Introduction Markov chains Application in insurance Application in nance

Stochastic Processes

Olivier Scaillet

University of Geneva and Swiss Finance Institute

Introduction Markov chains Application in insurance Application in nance

Outline

1 Introduction

2 Markov chains

3 Application in insurance

4 Application in nance
Introduction Markov chains Application in insurance Application in nance

Stochastic process

Stochastic processes are used to model the state of a time

dependent random system.

We work with a probability space (, A, P ) and a function

(t , ) X (t , ) of R+ in a space E giving the system state.

For xed, that is, a particular evolution of the system, the

successive states are given by the function t X (t , ) that we call
trajectory or path by analogy with a system modeling the position
of a particle.
Introduction Markov chains Application in insurance Application in nance

Denition: stochastic process

A stochastic process X = {Xt , t T } is a familly of rv dened on

the same probability space and indexed by the time t taking its
values in a set T of indices.

If T = {0, 1, 2, 3, . . . } or T = {. . . , 2, 1, 0, 1, 2, 3, . . . }, the
process is said to be a discrete-time process.
If the set T = R+ or T = R, it is said to be a continuous-time
process.
Introduction Markov chains Application in insurance Application in nance

Example: simple random walk

Let X1 , X2 , . . . be a family of independent rv, each taking value 1

with probability p and value 1 with probability 1 p .

Let the sum process St = S0 + ti=1 Xi . The series

P
S = {St , t N} is a random walk starting at S0 .
Say that St stands for the wealth of a gambler after t rounds,
earning $1 if she wins and paying $1 if she loses, and starting with
initial wealth $S0 .

We can consider the case P [St 0] (ruin probability), or

min t T : St 0 (stopping time), for instance.
Introduction Markov chains Application in insurance Application in nance

Example: simple random walk (cont'd)

We can rewrite the random walk by St = St 1 + Xt . The particle

being at the position St 1 at time t 1 can either go left 1 unit
with probability p or go right 1 unit with probability 1 p .

Properties:
1 Space homogeneity:
P [St = j |S0 = a] = P [St = j + b|S0 = a + b]
2 Time homogeneity: P [St = j |S0 = a] = P [St +h = j |Sh = a]
3 Markov property:
P [St +1 = j |S0 = s0 , . . . , St = st ] = P [St +1 = j |St = st ], t
(conditionally on present, future does not depend on past)

The process is said to be memory-less. The random walk is a

particular case of Markov chains.
Introduction Markov chains Application in insurance Application in nance

Denition: Markov chain

We consider a discrete-time process X taking values in a countable

space called state space.
The process X is a Markov chain if it satises the Markov condition
P [Xt = x |X0 = x0 , X1 = x1 , . . . , Xt 1 = xt 1 ] = P [Xt = x |Xt 1 = xt 1 ]
for any t .
This Markov property is equivalent to
P [Xt +1 = x |Xt1 = xt1 , Xt2 = xt2 , . . . , Xtk = xtk ] = P [Xt +1 = x |Xtk = xtk ]
with 0 t1 tk t , or also
P [Xt +h = x |X0 = x0 , X1 = x1 , . . . , Xt = xt ] = P [Xt +h = x |Xt = xt ]
for any t , h.
Introduction Markov chains Application in insurance Application in nance

Denition: Markov chain (cont'd)

We can generalize this concept to order r Markov chains:

P [Xt = x |X0 = x0 , X1 = x1 , . . . , Xt 2 = xt 2 , Xt 1 = xt 1 ]
= P [Xt = x |Xt 1 = xt 1 , Xt 2 = xt 2 , . . . , Xt r = xt r ]

As the state space is countable, we can map it to the set of

integers. We say that when Xt = i , the chain is in state i at time t ,
or also visits i or takes value i .

The chain evolution depends on the probabilities

P [Xt = j |Xt 1 = i ] called transition probabilities. In general, they
depend on t , i , and j . However, we often consider a simple case
where they depend on i and j , but are independent on t .
Introduction Markov chains Application in insurance Application in nance

Homogeneity

We consider a simple case where transition probabilities depend on

i and j , but are independent on t . A Markov chain X is said
homogenous if P [Xt +1 = j |Xt = i ] = P [X1 = j |X0 = i ], t , i , j . It
is a stability condition on the probabilities in time.

