John H.

Cochrane
1
. Revised October 4, 2013
2 Continuous Time Summary/Review
We need a statistical model that can apply to stock prices, returns, etc. The point of these notes
is to quickly present the standard diusion model that we use in asset pricing.
We use continuous time when its simpler. In point of fact all our data are discretely sampled.
Whether time is really continuous or discrete is an issue well leave to physicists and philosophers.
Most of the approximations that took a lot of eort in Chapter 1 of Asset pricing work much more
easily in continuous time.
2.1 Preview/summary:
2.1.1 Continuous time
1. Brownian motion
.
t+
.
t
(0, )
2. Dierential
d.
t
= lim
&0
(.
t+
.
t
)
3. dz and dt
d.
t
O(

dt)
d.
2
t
= dt
1
t
(d.
t
) = 0
ar
t
(d.
t
) = 1
t
(d.
2
t
) = d.
2
t
= dt
4. Diusions
dr
t
= j(r
t
, t)dt + o(r
t
, t)d.
t
1
t
(dr
t
) = j(r
t
, t)dt
o
2
t
(dr
t
) = o
2
(r
t
, t)dt
5. Examples
dj
t
j
t
= jdt + od.
t
dr
t
= c(r
t
j)dt + od.
t
1
Please report typos to john.cochrane@chicagobooth.edu. Make sure you have the latest version of the document.
If you can report the context of the typo and not just the page number that will help me to nd it more quickly.
9
6. Itos lemma
dr
t
= jdt + od.
t
; j
t
= )(t, r
t
) =
dj
t
=
0)
0t
dt +
0)
0r
dr
t
+
1
2
0
2
)
0r
2
dr
2
t
dj
t
=

0)
0t
+
0)
0r
j +
1
2
0
2
)
0r
2
o
2

dt +

0)
0r
o

d.
7. Stochastic calculus: do second-order expansions, keep the d. and dt terms.
2.1.2 Solving stochastic dierential equations
1. Stochastic integral add up the changes
.
T
.
0
=
Z
T
t=0
d.
t
.
T
.
0
=
T
X
t=1
-
t
the result is just a way of saying .
T
is a normally distributed random variable
Z
T
t=0
d.
t
= .
T
.
0
(0, T)
2. Example 1: diusion
dr
t
= jdt + od.
t
Z
T
t=0
dr
t
= j
Z
T
t=0
dt + o
Z
T
t=0
d.
t
r
T
r
0
= jT + o (.
T
.
0
)
3. Example 2: lognormal diusion
dr
t
r
t
= jdt + od.
t
d lnr
t
=

j
1
2
o
2

dt + od.
t
lnr
T
lnr
0
=

j
1
2
o
2

T + o
Z
T
0
d.
t
r
T
= r
0
c
(j
1
2
o
2
)T+o

0
o:
4. Example 3: AR(1)
dr
t
= c(r
t
j)dt + od.
t
r
T
j = c
T
(r
0
j) + o
Z
T
t=0
c
(Tt)
d.
Tt
10
5. Continuous-time MA processes
r
T
=
Z
T
t=0
n(T t)d.
t
6. With solutions, you can nd moments. Examples
(a) Lognormal diusion
r
T
= r
0
c
(j
1
2
o
2
)T+o

0
o:

1(r
T
) = r
0
c
(j
1
2
o
2
)T+
1
2
o
2
T
= r
0
c
jT
(b) AR(1)
(r
t
j) = c
t
(r
0
j) +
Z
t
c=0
c
(tc)
od.
c
1
0
(r
t
j) = c
t
(r
0
j)
o
2
(r
t
j) =
Z
t
c=0
c
2(tc)
o
2
d: =
1 c
2t
2c
o
2
(c) Continuous-time MA.
r
t
=
Z
t
c=0
n(t :)d.
c
,
1
0
(r
t
) = 1
0
Z
t
c=0
n(t :)d.
c

=
Z
t
c=0
n(t :)1
c
(d.
c
) = 0
1
0
(r
2
t
) = 1
0
Z
t
c=0
n(t :)d.
c

2
=
Z
t
c=0
n(t :)
2
d:.
2.1.3 Finding Moments and Densities
1. Backward equation. What is
)(r
t
, t) = 1
t
[c(r
T
)]?
We can work backwards using Itos lemma to nd
0)(r
,
t)
0t
+
0)(r
,
t)
0r
j(r
,
t) +
1
2
0
2
)(r
,
t)
0r
2
o
2
(r
,
t) = 0
Then solve this partial dierential equation with boundary condition
)(r, T) = c(r).
2. Densities and the Forward Equation.
11
(a) Dene )(r, t|r
0
) = the probability density of r
t
given r
0
at time 0. Its a delta function
at t = 0, r = r
0
. Looking forward it solves
0)(r, t|r
0
)
0t
+
0 [j(r))(r, t|r
0
)]
0r
=
1
2
0
2

o
2
(r))(r, t|r
0
)

0r
2
(b) This fact gives a nice formula for the stationary (unconditional) density. The formula is:
)(r) = /
c

2
()

