs at every date t ∈ {0, 1, 2 . . .} where the time unit is
arbitrary (it can be years, quarters, months, days, hours or seconds). The value of r will
depend on this unit. Assume the time unit is one month. If r is the monthly compound
interest rate, then investing 1 unit of numéraire at date t = 0, after one year we can
withdraw from the bank account the amount
j0
S12 = (1 + r)12 .
This implies that the corresponding yearly compound interest rate ry is such that
j0
1 + ry = S12 = (1 + r)12 .
Consider now the case where you want to withdraw your resources from your bank account
after 13 days. How much would you get? To answer this question, let’s assume that interest
in the bank account is compounded daily. The monthly compound interest rate r is then
computed as follows
1 + r = (1 + rd )30
where rd is the daily compound interest rate. We then get that the amount in the bank
account after 13 days is

j0 13 13 13
S 13 = (1 + rd ) = (1 + r) 30 = exp ln(1 + r) .
30 30
If interest is compounded continuously on the bank account, the amount Stj0 can be
determined for any t ⩾ 0 by the following formula
Stj0 = exp(t ln(1 + r)).
The popular way to write the above formula is to pose ρ := ln(1 + r) such that
Stj0 = eρt , for all t ⩾ 0.
It is important to observe that the value ρ depends on the chosen time unit. In this
discussion, the time unit is a month. If r is the monthly interest paid by the money
market, then the corresponding continuous-time compound interest rate is ρ = ln(1 + r).
To figure out the “correct” ρ, remember that exp(ρ) is the money you can withdraw from
your account after one time period (if you invested one unit of the numéraire).
2.2 Portfolios
At t = 0, an agent chooses an initial portfolio (or trading strategy) θ1 = (θ1j )j∈J ∈ RJ
where θ1j ∈ R represents the post-trade position (long if positive and short if negative)
in units of asset j at date 0. Instead of using the terms “pre-trade” and “post-trade”,
some authors use the expression “at the beginning of the period” and “at the end of the
period”, where the interpretation of “beginning” means “before trade” and “end” means
“after the trade occurred”.
The cost of the portfolio θ1 is
X j j
θ1 · S0 = θ1 S 0 .
j∈J
Although the portfolio is chosen at t = 0, we use the notation θ1 because this corresponds
to the agent’s position on each asset at the beginning of period 1, before trade occurs. In
particular, at date t = 1 and contingent to state ω, the agent can liquidate his position
and obtain the following proceeds
X j j
V1 (θ)(ω) := θ1 · [S1 (ω) + δ1 (ω)] = θ1 [S1 (ω) + δ1j (ω)].
j∈J
He can then choose a new portfolio θ2 (ω) ∈ RJ that represents the holdings of each asset
he will have at the beginning of period t = 2. The cost of the new portfolio θ2 (ω) is
θ2 (ω) · S1 (ω). Recall that S1 is F1 -measurable and θ2 is also F1 -measurable since it is
chosen at t = 1 where the only available information is described by F1 .
At the beginning of date t, the agent’s holding of assets is θt (ω) before the trade occurs.
He can liquidate his current positions and get the proceeds
Vt (ω) = θt (ω) · [St (ω) + δt (ω)].
Still, at date t, he can choose another portfolio

j
θt+1 (ω) := (θt+1 (ω))j∈J ∈ RJ
j
where θt+1 (ω) ∈ R represents the position (long if positive and short if negative) in units
of asset j after trade at date t contingent to history ω. The cost of the portfolio θt+1 (ω) is
X j
θt+1 (ω) · St (ω) := θt+1 (ω)Stj (ω).
j∈J
It corresponds to the resources needed to acquire the portfolio θt+1 (ω). As usual, if
θt+1 (ω) · St (ω) < 0, then this is interpreted as resources the agent receives.
To simplify the presentation, we fix a time interval [0, T ] and assume from now on that
assets do not pay dividends when t ∈ [0, T ].
A trading strategy θ = (θt )t⩾1 is a vector-valued (in RJ ) stochastic process such that
θt+1 is Ft -measurable. Such a process is called predictable. Given a trading strategy
(θt )t⩾1 , the value process (Vt (θ))t⩾0 is defined by
V0 (θ) := θ1 · S0 and Vt (θ)(ω) := θt (ω) · St (ω), ∀t ⩾ 1.

2.2. PORTFOLIOS 15
Fix an investment horizon T and assume an agent trades portfolios until date T − 1. At
every date t, the agent is endowed with some exogenous amount et (ω) ⩾ 0 of numéraire
and may decide to consume ct (ω) ⩾ 0. The flow budget constraint at the initial date is
c0 + θ1 · S0 ⩽ e0 .
At any intermediate date t ∈ {1, 2, . . . , T − 1}, the flow budget constraint is
ct (ω) + θt+1 (ω) · St (ω) ⩽ θt (ω) · St (ω) + et (ω).
Finally, at the final period T , we have
cT (ω) ⩽ θT (ω) · ST (ω) + eT (ω).
The agent’s preference relation on the consumption processes (ct )t⩾0 is numerically repre-
sented by a function c 7−→ U (c). Two polar cases are interesting:
1. the function U is strictly increasing as it is the case when

T
X
U (c) = β t EP [u(ct )]
t=0
for some strictly increasing Bernoulli function u : R+ → R and time-preference factor

β ∈ (0, 1);
2. the function U only depends on the final consumption cT and is strictly increasing
in this argument as it is the case when
U (c) = EP [u(cT )].
The second case is usually associated to a situation where et = 0 for every t ∈

{1, 2, . . . , T }.
An agent chooses a trading strategy θ that is budget feasible and maximizes his utility.
Such a trading strategy is said to be optimal.
Fix an optimal trading strategy θ. In the first case, the flow budget constraints are all
binding, and the optimal consumption levels are given by:
• at the initial period

c0 = e0 − θ1 · S0 ;
• at any intermediate date t ∈ {1, 2, . . . , T − 1}
ct (ω) = et (ω) + θt (ω) · St (ω) − θt+1 (ω) · St (ω);
• and at the final period T ,
cT (ω) = θT (ω) · ST (ω) + eT (ω).

2.3 Self-Financing Trading Strategies

In the second case, consuming zero at every t < T is optimal. Moreover, since the agent
can always invest in the riskless asset j0 , at any optimal trading strategy, we must have
• at the initial period

θ1 · S 0 = e 0 ;
• at any intermediate date t ∈ {1, 2, . . . , T − 1}
θt+1 (ω) · St (ω) = θt (ω) · St (ω);
• and at the final period T , we have
cT (ω) = θT (ω) · ST (ω).
A trading strategy θ = (θt )t⩾1 is said to be self-financing in the interval [0, T ] when
θt+1 · St = θt · St = Vt (θ), for all t ∈ {1, 2, . . . , T − 1}.
Observe that a trading strategy is self-financing when the total value of the portfolio at
the beginning of each period t has been used to purchased the new position.
For any stochastic process X = (Xt )t⩾0 , we pose
∆Xt := Xt − Xt−1
to represent the variation of X between t − 1 and t. If a trading strategy θ is self-financing

in the interval [0, T ], then, for every t ∈ {1, . . . , T }
∆Vt (θ) := θt · St − θt−1 · St−1 = θt · St − θt · St−1 = θt · ∆St .
θt ∆St is the gain obtained with portfolio θt due to the price change ∆St . We then have
that a trading strategy θ is self-financing in the interval [0, T ] if, and only if, for every
t ∈ {1, . . . , T } the value of the portfolio is the sum of all previous gains, i.e.,
Vt (θ) = V0 + θ1 ∆S1 + θ2 · ∆S2 + . . . + θt · ∆St .

| {z }
=: Gt (θ)
The process (Gt (θ))t⩾1 is called the gain process associated with the trading strategy θ.
Observe that for an arbitrary strategy θ, we have
∆Vt (θ) = Vt (θ) − Vt−1 (θ)

= θt · St − θt−1 · St−1
= θt · (St − St−1 ) + (θt − θt−1 ) · St−1
= θt · ∆St + (∆θt ) · St−1 .
Therefore, a trading strategy θ is self-financing if, and only if,
(∆θt ) · St−1 = 0.
2.4. HEDGING 17
2.4 Hedging
A contingent claim at maturity T is an FT -measurable random variable H : Ω → R. A
derivative on asset j can be seen as a contingent claim H that is a deterministic function
of the asset’s price at maturity T , i.e.,
H = f (STj )
j
for some function f : R → R. For instance, the contingent claim CT,K associated to a
European Call with maturity T and strike K written on asset j satisfies
h i+
j
CT,K := STj − K = max{STj − K, 0}.2
Since the price STj is random, the continent claim f (STj ) is typically a random variable.
A contingent claim H at maturity T is said to attainable when there exists a self-
financing trading strategy θ such that
VT (θ) = H.
Such a strategy is called a hedging or replicating strategy.

Assume that an infinite sequence of coin tosses represents uncertainty. Consider the
simpler case there are only two assets J = {j0 , j1 }. The prices S j0 and S j1 are denoted
respectively by S 0 and S since there is no ambiguity. A trading strategy (θtj0 , θtj1 )t⩾1 is
denoted by (θt0 , ∆t )t⩾0 .
Assume that the evolution of the price of the risky asset is as follows
St+1 − St
= µ + σZt+1
St
where (
+1, if ωt+1 = H
Zt+1 (ω) =
−1, otherwise.
This implies that condition to time t information, the price of the risk asset can take two
values at t + 1, which is exactly the number of possible shocks. This property is sufficient
to show that markets are complete in that for any contingent claim X at maturity T , a
self-financing trading strategy exists that replicates X. The argument can be made by
backward induction. Fix a date T −1 event ω T −1 . Contingent to this event, the contingent
claim can take at most two values at T , namely X(ω T −1 , H) and X(ω T −1 , L). We look
for two real numbers (θT0 (ω T −1 ), ∆T (ω T −1 ) such that
θT0 (ω T −1 )ST0 (ω T −1 , H) + ∆T (ω T −1 )ST (ω T −1 , H) = X(ω T −1 , H)
and
θT0 (ω T −1 )ST0 (ω T −1 , L) + ∆T (ω T −1 )ST (ω T −1 , L) = X(ω T −1 , L).
This is a linear system of two equations and two unknowns. There is a solution under the
assumption that σ > 0.
2
In other words, we have f (x) := [x − K]+ .
Fix a predecessor event ω T −1 . Since we look for a self-financing trading strategy, we

should find a pair (θT0 −1 (ω T −2 ), ∆T −1 (ω T2 )) such that
θT0 −1 (ω T −2 )ST0 −1 (ω T −2 , H) + ∆T −1 (ω T −2 )ST −1 (ω T −2 , H) =

θT0 (ω T −1 )ST0 −1 (ω T −2 , H) + ∆T (ω T −1 )ST −1 (ω T −1 , H)
together with
θT0 −1 (ω T −2 )ST0 −1 (ω T −2 , L) + ∆T −1 (ω T −2 )ST −1 (ω T −2 , L) =

θT0 (ω T −1 )ST0 −1 (ω T −2 , L) + ∆T (ω T −1 )ST −1 (ω T −1 , L).
We have again a system of two linear equations with two unknowns. There is a unique
solution under the assumption that σ > 0.
Repeating the above arguments backward, we can construct a self-financing trading
strategy that replicates X.
2.5 Arbitrage Opportunities

Recall that we consider an environment where agents can trade a fixed set J = {j0 , j1 , . . . , jd }
of d + 1 assets. Prices of each asset j are represented by a stochastic process (Stj )t⩾0 of
random variables Stj : Ω → R measurable with respect to the σ-algebra Ft .3 Asset j0 is
riskless and its price is Stj0 = eρt .
A trading strategy θ = (θt )t⩾1 is a vector valued stochastic process such that θt+1 :
Ω → RJ is chosen at date t and is therefore Ft -measurable.
An agent with horizon T chooses an optimal trading strategy θ = (θt )t⩾1 to maximize
his expected utility
T
X
U (c) = β t EP [u(ct )]
t=0
where θ is chosen to finance the non-negative consumption process (ct )t⩾0 , i.e., at t = 0
c0 + θ1 · S0 ⩽ e0
at every intermediate date t ∈ {1, 2, . . . , T − 1},
ct (ω) + θt+1 (ω) · St (ω) ⩽ θt (ω) · St (ω) + et (ω)
and at the terminal date
cT (ω) ⩽ θT (ω) · ST (ω) + eT (ω).
Let’s denote an optimal trading strategy by θ̄.

Proposition 2.1. Assume that the maximization problem of the agent has a solution θ̄.
Then, there does not exist another trading strategy θ such that
θ1 · S 0 ⩽ 0 (2.5.1)
3
This means that Stj (ω) only depends on the history of shock ω t .
2.5. ARBITRAGE OPPORTUNITIES 19
for every t ∈ {1, 2, . . . , T − 1} and every ω ∈ Ω
0 ⩽ [θt (ω) − θt+1 (ω)] · St (ω) (2.5.2)
and for every ω ∈ Ω

0 ⩽ θT (ω) · ST (ω) (2.5.3)
with at least one strict inequality in one of the inequalities in (2.5.1), (2.5.2) or (2.5.3).
A trading strategy θ satisfying (2.5.1), (2.5.2) and (2.5.3) with at least one strict
inequality is called an arbitrage opportunity.
Proof. If θ is an arbitrage opportunity, the agent can implement a consumption c̃ > c by

replacing the trading strategy θ̄ by the new strategy θ̄ + θ. This contradicts the optimality
of θ̄.
In a self-financing trading strategy, the inequalities in (2.5.2) are satisfied with equality.
This implies that θ is a self-financing arbitrage opportunity if, and only if, θ is self-
financing and satisfies
θ1 · S0 ⩽ 0 and θT (ω) · ST (ω) ⩾ 0, for all ω ∈ Ω
with at least one strict inequality.