The matrix P = (pij ) with pij = P [Xt +1 = j |Xt = i ] is the

transition matrix of the chain. The transition matrix P is a
stochastic matrix as it satises:
1 all entries of P
are non negative: pij 0, i , j .
2 for each row of P , the sum of the entries gives 1: j pij = 1,
P
i .
Introduction Markov chains Application in insurance Application in nance

Example: simple random walk

We move right 1 unit (go from i to i + 1) with probability p and

left 1 unit (go from i to i 1) with probability 1 p .

The state space is {0, 1, 2, . . . } and the transition probabilities

p
if j = i + 1,
are pij = P [St = j |St 1 = i ] = 1 p if j = i 1, .
0 else .


Introduction Markov chains Application in insurance Application in nance

Transitions

We consider the evolution of the chain in h steps (horizon h). The

h steps transition matrix P (t , t + h) is the matrix whose entries
give the transition probabilities in h steps:
pij (t , t + h) = P [Xt +h = j |Xt = i ].

The homogeneity hypothesis is equivalent to the condition

P (t , t + 1) = P (the matrix is not time dependent).

Using Chapman-Kolmogorov equations, we get:

pij (t , t + h + l ) = k pik (t , t + h)pkj (t + h, t + h + l ), or using
P
matrix notation, P (t , t + h + l ) = P (t , t + h)P (t + h, t + h + l ).
Introduction Markov chains Application in insurance Application in nance

Transitions (cont'd)

In the homogenous case, we have P (t , t + 2) = P (t , t + 1 + 1) =

P (t , t + 1)P (t + 1, t + 1 + 1) = PP = P 2 . By recurrence
P (t , t + h) = P h and P (t , t + h + l ) = P h P l = P h+l .

In particular, P (t , t + h) = P (0, 0 + h) = P h .

The evolution of the chain depends on the initial state at time 0

(t )
and the transition matrix P . Let i = P [Xt = i ] and the
(t )
corresponding row vector (t ) = (i ).

We can show (t ) = (0) P t , that is, the random evolution of the

chain characterized by the transition matrix P and the initial
probability mass function of (0) .
Introduction Markov chains Application in insurance Application in nance

State classication

Let us consider the chain evolution as a particule movement

between multiple states. We consider the time (possibly innite)
that the particle needs to come back to its starting point.

A state i is said recurrent (or persistent) if

P [Xt = i |X0 = i ] = pii (t ) = 1 for some t 1.

A state i is said transitory (or transient) if this probability is strictly

smaller than 1.
Introduction Markov chains Application in insurance Application in nance

First passage time

We can also consider the probabilities of rst passage

fij (t ) = P [X1 6= j , X2 6= j , . . . , Xt 1 6= j , Xt = j |X0 = i ], that is, the
probability that the rst visit of state j needs exactly t steps
starting at i .

Let us dene the probability that the chainPvisits state j at some

t =1 fij (t ).
point in time, starting at state i , by fij =

A state j is recurrent if and only if fjj = 1.

Let Tj = min {t 1 : Xt = j } be the time of rst passage of state

j , with Tj = if the state is never visited.

The state j is transitory if and only if P [Ti = |X0 = i ] > 0.

Introduction Markov chains Application in insurance Application in nance

Mean time of recurrence

We also compute the mean time of recurrence mi of state i dened

t tfii (t ) if i recurrent ,
(P
by mi = E [Ti |X0 = i ] =
if i transitory .
mi can nonetheless be innite for a recurrent state i and we say
that:
i is null recurrent if mi =
i is non-null recurrent (or positive recurrent) if mi <
Introduction Markov chains Application in insurance Application in nance

Periodicity

The period d (i ) of a state i is dened as the largest common

divider of all values t such that pii (t ) > 0.