2
()
oc
o
2
(r)
where / is a constant, so we integrate to one. Evaluating /,
)(r) =

Z
c

2
()

2
()
oc
o
2
(:)
d:

1
c

2
()

2
()
oc
o
2
(r)
2.1.4 Asset pricing in continuous time
1. Present value formula
j
t
= 1
t
Z

c=0
c
cc
n
0
(c
t+c
)
n
0
(c
t
)
r
t+c
d: = 1
t
Z

c=0

t+c

t
r
t+c
d:
2. Whats a return?
d1
t
=
dj
t
j
t
+
r
t
j
t
dt
3. Whats a cumulative value process
d\
t
\
t
= d1
t
4. Whats a riskfree rate?
r
)
t
dt
5. Whats an excess return?
d1
t
r
)
t
dt =
dj
t
j
t
+
r
t
j
t
dt r
)
t
dt
6. Whats the equivalent of 1 = 1(:1)?
1
t

d (
t
j
t
)

t
j
t

r
t
j
t
dt = 0
or 1
t
[d (
t
\
t
)] = 0
12
7. Whats the equivalent of 1
)
= 1,1(:)?
r
)
dt = 1
t

d
t

8. Whats the equivalent of 1(1

c
) = 1
)
co(:, 1
c
)?
1
t
(d1
t
) r
)
dt = 1
t

d
t

t
d1
t

9. Whats the equivalent of :

t+1
1 c c
t+1
?
d
t

t
= cdt
dc
t
c
t
+
1
2
( + 1)
dc
2
t
c
t
d
t

t
= cdt d [log c
t
] +
1
2

2
d [log c
t
]
2
10. Whats the equivalent of 1
)
c + 1(c)?
r
)
= c + j
c

1
2

2
o
2
c
where d log c
t
= j
c
dt + o
c
d.
t
11. Whats the equivalent of 1(1
c
) co(1
c
, c)?
1
t
(d1
t
) r
)
dt = 1
t

dc
t
c
t
d1
t

= co
t

dc
t
c
t
, d1
t

2.2 A quick discrete time review

Ill explain continuous time by analogy with discrete time. This is a quick review of the essential
concepts cant hurt.
1. A sequence of i.i.d. random variables (like a coin ip) forms the basic building block of time
series.
-
t
i.i.d.
i.i.d. means independent and identically distributed. In particular, it implies that the
conditional mean and variance are constant through time. Our building block series has a
zero mean,
1
t
(-
t+1
) = 0
o
t
(-
t+1
) = o
2
= const.
The concept of conditional mean and variance is really important here! 1
t
means expec-
tation conditional on all information at time t. We sometimes specify that -
t
are normally
distributed, but they dont have to be.
13
2. We build more interesting time series models from this building block. The AR(1) is the
rst canonical example
r
t+1
= jr
t
+ -
t+1
.
The sequence of conditional means follows
1
t
r
t+1
= 1
t
(jr
t
+ -
t+1
) = jr
t
1
t
r
t+2
= 1
t
(jr
t+1
+ -
t+2
) = j
2
r
t
1
t
r
t+I
= j
I
r
t
.
The AR(1) displays some persistence since after a shock it is expected to stay up for a while,
and then decay gradually back again. A good exercise at this point is to work out the sequence
of conditional variances of the AR(1). The answer is
o
2
t
(r
t+1
) = o
2
.
o
2
t
(r
t+2
) = (1 + j
2
)o
2
.
o
2
t
(r
t+I
) = (1 + j
2
+ ... + j
2(I1)
)o
2
.
3. We can also write the AR(1) as
r
t+1
= jr
t
+ o-
t+1
; o
2
.
= 1
The earlier notation is more common in discrete time; the latter notation is more common
as we move to continuous time. In either case, note 1
t
-
t+1
= 0 so its uncorrelated with r
t
.
Thus you can estimate an AR(1) by an OLS forecasting regression.
4. The canonical example 2 is the MA(1).
r
t+1
= -
t+1
+ 0-
t
.
Its sequence of conditional means follows
1
t
(r
t+1
) = 0-
t
1
t
(r
t+2
) = 0
1
t
(r
t+I
) = 0.
It displays some persistence too but only for one period. Problem: work out its conditional
variances.
5. You can transform from AR to MA representations. You can think of this operation as
solving the AR(1)
r
t
= jr
t1
+ -
t
r
t
= j (jr
t2
+ -
t1
) + -
t
= j
2
r
t2
+ j-
t1
+ -
t
r
t
= j
3
r
t3
+ j
2
-
t2
+ j-
t1
+ -
t
r
t
= j
t
r
0
+
t1
X
)=0
j
)
-
t)
14
Solving the dierence equation means, really, expressing r
t
as a random variable, given
time 0 information. Now we know r
t
has
1
0
(r
t
) = j
I
r
0
;
o
2
0
(r
t
) = (1 + j
2
+ .. + j
2(I1)
)o
2
.
.
If the - are normal, so is r
t
.
6. Continuing, we can write the AR(1) as
r
t
=

