Probabilistic Methods in Finance 1

PROBABILISTIC METHODS IN FINANCE

Q.E.M., M.M.E.F., M.A.E.F., I.M.-M.A.E.F.

Isabelle Nagot
Université PARIS I Panthéon-Sorbonne

Bibliography
Finance, but mathematical tools not sufficient
Hull, J. Options, futures, and other derivative securities, Prentice-Hall, 10th edition, 2018.
Goffin, R. Principes de finance moderne. Economica, 6e edition, 2012.
More mathematics
Baxter, M. and Rennie, A. Financial calculus. Cambridge University Press, 1996.
Kwok, Y.K. Mathematical models of financial derivatives, Springer, 2nd edition, 2008 (3 first chapters).
Jacod, J., Protter, P. (2000) Probability Essentials. Springer.

Chapter I. Introduction. Preliminaries

Aim = Option pricing.

Model in discrete time (Cox-Ross-Rubinstein model), or continuous time (Black-Scholes model).
Pricing by arbitrage.
Very simple market:
one risk-free asset (ie deterministic return, one value only at the end of any period:
discrete time: (1 + r)n after n periods, continuous time: ert at time t),
one risky asset = underlying asset of the option. St denotes its price at time t.

Options belong to derivative products, we will be quick on description and use (forwards/futures,
options), can be discussed in the tutorial.
A derivative is determined by its payoff, i.e. what it pays at time T (expiration or maturity date),
generally f (ST ).
The market will be complete, i.e. any asset is replicable
(example in discrete time: 2 states of the world at the end of each period).
Maths tools: martingales, for continuous time: Brownian motion.

1 Derivative products: Description and use

A derivative security is a security whose cash flows (hence value) depend on the values of other
more basic underlying variables, which may be the prices of traded securities, prices of commodities,
stock indices, exchange rates, or any observable variable (temperature,...).
Derivative securities are also known as contingent claims.
 Probabilistic Methods in Finance 2

1.1 Forward contracts

A forward contract is an agreement to buy or sell an asset (called the underlying asset, ”U.A.”)
at a certain future time T (called the maturity) for a certain price (called the delivery price).
Ex1: a wheat producer can specify in advance the price at which he will sell a given part of its
production. This cancels his price risk (low price when the harvest is ready).
Ex2: A company using oil for its production can fix the price it will pay for it.

One of the parties to the contract assumes a long position and agrees to buy the underlying asset
on T for the specified price.
The other party assumes a short position and agrees to sell the asset on T for the same price.
The contract must specify: the quantity of the asset to be delivered, the quality, place, way of
delivery.

A very wide range of commodities and financial assets form the underlying assets.
At the beginning, it was for seasonal products, products that are stored to satisfy the demand:
agricultural products...
Then any storable product (coffee, sugar, wood, live cattle...)
Now there exists contracts on all sort of products, even non storable like electricity.
Financial products, interest rates, stock indices (⇒ cash delivery), currencies, bonds...
commodities: gold, silver, copper, aluminum...
Ex: FRA (Forward Rate Agreement) = Forward on an interest rate, eg between a bank and its
client. It allows to fix now, the rate for a loan that the client will take in the future.
Contracts have been created on anything whose value fluctuates, to provide an hedge.
Eg: weather derivatives, on temperature, quantity of snow, rain (example: for a beer producer),...
A longevity swap allows a pension scheme to remove the risk that members live longer than ex-
pected.

Payoff: Let ST denote the price of the underlying asset on the date of expiration T and K the
delivery price. The terminal payoff for the long position is ST − K.

If the contract is exchanged on a market:

At the time the contract is entered into, the delivery price is chosen so that the value of the forward
contract to both parties is zero: it costs nothing to take either a long or a short position.
Later the contract can have a positive or negative value depending on movements in the price of the
underlying asset: if this price rises after the initiation of the contract, the value of a long position
becomes positive (because it gives you the opportunity to buy - forward - at a better price than
the current market forward price).

1.2 Future contract

Forward contracts are OTC = private contract between 2 counterparties, difficult to sell back +
counterparty risk (default for delivery or for payment).
Future contracts are the same products, but exchanged on a market.
⇒ standardization (to get liquidity): in amount, in term (ex: Mar, Jun, Sep, Dec), in quality.
 Probabilistic Methods in Finance 3

1st future contract = on the CBOT for the wheat.

One of the key role of the exchange is to organize trading so that contract defaults are minimized.
When an investor enters a contract through a broker, the broker will require the investor to deposit
funds in what is termed a margin account.
1. initial margin computed to cover the biggest loss possible for one day (depends on the underlying
asset)
2. at the end of each trading day, the margin account is adjusted to reflect the investor’s gain or
loss.
Equivalent to terminate the contract every evening.
F (t, T ) forward price at date t for the maturity T .
You enter a contract at time 0. At time 1, the new forward price is F (1, T ), if for example
F (1, T ) < F (0, T ), you could enter the same contract with a lower price F (1, T )
⇒ potential loss = F (0, T ) − F (1, T ) = margin call.
If negative, money is put back in your account.
At maturity, you have paid F (0, T ) − F (T, T ) = F (0, T ) − ST which is what you ”owe” (can be
negative) at time T , since you have to pay F (0, T ) to receive the underlying asset (in case of cash
delivery, after the last margin call, you owe nothing).
Note that interest is paid on the margins.
The Clearinghouse acts as an intermediary between buyers and sellers. Members = brokers. It
determines the close price depending on all the positions.
+ it organizes the exchange (daily price movement limits on the exchanged volumes..)

The future price reflects the investors expectations about future supply and demand for the prod-
uct. It gives some insight for the decisions of production, storage...

Note that most contracts are terminated before maturity. You can buy futures on the live cat-
tle and terminate them before receiving the animals...
Sometimes cash settlement anyway.

1.3 Options
The underlying assets include those of futures contract, and derivatives themselves.
A call option gives the holder the right to buy the underlying asset by a certain date (called the
maturity) for a certain price K (exercise price or strike price).
A put option gives the holder the right to sell the underlying asset by a certain date for a certain
price.
American options can be exercised at any time up to T (the maturity is called the expiration date).
European options can only be exercised on the maturity itself.
(these terms do not refer to the location of the option!)

Payoff: We consider a European call with maturity T and exercise price K.

Let ST denote the price of the underlying asset at time T .
The terminal payoff from the long position in this call is max(ST − K, 0) = (ST − K)+ (where, for
any x ∈ R, x+ denotes max(x, 0)). Indeed, at T , if ST ≥ K, the call is exercised, and its owner
 Probabilistic Methods in Finance 4

receives, in value: ST − K, else the call is not exercised and the payoff is 0.
For the corresponding put, the payoff is (K − ST )+ .

Note that the cash-flows for a call buyer are:

- premium at the time the call is bought, (ST − K)+ at time T .

An option gives the holder the right to do something. The holder does not have to exercise this
right. Whereas it costs nothing to enter into a forward contract, an investor must pay to purchase
an option contract. But he can choose the strike price of the contract.

Example (exercise 1): Consider a US company due to receive M euros at a known future time
T (because it exports in Europe). It costs are in dollars and it will have to change euros against
dollars.
If this exchange takes place at time T , the company doesn’t know how much dollars it will get
because the future rate (ST = number of dollars for 1=C at T ) is not known today.
The company is therefore exposed to a foreign exchange risk.

It has two solutions to hedge this risk:

1. A forward contract locks in the exchange rate for a future transaction. The company will assume
a short position: agrees to sell the asset (each euro) at time T for the given price.
(currency future on EUR: contract size = 125, 000=C).
2. Foreign currency options are an interesting alternative. The company can hedge its risk by
buying M put options on euro which mature at time T .
At time T , the company receive M euros and (K − ST )+ for each put, where ST is the value of one
euro at time T .
The resulting cash-flow is
(
M K if ST ≤ K (the puts are exercised)
M [ST + (K − ST )+ ] = = max(K, ST ) $
M ST if ST ≥ K (the puts are not exercised)

This guarantees that the value of the euro will not be less than the exercise price K, while allowing
the company to benefit from any favorable exchange-rate movements.

Whereas a forward contract locks in the exchange rate for a future transaction, an option pro-
vides a type of insurance (against losses on the exchange).
Of course, insurance is not free. It costs nothing to enter into a forward transaction, while options
require a premium to be paid up front.

Note the difference: option pricing is about finding the initial value of the option contract (for
a given exercise price, that is chosen in the contract), while future pricing means computing the
fair delivery price that ensures an initial value of 0 for the future contract.
 Probabilistic Methods in Finance 5

2 Rates and discounting

An interest rate is the cost of borrowing or the price paid for the rental of funds (usually expressed
as a percentage of the rental of $100 per year, ie annualized). The rate of return (or return) is the
reward that investors demand for accepting delayed payment.
1 euro today is worth more than 1 euro tomorrow because the euro today can be used/ invested
today.
It should be better to have it today than in the future, as you can just keep it, so an interest rates
should be positive.

But not in particular cases (and currently very frequent!):

Ex: in June 2014, the European Central Bank (ECB) cut its deposit rate to -0.1%, to encourage
eurozone banks to lend to small firms rather than to accumulate cash (the hope is to boost the
economy). Instead of earning interest on money left with the ECB, banks are charged by the central
bank to park their cash with it.
Such policies act on the short end of the yield curve.
Some regulations (e.g. for insurance companies) act on the remaining of the curve: companies are
not allowed to keep large amounts of cash, then it has to be invested in some secure assets.
Ex: in August 2019, German government rates up to maturity 30 years were all negative.

A bond is a debt security that promises to pay a certain stream of payments (coupons and principal)
in the future, until a final date called maturity: the basic bond pays its owner a fixed-interest pay-
ment (coupon) every period (ex: every year) until the maturity date, when a specified final amount
(face value or par value) is repaid.

Note: an interest rate depends on the maturity, for derivatives pricing we will assume the rates to
be constant most of the time.

2.1 Future and present value

1. Future value → capitalization → compounding interest rates
r interest rate, supposed to be constant.

Definition: the future value in n years of a capital M is the value that one obtains when investing
this capital on n years.

Ex: once per year.

Consider an amount M invested for t years at an interest rate of rd per year (discrete rate for one
year). After n years, you get: M (1 + rd )n .

We consider the general case where an interest is paid several times per year and compounded: the
already paid interests are themselves invested. We have m periods in 1 year, the interest is com-
pounded m times per year. The length of one period is 1/m (eg, m = 12: compounded monthly,
m = 24 for the saving account ”livret A” in France, i.e. interest compounded half-monthly).

We want to compute what should be the interest rate on each period, knowing that after 1 year
the total interest has to correspond to the annual rate rd , as announced by the bank.
Ex: the bank promises a rate of 10% for money invested for 1 year. But an interest is paid each
 Probabilistic Methods in Finance 6

month, so you wonder what the rate for a period of one month is. The answer is: 0,797%. But this
number cannot be easily compared to other rates. The habbit is to annualise any rate to facilitate
comparison. For that you multiply by 12 to know what the rate would be over a one year period
(without any compounding). You get 9,569% (values checked in the tutorial, exercise 2.1).

General case: let rm be the annualised rate for 1 period, i.e. the interest on M invested over 1
period is rm M × period length = rm M m , meaning that over one period, the wealth gets multiplied
rm
by 1 + m .
On one year, the interest rate (obtained by compounded the intermediate interests) is rd .
 rm m
Then rm is obtained by 1+ = 1 + rd , where rd is the announced annual rate for one
m
year.

Exercise 2.2: prove that you get too much if you use rd instead of rm in above formula (because
of the compounding).

An amount M being invested at time 0, let M (t) be the value of the investment at time t.

For any k ∈ N, the future value of M after k sub-periods (hence k interest payments), is:
k k k
k rm k
M(m ) = M (1 + m) = M (1 + rd ) m = M e m ln(1+rd ) = M er m where r = ln(1 + rd ).
k
So M (t) = M (1 + rd )t = M ert holds not only for t integer, but also for any t = m, k ∈ N,
k k+1
while M (t) is constant on [ m ; m [.

We observe that after each interest payment, the value of the investment belongs to the curve t 7→ ert ,
where t is the time (in years), curve that is independent of m.

Most mathematical models for option pricing (ex Black-Scholes) are in continuous time.
Hence a continuous interest rate is used: the interests are assumed to be continuously paid and
compounded, instead of discretely. That modeling can be seen as the limit when m goes to infinity
in the above discrete compounding (first approach).
When the number of periods m gets large, the future value dynamics is close to t 7→ M ert , with
r = ln(1 + rd ), which is the future value dynamics for a continuous interest rate. We have also:
Th: When m → +∞, rm converges toward r = ln(1 + rd ).
r
Proof: we have (1+ rmm )m = er , hence rm = m(e m −1) ∼ m(1+ m
r
−1) = r, ie rm → r = ln(1+rd ).
(we used ∀x > 0, ∀y ∈ R, xy = ey ln x and ex ∼ 1 + x when x → 0).

Second approach for a continuous rate: (modeling directly in continuous time)

M (t) wealth at t.
Continuous interest rate: at any period of time (even very short) an interest is paid.
In the continuous compound model, it is assumed that the instantaneous rate of change of M is
dM (t)
proportional to M : ∀t ≥ 0, = rM (t) (*)
dt
Another writing is: ”dM (t) = rM (t)dt”. This is a notation for (*), that can be interpreted as
follows: for an infinitely small change in t, dt, the corresponding change in the wealth at time t is
dM (t) = M (t + dt) − M (t) = rM (t)dt. This variation corresponds to the interest paid between
 Probabilistic Methods in Finance 7

t and t + dt. This means in fact that for h infinitely small, M (t + h) − M (t) ∼ rM (t)h, hence
M (t + h) − M (t)
lim = rM (t) i.e. (*).
h→0 h
1
Note that this is consistent with what we have in discrete time: on one period of length m , the
1
change in the wealth corresponds to the paid interest: ∆M = rm M m , therefore the rate of change
in the wealth on one period is: ∆M
1/m = rm M , proportional to the wealth.

From (*) we get: ∀t ≥ 0, M 0 (t) = rM (t), then the relative rate of change of the function M is
constant:
0 (t)
∀t ≥ 0, M 0
M (t) = r, which implies: ∀t ≥ 0, [ln M ] (t) = r.
Integrating between 0 and t, we obtain ln M (t) − ln M (0) = rt thus M (t) = M (0)ert .