The state is said to be

periodic if d (i ) > 1
aperiodic if d (i ) = 1
We thus have that pii (t ) = 0 unless t is a multiple of d (i ).
A state is said to be ergodic if it is recurrent, non-null, and
aperiodic.
Introduction Markov chains Application in insurance Application in nance

Chain classication

We dene the class of a chain by looking at the links of its states:

State j is accessible from state i if there exists some t such
that pij (t ) > 0
State i communicates with state j , or i j , if j is accessible
from i
States i and j intercommunicate, or i j , if i communicates
with j and j communicates with i .
If i j , then
1 i and j have the same period
2 i is transitory if and only if j is transitory
3 i is null recurrent if and only if j is null recurrent.
Introduction Markov chains Application in insurance Application in nance

Chain classication (cont'd)

A set C of states is said

closed if pij = 0 for all i C , j /C
irreducible if i j for all i , j C

If the chain reaches a state that belongs to a closed set, it never

leaves this set.
A closed set containing exactly one element (singleton) is said to
be absorbing.
Introduction Markov chains Application in insurance Application in nance

Stationary distribution

What is the behavior of a Markov chain when we wait a large

number of periods? We are interested in the distribution of Xt when
t is large. The existence of a limit distribution for Xt as t is
strongly linked to the existence of a stationary distribution.

The vector is called stationary distribution of the chain if has

entries (j ) such that
1 j 0 for all j , and
j j = 1
P

2 = P , that is, j =
i i pij for all j
P
Introduction Markov chains Application in insurance Application in nance

Stationary distribution (cont'd)

By iterating, we get that = P t , for any t . Additionally, we have

seen that homogenous chains have the property(t ) = (0) P t with
(t ) (t )
(t ) = (i ) dened by i = P [Xt = i ].

Therefore, if property (0) = , that is, X0 follows the distribution

given by , then Xt also follows distribution .
The distribution of Xt is stationary as time goes, which implies that
is also the limit distribution of Xt as t .
Introduction Markov chains Application in insurance Application in nance

Stationary distribution (cont'd)

An irreducible chain has a stationary distribution if and only if all

of its states are non-null recurrents.

In this case, is the unique stationary distribution and is given by

i = mi1 with mi the mean time of recurrence of i .
For irreducible and aperiodic chains, we have that pij (t ) mj1 ,
i , j as t .
Introduction Markov chains Application in insurance Application in nance

Application in insurance, reinsurance chain

Markov chains have applications in insurance:

bonus-malus system
reassuring chain
Introduction Markov chains Application in insurance Application in nance

Bonus-malus system

Use of Markov chains to model a problem of car insurance:

bonus-malus system.

We set a nite number L of classes that dene the premium paid by

the insured.
Each year, the class of the insured is updated depending on the
class of the previous year and the number of sinisters (or claims)
during the year.
If no sinister is recorded, the insured is moved to a class whose
premium is lower; conversely, if sinisters have been reported, she is
moved to a class whose premium is higher.
Introduction Markov chains Application in insurance Application in nance

Model: bonus-malus system

The parameters of the model are the following:

L classes indexed by 1 to L. Classe 1 is called superbonus and
class L is called supermalus. The annual premium depends on
the number of insured people in each class and is computed
using a fee scale.
a fee (or premium) scale b = (b1 , . . . , bL ) with
b1 b2 bL
a transition rule that denes the transfer from one class to the
other as a function of the number f sinisters. If k sinisters are
reported tij (k ) = 1 if the class changes from i to j, and
tij (k ) = 0 otherwise
i0 for any new customer.
an initial class
Let T (k ) = (tij (k )), each T (k ) is a 0 1 matrix for which each
row has exactly one1.
Introduction Markov chains Application in insurance Application in nance

Model: bonus-malus system (cont'd)

Suppose that the number of sinisters per year forms a

sequenceY1 , Y2 , . . . of i.i.d. rv and with mass function
P [Y = k ] = qk .
We denote by X1 , X2 , . . . the successive classes through which the
insured has gone, and suppose that the class of a particular year
only depends on the class of the previous year and the number of
sinisters during the year. We can descrive the evolution of {Xt }
using the recurrence equation Xt = (Xt 1 , Yt ), with (i , k ) = j if
and only if tij (k ) = 1.
Introduction Markov chains Application in insurance Application in nance

Model: bonus-malus system (cont'd)

This evolution corresponds toPa Markov chain whose probabilities of

transition are given by pij = k =1 qk tij (k ).