X
)=0
j
)
-
t)
Thus, an AR(1) is the same as an MA(). You can similarly write an MA(1) as an AR().
Choose which representation is easiest for what youre doing. An example: Lets nd the
unconditional mean and variance of the AR(1). One way to do this is with the MA()
representation
1(r
t
) =

X
)=0
j
)
1 (-
t)
) = 0.
ar (r
t
) = ar

X
)=0
j
)
-
t)

=

X
)=0
j
2)
o
2
.
=
o
2
.
1 j
2
Notice: the - are all uncorrelated with each other. Thats how we got rid of the covariance term
and all o
2
are the same. We can also nd the same quantities from the AR(1) representation
1(r
t
) = j1(r
t1
) + 0 1(r
t
) = 0
ar(r
t
) = j
2
ar(r
t1
) + o
2
.
ar(r
t
) =
o
2
.
1 j
2
.
To come to the latter conclusions, you have to understand that 1(r
t
) = 1(r
t1
) and ar(r
t
) =
ar(r
t1
).
7. All of the familiar tricks for linear ARMA models, including lag operator notation, have con-
tinuous time counterparts. It takes a little work to make the translation. If youre interested,
see my Continuous-time linear models Foundations and Trends in Finance 6 (2011), 165
219 DOI: 10.1561/0500000037. Its on my webpage,
http://faculty.chicagobooth.edu/john.cochrane/research/papers/continuous_time_linear_models_FT.pdf
2.3 Continuous time series
1. Preview. In discrete time, our building block is the i.i.d. (normal) shock -
t
with variance
o
2
(-
t
) = 1. We build up more complex time series from this building block with dierence
equation models such as the AR(1),
r
t+1
= jr
t
+ o-
t+1
15
We solve such dierence equations to represent r
t
as a function of its past shocks,
r
t
=
t1
X
)=0
j
)
o-
t)
+ j
t
r
0
.
Our task is to do the exact same sort of thing in continuous time.
2. Random walk-Brownian motion .
t
.
(a) In discrete time, a random walk .
t
is dened as the sum of independent shocks,
.
t
.
0
=
t
X
)=1
-
)
.
Obviously, we recover -
t
as dierences of .
t
.
.
t
= (.
t
.
t1
) = -
t
.
It will turn out to be easier to think rst about .
t
in continuous time and then take
dierences than to think directly about -
t
(b) Now, lets examine the properties of .
t
. The variance of . grows linearly with horizon.
.
t+2
.
t
= -
t+1
+ -
t+2
1 (.
t+2
.
t
) = 0
ar(.
t+2
.
t
) = 2
.
t+I
.
t
= -
t+1
+ -
t+2
+ .. + -
t+I
1 (.
t+I
.
t
) = 0
ar(.
t+I
.
t
) = /
(remember, I set o
.
= 1.)
(c) To think about continuous time, just generalize this idea for any time interval. Lets
dene the process .
t
as one with normal increments,
.
t+
.
t
(0, )
for any , not just integers. This is a Brownian motion, the continuous-time version of
a random walk.
(d) I.i.d. property In discrete time 1(-
t
-
t+1
) = 0, i.e. 1(.
t+2
.
t+1
, .
t+1
.
t
) = 0. The nat-
ural generalization is 1(.
t+
.
t
, .
t++j
.
t+
) = 0, or more generally Nonoverlapping
dierences of {.
t
} are uncorrelated with each other.
3. The dierential dz
16
(a) Now lets take the derivative as we took the dierence -
t
= .
t
.
t1
. Dene
d.
t
= lim
&0
(.
t+
.
t
)
d.
t
-
t
fundamental building block
d is a tiny forward-dierence operator.
(b) d. is of size

dt. What does this mean? Note

o
2
(.
t+
.
t
) = ;
o(.
t+
.
t
) =

Thus, d.
t
is of order (typical size)

dt! This means that sample paths of .

t
are contin-
uous but not dierentiable. d. makes sense, but d.,dt does not. d.
t
is of order

dt so
d.,dt +, , moving innitely fast (up and down). The .
t
process is fractal; its
jumpy at any time scale; it has an innite path length.
(c) d., d.
t
, and d.(t) are equivalent notations.
(d) Moments:
1
t
(d.
t
) = 0; 1
t
(.
t+
.
t
) = 0
ar
t
(d.
t
) = dt; ar(.
t+
.
t
) = ,
co(d.
t
, d.
c
) = 0 : 6= t co(.
t+
.
t
, .
c+
.
c
) = 0
Watch out for notation. 1
t
(d.
t
) 6= d.
t
! d is a forward dierence operator, so 1
t
(d.
t
)
means expected value of how much d. will change in the next instant. It really means
1
t
(.
t+
.
t
), which is obviously not the same thing as .
t+
.
t
(e) d.
2
t
= dt. Variance and second moment are the same.
ar
t
(d.
t
) = 1
t
(d.
2
t
) = d.
2
= dt.
Second moments are nonstochastic! Its not just d.
2
is of order dt, but in fact d.
2
=
dt. We often write 1(d.
2
) to remind ourselves that for any discrete interval these are
second moments of random variables. But in the limit, in fact, second moments are
nonstochastic. Similarly if we have two Brownians d. and dn, they can be correlated
and the cross product is
co
t
(d.
t
, dn
t
) = 1
t
(d.
t
dn
t
) = d.
t
dn
t
.
(f) (Optional note.) The fact that squares of random variables are deterministic is hard to
swallow. If d.
t
is a normal with standard deviation

dt, why is d.
2
t
a number, not a

2
distributed random variables? To see why, compare d.
t
and d.
2
t
. .
t+
.
t
is (0, ),
(.
t+
.
t
) ,