(
(1 + r)t if r = discrete interest rate
Conclusion: 1 euro today is worth at t :
ert if r = continuous interest rate

2. Present value → Pricing of a future cash flow

reverse of Future Value.

An asset is hold for the future cash flows it will give. We need to calculate the present value of a
future cash flow.

a. Deterministic cash flow (ie known at the time of pricing)

= ”risk-free asset”: can have only one value in the
 future
1
 if r = discrete interest rate
present value (today, at time 0) of 1 euro at t = (1 + r)t
 −rt
e if r = continuous interest rate
A zero-coupon bond (for example in dollar) with maturity date T is a contract which guarantees
the holder 1$ to be paid at time T .

Its price = discount factor.

Price of a 0-coupon bond with maturity T (a bond that pays no coupons and pays 1 euro at T )
= discount factor at t for time T , denoted by B(t, T ).
→ Discounting (= calculating what dollars received in the future are worth today)

allows to: bring back a future flow at date 0, compare flows at different times.

Obviously, the rate used to discount a cash-flow paid at time t is the rate that prevails, at the
pricing time 0, for the maturity t. We will denote it by r(0, t).

b. General case: stochastic cash flow

Imagine now that the flow is not known at the date of the pricing: its value at future date of
payment is random.
E(C)
Present value of a stochastic cash flow C: should we take: ?
(1 + r)t

Imagine you can choose between two flows to be received in one year:
- first cash flow is worth 500 in any case,
 Probabilistic Methods in Finance 8

(
1000 with proba 1/2
- second cash flow (lottery) is worth
0 with proba 1/2
Both flows have the same expectation but you certainly prefer the first one.
That means that you are ”risk averse”.
If you have to pay today to receive these flows in 1 year, you will be ready to pay more for the first
500
one (which is worth 1+r ).
An interpretation is that your preferences (the values you give for a future cash flow) can be repre-
sented by a utility function U which is concave (U (x) is the value attributed to the certain flow x):

To win 1 million is much better than 0 but to win 2 millions is not as much better than 1 million.
A concave utility function means risk aversion: you prefer to receive E(C) for sure than to receive
C. Indeed, since U is concave, the inequality of Jensen states: E(U (C)) ≤ U (E(C)), i.e.: the
expected utility of a flow is less than the utility of its expected value.
E(C)
is the PV of the cash flow E(C) at t (deterministic cash flow), then the PV today of the
(1 + r)t
E(C)
cash flow C at t is less than .
(1 + r)t

Generally the utility function of the investor is not estimated, only summarized by a discount
rate taking it into account with the investment risk as well: the Present Value of a stochastic cash
flow C is then expressed as:

 E(C)

if µ = discrete interest rate
(1 + µ)t
 E(C)e−µt if µ = continuous interest rate


We have most often µ > r and µ − r increases with the risk on the asset.
The method of Discounted Cash Flows (DCF) is the classical way to evaluate companies/projects
in corporate finance.

For quoted assets, the prices settle at levels that reflect an average of the preferences of the in-
vestors. The implied utility function is the one of the market (notion of ”representative investor”).
When the market risk aversion increases, an equity price generally goes down even if nothing has
changed for the company itself.

µ depends on the risk of the cash flow. Eg: stock.

The riskier the stock is (ie the more dispersed its future values are), the larger µ:
the investors require compensation for the risk they assume by buying this stock.

More precisely, the value of µ depends on the systematic risk of the investment.
See the Capital Asset Pricing Model (course on Portfolio Choice):
An investor should not require a higher expected return for bearing nonsystematic risk, as this
risk can be almost completely eliminated by holding a well-diversified portfolio. But he generally
requires a higher expected return than the risk-free interest rate for bearing positive amounts of
systematic risk.
 Probabilistic Methods in Finance 9

3 Arbitrage methods
We make the following assumptions on the financial markets for derivatives pricing (”usual assump-
tions”):

1. There are no transactions costs.

2. All trading profits are subject to the same tax rate (or no taxes).

3. The market participants can borrow and lend money at the same risk-free rate of interest. This
interest rate is denoted by r. A government bond is considered as a risk-free investment. Its return
equals therefore r. We call it a risk-free asset.

4. Short selling is allowed. This is a trading strategy that yields a profit when the price of a security
goes down. It involves selling securities that are not owned and buying them back later (reverse of
buy/sell).

5. Assets are divisible: one can hold α assets, with α ∈ R.

6. NAO: there are no arbitrage opportunities

An arbitrage opportunity is a situation where you can make a profit without any risk, by choosing
a given portfolio (ie a combination of assets) that requires no initial investment, and yields an
amount in the future that is non-negative under all possible circumstances (”states of the world”),
and not identically zero.

Arbitrage is defined as a strategy that allows to make a profit out of nothing without taking any
risk.
NAO: no sure gain at T can be made when investing 0 at time 0.

Mathematical formulation:
The uncertainty in the market on a period T = [0, Tf ] or {0, 1, ..., Tf } can be described as ’ran-
domness’ interpreted in the context of some probability space (Ω, F, P ). The unknown future is
just one of many possible outcomes, called states of the world (or economy).
One elementary event ω = one possible state of the world
= { realised prices between 0 and Tf , for any economical data }.
Ω = set of elementary events ω.
P a probability measure defined on a σ-algebra (or σ-field) F of Ω.
Seen from time 0, the prices at any future time t ∈ T are random variables.
(for some reminders in probability theory, see for example Jacod, J., Protter, P. (2000) Probability
Essentials. Springer).

Definition: An arbitrage opportunity (at time 0) is a portfolio, whose value at time t is Vt , with
(
P (VT ≥ 0) = 1
V0 = 0 and ∃T > 0 such that
P (VT > 0) > 0
NAO assumes that this does not exist (”no free-lunch”): the idea is that there are enough investors
such that the prices reach an equilibrium immediately. If there exists an AO, some product is
undervalued, the demand for it will increase, its price also. The AO disappears.
 Probabilistic Methods in Finance 10

Since the market participants take advantage of arbitrage opportunities as they occur, arbitrage
opportunities disappear quite quickly: indeed an arbitrage opportunity means that an asset is not
expensive enough when taking into account the price of other assets. Then the investors buy this
asset (and sell the other ones), its price increases and the arbitrage opportunity disappears.

That is why we can make this assumption of NAO.

Note: short selling is done through a broker who borrows the security from another client and sell
it in the market, depositing the sale proceeds to the investor’s account. An investor with a short
position must pay to his broker any income, such as dividends or interest, that would normally be
received on the securities that have been shorted.

We are going to prove some relationships involving derivatives prices, by showing that if they
were not satisfied arbitrage opportunities would exist (arbitrage pricing).
These relationships come from links between prices of the derivatives and of their underlying asset
and between derivatives prices themselves (redundancy of markets). Ex: relationship between Eu-
ropean call and put (same underlying asset, strike price and maturity).

Consequence of the assumption of NAO

If two portfolios (a portfolio is a list of financial positions) have same value at a future time T
(equality of random variables), they must have same value at any earlier time, t (with no addi-
tion/substraction to the portfolio, eg dividends stay in the portfolio).

Proof by contradiction: If this were not true, an investor could make a riskless profit by short
selling the most expensive one and buying the cheapest portfolio.
Indeed, let Vti the value at t of portfolio i, for i = 1 or 2. We have VT1 = VT2 .
If for example Vt1 > Vt2 > 0:
at t we sell short one portfolio 1, we buy one portfolio 2.
We have Vt1 − Vt2 $ (can be invested at the risk-free rate).
at T , we sell the portfolio 2 for VT2 $, we use this cash to buy one portfolio 1 and reimburse the
short sale. Our gain is at least Vt1 − Vt2 $ (more if we invested this cash).

Some alternative definitions are given on exercice 5.

Some applications:

1. A risk-free portfolio can only have a return equal to r.

2. Call-put parity
Without any a priori model of the evolution of the underlying asset, we can derive an important
relationship between the prices of European call and put on a non-dividend paying stock.
We consider European call and put with a same strike price K and same maturity T on a non-
dividend paying stock. Let Ct be the call price at time t and Pt the put price.
(
1
(1+r)T −t
if r is a discrete rate
Value at t of the 0-coupon bond maturing at T : B(t, T ) =
e−r(T −t) if r is a continuous rate
 Probabilistic Methods in Finance 11

Consider the two following portfolios:

at t at T
Portfolio A: 1 call and K 0-coupon bond with maturity T (ST − K)+ + K
(or an amount of cash equal to Ke−r(T −t) )
= max(ST , K)
Portfolio B: 1 put and 1 stock (K − ST )+ + ST

(we always assume that cash is invested at the risk-free rate)

Both are worth max(ST , K) at expiration of the options
(since the options are European, they cannot be exercised prior to the expiration date).

The portfolios must therefore have the same value at time t, then:

Ct + KB(t, T ) = Pt + St eg Pt = Ct − St + Ke−r(T −t) if r is a continuous rate

This relationship is known as call-put parity. It shows that the value of a European put with a
certain exercise price and exercise date can be deduced from the value of a European call with the
same exercise price and date (and vice versa).
Of course, if this relationship does not hold, there are arbitrage opportunities.

3. Condition on the parameters in the binomial model (see beginning of chapter III.)

4. Forward prices computation (see chapter II.)

 Probabilistic Methods in Finance 12

Chapter II. Futures / forward contracts pricing

We use arbitrage arguments to provide a relationship between the forward price for the maturity
T , F (t, T ), and the spot price at time t, St .
We compute the fair delivery price for a forward contract that would be exchanged on a market
(or equivalently the fair delivery price for a future contract assuming that there are no margin calls).

1. Forward contracts on a security that provides no income and has no storage cost
Examples: non-dividend paying stock or 0-coupon.

For there to be no arbitrage opportunities, at time t, the relationship between the forward price
and the spot price must be, for a no-income security:
St
F (t, T ) = = St er(T −t) , where r = r(t, T ) is the continuous risk-free interest rate.
B(t, T )

Proof 1: consider the two following portfolios:

(
one long forward contract on the security
Portfolio A:
F (t, T ) 0 − coupons with maturity T (or an amount of cash equal to F (t, T )e−r(T −t) )

Portfolio B: one unit of the security.

In portfolio A, the cash will grow to an amount F (t, T ) at time T . It can be used to pay for the
security at the maturity of the forward contract. Both portfolios will therefore consist of one unit
of the security at time T . It follows that they must be equally valuable at the earlier time, t:
0 + F (t, T )B(t, T ) = St , since when a forward contract is initiated, the delivery price specified in
the contract is chosen so that the value of the contract is zero. Thus
St
F (t, T ) = = St er(T −t)
B(t, T )

Proof 2 (by contradiction):

St
? assume first that F (t, T ) > B(t,T ) .
An investor can borrow St dollars for a period of time T −t at the risk-free rate of r (or equivalently
St
sell short B(t,T ) 0-coupons with maturity T ), buy the asset, and take a short position in the forward
St
contract. At time T, the asset is sold under the terms of the forward contract for F (t, T ), and B(t,T )
St
is used to repay the loan. A profit of F (t, T ) − B(t,T ) in $ is therefore realized at time T .
St
? assume next that F (t, T ) < B(t,T ) .
An investor can take a long position in the forward contract and short the asset. The short position
leads to a cash inflow of St that can be invested in 0-coupons with maturity T . At time T , the
asset is purchased under the terms of the forward contract for F (t, T ) and used to close the short
St
position, and a profit of B(t,T ) − F (t, T ) in $ is realized.

Note: such contracts do not normally arise in practice: forward contract exist on securities that
provide an income to the holder, or that represent a cost to the holder (example: some commodities).
 Probabilistic Methods in Finance 13

Below examples will be seen in tutorial, to get used with the arbitrage arguments.

2. Forward contracts on a security that provides a known cash income

dividend assumed to be known at t (generally announced in advance, OK if T − t is small).

· Discrete income: ex: Di at ti .

We want to use the same argument as above. Portfolio A is not changed by the existence of divi-
dends on the U.A.. But portfolio B will contain the dividends, needs to be modified.
St = value at t of this portfolio at T and not of the portfolio containing 1 underlying asset

Define D as the present value, using the risk-free discount rate, of income to be received during
the life of the forward contract.
X
Ex: D= Di e−r(ti −t)
i

Portfolio B modified: one unit of the security plus borrowings of amount D at the risk-free rate
(discrete).
The income from the security is used to repay the borrowings so that the portfolio contains exactly
one unit of the security at time T . Then both portfolios A and B have same content at T . As
before, they need to have the same value at t to ensure that we have NAO.

The value of portfolio B at t is changed in St − D (instead of St ).

St − D
Using the NAO assumption as above, we get: F (t, T ) = .
B(t, T )

· Continuous income: (stock indices can be regarded as securities that provide known dividend
yields in continuous time)

We assume that the underlying asset pays a continuous dividend proportional to the U.A. value,
with an annualised rate δ: for any date s, between s and s + ds, 1 unit of the U.A. pays a dividend
equal to δSs ds.
Portfolio B modified: n(t) units of the underlying asset; between t and T , all income continuously
paid is immediately reinvested in the security. We choose n(t) such that portfolio B contains ex-
actly one unit of the security at time T .

Let n(s) be the number of securities held in portfolio B at time s ∈]t, T [.

between s and s + ds, each security held in portfolio B provides a dividend equal to δSs ds,which
is reinvested in δds securities. Then dn(s) = n(s)δds: the number of securities held in portfolio B
grows exponentially at rate δ: n(s) = n(t)eδ(s−t) . In order to get n(T ) = 1, we take n(t) = e−δ(T −t) .

Portfolios A and B are therefore worth the same at time T .

From equating their values at time t, we obtain, with r = r(t, T ) the continuous risk-free interest
rate:
0 + F (t, T )e−r(T −t) = St e−δ(T −t) . Thus

F (t, T ) = St e(r−δ)(T −t)

 Probabilistic Methods in Finance 14

3. Commodity forward contracts

For a commodity forward contract, holding the security represents a cost (storage cost: from having
to rent a warehouse, or from losses: part of the product may get spoiled during the storage).
Assuming that the storage cost on [t, T ] is known at t, the previous arbitrage argument holds, with
the storage cost considered as a negative dividend. The portfolio B in the above argument must
be changed in:
one unit of the security plus an amount of cash (or 0-coupon) equal to the present value
of all the storage costs that will be incurred during the life of the forward contract.
Discrete cost: the value of portfolio B at t is changed in St + U where U = i Ui e−r(ti −t) sum of
P
St + U
the discounted payments for storage. We get F (t, T ) = .
B(t, T )

Continuous cost: if the storage costs incurred at any time are proportional to the price of the
commodity (with the rate u), the forward price is F (t, T ) = St e(r+u)(T −t) .
Remark: the previous results hold for investment commodities, i.e. for commodities that are held
solely for investment (ex: gold). Some other commodities are held primarily for consumption (ex:
oil). For the 2nd type, there can be some incentive to keep the commodity in inventory to avoid
shortage and our previous arguments have to be modified (see below).

