Itô integrals

A crash course

Suppose that W is a Brownian motion and that H is a right-continuous, adapted, and

locally bounded process. The Itô integral of H on [a, b] w.r.t. W is denoted by
Z b
H dWs .
a

Consider a partitions P of the interval [a, b]:

π = {a = x0 < x1 < · · · < xn = b}.

The Itô integral is the limit in probability of the approximating sums

n−1
X
S(π, H, W ) = Hti−1 [Wti − Wti−1 ]
i=0

as the length of the longest subinterval of π goes to 0.

Itô processes

Itô processes are adapted stochastic process that may be decomposed as the sum of an
integral with respect to Brownian motion and an integral with respect to time.

Z t Z t
Xt = X0 + σs dWs + µs ds;
0 0

here σ is a predictable W -integrable process and µ is predictable integrable, e.g.,

Z t
(σs2 + |µs |) ds < ∞ (t > 0).
0

The stochastic integral w.r.t. such a process is defined as:

Z t Z t Z t
H dX = Hs σs dWs + Hs µs ds.
0 0 0
SDEs
Typically an SDE is of the form

dXt = µ(Xt , t) dt + σ(Xt , t) dWt ,

where W is a standard Brownian motion; it means

Z t+s Z t+s
Xt+s − Xt = µ(Xu , u)du + σ(Xu , u) dBu .
t t

This characterises the behaviour of the continuous time stochastic process X as the
sum of a Lebesgue integral and an Itô integral.
Another interpretation of SDE is that in a small time interval of length δ the process X
changes its value by an amount that is normally distributed with expectation µ(Xt , t)δ and
variance σ(Xt , t)2 δ and is independent of the past behaviour (Markov property!). This is so
because the increments of a Brownian motion are independent and normally distributed.
The function µ is referred to as the drift coefficient, while σ is the diffusion coefficient.
Ito’s lemma
Theorem (Ito’s lemma)
For any twice continuously differentiable function f : R2 → R and an Itô process X ,
f (X ) is itself an Itô process satisfying

1
df (Xt ) = f ′ (Xt ) dXt + f ′′ (Xt )σt2 dt.
2

A process S follows a geometric Brownian motion with constant volatility σ and

constant drift µ if it satisfies the SDE
dS = S(σdW + µdt).

Applying Ito’s lemma, with f (s, t) = log s gives

d log(S) = f ′ (S) dS + 12 f ′′ (S)S 2 σ 2 dt
= S1 (σS dW  + µS dt)
1 2
 − 2 σ dt
σ2
= σ dW + µ − 2 dt.
Ito’s lemma

It follows that

 
σ2
log(St ) = log(S0 ) + σBt + µ − 2 t,

exponentiation gives the expression for S,

   
σ2
St = S0 exp σBt + µ − 2 t .
 The Black–Scholes model.

Fischer Black, Myron Scholes, and Robert Merton

Black–Scholes
Assumptions

The Black Scholes Model has some underlying assumptions:

No Dividends: the underlying security does not pay any dividends during its life.
Efficient Market: The market is assumed to be liquid has price continuity and all
the players have equal access to all information available.
No commission or transaction cost: The model assumes that there is no
transaction cost involved in buying/selling the asset.
Constant Volatility: Volatility can be constant relatively in the short run but it can
never be constant in the long run.
As the options are considered European, they can only be exercised at their
expiration.
The risk-free rate of investment and underlying volatility is assumed to be constant.
The returns are distributed evenly throughout the period of the option.
Black–Scholes
Inputs