Since there is a riskless asset, we can restrict attention to self-financing arbitrage
opportunities without any loss of generality. Formally, we can show that there exists an
arbitrage opportunity if, and only if, there exists a self-financing arbitrage opportunity.
One direction is obvious: if θ is a self-financing arbitrage opportunity, then it is an arbitrage
opportunity. To prove the converse, let θ be an arbitrage opportunity. Since θ1 · S0 ⩽ 0,
we can choose to invest the extra money in the riskless asset. We let φ1 ∈ RJ be defined
by
φj10 := θ1j0 − θ1 · S0 and φj1 := θ1j , for all j ̸= j0 .
Observe that φj10 ⩾ θ1j0 and
φ1 · S0 = 0.
At t = 1, we have (implicitly for every ω ∈ Ω)4
φ1 · S1 (ω) − θ2 (ω) · S1 (ω) = θ1 · S1 (ω) − θ2 (ω) · S1 (ω) + (−θ1 · S0 )S1j0

⩾ θ1 · S1 (ω) − θ2 (ω) · S1 (ω) ⩾ 0.
Again, we put the extra money in the bank account, i.e., we let φ2 (ω) ∈ RJ be defined by
φj20 (ω) := θ2j0 (ω) + [φ1 · S1 (ω) − θ2 (ω) · S1 (ω)] /S1j0
and we let unchanged the positions on the risky asset, i.e., φj2 (ω) := θ2j (ω) for every j ̸= j0 .
We have φj20 (ω) ⩾ θ2j0 (ω) and
φ1 · S1 (ω) = φ2 (ω) · S1 (ω).

4
We do not need to specify the state of nature when writing the price of the riskless asset j0 since its
price is deterministic. Recall that Stj0 = eρt .
We can continue this argument at every t ∈ {2, . . . , T − 1}:
φjt+1
0 j0
(ω) := θt+1 (ω) + [φt (ω) · St (ω) − θt+1 (ω) · St (ω)] /Stj0
and we let unchanged φjt+1 (ω) = θt+1

j
(ω) for every j ̸= j0 . We then have φjt+1
0 j0
⩾ θt+1 and
φt+1 (ω) · St (ω) = φt (ω) · St (ω).
At the terminal date T , we have
φT (ω) · ST (ω) ⩾ 0, for all ω ∈ Ω
with a strict inequality for at least one state ω.

We have thus proved that φ is a self-financing trading strategy that is also an arbitrage
opportunity. Actually, φ is a self-financing strategy that satisfies
φ1 · S0 = 0, φT · ST ⩾ 0 and EP [φT · ST ] > 0.
Equivalently,
V0 (φ) = 0, VT (φ) ⩾ 0 and EP [VT (φ)] > 0.
2.6 Stochastic Discount Factor

Theorem 2.1 (Fundamental Theorem of Finance). There does not exist an arbitrage
opportunity if, and only if, there exists a strictly positive adapted process π = (πt )t⩾0 such
that (πt St )t⩾0 is a martingale, that is,
πt St = EP [πt+1 St+1 |Ft ], for all t ⩾ 0. (2.6.1)
Since π0 is a positive constant, we can assume without any loss of generality that
π0 = 1. A strictly positive adapted process π satisfying π0 = 1 and (2.6.1) is called a
stochastic discount factor associated to the process of prices (St )t⩾0 .
It is important to observe that the pricing equation (2.6.1) is an equality in RJ , that
is, for every asset j, we have
πt Stj = EP [πt+1 St+1

j
|Ft ].
Proof of the Fundamental Theorem of Finance. We only prove the easy direction. As-
sume a strictly positive adapted process exists π satisfying (2.6.1). Assume, by way of
contradiction, that there exists an arbitrage opportunity. We have proved that this implies
that there exists a self-financing trading strategy θ such that
θ1 · S0 ⩽0 and θT (ω) · ST (ω) ⩾ 0, for all ω ∈ Ω
with at least one strict inequality. Since π is a stochastic discount factor, we must have
πT −1 (θT · ST −1 ) = EP [πT (θT · ST ) |FT −1 ] .
Since θ is self-financing, we have
θT · ST −1 = θT −1 · ST −1 .
2.7. EQUIVALENT MARTINGALE MEASURE 21
Combining the above two results, we have
πT −1 (θT −1 · ST −1 ) = EP [πT (θT · ST ) |FT −1 ] .
Since θT −1 is FT −2 -measurable, we have
πT −2 (θT −1 · ST −2 )) = EP [πT −1 (θT −1 · ST −1 ) |FT −2 ] = EP [πT (θT · ST ) |FT −2 ] .
Repeating the above arguments, we get
π0 (θ1 · S0 ) = EP [πT (θT · ST ) |F0 ] .
Since πT is strictly positive and θT · ST is non-negative and non-zero, we must have

EP [πT (θT · ST )] > 0. This contradicts the fact that θ1 · S0 ⩽0.
Fix a state of nature ω = (ω1 , ω2 , . . .). If π is a stochastic discount factor, then

1 X
πt (ω t )St (ω t ) = πt+1 (ω t , z)St+1 (ω t , z)P(Ωt+1 (ω t , z))
P(Ωt (ω t ))
z∈Z
which can be written as follows

X
St (ω t ) = At+1 (ω t , z)St+1 (ω t , z)
z∈Z
where
πt+1 (ω t , z)P(Ωt+1 (ω t , z))
At+1 (ω t , z) :=
πt (ω t )P(Ωt (ω t ))
can be interpreted as the price at date t contingent to the partial history ω t of the one-
period ahead contract paying 1 unit of numéraire at t + 1 contingent to outcome z.5
Similarly, the term
A1 (ω1 )A2 (ω2 ) . . . At (ωt ) = πt (ω t )P(Ωt (ω t ))
can be interpreted as the price at t = 0 of the contract that delivers one unit of numéraire
at date t contingent to the partial history ω t .
2.7 Equivalent Martingale Measure

Fix a state of nature ω = (ω1 , ω2 , . . .). If π is a stochastic discount factor, then
1 X
πt (ω t )St (ω t ) = πt+1 (ω t , z)St+1 (ω t , z)P(Ωt+1 (ω t , z)). (2.7.1)
P(Ωt (ω t ))
z∈Z
Analyzing the above equation for asset j0 , we have

1 X
πt (ω t )eρt = eρ(t+1) πt+1 (ω t , z)P(Ωt+1 (ω t , z)).
P(Ωt (ω t ))
z∈Z
5
Such a short-term asset is called an Arrow security.
S
We use the above property to construct a probability measure Q on t⩾1 Ft recursively:
Q(Ω1 (ω 1 )) := eρ π1 (ω 1 )P(Ω1 (ω 1 )).
Observe that
X
Q(Ω1 (ω 1 )) = 1.
ω 1 ∈Z
For every t ⩾ 0,
Q(Ωt (ω t )) := πt (ω t )eρt P(Ωt (ω t )).
Observe that
X X X
Q(Ωt+1 (ω t+1 )) = Q(Ωt (ω t )) = . . . = Q(Ω1 (ω 1 )) = 1.
ω t+1 ∈Z t+1 ω t ∈Z t ω 1 ∈Z
We omit the details to verify that the definition of Q can be extended the σ-algebra
σ(∪t⩾1 Ft ).
Observe that the conditional probability
Q(Ωt+1 (ω t+1 )) ρ πt+1 (ω

t+1 ) P(Ωt+1 (ω t , z))
Q(Ωt+1 (ω t+1 )|Ωt (ω t )) = = e .
Q(Ωt (ω t ) πt (ω t ) P(Ωt (ω t ))
This implies that the asset pricing equation (2.7.1) can be written as follows
X
e−ρt St (ω t ) = e−ρ(t+1) St+1 (ω t , z)Q(Ωt+1 (ω t , z)|Ωt (ω t )). (2.7.2)
z∈Z
Or, equivalently,
h i
e−ρt St = EQ e−ρ(t+1) St+1 |Ft , for all t ⩾ 0.
This means that the discounted price process (Sbt )t⩾0 defined by
Sbt := e−ρt St
is a martingale with respect to the probability Q. We also use the alternative terminology:
(Sbt )t⩾0 is a Q-martingale.
A probability measure Q such that (e−ρt St )t⩾0 is a Q-martingale is called an equiva-
lent martingale measure. The term equivalent comes from the property that Q and P
assign zero probability to the same set of events. The probability Q is also called a risk-
neutral probability measure. The Fundamental Theorem of Finance can be restated
as follows.
Theorem 2.2 (Fundamental Theorem of Finance). There does not exist an arbitrage
opportunity if, and only if, there exists an equivalent martingale measure.
2.8. DISCOUNTED VALUE OF SELF-FINANCING TRADING STRATEGIES 23
2.8 Discounted Value of Self-Financing Trading Strategies

We have seen in the proof of the Fundamental Theorem of Finance that if θ is a self-
financing strategy, then
πt (θt · St ) = EP [πt+1 (θt+1 · St+1 ) |Ft ], for all t ⩾ 0.
Using the probability Q constructed with the stochastic discount factor, we deduce that
eρt θt · St = EQ [e−ρ(t+1) θt+1 · St+1 |Ft ], for all t ⩾ 0.
In other words, the discounted value process (Vbt (θ))t⩾0 defined by
Vbt (θ) := e−ρt Vt (θ) = eρt θt · St
is a Q-martingale.
Fix an arbitrary contingent claim H at date T ; that is, H : Ω → R is an FT -measurable
random variable. Assume that H is attainable. Recall that this means that there exists a
self-financing strategy such θ such that
VT (θ) = H.
Since θ is self-financing, we get that (Vbt (θ))t⩾0 is a Q-martingale. This implies that
θt · St = eρt Vbt (θ) = eρt EQ [e−ρT H|Ft ] = e−ρ(T −t) EQ [H].
The amount θt · St can be interpreted as the resources the agent needs to spend at date t
to replicate the payoff H at date T . Assume there exists another self-financing strategy φ
that replicates H, i.e., VT (φ) = H. Following the above arguments, we have
φt · St = e−ρ(T −t) EQ [H] = θt · St .
The above property states that two self-financing trading strategies having the same value
at some date T , must have the same value at any previous date t < T . This is called the
Law of One Price.
2.9 The Cox-Ross-Rubinstein Binomial Market Model

Consider a market with two assets J = {j0 , j1 } where asset j1 can be interpreted as the
price of a stock. The price Stj0 of the riskless asset at date t is denoted by Dt . Recall that
Dt = eρt . The price Stj1 of the risky asset is denoted by St . We assume that
St+1 − St = αSt + σSt Zt+1
where α, σ ⩾ 0 and Zt : Ω → R is a random variable that takes two values +1 and −1. If
uncertainty is only driven by the stock price, the natural candidate for the sample space
is Ω := Z N with Z = {H, T }. Indeed, it suffices to pose
(
+1, if ωt = H
Zt (ω) :=
−1, otherwise.
Observe that
Dt+1 − Dt = rDt .
where ρ = ln(1 + r).
If σ = 0, then both assets j0 and j1 are riskless. In that case, there are no arbitrage
opportunities if, and only if, α = r. Moreover, markets are incomplete. Indeed, if W is a
contingent claim at t = 1 with W (H) ̸= W (L), we cannot find a portfolio θ1 ∈ RJ such
that (
j0 j1 j1 W (H), if ω1 = H
θ1 (1 + r) + θ1 (1 + α)S0 =
W (L), otherwise.
From now on, we assume that σ > 0. This implies that markets are complete.6
2.9.1 Equivalent Martingale Measure

Recall that (Sbt )t⩾0 is the process of discounted prices defined by Sbt = e−ρt St . There are
no arbitrage opportunities in this market if, and only if, there exists a probability measure
Q such that
Sbt = EQt [St+1 ],
b for all t ⩾ 0.
Equivalently, we should have
(1 + r)St = EQ Q
t [St+1 ] = (1 + α)St + σSt Et [Zt+1 ].
We then get that there are no arbitrage opportunities if, and only if,
r = α + σ EQ
t [Zt+1 ], for all t ⩾ 0.
Recall that EQ
t [Zt+1 ] is Ft -measurable and satisfies
t Q(Ωt+1 (ω t , H)) t Q(Ωt+1 (ω t , L))

EQ
t [Zt+1 ](ω) = Zt+1 (ω , H) + Z t+1 (ω , L) .
Q(Ωt (ω t )) Q(Ωt (ω t ))
Since Zt+1 (ω t , H) = 1 and Zt+1 (ω t , L) = −1, we have that
Q(Ωt+1 (ω t , H)) Q(Ωt+1 (ω t , L))

= q(H) and = q(L)
Q(Ωt (ω t )) Q(Ωt (ω t ))
where the pair (q(H), q(L)) ∈ (0, 1)2 is the unique solution of the system
q(H) + q(L) = 1 and r = α + σ(q(H) − q(L)).
The above system has a solution if, and only if,
α − σ < r < α + σ.
The solution is
r−α+σ σ−r+α
q(H) = and q(L) = .
2σ 2σ
6
At any partial history ω t , the agent can trade two assets with non-colinear future prices and face
uncertainty regarding t + 1 represented by two shocks, H or L.
2.9. THE COX-ROSS-RUBINSTEIN BINOMIAL MARKET MODEL 25
Observe that the random experiments (or, equivalently, the random variables (Zt )t⩾1 ) are
independent with respect to the probability Q since
Q(Ωt (ω t )) = q(ω1 )q(ω2 ) . . . q(ωt ).
It is worth pointing out that Q is unique and we do not need to specify the physical
probability P underlying the random experiment. Actually, the only relevant property
about P is that it assigns strictly positive probability to every finite time event of the form
Ωt (ω t ).
2.9.2 Asset Pricing: European Call

Assume now that there is a third asset j2 that is a European call option on the stock with
strike K and maturity T . We replace the notation Stj3 by Ct . The price at maturity is
given by
CT = [ST − K]+ = max{ST − K, 0}.
Assuming that there are no arbitrage opportunities in this market, we must have
bt = EQ [C
C t
bT ], or, equivalently, Ct = e−ρ(T −t) EQ +
t [(ST − K) ].
Let’s denote by Rt+1 the gross return of the stock from t to t + 1, i.e., St+1 = Rt+1 St .
Observe that the gross return is random since
Rt+1 = (1 + α) + σZt+1 .
Using the definition of gross return recursively, we have
ST = St Rt+1 Rt+2 . . . RT .
This implies that
Ct = e−ρ(T −t) EQ +

t (St Rt+1 Rt+2 . . . RT − K) .
Since St is Ft -measurable and Rt+1 . . . RT is independent of Ft , we deduce from the prop-
erties of conditional expectations (see Lemma 2.3.4 in Shreve (2004b)) that
Ct = vt (St ) where vt (x) := e−ρ(T −t) EQ (xRt+1 Rt+2 . . . RT − K)+ .