4. General result (in continuous time)

The relationship between forward prices and spot prices can be summarized in terms of what is
known as the cost of carry. This measures (the interest that is paid to finance the asset) plus (the
storage cost) minus (the income earned on the asset). Cost of carry: c = r + u − δ. Then

F (t, T ) = St ec(T −t)

Example: c = r − δ for a stock.

Conclusion: when the cost of carry between t and T is known at t, the forward price can be
precisely determined from the spot price. We used the fact that the forward contract can be repli-
cated by a position in its underlying asset that is ”carried” between t and T . The forward contract
is ”redundant”.
Notes:
- the probability distribution of ST is not involved (no need of a model for it).
- we made no difference between forwards and futures for the pricing, but future and forward
prices can differ, specially for long (in time) contracts. Indeed, taxes and transaction costs are miss-
ing in our calculations, liquidity can differ (futures more liquid, easier to trade...), interest rates are
random (it matters if margin calls are taken into account as a future contract is terminated every
day through the margin call, so rates at intermediary dates are involved).

In the general case, the cost of carry between t and T is not known at time t, the relationships be-
tween forward and spot prices are more complex to establish, the cost of carry has to be modeled.
Then forward prices as observed on the market give some information on cost of carry expectations
(example on dividends for an equity).
 Probabilistic Methods in Finance 15

An upward sloping forward curve is said to be ”in contango”: it is the situation where the price of a
security for future delivery is higher than the spot price, and a far future delivery price higher than
a nearer future delivery. The opposite market condition to contango is known as backwardation.
For example, a contango is normal for a (non-perishable) commodity which has a positive cost of
carry.

For a consumption asset, some investors can prefer to hold the asset (ex: oil).
Indeed, users of a consumption asset may obtain a benefit from physically holding the asset (as
inventory) prior to T (maturity) which is not obtained from holding the futures contract. These
benefits include the ability to profit from temporary shortages, and the ability to keep a production
process running.
Then arbitrage arguments can only be used to give an upper bound to the futures prices. But,
because of this preference for the spot position, the futures prices can stay lower than expected, as
if the UA was providing an income (cf formula F (t, T ) = St e(r−δ)(T −t) ).
Therefore futures prices (observed value of T → F (t, T )) show an implied return from holding of
the asset, called the ”convenience yield” . It measures the benefit from physically holding the asset.
We get (F (t, T ) = St e(r+u−y)(T −t) , with y=convenience yield).
The convenience yield reflects the market’s expectations concerning the future availability of the
commodity. The greater the possibility that shortages will occur during the life of the futures
contract, the higher the convenience yield. In particular, the convenience yield is inversely related
to inventory levels. Low inventories tend to lead to high convenience yields.

Example: currency forward

The underlying variable is, for example for a $-based investor, the current price in $ of one unit of
the foreign currency (for example the euro).
A foreign currency has the property that the holder of the currency can earn interest at the risk-free
interest rate prevailing in the foreign country (case of a security that provides an income). This
interest can be regarded as a dividend yield. We denote by rf the value of this foreign risk-free
interest rate with continuous compounding.
Consider the two following portfolios:
Portfolio A: one long forward contract on the security and F (t, T ) 0-coupons in $, with maturity T
Portfolio B: 1 0-coupon in the currency.

The value of portfolio B at time t is Bf (t, T ) in the currency (with Bf (t, T ) price at t of the
0-coupon in the foreign currency), or St Bf (t, T ) $ where St is the exchange rate at time t.
Both portfolios will become worth the same as one unit of the foreign currency at time T . Then
Bf (t, T )
F (t, T ) = St or
B$ (t, T )
F (t, T ) = St e(r−rf )(T −t)
This is called the interest rate parity relationship. It involves the interest rate differential (domestic
minus foreign r − rf ).

Exercise: find an arbitrage opportunity if this relationship doesn’t hold.

See Exercise 7. to explain the name interest rate parity relationship.
 Probabilistic Methods in Finance 16

Chapter III. Option pricing in discrete time: one-period binomial

model

Why a model is needed to price an option :

Roughly speaking, the volatility of a stock price is a measure of how uncertain we are about future
stock price movements.
As volatility increases, the chance that the stock will do very well or very poorly increases.
For the owner of a stock (spot or future), these two outcomes tend to offset each other. However,
this is not so for the owner of a call or put. The owner of a call benefits from price increases but
has limited downside risk in the event of price decreases since the most that he can lose is the price
of the option (equivalent result for a put).
Therefore, the values of both calls and puts increase as volatility increases.
Whereas the forward price F (t, T ) = St ec(T −t) is independent of the distribution of the possible
values for ST , an option price depends on this distribution. Therefore, we need to make a supple-
mentary assumption on the evolution of the stock value to price an option.
The objective of this chapter is to present the main ideas related to option theory within the very
simple framework of discrete-time models, with the example of Cox, Ross and Rubinstein’s model.

One period binomial model

2 dates 0 and 1. At t = 1, the risk-free asset is worth 1 + r for any state of the world. We assume
that the risky asset can take 2 values: Su
%
S with S d < S u
&
Sd
That means that there are 2 states of the world ω0 and ω1 :
the risky asset is worth S1 at t = 1, with S1 (ω1 ) = S u and S1 (ω0 ) = S d .
Let Ω = {ω0 , ω1 }. We assume P ({ω1 }) = p and P ({ω0 }) = 1 − p. P is a probability on (Ω, F)
where F is the σ-algebra {∅, {ω0 }, {ω1 }, Ω}.

We make the assumptions of chapter 1 on the market made of the basis assets and their derivatives
(in particular, short selling of the risk-free asset is equivalent to borrowing at rate r).

Due to the NAO assumption, we have

S d < S(1 + r) < S u (1)

Proof: we argue by contradiction.
If for example S(1 + r) ≤ S d , at time t = 0, you find the following ”free-lunch”:
borrow S $ at rate r, buy the stock. At time t = 1, you get S u − S(1 + r) $ or S d − S(1 + r) $.
Both are non-negative values, and the first one is positive. This is an AO.
(S(1 + r) ≤ S d implies that everybody will buy stocks instead of saving money on the bank
account, whereas S(1 + r) ≥ S u nobody will buy the stock.)
(
+1 U.A.
This strategy corresponds to the portfolio
−S risk − free asset
 Probabilistic Methods in Finance 17

Exercise: write the argument when S u ≤ S(1 + r). (

−1 U.A.
It involves a short sale of the U.A., the corresponding portfolio is
+S risk − free asset.

We consider a European option on the risky asset with maturity T = 1. Such a derivative product
is determined by F1 , its payoff at T : let F1 (ω1 ) = F u , F1 (ω0 ) = F d .

Fu
F ex: call with strike price K ∈]S d , S u [: F u = S u − K (exercise)
Fd and F d = 0 (no exercise).

F1 is F1 -measurable where F1 = {∅, {ω0 }, {ω1 }, Ω} is the information at date 1 (note that F1 = F).

To price the option, we will replicate its payoff at T = 1 with a portfolio consisting of basis
assets. Here, any such derivative asset is replicable (attainable):
there exists a portfolio portfolio consisting of basis assets (α risk-free asset, ∆ risky asset) whose
value at t = 1 is F1 , ie α(1 + r) + ∆S1 = F1

(α, ∆ ∈ R, cf pricing assumptions; ex: α < 0 means borrowing at rate r, ∆ < 0 means short sale
of the risky asset).
Called replicating portfolio.
(
α(1 + r) + ∆S u = F u
Proof: looking for (α, ∆) s.t.
α(1 + r) + ∆S d = F d
Fu − Fd F u − ∆S u
∃ a unique solution: ∆ = and α = .
Su − Sd 1+r

NAO assumptions ⇒ the asset price is:

Fu Su
 
F = α + ∆S = +∆ S−
1+r  1+r 
1 u u d u d u
= u F (S − S ) + (F − F )[S(1 + r) − S ]
(S − S d )(1 + r)  
1 u d d u
= u F [S(1 + r) − S ] + F [S − S(1 + r))
(S − S d )(1 + r)

S(1 + r) − S d
 
(∗)1 ∗ u ∗ d
We get F = p F + (1 − p )F with p∗ = .
1+r Su − Sd
p∗ ∈]0, 1[ since S d < S(1 + r) < S u .

Notes: 1. We are able to price any derivative asset. Would not be the case if the risky asset
had 3 possible values at time 1 (3 equations, 2 unknowns).
General result: as many assets needed as the number of states of the world.
(
−1 option
2. The same equation (*) would be obtained by considering the portfolio: .
∆ U.A.
This portfolio is risk-free, indeed it can take only one value at time 1,
as its ”two” possible values, ∆S u − F u and ∆S d − F d are equal with our choice of ∆.
 Probabilistic Methods in Finance 18

Hence its return can only be r, that leads to the equation (*).

3. You will check easily that for a call with strike K with S d < K < S u (the other calls are always,
or never, exercised, so they are not options), the delta belongs to ]0, 1[.
0 − (K − S d )
For a put with strike K, the delta belongs to ] − 1, 0[, indeed ∆ = .
Su − Sd

Let P ∗ ({ω1 }) = p∗ et P ∗ ({ω0 }) = 1 − p∗ .

P ∗ is a probability on (Ω, F1 ), equivalent to P (ie > 0 on the same events).

1
Equation (*) can be written: F = I ∗ (F1 )
E
1+r
expectation under P ∗ of the future value, discounted at the risk-free rate.

I.e., when P is replaced by P ∗ , we can do as if the investors were risk neutral (indifferent between
receiving F1 , which is random, or receiving EI (F1 ), which is certain, at t = 1): then we get rid of
the question of determining the correct discount rate for the flow F1 .

P ∗ is called risk-neutral probability, and P real or historical probability.

Therefore to price the derivatives, you just have to ”set yourself in the (imaginary, fictitious)
risk-neutral universe”: you attribute the probability p∗ at state ω1 , and 1 − p∗ at state ω1 and you
price all the assets as if the investors were risk-neutral. = Risk-neutral valuation
We will refer to a world where everyone is risk neutral as a risk neutral world. In such a world
investors require no compensation for risk, and the expected return on all securities is the risk-free
rate.

Notes: 1. use the formula for the 2 basis assets. We have in particular S = 1+r 1
I ∗ (S1 ): set-
E
ting the probability of an up movement equal to p∗ is equivalent to assuming that the return on
the stock equals the risk-free rate.
2. we obtain a price formula that does not depend on the real probability! In fact, the price de-
pends on S, and S depends itself on the real probability. The real probability and the risk-aversion
are embedded in the spot price of the U.A., and thanks to the replication, we get the derivative
price as a function of S.

The option pricing formula does not involve the probabilities of the stock price moving up or down.
This is surprising and seems counterintuitive. It is natural to assume that as the probability of an
upward movement in the stock price increases, the value of a call option on the stock increases and
the value of a put option decreases. This is not the case. The key reason for that is that we are
not valuing the option in absolute terms. We are calculating its value in terms of the price of the
underlying stock. The probabilities of future up or down movements are already incorporated into
the price of the stock.

The strategies in this model are static. In a model with several periods, dynamic strategies are
used. Fo this purpose, we need some concepts like filtrations to describe the information structure,
and particular stochastic processes (martingales) to model the prices.
 Probabilistic Methods in Finance 19

Chapter IV. Mathematical tools

(Ω, F, P ) is a given probability space. We will use the following definitions:

• Stochastic process:
(Xt )t∈T family of random variables on (Ω, F, P ) with values in (E, E) measurable space (will
most often be (R, B(R)) ie R equipped with the Borel σ-algebra),
with T = {0, ..., T } in discrete-time or [0, T ] in continuous-time.

Ex: Xt an asset price at time t, then (Xt )t≥0 is a stochastic process.

• Filtration on (Ω, F, P ): non-decreasing family of sub-σ-algebras of F: (Ft )t∈T , s ≤ t ⇒ Fs ⊂ Ft .

(Ft ) = ”information structure”. (Ω, F, P, (Ft )) is called a ”filtered probability space”.

used to model a flow of information, for example corresponding to the observations of an asset
price: the σ-algebra Ft usually models the events which can be observed up to time t.
Eg: with Ss the asset price at time s,
let, for t ≥ 0, Ft = σ{Ss , s ≤ t}, i.e. smallest σ-algebra that contains all pre-images of measurable
subsets of R for times s up to t (or all sets of the form {a ≤ Ss ≤ b} for 0 ≤ s ≤ t, a, b ∈ R).
(Ft )t∈T is obviously non-decreasing. It is called the natural filtration of the process (St )t∈T .
It records the ”past behaviour” of the process, and only that information.

More generally, Ft = information available to investors at time t,

which consists of assets prices before and at time t,
i.e., for N assets: Ft = σ{Ssi , s ≤ t, 1 ≤ i ≤ N }, where Ssi is the price at time s of asset i.

In this set-up, we will always assume F0 = {∅, Ω} and FT = F. Then ”Z F0 -measurable” means
that Z is constant.
As time passes, an observer knows more and more information, that is, finer and finer partitions
of Ω.

Example: just one risky asset is observed, in discrete time, price taking discrete values. Then
its price dynamics can be described in a tree, a node being a set {ω|Xt (ω) = a} (i.e. the pre-image
Xt−1 ({a})) for some time t and some value a.
A state of the world ω is a whole path in the tree (not a terminal node).
Note that this tree corresponds to a family (Pt )t∈T of partitions of Ω, satisfying:
P0 = {Ω}, PT = {{ω1 } , ..., {ωN }}, and ∀A ∈ Pt+1 , ∃A− ∈ Pt such that A ⊂ A− .

The filtration describing the information structure is then given by (Ft ) such that for any t ∈ T ,
Ft is generated by Pt , or, equivalently: Ft = σ{Ss , s ≤ t}.
It is the information available to an observer of the assets prices S· up to time t.

