STOCHASTIC PORTFOLIO

THEORY
IOANNIS KARATZAS
Mathematics and Statistics Departments
Columbia University
ik@math.columbia.edu
Joint work with
Dr. E. Robert FERNHOLZ, C.I.O. of INTECH
Enhanced Investment Technologies, Princeton NJ
Talk in Vienna, September 2007
CONTENTS
1. The Model
2. Portfolios
3. The Market Portfolio
4. Diversity
5. Diversity-Weighting and Arbitrage
6. Performance of Diversity-Weighted Portfolios
7. Strict Supermartingales
8. Completeness & Optimization without EMM
9. Concluding Remarks
1
SYNOPSIS
The purpose of this talk is to oer an overview of
Stochastic Portfolio Theory, a rich and exible
framework for analyzing portfolio behavior and eq-
uity market structure. This theory is descriptive as
opposed to normative, is consistent with observ-
able characteristics of actual markets and portfo-
lios, and provides a theoretical tool which is useful
for practical applications.
As a theoretical tool, this framework provides fresh
insights into questions of market structure and ar-
bitrage, and can be used to construct portfolios
with controlled behavior. As a practical tool, Sto-
chastic Portfolio Theory has been applied to the
analysis and optimization of portfolio performance
and has been the basis of successful investment
strategies for close to 20 years.
2
REFERENCES
Fernholz, E.R. (2002). Stochastic Portfolio The-
ory. Springer-Verlag, New York.
Fernholz, E.R. & Karatzas, I. (2005). Relative
arbitrage in volatility-stabilized markets. Annals of
Finance 1, 149-177.
Fernholz, E.R., Karatzas, I. & Kardaras, C. (2005).
Diversity and arbitrage in equity markets. Finance
& Stochastics 9, 1-27.
Karatzas, I. & Kardaras, C. (2007). The numeraire
portfolio and arbitrage in semimartingale markets.
Finance & Stochastics 11, 447-493.
Fernholz, E.R. & Karatzas, I. (2008) Stochastic
Portfolio Theory: An Overview. To appear.
3
1. THE MODEL. Standard Model (Bachelier,
Samuelson,...) for a nancial market with n stocks
and d n factors: for i = 1, . . . , n,
dX
i
(t) = X
i
(t)
_
_
b
i
(t)dt +
d

=1

i
(t)dW

(t)
_
_
. .
.
Vector of rates-of-return: b() = (b
1
(), . . . , b
n
())

.
Matrix of volatilities: () = (
i
())
1in, 1d
will be assumed bounded for simplicity.
Assumption: for every T (0, ) we have
n

i=1
_
T
0

b
i
(t)

2
dt < , a.s.
All processes are adapted to a given ow of infor-
mation (or ltration) F = {F(t)}
0t<
, which
satises the usual conditions and may be strictly
larger than the ltration generated by the driving
Brownian motion W() = (W
1
(), . . . , W
d
())

.
No Markovian, or Gaussian, assumptions...
4
Suppose, for simplicity, that the variance/covariance
matrix a() = ()

() has all its eigenvalues bounded

away from zero and innity: that is,
|| ||
2

a(t) K|| ||
2
, R
d
holds a.s. (for suitable constants 0 < < K < ).
Solution of the equation for stock-price X
i
():
d (logX
i
(t)) =
i
(t) dt +
d

=1

i
(t) dW

(t)
. .
with

i
(t) := b
i
(t)
1
2
a
ii
(t)
the growth-rate of the i
th
stock, in the sense
lim
t
1
T
_
logX
i
(t)
_
T
0

i
(t)dt
_
= 0 a.s.
5
2. PORTFOLIO. An adapted vector process
(t) = (
1
(t), ,
n
(t))

;
fully-invested, no short-sales, no risk-free asset:

i
(t) 0,
n

i=1

i
(t) = 1 for all t 0 .
Value Z

() of portfolio:
dZ

(t)
Z

(t)
=
n

i=1

i
(t)
dX
i
(t)
X
i
(t)
= b

(t)dt +
d

=1

(t)dW

(t)
with Z

(0) = 1. Here
b

(t) :=
n

i=1

i
(t)b
i
(t) ,

(t) :=
n

i=1

i
(t)
i
(t) ,
are, respectively, the portfolio rate-of-return, and
the portfolio volatilities.
6
Solution of this equation:
d
_
logZ