It is important to stress that the expectation in the definition of vt (x) is unconditional.

In particular, the price Ct is random, but the uncertainty involved in Ct is only driven by
the possible values of St .
We can solve explicitly for vt (x) since
τ
−ρτ
X τ +
q(H)u q(L)τ −u x(1 + α + σ)u (1 + α − σ)τ −u − K

vt (x) = e
u
u=0
where τ := T − t is time to maturity and

τ τ −u+1τ −u+2 τ
= ...
u 1 2 u
is the number of subsets of {t, . . . , T } with u elements indicating the dates s ∈ {t, . . . , T }
where ωs = H.
2.9.3 Replicating Strategy

We propose to show that we can construct a self-financing trading strategy (φ0t , ∆t , 0)t⩾1
combining the riskless asset and the stock that replicates the European Call with maturity
T and strike K. The standard approach is to solve by backward induction.
Fix an arbitrary partial history ξ := ω T −1 . We would like to find (φ0T (ξ), ∆T (ξ)) such
that
φ0T (ξ)(1 + r)DT −1 + ∆T (ξ)ST (ξ, H) = [ST (ξ, H) − K]+ = vT (ST (ξ, H))
and
φ0T (ξ)(1 + r)DT −1 + ∆T (ξ)ST (ξ, L) = [ST (ξ, L) − K]+ = vT (ST (ξ, L)).
We have two equations and two unknowns. The equation determines the unique solution
v(T, ST (ξ, H)) − v(T, ST (ξ, L))

∆T (ξ) = .
ST (ξ, H) − ST (ξ, L)
Since
ST (ξ, H) = ST −1 (ξ)[1 + α + σ] and ST (ξ, L) = ST −1 (ξ)[1 + α − σ]
we deduce that ∆T (ξ) can be written as δT (ST −1 (ξ)) where δT : R → R is defined by
vT (x(1 + α + σ)) − vT (x(1 + α − σ))

δT (x) = .
2xσ
Because there are no arbitrage opportunities, we must have
φ0T (ω T −1 )DT −1 (ω T −1 ) + ∆T (ω T −1 )ST −1 (ω T −1 ) = CT −1 (ω T −1 ) = vT −1 (ST −1 (ω T −1 )).
Fix now ω T −2 . The problem we should solve to find (φ0T −1 (ω T −2 ), ∆T −1 (ω T −2 )) is exactly

the same as the previous problem. We then get
∆T −1 (ω T −2 ) = δT −1 (ST −2 (ω T −2 ))
where
vT −1 (x(1 + α + σ)) − vT −1 (x(1 + α − σ))
δT −1 (x) := .
2xσ
Using a backward induction argument, we deduce that
∆t (ω t−1 ) = δt (St−1 (ω t−1 ))
where
vt (x(1 + α + σ)) − vt (x(1 + α − σ))
δt (x) := .
2xσ
Chapter 3
Continuous Time
Congratulations on successfully navigating through the intricacies of discrete-time dynamic

trading in financial markets. Your dedication and understanding of key concepts have laid
a solid groundwork for the next phase of our journey. In this chapter, we will embark on
an exciting exploration of ”Continuous-Time Dynamic Trading.”
As we transition from discrete to continuous time, we embrace a more refined frame-
work to model the dynamic evolution of financial markets. This continuous-time setting
allows us to capture the ever-changing nature of asset prices and the continuous flow of
information that influences market behavior. We will delve into a realm where mathemat-
ical tools like Brownian motion and Itô calculus become indispensable for modeling and
analyzing financial phenomena.
Key Topics to be Explored:
1. Scaled Random Walk: We will begin by connecting our discrete-time understand-

ing to continuous time through a scaled random walk. We bridge the gap between
discrete and continuous worlds by taking finer and finer time intervals.
2. Brownian Motion: The centerpiece of continuous-time modeling, Brownian mo-

tion, is a key building block for stochastic calculus. We will examine its properties
and understand its role as the driving force behind various financial models.
3. Itô’s Integral: Armed with Brownian motion, we can now introduce Itô’s integral,
a powerful tool for integrating stochastic processes with respect to Brownian motion.
This integral is vital in formulating the market value of dynamic trading strategies.
4. Itô-Doeblin Formula: We will explore the Itô-Doeblin formula, which provides a

means to differentiate functions involving stochastic processes, enabling us to derive
essential results in finance.
5. Itô Processes: Building upon the integral and formula, we will define Itô processes,
a class of stochastic processes that incorporate the dynamic behavior of financial
quantities, such as asset prices.
6. Risk-Neutral Pricing: An intriguing concept that simplifies the valuation of

derivatives, risk-neutral pricing allows us to price complex financial instruments
easily.
27
28 CHAPTER 3. CONTINUOUS TIME
7. Black-Scholes-Merton Formula: Among the most celebrated results in finance,

the Black-Scholes-Merton formula offers a powerful tool for option pricing, revealing
insights into market dynamics and implied volatilities.
8. Martingale Representation Theorem: We will explore the Martingale Repre-

sentation Theorem. This profound result connects martingales to Itô processes and
provides a deeper understanding of pricing and hedging strategies.
9. Complete Markets: Finally, we will examine complete markets, where all contin-
gent claims can be perfectly replicated. This concept plays a crucial role in asset
valuation by means of hedging strategies.
As we venture further into the realm of continuous-time dynamic trading, you will
witness the elegance and power of stochastic calculus in financial modeling. This chapter
marks a significant turning point in your understanding of the complexities inherent in
financial markets and the methods employed to make informed investment decisions.
Throughout this chapter, I encourage you to actively engage with the material, work
through examples, and embrace the challenge of mastering these advanced concepts. A
thorough grasp of continuous-time dynamic trading will empower you to analyze real-world
financial scenarios with precision and insight.
3.1 Scaled Random Walk

Uncertainty is represented by repeatedly tossing a fair coin: Z = {H, T }, Ω = Z N and
P(Ωt (ω t )) = 2−t . The σ-algebra Ft represents the information available up to date t.1
3.1.1 Symmetric Random Walk

Let Xt be the random variable defined by
(
1, if ωt = H
Xt (ω) =
−1, if ωt = L.
The random variable Xt is independent of all tosses preceding t, i.e., Xt and Ft−1 are
independent. We can interpret Xt as a possible unit increment, positive or negative.
Let (Mt )t⩾0 be the stochastic process defined by M0 = 0 and for every n ∈ N
Mt := Mt−1 + Xt = X1 + X2 + . . . + Xt .
At every period t, the value of the process can step up one unit (if Xt (ω) = 1) or down
one unit (if Xt (ω) = −1), and each of the two possibilities is equally likely. This process
is also called the symmetric random walk. When Xt (ω) = 1, we step forward. When
Xt (ω) = −1, we step backward.
1
Formally, Ft is the class of unions of the events of the form Ωt (z t ) when z t varies in Z t . It is also the
σ-algebra generated by the projection ω 7−→ ω t = (ω1 , ω2 , . . . , ωt ).
3.1. SCALED RANDOM WALK 29
Figure 3.1: 4 paths of the random walk t 7→ Mt (ω) with 20 periods
Increments. Fix n arbitrary time periods tk ∈ {0, 1, . . .} satisfying
0 = t0 < t1 < . . . < tn
The increments
(Mt1 − Mt0 , Mt2 − Mt1 , . . . , Mtn − Mtn−1 )
Figure 3.2: 10 paths of the random walk t 7→ Mt (ω) with 40 periods
form a family of independent random variables. Moreover for any two dates t > s, we
have
E(Mt − Ms ) = 0 and var(Mt − Ms ) = t − s.
The variance of the random walk accumulates at a rate of one per unit of time. Indeed,
E(Mt − Ms ) = E(Xs+1 ) + . . . + E(Xt ) = 0.
Moreover, since the random variables (Xs+1 , Xs+2 , . . . , Xt ) are independent, we have
var(Mt − Ms ) = var(Xs+1 ) + var(Xs+2 ) + . . . + var(Xt ) = 1 + 1 + . . . + 1 = t − s.
Martingale Property. We have already seen that the symmetric random walk is a
martingale. Indeed, fix two dates ti+1 > ti , then
Mti+1 − Mti = Xti +1 + . . . + Xti+1 .
For every t > ti , the random variable Xt is independent of the information Ft . This
implies that
E Mti+1 − Mti |Fti = E[Xti +1 ] + . . . + E[Xti+1 ].
Since the coin is fair, we have E[Xt ] = 0 for every t, and we get the desired result:

E Mti+1 |Fti = Mti .
Quadratic Variation. Fix an arbitrary stochastic process (Yt )t⩾0 . The quadratic
variation is another stochastic process ([Y, Y ]t )t⩾0 defined by
t
X
[Y, Y ]t := (Yτ − Yτ −1 )2 .
τ =1
Be careful, there is no expectation in the definition of [Y, Y ]t . The computation is made

path by path, that is, the quadratic variation is a non-negative stochastic process.
In the case of the symmetric random walk, we have
t
X t
X
∀ω ∈ Ω, [M, M ]t (ω) = (Mτ (ω) − Mτ −1 (ω))2 = (Xτ (ω))2 = 1.
τ =1 τ =1
It turns out that the quadratic variation of the random walk is not stochastic. This is a
very specific result.
3.1.2 Increasing the Frequency of Tosses

Up to now, we considered time periods described by
t ∈ {1, 2, . . .}
without specifying the time unit. It could be days, months, quarters, years. Fix once for
all a time unit, say one month.
Instead of tossing a coin every month, fix an integer n ∈ N and assume we toss the
coin n times during the month. Now, the coin is tossed at every date

(n) k
t ∈ Q := : k ∈ {0, 1, . . .} .
n
(n)
Let (Wt )t⩾0 the continuous time stochastic process defined as follows:
(a) for every date t of the form k/n with k ∈ {0, 1, . . .}, we pose
(n) 1 1
Wt := √ Mnt = √ Mk
n n
k k+1

(b) for any other date ∈ n, n , we let
(n)
Wt := at + b
where a and b are defined such that

(n) k k+1
t 7−→ Wt (ω) is continuous on ,
n n
There is a substantial difference between the sample paths of W (10) and W (100) . How-
ever, the sample paths of W (100) and W (200) are very similar.
Increments are independent at dates in Q(n) . Formally, for any
0 = t0 < t1 < . . . < tm with ti ∈ Q(n)
the family of increments

(n) (n) (n) (n) (n) (n)
Wt1 − Wt0 , Wt2 − Wt1 , . . . , Wtm − Wtm−1
(n)
Figure 3.3: Scaled Random Walk t 7→ Wt (ω) with n = 10
(n)
(n)
are independent.
For any tossing dates s, t ∈ Q(n) with s < t, we have

(n) (n)
E Wt − Ws(n) = 0 and var Wt − Ws(n) = t − s.
The variance of the scaled random walk accumulates (on special dates) at rate one per unit
of time. This is the main motivation for scaling the random walk. Otherwise, by increasing
the frequencies of tosses, the variance would be explosive when n tends to infinite.
(n)
The process (Wt )t∈Q(n) (indexed by the tossing dates) is a martingale. Indeed, fix
t > s two numbers in Q(n) and observe that
(n) (n)
E[Wt |Fs ] = E[(Wt − Ws(n) ) + Ws(n) |Fs ]
(n)
= E[Wt − Ws(n) |Fs ] + E[Ws(n) |Fs ]
(n)
= E[Wt − Ws(n) ] + Ws(n)
= Ws(n) .
For any tossing date t ∈ Q(n) and any ω ∈ Ω, we have
[W (n) , W (n) ]t (ω) = t.