For example, below: P1 = {{ω1 , ω2 , ω3 , ω4 , ω5 } , {ω6 , ω7 , ω8 , ω9 , ω10 , ω11 }}.

P2 = {{ω1 , ω2 } , {ω3 , ω4 } , {ω5 } , {ω6 , ω7 , ω8 } , {ω9 , ω10 , ω11 }}.
 Probabilistic Methods in Finance 20

{ω1 }
.
{ω1 , ω2 }
-
{ω2 }
.
{ω3 }
.
{ω1 , ω2 , ω3 , ω4 , ω5 } ← {ω3 , ω4 }
-
- {ω4 }
.
{ω5 } ← {ω5 }

{ω6 }
.
- {ω6 , ω7 , ω8 } ← {ω7 }
-
. {ω8 }

{ω6 , ω7 , ω8 , ω9 , ω10 , ω11 }

- {ω9 }
.
{ω9 , ω10 , ω11 } ← {ω10 }
-
{ω11 }

This tree describes how the information is revealed:

At t = 0, the available information is: we are in Ω.
At t = 1, we know if we have event {ω1 , ω2 , ω3 , ω4 , ω5 } or event {ω6 , ω7 , ω8 , ω9 , ω10 , ω11 },
and so on...
At t = 3, we know exactly which element of our universe of possible states has been realised.

• A stochastic process (Xt )t∈T is called adapted to the filtration (Ft )t∈T iff for all t ∈ T ,
Xt is Ft -measurable.
X Ft -measurable means: observable at time t given the information Ft . Note that a stochastic
process is always adapted to its natural filtration.

Ex: Ft1 σ-algebra generated by the r.v. representing the prices of all the assets until time t,
Ft2 same, but for quoted assets only.
Xt price at t of a non quoted asset, then (Xt ) is adapted to filtration (Ft1 ), but not to (Ft2 ), as
for a given t, Xt is not Ft2 -measurable.

In the computations at different times, the conditional expectation will be an important tool:
for X ∈ L1 (Ω, F, P ) (real-valued r.v. F = FT -measurable such that EI (|X|) < +∞),
E
I (X|Ft ) is the expectation of X given the information available at time t.
In particular, if X is Ft -measurable, E
I (X|Ft ) = X and,
if X is Ft -measurable and bounded, E I (XY |Ft ) = XEI (Y |Ft ) for any Y ∈ L1 :
ie X is considered as a constant at time t (i.e. known).
 Probabilistic Methods in Finance 21

Reminder on the conditional expectation:

Definition : B sub-σ-algebra of F, X ∈ L1 (Ω, F, P ) (i.e. E

I (|X|) < +∞),
EI (X|B) is the unique integrable r.v. B-measurable such that
∀Y B-mesurable and bounded, E I (XY ) = EI (Y E
I (X|B)).
”Conditional expectation of Y given B”.
For X ∈ L2 (Ω, F, P ), E
I (X|B) is the unique integrable r.v. B-mesurable such that
2
∀Y ∈ L (Ω, F, P ), EI (XY ) = EI (Y E
I (X|B)).
Properties :
· linearity
For any B sub-σ-algebra of F and X ∈ L1 (all the inequalities and equalities below hold P -almost
surely only, except first one):
·E
I (E
I (X|B)) = E
I (X),
· X B-measurable ⇒ E
I (X|B) = X,
· X independent of B ⇒ E
I (X|B) = E
I (X),
· for Y ∈ L1 B-measurable s.t. XY ∈ L1 , E
I (XY |B) = Y E
I (X|B),
· for X, Y ∈ L2 , if Y B-measurable, then E
I (XY |B) = Y E
I (X|B)
· A sub-σ-algebra of B ⇒ E
I (E
I (X|B)|A) = E
I (E
I (X|A)|B) = E
I (X|A) (”tower property”)
· positivity: X ≥ 0 ⇒ EI (X|B) ≥ 0,
1
then if X, Y ∈ L , X ≤ Y ⇒ E I (X|B) ≤ E
I (Y |B).
See also exercise 11.

• Martingale:
(Ω, F, P, (Ft )) filtered probability space.

1. In discrete-time filtration (Fn )n∈N

Def: (Fn )-martingale in discrete-time: (Mn )n∈N (Fn )n∈N -adapted process with values in R, s.t.
∀n ∈ N, Mn ∈ L1 (Ω, F, P ) and EI (Mn+1 |Fn ) = Mn P -a.s. (depends on the probability P ).

Prop: ∗ (Mn ) (Fn )-martingale then ∀m ≥ n, E I (Mm |Fn ) = Mn P -a.s. (from the ”tower prop-
erty”).
∗ E I (Mn ) is constant (cf E
I (E
I (X|B) = E
I (X)).

martingale = mathematical object with good properties (applications in finance and in game theory)

Example of use (tutorial):

(Xn )n∈N∗ sequence of independent r.v. in L2 such that ∀n ∈ N∗ , E
I (Xn ) = 0 (fair play).
Fn generated by {X1 , ..., Xn } ((Fn ) = ”natural filtration”).
Mn = X1 + ... + Xn (gain after n games). (Mn )n∈N∗ is an (Fn )-martingale.
Then expected gain after n games: E
I (Mn ) = E
I (M1 ) = 0.
 Probabilistic Methods in Finance 22

”Strategy”: we can stop at a given date, ex: after gain G

ie replace n by a r.v. function of the obtained ”gain”.

Def: (Fn )-stopping time: τ : (Ω, F, P ) 7→ N ∪ {∞} such that: ∀n ∈ N, {τ = n} ∈ Fn .

It can be proved that: E I (Mτ ) = EI (M1 ) = 0, for any bounded stopping time τ (see tutorial):
i.e. if the player decides to play until a date defined by its gain, its expected gain is still 0.

2. In continuous time filtration (Ft )t≥0

Def: (Ft )t≥0 -martingale: (Mt )t≥0 (Ft )t≥0 -adapted process with values in R such that
∀t ≥ 0, Mt ∈ L1 (Ω, F, P ) and if ∀s ≤ t, E I (Mt |Fs ) = Ms P -a.s.

Ex: For any X ∈ L1 (Ω, F, P ), (E

I (X|Ft ))t≥0 is an (Ft )t≥0 -martingale (same in discrete time).

Proof: (E
I (X|Ft ))t≥0 obviously (Ft )-adapted.
|E
I (X|B)| ≤ EI (|X| |B) (from positivity: X and −X ≤ |X| or using Jensen with ϕ(x) = |x|).
I (|X| |B) is integrable then |E
E I (X|B)| is integrable.

For s ≤ t, Fs ⊂ Ft thus E
I (E
I (X|Ft )|Fs ) = E
I (X|Fs ).

See classical exercices 13. to 15.

 Probabilistic Methods in Finance 23

Chapter V. Option pricing in discrete time: general N periods

model, with d risky assets

I. Market model
A discrete-time financial model is built on a finite probability space (Ω, F, P ) equipped with a
filtration (Fn )0≤n≤N . The horizon N will correspond to the maturity of the options.
Again, for n ≤ N , Fn is the σ-algebra of events up to time n.
We will always assume F0 = {∅, Ω}, FN = F = P(Ω), and ∀ω ∈ Ω, P ({ω}) > 0.

The market consists in (d + 1) financial assets, whose prices at time n are given by the non-negative
random variables Sn0 , Sn1 , ...Snd , Fn -measurable (investors know past and present prices but obvi-
ously not the future ones).
The vector Sn = (Sn0 , Sn1 , ...Snd ) is the vector of prices at time n, (Sn )0≤n≤N is (Fn )0≤n≤N -adapted.
The asset 0 is the risk-free asset, its return over one period is constant and equal to r, taking
S00 = 1, we get Sn0 = (1 + r)n or, more conveniently (then r does not depend on n), ern∆t after n
periods, where ∆t is the length of 1 period.
The coefficient S10 corresponds to the discount factor between time 0 and time n.
n

Remark: even if the interest rate was random, asset 0 would be considered as risk-free as Sn0 is known (unique
value) as soon as rk , k ≤ n is known.
i
Sn
The assets 1 to d are risky assets. We will use below their discounted prices 0.
Sn

We do the usual technical assumptions on the market regarding transaction costs, taxes, short
selling allowed - including for the risk-free asset -, assets divisible (”frictionless market”).

II. Strategies
A portfolio strategy is defined by a stochastic process Θ = ((θn0 , θn1 , ...θnd ))0≤n≤N , where θni ∈ R
denotes the quantity of asset i in the portfolio at time n.
This process is supposed to be predictable, i.e., Θ0 is F0 -measurable, and, for 1 ≤ n ≤ N , Θn is
Fn−1 -measurable.
This assumption means that the positions in the portfolio at time n are decided with respect to
the information available at time n − 1 and kept until time n when new quotes are available for
the assets.
d
X
The value of the portfolio at time n is the scalar product VnΘ = Θn · Sn = θni Sni .
i=0

We define the notion of self-financing portfolio strategy, meaning that at any time n, once the new
prices Sn0 , Sn1 , ...Snd are quoted, the investor readjusts (or rebalances) his position from Θn to Θn+1
without bringing or consuming any wealth external to the portfolio.
Def: the portfolio strategy Θ is said to be self-financing iff ∀0 ≤ n ≤ N − 1, Θn · Sn = Θn+1 · Sn .
The self-financing condition at time n, Θn · Sn = Θn+1 · Sn , is equivalent to:
Θn+1 · Sn+1 − Θn · Sn = Θn+1 · (Sn+1 − Sn ) i.e. Θ −VΘ =Θ
Vn+1 n n+1 · (Sn+1 − Sn )

Then the variation of value of the portfolio between time n and time n + 1 is due to the assets price
moves only.
 Probabilistic Methods in Finance 24

Note: this notion is obviously useless when one period only is considered like when deriving forward prices,
the call-put parity relationship, or when replicating the option in the 1-period binomial model: the portfo-
lios involved in our arguments were static strategies hence obviously self-financing strategies, it is useful for
dynamic strategies only.

We write the self-financing condition on the discounted prices w.r.t. time 0:

Yn being an asset price (or a vector of asset prices) at time n, let Ỹn = SYn0 (e.g. e−rn∆t Yn ).
n

Ex: S̃n0 is constant, equal to 1.

For any n, Θn · Sn = Θn+1 · Sn ⇐⇒ Θn · S̃n = Θn+1 · S̃n
⇐⇒ Θn+1 · S̃n+1 − Θn · S̃n = Θn+1 · [S̃n+1 − S̃n ] ie Θ − Ṽ Θ = Θ
Ṽn+1 n n+1 · (S̃n+1 − S̃n ).

Therefore, the portfolio strategy Θ is self-financing iff:

n−1
X
Θ
∀1 ≤ n ≤ N, Ṽn = Ṽ0 + Θ Θk+1 · (S̃k+1 − S̃k )
k=0

Knowing S̃k0 = 1 for any k, we can see that a self-financing strategy is determined by its initial
value V0Θ (= Ṽ0Θ ) and the quantity of risky assets at any time (the proceeds of the trading in the
risky assets are put in cash).

III. No arbitrage opportunity property

Short-selling and borrowing are allowed, but we add a contraint on the sign of the portfolio’s values
at all times:
Definition: A strategy Θ is admissible if it is self-financing and if VnΘ ≥ 0 for any n ∈ {0, 1, ..., N }.
i.e. the investor must be able to pay back his debts (in risk-free or risky asset) at any time.
In this multi-periods framework, the notion of arbitrage opportunity is formalised as follows:
Def: an arbitrage opportunity is:
an admissible strategy with zero initial value and non-zero final value.
Hence an arbitrage opportunity is a a self-financing strategy such that:
V0Θ = 0, ∀n ≤ N , VnΘ ≥ 0 (including at time N ), and P (VNΘ > 0) > 0.

Th1: The market is without arbitrage (NAO property, or ”viable market”) iff there
exists a probability P ∗ equivalent to P such that the discounted prices of the d + 1
basis assets are martingales under P ∗ .
P ∗ is called an Equivalent Martingale Measure (EMM) or a Risk-neutral probability as we
will have the risk-neutral valuation under this probability.

Reminder: two probability measures P1 and P2 are equivalent if and only if for any event A ∈ F, P1 (A) = 0
iff P2 (A) = 0. Here, P ∗ equivalent to P means that, for any ω ∈ Ω, P ? ({ω}) > 0.

Lemma:
if there exists an EMM P ∗ , then, for any self-financing strategy Θ,
(ṼnΘ ) is a martingale under P ∗ .
n−1
X
Θ
Indeed, we have: ∀n ≤ N , Ṽn = Ṽ0 + Θ Θk+1 · (S̃k+1 − S̃k ).
k=0
 Probabilistic Methods in Finance 25

Thus (exercise 15.), (ṼnΘ ) is a martingale under P ∗ (sum of d martingale transforms: (S̃ni ) by
i
(θn+1 ), for i = 1, ..., d, with all the θni bounded as Ω is finite).

Proof of Th1: ⇐

I ∗ being the expectation under probability measure P ∗ , for any self-financing strategy Θ, we have
E
I ∗ (ṼNΘ ) then the value at time 0 of the strategy is: V0Θ = S10 E
Ṽ0Θ = E I ∗ (VNΘ ) .
N

∗
If the strategy is admissible and its initial value is zero, then E
I (VNΘ )
= 0, with VNΘ ≥ 0 P a.s. then
P ∗ a.s.. Hence VNΘ = 0 since P ∗ ({ω}) > 0 for all ω ∈ Ω (like for P ). Hence there is no arbitrage
opportunity.
(otherwise stated: when an EMM P ∗ exists, [VNΘ ≥ 0 P a.s. and P (VNΘ > 0) > 0] ⇒ same for P ∗ ,
then V0Θ = EI ∗ (VNΘ ) > 0, hence Θ cannot be an AO).

⇒ The proof of the converse implication is more tricky (the probability P ∗ is built using the convex sets
separation theorem). We will build P ∗ in the example of the binomial tree only.

IV. Complete markets and option pricing

A European option with maturity T is given by its payoff at time T (after N periods):
FN = f (S1 , ..., SN ) ≥ 0, with f measurable (borelian). FN is FN -measurable.
Note that the filtration (Fn )0≤n≤N will be generally the natural filtration of the prices process.
Reminder: X is σ(S1 , ..., SN )-measurable iff X = f (S1 , ..., SN ) with f measurable.

