MIT15 450F10 Lec02

Stochastic Integral Its Lemma Black-Scholes Model Multivariate It Processes SDEs SDEs and PDEs Risk-Neutral Probability Risk-Neutral
Neutral Pricing
Stochastic Calculus and Option Pricing
Leonid Kogan
MIT, Sloan
15.450, Fall 2010
Leonid Kogan ( MIT, Sloan ) Stochastic Calculus 15.450, Fall 2010 1 / 74 c
Stochastic Integral Its Lemma Black-Scholes Model Multivariate It Processes SDEs SDEs and PDEs Risk-Neutral Probability Risk-Neutral Pricing
Outline
1
Stochastic Integral
2
Its Lemma
3
Black-Scholes Model
4
Multivariate It Processes
5
SDEs
6
SDEs and PDEs
7
Risk-Neutral Probability
8
Risk-Neutral Pricing
c Stochastic Calculus 2 / 74 Leonid Kogan ( MIT, Sloan ) 15.450, Fall 2010
Outline
1
Stochastic Integral
2
Its Lemma
3
Black-Scholes Model
4
5
SDEs
6
SDEs and PDEs
7
8
Brownian Motion
Consider a random walk x
t +nt
with equally likely increments of
t .
Let the time step of the random walk shrink to zero: t 0.
The limit is a continuous-time process called Brownian motion, which we
denote Z
t
, or Z (t ).
We always set Z
0
=0.
Brownian motion is a basic building block of continuous-time models.
c Leonid Kogan ( MIT, Sloan ) Stochastic Calculus 15.450, Fall 2010 4 / 74
Properties of Brownian Motion
Brownian motion has independent increments: if t <t
<t
, then Z
t
Z
t
is
independent of Z
t
Z
t
.
Increments of the Brownian motion have normal distribution with zero mean,
and variance equal to the time interval between the observation points
Z
t
Z
t
N(0,t
t )
Thus, for example,
E
t
[Z
t
Z
t
] =0
Intuition: Central Limit Theorem applied to the random walk.
Trajectories of the Brownian motion are continuous.
Trajectories of the Brownian motion are nowhere differentiable, therefore
standard calculus rules do not apply.
Ito Integral
Ito integral, also called the stochastic integral (with respect to the Brownian
motion) is an object
u
dZ
u
0
where
u
is a stochastic process.
Important:
u
can depend on the past history of Z
u
, but it cannot depend on
the future.
u
is called adapted to the history of the Brownian motion.
Consider discrete-time approximations
N
(i 1)t
(Z
i t
Z
(i 1)t
), t =
N
i =1
and then take the limit of N (the limit must be taken in the
mean-squared-error sense).
The limit is well dened, and is called the Ito integral.
Properties of It Integral
It integral is linear
t
(a
u
+c b
u
) dZ
u
= a
u
dZ
u
+c b
u
dZ
u
0 0 0
t
X
t
=
0
u
dZ
u
, is continuous as a function of time.
Increments of X
t
have conditional mean of zero (under some technical
restrictions on
u
):
E
t
[X
t
X
t
] =0, t
>t
Note: a sufcient condition for E

t
t
u
dZ
u
=0 is E
0
0
2
u
du <.
Increments of X
t
are uncorrelated over time.
If
t
is a deterministic function of time and
0
t
2
du <, then X
t
is normally
u
distributed with mean zero, and variance
t
E
0
[X
t
2
] =
u
2
du
0

It Processes
An It process is a continuous-time stochastic process X
t
, or X(t ), of the form
u
du +
u
dZ
u
0 0
u
is called the instantaneous drift of X
t
at time u, and
u
is called the
instantaneous volatility, or the diffusion coefcient.
t
dt captures the expected change of X
t
between t and t +dt .
t
dZ
t
captures the unexpected (stochastic) component of the change of X
t
between t and t +dt .
Conditional mean and variance:
E
t
(X
t +dt
X
t
) =
t
dt +o(dt ), E
t
(X
t +dt
X
t
)
2
=
t
2
dt +o(dt )
Outline
1
Stochastic Integral
2
Its Lemma
3
Black-Scholes Model
4
5
SDEs
6
SDEs and PDEs
7
8
Quadratic Variation
Consider a time discretization
0 =t
1
<t
2
<...<t
N
=T , max |t
n+1
t
n
|<
n=1,...,N1
Quadratic variation of an It process X (t )between 0 and T is dened as
N1
[X ]
T
= lim |X(t
n+1
) X (t
n
)|
2
0
n=1
For the Brownian motion, quadratic variation is deterministic:
[Z ]
T
=T
To see the intuition, consider the random-walk approximation to the Brownian
motion: each increment equals
t
n+1
t
n
in absolute value.
Quadratic Variation
Quadratic variation of an It process
t
X
t
=
u
du +
u
dZ
u
0 0
is given by
T
[X ]
T
=
2
t
dt
0
Heuristically, the quadratic variation formula states that
(dZ
t
)
2
=dt , dt dZ
t
=o(dt ), (dt )
2
=o(dt )
Random walk intuition:
|dZ
t
|=
dt , |dt dZ
t
|= (dt )
3/2
=o(dt ), dZ
t
2
=dt
Conditional variance of the It process can be estimated by approximating its
quadratic variation with a discrete sum. This is the basis for variance
estimation using high-frequency data.