In practice, we suppose that the number of sinisters follows a

Poisson law with parameter , which leads to
k
pij = k =1 k ! e tij (k ).
P
Introduction Markov chains Application in insurance Application in nance

Model: bonus-malus system (cont'd)

In a bonus-malus system, the fee bi is applied every time an insured

visits class i . He may then consider
the cumulative sum of the premia paid until t : Rt = tk=10 bXk
P

the discounted cumulative sum of the premia paid until t :

Rta = tk=10 (1b+Xrk)k , with interest rate r
P

the discounted cumulative sum of the premia paid with an

bxk
innite horizon: R
a = P
k =0 (1+r )k .
Introduction Markov chains Application in insurance Application in nance

Model: bonus-malus system (cont'd)

We may show that when the chain has an initial class X0 = i0 , we

have
E [Rt ] = ei0 tk=10 P k b0
P

E [Rta ] = ei0 (Id (1 + r )1 P )1 (Id (1 + r )n P n )b0

E [Ra ] = e (Id (1 + r )1 P )1 b 0 ,
i0
where ei0 denotes the row vector with zeros everywhere except on
the i0 th entry which which is equal to 1.
We may also show that if the chain has a stationary distribution ,
E [Rta ] is a linear combination of the stationary fee, i.e. b = b0 .
Introduction Markov chains Application in insurance Application in nance

Chain of reinsurance

Suppose that an asset such as a plane needs insurance. Its value is

too high for a unique company to bear the risk.
We start at time 0 with an insurer of rst rank. Then, the insurer
negotiates with second rank insurers in order to sell them part of
the original insurance. Each on the second rank companies does the
same and negotiates part of its own insurance policy share with a
third rank insurer. This sequence corresponds to a (simplied)
chain of reinsurance.
Introduction Markov chains Application in insurance Application in nance

Model: chain of reinsurance

Suppose that each company needs a certain time interval (e.g.

administrative tasks) with distribution function F before signing
simultaneously all its contrats of reinsurance.
The number of reinsurance companies is assumed to be randomly
distributed with some probability mass function.
Each company initiates its chain of searches, independently on the
history of other companies. All administrative durations are i.i.d.
following distribution F .
Introduction Markov chains Application in insurance Application in nance

Model: chain of reinsurance (cont'd)

The sequence {Xt } counting the number of companies in each

node of the chain at time t (t th rank) is called branching process of
Galton-Watson-Bienaym.
Let Rt ,i be the number of companies linked to company i in node t
and suppose that all Rt ,i are i.i.d. The process {Xt } dened by
X1 = 1, Xt +1 = Xi =t 1 Rt ,i is a Markov chain.
P

We may show that the average number of companies at link t = n ,

that is, E [Xn ], is equal to E [Xn ] = mn where m is the average
number of companies contacted at any link and by any company.
Introduction Markov chains Application in insurance Application in nance

Application in nance

Markov chains have multiple applications in nance, e.g.:

credit ratings
binomial model for derivatives pricing
Introduction Markov chains Application in insurance Application in nance

Change of credit rating

A change in the credit rating of an issuer relates to the increase or

decrease of the probability of default. It has an impact on the
assets (debts) it might issue.
We may go bankrupt, or have the credit quality increase or
decrease. It is important to correctly estimate these changes in a
context of quantitative risk management with risky assets.
If returns follow approximately a normal distribution, mean and
variance are sucient to identify risk-return arbitrages. Returns are
no longer normally distributed when there is a risk of default, as a
default event might wipe out the entire value of the assets.
Introduction Markov chains Application in insurance Application in nance

Change of credit rating (cont'd)

The distribution is skewed because of the increased risk of large loss

and has more extreme events than the normal distribution
(kurtosis).
The standard deviation (or variance) is a symmetrical risk measure
and is not adapted to this context. We prefer using quantiles of the
portfolio return distribution (or the value distribution).
We estimate quantiles using simulations of the assets values across
time, taking into account market conditions (interest rates,
exchange rates, ...) and also changes in the credit risk (default,
credit rating change)
Introduction Markov chains Application in insurance Application in nance

Model: change of credit rating

We use a Markov chain whose states correspond to credit ratings.