For instance, FN = (SN 1 − K)+ for a call on the asset 1 with strike price K. In this example, F
N
is a function of SN only. There are some options dependent on the whole path of the underlying
asset, i.e. FN is a function of S0 , S1 , ..., SN , they are said to be path-dependent (e.g.: Asian options
involve the average of the stock prices observed during a certain period of time before maturity).

Def: an option with payoff FN FN -measurable is said to be replicable (or attainable) if there
exists an admissible portfolio strategy Θ whose value at time N is: VNΘ = FN P a.s. (the corre-
sponding portfolio is called a replicating portfolio).

Remark: In a market without arbitrage, we just need to find a self-financing strategy worth FN at
maturity to say that FN is replicable (as FN ≥ 0). Indeed, as an EMM exists, this strategy will
always be admissible:
if Θ is a self-financing strategy and if P ∗ is a probability equivalent to P under which discounted
prices are martingales, then (ṼnΘ ) is also a martingale under P ∗ (like in proof of Th1). Hence, for
n ≤ N , ṼnΘ = E I ∗ (ṼNΘ |Fn ). Therefore, if VNΘ ≥ 0 P a.s. (which is the case for an option), then, for
any n, ṼnΘ ≥ 0 P a.s., hence the strategy is admissible.
(retain: Θ self-financing with VNΘ ≥ 0 P ∗ a.s. ⇒ Θ admissible).

Def: the market is complete if any option written on the risky assets is replicable.
Note: to assume that a financial market is complete is a rather restrictive assumption that does not have
such a clear economic justification as the no-arbitrage assumption. The interest of complete markets is that
it allows us to derive a simple theory of derivatives pricing and hedging.
 Probabilistic Methods in Finance 26

The Cox-Ross-Rubinstein model is a very simple example of complete market modeling.

The following theorem gives a precise characterisation of complete markets without arbitrage.

Th2: A market without arbitrage is complete if and only if there exists a unique
probability P ∗ equivalent to P under which discounted prices are martingales.

The probability P ∗ will appear to be the computing tool whereby we can derive closed-form pricing
formulae and hedging strategies.

Proof: ⇒ We assume that the market is without arbitrage and complete. Then there exists some
EMM (Th1).
Consider P1 and P2 two probability measures under which discounted prices are martingales. We
want to prove they are equal.
For any ∀B ∈ FN , 1IB is FN -measurable, then replicable: can be written as 1IB = VNΘ , where Θ
is an admissible strategy that replicates the payoff 1IB . Since Θ is self-financing, we know that
n−1
1IB X
0 = Ṽ Θ
N = V0
Θ
+ Θk+1 · (S̃k+1 − S̃k ).
SN
k=0

(ṼnΘ ) is a martingale under both P1 and P2 . It follows that, for i = 1 or 2, E

I i (ṼNΘ ) = E
I i (V0Θ ) = V0Θ ,
the last equality coming from the fact that F0 = {∅, Ω}.

Then Pi (B) = E I i (1IB ) = E 0 Ṽ Θ ) = S 0 V Θ , where Θ is a replicating strategy. Then P (B) is

I i (SN N N 0 i
uniquely determined (if two strategies exist, they must have same value at 0, by NAO).
Therefore P1 (B) = P2 (B) and since B is arbitrary, P1 = P2 on the whole σ-algebra FN (assumed
to be equal to F).

⇐ By contradiction (admitted): if there exists a random variable FN ≥ 0, FN -measurable, which is not

replicable, it is possible to build another probability equivalent to P under which discounted prices are
martingales.

Conclusion: pricing and hedging derivatives in complete markets

The market is assumed to be without arbitrage and complete. We denote by P ∗ the unique
probability under which the discounted prices of financial assets are martingales. Let FN be an
FN -measurable, non-negative random variable and Θ be an admissible strategy replicating the
derivative hence defined, i.e. VNΘ = FN .

I ∗ (ṼNΘ |Fn ).
The sequence (ṼnΘ ) is a P ∗ -martingale, then for n = 0, 1, ..., N , ṼnΘ = E
In particular, V0Θ = Ṽ0Θ = E I ∗ ( SN0 ) (as stated in the proof of theorem 1).
F
N

We get the same results as in the one-period model: V0Θ = 1

I ∗ (FN )
0 E
SN
.

Hence we have the risk-neutral valuation under P ∗ like in the one-period model.
P ∗ is called again the risk-neutral probability.

Here again, the computation of the option price does not require the knowledge of the probability
P , but of P ∗ only.

Note that the value of an admissible strategy replicating FN is completely determined by FN

 Probabilistic Methods in Finance 27

at any time. VnΘ corresponds to the price of the option: that is the wealth needed at time n to
replicate FN at time N by following the strategy Θ.
I ∗ ( FSN0 ), he can follow a replicating strategy Θ in order
If, at time 0, an investor sells the option for E
N 


 −1 option
 θ1 assets 1

n
to generate an amount FN at time N . His portfolio at time n = 1− , ..., N − is:


 ...
 θd assets d

n
This strategy allows him to be perfectly hedged.

V. Example: N periods binomial model (Cox-Ross-Rubinstein)

1. Model
Pricing of an option with maturity T on a risky asset. The U.A. price is modeled in discrete time.
T
We divide [0, T ] in N sub-periods with duration ∆t = N .
We assume that the risky asset price changes only at the discrete times ∆t, 2∆t,..., (N − 1)∆t
(dynamic model in discrete time).
2 basis assets:
- a risk-free asset whose price is 1 at time 0. Its return is r (continuous interest rate chosen as
we will take N large later). Then this asset is worth Sn0 = ern∆t after n periods,
T
- a non-dividend-paying stock whose price is Sn after n periods (corresponds to time n N ). The
initial price S0 is denoted by S.
The most elementary model is to suppose that, given Sn , Sn+1 has two possible values.
The price process is then described by a tree. For programming purpose, we need a recombining
tree (n + 1 possible values at time n, 2n for the non-recombining tree).
We will assume that, with 0 < d < u, at each period: Sn+1 = uSn or dSn , then the tree is
recombining.
Su2
Su
S Sdu
Sd
Sd2

It is equivalent to assume that the rate of return on the asset over each period of time inter-
val ∆t is worth u − 1 or d − 1.

To calculate the option price at the initial node of the tree we will apply repeatedly the prin-
ciples established for one period: the option will be replicated by a portfolio rebalanced at each
period. Then the dynamics of the U.A. price matters: the moves of the U.A. give the relevant
”events” (no need to consider other assets).
Definition of Ω: an elementary event ω corresponds to a path in the tree (observation of the risky
asset prices S1 , ..., SN , or of (X1 , ..., XN ) where Xn = SSn−1
n
, e.g.: ω = (u, d, d, ..., u)).
Ω = {d, u}N : each N-tuple represents the successive values of Xn .
 Probabilistic Methods in Finance 28

Information at time n: path until the current node. It corresponds to the natural filtration of the
U.A. price process. Let F0 = {∅, Ω} and for n ≥ 1, Fn = σ(S1 , ..., Sn ).
We assume FN = F, ie no other uncertainty than the risky asset price.
No assumption on the real probability of each path, as the probability of d and u was not involved in the
one-period option price formula. In particular, the probabilities of each move on the sub-periods are not
necessarily constant in the tree.
Note: the knowledge of a probability on F is equivalent to the knowledge of the law of (X1 , ..., XN ).

The market is assumed frictionless.

2. Self-financing strategies (same as in the general case)

A portfolio strategy is defined by a predictable stochastic process Θ = ((θn0 , θn ))0≤n≤N :
at time n, there are θn0 risk-free asset, θn risky asset, with θn0 , θn Fn−1 -measurable for 1 ≤ n ≤ N .
The value of the portfolio at time n is VnΘ = Θn · Sn (scalar product), with Sn = (Sn0 , Sn ).
Yn being an asset price at time n, the discounted price is computed as Ỹn = e−rn∆t Yn .
The portfolio strategy Θ is self-financing iff ∀0 ≤ n ≤ N − 1, Θn · S̃n = Θn+1 · S̃n or
n−1
X
iff ∀1 ≤ n ≤ N, ṼnΘ = Ṽ0Θ + θk+1 (S̃k+1 − S̃k ) 0
(indeed S̃k+1 = S̃k0 for any k).
k=0

i.e. a self-financing strategy is determined by its initial value V0Θ and its quantity of risky asset at
any time (the proceeds of the trading in the risky asset are put in cash).

3. Risk-neutral probability
We assume that the market is without arbitrage, then there exists a probability P ∗ equivalent to
P such that (S̃n )0≤n≤N is a martingale under P ∗ ((S̃n0 )0≤n≤N is constant).
We have for 0 ≤ n ≤ N − 1, E I ∗ (S̃n+1 |Fn ) = S̃n , with S̃n+1 = e−r∆t S˜n Xn+1 and S˜n Fn -measurable.
I ∗ (Xn+1 |Fn ) = er∆t and E
Therefore E I ∗ (Xn+1 ) = er∆t .
er∆t −d
It gives P ∗ (Xn+1 = u)u + [1 − P ∗ (Xn+1 = u)]d = er∆t hence P ∗ (Xn+1 = u) = u−d .

er∆t −d
We get ∀0 ≤ n ≤ N − 1, P ∗ (Xn+1 = u) = p∗ , where p∗ = u−d .

The condition P ∗ (Xn+1 = u) ∈]0, 1[ (as P ∗ is equivalent to P ) implies d < er∆t < u, which makes
sense as it is needed to avoid the existence of arbitrage opportunities on the sub-periods (which
would lead to AO between 0 and T , using the risk-free asset). And we find on each mesh the same
r∆t d
coefficient p∗ = e u−d−d (corresponding to what we found in the 1-period tree ( S(1+r)−S
S u −S d
).

We build the tree with these new values p∗ (for the up move) and 1 − p∗ (for the down move) at
each step. This leads to the following definition: let P ∗ the probability defined by:
for a given path ω, P ∗ (ω) = p∗k (1 − p∗ )N −k , where k is the number of moves ×u in the path.
P ∗ is equivalent to P .

We assume that the successive moves are independent under P ∗ .

(
u with probability p∗
Under P ∗ , the r.v. (Xn ) are i.i.d. with law: Xn = .
d with probability 1 − p∗
 Probabilistic Methods in Finance 29

Then for 0 ≤ n ≤ N − 1, E I ∗ (S̃n+1 |Fn ) = e−r∆t S˜n E

I ∗ (Xn+1 |Fn ) = e−r∆t S˜n E
I ∗ (Xn+1 ) = S˜n , thus
(S̃n )0≤n≤N is a martingale under P ∗ . And P ∗ is the unique probability equivalent to P having that
property.

4. Option pricing
Consequence: the market is complete. Then, as in the general model, to price any derivative prod-
uct given by its payoff FN = f (S1 , ..., SN ) ≥ 0 at time T (after N periods), we just need to replicate
it with a self-financing strategy Θ. We have that (ṼnΘ )0≤n≤N is a martingale under P ∗ , and then,
the value at time 0 of the derivative is:

I ∗ (VNΘ )= e−rT E
F0 =V0Θ = e−rT E I ∗ (FN ).

Note that for a standard option (non path-dependent), FN is a function of SN only. We compute
the price using the law of SN under P ∗ , given by:
P ∗ (SN = Suk dN −k ) = CN
k p∗k (1 − p∗ )N −k for 0 ≤ k ≤ N .

The replicating strategy is built through a recursive backward procedure: starting at the expi-
ration date and working backwards, using the result on one period (=what is done by programs).

On any mesh between time N − 1 and N , we can replicate the payoff at N (see one-period model).
The replicating strategy is then computed by backward induction. The part in cash is uniquely
determined (self-financing strategy).
Note that the discounted price of the option is a martingale under P ∗ (indeed it is (ṼnΘ )).

Th: the option price at time t = n∆t is given by: I ∗ (FN |Fn ) .
Fn = e−r(T −t) E

expectation under P ∗ of the future value, discounted at risk-free rate.

I.e.: replacing P by P ∗ , we can compute prices as if the investors were risk-neutral.

P ∗ is called the risk-neutral probability and P real or historical probability (ie as anticipated).

We still have the principle of risk-neutral valuation, like in the 1-period model:
the option price is equal to its expected payoff in a risk-neutral world (the tree is built
under P ∗ ), discounted at the risk-free rate.

The real probability does not appear in the price formula.

5. Delta hedging
The price is obtained by replicating the option.
At each node, by holding a given number of U.A., we know how to replicate the option on the next
period.
This number depends on the date n and on the state of the world at that date. It is called the
delta.
For n ∈ {1, ..., N }, let Snk be the risky asset price at time n after k up moves, and Fnk the corre-
sponding option price.
 Probabilistic Methods in Finance 30

At time n − 1 after k up moves:

Snk+1 Fnk+1
% %
k
Sn−1 k
Fn−1
& &
Snk Fnk

Prices of time n − 1 are known, then, for the U.A. price or for the option price, the 2 possible
values are known as well. We compute the quantity of U.A. to be held until time n (included)
F k+1 − Fnk
as: ∆kn = nk+1 (value known at n − 1). It corresponds to the coefficient θn of the above
Sn − Snk
replicating strategy Θ. The other coefficient θn0 is chosen in order to have a self-financing strategy
Θ. The strategy Θ is determined by (∆1 , ∆2 , ..., ∆n ).
The predictability is clear from above computation.

The option can then be replicated by holding a portfolio that contains delta U.A. and some cash,
which implies buying and selling continuously, to maintain the correct quantity of U.A. (self-
financing portfolio strategy).
This portfolio is worth exactly as the option at any time / state of the world.
The option price does not depend on the real probability because it is a replication price.

The bank that sells the option is exposed to a risk: it has to pay FN at T , potentially large amount.
Generally, it does not speculate: as soon as the option is sold, with the premium F , it builds the
replicating portfolio, called hedging portfolio: the premium is invested in that portfolio, then the
correct quantity of U.A. is maintained during the whole life of the option (dynamic hedge).
For the bank, the final outcome is then null: the risk is cancelled.

The bank becomes indifferent to the U.A. price variations (up or down moves), indeed:
(
−1 option
the portfolio is risk-free between n − 1 and n.
∆kn U.A.
which is the main property of the delta.

This trading is equivalent to building (manufacturing) the option.

In fact, the bank sells at price F + margin and it is this service that the buyer pays with the margin.
Moreover: theoretically, there is a perfect replication, but not in reality, then part of the margin
exists to cover the residual risk.