(0, 1) so the probability that, say, k.

t+
.
t
k 2

is 5%. The
typical size of a movement .
t+
.
t
goes to zero at rate

On the other hand, (.

t+
.
t
)
2
,
2
1
. (
2
1
or chi-squared with one degree of
freedom just means the distribution of the square of a standard normal.) The fact
that the distribution (.
t+
.
t
)
2
, stays the same as gets small means that the
17
probability that, say, (.
t+
.
t
)
2
2 is xed as gets small. Thus (.
t+
.
t
)
2
goes to zero at rate . The typical size of a movement (.
t+
.
t
)
2
or the probability
that it exceeds any value goes to zero at rate . Things whose movements go to zero at
rate are dierentiable, and hence nonstochastic. (.
t+
.
t
) is of order dt and hence
nonstochastic.
4. Diusion processes. We have the building block, the analog to -
t
. Now, we build more
complex process like 1(1), ARMA, etc. in the same way as we do in discrete time.
(a) Random walk with drift.
discrete: r
t+1
= j + r
t
+ o-
t+1
r
t+1
r
t
= j + o-
t
continuous: dr
t
= jdt + od.
t
i. Moments of the random walk with drift, which gives you some practice in d. and
dt. The mean is
1
t
(dr
t
) = jdt + o1
t
(d.
t
) = jdt
For the variance, notice
dr
2
t
= (dr
t
)
2
= j
2
dt
2
+ o
2
d.
2
+ 2jdtd.
t
= o
2
dt
The rst parenthesis clears up notation. In calculus, you learned to keep terms of
order dt, and ignore terms of order dt
2
and so on, because they get arbitrarily small
as the time interval shrinks. The new rule is, we keep terms of order d. =

dt and
dt, but we ignore terms of order dt
32
, dt
2
, etc. just as in regular calculus. So terms
like dt
2
and dtd. drop out. Only the o
2
dt term remains. So, the variance is
ar
t
(dr
t
) = 1
t
(dr
2
t
) = dr
2
t
= o
2
dt
ii. The mean drops out of the variance. Loosely, diusions vary so much that variances
overwhelm means and it doesnt matter if you take the mean out when you take
variance or not. You might have objected, and wanted to dene
ar
t
(dr
t
) = 1
t
n
[dr
t
1
t
(dr
t
)]
2
o
But watch:
1
t
n
[dr
t
1
t
(dr
t
)]
2
o
= 1
t
n
[dr
t
jdt]
2
o
= 1
t
n
[od.
t
]
2
o
= o
2
1
t

d.
2
t

= o
2
dt
We get the same result whether we take the mean out or not. The d.
t
term is of
order

dt, so almost always innitely bigger than the dt term.

(b) Geometric Brownian motion with drift. This is the standard asset price process, so
worth careful attention.
dj
t
j
t
= jdt + od.
t
dj
t
is the change in price, so dj
t
,j
t
is the instantaneous percent (not annualized) change
in price or instantaneous rate of return if there are no dividends. Well see how to add
dividends in a moment.
18
i. The moments are
1
t

dj
t
j
t

= jdt
ar
t

dj
t
j
t

= o
2
dt
so j and o can represent the mean and standard deviation of the arithmetic percent
rate of return. Since they multiply dt, they have annual units. j = 0.05 and
o
2
= 0.10
2
are numbers to use for a 5% mean return and 10% standard deviation of
return.
ii. Dont forget the dt at the end of these expressions
1
t