Its Lemma
Its Lemma states that if X
t
is an It process,
t
X
t
=
u
du +
u
dZ
u
0 0
then so is f (t ,X
t
), where f is a sufciently smooth function, and
df (t ,X
t
) =
f (t ,X
t
)
+
f (t ,X
t
)
t
+
1
2
f (t ,X
t
)
2
t
dt +
f (t ,X
t
)
t
dZ
t
t X
t
2 X
t
2
X
t
Its Lemma is, heuristically, a second-order Taylor expansion in t and X
t
,
using the rule that
(dZ
t
)
2
=dt , dt dZ
t
=o(dt ), (dt )
2
=o(dt )
Its Lemma
Using the Taylor expansion,
df (t ,X
t
)
f (
t ,
t
X
t
)
dt +
f (
t
X
,X
t
t
)
dX
t
+
2
1
2
f
(t
t
,
2
X
t
)
dt
2
+
2
1
2
f
(
X
t ,
t
2
X
t
)
(dX
t
)
2
+
f
t
(
t ,
X
X
t
t
)
dt dX
t
f (t ,X
t
) f (t ,X
t
) f (t ,X
t
)
= dt +
t
dt +
t
dZ
t
t X
t
X
t
+
2
1
2
f
(
X
t ,
t
2
X
t
)
2
t
dt +o(dt )
Short-hand notation
df (t ,X
t
) =
f (
t ,
t
X
t
)
dt +
f (
t
X
,X
t
t
)
dX
t
+
2
1
2
f
(
X
t ,
t
2
X
t
)
(dX
t
)
2

Its Lemma
Example
Let X
t
=exp(at +bZ
t
).
We can write X
t
=f (t ,Z
t
), where
f (t ,Z
t
) =exp(at +bZ
t
)
Using
f (t ,Z
t
)
=af (t ,Z
t
),
f (t ,Z
t
)
=bf (t ,Z
t
),
2
f (t ,Z
t
)
=b
2
f (t ,Z
t
)
t Z
t
Z
t
2
Its Lemma implies
b
2
dX
t
= a + X
t
dt +bX
t
dZ
t
2
Expected growth rate of X
t
is a +b
2
/2.
Outline
1
Stochastic Integral
2
Its Lemma
3
Black-Scholes Model
4
5
SDEs
6
SDEs and PDEs
7
8
The Black-Scholes Model of the Market
Consider the market with a constant risk-free interest rate r and a single risky
asset, the stock.
Assume the stock does not pay dividends and the price process of the stock
is given by

S
t
=S
0
exp
1
2
t +Z
t
2
Because Brownian motion is normally distributed, using
E
0
[exp(Z
t
)]=exp(t /2), nd
E
0
[S
t
] =S
0
exp(t )
Using Its Lemma (check)
dS
t
=dt +dZ
t
S
t
is the expected continuously compounded stock return, is the volatility of
stock returns.
Stock returns have constant volatility.

Dynamic Trading
Consider a trading strategy with continuous rebalancing.
At each point in time, hold
t
shares of stocks in the portfolio.
Let the portfolio value be W
t
. Then W
t
t
S
t
dollars are invested in the
short-term risk-free bond.
Portfolio is self-nancing: no exogenous incoming or outgoing cash ows.
Portfolio value changes according to
dW
t
=
t
dS
t
+ (W
t
t
S
t
)r dt
Discrete-time analogy
B
t +t
W
t +t
W
t
=
t
(S
t +t
S
t
) + (W
t
t
S
t
) 1
B
t
Option Replication
Consider a European option with the payoff H(S
T
).
We will construct a self-nancing portfolio replicating the payoff of the option.
Look for the portfolio such that
W
t
=f (t ,S
t
)
for some function f (t ,S
t
).
By Law of One Price, f (t ,S
t
)must be the price of the option at time t , being
the cost of a trading strategy with an identical payoff.
Note that the self-nancing condition is important for the above argument: we
do not want the portfolio to produce intermediate cash ows.

Option Replication
Apply Its Lemma to portfolio value, W
t
=f (t ,S
t
):
f f 1
2
f
dW
t
=
t
dS
t
+ (W
t
t
S
t
)r dt =
t
dt +
S
t
dS
t
+
2 S
t
2
(dS
t
)
2
where (dS
t
)
2
=
2
S
t
2
dt
The above equality holds at all times if
t
=
f (
t
S
,S
t
t
)
,
f (
t ,
t
S
t
)
+
2
1
2
f
(
S
t ,
t
2
S
t
)
2
S
t
2
r f (t ,S
t
)
S
f
t
S
t
=0
If we can nd the solution f (t ,S)to the PDE
f (t ,S) f (t ,S) 1
2
S
2
2
f (t ,S)
r f (t ,S) + +rS + =0
t S 2 S
2
with the boundary condition f (T ,S) =H(S), then the portfolio with
f (t ,S
t
)
W
0
=f (0,S
0
),
t
=
S
t
replicates the option!
Black-Scholes Option Price
We conclude that the option price can be computed as a solution of the
Black-Scholes PDE
f f 1
2
f
r f + +rS +
2
S
2
=0
t S 2 S
2
If the option is a European call with strike K , the PDE can be solved in closed
form, yielding the Black-Scholes formula:
C(t ,S
t
) =S
t
N(z
1
) exp(r (T t ))KN(z
2
),
where N(.)is the cumulative distribution function of the standard normal
distribution,