So the writer of the option (for example a bank that has sold a put on a currency) can cover his
position by purchasing the underlying so that the loss in the short position in the option is offset by
the long position in the stock. This construction of a risk-free hedge is referred to as delta hedging.
 Probabilistic Methods in Finance 31

Chapter VI. Option pricing in continuous time: Black-Scholes

model

More complex.
Cox, Ross, Rubinstein (1979) Option pricing: a simplified approach. J of Financial Economics.
Abstract: ”This paper presents a simple discrete-time model for valuing options. The fundamental
economic principles of option pricing by arbitrage methods are particularly clear in this setting.
Its development requires only elementary mathematics, yet it contains as a special limiting case
the celebrated Black-Scholes model, which has previously been derived only by much more difficult
methods.”

Mathematical models of financial derivatives, Y.K. Kwok, Springer (2nd ed 2008) (pages 1-130).
Baxter, M. and Rennie, A. Financial calculus. Cambridge University Press, 1996.

I. Brownian motion and Ito processes

(Ω, F, P ) probability space.

The risky asset price is now modeled in continuous time, and taking continuous values.
(real world: discrete values, ex: cents, and discrete variations, only when market open...).

Stochastic process: sequence of r.v. (Xt )t∈R+ . For a given ω, t 7→ Xt (ω) = path.

Def: (Xt )t≥0 is said to be continuous, or to have continuous paths

iff P -as in ω, t 7→ Xt (ω) is continuous.

1. Brownian motion
In 1828, the botanist Robert Brown observed the motion of a pollen particle suspended in water
under a microscope and described it as a continuous jittery motion: the particles moved in an
irregular, random fashion. Ceaseless (incessant) changes of direction. The phenomenon is now
known as Brownian motion.
1905: Einstein argued that the jiggling of the pollen grains seen in Brownian motion was due to
molecules of water hitting the tiny pollen grains.
1923: Wiener (and Einstein) proposes a rigorous mathematical study of Brownian motion (in par-
ticular proves the existence).

Existence theorem (admitted): (Ω, F, P ) probability space.

There exists a stochastic process (Bt )t≥0 with real values and continuous paths (P a.s.
t 7→ Bt is continuous), called Brownian motion s.t. B0 = 0 and:
(
Bt − Bs ∼ N (0, t − s)
∀0 ≤ s ≤ t,
Bt − Bs is independent of σ(Bu , u ≤ s) (the past)

Note: ∀t ≥ 0, Bt ∼ N (0, t) (as it is Bt − B0 ).

In fact, for any division in sub-intervals, Bt is the sum of some independent Normal r.v..
Consistency: t1 < t2 < t3 ,
Bt2 − Bt1 ∼ N (0, t2 − t1 ), Bt3 − Bt2 ∼ N (0, t3 − t2 ) and independent r.v.;
allows indeed Bt3 − Bt1 ∼ N (0, t3 − t1 ).
 Probabilistic Methods in Finance 32

Properties: Let (Bt ) be Brownian motion. Let Ft = σ(Bs , s ≤ t).

λ2
(Bt )t≥0 , (Bt2 − t)t≥0 and, for λ ∈ R, (eλBt − 2
t
)t≥0 are (Ft )-martingales with continuous paths.

Proof (tutorial): 2 main properties of the conditional expectation are used: for X ∈ L1 ,
· X B-measurable ⇒ E I (X|B) = X
· X independent of B ⇒ E I (X|B) = E I (X)

The 3 processes are obviously (Ft )-adapted.

∀t ≥ 0, Bt ∼ N (0, t) therefore the r.v. are integrable: Gaussian variables have moments of any
λ2 2
order and X ∼ N (m, σ 2 ) ⇒ E I (eλX ) = eλm+ 2 σ
(see characteristic function, generating function, or Laplace transform of the Gaussian law)
λ2 λ2
I (eλBt ) = e
then E 2
t
I (eλBt −
or E 2
t
) = 1.

Let s ≤ t. We use below: (∗) Bt − Bs is independent of Fs and (∗∗) Bt − Bs ∼ N (0, t − s).

∗ ∗∗
·E
I (Bt |Fs ) = E I (Bt − Bs ) + Bs = Bs (uses Bs Fs -measurable),
I (Bt − Bs + Bs |Fs ) = E

∗ ∗∗
I ((Bt − Bs )2 |Fs ) = E
·E I ((Bt − Bs )2 ) = t − s and
=E I (Bt2 |Fs ) − Bs2 from the martingale property of (Bt ) and Fs -measurability of Bs .
I (Bt2 − t|Fs ) = Bs2 − s.
Then E

λ2 ∗ λ2 ∗∗
I (eλ(Bt −Bs )−
·E 2
(t−s)
I (eλ(Bt −Bs )−
|Fs ) = E 2
(t−s)
) = 1.
λ2 2 λ2
λBt − λ2 t
eλBs − 2
s
is Fs -measurable, then E
I (e |F ) = eλBs −
s 2
s
.

Note that Cov(Bs , Bt ) = s ∧ t indeed if s ≤ t, E

I (Bs Bt ) = E I (Bs2 ) = s = s ∧ t.
I (Bt |Fs )) = E
I (Bs E

The Markov property for a process (Xt )t≥0 on a filtered probability space (Ω, F, P, (Ft≥0 ))
means that at any given time t (present), its future behavior is independent of the past:

for h ≥ 0, the law of Xt+h (future) depends on Xt only (present) and not on the Xs , s < t (past).

Rigorous definition:
Def: (Xt )t≥0 , an (Ft )t≥0 -adapted process with values in Rk , is a Markov process with respect to
(Ft ) iff for any Borelian bounded function f : Rk → R, we have:

∀s ≤ t, E
I (f (Xt )|Fs ) = E
I (f (Xt )|σ(Xs ))
If no filtration is mentioned, it may be assumed to be the natural one generated by (Xt )t≥0 .

In finance: the assets prices are frequently modeled as Markov processes (ex: binomial model).
It is an efficient-market hypothesis (EMH): the current price reflects already all the available in-
formation (including the information embedded in past prices).
Note that Markovian prices would not be compatible with the possibility of ”chartism” (or tech-
nical analysis): indeed this assumption implies the same dynamics after St , is it after an upward
move or after a downward move, while chartists use past values of stock prices (and volume) to try
to forecast future stock prices. They try to detect presence of geometric shapes in historical price
charts, ex: head-and-shoulders.
According to the weak-form efficient-market hypothesis, such forecasting methods are valueless,
 Probabilistic Methods in Finance 33

since prices follow a random walk or are otherwise essentially unpredictable.

Interpretation of EMH: there are many investors; if the price was forecast to increase with a high
probability, everybody would buy, the price would increase instantaneously, deleting the pattern:
the current price contains reflects already this anticipation).

Note: in our both models (discrete and continuous), the U.A. price is Markovian and the price of
an option depends only on the U.A. price at that date, not on its past values (except obviously for
path-dependent options, i.e. whose payoff depends on the past values of the U.A. price, for exemple options
involving an average of the U.A. prices).

Th: The Brownian motion is a Markov process.

idea: Bs+h = Bs + Xh , with Xh ∼ N (0, h) independent of Fs = σ(Bu , u ≤ s).

Proof: f : R → R Borelian bounded function, s ≤ t, we have E I (f (Bs + Bt − Bs )|Fs )

I (f (Bt )|Fs ) = E
with Bs Fs -measurable and Bt − Bs independent of Fs .

To compute this term, we need the following result:

Lemma: for B sub-σ-algebra of F,

let X, Y r.v. with values in (R, B(R)), with X B-measurable and Y independent of B.
For any f : R2 → R Borelian bounded (or non-negative), we have:
I [f (X, Y )|B] = ϕ(X) P -as, with ϕ Borelian, defined on E by ϕ(x) = E
E I [f (x, Y )],
i.e. can be computed as if X was a constant.

This is written: E
I [f (X, Y )|B] = E
I [f (x, Y )]/x=X .
R
Proof of the lemma: by definition, ϕ(x) = F f (x, y)dPY (y) where PY is the distribution of Y .
The measurability of ϕ results from the Fubini theorem? (recalled below), as f is bounded.
Then ϕ(X) is B-measurable.
Let Z B-measurable bounded, let PX,Z denote the law of (X, Z).
Y and (X, Z) being independent, we have:
RR R R

E
I [f (X, Y )Z] = f (x, y)z dPX,Z (x, z)dPY (y) = f (x, y)dPY (y) z dPX,Z (x, z)
R
= ϕ(x)z dPX,Z (x, z) = E I [ϕ(X)Z]. 

The Fubini theorem? states that it is possible to compute a double integral (assumed to be finite when the integrand is
replaced by its absolute value) by using an iterated integral, and that one may switch the order of integration:
R
f : (E × F, E ⊗ F) → R measurable and s.t. |f |dPX,Y < ∞,
R
then x 7→ f (x, y)dPY (y) is measurable and integrable w.r.t. dPX .

I (f (Bs + Bt − Bs )|Fs ) as if Bs was constant:

Back to the theorem’s proof: we compute E
E I (f (x + Bt − Bs )|Fs )/x=Bs = E
I (f (Bt )|Fs ) = E I (f (x + Bt − Bs ))/x=Bs
E
I (f (Bt )|σ(Bs )) is computed on the same way as:
I (f (x + Bt − Bs )|σ(Bs ))/x=Bs = E
E I (f (x + Bt − Bs ))/x=Bs , both are then equal.

Quadratic variation of the Brownian motion

 Probabilistic Methods in Finance 34

Th: For a given T > 0, let t0 = 0 < t1 < ... < tn = T a subdivision of [0, T ].
n−1
L2
X
We have: (Btk+1 − Btk )2 −→ T when max |tk+1 − tk | → 0.
k
k=0

Proof: see exercise 20.

Note that max |tk+1 − tk | is the length of the longest of the subintervals of the partition, it is called
k
the norm (or step, or mesh) of the partition, or ’subdivision step’, sometimes denoted by |{tk }|.

The quadratic variation of the Brownian motion (or of any stochastic process) over [0, T ] is defined
as the limit in probability of the sum above, and denoted by [B, B]T ).
Here the convergence is in L2 , which implies convergence in probability. We get:

The quadratic variation of the Brownian motion over [0, T ] is T .

n−1
X
For a class C 1 function f , we have on contrary: [f (tk+1 ) − f (tk )]2 −→0 when |{tk }| → 0.
k=0

Ie classC1 implies vanishing of the quadratic variation of the function.

So, for a given ω, how smooth can t 7→ Bt (ω) be?

It can be proved that, as soon as a continuous process is of bounded variation on [0, T ], then
it has a quadratic variation on [0, T ] equal to 0 (see Appendix on the EPI).
So, it is not the case for a Brownian motion.

In fact, a Brownian motion has very irregular paths (between any two times, a Gaussian vari-
able is drawn):
for almost every ω ∈ Ω, the Brownian motion path t 7→ Bt (ω) is continuous,
but for almost every ω, t 7→ Bt (ω) is nowhere differentiable.

Precisely, the set below contains an event with probability 1:

Bt+h (ω)−Bt (ω) Bt+h (ω)−Bt (ω)
{ω ∈ Ω | ∀t ≥ 0, either limh→0+ h = +∞, either limh→0+ h = −∞}
(while lim = lim < +∞ is needed to have differentiability at t)

In the option pricing model in continuous time, the instantaneous variation of the U.A. price
dSt will involve the instantaneous variation of the Brownian motion dBt .
P
In discrete time the value of a self-financing
Z strategy has been expressed asZ k θk+1 (Sk+1 − Sk ).
In continuous time this is replaced by θt dSt , which involves an integral θt dBt .

But
Z for a given ω, t 7→ Bt (ω) is not of bounded variation, hence we cannot define the integral
θt (ω)dBt (ω) as a Stieltjes integral.
Z
θt dBt cannot be defined pathwise (i.e. for each ω separately).

Therefore Ito integral = new concept.

We define it on an interval [0, T ] (T will be the maturity of the options in the applications).
 Probabilistic Methods in Finance 35

Z T
Problem: defining Ht dBt with (Ht ) stochastic process, cannot be done ω by ω.
0

Reminder: the Riemann integral is defined first for step functions, then extended to a larger class
of functions (the Riemann-integrable functions) by approximation: the integral of a function f is
defined to be the limit of the integrals of step functions which converge (in a certain sense) to f .
We do the same to define the Ito integral. Step functions are replaced by simple processes, which
are random step functions. The integral is then extended to larger classes of processes by approxi-
mation.

• def: (Ht ) is a simple process (or elementary process) if for some finite sequence of times t0 = 0 <
t1 < ... < tn = T , we have: ∀t ∈ [0, T ], P -as in ω,
n−1
X
Ht (ω) = H k (ω)1I[tk ,tk+1 [ (t) where for 0 ≤ k ≤ n−1, H k ∈ L2 (Ω, F, P ) and is Ftk -measurable.
k=0
Rb
To ensure adBt = Bb − Ba , we take the following definition for H simple process as above:
Z T n−1
X
def: Ht dBt = H k (Btk+1 − Btk ).
0 k=0
Note that for a simple process, Ht (ω) is evaluated at the left-hand point of the interval in which t
falls. This is a key component in the definition of the stochastic integral and it makes the resulting
theory suitable for financial applications. In particular, if we interpret Ht (ω) as a trading strategy
and the stochastic integral as the gains or losses from this trading strategy, then evaluating Ht (ω)
at the left-hand point is equivalent to imposing the non-anticipativity of the trading strategy, a
property that we had in discrete time as well.
The function I: E = {simple processes} → L2 (Ω)
Z T
H 7−→ I(H) = Ht dBt satisfies: ||I(H)||L2 (Ω) = ||H||L2 (Ω×]0,T [)
0

i.e. it is an isometry from E, equipped with norm L2 (Ω×]0, T [, F × B]0,T [ , P × dt), to L2 (Ω, F, P ) .
| {z }
complete space
Z t 2  Z t 
2
Proof of E
I Hs dBs =E
I Hs ds for H simple process: see exercise 23.
0 0

• extension by taking limits for other processes:

R 
T 2 ds < +∞. Indeed,
- first extension to processes H. (.) measurable, (Ft ) adapted s.t. E I 0 Hs
such a process is the limit in L2 (Ω×]0, T [, F × B]0,T [ , P × dt) of a sequence of simple processes;
RT 2
0 Hs dBs is then defined as the limit in L (Ω, F, P ) of the sequence of corresponding integrals.