(100) 2
Figure 3.6: Distribution of W1/4 and the normal curve y = √1 e−2x
2π
r
Indeed, fix t = n for some r ∈ N and observe that
r h i2
(n) (n)
X
[W (n) , W (n) ]t = Wk/n − W(k−1)/n
k=1
r 2
X 1
= √ Xk
n
k=1
r
X 1 r
= = = t.
n n
k=1
(n) (n)
Theorem 3.1 (Central Limit). Fix t ⩾ s. As n → ∞, the distribution of Wt − Wt
converges to the normal distribution with mean zero and variance t − s
(n) (n)
−Ws
lim PWt = N (0, t − s)
n→∞
The convergence is defined in terms of cumulative distributions

(n)
∀x ∈ R, lim P({Wt ⩽ x}) = Φt (x)
n→∞
where Φt is the cumulative distribution of N (0, t).

Equivalently, we have convergence of moment-generating functions
h (n)
i Z
∀u ∈ R, lim E euWt = eux ft (x)dx
n→∞ R
3.2. BROWNIAN MOTION 35
Figure 3.7: Cumulative Density Function
Figure 3.8: Probability Density Function
where
1 2
ft (x) := √ e−x /(2t) .
2π
3.2 Brownian Motion

The Brownian motion is obtained as the limit of the scaled random walks
(n)
Wt = lim Wt .
n→∞
The above limit is informal. Indeed, we have not defined the adequate sample space Ω,
the filtration (Ft )t⩾0 and the probability measure P on F, the σ-algebra generated by all
Ft . The construction of the scaled random motion is an artifact to motivate the following
proposition-definition.
Proposition 3.1. There exists a probability space (Ω, F, P) and, for every t ∈ [0, ∞), a
random variable
Wt : Ω → R
such that
(i) for each ω, the function t 7→ Wt (ω) is continuous with W0 (ω) = 0;
(ii) for any 0 ⩽ s < t, the increment Wt − Ws is a random variable that is normally
distributed with
E[Wt − Ws ] = 0 and var(Wt − Ws ) = t − s;
(iii) for any finite family 0 = t0 < t1 < . . . < tn , the increments in the family

Wt1 − Wt0 , Wt2 − Wt1 , . . . , Wtn − Wtn−1 ;
are P-independent.2
Recall that a filtration (Ft )t⩾0 represents the time evolution of observable information.
We assume that, at every date t, the realization of the random variable Wt is observed.
In other words, we assume that each Wt is Ft -measurable. Since Fs ⊆ Ft for every s < t,
we also have that Ws is Ft -measurable. Formally, we say that Ft contains the σ-algebra
σ((Ws )s∈[0,t] ) generated by all process (Ws )s∈[0,t] . We assume further that the σ-algebra
Ft does not carry any information that would allow us to predict future movements of the
Brownian motion, in the sense that the increment Wu − Wt for u > t is P-independent of
Ft . All these properties are collected in the following definition.
Definition 3.1. A filtration for the Brownian motion is a family (Ft )t⩾0 of σ-algebra
satisfying
(i) Information accumulates: If t > s, then Fs is a subset of Ft ;
(ii) Adaptivity: For each t ⩾ 0, the random variable Wt is Ft -measurable;
(iii) Independence of future increments: For any 0 ⩽ s < t, the increment Wt − Ws

is independent of Fs .
From now on, we fix a probability space (Ω, F, P) that carries a Brownian motion
(Wt )t⩾0 and we fix a filtration (Ft )⩾0 for (Wt )t⩾0 .
Proposition 3.2. The process (Wt )t⩾0 is a martingale.
2
Formally, a family (X 1 , . . . , X n ) of random variables are P-independent when
n
! n
\ Y
P {X i ⩽ xi } = P({X i ⩽ xi })
i=1 i=1
for any arbitrary family (x1 , . . . , xn ) ∈ Rn .

Proof. Fix t > s two numbers in [0, ∞). We have
E[Wt |Fs ] = E[(Wt − Ws ) + Ws |Fs ]

= E[Wt − Ws |Fs ] + E[Ws |Fs ]
= E[Wt − Ws ] + Ws
= Ws .
The steps in the above sequence of equalities are explained below:
(1) We introduce the increment Wt − Ws to use the independence assumption.
(2) We use linearity.
(3) We observe Ws at time s and Wt − Ws is independent of Fs .
(4) Increments have zero means.
It is important to be able to write and justify correctly the above proof since it combines
simply the important properties of conditional expectations.
First-Order Variation. Fix a date T > 0. A dissection of [0, T ] is a family Π =

(t0 , t1 , . . . , tn ) for some n ∈ N such that
0 = t0 < t1 < . . . < tn−1 < tn = T.
We define
∥Π∥ := max{ti − ti−1 : i ∈ {1, . . . , n}}.
This captures the “size” of the dissection Π.
The first-order variation (or total variation) of a function f : [0, ∞) → R at T is
defined by
n−1
X
FVT (f ) := lim |f (tj+1 ) − f (tj )|
∥Π∥→0
i=0
where Π = (t0 , t1 , . . . , tn ) is a dissection of [0, T ].
Proposition 3.3. If f is continuous differentiable on (0, ∞), then for any T ⩾ 0, we have
Z T
FVT (f ) = |f ′ (t)|dt.
0
Proof. Fix T > 0 and a dissection Π = (t0 , t1 , . . . , tn ) of [0, T ]. The idea is to apply the
Mean Value Theorem. For every i ∈ {0, . . . , n − 1}, there exists t⋆i ∈ (ti , ti+1 ) such that
f (tj1 ) − f (tj ) = f ′ (t⋆j )(tj+1 − tj ).
The rest follows from the definition of the integral.

Figure 3.9: Mean Value Theorem
Quadratic Variation. Fix a function f : [0, ∞) → R and T ⩾ 0. The quadratic

variation of f up to time T is defined to be
n−1
X
[f, f ]T = lim |f (ti+1 ) − f (ti )|2 .
∥Π∥→0
i=0
Proposition 3.4. If f is continuously differentiable on (0, ∞), then for any T ⩾ 0, we

have
[f, f ]T = 0.
Proof. Applying the Mean Value Theorem
|f (ti+1 ) − f (ti )|2 = |f ′ (t⋆i )(ti+1 − ti )|2 ⩽ ∥Π∥ |f ′ (t⋆i )|2 (ti+1 − ti ).
The last inequality follows from the following obvious observation:
(ti+1 − ti )2 ⩽ ∥Π∥ (ti+1 − ti ).
We then have
n−1
" #
X
[f, f ]T ⩽ lim ∥Π∥ |f (t⋆i )|2 (ti+1 − ti )
∥Π∥→0
i=0
n−1
X
= lim ∥Π∥ × lim |f (t⋆i )|2 (ti+1 − ti )
Π→0 ∥Π∥→0
i=0
Z T
= lim ∥Π∥ × |f ′ (t)|2 dt = 0
Π→0 0
Intuitively, it is difficult to see where the term

n−1
X
|f (ti+1 ) − f (ti )|2
i=0
goes when ∥Π∥ converges to 0. Indeed, since n must go to infinite, we are adding more
and more terms. At the same, the terms are smaller and smaller since f is continuous.
There is a race. For the first-order variation, we saw that the infinite number of sums
exactly compensates the smaller and smaller size of the variation |f (ti+1 ) − f (ti )| such
RT
that we converge to the integral 0 |f ′ (t)|dt. The term |f (ti+1 ) − f (ti )|2 of the quadratic
variation is much smaller than |f (ti+1 ) − f (ti )|. It is so small that, although we make an
infinite sum, the aggregate term converges to zero. This is because the quadratic increment
|f (ti+1 ) − f (ti )|2 is bounded above by M |ti+1 − ti |2 where M is an upper bound of |f ′ | on
the compact interval [0, T ]. The infinite sum of terms of the form |ti+1 − ti |2 converges to
zero. This is sufficient to get the desired result.
Instead of writing
n−1
X
lim (ti+1 − ti )2 = 0 (3.2.1)
∥Π∥→0
i=0
we could use the intuitive notation

Z T
dtdt = 0.
0
Actually, we use the more concise notation
dtdt = 0
Fix now a state ω ∈ Ω and consider the function t 7−→ Wt (ω). If this function were
continuously differentiable for every ω, we could apply the above result and conclude that
[W, W ]T = 0 for every T ⩾ 0. It turns out that for the Brownian motion, we get a
completely different result. Recall that for the scaled random walk, we proved that
[W (n) , W (n) ]T (ω) = T, ∀ω ∈ Ω.
Passing to the limit, we expect to get [W, W ]T (ω) = 1. We can provide formal proof of
this property directly from the definition.
Proposition 3.5. For all T ⩾ 0, we have
P{[W, W ]T = T } = 1
that is,
[W, W ]T (ω) = T for P-almost all ω.
We can conclude from the above result that for P-almost all ω, the sample path t 7→
Wt (ω) is not continuously differentiable. In a “typical” sample path, there are too many
oscillations.
Proof. To prove the above proposition, we fix T > 0 and a dissection Π = (t0 , t1 , . . . , tn )
of [0, T ]. Define
n−1
X 2
QΠ := Wti+1 − Wti .
i=0
Observe that QΠ is a random variable. To get the desired result, it is sufficient to show
that
lim E(QΠ ) = T and lim var(QΠ ) = 0.
∥Π∥→0 ∥Π∥→0
Since E[Wti+1 − Wti ] = 0, we have
E (Wti+1 − Wti )2 = var[Wti+1 − Wti ] = tj+1 − tj .

This implies that

n−1
X
E(QΠ ) = (tj+1 − tj ) = T.
i=0
We expect students to be able to prove the above property.
Observe that
var (Wti+1 − Wti )2 = E (Wti+1 − Wti )4 + (ti+1 − ti )2 −

− 2(ti+1 − ti ) E (Wti+1 − Wti )2 .

We can check that E[X 4 ] = 3 var(X)2 for any X normally distributed with zero mean.
This implies that
var (Wti+1 − Wti )2 = 3(ti+1 − ti )2 + (ti+1 − ti )2 − 2(ti+1 − ti )2

= 2(ti+1 − ti )2 .
We deduce that
n−1
X
var (Wti+1 − Wti )2

var(QΠ ) =
i=0
n−1
X
= 2(ti+1 − ti )2
i=0
n−1
X
⩽ 2 ∥Π∥ (ti+1 − ti ) = 2 ∥Π∥ T.
i=0
This is sufficient to get the desired result.
We proved that
n−1
X 2
lim Wti+1 − Wti = T, P-a.s.
Π→0
i=0
We write informally
dWt dWt = dt
We can show that
dWt dt = 0
The meaning of the above notation is
n−1
X
lim [Wti+1 − Wti ](ti+1 − ti ) = 0.
∥Π∥→0
i=0
3.3. ITÔ’S INTEGRAL 41
Proof. Let
∆Π (W ) := max |Wti+1 − Wti |.
0⩽i⩽n−1
We have
n−1
X
[Wti+1 − Wti ](ti+1 − ti ) ⩽ ∆Π (W )T.

i=0
Since W is continuous on [0, T ], it is uniformly continuous and we have
lim ∆Π (W ) = 0.
Π→0
3.2.1 Exercises
1. Fix a date T > 0 and a dissection Π = (t0 , t1 , . . . , tn ) of [0, T ]. Define QΠ :=
Pn−1 2
i=0 Wti+1 − Wti . Show that lim∥Π∥→0 var(QΠ ) = 0.
2. Fix a constant σ > 0. Let (Zt )t⩾0 be the process defined by

1 2
Zt := exp σWt − σ t .
2
Prove that the process (Zt )t⩾0 is a martingale.
3.3 Itô’s Integral

Consider a probability space (Ω, F, P), a Brownian motion (Wt )t⩾0 and a filtration (Ft )t⩾0
for the Brownian motion. Fix a maturity date T and consider an adapted process (∆t )t⩾0 .
What could be the formal definition of
Z T
∆t dWt ?
0
Itô’s Integral for Simple Processes. We shall start with simple processes. Fix a
dissection (t0 , t1 , . . . , tn ) of [0, T ], i.e.,
0 = t0 < t1 < . . . < tn = T.
Assume that the function t 7→ ∆t is constant on each [ti , ti+1 ). Since the process, (∆t )t⩾0 ,
is adapted, the value of the first step does not depend on ω. However, the value of the
second step can be stochastic but should be Ft1 -measurable
Definition 3.2. If (∆t )t⩾0 is a simple process associated with a dissection Π = (t0 , t1 , . . . , tn ),
and t is a time period in [tk , tk+1 ), then we pose
k−1
X
It = I(t) := ∆ti [Wti+1 − Wti ] + ∆tk [Wt − Wtk ]
i=0
Figure 3.10: A path of a simple process
Observe that the last term involves Wt and not Wtk+1 . This is important to get the
following property
Proposition 3.6. (It )t⩾0 is a stochastic process that is adapted to the filtration (Ft )t⩾0 .
The martingale property of the Brownian motion implies the following important re-
sult.
Theorem 3.2. The process (It )t⩾0 is a martingale.
Proof. Fix 0 ⩽ s < t ⩽ T . Recall that t 7→ ∆t is constant on eah interval [ti , ti+1 ) of the
dissection (t0 , t1 , . . . , tn ) of [0, T ]. Since 0 ⩽ s < t ⩽ T , there exists ℓ < k such that
s ∈ [tℓ , tℓ+1 ) and t ∈ [tk , tk+1 ).
This implies that
ℓ−1
X
It = ∆ti [Wti+1 − Wti ] + ∆tℓ [Wtℓ+1 − Wtℓ ]
i=0
k−1
X
+ ∆ti [Wti+1 − Wti ] + ∆tk [Wt − Wtk ].
i=ℓ+1
Since s > tℓ , we have