log
S
K
t
+ r +
2
1
2
(T t )
z
1
= ,
T t
and
z
2
=z
1
T t .
Note that does not enter the PDE or the B-S formula. This is intuitive from
the perspective of risk-neutral pricing. Discuss later.
1
2
3
Black-Scholes Option Replication
The replicating strategy requires holding
f (t ,S
t
)
t
=
S
t
stock shares in the portfolio.
t
is called the options delta.
It is possible to replicate any option in the Black-Scholes setting because
Price of the stock S
t
is driven by a Brownian motion;
Rebalancing of the replicating portfolio is continuous;
There is a single Brownian motion affecting the payoff of the option and the price
of the stock. More on this later, when we cover multivariate It processes.
Single-Factor Term Structure Model
Consider a model of the term structure of default-free bond yields.
Assume that the short-term interest rate follows
dr
t
=(r
t
)dt +(r
t
)dZ
t
Let P(t ,)denote the time-t price of a discount bond with unit face value
maturing at time .
We want to construct an arbitrage-free model capturing, simultaneously, the
dynamics of bond prices of many different maturities.

Assume that bond prices can be expressed as a function of the short rate
only (single-factor structure)
P(t ,) =f (t ,r
t
,)
It formula implies that
dP(t ,) =
f
+
f
(r
t
) +
1
2
f
2
(r
t
) dt +
f
(r
t
) dZ
t
t r 2 r
2
t
t
Consider a self-nancing portfolio which invests P(t ,)/
t
dollars in the
bond maturing at , and P(t ,
)/
dollars in the bond maturing at
.
Self-nancing requires that the investment in the risk-free short-term bond is
W
t
P(
,)
+
P(
t ,
)
t t

Portfolio value evolves according to
dW
t
= W
t
P(
t ,
)
+
P(
t ,
)
r
t
+
dt +
t t t t
t t
dZ
t
t t
= W
t
P(
t ,
)
+
P(
t ,
)
r
t
+
dt
t t t t
Portfolio value changes are instantaneously risk-free.
To avoid arbitrage, the portfolio value must grow at the risk-free rate:
W
t
P(
t ,
)
+
P(
t ,
)
r
t
+
=W
t
r
t
t t t t
We conclude that
r
t
P(t ,)
=
r
t
P(t ,
)
, for any and
t t
Assume that, for some
r
t
P(t ,
)
=(t ,r
t
)
t
Then, for all bonds, must have
r
t
P(t ,) =(t ,r
t
)
t
Recall the denition of
t
t
to derive the pricing PDE on P(t ,) =f (t ,r
t
,)
f f 1
2
f f
+ (r ) +
2
(r ) rf =(t ,r )(r ) , f (,r ,) =1
t r 2 r
2
r
The solution, indeed, has the form f (t ,r ,).
What if
t
r
t
P(t ,)
t
=(t ,r
t
)
for any function , i.e., the LHS depends on something other than r
t
and t ?
Then the term structure will not have a single-factor form.
We have seen that, for each choice of (t ,r
t
), absence of arbitrage implies
that bond prices must satisfy the pricing PDE.
The reverse is true: if bond prices satisfy the pricing PDE (with well-behaved
(t ,r
t
)), there is no arbitrage (show later, using risk-neutral pricing).
The choice of (t ,r
t
)determines the joint arbitrage-free dynamics of bond
prices (yields).
(t ,r
t
)is the price of interest rate risk.
Outline
1
Stochastic Integral
2
Its Lemma
3
Black-Scholes Model
4
5
SDEs
6
SDEs and PDEs
7
8
Multiple Brownian motions
Consider two independent Brownian motions Z
t
1
and Z
t
2
. Construct a third
process

X
t
=Z
t
1
+ 1
2
Z
t
2
X
t
is also a Brownian motion:
X
t
has IID normal increments;
X
t
is continuous.
X
t
and Z
t
1
are correlated:

E
0
X
t
Z
t
1
=t
Correlated Brownian motions can be constructed from uncorrelated ones,
just like with normal random variables.
Cross-variation
N1
[Z
t
1
,Z
t
2
]
T
= lim (Z
1
(t
n+1
) Z
1
(t
n
))(Z
2
(t
n+1
) Z
2
(t
n
))=0
0
n=1
Short-hand rule
dZ
t
1
dZ
t
2
=0 dZ
t
1
dX
t
=dt
A multivariate It process is a vector process with each coordinate driven by
an It process.
Consider a pair of processes
dX
t
=
t
X
dt +
X
t
dZ
t
X
,
dY
t
=
t
Y
dt +
Y
t
dZ
t
Y
,
dZ
t
X
dZ
t
Y
=
t
dt
Its formula can be extended to multiple process as follows:
f f f
df (t ,X
t
,Y
t
) = dt + dX
t
+ dY
t
+
t X
t
Y
t
1
2
f 1
2
f
2
f
(dX
t
)
2
+ (dY
t
)
2
+ dX
t
dY
t
2 X
t
2
2 Y
t
2
X
t
Y
t
Example of Its formula
Consider two asset price processes, X
t
and Y
t
, both given by It processes
dX
t
=
X
t
dt +
X
t
dZ
t
X
dY
t
=
t
Y
dt +
Y
t
dZ
t
Y
dZ
t
X
dZ
t
Y
=0
Using Its formula, we can derive the process for the ratio f
t
=X
t
/Y
t
(use
f (X,Y ) =X/Y ):