(t)
_
=

(t) dt +
n

=1

(t) dW

(t)
. .
.
Portfolio growth-rate

(t) :=
n

i=1

i
(t)
i
(t) +

(t) .
Excess growth-rate

(t) :=
1
2
_
_
n

i=1

i
(t)a
ii
(t)
n

i=1
n

j=1

i
(t)a
ij
(t)
j
(t)
_
_
. .
.
Portfolio variance
a

(t) :=
d

=1
(

(t))
2
=
n

i=1
n

j=1

i
(t)a
ij
(t)
j
(t) .
For a given portfolio (), let us introduce the
order statistics notation, in decreasing order:

(1)
:= max
1in

i

(2)
. . .
(n)
:= min
1in

i
.
7
3. MARKET PORTFOLIO: Look at X
i
(t) as
the capitalization of company i at time t . Then
X(t) := X
1
(t)+. . .+X
n
(t) is the total capitalization
of all stocks in the market, and

i
(t) :=
X
i
(t)
X(t)
=
X
i
(t)
X
1
(t) +. . . +X
n
(t)
> 0
the relative capitalization of the i
th
company. Clearly

n
i=1

i
(t) = 1 for all t 0, so () is a portfolio
process, called "market portfolio".
. Ownership of () is tantamount to ownership of
the entire market, since Z

() c.X().
FACT 1:

() (/2)
_
1
(1)
()
_
.
Moral: Excess-rate-of growth is non-negative, strictly
positive unless the portfolio concentrates on a sin-
gle stock: diversication helps not only to reduce
variance, but also to enhance growth.
8
HOW? Consider, for instance, a xed-proportion
portfolio
i
() p
i
0 with

n
i=1
p
i
= 1 and
p
(1)
= 1 < 1. Then
log
_
Z
p
(T)
Z

(T)
_

i=1
p
i
log
i
(T) =
_
T
0

(t) dt

2
T .
And if lim
T
1
T
log
i
(T) = 0 (no individual stock
collapses very fast), then this gives almost surely
lim
T
1
T
log
_
Z
p
(T)
Z

(T)
_


2
> 0 :
a signicant outperforming of the market.
Remark: Tom Covers universal portfolio

i
(t) :=
_

n p
i
Z
p
(t) dp
_

n Z
p
(t) dp
, i = 1, , n
has value
Z

(t) =
_

n Z
p
(t) dp
_

n dp
max
p
n
Z
p
(t) .
FACT 2:

() 2K
_
1
(1)
()
_
.
4. DIVERSITY. The market-model M is called
Diverse on [0, T], if there exists (0, 1) such
that we have a.s.:
(1)
(t) < 1
. .
, 0 t T.
Weakly Diverse on [0, T], if for some (0, 1):
1
T
_
T
0

(1)
(t) dt < 1 , a.s.
. .
FACT 3: If M is diverse, then

() for
some > 0; and vice-versa.
FACT 4: If all stocks i = 1, . . . , n in the market
have the same growth-rate
i
() (), then
lim
T
1
T
_
T
0

(t) dt = 0, a.s.
In particular, such a market cannot be diverse
on long time-horizons: once in a while a single
stock dominates such a market, then recedes; sooner
or later another stock takes its place as absolutely
dominant leader; and so on.
10
Here is a quick argument: from
i
() () and
X() = X
1
() + +X
n
() we have
lim
T
1
T
_
logX(T)
_
T
0

(t)dt
_
= 0,
lim
T
1
T
_
logX
i
(T)
_
T
0
(t)dt
_
= 0.
for all 1 i n. But then
lim
T
1
T
_
logX
(1)
(T)
_
T
0
(t)dt
_
= 0, a.s.
for the biggest stock X
(1)
() := max
1in
X
i
() ,
and note X
(1)
() X() nX
(1)
() . Therefore,
lim
T
1
T
_
logX
(1)
(T) logX(T)
_
= 0, thus
lim
1
T
_
T
0
_