" ℓ−1 #
X ℓ−1
X
∆ti [Wti+1 − Wti ] Fs = ∆ti [Wti+1 − Wti ]

E

i=0 i=0
and
E[∆tℓ [Wtℓ+1 − Wtℓ ]|Fs ] = ∆tℓ [E[Wtℓ+1 |Fs ] − Wtℓ ] = ∆tℓ [Ws − Wtℓ ]
where the last equality follows from the martingale property of the Brownian motion. The
last term can be analyzed as follows. Fix i ∈ {ℓ + 1, . . . , k − 1}. Using the Tower Property
and the fact that s < ti , we get that

E ∆ti [Wti+1 − Wti ]|Fs = E E ∆ti [Wti+1 − Wti ]|Fti |Fs

= E ∆ti [E Wti+1 |Fti − Wti ]|Fs
= E [∆ti [Wti − Wti ]|Fs ]
where the last inequality follows from the martingale property of (Wτ )τ ⩾0 .
We have proved that the process (It )t⩾0 is a martingale. Since I0 = 0, we deduce that
E(It ) = 0, for all t ⩾ 0.
This implies that var(It ) = E(It2 ). We can compute the variance as follows.
Proposition 3.7. Z t
E(It2 ) = E (∆u )2 du.
0
Proof. Recall that ∆ is constant on eah interval [ti , ti+1 ) of the dissection (t0 , t1 , . . . , tn )
of [0, T ]. For each t < T , there exists k < n such that t ∈ [tk , tk+1 ). Recall that
k−1
X
It = ∆ti [Wti+1 − Wti ] +∆tk [Wt − Wtk ] .
| {z } | {z }
i=0
=:Di =:Dk
We have
k
X
It = ∆ti Di
i=0
which implies that
k
X X
(It )2 = (∆ti )2 (Di )2 + 2 ∆ti ∆tj Di Dj .
i=0 0⩽i<j⩽k
Take i < j. Observe that ∆ti ∆tj Di is Ftj . Since Dtj and Ftj are independent, we have
E[∆ti ∆tj Di Dj ] = E[∆ti ∆tj Di ] E[Dj ] = E[∆ti ∆tj Di ]0 = 0.
For i < k, use the independence of Di and Fti to show that
E[(∆ti )2 (Di )2 ] = E[(∆ti )2 ](ti+1 − ti ).
Finally, show that

E[(∆tk )2 (Dk )2 ] = E[(∆tk )2 ](t − tk ).
We have thus proved that
k−1
X Z t
E[(It )2 ] = E[(∆ti )2 ](ti+1 − ti ) + E[(∆tk )2 ](t − tk ) = E[(∆u )2 ]du.
i=0 0
We will assume at this point that we can switch the integrals (Fubini’s Theorem) to get
that Z t
2 2
var(It ) = E[(It ) ] = E (∆u ) du .
0
Theorem 3.3. The quadratic variation accumulated up to time t by the Itô integral sat-
isfies Z t
[I, I]t = (∆u )2 du.
0
Since (∆u )u⩾0 is a stochastic process, the quadratic variation [I, I]t is random variable
with Z t
∀ω ∈ Ω, [I, I]t (ω) = (∆u )2 (ω)du.
0
Proof. Fix an interval [ti , ti+1 ) on which t 7→ ∆t is constant. We compute the quadratic
variation accumulated by t 7→ It on [ti , ti+1 ). Fix an arbitrary dissection Π = (s0 , s1 , . . . , sm )
of [ti , ti+1 ]. Observe that
m−1 m−1
X 2 X 2
= (∆ti )2

Isk+1 − Isk Wsk+1 − Wsk .
k=0 k=0
We then get
m−1 Z ti+1
X 2
Isk+1 − Isk = (∆ti )2 (ti+1 − ti ) = (∆u )2 du.

lim
∥Π∥→0 ti
k=0
Summing over the intervals of the partition where t 7→ ∆t is constant, we get the desired
result.
Differential Equation. Fix a simple process (∆t )t⩾0 . When a stochastic process (Jt )t⩾0
satisfies Z t
Jt = J0 + ∆u dWu (3.3.1)
0
we use the following notation
dJt = ∆t dWt (3.3.2)
When J0 = 0, we get Jt = It . Eq. (3.3.1) is called the integral form while Eq. (3.3.2) is
called the differential form.
Informal Notation. We use the notation

dWt dWt = dt
to represent the property that [W, W ]t = t. Using the differential form equation, we have
dIt dIt = (∆t )2 dWt dWt = (∆t )2 dt.
This represents the property that (It )t⩾0 accumulates quadratic variation at rate (∆t )2
per unit of time.
Figure 3.11: Approximating a continuously varying path
Itô’s Integral for General Processes. Consider an adapted stochastic process

(∆t )t⩾0 such that
Z T
E (∆t )2 dt < ∞.
0
We would like to define Z T
∆t dWt .
0
Given a dissection Π = (t0 , t1 , . . . , tn ) of [0, T ], we consider the simple Process (∆Π

t )t⩾0
defined by
∆Π t := ∆ti where t ∈ [ti , ti+1 ).
There exists a sequence (Πn )n∈N of dissections such that

Z T 2
Πn
lim E ∆t − ∆t dt = 0.

n→∞ 0
We then pose Z T Z T
∆t dWt := lim ∆Π n
t dWt .
0 n→∞ 0
This integral inherits the properties of Itô’s integrals for simple stochastic processes.
Theorem 3.4. Fix an adapted stochastic process (∆t )t⩾0 such that
Z T
E (∆t )2 dt > ∞.
0
The Itô integral Z t

It := ∆u dWu
0
is well-defined and satisfies the following properties:
(i) the mapping t 7→ It is continuous for every path;
(ii) the process (It )t⩾0 is adapted;

Rt
(iii) the mapping ∆ 7→ 0 ∆u dWu is linear;
(iv) the process (It )t⩾0 is a martingale;

Rt
(v) E(It2 ) = E 0 ∆2u du;
Rt
(vi) [I, I]t = 0 ∆2u du.
3.4 Itô-Doeblin Formula

Let f : R → R be continuously differentiable. Standard analysis shows that
Z x
f (x) − f (0) = f ′ (u)du.
0
Naively, we may think that3

Z t
f (Wt ) − f (W0 ) = f ′ (Wu )dWu .
0
Because the quadratic variation is not zero, the above equation is not valid. We need an
extra term.
As a first step, we consider the simple function f : x 7→ (1/2)x2 . Taylor’s formula is
an equality because the nth -derivative satisfies f (n) = 0 for every n ⩾ 3:
1
f (y) − f (x) = f ′ (x)(y − x) + f ′′ (x)(y − x)2 .
2
Fix a dissection Π = (t0 , t1 , . . . , tn ) of [0, T ].
n−1
X
f (WT ) − f (W0 ) = f (Wti+1 ) − f (Wti )
i=0
n−1
X
= f ′ (Wti )[Wti+1 − Wti ]
i=0
n−1
1 X ′′ 2
+ f (Wti ) Wti+1 − Wti .
2
i=0
3
Be careful with notations. The left-hand side (LHS) term f (Wt ) − f (W0 ) is a random variable since,
for every state of nature ω ∈ Ω, we can compute f (Wt (ω)) − f (W0 ). The right-hand side (RHS) is also a
random variable since Itô’s integral is computed path by path. Formally, we have by definition
Z t Z t
f ′ (Wu )dWu (ω) = f ′ (Wu (ω))dWu (ω).
0 0
3.4. ITÔ-DOEBLIN FORMULA 47
Observe that f ′′ (Wt ) is constant and equal to 1 on the interval [0, T ]. This implies that
n−1 Z T
X
′′
2
lim f (Wti ) Wti+1 − Wti = [I, I]T = T = dt.
∥Π∥→0 0
i=0
Therefore, passing to the limit when ∥Π∥ → 0, we get

Z T
1 T ′′
Z
′
f (WT ) − f (W0 ) = f (Wt )dWt + f (Wt )dt.
0 2 0
The above formula is valid for general functions.
Theorem 3.5 (Itô-Doeblin Formula). Let f : [0, ∞)×R → R be a function (t, x) 7→ f (t, x)
such that the partial derivatives ft , fx and fxx exist and are continuous functions. Then,
for every T ⩾ 0,
Z T Z T Z T
1
f (T, WT ) − f (0, W0 ) = ft (t, Wt )dt + fx (t, Wt )dWt + fxx (t, Wt )dt. (3.4.1)
0 0 2 0
Proof. We use Taylor’s expansion
f (τ, y) − f (t, x) = ft (t, x)(τ − t) + fx (t, x)(y − x)

1
+ fxx (t, x)(y − x)2 + ftx (t, x)(τ − t)(y − x)
2
1
+ ftt (t, x)(τ − t)2 + high-order terms.
2
Fix an arbitrary dissection Π = (t0 , t1 , . . . , tn ) of [0, T ]. We do have
n−1
X
lim ftx (ti , Wti )(ti+1 − ti )[Wti+1 − Wti ] = 0
∥Π∥→0
i=0
because
lim max |Wtk+1 − Wtk | = 0.
∥Π∥→0 0⩽k⩽n−1
We use again the informal notation dWt dt = 0

We also have
n−1
X
lim ftt (ti , Wti )(ti+1 − ti )2 = 0.
∥Π∥→0
i=0
Indeed,
|ftt (ti , Wti )(ti+1 − ti )2 | ⩽ ∥Π∥ |ftt (ti , Wti )(ti+1 − ti )|.
We then get
n−1
X Z T
2
lim ftt (ti , Wti )(ti+1 − ti ) ⩽ lim ∥Π∥ ftt (t, Wt )dt.
∥Π∥→0 ∥Π∥→0 0
i=0
We use again the informal notation dtdt = 0.

By definition of the Riemann sums, we have

n−1
X Z T
lim ft (ti , Wti )(ti+1 − ti ) = ft (t, Wt )dt.
∥Π∥→0 0
i=0
By definition of Itô’s integral, we have

n−1
X Z T
lim fx (ti , Wti )(Wti+1 − Wti ) = fx (t, Wt )dWt .
∥Π∥→0 0
i=0
Using the property dWt dWt = dt, we have

n−1
X Z T
2
lim fxx (ti , Wti )(Wti+1 − Wti ) = fxx (t, Wt )dt.
∥Π∥→0 0
i=0
Informal Stochastic Calculus. Pose Xt = f (t, Wt ). Itô’s formula states that

Z T
1
XT − X0 = ft (t, Wt )dt + fx (t, Wt )dWt + fxx (t, Wt )dt .
0 2
In the same way, we have defined Itô’s integral
Z T
∆t dWt
0
we could also define an integral with respect to the process X and use the notation
Z T
∆t dXt .
0
Choosing ∆t = 1 for every t (and every ω), we have

Z T Z T
1
dXt = XT − X0 = ft (t, Wt )dt + fx (t, Wt )dWt + fxx (t, Wt )dt .
0 0 2
Removing the integral symbol from both sides of the equation and replacing Xt by f (t, Wt ),
we get the following “informal” notation
1
d[f (t, Wt )] = ft (t, Wt )dt + fx (t, Wt )dWt + fxx (t, Wt )dt. (3.4.2)
2
This is simply a notation and has no other meaning than Itô’s formula in the integral
form. Equation (3.4.2) is called the differential form of Itô’s formula.
The above differential form can be obtained informally. Indeed, Taylor’s formula im-
plies
1
d[f (t, Wt )] = ft (t, Wt )dt + fx (t, Wt )dWt + fxx (t, Wt )dWt dWt
2
1
+ ftx (t, Wt )dtdWt + ftt (t, Wt )dtdt.
2
Using the informal formulas dtdWt = 0 and dtdt = 0, we get Itô’s formula in differential
form.
3.5. ITÔ PROCESSES 49
Example 3.1. We may believe that

Z T
1
Wt dWt = (Wt )2 .
0 2
Applying Itô’s formula for f : x 7→ (1/2)x2 , we can show that

Z T
1 1
Wt dWt = (Wt )2 − T.
0 2 2
3.5 Itô Processes

A process (Xt )t⩾0 is said to be an Itô process when
Z t Z t
Xt = X0 + ∆u dWu + Θu du (3.5.1)
0 0
where X0 ∈ R and (∆t )t⩾0 and (Θt )t⩾0 are adapted processes satisfying
Z t Z t
2
E (∆u ) du < ∞ and |Θu |du < ∞, P-a.s.
0 0
Instead of Eq. (3.5.1), we can use the differential form expression
dXt = ∆t dWt + Θt dt.
Proposition 3.8. The quadratic variation of the Itô process (Xt )t⩾0 defined by dXt =
∆t dWt + Θt dt is Z t
[X, X]t = (∆u )2 du.
0
The differential form of the above result is
d[X, X]t = (∆t )2 dt.
Proof. For an informal proof, start with the differential form dXt = ∆t dWt + Θt dt. Write
dXt dXt and use the properties dWt dWt = dt, dtdWt = 0 and dtdt = 0. For a formal proof,
let Z t Z t
It := ∆u du and Rt = Θu du.
0 0
To compute [X, X]t , fix a dissection Π = (t0 , t1 , . . . , tn ) of [0, t].
n−1
X n−1
X n−1
X
[Xti+1 − Xti ]2 = [Iti+1 − Iti ]2 + [Rti+1 − Rti ]2
i=0 i=0 i=0
n−1
X
+2 [Iti+1 − Iti ][Rti+1 − Rti ].
i=0
Passing to the limit in the first term, we have

n−1
X Z t
2
lim [Iti+1 − Iti ] = [I, I]t = (∆u )du.
∥Π∥→0 0
i=0
To analyze the other terms, consider a function f : [0, ∞) → R and let
ζΠ (f ) := max |f (ti+1 ) − f (ti )|.