2
df
t
dX
t
dY
t
dX
t
dY
t
dY
t
= +
f
t
X
t
Y
t
X
t
Y
t
Y
t
X Y
Y
=
t
+
t
dt +
t
dZ
t
X
t
dZ
t
Y
X
t
Y
t
Y
t
2
X
t
Y
t
Example of Its formula
We nd that the expected growth rate of the ratio X
t
/Y
t
is
X Y
t
t
+
X
t
Y
t
Y
t
2
Assume that
t
X
=
t
Y
. Then,
df
t
Y
t
2
E
t
= dt
f
t
Y
t
2
Repeating the same calculation for the inverse ratio, h
t
=Y
t
/X
t
, we nd

2
E
t
dh
t
=
t
X
dt
h
t
X
t
2
It is possible for both the ratio X
t
/Y
t
and its inverse Y
t
/X
t
to be expected to
grow at the same time. Application to FX.
Outline
1
Stochastic Integral
2
Its Lemma
3
Black-Scholes Model
4
5
SDEs
6
SDEs and PDEs
7
8
Stochastic Differential Equations
Example: Hestons stochastic volatility model
dS
t
=dt +
v
t
dZ
S
S
t
t
dv
t
= (v
t
v)dt +
v
t
dZ
t
v
dZ
t
S
dZ
t
v
=dt
Conditional variance v
t
is described by a Stochastic Differential Equation.
Denition (SDE)
The It process X
t
satises a stochastic differential equation
dX
t
=(t ,X
t
)dt +(t ,X
t
)dZ
t
with an initial condition X
0
if it satises
X
t
=X
0
+
t
0
(s,X
s
)ds +(s,X
s
)dZ
s
.
Existence of Solutions of SDEs
Assume that for some C,D >0
|(t ,X)|+|(t ,X )|C(1 +|X |)
and
|(t ,X ) (t ,Y )|+|(t ,X) (t ,Y )|D|X Y |
for any X and Y (Lipschitz property).
Then, the SDE
dX
t
=(t ,X
t
)dt +(t ,X
t
)dZ
t
, X
0
=x,
has a unique continuous solution X
t
.
Common SDEs
Arithmetic Brownian Motion
The solution of the SDE
dX
t
=dt +dZ
t
is given by
X
t
=X
0
+t +Z
t
.
The process X
t
is called an arithmetic Brownian motion, or Brownian motion
with a drift.
Guess and verify.
We typically reduce an SDE to a few common cases with explicit solutions.
Common SDEs
Geometric Brownian Motion
Consider the SDE
dX
t
=X
t
dt +X
t
dZ
t
Dene the process
Y
t
=ln(X
t
).
By Its Lemma,
dY
t
=
X
1
t
X
t
dt +
X
1
t
X
t
dZ
t
+
1
2
(
X
1
t
2
)
2
X
t
2
dt = (
2
/2)dt +dZ
t
.
Y
t
is an arithmetic Brownian motion, given in the previous example, and
Y
t
=Y
0
+ (
2
/2)t +Z
t
.
Then

X
t
=e
Y
t
=X
0
exp (
2
/2)t +Z
t
Common SDEs
Ornstein-Uhlenbeck process
The mean-reverting Ornstein-Uhlenbeck process is the solution X
t
to the
stochastic differential equation
dX
t
= (X X
t
)dt +dZ
t
We solve this equation using e
t
as an integrating factor.
Setting Y
t
=e
t
and using Its lemma for the function f (X ,Y ) =X Y , we nd
d(e
t
X
t
) =e
t
(X dt +dZ
t
).
Integrating this between 0 and t , we nd
t
e
t
X
t
X
0
= e
s
Xds + e
s
dZ
s
,
0 0
i.e.,
t
X
t
=e
t
X
0
+ (1 e
t
)X + e
st
dZ
s
.
0
Option Replication in the Heston Model
Assume the Heston stochastic-volatility model for the stock.
Attempt to replicate the option payoff with the stock and the risk-free bond.
Can we nd a trading strategy that would guarantee perfect replication?
It is possible to replicate an option using a bond, a stock, and another option.