We model the probabilities of transition between ratings, as well as

the default probability, e.g. probability of going from rating AAA to
rating AAB, or probability to go from rating BBB to bankruptcy.
These parameters are estimated using historical data on the change
of ratings and defaults of rms.
Introduction Markov chains Application in insurance Application in nance

Model: change of credit rating(cont'd)

We then simulate:
1 the evolution of the Markov Chain to obtain the change of
ratings of the portfolio assets
2 the evolution of the nancial variables in order to keep track of
the changes in nancial markets
We build a histogram of the portfolio PnL with a maturity of 1
year, 2 years, ... We compute the associated portfolio quantiles.
We dene the economic capital as the dierence between an
extreme loss quantile (99.9%) and the average loss. This value is a
buer used to absorb the impact of large unexpected losses
(defaults).
Introduction Markov chains Application in insurance Application in nance

Options valuation
An option is a nancial asset giving the right (but not the
obligation) to its holder to buy or sell some predined amount of
another asset, at some particular date and price.
Multiple factors dene an option:
option kind: call (option to buy), put (option to sell)
underlying asset (stock, debt, interest rate, exchange rate, ...)
whose value at time t is denoted by St )
quantity
maturity: expiration date (denoted by T)
exercise price: price at which the exercise takes place (denoted
by K )
kind of exercise: European (at maturity), American (until
maturity)
premium: price of the option
Introduction Markov chains Application in insurance Application in nance

Payo of a European option

Consider a call option:

if ST > K , we exercise the option and the prot is ST K as
we pay a price K for an asset that has value ST
if ST K , we do not exercise the option and there is no
prot, as we do not buy the underlying at a price higher than
its market price.
Thus, the payo (or value of the call option) at maturity is
max(ST K , 0) = (ST K )+ . Symmetrically, the payo of a put
option at maturity is given by (K ST )+ .
Introduction Markov chains Application in insurance Application in nance

Pricing and hedging

The option seller faces two problems:

1 a pricing problem: determine the value today of the future
random payo
2 hedging problem: be sure to be able to honor his part of the
contract at maturity
The basic assumption in derivatives pricing is absence of arbitrage
(no free lunch), that is, we cannot make a prot (above riskfree
rate) without investing an initial capital and taking risks.
Introduction Markov chains Application in insurance Application in nance

Put-call parity

We can deduce the put-call parity relationship from the absence of

arbitrage assumption. It takes the form Ct Pt = St Ke r (T t ) ,
i.e., the dierence between the price of a call and a put option
(with same strike price and maturity) is equal to the dierence of
the underlying price and the discounted strike price.
We may also deduce price boundaries:
St > Ct > max(St Ke r (T t ) )
Ke r (T t ) > Pt > max(Ke r (T t ) St , 0)
These results are only due to the absence of arbitrage hypothesis,
in particular there is no assumption on the stochastic process
followed the the assets.
Introduction Markov chains Application in insurance Application in nance

European call pricing

Unfortunately, this assumption is not sucient to compute a unique

price, and boundaries are large.
We need to model the evolution of the underlying asset. The most
famous model is Black-Scholes model (1973). It is a
continuous-time model which can be dened as the limit of a
discrete-time binomial model as the time interval between price
changes goes to zero.
Introduction Markov chains Application in insurance Application in nance

One-period binomial model

One period (two dates, t et t + 1),

Two states: high (St St +1 = uSt ) and low
(St St +1 = dSt ), where d < u are multiplicative factors,
Two nancial assets: a stock (St ), and a risk free asset
(1 1 + r in any state).