0⩽i⩽n−1
If f is continuous, we have ([0, t] is compact)
lim ζΠ (f ) = 0.
∥Π∥→0
Pn−1
We let AΠ := i=0 [Rti+1 − Rti ]2 .
n−1
X
AΠ ⩽ ζΠ (R) |Rti+1 − Rti |
i=0
n−1
X Z ti+1

⩽ ζΠ (R)
Θu du
i=0 ti
n−1
X ti+1
Z
⩽ ζΠ (R) |Θu | du
i=0 ti
Z t
⩽ ζΠ (R) |Θu | du.
0
Rt
Passing to the limit, we get lim∥Π∥→0 AΠ = 0 since 0 |Θu | du is finite.
We let BΠ := n−1
P
i=0 [Iti+1 − Iti ][Rti+1 − Rti ].
n−1
X
BΠ ⩽ ζΠ (I) |Rti+1 − Rti |
i=0
n−1
X Z ti+1

⩽ ζΠ (I)
Θu du
i=0 ti
n−1
X ti+1
Z
⩽ ζΠ (I) |Θu | du
i=0 ti
Z t
⩽ ζΠ (I) |Θu | du.
0
Rt
Passing to the limit, we get lim∥Π∥→0 BΠ = 0 since 0 |Θu | du is finite.
Integral with Respect to an Itô Process. Let (Xt )t⩾0 be an Itô process defined by
dXt = Θt dt + ∆t dWt
3.5. ITÔ PROCESSES 51
where the processes (Θt )t⩾0 and (∆t )t⩾0 satisfy

Z t Z t
2
E (∆u ) du < ∞ and |Θu |du < ∞, P-a.s.
0 0
Fix an adapted process (Γt )t⩾0 .
Definition 3.3. If
Z t Z t
E (Γu )2 (∆u )2 du < ∞ and |Γu Θu |du < ∞, P-a.s.
0 0
then the integral of (Γt )t⩾0 with respect to (Xt )t⩾0 is defined by
Z t Z t Z t
Γu dXu := Γu ∆u dWu + Γu Θu du.
0 0 0
Itô-Doeblin Formula for Itô Processes.
Theorem 3.6. Let f : (t, x) 7→ f (t, x) from [0, ∞)×R to R be a function twice continuously
differentiable and let (Xt )t⩾0 be an Itô process. For any T ⩾ 0,
Z T Z T Z T
1
f (T, XT ) = f (0, X0 ) + ft (t, Xt )dt + fx (t, Xt )dXt + fxx (t, Xt )d[X, X]t .
0 0 2 0
Recall that d[X, X]t = (∆t )2 dt when dXt = ∆t dWt + Θt dt. The differential form
expression of the Itô-Doeblin formula is
1
df (t, Xt ) = ft (t, Xt )dt + fx (t, Xt )dXt + fxx (t, Xt )dXt dXt .
2
Or, equivalently,

1 2
df (t, Xt ) = ft (t, Xt ) + fx (t, Xt )Θt + fxx (t, Xt )(∆t ) dt + fx (t, Xt )∆t dWt .
2
Generalized Geometric Brownian Motion. Let (αt )t⩾0 and (σt )t⩾0 be two adapted
processes satisfying for every T ⩾ 0
Z T Z T
E (σt )2 dt < ∞ and |αt |dt < ∞, P-a.s.
0 0
Let (Xt )t⩾0 be the Itô process satisfying X0 = 0 and defined by

1 2
dXt = σt dWt + αt − σt dt.
2
Let (St )t⩾0 be the process (asset price) defined by
St = S0 exp(Xt ) = S0 eXt
for some S0 > 0. Applying Itô’s formula, we get
dSt = αt St dt + σt St dWt .
The asset price St has an instantaneous mean rate of return αt and volatility σt . When α
and σ are constant, we have the usual geometric Brownian motion model
St = S0 exp σWt + (α − σ 2 /2)t

Observe that
St = S0 eαt Mt
where (Mt )t⩾0 is a martingale.
3.6 Itô Processes: Examples

Recall that a process (St )t⩾0 is said to be an Itô process when
Z t Z t
St = S0 + au du + bu dWu , for all t ⩾ 0 (3.6.1)
0 0
where S0 ∈ R is the initial value of the process and (at )t⩾0 and (bt )t⩾0 are adapted processes
satisfying, for every t ⩾ 0,
Z t Z t
2
E (bu ) du < ∞ and |au |du < ∞, P-a.s.
0 0
Instead of Eq. (3.6.1), we can use the differential form expression
dSt = at dt + bt dWt .
We analyze the case where the processes (at )t⩾0 and (bt )t⩾0 take the following form
at = α(t, St ) and bt = β(t, St )
for two functions α : [0, ∞) × R → R and β : [0, ∞) × R → R. We propose to analyze

several examples.
3.6.1 Linear constant coefficient

Assume there are two numbers ā and b̄ such that
α(t, x) = ā and β(t, x) = b̄, ∀(t, x) ∈ [0, ∞) × R.
In this case, the behavior of (St )t⩾0 seems to fluctuate around a straight line with
slope equal to ā. The random variable St can, in principle, become negative even when ā
is positive. This process should not be used to represent stock prices, for instance.
It is trivial to solve this SDE since we have
St = S0 + āt + b̄Wt , for all t ⩾ 0.
To simulate several sample paths of this stochastic process, we can code in Python and
LATEX. See the example below.
1 # library for plotting
2 import matplotlib.pyplot as plt
3 plt.rcParams.update({'font.size': 16})
4 # choosing the font and allowging for the use of LaTeX
5 from matplotlib import rc
6 rc('font',**{'family':'serif','serif':['Calibri']})
7 rc('text', usetex=True)
8 # defining tick size
9 plt.rc('xtick', labelsize=16)
10 plt.rc('ytick', labelsize=16)
11 # defining figure size
12 plt.rcParams["figure.figsize"] = (7,4.5)
13 # library for computations
14 import numpy as np
15 # horizon = number of time units (could be years, months, quarters, weeks, days)
16 T = 2
17 # number of steps within a time period
18 N=300
19 #initial value
20 S_0 = 3
21 #trend coefficient
22 mu = 1
23 #drift coefficient
24 sigma = 1
25 # number of sample paths
26 k = 4
27 # defining a function that constructs an array containing the values of a sample path
28 def generate_data():
29 # define a variable that will change
30 S = S_0
31 # create an array with S as a single value
32 S_values = [S]
33 # we shall collect of the values of process in the array S_value
34 for i in range(N*T): # steps
35 # construct the new value of S from the old value
36 S = S + mu*1/N + sigma*np.random.normal(0,np.sqrt(1/N),size=None)
37 # add the new value as an additional element in the array S_values
38 S_values.append(S)
39 return S_values
40 # define the figure
41 fig, axes = plt.subplots()
42 # compute several sample paths
43 for i in range(k):
44 axes.plot(generate_data(), label = f'$\omega_{i}$')
45 # defining the title
46 axes.set(title = f'$dS_t= \mu dt + \sigma dW_t$ with $\mu={mu}$ and $\sigma={sigma}$')
47 # defining the x ticks
48 axes.set_xticks([0,N*T])
49 # defining the x ticks labels
50 axes.set_xticklabels([0,f'{T}'])
51 # include the legend
52 plt.legend()
53 # show the plots in a separate file
54 plt.show()
Figure 3.12: dSt = ādt + b̄dWt with ā = 1.5 and b̄ = 0.5
Figure 3.13: dSt = ādt + b̄dWt with ā = 1 and b̄ = 1

3.6. ITÔ PROCESSES: EXAMPLES 55
Figure 3.14: dSt = ādt + b̄dWt with ā = 0.2 and b̄ = 2
3.6.2 Geometric
Assume that there are two numbers µ and σ such that
α(t, x) = µx and β(t, x) = σx, ∀(t, x) ∈ [0, ∞) × R.
This implies that

dSt = µSt dt + σSt dWt . (3.6.2)
In this case, both drift (at ) and diffusion (bt ) terms depend on the information set available
at time t through change in St itself; both drift and diffusion will change proportionally
to the price levels. To see this, note that dSt /St represents percentage changes in St . In
the geometric process, the behavior of St seems to fluctuate around an exponential trend.
Differently from the linear constant case, St cannot become negative (if it starts positive);
as St gets close to zero, the innovation term also becomes smaller (and at a faster rate,
square root of dt), which impedes St to turn negative.
To solve the SDE (3.6.2), we assume that St > 0 and introduce the auxiliary process
(Xt )t⩾0 , defined by Xt = ln(St ). Applying Itô’s Lemma, we have
1
dXt = f ′ (St )dSt + f ′′ (St )dSt dSt .
2
Since f ′ (x) = 1/x and f ′′ (x) = −1/x2 , we get
1
dXt = µdt + σdWt − σ 2 dt = (µ − σ 2 /2)dt + σdWt .
2
The above expression means
Z t Z t
Xt = X0 + (µ − σ 2 /2)dt + σdWt = X0 + (µ − σ 2 /2)t + σWt .
0 0
Figure 3.15: dSt = µSt dt + σSt dWt with µ = 0.2 and σ = 0.2
We then deduce that
St = exp(Xt ) = S0 exp (µ − σ 2 /2)t + σWt .

The process (St )t⩾0 is called a geometric Brownian motion. Each random variable St has
a log-normal distribution.
3.6.3 Square-root
Assume that there are two numbers µ and η such that
√
α(t, x) = µx and β(t, x) = η x, ∀(t, x) ∈ [0, ∞) × R.
This implies that p

dSt = µSt dt + η St dWt .
As with the case of geometric process, both drift (at ) and diffusion (bt ) terms depend
on St . Also, St will behave as fluctuating around an exponential trend. However, now the
variance (and not the standard deviation) remains proportional to prices. This process
should be applied to processes that require St to remain positive throughout but with a
variance that does not increase too much when St increases.
3.6.4 Mean-Reverting Process

Assume that there are four numbers S̄, λ, η and σ such that
√
α(t, x) = λ(S̄ − x) and β(t, x) = η x, or β(t, x) = σx ∀(t, x) ∈ [0, ∞) × R.
Figure 3.16: dSt = µSt dt + µSt dWt with µ = 0.4 and µ = 0.2
Figure 3.17: dSt = µSt dt + σSt dWt with µ = 0.2 and σ = 0.6
Figure 3.18: dSt = µSt dt + σSt dWt with µ = 0.05 and σ = 1
Figure 3.19: dSt = µSt dt + σSt dWt with µ = 1 and σ = 0.25

√
Figure 3.20: dSt = µSt dt + η St dWt with µ = 0.07 and η = 0.25
This implies that

p
dSt = λ(S̄ − St )dt + η St dWt or dSt = λ(S̄ − St )dt + σSt dWt .
The main difference now is that asset prices tend to revert to the mean value S̄ (long-
run mean). Let λ > 0. Note that whenever prices are above the mean, the drift will be
negative, forcing prices to fall eventually. In contrast, if prices are below the mean, the
drift will be positive which forces prices to rise. The speed of mean reversion is controlled
for by the parameter λ; the higher its value the faster the convergence to µ. Note that in
this case, St can become negative.
3.6.5 Ornstein-Uhlenbeck Process

Assume that there are two numbers θ and σ such that
α(t, x) = −θx and β(t, x) = σ ∀(t, x) ∈ [0, ∞) × R.
This implies that

dSt = −θSt dt + σdWt .
To get an explicit expression of St , we let Xt = f (t, St ) where
f (t, x) = exp(θt)x.
Applying Itô’s formula, we have

1
dXt = ft (t, St )dt + fx (t, St )dSt + fxx (t, st )dSt dSt .
2
We then get
dXt = θeθt St dt + eθt dSt = eθt σdWt .
Figure 3.21: dSt = λ(S̄ − St )dt + σSt dWt with S̄ = 10, λ = 1.5 and σ = 0.25
Figure 3.22: dSt = λ(S̄ − St )dt + σSt dWt with S̄ = 2, λ = 1.5 and σ = 0.25
Figure 3.23: dSt = −θSt dt + σdWt with θ = 1 and σ = 0.25
This means that

Z t
Xt = X0 + eθu σdWu
0
and we get
Z t
−θt
St = e S0 + σ e−θ(t−u) dWu .
0
3.6.6 Stochastic Volatility Process
Consider the following SDE
dSt = µSt dt + σt St dWt1
where the coefficient σt is itself a random process defined with respect to another Brownian
motion (Wt2 )t⩾0 that is not correlated with the Brownian motion of the innovation in stock
prices Wt1 . For example, we could instead assume that (σt )t⩾0 is a mean reverting process
specified by
dσt = λ(σ0 − σt )dt + γσt dWt2 .