Recap of APT
Recall the logic of APT.
Suppose we have N assets with two-factor structure in their returns:
R
t
i
=a
i
+b
i
1
F
t
1
+b
i
2
F
t
2
Interest rate is r .
While not stated explicitly, all factor loadings may be stochastic.
Unlike the general version of the APT, we assume that returns have no
idiosyncratic component.
At time t , consider any portfolio with fraction
i
in each asset i that has zero
exposure to both factors:
b
1
1
1
+b
1
2
2
+...+b
1
N
N
=0
b
2
1
1
+b
2
2
2
+...+b
2
N
N
=0
This portfolio must have zero expected excess return to avoid arbitrage
1
E
t
[R
t
1
] r +
2
E
t
[R
t
2
] r +...+
N
E
t
[R
t
N
] r =0

Recap of APT
To avoid arbitrage, any portfolio satisfying
b
1
1
1
+b
1
2
2
+...+b
1
N
N
=0
b
2
1
1
+b
2
2
2
+...+b
2
N
N
=0
must satisfy
1
E
t
[R
t
1
] r +
2
E
t
[R
t
2
] r +...+
N
E
t
[R
t
N
] r =0
Restating this in vector form, any vector orthogonal to
(b
1
1
,b
1
2
,...,b
1
N
)and (b
2
1
,b
2
2
,...,b
2
N
)
must be orthogonal to (E
t
[R
t
1
] r ,E
t
[R
t
2
] r ,...,E
t
[R
t
N
] r ).
Conclude that the third vector is spanned by the rst two: there exist
constants (prices of risk) (
1
t
,
2
t
)such that
E
t
[R
t
i
] r =
1
t
b
1
i
+
t
2
b
2
i
, i =1,...,N
Option Pricing in the Heston Model
Suppose there are N derivatives with prices given by
f
i
(t ,S
t
,v
t
)
where the rst option is the stock itself: f
1
(t ,S
t
,v
t
) =S
t
.
Using Itos lemma, their prices satisfy
df
i
(t ,S
t
,v
t
) =a
i
t
dt +
f
i
(t ,S
t
,v
t
)
dS
t
+
f
i
(t ,S
t
,v
t
)
dv
t
S
t
v
t
Compare the above to our APT argument
Conclude that there exist
S
t
and
v
t
such that
E
df
i
(t ,S
t
,v
t
) rf
i
(t ,S
t
,v
t
)dt
=
f
i
(t
,
S
S
t
t
,v
t
)
t
S
dt +
f
i
(t
,
v
S
t
t
,v
t
)
t
v
dt
We work with price changes instead of returns, as we did in the APT,
because some of the derivatives may have zero price.

Option Pricing in the Heston Model
The APT pricing equation, applied to the stock, implies that
E [dS
t
rS
t
dt ] = (r )S
t
dt =
t
S
dt
v
t
is the price of volatility risk, which determines the risk premium on any
investment with exposure to dv
t
.
Writing out the pricing equation explicitly, with the Itos lemma providing an
expression for E df
i
(t ,S
t
,v
t
) ,
f
i
f
i
f
i
+ S + ()(v v)+
t S v
1
2
f
i
1
2
f
i
2
f
i
f
i
f
i
vS
2
+
2
v + Sv rf
i
= (r )S +
v
2 S
2
2 v
2
Sv S v
t
As long as we assume that the price of volatility risk is of the form
v
=
v
(t ,S
t
,v
t
)
t
the assumed functional form for option prices is justied and we obtain an
arbitrage-free option pricing model.

Numerical Solution of SDEs
Except for a few special cases, SDEs do not have explicit solutions.
The most basic and common method of approximating solutions of SDEs
numerically is using the rst-order Euler scheme.
Use the grid t
i
=i .
X
i +1
=X
i
+(t
i
,X
i
)+(t
i
,X
i
)
i
,
where
i
are IID N(0,1)random variables.
Using a binomial distribution for
i
, with equal probabilities of 1, is also a
valid procedure for approximating the distribution of X
t
Outline
1
Stochastic Integral
2
Its Lemma
3
Black-Scholes Model
4
5
SDEs
6
SDEs and PDEs
7
8
Moments of Diffusion Processes
Often need to compute conditional moments of diffusion processes:
Expected returns and variances of returns on nancial assets over nite time
intervals;
Use the method of moments to estimate a diffusion process from discretely
sampled data;
Compute prices of derivatives.
One approach is to reduce the problem to a PDE, which can sometimes be
solved analytically.

Kolmogorov Backward Equation
Diffusion process X
t
with coefcients (t ,X)and (t ,X ).
Objective: compute a conditional expectation
f (t ,X) =E[g(X
T
)|X
t
=X ]
Suppose f (t ,X )is a smooth function of t and X . By the law of iterated
expectations,
f (t ,X
t
) =E
t
[f (t +dt ,X
t +dt
)] E
t
[df (t ,X
t
)]=0
Using Itos Lemma,
E
t
[df (t ,X
t
)]=
f (t ,X )
+(t ,X )
f (t ,X )
+
1
(t ,X)
2
2
f (t ,X )
dt =0
t X 2 X
2
Kolmogorov backward equation
f (t ,X )
+(t ,X)
f (t ,X)
+
1
(t ,X )
2
2
f (t ,X )
=0,
t X 2 X
2
with boundary condition
f (T ,X) =g(X)
Example: Square-Root Diffusion
Consider a popular diffusion process used to model interest rates and
stochastic volatility
dX
t
= (X
t
X )dt + X
t
dZ
t
We want to compute the conditional moments of this process, to be used as
a part of GMM estimation.
Compute the second non-central moment
f (t ,X) =E(X
T
2
|X
t
=X)
Using Kolmogorov backward equation,
f (t ,X )
(X X )
f (t ,X )
+
1
2
X
2
f (t ,X )
=0,
t X 2 X
2
with boundary condition
f (T ,X) =X
2