dj
t
j
t

= j, ar
t

dj
t
j
t

= o
2
is tempting but incorrect. Its ok to write
1
dt
1
t

dj
t
j
t

= j,
1
dt
ar
t

dj
t
j
t

= o
2
but its better to write
1
t

dj
t
j
t

= jdt, ar
t

dj
t
j
t

= o
2
dt
iii. Sticklers like to write this process
dj
t
= jj
t
dt + oj
t
d.
t
so it is in the canonical form
dj
t
= j(j
t
, t, ..)dt + o(j
t
, t, ..)d.
t
.
iv. You can also see here the reason we call it a stochastic dierential equation.
Without the d.
t
term, we would solve this equation to say
j
t
= j
0
c
jt
.
i.e. value grows exponentially. We will come back and similarly solve the stochastic
dierential equation with the d.
t
term.
(c) AR(1), Ornstein Uhlenbeck process
discrete: r
t+1
= (1 j)j + jr
t
+ o-
t+1
r
t+1
r
t
= (1 j)(r
t
j) + o-
t+1
1
t
(r
t+1
r
t
) = (1 j)(r
t
j)
ar
t
(r
t+1
r
t
) = o
2
19
continuous : dr
t
= c(r
t
j)dt + od.
t
1
t
dr
t
= c(r
t
j)dt
ar
t
dr
t
= o
2
dt
The part in front of dt is called the drift and the part in front of d.
t
is called the
diusion coecient.
(d) In general, we build complex time series processes from the stochastic dierential equa-
tion
dr
t
= j(r
t
, t, ...)dt + o(r
t
, t, ...)d.
t
.
j and o may be nonlinear functions (ARMA in time series are just linear). Note we
often suppress state and time dependence and just write j, o or j(), o() or j
t
and o
t
.
1
t
(dr
t
) = j(r
t
, t, ..)dt
ar
t
(dr
t
) = o
2
(r
t
, t, ..)dt
(e) Simulation/approximation. We often use continuous time to derive approximations for
discrete time. We also will often want to simulate forward continuous time processes on
a computer. The Euler approximation is a natural way to do this. From
dr
t
= j(r
t
, t)dt + o(r
t
, t)d.
t
write
r
t+
r
t
= j(r
t
, t)+ o(r
t
, t)

-
t+
where is a short time interval and -
t+
(0, 1). For example, you could use
= 1,365 to simulate daily data.
5. Itos lemma. Given a diusion process for r
t
, we can construct a new process j
t
= )(r
t
).
How can we nd a diusion representation for j
t
given the diusion representation of r
t
?.
For example, what does the log price process log(j
t
) look like?
(a) Nonstochastic answer: Given
dr
t
dt
= j
we use the chain rule.
dj
t
=
0)
0r
dr
t
=
0)
0r
jdt
But that assumes dr,dt is meaningful, i.e. dr
t
= jdt. What if dr
t
= j
t
dt + od.
t
?
(b) So our question:
dr
t
= jdt + od.
t
j
t
= )(r
t
)
dj
t
= ?
Answer:
20
Do a second order Taylor expansion. Keep terms of order dt and d.
t
=

dt, ignore
higher order terms. Dont forget d.
2
t
= dt.
The answer is called Itos lemma:
dj
t
=
0)
0r
dr
t
+
1
2
0
2
)
0r
2
dr
2
t
dj
t
=

0)
0r
j +
1
2
0
2
)
0r
2
o
2

dt +

0)
0r
o

d.
Intuition: Jensens inequality says, 1 [j(r)] < j [1(r)] if j (r) is concave. 1 [n(c)] <
n[1(c)] for example. The second derivative term takes care of this fact. You will see
this in the convexity terms of option and bond pricing.
(c) An example. Geometric growth
dj
t
j
t
= jdt + od.
t
j
t
= lnj
t
dj
t
=
1
j
t
dj
t

1
2
1
j
2
dj
2
t
=

j
1
2
o
2

dt + od.
t
(d) Conversely,
dj
t
= jdt + od.
t
j
t
= c
j

dj
t
= c
j

dj
t
+
1
2
c
j

dj
2
t
dj
t
j
t
=

j +
1
2
o
2

dt + od.
t
(e) Another example. We will often multiply two diusions together. What is the product
diusion? Use the chain rule but go out to second derivatives.
d(r
t
j
t
) =
0(rj)
0r
dr
t
+
0(rj)
0j
dj
t
+
1
2
2
0
2
(rj)
0r0j
dr
t
dj
t
= j
t
dr
t
+ r
t
dj
t
+ dr
t
dj
t
The second partials with respect to x and y 0
2
(rj),0r
2
are zero. Notice the extra term
dr
t
dj
t
relative to the usual chain rule.
(f) Itos lemma more generally. If
j = )(r
t
, t)
then
dj
t
=
0)
0t
dt +
0)
0r
dr
t
+
1
2
0
2
)
0r
2
dr
2
t
dj
t
=

0)
0t
+
0)
0r
j +
1
2
0
2
)
0r
2
o
2

dt +

0)
0r
o

d.
The version with partial derivatives starting from j
t
= )(r
1t
, r
2t
, t) is obvious enough
and long enough that I wont even write it down.
21
2.4 Solving stochastic dierential equations.
1. The problem: We want to do the analogue of solving dierence equations. For example,
when we write an AR(1),
r
t
= jr
t1
+ -
t
we solve backward to
r
t
= j
t
r
0
+
t1
X
)=0
j
)
-
t)
.
How do we do this in continuous time?
2. Another motivation. If you have a dierential equation
dr
t
= jdt
you know how to solve this. Integrate both sides from 0 to T,
r
T
= r
0
+ j
Z
T
c=0
dt = r
0
+ jT.
So, how do you solve a stochastic dierential equation,
dr
t
= jdt + od.
t
?
3. A natural idea: Why dont we just integrate both sides,
Z
T
t=0
dr
t
= j
Z
T
t=0
dt + o
Z
T
t=0
d.
t
.
What meaning can we give to the integrals? Why not just add up all the little changes.
Z
T
0
d.
t
= (.