Spot Price: The market price of the asset as on the day of valuation is the spot
price. This price is difficult to estimate however for the simplicity of use, the
closing market price is used.
Strike price: This is the price at which the option holder has the right to buy or
sell the underlying security. It is not difficult to get this input as it is mentioned in
the option contract.
Time to maturity: The time (years) till the option expires and after which the
option holder do not have any right to exercise it.
Risk-free interest rates: This rate is normally the risk-free rate of the zero coupon
government bond.
Volatility: It is one of the most important inputs in the option pricing model.
There are several ways to estimate volatility. The most common method to
estimate volatility is to collect data from the previous term of the security.
Theorem (Cameron–Martin–Girsanov)
Suppose that X = (Xt )t∈[0,T ] is an (Ft )-predictable process on (Ω, F, P) such that
Z T
E[exp( 12 Xt2 )] < ∞.
0

Then there exists a measure Q on (Ω, F) such that

1 Q is equivalent to P;
2 the Radon–Nikodym derivative (density) dQ/dP has the form
Z T Z T
dQ
= exp[− Xt dWt − 1
2 Xt2 dt],
dP 0 0

3 The process Z t
W̃t = Wt + Xs ds
0
is a Brownian motion with respect to Q and (Ft ).
The Black–Scholes model
Framework

The time horizon: [0, T ].

The price of the riskless asset B is given by Bt = e rt (0 ⩽ t ⩽ T ).
The price of the risky asset S is driven by the SDE

dSt = µSt dt + σSt dWt (t ∈ [0, T ])

with S0 > 0.
As such
σ2
 
St = S0 exp (µ − )t + σWt .
2
(St is geometric Brownian motion with drift µ = σ 2 /2 and volatility σ.
Black–Scholes

Consider the discounted process St∗ = e −rt St (t ∈ [0, T ]).

Note that for s < t we have
h ∗ i h
i
σ2
E SSt∗ |Fs = E exp (µ − − r )(t − s) + σ(Wt − Ws ) |Fs
2
s h i
σ2
= E exp (µ − − r )(t − s) + σ(Wt − Ws )
2 h i
σ2
= exp (µ − 2 − r )(t − s) · E exp(σWt−s )
 2 2

= exp (µ − σ2 − r )(t − s) + σ2 (t − s)
= e (µ−r )(t−s) ,

θ2 σ2
where we used that E[e θZ ] = e θµ+ 2 , whenever Z ∼ N(µ, σ 2 ).
Hence S ∗ is a martingale under P if and only if µ = r .
Does there exist a probability measure Q under which S ∗ is a martingale?
Black–Scholes

The answer is given by the (Cameron–Martin–)Girsanov theorem. Let Q by given

by the Radon–Nikodym density
 !
1 µ−r 2

dQ µ−r
= exp − WT − T .
dP σ 2 σ

µ−r
Then the process W̃t = σ t + Wt is a Brownian motion under Q.
Moreover, S∗ is a martingale under Q.

Theorem (Risk-neutral pricing principle)

Let X be a contingent claim such that EQ [|X |] < ∞. Then the risk-neutral price at
t ∈ [0, T ] is given by
PX (t) = e −r (T −t) EQ [X |Ft ].
Quadratic variation
Let X = (Xt ) be a stochastic process on (Ω, F, P). The quadratic variation of X at t
is defined as
Xn
[X ]t = lim (Xtk − Xtk−1 )2
∥π∥→0
k=1

where π is a partition of [0, t] should the limit exist in the sense of convergence in
probability.
More generally, the covariance of two processes is defined as
n
X


[X , Y ]t = lim Xtk − Xtk−1 Ytk − Ytk−1
∥π∥→0
k=1

We have
[X , Y ]t = 21 ([X + Y ]t − [X ]t − [Y ]t )
(if all variations exist; think of the covariance as an inner product).
Girsanov’s theorem (toy version)

Theorem (Girsanov’s theorem (toy version))

Let X be a process continuous in mean-squarea on (Ω, F, P). Then there exists and
equivalent measure Q such that

W̃t = Wt − [W , X ]t

is a Brownian process under Q.

h i
a 2
E|Xt |2 < ∞ and lims→t E Xs − Xt = 0.