3.7 Risk-Neutral Pricing

Consider a financial market with two assets: a riskless asset and a stock. At date t, the
price of the riskless asset is denoted by Dt and the price of the stock is St . A trading
strategy is a process (θt )t⩾0 with θt = (Γt , ∆t ) ∈ R2 . The amount ∆t represents the
holdings of the stock and Γt those of the riskless asset. The value Vt (θ) at date t of the
trading strategy is
Vt (θ) := Γt Dt + ∆t St .
A trading strategy (θt )t⩾0 is self-financing when
Z t Z t
Vt (θ) = V0 (θ) + Γs dDs + ∆s dSs .
0 0
We can use the differential form
dVt (θ) = Γt dDt + ∆t dSt .
Assume that the dynamics of asset prices are driven by the following stochastic differ-
ential equation
dDt = rDt dt and dSt = αSt dt + σSt dWt .
We abuse notations and write Vt for Vt (θ). Observe that
dVt = Γt dDt + ∆t dSt

= rΓt Dt dt + ∆t dSt
= r[Vt − ∆t St ]dt + ∆t dSt
= r[Vt − ∆t St ]dt + ∆t [αSt dt + σSt dWt ]
= rVt dt + ∆t (α − r)St dt + ∆t σSt dWt
• The term rVt dt represents the riskless part of the return.

3.7. RISK-NEUTRAL PRICING 63
• The term ∆t (α − r)St dt is the risk-premium for investing in the stock (proportional
to stock holdings).
• The term ∆t σSt dWt is the volatility term (proportional to stock holdings).
3.7.1 Discounted Prices and Values

In the discrete-time model, we have seen that there are no arbitrage opportunities if, and
only if, the discounted price process is a martingale under some appropriate probability
measure. To show that our market does not allow for arbitrage opportunities, we start by
analyzing the dynamics of the discounted price.
Let (Sbt )t⩾0 be the discounted process defined by
Sbt := e−rt St
Applying the Itô-Doeblin formula for f (t, x) = e−rt x, we get that
1
dSbt = ft (t, St )dt + fx (t, St )dSt + fxx (t, St )dSt dSt
2
= −re−rt St dt + e−rt dSt
= σ Sbt [µdt + dWt ]
where
α−r
µ :=
σ
is called the market price of risk.
We can obtain a similar result for the discounted value of a self-financing strategy. Let
(Vbt )t⩾0 be the discounted process defined by
Vbt := e−rt Vt .
Apply the Itô-Doeblin formula for f (t, x) = e−rt x.
1
dVbt = ft (t, Vt )dt + fx (t, Vt )dSt + fxx (t, Vt )dVt dVt
2
−rt −rt
= −re Vt dt + e dVt
= (α − r)∆t e−rt St dt + σ∆t e−rt St dWt
= ∆t dSbt .
3.7.2 Martingale Equivalent Measure

Recall that the discounted price process (Sbt )t⩾0 satisfies the stochastic differential equation
dSbt = σ Sbt [µdt + dWt ] .
To prove that (Sbt )t⩾0 is a martingale for a suitably chosen probability measure, we use
the following result.
Theorem 3.7 (Girsanov). Let (Wt )t⩾0 be a Brownian motion on a probability space
(Ω, F, P) and let (Ft )t⩾0 be a filtration for this Brownian motion. The process (W
ft )t⩾0
defined by
W
ft := Wt + µt
is a Brownian motion with respect to some probability measure Q.
To characterize the probability measure Q, we let (Zt )t∈[0,T ] be the strictly positive
process
1 2 µ2
Zt := exp −µWt − µ t = e−µWt − 2 t .
2
Apply Itô-Doeblin’s formula to show that
dZt = −µZt dWt .
This implies that (Zt )t∈[0,T ] is a martingale and
E[Zt ] = 1.
For every random variable Xt that is Ft -measurable, we have
EQ [Xt ] = EP [Xt Zt ].
3.7.3 Risk-Neutral Measure

Recall that the process (Sbt )t∈[0,T ] of discounted prices satisfies
dSbt = σ Sbt [µdt + dWt ]
where µ = (α − r)/σ is the market price of risk. Applying Girsanov’s Theorem, we have
dSbt = σ Sbt dW
ft
where (Wft := Wt + µt)t∈[0,T ] is a Brownian motion under Q. Since the process (Sbt )t∈[0,T ]
of discounted prices satisfies Z t
St = S0 +
b b σ Sbu dW
fu ,
0
we can deduce that (Sbt )t∈[0,T ] is a martingale under Q. In particular, we have
Sbt = EQ [SbT |Ft ].
The probability Q is known as a risk-neutral measure.

Recall that
dSt = αSt dt + σSt dWt .
Using dWt = −µdt + dW
ft , we get
dSt = rSt dt + σSt dW

ft .
3.8. BLACK-SCHOLES-MERTON FORMULA 65
The mean rate of return of the stock price is equal to the risk-less interest rate under Q:

1 2 ft = S0 ert eσW ft − σ2 t
St = S0 exp r − σ t + σW 2 .
2
Recall that the discounted value (Vbt )t⩾0 of the portfolio process where the stock holdings
are given by (∆t )t∈[0,T ] satisfies
dVbt = ∆t dSbt .
Since dSbt = σ Sbt dW
ft , we get that
dVbt = σ∆t Sbt dW

ft .
This means that Z t

Vbt = Vb0 + σ∆u Sbu dW
fu .
0
We deduce that (Vbt )t⩾0 is a martingale under Q.
3.7.4 Hedging
Let Y be an FT -measurable random variable. It can represent the payoff at time T of
derivative security. Assume that Y can be replicated in the sense that there exists a
self-financing portfolio strategy θ such that
VT (θ) = YT .
We omit from now on the dependence on θ. Since (Vbt )t∈[0,T ] is a martingale process under
Q, we have
Vbt = EQ [VbT |Ft ]
or, equivalently,
Vt = e−r(T −t) EQ [Y |Ft ].
3.8 Black-Scholes-Merton Formula

Consider a financial market with three assets: the riskless asset, the stock and a European
call on the stock with maturity T and strike K. We (always) assume that this market is
arbitrage-free.
We assume that the prices of the riskless and the stock satisfy the following stochastic
differential equations
dDt = rDt dt and dSt = αSt dt + σSt dWt .
where µ := (α − r)/σ. We have seen that applying Girsanov’s Theorem, the process
(W
ft )t⩾0 defined by
W
ft := Wt + µt
is a Brownian motion with respect to some probability measure Q.4 Recall that

ft and dSbt = σ Sbt dW
ft .
This implies that the process of discounted stock prices (Sbt )t∈[0,T ] is a martingale under
Q.5 In other words, Q is an equivalent martingale measure for the market composed of
the riskless asset and the stock.
We denote by (Ct )t⩾0 the process representing the prices of the European call. We
would like to show that, under the assumption of no-arbitrage, we can derive properties
of the process (Ct )t⩾0 from those of the underlying asset (the stock). At this point, it is
important to stress that we do not know whether the discounted price of call (C) b t⩾0 is a Q-
martingale. Indeed, from the assumption of no-arbitrage, we know from the Fundamental
Theorem of Finance that there exists a probability measure Q e such that (Sbt )t⩾0 and
(Ct )t⩾0 are Q-martingales. However, we do not know whether Q = Q.
b e e Observe that Q has
been constructed from the assumptions we made on the stock prices and using Girsanov’s
Theorem. The probability Q e comes from the Fundamental Theorem of Finance. A priori,
we may have Q ̸= Q. This is a problem if we want to compute conditional expectations.
e
Indeed, since (Cbt )t⩾0 is a Q-martingale,
e we have
bt = EQe [C
C bT |Ft ].
The above property is important because we know the price of the derivative at maturity:
CT = [ST − K]+ . Recalling that C
bt = exp(−rt)Ct , we derive that
Ct = e−r(T −t) EQ [(ST − K)+ |Ft ].

e
The problem with the above equation is that it is not fruitful for computation. Indeed,
we do not know how to write ST as an explicit function of some variables with interesting
properties concerning the probability Q.
e Recall that
ST = S0 exp{(r − σ 2 /2)t + σ W
fT }
The process (W ft )t⩾0 is a martingale with respect Q, but we have no information regarding
Q.
e
There is a way to overcome this difficulty. The idea is to hedge the derivative using the
two initial assets: the riskless and the stock. Formally, assume there exists a self-financing
strategy (θt )t⩾0 with θt = (Γt , ∆t , 0) where the first coordinate represents holdings to
the riskless asset, the second represents the stock holdings, while the third represents the
holdings of the derivative6 such that
VT (θ) = CT = [ST − K]+ .
The trading strategy η defined by ηt := (0, 0, 1) is also self-financing and satisfies VT (η) =
CT . We have assumed that the market with three assets is arbitrage-free. This implies
4
Girsanov’s Theorem also provides explicitly the density of Q with respect to P.
5
Since D
b t = 1 for every t ⩾ 0, it is obviously that this process is a Q-martingale.
6
Holdings are in the same unit as the price. If St is the price of one stock of a firm, then ∆t = 1 means
one stock.
that there exists a probability Q

e such that the processes (Vbt (θ))t⩾0 and (Vbt (η))t⩾0 are
Q-martingales. We deduce that
e
Vbt (θ) = EQ [VbT (θ)|Ft ] = EQ [VbT (η)|Ft ] = Vbt (η).

e e
This leads to
Ct = Vt (η) = Vt (θ).
Now, we use the fact that θ is a portfolio using only the riskless and stock. Therefore, we
can use Q (and not Qe to get that
Ct = Vt (θ) = ert Vbt (θ) = e−r(T −t) EQ [VT (θ)|Ft ] = e−r(T −t) EQ [(ST − K)+ |Ft ].
Recall that
ft .
This implies that

1 2
St = S0 exp r − σ t + σWt
2
If we pose τ := T − t for time-to-maturity, then we have

1 2
ST = St exp σ(WT − Wt ) + r − σ τ
f f
2
√

1 2
= St exp −σ τ Y + r − σ τ
2
where
fT − W
W ft
Y := − √
τ
is a random variable with standard normal distribution.
We have shown that
" + #
√

1
Ct = EQ e−rτ St exp −σ τ Y + r − σ 2 τ − K Ft

2
Brutal Computation. Since the random variable

√

1
exp −σ τ Y + r − σ 2 τ
2
is independent of Ft , we deduce that Ct = c(t, St ) where

" + #
√

Q −rτ 1 2
c(t, x) := E e x exp −σ τ Y + r − σ τ − K
2
We shall compute
Z ∞ +
√

1 −rτ 1 2 2
√ e x exp −σ τ y + r − σ τ − K e−y /2 dy.
2π −∞ 2
The term +
√

1 2
x exp −σ τ y + r − σ τ − K
2
is non-zero if, and only if,

1 x 1 2
y < d− (τ, x) := √ ln + r− σ τ .
σ τ K 2
We then have
Z d− (τ,x) n √ 2 o +
1 −rτ −σ τ y+ r− σ2 τ 2
c(t, x) = √ e xe −K e−y /2 dy
2π −∞
Z d− (τ,x) y2 √ 2

Z d− (τ,x)
1 − 2 −σ τ y− σ 2 τ 1 2
= √ xe dy − √ e−rτ Ke−y /2 dy
2π −∞ 2π −∞
Z d− (τ,x) √
x
e{− 2 (y+σ τ ) } dy − e−rτ KN (d− (τ, x))
1 2
= √
2π −∞
Z d− (τ,x)+σ√τ
x 2
= √ e−z /2 dz − e−rτ KN (d− (τ, x)).
2π −∞
Finally, we proved that
c(t, x) = xN (d+ (τ, x)) − e−rτ KN (d− (τ, x))
where
1 x 1 2
d− (τ, x) = √ ln + r− σ τ
σ τ K 2
and
√

1 x 1 2
d+ (τ, x) = d− (τ, x) + σ τ = √ ln + r+ σ τ .
σ τ K 2
Partial Differential Equation. Let θ = (θt )t∈[0,T ] be a self-financing trading strategy

that replicates the call using the riskless asset and the stock. If we let θt =: (Γt , ∆t , 0),
then we have
c(t, St ) = Ct = Vt (θ) = Γt Dt + ∆t St
Since θ is self-financing, we have
dCt = Γt dDt + ∆t dSt

= Γt rDt dt + ∆t dSt
= r(Ct − ∆t St )dt + ∆t dSt .
Applying Itô’s Lemma, we have

1
dCt = ct (t, St )dt + cx (t, St )dSt + cxx (t, St )dSt dSt
2
1 2 2
= ct (t, St ) + rcx (t, St )St + cxx σ St + cx (t, St )σSt dW
ft
2
We then deduce that
∆t = cx (t, St )
and, for every t ∈ [0, T ] and every x ∈ R,
1
ct + rcx x + cxx σ 2 x2 = rc(t, x).
2
Moreover, the function c satisfies the following border condition
c(T, x) = [x − K]+ .
3.9 Martingale Representation Theorem

Let (Ω, F, P) be a probability space on which is defined a Brownian motion (Wt )t⩾0 . Fix a
filtration (Ft )t⩾0 for the Brownian motion. Fix an arbitrary process x = (xt )t⩾0 adapted
to the filtration (Ft )t⩾0 . Fix some arbitrary value I0 . For every t ⩾ 0, denote by It (x) the
Itô’s integral Z t
It (x) := I0 + xu dWu .
0
The process (It (x))t⩾0 is a P-martingale. This is an important property satisfied by the
Itô’s integral. Actually, we have the converse property: any process that is a P-martingale
can be written as an Itô integral. Formally, we have the following result.
Theorem 3.8 (Martingale Representation Theorem). Let (Ω, F, P) be a probability space
on which is defined a Brownian motion (Wt )t⩾0 . Fix a filtration (Ft )t⩾0 for the Brownian
motion. If (Mt )t⩾0 is a P-martingale, then there exists an adapted process (xt )t⩾0 such
that Z t
Mt = M 0 + xu dWu .
0
3.10 Complete Markets

Let (Ω, F, P) be a probability space on which is defined a Brownian motion (Wt )t⩾0 . Fix
a filtration (Ft )t⩾0 for the Brownian motion. Consider a financial market with two assets:
a riskless asset and a stock. The prices of the riskless asset and the stock satisfy the
following stochastic differential equations
dDt = rDt dt and dSt = αSt dt + σSt dWt
for some parameters r, α and σ ̸= 0 where we assume that D0 = 1.