(t) (t)
_
dt = 0.
But

(t) =

n
i=1

i
(t)(t) +

(t) = (t) +

(t) ,
because all growth rates are equal. 2
11
FACT 5: In a Weakly Diverse Market there exist
potfolios () that lead to arbitrage relative to
the market-portfolio: with common initial capital
Z

(0) = Z

(0) = 1, and some T (0, ), we have

P[Z

(T) Z

(T)] = 1, P[Z

(T) > Z

(T)] > 0.
And not only do such relative arbitrages exist; they
can be described, even constructed, fairly explicitly.
5. DIVERSITY-WEIGHTING & ARBITRAGE
For xed p (0, 1), set

i
(t)
(p)
i
(t) :=
(
i
(t))
p

n
j=1
(
j
(t))
p
, i = 1, . . . , n .
Relative to the market portfolio (), this () de-
creases slightly the weights of the largest stock(s),
and increases slightly those of the smallest stock(s),
while preserving the relative rankings of all stocks.
12
We shall show that, in a weakly-diverse market M,
this portfolio

i
(t)
(p)
i
(t) :=
(
i
(t))
p

n
j=1
(
j
(t))
p
. .
, i = 1, . . . , n,
for some xed p (0, 1), satises
P[Z

(T) > Z

(T)] = 1, T T

:=
2
p
logn .
In particular,
(p)
() represents an arbitrage oppor-
tunity relative to the market-portfolio ().
Suitable modications of
(p)
() can generate
such arbitrage over arbitrary time-horizons.
The signicance of such a result, for practical long-
term portfolio management, cannot be overstated.
Discussion and performance charts can be found in
the monograph Fernholz (2002).
13
6: Performance of Diversity-Weighting
Indeed, for this diversity-weighted portfolio

i
(t)
(p)
i
(t) :=
(
i
(t))
p

n
j=1
(
j
(t))
p
, i = 1, . . . , n
with xed 0 < p < 1 and D(x) :=
_

n
j=1
x
p
j
_
1/p
,
we have
log
_
Z

(T)
Z

(T)
_
= log
_
D((T))
D((0))
_
+(1 p)
_
T
0

(t)dt .
First term on RHS tends to be mean-reverting,
and is certainly bounded:
1 =
n

j=1
x
j

n

j=1
(x
j
)
p

_
D(x)
_
p
n
1p
.
Measure of Diversity: minimum occurs when one
company is the entire market, maximum when all
companies have equal relative weights.
14
We remarked already, that the biggest weight of
() does not exceed the largest market weight:

(1)
(t) := max
1in

i
(t) =
_

(1)
(t)
_
p

n
k=1
_

(k)
(t)
_
p

(1)
(t) .
By weak diversity over [0, T], there is a number
(0, 1) for which
_
T
0
(1
(1)
(t)) dt > T holds;
thus, from Fact #1:
2

_
T
0

(t) dt
_
T
0
_
1
(1)
(t)
_
dt

_
T
0
_
1
(1)
(t)
_
dt > T , a.s.
From these two observations we get
log
_
Z

(T)
Z

(T)
_
> (1 p)
_
T
2

1
p
logn
_
,
so for a time-horizon
T > T

:= (2logn)/p
suciently large, the RHS is strictly positive. 2
15
Remark: It can be shown similarly that, over su-
ciently long time-horizons T > 0, arbitrage relative
to the market can be constructed under

(t) > 0, 0 t T (1)

for some real , or even under the weaker condition
_
T
0

(t) T > 0.
Open Question, whether this can also be done
over arbitrary horizons T > 0.
This result does not presuppose any condition
on the covariance structure (a
ij
()) of the mar-
ket, beyond (1). There are examples, such as the
volatility-stabilized model (with 0)
d logX
i
(t) =

2
i
(t)
dt +
dW
i
(t)
_

i
(t)
, i = 1, , n
for which variances are unbounded, diversity fails,
but (1) holds:

() ((1+)n1)/2, a

() 1.
In this example, arbitrage relatively to the market
can be constructed over arbitrary time-horizons.
16
7: STRICT SUPERMARTINGALES.
The existence of relative arbitrage precludes the ex-
istence of an equivalent martingale measure (EMM)
at least when the ltration F is generated by the
Brownian motion W itself, as we now assume.
In particular, if we can nd a market-price-of-
risk process () with
() () = b() and
_
T
0
||(t)||
2
dt < a.s. ,
then it can be shown that the exponential process
L(t) := exp
_

_
t
0

(s) dW(s)
1
2
_
t
0
||(s)||
2
ds
_
. .
is a local (and super-)martingale, but not a mar-
tingale: E[L(T)] < 1.
Same for L()X
i
() : E[L(T)X
i
(T)] < X
i
(0) .
Typically: lim
T
E[L(T)X
i
(T)] = 0, i = 1, , n.
Examples of diverse and volatility-stabilized mar-
kets satisfying these conditions can be constructed.
17
In terms of this exponential supermartingale L() ,
we can answer some basic questions, for d = n:
Q.1: On a given time-horizon [0, T], what is the
maximal relative return in excess of the market
R(T) := sup{ r > 1 : h() s.t. Z
h
(T)/Z

(T) r , a.s. }
that can be attained by trading strategies h()?
(These can sell stock short, or invest/borrow in
a money market at rate r(), but are required to
remain solvent: Z
h
(t) 0, 0 t T .)
Q.2: Again using such strategies, what is the
shortest amount of time required to guarantee
a return of at least r > 1, times the market?
T(r) := inf{ T > 0 : h() s.t. Z
h
(T)/Z

(T) r , a.s. }
Answers: R(T) = 1/f(T) and f
_
T(r)
_
= r ,
where f(t) := E
_
e

_
t
0
r(s)ds
L(t)
X(t)
X(0)
_
0.
18
8. COMPLETENESS AND OPTIMIZATION
WITHOUT EMM
In a similar vein, given an F(T)measurable ran-
dom variable Y : [0, ) (contingent claim),
we can ask about its hedging price
H
Y
(T) := inf{ w > 0 : h() s.t. Z
w,h
(T) Y , a.s. } ,
. .
the smallest amount of initial capital needed to
hedge it without risk.
With D(T) := e

_
T
0
r(s)ds
, this can be computed
as
H
Y
(T) = y := E[ L(T)D(T) Y ]
(extended Black-Scholes) and an optimal strategy

h() is identied via Z

y,

h
(T) = Y , a.s.
To wit: such a market is complete, despite the
fact that no EMM exists for it.
19
Take Y = (X
1
(T) q)
+
as an example, and as-
sume r() r > 0. Simple computation Jensen:
X
1
(0) > H
Y
(T)
_
E[L(T)D(T)X
1
(T)] qe
rT
_
+
.
Letting T we get, as we have seen:
H
Y
() := lim
T
H
Y
(T)
= lim
T
E[L(T)D(T)X
1
(T)] = 0.
. Please contrast this, to the situation whereby
an EMM exists on every nite time-horizon [0, T].
Then at t = 0, you have to pay full stock-price for
an option that you can never exercise!
H
Y
() = X
1
(0) .
Moral: In some situations, particularly on long
time-horizons, it might not be such a great idea to
postulate the existence of EMMs.
20
Ditto with portfolio optimization. Suppose we
are given initial capital w > 0, nite time-horizon
T > 0, and utility function u : (0, ) R (strictly
increasing and concave, of class C
1
, with u

(0+) =
, u

() = 0.) Compute the maximal expected

utility from terminal wealth
U(w) := sup
h()
E
_
u
_
Z
w,h
(T)
_ _
,
decide whether the supremum is attained and, if
so, identify an optimal trading strategy

h() .
. Answer: replicating trading strategy

h() for the
contingent claim
Y = I
_
(w) D(T)L(T)
_
, i.e., Z
w,

h
(T) = Y.
Here I() is the inverse of the strictly decreasing
marginal utility function u

() , and () the in-

verse of the strictly decreasing function
W() := E
_
D(T)L(T) I ( D(T)L(T))
_
, > 0.
No assumption at all that L() should be a mar-
tingale, or that an EMM should exist !
21
Here are some other problems, in which no EMM
assumption is necessary:
#1: Quadratic criterion, linear constraint (Mar-
kowitz, 1952). Minimize the portfolio variance
a