Observe that
h i
W̃ = [Wt , Wt ] − 2 [Wt , [W , X ]t ] + [[W , X ]t , [W , X ]t ] = [W ]t = t.
t
Quadratic variation – problems

Show that for each t, [·, ·]t is an inner product on the space X2t of processes
X = (Xs ) such that E|Xs |2 < ∞ for all s ⩽ t. Use bilinearity of inner products to
justify the calculation under statement of the toy version of Girsanov’s theorem.
Find the covariance function f (r , s) = [W , W ]t of the standard Brownian motion
W.
Try to find in the literature a relevant theorem for Gaussian processes and their
determination by their their covariance functions.
Stochastic processes: recap
A (real-valued) stochastic process X indexed by [0, T ] is a family of random variables
X = {Xt }t∈[0,T ] defined on the same probability space (Ω, F, P).

We can think of a stochastic process as a function

X : [0, T ] × Ω → R

which maps (t, ω) to Xt (ω).

For every ω ∈ Ω fixed, the process X defines a function

X(ω) : [0, T ] → R

which maps t to Xt (ω), and is called a trajectory or a sample path of the process.
Hence, we can look at X as a mapping

X : Ω → R[0,T ]

which maps ω to X(ω) , where R[0,T ] is the Cartesian product of [0, T ] copies of R
which is the set of all functions from [0, T ] to R. That is, we can see X as a
mapping from Ω to a space of functions.
The Black–Scholes model – pricing formulae — proof
The Black–Scholes model – pricing formulae — proof
The Black–Scholes model – pricing formulae — proof
Greeks from the B–S model
Let us close the topic CRR vs BS for good: recap on characteristic functions

For a random variable X , the characteristic function of X is φ(t) = E exp(itX ).

Let us close the topic CRR vs BS for good: recap on characteristic functions

For a random variable X , the characteristic function of X is φ(t) = E exp(itX ).

It encodes all information on the distribution of X !
Let us close the topic CRR vs BS for good: recap on characteristic functions

For a random variable X , the characteristic function of X is φ(t) = E exp(itX ).

It encodes all information on the distribution of X !
For (absolutely) continuous variables, φX is just the Fourier transform of the density!
(In general, it is Fourier–Stjelties transform of the distribution.)
Let us close the topic CRR vs BS for good: recap on characteristic functions

For a random variable X , the characteristic function of X is φ(t) = E exp(itX ).

It encodes all information on the distribution of X !
For (absolutely) continuous variables, φX is just the Fourier transform of the density!
(In general, it is Fourier–Stjelties transform of the distribution.)
We have the inversion formula for the points of continuity of the c.d.f. F of X :
Za
1 e −ity − e −itx
F (x) − F (y ) = lim φX (t) dt.
a→∞ 2π it
−a
Convergence in distribution

A sequence (Xn ) of random variables converges in distribution (weakly, in law) to a

random variable X , whenever

lim FXn (t) → FX (t)

n→∞

for every point of continuity t of FX .

Convergence in distribution

A sequence (Xn ) of random variables converges in distribution (weakly, in law) to a

random variable X , whenever

lim FXn (t) → FX (t)

n→∞

for every point of continuity t of FX .

Xn → X in distribution, if and only if g(Xn ) → g(Xn ) for every bounded Lipschitz
function g .
Theorem (Levy continuity theorem)
A sequence (Xn ) of random variables converges in distribution to X if and only if the
sequence (φXn ) converges pointwise to φX .
Convergence CRR → BS

Consider a family of CRR market models indexed by n;

T T
Partition [0, T ) into [(j − 1) · n ,j · n ), j = 1, . . . , n,
T
Let Sn (j) stand for the stock price at j · n in the nth model,
T
Let Bn (j) stand for the bond price at j · n in the nth model,
T
Let rn = r · n be the interest rate; recall that
1 n
lim (1 + ) = e rT ,
n→∞ rn
p
Set an = σ · T /n, where σ is interpreted as the instantaneous volatility;
Set up the factor up/down factors by

un = e an (1 + rn )

dn = e −an (1 + rn );
they are separated by 1 for large enough n.
Convergence CRR → BS
The risk-neutral measure parameter in the nth model is
1 + rn − dn 1 − e −an an − 12 an2 + o(an2 ) 1
qn = = an = = 2 − 14 an + o(an ).
un − dn e − e −an 2an + 13 an3 + o(an3 )
Convergence CRR → BS
The risk-neutral measure parameter in the nth model is
1 + rn − dn 1 − e −an an − 12 an2 + o(an2 ) 1
qn = = an = = 2 − 14 an + o(an ).
un − dn e − e −an 2an + 13 an3 + o(an3 )