We admit that this limit is the same for any sequence of simple processes converging to H. (.).
RT
(- next extension: to processes H. (.) measurable, (Ft ) adapted s.t. 0 Hs2 ds < +∞ P -as).

Z T
Example: compute Bt dBt (question 2 of exercise 22.).
0
n−1
X
kT
For n ∈ N, we set tk = n for k = 0, ..., n, and Btn = Btk 1I[tk ,tk+1 [ (t).
k=0
(Btn )0≤t≤T is a simple process as Btk ∈ L2 (Ω, F, P ) and is Ftk -measurable.
 Probabilistic Methods in Finance 36

? When n → +∞, (Btn )0≤t≤T converges in L2 (Ω×]0, T [, F × B]0,T [ , P × dt) toward (Bt )0≤t≤T ,
indeed we have, for a given n (note that the dates t0 , t1 , ..., tn depend on n):
n−1
Z T ! n−1 Z
 X Z tk+1 X tk+1
[Btn − Bt ]2 dt = E [Btn − Bt ]2 dt = I [Btk − Bt ]2 dt


EI I E
0 k=0 tk k=0 tk
n−1
X tk+1 n−1
X Z tk+1 T2
Z
T
= (t − tk )dt ≤ 1dt = → 0.
tk n tk n
k=0 k=0

Z T X
? Then, with ∆Bk = Btk+1 − Btk , Bt dBt = lim (in L2 ) Btk ∆Bk (by definition).
0 n→+∞
k
X X
But BT2 = B02 + 2 − Bt2k ) = [(Btk + ∆Bk )2 − Bt2k ] = [2Btk ∆Bk + (∆Bk )2 ]
P
k (Btk+1
k k
Z T X Z T
=2 Btn dBt + (Btk+1 − Btk ) 2
converges to 2 Bt dBt + T when |{tk }| → 0.
0 k 0

Z T
We get BT2 = 2 Bt dBt + T . That is denoted as ”d(Bt2 ) = 2Bt dBt + dt ”,
0

while for a C 1 function f , we have: ”d(f (t)2 ) = 2f (t)df (t)” i.e. (f 2 )0 (t) = 2f (t)f 0 (t), hence
Z T
2
[f (T )] = 2 f (t)df (t) (assuming f (0) = 0).
0

When using the Ito integral, the ”differential calculus rules” (giving the way to dif-
ferentiate t 7→ f (Bt )) are modified by an additional term linked to the fact that the
quadratic variation of the Brownian motion on [0, T ] is not 0.

2. Ito processes
(Bt )t≥0 Brownian motion. ∀t ≥ 0, E
I (Bt ) = 0
but we want to model processes whose expectation increases with time (equity price).
→ we build more general processes from (Bt ).

def: Ito process: (Xt )t≤T with real values s.t. its variation between t and t + dt can be written:
dXt = a(t, Xt )dt + b(t, Xt )dBt , with a, b : [0, T ] × R → R C 0 .

To be understoood as a Znotation for the Zequation:

t t
∀t ≤ T, Xt = X0 + a(s, Xs )ds + b(s, Xs )dBs P -as.
0 0
Note that for a and b having good properties (that we assume), this equation has a unique solution.
a(t, Xt ) is called the drift (instantaneous trend).

An Ito process is a Markov process, heuristic proof:

(Bt ) is a Markov process, then dBt , interpreted as Bt+dt − Bt , is independent of Ft .
From Xt+dt − Xt = a(t, Xt )dt + b(t, Xt )dBt , we deduce that the law of Xt+dt depends
only on Xt and not on the past.

Ex: geometric Brownian motion (model for the equity price in Black-Scholes model):
dXt = Xt (µdt + σdBt ). (Xt ) is an Ito process.
dXt
Note: we do not deduce that d(ln Xt ) = µdt + σdBt indeed d(ln Xt ) 6= Xt :
 Probabilistic Methods in Finance 37

we cannot integrate in a usual way, an additional term should arise.

3. Ito lemma
(Xt )t≤T Ito process: dXt = a(t, Xt )dt + b(t, Xt )dBt .
For a given ω, t 7→ Xt (ω) is a function g. If g was C 1 , we would write, for any C 1 function f :
d[f (g(t))] = f 0 (g(t))dg(t), meaning:
Z t
∀t ≤ T , f (g(t)) = f (g(0)) + f 0 (g(s))g 0 (s)ds.
0
Z t
1
But t 7→ Xt (ω) is not C and f (Xt ) = f (X0 ) + f 0 (Xs )dXs is wrong for an Ito process.
0
• f : R → R C 2 ; the variation of f (Xt ) between t and t + dt is (we expand d[f (Xt )] in a Taylor
series up to second-order term):
1
d[f (Xt )] = f (Xt+dt ) − f (Xt ) = f 0 (Xt )dXt + f 00 (Xt )(dXt )2 + o[(dXt )2 ].
2
The last term contains a term of 1st order (heuristic proof): (dXt )2 = [a(t, Xt )dt + b(t, Xt )dBt ]2
contains:
terms in dtdBt and (dt)2 , negligible compared with 1st order terms (in dt and dBt )
a term in (dBt )2 ; but E I [(dBt )2 ] = dt, since Bt+dt − Bt ∼ N (0, dt),
and V [(dBt )2 ] is in (dt)2 , negligible compared with dt (cf E
I ((dBt )4 ) = 3(dt)2 .
Then we identify (dBt )2 to its expectation, dt.

We get that (dXt )2 is equivalent to [b(t, Xt )]2 dt.

We obtain that the variation of f (Xt ) between t and t + dt is:
1
d[f (Xt )] = f 0 (Xt )dXt + f 00 (Xt )[b(t, Xt )]2 dt
2
The correct meaning is the integrated version:

Th: Ito Formula.

(Xt )t≤T Ito process: dXt = a(t, Xt )dt + b(t, Xt )dBt . f : R → R class C 2 ;
Z t
1 t 00
Z
0
we have: ∀t ≤ T , f (Xt ) = f (X0 ) + f (Xs )dXs + f (Xs )[b(s, Xs )]2 ds P -as.
0 2 0
i.e. there is an additional term.

Notes: 1. the rigorous proof of this result relies on the same steps as for the computation of
RT
0 Bt dBt (which corresponds to a particular case of the lemma).

2. the Ito Formula proves that (f (t, Xt ))t≤T is still an Ito process (sum of terms in dt and in dBt ).

Examples of use: dBt = 0 · dt + 1 · dBt (the Brownian motion is obviously an Ito process), then:
1
? using f (x) = x2 , we get d(Bt )2 = 2Bt dBt + 2dt (see previous page for a rigorous proof).
2
? definition: geometric Brownian motion: stochastic process for which:
∃ µ, σ ∈ R, with σ > 0 such that dSt = St (µdt + σdBt ), with S0 > 0 known.
In the Black Scholes model, the stock price is assumed to be a geometric Brownian motion.
dSt dSt
The instantaneous return satisfies: St = µdt + σdBt then St ∼ ”N (µdt, σ 2 dt)”.
 Probabilistic Methods in Finance 38

dSt
Compare to St = rdt for a risk-free asset (return = r, deterministic): a (Gaussian) randomness is
added.

We compute d[ln St ] (Ito on an open set, since ln not defined on R).

σ2 σ2
Rt Rt
d[ln St ] = dS 1 2 2
St − 2S 2 σ St dt = (µ− 2 )dt+σdBt , then ∀t ≥ 0, ln St = ln S0 + 0 (µ− 2 )ds+ 0 σdBs
t
t
2
(µ− σ2 )t+σBt
hence St = S0 e .

In fact, we saw Ito on R only ⇒ better to start with the solution (and use the unicity of the solution
of the Stochastic Differential Equation dSt = St (µdt + σdBt ), for S0 given):
2 σ2
let Xt = (µ − σ2 )t + σBt and St = S0 eXt . Ito lemma applied to dXt = (µ − 2 )dt + σdBt with
x 7→ S0 ex :
2
dSt = St dXt + σ2 St dt = St (µdt + σdBt ).

Properties of the geometric Brownian motion: continuous process, with positive values,
2
lognormal: each r.v. ln St is gaussian. Indeed ln St = ln S0 + N ((µ − σ2 )t, σ 2 t).

More general Ito Formula:

• f : [0, T ] × R → R C 1,2 (twice differentiable in x, once in t, and continuous derivatives in (t, x)), we
have:
∂f ∂f 1 ∂2f
df (t, Xt ) = (t, Xt )dt + (t, Xt )dXt + (t, Xt )[b(t, Xt )]2 dt.
∂t ∂x 2 ∂x2

II. Black-Scholes model assumptions

= basis model in continuous time.

financial market consisting of 2 basis assets and their derivative products:

- one risk-free asset, whose value grows at the interest rate r > 0 constant: price at time t: ert .
- the other asset is risky (equity paying no dividend), with price St at time t.

Let (Bt )t∈[0,T ] be a Brownian motion,

we assume dSt = St (µdt + σdBt ), µ, σ constants, σ > 0, ie (St ) geometric Brownian motion.
dSt
We have: = µdt + σdBt ∼ ”N (µdt, σ 2 dt) ”:
St
µ is interpreted as the instantaneous expected rate of return, and σ 2 as its instantaneous
variance.
The successive shocks dBt are independent and Gaussian.
σ measures the sensitivity to these shocks, called ”volatility”.
σ2
∀t ≥ 0, St = S0 e(µ− 2
)t+σBt
. Then ln St is normal (St lognormal) and (St ) is a Markov process.

I (St ) = S0 eµt grows like the price of an asset with constant instantaneous return rate µ.
E
S(j+1)δ 2
For a given time length δ (ex 1 day), the log-returns ln Sjδ are i.i.d., with law N ((µ − σ2 )δ, σ 2 δ).

It is then easy to test the validity of the model on real data.

S(j+1)δ −Sjδ −Sjδ
 
S S
If δ small, ln (j+1)δ
Sjδ = ln 1 + Sjδ ' (j+1)δ
Sjδ = successive returns of the equity.
 Probabilistic Methods in Finance 39

Test = are the successive returns i.i.d. Gaussian?

Conclusion: not perfect.

Note about µ: µ − r corresponds to a premium to compensate for the risk of the asset. The
higher σ, the riskier the asset and the higher µ (generally). The higher the level of risk aversion of
the investors, the higher µ will be.

We make the usual assumptions on the financial markets for derivatives pricing:
1. There are no transactions costs and no taxes.
2. Short selling is allowed, with no limit. In particular, with the risk-free asset:
The market participants can borrow and lend money at the same risk-free rate of interest r.
3. The assets are divisible.
Then one can hold α assets, with α ∈ R.
4. Trading takes place continuously in time: the quantities of assets being held can change at any
time.
5. NAO: there are no arbitrage opportunities.

III. Partial differential equation approach

I (h(ST )2 ) finite).
Pricing of a European option with maturity T and payoff h(ST ) at T (with E

1. The option price at time t can be written F (t, St )

The price at time t depends on t, St , and not on Ss , s < t, since the future variations of the U.A.
price are functions of St only (Markov process), denoted by F (t, St ) where F : [0, T ] × R+ → R,
(t, x) 7→ F (t, x).
We assume that F has class C 1,2 (that is satisfied when h is C 2 , with E I [h(ST )2 ] finite; or for
h(x) = (x − K)+ , or (K − x)+ :
for the call and the put with strike K, h is not differentiable at K,
σ2
but F will have class C 1,2 as P (S0 e(µ− 2 )T +σBT
= K) = 0).

2. Partial differential equation satisfied by F

We consider locally a portfolio constituted of
(
−1 option
nt U.A.
where we choose nt s.t. the portfolio is risk-free between t and t + dt (short sale if nt < 0).
Let Vt be the value of the portfolio at time t: Vt = −F (t, St ) + nt St .
Between t and t + dt, the quantity of U.A. remains equal to nt (this is linked to some notion of
self-financing strategy like in discrete time). The variation of the portfolio value between t and
t + dt is: dVt = −dF (t, St ) + nt dSt with (Ito lemma):

∂2F σ2
 
∂F 2 ∂F
dF (t, St ) = (t, St ) + 2
(t, St ) (St ) dt + (t, St )dSt
∂t ∂x 2 ∂x
 Probabilistic Methods in Finance 40

We take nt = ∂F ∂x (t, St ). Then dVt contains terms in dt only and none in dSt . No randomness: the
portfolio is risk-free between t and t+dt; then its return can only be r: dVt = rVt dt (else AO locally):

∂2F σ2
   
∂F 2 ∂F
dVt = − (t, St ) + (t, St ) (St ) dt = r −F (t, St ) + (t, St )St dt
∂t ∂x2 2 ∂x

Note: the portfolio is risk-free for an infinitesimally short period of time. To keep it risk-free
between 0 and T , it is necessary to change continuously the proportions of the derivative security
and the stock in the portfolio.

We get the PDE satisfied by the price function F :

∂F ∂F ∂2F σ2
(t, St ) + rSt (t, St ) + (t, St ) (St )2 = rF (t, St )
∂t ∂x ∂x2 2

Changing S0 , we see that St can take any value of ]0, +∞[, then we have:
∂F ∂F σ2 ∂ 2F
∀x ∈]0, +∞[, ∀t ∈]0, T [, (t, x) + rx (t, x) + x2 2 (t, x) = rF (t, x) (”parabolic equation”)
∂t ∂x 2 ∂x
with: ∀x ∈]0, +∞[, F (T, x) = h(x) (”boundary condition”).

Notes: these two equations are independent of µ.

The equation is the same for any European derivative, only the boundary condition changes.
Any derivative price satisfies this PDE (precisely: any asset with a price FT -measurable ie whose
randomness comes from a same risk factor, the BM driving the risky asset price S; in particular
any derivative with U.A. = the risky asset).

Ex: forward contract, maturity T , future price K.

value at t = St − Ke−r(T −t) . We check that F (t, x) = x − Ke−r(T −t) satisfies the equation.

Resolution of the equation: seen later (some numerical methods can be used anyway. Ex: finite
difference method = discretisation of derivative functions).

3. Hedging
∂F
Let ∆t = (t, St ). The portfolio {−1 option, ∆t U.A.} is locally risk-free.
∂x
Then the portfolio {1 option,−∆t U.A.} is equivalent to cash (its annualised return between t and
t + dt is r).
(
∆t U.A.
That means also that we can build a portfolio replicating the option,
+ some risk free asset
i.e. a portfolio consisting of the basis assets that is worth the same than the option at any time.