(t) =
n

i=1
n

j=1

i
(t)a
ij
(t)
j
(t)
among all portfolios () with rate-of-return
b

(t) =
n

i=1

i
(t)b
i
(t) b
0
at least equal to a given constant.
#2: Quadratic criterion, quadratic constraint.
Minimize the portfolio variance
a

(t) =
n

i=1
n

j=1

i
(t)a
ij
(t)
j
(t)
among all portfolios () with growth-rate at least
equal to a given constant
0
:
n

i=1

i
(t)b
i
(t)
0
+
1
2
n

i=1
n

j=1

i
(t)a
ij
(t)
j
(t) .
22
#3: Maximize the probability of reaching a given
ceiling c before reaching a given oor f , with
0 < f < 1 < c < . More specically, maximize
P[ T
c
< T
f
] , with T
c
:= inf{ t 0 : Z

(t) = c } .
In the case of constant coecients
i
and a
ij
, the
solution to this problem is nd a portfolio that
maximizes the mean-variance, or signal-to-noise,
ratio (Pestien & Sudderth, MOR 1985):

n
i=1

i
(
i
+
1
2
a
ii
)

n
i=1

n
j=1

i
a
ij

1
2
,
#4: Minimize the expected time E[ T
c
] until a
given ceiling c (1, ) is reached.
Again with constant coecients, it turns out that
it is enough to maximize the drift in the equation
for logZ
w,
(), namely

=

n
i=1

i
_

i
+
1
2
a
ii
_

1
2

n
i=1

n
j=1

i
a
ij

j
,
the portfolio growth-rate (Heath, Orey, Pestien &
Sudderth, SICON 1987).
23
#5: Maximize the probability P[ T
c
< T T
f
] of
reaching a given ceiling c before reaching a given
oor f with 0 < f < 1 < c < , by a given
deadline T (0, ).
Always with constant coecients, suppose there is
a portfolio = (
1
, . . . ,
n
)

that maximizes both

the signal-to-noise ratio and the variance,

n
i=1

i
(
i
+
1
2
a
ii
)

n
i=1

n
j=1

i
a
ij

1
2
and a

,
over all
1
0, . . . ,
n
0 with

n
i=1

i
= 1. Then
this portfolio is optimal for the above criterion
(Sudderth & Weerasinghe, MOR 1989).
This is a big assumption; it is satised, for instance,
under the (very stringent) condition that, for some
G > 0, we have
b
i
=
i
+
1
2
a
ii
= G, for all i = 1, . . . , n.
Open Question: As far as I can tell, nobody seems
to know the solution to this problem, if such si-
multaneous maximization is not possible.
24
9. SOME CONCLUDING REMARKS
We have surveyed a framework, called Stochastic
Portfolio Theory, for studying the behavior of port-
folio rules and exhibited simple conditions, such
as diversity (there are others...), which can lead
to arbitrages relative to the market.
All these conditions, diversity included, are descrip-
tive as opposed to normative, and can be tested
from the predictable characteristics of the model
posited for the market. In contrast, familiar as-
sumptions, such as the existence of an equivalent
martingale measure (EMM), are normative in na-
ture, and cannot be decided on the basis of pre-
dictable characteristics in the model; see example
in [KK] (2006).
The existence of such relative arbitrage is not the
end of the world; it is not heresy, or scandal, either.
25
Under reasonably general conditions, one can still
work with appropriate deators L()D() for the
purposes of hedging derivatives and of portfolio op-
timization.
Considerable computational tractability is lost, as
the marvelous tool that is the EMM goes out of
the window; nevertheless, big swaths of the eld
of Mathematical Finance remain totally or mostly
intact, and completely new areas and issues thrust
themselves onto the scene.
There is a lot more scope to this Stochastic Port-
folio Theory than can be covered in one talk. For
those interested, there is the survey paper with R.
Fernholz, at the bottom of the page
www.math.columbia.edu/ ik/preprints.html
It contains a host of open problems.
Please let us know if you solve some of them!
26