.
0
) + (.
2
.

) + (.
3
.
2
) + .... + (.
T
.
T
) = .
T
.
0
.
Then our solution reads
r
T
r
0
= jT + o (.
T
.
0
) .
Now, we dened (.
T
.
0
) to mean a normally distributed random variable with mean 0 and
variance T. So this solution means, r
T
r
0
is a normally distributed random variable with
mean jT and variance o
2
T. Similarly, if d.
t
are the little changes, you get . back if you add
them up.
4. In sum, the building block of solutions is the opposite of our d.
t
dierencing operator, and
completes the analogy we started when dening .
T
in the rst place
.
T
.
0
=
Z
T
t=0
d.
t
.
T
.
0
=
T
X
t=1
-
t
22
R
T
t=0
d.
t
is called a Stochastic integral. This is a fundamentally dierent denition of integral
which gives mathematicians a lot to play with. For us, it just means add up all the little
changes. Recall that .
t
is not dierentiable. Thats why we do not write the usual integral
notation
Yes: .
T
.
0
=
Z
T
0
d.
t
No : .
T
.
0
=
Z
T
0

d.(t)
dt

dt
We do not dene the integral as the area under the curve as you do in regular calculus.
What we mean by
R
T
0
d.
t
in the end is just a normally distributed random variable with
mean zero and variance t.
Z
T
0
d.
t
= .
T
.
0
(0, T)
5. Now, to more complex stochastic dierential equations. In discrete time, we solve the
dierence equation:
r
t
= (1 j)j + jr
t1
+ -
t
r
t
j = j (r
t1
j) + -
t
r
T
j = j
T
(r
0
j) +
T1
X
)=0
j
)
-
T)
.
This means we know the conditional distribution of the random variable r
T
, and really thats
what it means to have solved the stochastic dierence equation.
r
T
|r
0

j + j
T
(r
0
j) , o
2
.
T1
X
)=0
j
2)

Lets do the same thing in continuous time.

6. Ill just do some examples. I wont use anything fancier than the basic idea: plop
R
T
t=0
on
both sides of an expression and interpret the results.
(a) Random walk. If we start with d.
t
and integrate both sides,
.
T
.
0
=
Z
T
t=0
d.
t
.
T
.
0
(0, T)
23
(b) Random walk with drift
dr
t
= jdt + od.
t
Z
T
t=0
dr
t
=
Z
T
t=0
jdt +
Z
T
t=0
od.
t
r
T
r
0
= jT + o
Z
T
t=0
d.
t
r
T
r
0
= jT + -; - (0, To
2
)
r
T
r
0
(jT, To
2
)
As in discrete time,
r
T
= r
0
+ jT +
T
X
t=1
-
t
(c) Lognormal price process.
dj
t
j
t
= jdt + od.
t
By Itos lemma
d lnj
t
=

j
1
2
o
2

dt + od.
t
now integrate both sides,
Z
T
0
d lnj
t
=
Z
T
t=0

j
1
2
o
2

dt +
Z
T
t=0
od.
t
lnj
T
lnj
0
=

j
1
2
o
2

T + o
Z
T
t=0
d.
t
j
T
= j
0
c
(j
1
2
o
2
)T+o

=0
o:
i.e,
j
T
= j
0
c
(j
1
2
o
2
)T+o

T.
- (0, 1)
j
T
is lognormally distributed. (Lognormal means log of j
T
is normally distributed.)
A geometric diusion, where the arithmetic return is instantaneously normal, means
that the nite-period return is lognormally distributed.
(d) AR(1)
dr
t
= c(r
t
j)dt + od.
t
The solution is,
r
T
j = c
T
(r
0
j) + o
Z
T
c=0
c
c
d.
Tc
this looks just like the discrete time case
r
T
j = j
T
(r
0
j) +
T1
X
)=0
j
)
-
t)
.
24
i. To derive this answer, you have to be a little clever. Find
d

c
t
(r
t
j)

= cc
t
(r
t
j) dt + c
t
dr
t
= cc
t
(r
t
j) dt + c
t
[c(r
t
j)dt + od.
t
]
= oc
t
d.
t
Now integrate both sides,
Z
T
t=0
d

c
t
(r
t
j)

=
Z
T
t=0
oc
t
d.
t
c
T
(r
T
j) (r
0
j) = o
Z
T
t=0
c
t
d.
t
(r
T
j) = c
T
(r
0
j) + o
Z
T
t=0
c
(tT)
d.
t
In the continuous time tradition, we usually express integrals going forward in time,
as this one is. To connect with discrete time, you can also rearrange the integral to
go back in time,
Z
T
t=0
c
(tT)
d.
t
=
Z
T
c=0
c
c
d.
Tc
.
ii. To check this answer, write it as
r
t
j = c
t
(r
0
j) + o
Z
t
c=0
c
(tc)
d.
c
and take dr
t
. The rst term is just the usual time derivative. Then, we take the
derivative of the stu inside the integral as time changes, and lastly we account for
the fact that the upper end of the integral changes,
dr
t
= cc
t
(r
0
j) dt + o
Z
t
c=0
(c) c
(tc)
d.
c
+ od.
t
dr
t
= c