We show that the probability of the two states does not have any
impact on the option price.
Introduction Markov chains Application in insurance Application in nance

One-period binomial model (cont'd)

Hypothesis: dSt < K < uSt . Indeed, if K < dSt < uSt , the buyer
always exercises his option, thus nobody would like to sell such a
contract (and inversely for K > dSt > uSt ).
We want to determine the call price at time t .The payo function
at t + 1 is max(St +1 K , 0), that is
max(uSt K , 0) = uSt K in the up state
max(dSt K , 0) = 0 in the down state
Introduction Markov chains Application in insurance Application in nance

Pricing by replication

In order to evaluate the option price, we build a portfolio with the

underlying risky asset and the risk free asset that replicates the
option payo.
If there is no arbitrage between the option and such a portfolio,
their prices should be the same as they both have the same payo
at maturity in any state of the world.
If they do not have the same price, it means that there is an
arbitrage opportunity and it is protable to sell the expensive asset
and buy the cheap one. An immediate and risk free prot is made
as the payos of this strategy oset each other in any state of the
world.
Introduction Markov chains Application in insurance Application in nance

Pricing by replication (cont'd)

We denote by 1 the amount of stocks, 2 the amount of risk free

asset.
At time t , the portfolio value is 1 St + 2
At time t + 1, the portfolio value is1 St +1 + 2 (1 + r ), that
is, 1 uSt + 2 (1 + r ) in the up state, and 1 dSt + 2 (1 + r )
in the down state.
We determine the value of 1 and 2 such that the portfolio
replicates the option payo:
1 uSt + 2 (1 + r ) = uSt K
1 dSt + 2 (1 + r ) = 0.
Introduction Markov chains Application in insurance Application in nance

Pricing by replication (cont'd)

We obtain a system of equations with two equations and two

uSt K and
unknowns whose solution is 1 = uSt dSt
2 = (uS(uS t K )dSt .
t dSt )(1+r )

The current option price must be equal to the portfolio price whose
weights are given by 1 ,2 in order to avoid arbitrages. The option
price is therefore given by
Ct = 1 St + 2 = uSuSttdSKt St (uS(uSt tdS
K )dSt
t )(1+r )
.
Introduction Markov chains Application in insurance Application in nance

Pricing with risk-neutral expectation

r )St dSt (1+r )d

Dening = (1+ uSt dSt = u d , we may rewrite the option
pricing formula as Ct = 1+ 1
r [(uSt K ) + (1 )0].
This expression corresponds to the expectation of a binomial law
taking value uS1t+rK with probability and value 1+
0
r with
probability 1 . The option price is thus an expectation of the
discounted future payos, where the probabilities have been shifted.
We call risk-neutral probability of the up state, in contrast to the
true or historical probability p . The probability p is not involved in
the pricing equation and is therefore irrelevant in this context!
Introduction Markov chains Application in insurance Application in nance

Risk-neutral probability

Let (p , 1 p ) be the historical probabilities of an up and down

move of the stock. As there are only two possible states, St +1
follows a binomial law that takes value uSt with probability p and
value dSt with probability 1 p .
The expectation is E p [St +1 ] = puSt + (1 p )dSt and the expected
p
return is given by E [SStt +1 ] = pu + (1 p )d .
By adding 0 = (1 + r ) [p (1 + r ) + (1 p )(1 + r )] to this
expression, one gets:

E p [St +1 ]
= (1 + r ) [p (1 + r ) + (1 p )(1 + r )] + pu + (1 p )d
St
= (1 + r ) + [u (1 + r )]p + [d (1 + r )p ](1 p )
Introduction Markov chains Application in insurance Application in nance

Risk-neutral probability (cont'd)

Therefore, the factor

E p [ St + 1 ]
St ( 1 + r ) = [u (1 + r )]p + [d (1 + r )p ](1 p )
corresponds to a risk premium, that is, an additional return paid as
a compensation for risk.
Doing the same calculations with the risk-neutral probability
rather than p , we get E [St +1 ] = uSt + (1 )dSt = (1 + r )St .
The expected return is then E [SStt +1 ] = 1 + r , i.e., computing the

expected return with probability gives a risk free return.

The risk premium is thus zero: E [SStt +1 ] (1 + r ) = 0.