Look for the solution in the the form
f (t ,X ) =a
0
(t ) +a
1
(t )X +
a
2
(t )
X
2
2
Substitute f (t ,X )into the PDE
a
0
(t ) +a
1
(t )X +
a
2
(t )
X
2
(X X )(a
1
(t ) +a
2
(t )X ) +
2
Xa
2
(t ) =0
2 2
Collect terms with different powers of X , zero, one and two, to get
a
0
(t ) +Xa
1
(t ) =0
2
a
1
(t ) a
1
(t ) + +X a
2
(t ) =0
2
a
2
(t ) 2a
2
(t ) =0
with initial conditions
a
0
(T ) =a
1
(T ) =0, a
2
(T ) =2
We solve the system of equations starting from the third one and working up
to the rst:
a
0
(t )=XA
1
2
e
2(t T )
e
(t T )
+
1
2

a
1
(t )=A e
(t T )
e
2(t T )
a
2
(t )=2e
2(t T )
A =
2
+2X
Compare the exact expression above to an approximate expression, obtained

by assuming that T t =t is small.

Assume T t =t is small. Using Taylor expansion,
a
0
(t ) =o(t ), a
1
(t ) =At +o(t )
a
2
(t ) =2(1 2t ) +o(t )
Then

E
t
(X
t
2
+t
|X
t
=X )X
2
+t (
2
+2X )X 2X
2
Alternatively,
X
t +t
X
t
(X
t
X )t +
X
t
t
t
,
t
N(0,1)
Therefore
E
t
(X
t
2
+t
|X
t
=X )X
2
+t
2
X 2X (X X )
Outline
1
Stochastic Integral
2
Its Lemma
3
Black-Scholes Model
4
5
SDEs
6
SDEs and PDEs
7
8
Risk-Neutral Probability Measure
Under the risk-neutral probability measure Q, expected conditional asset
returns must equal the risk-free rate.
Alternatively, using the discounted cash ow formula, the price P
t
of an asset
with payoff H
T
at time T is given by

T

P
t
t
exp r
s
ds H
T
=E
Q
t
where r
s
is the instantaneous risk-free interest rate at time s.
The DCF formula under the risk-neutral probability can be used to compute
asset prices by Monte Carlo simulation (e.g., using an Euler scheme to
approximate the solutions of SDEs).
Alternatively, one can derive a PDE characterizing asset prices using the
connection between PDEs and SDEs.

Change of Measure
We often want to consider a probability measure Qdifferent from P, but the
one that agrees with Pon which events have zero probability (equivalent to
P) (e.g., Qcould be a risk-neutral measure).
Different probability measures assign different relative likelihoods to the
trajectories of the Brownian motion.
It is easy to express a new probability measure Qusing its density
dQ
T
=
dP
T
For any random variable X
T
,
E
Q
0
[X
T
] =E
P
0
[
T
X
T
], E
Q
t
[X
T
] =E
P
T
X
T
,
t
E
P
t
[
T
]
t
t
Qis equivalent to Pif
T
is positive (with probability one).
Change of Measure
If we consider a probability measure Qdifferent from P, but the one that
agrees with Pon which events have zero probability, then the P-Brownian
motion Z
t
P
becomes an Ito process under Q:
dZ
t
P
=dZ
t
Q
t
dt
for some
t
. Z
t
Q
is a Brownian motion under Q.
When we change probability measures this way, only the drift of the Brownian
motion changes, not the variance.
Intuition: a probability measure assigns relative likelihood to different
trajectories of the Brownian motion. Variance of the Ito process can be
recovered from the shape of a single trajectory (quadratic variation), so it
does not depend on the relative likelihood of the trajectories, hence, does not
depend on the choice of the probability measure.
Under the risk-neutral probability measure, expected conditional asset
Start with the stock price process under P:
dS
t
=
t
S
t
dt +
t
S
t
dZ
t
P
Under the risk-neutral measure Q,
dS
t
=r
t
S
t
dt +
t
S
t
dZ
t
Q
Thus, if dZ
t
P
=
t
dt +dZ
t
Q
,
t
=r
t
t
is the price of risk.
The risk-neutral measure Qis such that the process Z
t
Q
dened by
dZ
t
Q
=
t
r
t
dt +dZ
t
P
t