c
t
(r
0
j) dt + o
Z
t
c=0
c
(tc)
d.
c

+ od.
t
dr
t
= c(r
t
j) dt + od.
t
(e) This example adds an important tool to our arsenal. Notice we can weight sums of d.
t
terms, just as we weight sums of -
t
terms. We can produce or write results as continuous
time moving average processes
r
T
=
Z
T
t=0
n(T t)d.
t
7. Once you have solved an sde you have expressed r
T
as a random variable. Naturally, you
will want to know moments means and variances of r
T
. Here are some examples.
25
(a) Example: Diusion:
r
T
r
0
= jT + o
Z
T
t=0
d.
t
1 (r
T
r
0
) = jT + o
Z
T
t=0
1 (d.
t
) = jT
ar (r
T
r
0
) = o
2
1
Z
T
t=0
d.
t

2
= o
2
Z
T
t=0
1

d.
2
t

= o
2
Z
T
t=0
dt = o
2
T
I used d.
t
d.
c
= 0 for t 6= : in the last line.
(b) Example: Lognormal price process
dj
t
j
t
= jdt + od.
t
.
The solution is
j
t
= j
0
c
(j
1
2
o
2
)t+o

=0
o:
and then we can nd the expected value
1
0
(j
t
) = j
0
c
(j
1
2
o
2
)t+
1
2
o
2

=0
oc
= r
0
c
(j
1
2
o
2
)t+
1
2
o
2
t
= r
0
c
jt
(In the last line I use the fact that for normal j, 1(c
j
) = c
1(j)+
1
2
o
2
(j)
.)
(c) Example: AR(1). Start with the solution, then use the fact that 1(d.
t
) = 0, 1(d.
2
t
) =
dt, 1(d.
t
d.
c
) = 0, for t 6= :.
(r
t
j) = c
t
(r
0
j) +
Z
t
c=0
c
(tc)
od.
c
1
0
(r
t
j) = c
t
(r
0
j)
o
2
(r
t
j) = 1
"
Z
t
c=0
c
(tc)
od.
c

2
#
=
Z
t
c=0
c
2(tc)
o
2
d: =
1 c
2t
2c
o
2
as in discrete time
1
0
(r
t+)
j) = j
)
(r
0
j)
ar
0
(r
t
j) = 1

t1
X
)=0
j
)
-
t)

=
t1
X
)=0
j
2)
o
2
=
1 j
2t
1 j
2
o
2
.
(d) Example: Continuous-time MA. If you have a solution
r
t
=
Z
t
0
q(t :)d.
c
,
26
then
1
0
(r
t
) = 1
0
Z
t
0
q(t :)d.
c

=
Z
t
0
q(t :)1
c
(d.
c
) = 0
and
1
0
(r
2
t
) = 1
0
Z
t
0
q(t :)d.
c

2
=
Z
t
0
q(t :)
2
d:;
8. In general, solving SDEs is not so easy! As with regular dierential equation, you cant just
integrate both sides, because you might have an r on the right side as well that you cant
easily get rid of. The idea is simple, from
dr
t
= j(r
t
, t)dt + o(r
t
, t)d.
t
you can write
r
T
= r
0
+
Z
T
0
j(r
t
, t)dt +
Z
T
0
o(r
t
, t)d.
t
We end up describing a random variable )(r
T
|r
0
). The hard part is nding closed-form
expressions, when the j and o depend on r. The rst term is just a standard dierential
equation. (And remember how easy that is!)
9. Simulation. You can easily nd distributions and moments by simulating the solution to a
stochastic dierential equation. Use the random number generator and program
r
t+
= r
t
+ j(r
t
, t)+ o(r
t
, t)

-
t+
; -
t+
(0, 1)
10. A last example. The CIR (Cox Ingersoll Ross) square root process. Suppose we generate r
t
from .
t
by
dr
t
= c(r
t
j)dt + o

r
t
d.
t
.
The AR(1) process ranges from to . As r 0, volatility goes down, and the drift is
pulling it back, so r
t
cant cross zero. That fact, and the fact that its more volatile when the
level is higher, makes it a good process for nominal interest rates. Its a nonlinear process,
not in ARMA class. A strong point for continuous time is that we can handle many of these
nonlinear processes in continuous time, though we really cant really do much with them in
discrete time. For example, the discrete-time square root process is
r
t
= jr
t1
+

r
t
-
t
Now what do you do? There are closed form solutions for the CIR process and many other
nonlinear processes. I wont give those here.
2.5 Finding Moments more generally
(The remaining sections will not be used in Asset Pricing or my courses.)
Often you dont really need the whole solution, you only want moments. If r
T
is a cash payo
at time T, you might want to know its expected value, 1
t
(r
T
). You might want the expected value
27
of a function of r
T
, i.e. 1
t
[c(r
T
)]. This happens in option pricing, where c is the call option payo
function. The most basic asset pricing valuation formula is a moment 1
0
R
c
jt
:
t
r
t
dt