Let (Xn (j))nj=1 be the Bernoulli variables driving the nth CRR model:
Pj Pj
k=1 Xn (k) j− k=1 Xn (k)
Sn (j) = S(0)un dn
Convergence CRR → BS
The risk-neutral measure parameter in the nth model is
1 + rn − dn 1 − e −an an − 12 an2 + o(an2 ) 1
qn = = an = = 2 − 14 an + o(an ).
un − dn e − e −an 2an + 13 an3 + o(an3 )

Let (Xn (j))nj=1 be the Bernoulli variables driving the nth CRR model:
Pj Pj
k=1 Xn (k) j− k=1 Xn (k)
Sn (j) = S(0)un dn

The value at time zero of a put option with strike K is given by

 
n −n + K Yn
PP (0) = (1 + rn ) EQn [(K − S(n)) ] = EQn − S(0)e ,
(1 + rn )n
where
n n
!
X (j) 1−Xn (j)
X X un n dn
Yn = Yn (j) = log .
1 + rn
j=1 j=1
Convergence CRR → BS

For Yn (1), . . . , Yn (n) i.i.d. with

un dn
EQn [Yn (j)] = qn log( 1+r n
) + (1 − qn ) log 1+r n
= ( 12 − 14 an + o(an ))an + ( 12 + 41 an + o(an ))(−an )
= − 21 an2 + o(an2 )
2
EQn [Yn (j)] = an2 + o(an2 ),
m
EQn [|Yn (j)| ] = o(an2 ).
Convergence CRR → BS
2
Let Y ∼ N(− σT 2, σ 2 T ). Then

σ2T 2
 
2σ T
ϕY (θ) = exp −iθ −θ .
2 2
Convergence CRR → BS
2
Let Y ∼ N(− σT 2, σ 2 T ). Then

σ2T 2
 
2σ T
ϕY (θ) = exp −iθ −θ .
2 2

As Yn (j), . . . , Yn (n) are i.i.d., we have

EQn [e iθYn ] = nj=1 EQn [e iθYn (j) ] = E[e iθYn (1) ]n

Q
ϕYn (θ) =
2
= (1 + iθEQn [Yn (j)] − θ2 E[Yn2 (j)] + o(an2 ))n
2
= (1 − ( iθ+θ 2 2 2
2 )an + o(an ))
iθ+θ2 2 T
= (1 − ( 2 )σ n + o(1/n))n ,

which converges to ϕY (θ) as n → ∞.

Convergence CRR → BS
2
Let Y ∼ N(− σT 2, σ 2 T ). Then

σ2T 2
 
2σ T
ϕY (θ) = exp −iθ −θ .
2 2

As Yn (j), . . . , Yn (n) are i.i.d., we have

EQn [e iθYn ] = nj=1 EQn [e iθYn (j) ] = E[e iθYn (1) ]n

Q
ϕYn (θ) =
2
= (1 + iθEQn [Yn (j)] − θ2 E[Yn2 (j)] + o(an2 ))n
2
= (1 − ( iθ+θ 2 2 2
2 )an + o(an ))
iθ+θ2 2 T
= (1 − ( 2 )σ n + o(1/n))n ,

which converges to ϕY (θ) as n → ∞.