The seller of the option will build this portfolio, called hedging portfolio (in complement of the
short option position).
At T , this portfolio provides exactly h(ST ), i.e. what is due to the option buyer.

The seller makes money on the margin (he sells at F0 +margin) and eliminates completely its risk:
 Probabilistic Methods in Finance 41

there is a perfect replication of the final payoff.

Indeed the hedging portfolio is continuously adjusted (dynamic hedge), i.e. continuous trading of
the U.A., in order to hold the correct quantity of U.A. (∆t = ∂F ∂x (t, St ), the ”delta” of the option)
at any time.
The seller is then indifferent to the fact that the U.A. price goes up (even if he sold a call option)
or down (even if he sold a put option).

Note that the delta measures the sensitivity of the option price to the variations of the U.A. price.
To a change of 1 euro in the price at t, corresponds a change of ∆t euro in the option price.

In the theoretical model, the replication is perfect.

In reality, the hedge is done discretely:
- the physical transactions are necessarily discrete,
- existence of transaction costs that limit the frequency of re-hedging.
The hedging in then not perfect and the seller’s portfolio is in fact risky. The more frequently he
re-hedges, the closer to the option is his portfolio, but the more transaction costs he pays.
The hedging transactions can be less frequent when the delta changes slowly.
2
This is measured by the gamma: Γt = ∂∂xF2 (t, St ) measures the sensitivity of the delta to the varia-
tions of the U.A. price, then it measures the residual risk of the hedged position.

IV. Probabilistic approach for European options. Black Scholes formula

1. Price as an expectation
Pricing of a European option with maturity T and payoff h(ST ) at T .
This other approach provides the solution of the PDE, by a probabilistic method.
The coefficients of the PDE are independent of µ (linked to the risk aversion of investors), then the
price dynamics is independent of µ.
Option pricing can then be done as if investors were risk-neutral.
=”set yourself in the risk-neutral universe”, where any return or discount rate is equal to r. In this
world, investors do not demand extra returns above the risk-free interest rate for bearing risks.

In fact, this corresponds again to a change of probability:

Girsanov theorem (particular case)
λ2
For a given λ ∈ R, let Lt = e−λBt − 2 t for any t ≥ 0. We denote by P ∗ the probability
with density LT with respect to P . Then (Bt + λt)t∈[0,T ] is a Brownian motion under P ∗ .

dP ∗ λ2
(Bt + λt)t∈[0,T ] Brownian motion under P ∗ defined by the density: = e−λBT − 2 T
dP

I (LT ) = 1 then P ∗ is a probability.

Proof (see also tutorial, exercise 21): E
λ2
Let Wt = Bt + λt and Lt = e−λBt − 2
t
. We know that (Lt )t≥0 is a martingale under P .
(Wt )t≤T is an adapted process, with continuous paths, and W0 = 0.

We will prove that for s ≤ t, Wt − Ws ∼ N (0, t − s) and is independent of Fs = σ(Bu , u ≤ s) =

σ(Wu , u ≤ s) under P ∗ .
 Probabilistic Methods in Finance 42

u2
It is sufficient to prove that ∀u ∈ R, EI ∗ (eiu(Wt −Ws ) |Fs ) = e− 2
(t−s)
(*).

u2
I ∗ (eiu(Wt −Ws ) ) = e−
Indeed, that implies that ∀u ∈ R, E 2
(t−s)
,
then Wt − Ws and N (0, t − s) have same characteristic function, hence Wt − Ws ∼ N (0, t − s).
And we will use below lemma to conclude.
I (eiuX |B) = E
If a r.v. X satisfies: ∀u ∈ R, E I (eiuX ) P a.s., then X is independent of B.

I (eiuX P1I(B)
Proof: we have: ∀B ∈ B, E B
I (eiuX ).
)=E
Then X has same law under P than under the probability with density P1I(B) B
with respect to P
(the characteristic functions are the same). Then for any f : R → R measurable and bounded,
I (f (X) 1IB ) = E
E P (B) I (f (X)) i.e. E
I (f (X)1I ) = E
B I (f (X))E
I (1I ), which proves the independence.
B

To prove (*), we need to know how to compute a conditional expectation under P ∗ .

I ∗ (Z) = E
If Z ∈ L1 (Ω, F, P ∗ ) is Ft -meas., then E I (ZLT ) = E I (LT |Ft )) = E
I (ZE I (ZLt ).

I ∗ (Z|Fs ) for s ≤ t.
Now we want to compute V = E
I ∗ (Y Z) = E
V is Fs -measurable and satisfies: for any Y Fs -measurable in L1 , E I ∗ (Y V ).

Y Z and Y V are respectively Ft -measurable and Fs -measurable. Hence we need:

E I (Y V Ls ) for any Y Fs -measurable, i.e.: V Ls = E
I (Y ZLt ) = E I (Lt Z|Fs ).
1
I ∗ (Z|Fs ) =
We get: for Z Ft -meas, and s ≤ t, E I (ZLt |Fs )
E
Ls

Now let us prove (*): we take s ≤ t and u ∈ R.

1
I ∗ (eiu(Wt −Ws ) |Fs ) =
E I (Lt eiu(Wt −Ws ) |Fs )
E
Ls
λ2
Lt eiu(Wt −Ws ) = e−λBt − 2
t+iu[Bt −Bs +λ(t−s)]
. We group the terms in Bt :
(−λ+iu)2
(−λ+iu)Bt − t
let Mt = e 2 . We know that (Mt ) is a martingale under P .

1 u2
I ∗ (eiu(Wt −Ws ) |Fs ) =
We have E I (Mt |e−iuBs −iuλs−
E 2
t
} |Fs )
Ls {z
Fs −meas
1 −iuBs −iuλs− u2 t 2 2 (−λ+iu)2 u2
λBs + λ2 s −iuBs −iuλs− u2 t (−λ+iu)Bs −
= e 2 M = e
s e e 2
s
= e− 2 (t−s) . 
Ls

Then the theorem is proved: we get that (Wt )t∈[0,T ] is a Brownian motion under P ∗ , the probability
λ2
with density LT = e−λBT − 2
T
w.r.t. P .

We rewrite the dynamics of the U.A. price in order to have an instantaneous return equal to r
instead of µ:
dSt = St (µdt + σdBt ) = St [rdt + σ(dBt + µ−r
σ dt)].
µ−r
Let λ = σ and for t ≥ 0, Wt = Bt + λt.
λ2
(Wt )t∈[0,T ] is a Brownian motion under P ∗ , the probability with density e−λBT − 2
T
w.r.t. P .

We have dSt = St (rdt + σdWt ), with Wt = Bt + λt.

σ2
In particular, for t ≥ 0, St = S0 e(r− 2
)t+σWt
, then (St e−rt )t≤T is a martingale under P ∗
 Probabilistic Methods in Finance 43

i.e. the discounted asset prices are martingales under P ∗ (making P ∗ an E.M.M., with a similar
definition as in discrete time).
Under P ∗ , we have:
I ∗ (ST |Ft ) = value of the future flow, discounted at the risk-free rate r.
St = e−r(T −t) E
σ2
to be compared to St = e−µ(T −t) EI (ST |Ft ) (using St = S0 e(µ− 2 )t+σBt ):
under the real probability, the price is discounted at rate µ,
rate taking into account the asset risk, and the risk aversion of the investors.

Th: Price at time t of the asset paying h(ST ) at T : I ∗ (h(ST )|Ft ).

Ft = e−r(T −t) E
I ∗ (h(ST )). We get:
therefore (Ft e−rt )t≤T is a martingale under P ∗ and F0 = e−rT E
σ2
F0 = e−rT E
I ∗ [h(S0 e(r− 2
)T +σWT
)]

Under P ∗ , the future cash-flows are discounted at rate r, as if the investors were ”risk neutral”.
Thus the name ”risk neutral probability”.
By contrast, P is the ”real probability, or historical or objective”.
Not that the complete proof involves, like in discrete time, the fact that the market is complete:
any option is replicable by a self-financing strategy. This argument is used here when establishing
the PDE.

Computation easy in the standard cases: the law of h(ST ) is known (ST lognormal).
Note: there exists some numerical methods to compute an expectation = Monte-Carlo method
(simulation of n r.v. with law = law of X and approximation of E
I (X) by the mean of the n r.v.).

2. Computation for a call (the put option price is then obtained from put-call parity)
Price at time t: Ct = C(t, St ). We compute C0 .

I ∗ [e−rT (ST − K)+ ] = E

C0 = E I ∗ [ST e−rT 1I{ST ≥K} ] − Ke−rT P ∗ (ST ≥ K)
σ2
σ2
ST ≥ K ⇔ S0 e(r− 2
)T +σWT
≥ K ⇔ σWT ≥ ln SK0 − (r − 2 )T

2
ln S0 + (r − σ2 )T

 d2 = K √


then P ∗ (ST ≥ K) = P ∗ (− W
√ T ≤ d2 ) = N (d2 ) where σ ZT
T d
1 x2
 N (d) = √ e− 2 dx


2π −∞
N is the cumulative probability distribution function for a standardized normal variable.
σ2 ? √ √
I ∗ [ST e−rT 1I{ST ≥K} ] = E
E I ∗ [S0 eσWT − 2 T 1I{WT ≥−d2 √T } ] = S0 P̃ [B̃T ≥ −d2 T −σT ] = S0 N (d2 + σ T )
| {z }
d1
? from Girsanov:
σ2
(B̃t = Wt − σt)t∈[0,T ] is a Brownian motion under P̃ , probability with density eσWT − 2 T w.r.t. P ∗ .

then C0 = S0 N (d1 ) − Ke−rT N (d2 )

and, replacing S0 by St and T by T − t, the price at time t ≤ T is: (with di = di (0, S0 ))
 Probabilistic Methods in Finance 44

Ct = St N (d1 (t, St )) − Ke−r(T −t) N (d2 (t, St )) , where

 2
x
 ln K + (r + σ2 )(T − t)
 d1 (t, x) = √ Shape :


σ T −t
x 2
 ln K + (r − σ2 )(T − t)
d (t, x) = √

 2

σ T −t

Hedging: we compute the delta (slope of the previous curve).

σ2
I ∗ [e−rT (S0 XT − K)+ ] with XT =
C0 = E ST
S0 = e(r− 2
)T +σWT
.
| {z }
f (S0 ,ω)
We compute the first derivative w.r.t. S0 . We can differentiate inside the expectation if:
σ2
(1) P -as in ω, S0 7→ f (S0 , ω) is C 1 , fine since {ω|xe(r− 2 )T +σWT = K} is negligible.
(2) For S0 in a compact set, |f 0 (S0 , ω)| ≤ g(ω) with g ∈ L1 .
derivative of S0 7→ f (S0 , ω): XT 1IR+ (S0 XT − K) ≤ XT , then (2) is satisfied.

I ∗ [e−rT XT 1IR+ (S0 XT − K)] = E

Then ∆0 = E I ∗ [e−rT SST0 1IST ≥K ].

I ∗ [ST e−rT 1I{ST ≥K} ] = S0 N (d1 ). Therefore ∆0 = N (d1 ).

We proved E
Z d1 (t,St )
1 x2
And ∆t = N (d1 (t, St )) = √ e− 2 dx ∈ [0, 1].
2π −∞

3. PDE when the U.A. pays dividends (exercise 25)

Stock paying a continuous dividend yield at rate δ (known).
(
−1 option ∂F
We consider locally a portfolio consisting of , with ∆t = (t, St ).
∆t U.A. ∂x
Let Vt the portfolio value at date t. We have, between t and t + dt:
dVt = −dF (t, St ) + ∆t dSt + ∆t δdSt
(dividend received if ∆t < 0, paid in the short sale if ∆t > 0).

The equation is then changed in:

∂F ∂F σ2 ∂ 2F
(t, x) + (r − δ)x (t, x) + x2 2 (t, x) = rF (t, x)
∂t ∂x 2 ∂x
4. Use
All the parameters are observed on the market, except the volatility σ.
2 methods to get it:
S(j+1)∆t
Historical volatility: under Black-Scholes assumptions, the Xj = ln Sj∆t , 0 ≤ j ≤ n − 1 are i.i.d.
σ2 2
with law N ((µ − 2 )∆t, σ ∆t).
1 Pn
This is used to estimate σ on past series: n−1 j=1 (Xj − X̄)2 is an unbiased estimator of σ 2 ∆t.

Implicit volatility: the call price is an increasing function of the volatility.

Observed (quoted) call prices are used to infer the volatility, by reverting the Black-Scholes formula.
 Probabilistic Methods in Finance 45

= volatility as anticipated by the market.

Note that it has been assumed that r is constant. In practice the formulas are used with r equal
to the risk-free rate of interest on an investment lasting for T − t, the life of the option.

5. Choice of u and d in the binomial tree

They need to depend on the volatility of the U.A..

We build the tree directly under the risk-neutral probability. √

√ √ er∆t − e−σ ∆t
With u = eσ ∆t , d = e−σ ∆t , we have d < er∆t < u and p∗ = √ √ .
√ √ eσ ∆t − e−σ ∆t
We approximate p∗ by p∗ = √ ∆t
r∆t+σ
2σ ∆t
= 1
2 + r ∆t
2σ .
t S −S S
For ∆t small, the return on a period, t+∆t St , can be approximated by ln t+∆t
St (logarithmic re-
turn), from ln(1 + x) ∼ x.
S √ √
ln t+∆t
St is worth σ ∆t or −σ ∆t. Then the expected return on a period is:
√ √ √ √ √
p∗ σ ∆t − (1 − p∗ )σ ∆t = σ ∆t(1 − 2p∗ ) = σ ∆t r σ∆t = r∆t
√
while the expected variance is: p∗ (1 − p∗ )(2σ ∆t)2 ∼ ( 12 )2 4σ 2 ∆t = σ 2 ∆t.

converges to BS model when ∆t → 0

(ie the discrete process converges to a geometric Brownian motion with volatility σ).

ln S moves are sums of i.i.d. Bernouilli r.v. converging to a Gaussian variable on any inter-
S
val, while in the BS model, under the risk-neutral probability, the ln (j+1)∆t
Sj∆t are i.i.d. with law
σ2 2
N ((r − 2 )∆t, σ ∆t).