Introduction Markov chains Application in insurance Application in nance

Risk-neutral probability (cont'd)

Using the risk-neutral probability rather than the historical

probability p neutralizes the risk premium, i.e., investors ask for the
same return are truly risk-neutral agents. The general result follows:
In a one-period model with two states of the world, there exists a
unique probability called risk-neutral probability, such that for
each assetxt +1 R2 its price and at time t is equal to
Pt = 1+1 r E [xt +1 ], that is, the discounted risk-neutral expectation
of the future random payos.
Introduction Markov chains Application in insurance Application in nance

Risk-neutral probability (cont'd)

Note that there is no notion of utility function with this approach.

This approach uses the absence of arbitrage hypothesis to value
assets, which does not require notions of equilibrium and modeling
of economic agents with utility functions.
We make the assumption of a perfect market:
1 we can buy or sell an asset at the same price (no transaction
fees on the risky asset)
2 we can borrow or lend at the same interest rate (no fee on the
risk free asset)
In this model, we only have one period (or two dates), which is a
very short Markov chain! We generalize this result to a multi
periods binomial model.
Introduction Markov chains Application in insurance Application in nance

Multi period binomial model

Multiple periods: n periods, or n + 1 dates

Two states of the world from one period to the other: up and
down
Two nancial assets: a stock and a risk free asset

Let n = 2, that is, three dates 0, 1, 2. At the end of the rst period,
there are two possibilities: S0 S1 = uS0 or S0 S1 = dS0 .
At the end of the second period, there are three possibilities:
S1 S2 = uuS0 or S1 S2 = udS0 = duS0 or S1 S2 = ddS0 .
We thus have a recombining tree: the price is the same after an
upward and downward move that it is after a downward and upward
move. Also, the price only depends on the previous period (Markov
property).
Introduction Markov chains Application in insurance Application in nance

Pricing by replication

The payo function at maturity is

after two ups Cuu = max(S0 uu K , 0)
after one up and one down (and vice versa)
Cud = max(S0 ud K , 0)
after two downs Cdd = max(S0 dd K , 0)
Introduction Markov chains Application in insurance Application in nance

Pricing by replication (cont'd)

Consider the second period, and the asset having the following
payos:
in up state Cuu = max(S0 uu K , 0)
in down state Cud = max(S0 ud K , 0).
We see that we are back to a one-period model and that we can
apply the same tools as in the previous slides, using the risk-neutral
valuation.
We thus have:
Cu = 1+1 r [Cuu + (1 )Cud ] after an upward move in the
rst period
Cd = 1+1 r [Cud + (1 )Cdd ] otherwise.
Introduction Markov chains Application in insurance Application in nance

Pricing by replication (cont'd)

By applying the same method recursively for the rst period, we

getC = 1+1
r [ Cu + (1 )Cd ], that is:
C = (1+1r )2 [ 2 (S0 uu K )+ + 2(1 )(S0 ud K )+ + (1
)2 (S0 dd K )+ ].
This expectation corresponds to a binomial lawB (2, ) whose
events are (S0 uu K )+ , (S0 ud K )+ et (S0 dd K )+ with
probabilities 2 , 2(1 ) et (1 )2 , respectively.
Again, the asset price is given by the discounted expectation of the
future payos under the risk-neutral probability measure.
Introduction Markov chains Application in insurance Application in nance

Generalization

We may apply the same technique with an arbitrary number of

periods n. We get:
C = (1+1r )n nj=0 nj j (1 )nj (u j d nj S0 K )+ , avec
P

n
= n! .
j j !(nj )!
The expectation is taken with the risk-neutral distributionB (n, ).
Introduction Markov chains Application in insurance Application in nance

Generalization (cont'd)

Taking an horizonT and dividing the interval [0, T ] in n

subintervals of length T /n converges to the Black-Scholes formula
as n tends to innity, that is, a continuous-time model where the
underlying asset is assumed to follow a geometric Brownian motion.
This kind of numerical approximation of a continuous-time model is
extremely useful when computations are too dicult to perform
explicitely.