Under the risk-neutral probability measure, expected conditional asset
We conclude that for any asset (paying no dividends), the conditional risk
premium is given by
E
P
dS
t
r
t
dt =E
P
dS
t
E
Q
dS
t
t t t
S
t
S
t
S
t
Thus, mathematically, the risk premium is the difference between expected
returns under the Pand Qprobabilities.
If we want to connect Qto Pexplicitly, how can we compute the density,
dQ/dP?
In discrete time, the density was conditionally lognormal.
The density
T
= (dQ/dP)
T
is given by

t
t
=exp
u
dZ
u
P
2
u
du , 0 t T
2
0 0
The state-price density is given by
t
=exp r
u
du
t
0
The reverse is true: if we dene
T
as above, for any process
t
satisfying
certain regularity conditions (e.g.,
t
is bounded, or satises the Novikovs
condition as in Back 2005, Appendix B.1), then measure Qis equivalent to P
and
t
Z
t
Q
=Z
t
P
+
u
du
0
Risk-Neutral Probability and Arbitrage
If there exists a risk-neutral probability measure, then the model is
arbitrage-free.
If there exists a unique risk-neutral probability measure in a model, then all
options are redundant and can be replicated by trading in the underlying
assets and the risk-free bond.
A convenient way to build arbitrage-free models is to describe them directly
under the risk-neutral probability.
One does not need to describe the Pmeasure explicitly to specify an
arbitrage-free model.
However, to estimate models using historical data, particularly, to estimate
risk premia, one must specify the price of risk, i.e., the link between Qand P.
Outline
1
Stochastic Integral
2
Its Lemma
3
Black-Scholes Model
4
5
SDEs
6
SDEs and PDEs
7
8

Black-Scholes Model
Assume that the stock pays no dividends and the stock price follows
dS
t
=dt +dZ
t
P
S
t
Assume that the interest rate is constant, r .
Under the risk-neutral probability Q, the stock price process is
dS
t
=r dt +dZ
t
Q
S
t
Terminal stock price S
T
is lognormally distributed:
ln S
T
=ln S
0
+ r
2
T +
T
Q
,
Q
N(0,1)
2
Price of any European option with payoff H(S
T
)can be computed as
=E
Q
e P
t
t
r (T t )
H(S
T
)
Term Structure of Interest Rates
Consider the Vasicek model of bond prices.
A single-factor arbitrage-free model.
To guarantee that the model is arbitrage-free, build it under the risk-neutral
probability measure.
Assume the short-term risk-free rate process under Q
dr
t
= (r
t
r )dt +dZ
t
Q
Price of a pure discount bond maturing at T is given by

T
P(t ,T ) =E
Q
t
exp r
s
ds
t
Characterize P(t ,T )as a solution of a PDE.

Look for P(t ,T ) =f (t ,r
t
).
Using Itos lemma,
E
Q
t
[df (t ,r
t
)]=
f (t ,r
t
)
(r
t
r )
f (t ,r
t
)
+
1
2
f (t
2
,r
t
)
dt
t r
t
2 r
t
Risk-neutral pricing requires that
E
Q
t
[df (t ,r
t
)]=r
t
f (t ,r
t
)dt
and therefore f (t ,r
t
)must satisfy the PDE
f (t ,r )
(r r )
f (t ,r )
+
1
2
f (t ,r )
=rf (t ,r )
t r 2 r
2
with the boundary condition
f (T ,r ) =1

Look for the solution in the form
f (t ,r
t
) =exp (a(T t ) b(T t )r
t
)
Derive a system of ODEs on a(t )and b(t )to nd
a(T t ) =r (T t )
r
1 e
(T t )
2(T t ) e
2(T t )
+4e
(T t )
3
4
3
1
(T t )
b(T t ) = 1 e

Assume a constant price of risk . What does this imply for the interest rate
process under the physical measure Pand for the bond risk premia?
Use the relation
dZ
t
P
= dt +dZ
t
Q
to derive
dr
t
= (r
t
r )dt +dt +dZ
t
P
= r
t
r +
dt +dZ
t
P
Expected bond returns satisfy

E
P
dP(t ,T )
= (r
t
+
P
t
)dt
t
P(t ,T )
where
P
t
=
1 f (t ,r
t
)
= b(T t )
f (t ,r
t
) r
t
|b(T t )|is the volatility of bond returns.
Equity Options with Stochastic Volatility
Consider again the Hestons model. Assume that under the risk-neutral
probability Q, stock price is given by
d ln S
t
= (r
1
v
t
)dt +
v
t
dZ
t
Q,S
2
dv
t
= (v
t
v)dt +
v
t
dZ
t
Q,S
+
1
2
v
t
dZ
t
Q,v
dZ
t
Q,S
dZ
t
Q,v
=0
Z
t
Q,v
models volatility shocks uncorrelated with stock returns.
Constant interest rate r .
The price of a European option with a payoff H(S
T
)can be computed as
P
t
=E
Q
t
[exp (r (T t ))H(S
T
)]
We havent said anything about the physical process for stochastic volatility.
In particular, how is volatility risk priced?