3.10. COMPLETE MARKETS 71
where µ := (α − r)/σ. Applying Girsanov Theorem, the process (W

ft )t⩾0 defined by
W
ft := Wt + µt
is a Brownian motion with respect to some probability measure Q. Moreover, we have

dSbt = σ Sbt dW
ft .
Fix some arbitrary maturity date T and an arbitrary random variable Y that is FT -
measurable. We look for a self-financing portfolio θ = (Γt , ∆t )t⩾0 that replicates Y , in the
sense that VT (θ) = Y .
Necessary Condition. We start by assuming that Y can be replicated. Let θ =

(Γt , ∆t )t⩾0 be a self-financing trading strategy such that VT (θ) = Y . We abuse notations
and write Vt in place of Vt (θ). Since θ is self-financing, we have
dVt = Γt dDt + ∆t dSt = rΓt Dt dt + ∆t dSt .
Since Vt = Γt Dt + ∆t St , we deduce that
dVt = r(Vt − ∆t St )dt + ∆t dSt .
Applying Itô’s Lemma to f (t, St ) where (t, x) 7→ f (t, x) := e−rt x, we get
dSbt = −re−rt St dt + e−rt dSt = (α − r)Sbt dt + σ Sbt dWt = σ Sbt (µdt + dWt )
where µ := (α − r)/σ. It follows from Girsanov’s Theorem that there exists a probability
measure Q (equivalent to P) such that (W ft )t⩾0 is a Q-Brownian motion where Wft :=
Wt + µt. We then get
dSbt = σ Sbt dW
ft .
Applying Itô’s Lemma to f (t, Vt ) where (t, x) 7→ f (t, x) := e−rt x, we get
dVbt = −re−rt Vt dt + e−rt dVt
= −r∆t Sbt dt + ∆t e−rt dSt
h i
= ∆t (α − r)Sbt + σ Sbt dWt
= σ∆t Sbt dW
ft .
This means that Z t

Vbt = V0 + σ∆u Sbu dW
fu .
0
In particular, the process (Vbt )t⩾0 is a Q-martingale. Since VbT = e−rT Y , we deduce that
Vbt = EQ [e−rT Y |Ft ] and dVbt = σ∆t Sbt dW
ft .
We can derive two important properties from the above equations. First, we can compute
Vbt even if we do not know the portfolio θ simply by computing Mt := EQ [e−rT Y |Ft ].
Second, we can infer ∆t if we are able to write dMt as a function of dW
ft . Moreover, by
definition of Vt , we have
1
Mt = Vbt = Vt = Γt + ∆t Sbt
Dt
and we can infer Γt .
Sufficient Condition. Let M = (Mt )t⩾0 be the stochastic process defined by
Mt := EQ e−rT Y |Ft .

The process (Mt )t⩾0 is a Q-martingale by construction (apply the Tower Property of the
conditional expectation). Moreover, the filtration (Ft )t⩾0 is also a filtration for the Q-
Brownian motion (W ft )t⩾0 .7 We can apply the Martingale Representation Theorem to
deduce that there exists an adapted process (xt )t⩾0 such that
Z t
Mt = M 0 + xu dW
fu .
0
Equivalently, we have dMt = xt dW

ft . Give our analysis of the necessary condition, we
propose to let
xt
∆t := and Γt := Mt − ∆t Sbt .
σ Sbt
We can now define the candidate for the replicating portfolio θ = (θt )t⩾0 by posing
θt := (Γt , ∆t ).
We should prove that θ is self-financing and satisfies VT (θ) = Y . We abuse notations and
write Vt in place of Vt (θ). Since
Vt = Γt Dt + ∆t St ,
we have
Vbt = Γt + ∆t Sbt = Mt .
This implies that
e−rT VT = VbT = MT = EQ e−rT Y |FT = e−rT Y

and we deduce that VT = Y .

Let’s prove now that θ is self-financing. By construction of ∆t , we have
ft = xt dSbt = ∆t dSbt .
dVbt = dMt = xt dW (3.10.1)
σ Sbt
Since Vbt = e−rt Vt , we can apply Itô’s Lemma to show that
dVbt = −re−rt Vt dt + e−rt dVt . (3.10.2)
Combining (3.10.1) and (3.10.2), we have
e−rt dVt = ∆t dSbt + rVbt dt.
Since
dSbt = −re−rt St dt + e−rt dSt
7 fs − W
Since W
ft = Wt + µt, the process (W
ft )t⩾0 is adapted. For any 0 ⩽ t ⩽ s, the increment W ft =
Ws − Wt + µ(t − s) is also independent of Ft .
3.10. COMPLETE MARKETS 73
we deduce that
∆t (−re−rt St )dt + e−rt ∆t dSt + re−rt Vt dt = e−rt dVt .
Removing the term e−rt , we get that
dVt = r(Vt − ∆t St )dt + ∆t dSt .
By construction of Vt , we have
r(Vt − ∆t St ) = rΓt Dt .
This is sufficient to conclude that
dVt = Γt dDt + ∆t dSt .
This is exactly the definition of (Γt , ∆t )t⩾0 is self-financing.

Bibliography
Dana, R.-A. and Jeanblanc, M.: 2007, Financial Markets in Continuous Time, Springer-
Verlag, Berlin Heidelberg.
Duffie, D.: 2001, Dynamic Asset Pricing Theory, Princeton University Press.
Duffie, D.: 2003, Intertemporal Asset Pricing Theory, Vol. 1, Elsevier North-Holland,
Amsterdam, chapter 11, pp. 639–742. Available here.
Elliott, R. J. and Kopp, P. E.: 2005, Mathematics of Financial Markets, Springer-Verlag,

New York.
Föllmer, H. and Schied, A.: 2016, Stochastic finance, de Gruyter.
Granger, T. and Etner, F.: 2011, Economie financière, Economica.
Hull, J. C.: 2015, Options, Futures, and Other Derivatives, 9th edn, Pearson Education.
LeRoy, S. F. and Werner, J.: 2014, Principles of Financial Economics, 2 edn, Cambridge
University Press, Cambridge.
Pliska, S.: 1997, Introduction to mathematical finance, Blackwell, Oxford.
Project Jupyter: 2022. https://jupyter.org/index.html.
QuantEcon: 2022. https://quantecon.org/.
Shreve, S. F.: 2004a, Stochastic Calculus for Finance I, Springer-Verlag, New York.
Shreve, S. F.: 2004b, Stochastic Calculus for Finance II, Springer-Verlag, New York.
Williams, D.: 2013, Probability with Martingales, Cambridge University Press, New York.
75

Lecture Notes

Uploaded by

Lecture Notes

Uploaded by

Introduction to Stochastic Calculus for Economists

1 Basic Probability Theory 1

2 Discrete Time Dynamic Trading 11

3.6.4 Mean-Reverting Process . . . . . . . . . . . . . . . . . . . . . . . . . 56

Basic Probability Theory

1.1 Time and Uncertainty

The likelihood of each outcome is described by a probability p on Z. Recall that p is a

Figure 1.1: Tossing a coin at every period

Figure 1.2: Two possible states of nature ω and ω̂

Figure 1.3: Date-2 event

The mapping ω 7→ ω T from Z N to Z T is denoted by projT . It represents the available

ΩT (ω̄ T ) := {ω ∈ Ω : projT (ω) = ω̄ T } = {ω ∈ Ω : ωt = ω̄t , ∀t ∈ {1, . . . , T }}.

P(Ωt (ω̄ t )) := p(ω̄1 )p(ω̄2 ) . . . p(ω̄t ).

P ΩT (ω̄ T ) ∪ Ωτ (ω̂ τ ) = p(ω̄1 )p(ω̄2 ) . . . p(ω̄T ) + p(ω̂1 )p(ω̂2 ) . . . p(ω̂τ ).

2. For any date t and partial history ω t , the set Ωt (ω t ) is an event;

3. If A and B are two events, then A ∩ B and A ∪ B are also events;

4. If A is an event, then Ac := Ω \ A is also an event;

1.3 Random Variable and Stochastic Process

A random variable is a function X : Ω −→ R such that {X ⩽ x} is an event (i.e.,

Figure 1.4: The set of date-2 events forms a partition of Ω

information ω t observed up to date t, in the sense that X(ω) = X(ω

1.5 Expectation and Conditional Expectation

X(Ω) := {x ∈ R : ∃ω ∈ Ω, x = X(ω)} = {X(ω) : ω ∈ Ω}

for any Ft -measurable random variable Y : Ω → R. A random variable Xt satisfying the

The conditional expectation operator satisfies the following important properties:

(i) if X is already Ft -measurable, then EP [X|Ft ] = X;

(ii) positivity: if X ⩾ 0, then EP [X|Ft ] ⩾ 0;

(iv) tower property: if s < t, then EP [EP [X|Ft ]|Fs ] = EP [X|Fs ];

(v) taking out what is known: if Z is Ft -measurable, then EP [ZX|Ft ] = Z EP [X|Ft ];

(vi) Jensen: if u : R → R is concave, then EP [u(X)|Ft ]⩽u(EP [X|Ft ]);

(vii) role of independence: if X is independent of Ft , then EP [X|Ft ] = EP [X].

Consider a sub σ-algebra G ⊆ F. Intuitively, the σ-algebra G contains less information

Mti+1 − Mti = Xti +1 + . . . + Xti+1 .

Discrete Time Dynamic Trading

2. Portfolios: As investors, assembling portfolios is critical to our decision-making

3. Self-financing Trading Strategies: A vital concept in dynamic trading is the

4. Hedging: Managing risk is a paramount concern for investors. We will investi-

5. Arbitrage Opportunities: Arbitrage opportunities are unique instances where

7. Equivalent Martingale Measures: We will introduce equivalent martingale mea-

By mastering these concepts, you will develop a solid understanding of discrete-time

Vt (ω) = θt (ω) · [St (ω) + δt (ω)].

Still, at date t, he can choose another portfolio

V0 (θ) := θ1 · S0 and Vt (θ)(ω) := θt (ω) · St (ω), ∀t ⩾ 1.

At any intermediate date t ∈ {1, 2, . . . , T − 1}, the flow budget constraint is

ct (ω) + θt+1 (ω) · St (ω) ⩽ θt (ω) · St (ω) + et (ω).

Finally, at the final period T , we have

cT (ω) ⩽ θT (ω) · ST (ω) + eT (ω).

1. the function U is strictly increasing as it is the case when

for some strictly increasing Bernoulli function u : R+ → R and time-preference factor

U (c) = EP [u(cT )].

The second case is usually associated to a situation where et = 0 for every t ∈

• at the initial period

• at any intermediate date t ∈ {1, 2, . . . , T − 1}

ct (ω) = et (ω) + θt (ω) · St (ω) − θt+1 (ω) · St (ω);

• and at the final period T ,

cT (ω) = θT (ω) · ST (ω) + eT (ω).

2.3 Self-Financing Trading Strategies

• at the initial period

• at any intermediate date t ∈ {1, 2, . . . , T − 1}

θt+1 (ω) · St (ω) = θt (ω) · St (ω);

• and at the final period T , we have

cT (ω) = θT (ω) · ST (ω).

θt+1 · St = θt · St = Vt (θ), for all t ∈ {1, 2, . . . , T − 1}.

to represent the variation of X between t − 1 and t. If a trading strategy θ is self-financing

∆Vt (θ) := θt · St − θt−1 · St−1 = θt · St − θt · St−1 = θt · ∆St .

Vt (θ) = V0 + θ1 ∆S1 + θ2 · ∆S2 + . . . + θt · ∆St .

∆Vt (θ) = Vt (θ) − Vt−1 (θ)

Therefore, a trading strategy θ is self-financing if, and only if,

Such a strategy is called a hedging or replicating strategy.

θT0 (ω T −1 )ST0 (ω T −1 , H) + ∆T (ω T −1 )ST (ω T −1 , H) = X(ω T −1 , H)

Fix a predecessor event ω T −1 . Since we look for a self-financing trading strategy, we

θT0 −1 (ω T −2 )ST0 −1 (ω T −2 , H) + ∆T −1 (ω T −2 )ST −1 (ω T −2 , H) =

θT0 −1 (ω T −2 )ST0 −1 (ω T −2 , L) + ∆T −1 (ω T −2 )ST −1 (ω T −2 , L) =

2.5 Arbitrage Opportunities

at every intermediate date t ∈ {1, 2, . . . , T − 1},

ct (ω) + θt+1 (ω) · St (ω) ⩽ θt (ω) · St (ω) + et (ω)