.
We saw above how to nd moments after you have a full solution. but thats hard. But you
can often nd moments without rst nding the whole solution or distribution of the random
variable, i.e. r
T
or c(r
T
) itself.
1. But we can also nd moments directly because Moments follow nonstochastic dierential
equations. You dont have to solve stochastic dierential equations to nd the moments.
(a) Example.
dr
t
= jdt + o()d.
Lets suppose we want to nd the mean. Now,
1
0
(r
t+
) 1
0
(r
t
) = d [1
0
(r
t
)] = 1
0
(dr
t
) = jdt
Watch the birdie here. By d [1
0
(r
t
)] I mean, how does the expectation 1
0
(r
t
) move
forward in time. But now the point: Since the d. term is missing, you can nd the mean
of x without solving the whole equation.
(b) Example: lognormal pricing
dr
r
= jdt + od.
The mean follows
d1
0
(r
t
)
1
0
(r
t
)
= jdt
d1
0
(r
t
) = j1
0
(r
t
)dt
1
0
(r
t
) = r
0
c
jt
This is the same solution we had before. Look how much easier that was!
(c) Similarly, to nd moments 1
0
c(r), nd the diusion representation for 1
0
c(r) by Itos
lemma.
(d) Alas, this technique is limited. If j(r
t
) then 1
0
(dr
t
) = 1
0
(j(r
t
))dt. You need to know
the distribution of r
t
to get anywhere.
2. The Backward equation. What is 1
t
[c(r
T
)]? We know that 1
T
[c(r
T
)] = c(r
T
). We can
work backwards using Itos lemma to nd 1
t
[c(r
T
)]. We have a process
dr = j()dt + o()d.
and we want a moment, say
1
t
[c(r
T
)] .
In particular, 1
t
(r
T
) is c(r
t
) = r
t
, 1
t
(r
2
T
) is c(r
t
) = r
2
t
. Dene
)(r
t
, t) = 1
t
[c(r
T
)]
28
Since its a conditional expectation, 1
t
[1
t+
()] = 1
t
(), so
1
t
(d)
t
) = 0.
Applying Itos lemma to ),
d) =
0)
0t
dt +
0)
0r
dr +
1
2
0
2
)
0r
2
dr
2
1
t
(d))
dt
=
0)
0t
+
0)
0r
j() +
1
2
0
2
)
0r
2
o
2
() = 0
Thus, the conditional expectation solves the backward equation
0)(r
,
t)
0t
+
0)(r
,
t)
0r
j(r
,
t) +
1
2
0
2
)(r
,
t)
0r
2
o
2
(r
,
t) = 0
starting from the boundary condition
)(r, T) = c(r)
(a) This is a partial dierential equation (ugh). The usual way to solve it is to guess and
check you guess a functional form with some free parameters, and then you see what
the free parameters have to be to make it work. It is also the kind of equation you can
easily solve numerically. Use the spatial derivatives at time t to nd the time derivative,
and hence nd the function at time t . Loop. That may be easier numerically than
solving for 1
t
c(r
T
) by Monte-Carlo simulation of r
T
forward. Or, the Monte-Carlo
simulation may turn out to be an easy way to solve partial dierential equations.
3. Finding densities. You can also nd the density by an Itos lemma dierential equation
rather than solving the whole thing as we did above. This is called the Forward equation.
(a) The density at t + must be the density at t times the transition density to go from t
to t +.
)(r
t+
|r
0
) =
Z
)(r
t
|r
0
))(r
t+
|r
t
)dr
t
=
Z
)(r
t
|r
0
)

(r
t+
r
t
j(r
t
))
o(r
t
)

dr
t
...Itos lemma and lots of algebra, and allowing explicit )(r, t), ...
0)(r, t|r
0
)
0t
+
0 [j(r))(r, t|r
0
)]
0r
=
1
2
0
2

o
2
(r))(r, t|r
0
)

0r
2
Note the j and o
2
are inside the derivative. This is because its j(r
t
) not j(r
t+
).
(b) This fact gives a nice formula for the stationary (unconditional) density. The formula is:
)(r) = /
c

2
()

2
()
oc
o
2
(r)
29
where / is a constant, so we integrate to one. Evaluating /,
)(r) =

Z
c

2
()

2
()
oc
o
2
(:)
d:

1
c

2
()

2
()
oc
o
2
(r)
This is often pretty easy to use.
i. Derivation: The unconditional density satises 0),0t = 0 so from the forward equa-
tion,
d [j(r))(r)]
dr
=
1
2
d
2

o
2
(r))(r)

dr
2
j(r))(r) =
1
2
d

o
2
(r))(r)

dr
2
j(r)
o
2
(r)

o
2
(r))(r)

=
d

o
2
(r))(r)

dr
o
2
(r))(r) = /c

2
()

2
()
oc
)(r) = /
c

2
()

2
()
oc
o
2
(r)
Normalizing so
R
)(r)dr = 1,
)(r) =

Z
c

2
()

2
()
oc
o
2
(:)
d:

1
c

2
()

2
()
oc
o
2
(r)
30