In the Hestons model, assume that the price of volatility risk is constant,
v
,
and the price of stock price risk is constant,
S
.
Then, under P, stock returns follow
d ln S
t
= r +
S
v
t
2
1
v
t
dt +
v
t
dZ
t
P,S
dv
t
=
(v
t
v) +
v
t
S
+
1
2
dt +
v
t
dZ
t
P,S
+
1
2
v
t
dZ
t
P,v
dZ
t
P,S
dZ
t
P,v
=0
We have used
dZ
t
P,S
=
S
dt +dZ
t
Q,S
dZ
t
P,v
=
v
dt +dZ
t
Q,v
Conditional expected excess stock return is
E
P
dS
t
r dt
=E
P
d ln S
t
+
v
t
dt r dt
v
t
dt
t t
S
t
2
Our assumptions regarding the market prices of risk translate directly into
implications for return predictability.
For stock returns, our assumption of constant price of risk predicts a
nonlinear pattern in excess returns: expected excess stock returns
proportional to conditional volatility.
Suppose we construct a position in options with the exposure
t
to stochastic
volatility shocks and no exposure to the stock price:
dW
t
= [...]dt +
t
dZ
t
P,v
Then the conditional expected gain on such a position is
(W
t
r +
t
v
)dt
Hestons Model of Stochastic Volatility
Assume that under the risk-neutral probability Q, stock price is given by
d ln S
t
= (r
1
v
t
)dt +
v
t
dZ
t
Q,S
2
dv
t
= (v
t
v)dt +
v
t
dZ
t
Q,S
+
1
2
v
t
dZ
t
Q,v
dZ
t
Q,S
dZ
t
Q,v
=0
Z
Q,v
models volatility shocks uncorrelated with stock returns.
t
Constant interest rate r .
The price of a European option with a payoff H(S
T
)can be computed as
=E
Q
[exp (r (T t ))H(S
T
)] P
t
t
Assume that the price of volatility risk is constant,
v
, and the price of stock
price risk is constant,
S
.
Variance Swap in Hestons Model
Consider a variance swap, paying
T
(d ln S
u
)
2
K
t
2
t
at time T . What should be the strike price of the swap, K
t
, to make sure that
the market value of the swap at time t is zero?
Using the result on quadratic variation,
(d ln S
t
)
2
=v
t
dt
the strike price must be such that
e
r (T t )
E
Q
t
v
u
du K
t
2
=0
t
Need to compute
K
2
=E
t
Q
t
t
v
u
du

Variance Swap in Hestons Model
Since
v
u
=v
t
u
(v
s
v)ds +
u
v
s
dZ
s
Q,S
+
1
2
u
v
s
dZ
s
Q,v
t t t
we nd that
u u
E
Q
t
[v
u
] =v
t
E
Q
t
(v
s
v)ds =v
t
(E
Q
t
[v
s
] v)ds
t t
Solving the above equation for E
Q
t
[v
u
], we nd
E
Q
t
[v
u
] =v +e
(ut )
(v
t
v)
We obtain the strike price
1

K
t
2
=E
Q
t
v
u
du =v(T t ) + (v
t
v) 1 e
(T t )
t

Expected Prot/Loss on a Variance Swap
To compute expected prot/loss on a variance swap, we need to evaluate
E
P
(d ln S
u
)
2
K
2
t t
t
Instantaneous expected gain on a long position in the swap is easy to
compute in closed form.
The market value of a swap starts at 0 at time t , and at s >t becomes

s
P
s
E
Q
e
r (T s)
v
u
du + v
u
du K
2
s t
t s
s
=e
r (T s)
v
u
du +K
s
2
K
t
2
t
We conclude that the instantaneous gain on the long swap position at time s
equals
[...]ds +
1
e
r (T s)
1 e
(T s)
dv
s

Expected Prot/Loss on a Variance Swap
We conclude that the instantaneous gain on the long swap position at time s
equals
[...]ds +
1
e
r (T s)
1 e
(T s)
dv
s
Given our assumed market prices of risk,

S
and
v
, the time-s expected
instantaneous gain on the swap opened at time t is
r (T s) (T s)
P
s
r ds +
v
s
1
e
1 e
S
+
1
2
ds

Summary
Risk-neutral pricing is a convenient framework for developing arbitrage-free
pricing models.
Connection to classical results: risk-neutral expectation can be characterized
by a PDE.
Risk premium on an asset is the difference between expected return under P
and under Qprobability measures.
Readings
Back 2005, Sections 2.1-2.6, 2.8-2.9, 2.11, 13.2, 13.3, Appendix B.1.
MIT OpenCourseWare
http://ocw.mit.edu
15.450 Analytics of Finance
Fall 2010
For information about citing these materials or our Terms of Use, visit: http://ocw.mit.edu/terms .

MIT15 450F10 Lec02

Uploaded by

Copyright:

Available Formats

You might also like

MIT15 450F10 Lec02

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

MIT15 450F10 Lec02

Uploaded by

Copyright:

Available Formats

Stochastic Integral Its Lemma Black-Scholes Model Multivariate It Processes SDEs SDEs and PDEs Risk-Neutral Probability Risk-Neutral

Note: a sufcient condition for E

dollars in the bond maturing at

Compare the exact expression above to an approximate expression, obtained

c Leonid Kogan ( MIT, Sloan ) Stochastic Calculus 15.450, Fall 2010 63 / 74

Expected bond returns satisfy

c Leonid Kogan ( MIT, Sloan ) Stochastic Calculus 15.450, Fall 2010 70 / 74

c Leonid Kogan ( MIT, Sloan ) Stochastic Calculus 15.450, Fall 2010 71 / 74

Given our assumed market prices of risk,

c Leonid Kogan ( MIT, Sloan ) Stochastic Calculus 15.450, Fall 2010 72 / 74

You might also like