2
By Levy’s continuity theorem, Yn converges to N(− σ 2T , σ 2 T ) in distribution.
Convergence CRR → BS
A sequence {Yn }n≥1 of random variables converges in distribution to Y if and only if

En [g (Yn )] → E[g (Y )],

when n → +∞, for all g ∈ Cb (R).

One can check that
 + 
PPn (0) − EQ Ke −rT
− S(0)e Yn
⩽ K (1 + rn )−n − e −rT .

Therefore,  + 
−rT
lim P n (0) = lim EQ Ke − S(0)e Yn
n→∞ P n→∞

z2 +
+∞
e− 2 √
Z   2
−rT σ T
= √ Ke − S(0) exp − + σ Tz dz = PP (0),
−∞ 2π 2
 2  2 √
where we have used that Y ∼ N − σ 2T , σ 2 T if Y = − σ 2T + σ T Z with Z ∼ N (0, 1).
Convergence CRR → BS
One can directly check that

PP (0) = Ke −rT Φ(−d2 (S(0), T )) − S(0)Φ(−d1 (S(0), T )),

where Φ is the c.d.f. of N (0, 1).

Convergence CRR → BS
One can directly check that

PP (0) = Ke −rT Φ(−d2 (S(0), T )) − S(0)Φ(−d1 (S(0), T )),

where Φ is the c.d.f. of N (0, 1).

By using the put-call parity one gets that

lim PC′ (0) = PC (0) = S(0)Φ(d1 (S(0), T )) − Ke −rT Φ(d2 (S(0), T )),
n→∞

where
Pc′ (0) = (1 + rn )−n EQ (S(n) − K )+ .
 
Convergence CRR → BS
One can directly check that

PP (0) = Ke −rT Φ(−d2 (S(0), T )) − S(0)Φ(−d1 (S(0), T )),

where Φ is the c.d.f. of N (0, 1).

By using the put-call parity one gets that

lim PC′ (0) = PC (0) = S(0)Φ(d1 (S(0), T )) − Ke −rT Φ(d2 (S(0), T )),
n→∞

where
Pc′ (0) = (1 + rn )−n EQ (S(n) − K )+ .
 

One can modify the previous arguments to provide formulae for Pc (t) and Pp (t).
Convergence CRR → BS

Theorem
Let g ∈ Cb (R) and let X = g (S(T )) be a contingent claim in the Black–Scholes
model. Then the price process of X is given by

PX (t) = lim PXn (t), 0 ⩽ t ⩽ T,

n→+∞

where PXn (t), n ⩾ 1 are the price processes of X in the corresponding CRR models.
Convergence CRR → BS

Theorem
Let g ∈ Cb (R) and let X = g (S(T )) be a contingent claim in the Black–Scholes
model. Then the price process of X is given by

PX (t) = lim PXn (t), 0 ⩽ t ⩽ T,

n→+∞

where PXn (t), n ⩾ 1 are the price processes of X in the corresponding CRR models.

There are similar proofs of the previous results using the normal approximation to
the binomial law, based on the CLT.
However, note that here we have a triangular array of random variables
{Yn (j)}nj=1,...,n ⩾ 1. Hence, standard version of the CLT is not applicable here.
Moreover, the asymptotic distribution of Yn need not be Gaussian if we choose
−rT
suitably the parameters of the CRR model: if we set un = u and dn = e n , c < r
we have that Yn converges in law to a Poisson random variable.
This lead to consider the exponential of more general Lévy process as underlying
price process for the stock.
American options

In contrast to European contingent claims, an American claim may be exercised at

any time before the expiration date T .
American options

In contrast to European contingent claims, an American claim may be exercised at

any time before the expiration date T .
In the study of American claims, one is more concerned with the price process and
the ‘optimal’ exercise policy by its holder.
American options

In contrast to European contingent claims, an American claim may be exercised at

any time before the expiration date T .
In the study of American claims, one is more concerned with the price process and
the ‘optimal’ exercise policy by its holder.
If the holder of an American option exercises it at τ ⩽ T , τ is called an exercise
time.
What exercise times are ‘optimal’ ?