On Stochastic Portfolio Theory

by E. R. Fernholz

May 9, 2002

The research monograph Stochastic Portfolio Theory by E. R. Fernholz presents a

novel mathematical methodology for analyzing portfolio behavior and stock market
structure. Stochastic portfolio theory is descriptive rather than normative, and hence
it differs from the current doctrine of mathematical and quantitative finance. De-
scriptive theories offer explanations for observed phenomena and predictions for the
outcomes of future experiments; the theories of natural science are descriptive the-
ories. The central theories of finance are normative theories: general equilibrium,
dynamic asset pricing, no-arbitrage. These normative theories are based on assumed
ideal behavior regarding the interaction of the participants and agents in the markets
under consideration, and this ideal behavior is frequently far different from actual
observed behavior.
Since stochastic portfolio theory is descriptive, it is applicable under a wide range
of assumptions and conditions that may hold in real markets. Unlike current methods
used in mathematical finance, stochastic portfolio theory is consistent with either
equilibrium or disequilibrium and with either arbitrage or no-arbitrage. It can be
applied to the optimization of actual portfolios and to the analysis of real stock
markets.
Chapter 1 introduces the basic definitions and preliminary results that are used
throughout the rest of the book. The definitions of stocks and portfolios, although
technically equivalent to the conventional definitions, differ from a philosophical per-
spective. The model presented here is based on the growth rate, or logarithmic rate
of return, of the stocks and portfolios, and not on the ordinary (arithmetic) rate of
return. From this new perspective, features of stock and portfolio behavior that have
heretofore been obscure now stand out clearly.
As an elementary result, it is shown that the growth rate of a stock or portfolio
determines its long-term behavior, and that the rate of return becomes irrelevant over
time. It is then shown that the growth rate of a portfolio depends not only on the
growth rates of its component stocks, but also on their variances and covariances.
This result differs significantly from the conventional theory, where the portfolio rate
of return is simply the weighted average of the component stocks’ rates of return.
This difference leads to portfolio optimization in which the covariances of the stocks
play a greatly increased role. This form of optimization has a practical advantage,
since the variances and covariances are much more amenable to statistical analysis
than are the rates of return.

1
 The market, consisting of all the stocks, is studied as a portfolio. The market
weight of a stock is the ratio of the capitalization of that stock to the total capitali-
zation of the market, so all market weights are numbers between zero and one. The
market portfolio is the portfolio that holds each stock at its market weight.
Chapter 2, “Stock market behavior and diversity,” develops minimal conditions
for the structural stability of an equity market. Diversity measures how uniformly
capital is distributed among the stocks in the market. In a diverse market, the market
weight of any stock cannot approach one. In actual markets, this condition would
certainly be ensured by any credible antitrust law. In a weakly diverse market, the
time average of the maximum of all the market weights cannot approach one. It is
not difficult to see that a diverse market is also weakly diverse, but a weakly diverse
market need not be diverse. These diversity conditions are both quite weak, and are
obviously consistent with the structure of the U.S equity market, or any other equity
market of any significance.
Despite the fact that diversity and weak diversity are not very restrictive, many
simple hypothetical markets fail to satisfy either of these conditions. For example, if
the stocks in a market have constant growth rates, or if they all have the same growth
rate, then the market is not weakly diverse. In fact, these markets are the “opposite”
of diverse: the time average of the maximum of the market weights approaches one
with probabilistic certainty.
As a measure of diversity, the entropy of the market is introduced. Entropy has
been used in physics and mathematics as a measure of diversity for some time, and it
is natural to introduce it here. However, the interesting point here is that the entropy
measure generates a portfolio. The log-return on this portfolio, called the entropy-
weighted portfolio, is related to the market’s log-return by a stochastic differential
equation. This equation allows probability-one conclusions to be drawn about the
relative performance of the entropy-weighted portfolio based on the stability of the
entropy of the market. This establishes a connection between market diversity and
the relative return of a portfolio. This connection is generalized and expanded in the
next chapter.
In Chapter 3, “Functionally generated portfolios,” the portfolio construction tech-
nique used for the entropy-weighted portfolio is extended to more general functions
of the market weights. These functions are called portfolio generating functions and
the portfolios they generate are called functionally generated portfolios. The return
on such a portfolio satisfies a stochastic differential equation similar to the equation
for the entropy-weighted portfolios.
The stochastic differential equation associated with a functionally generated port-
folio decomposes the logarithmic relative return of the portfolio into two components.
The first component is the logarithmic change in the value of the generating function,
and this term can be controlled by bounds on the value of the generating function.
The second term is called the drift process, and this is usually an expression involving

2
the relative variances and covariances of the stocks in the portfolio.
The entropy function is a member of a class of functions called measures of diver-
sity. Measures of diversity generate portfolios called diversity-weighted portfolios, of
which the entropy-weighted portfolio is an example. Relative to the market portfo-
lio, diversity-weighted portfolios are overweighted in the smaller stocks, and under-
weighted in the larger stocks. The drift processes for diversity-weighted portfolios
are increasing, so these portfolios outperform the market over periods in which the
level of diversity of the market is maintained. Since in a stable market, the market
diversity is likely to be maintained, this condition is frequently met in practice. In
one example of this chapter, the increasing nature of the drift process for a partic-
ular measure of diversity is exploited to show that the no-arbitrage hypothesis of
mathematical finance fails to hold in a weakly diverse market. Another measure of
diversity defined here, Dp , has been used in practice to construct portfolios for the
management of institutional equity portfolios.
In Chapter 4, “Portfolios of stocks selected by rank,” the idea of considering stocks
by their rank rather than by their name is introduced. The capital distribution of the
market is defined to be the family of ranked market weights, starting with the largest
weight and going to the smallest. The results of the previous chapter are extended
here to show that certain functions of the ranked market weights generate portfolios.
Again the relative return of a functionally generated portfolio satisfies a stochastic
differential equation, but now the relative return has an additional component. The
new component depends on the local times that measure the frequency of changes in
rank among the stocks.
This additional component provides new insight in at least two situations. First,
for portfolios composed of stocks from an index, e.g., an index composed of the
largest 1000 stocks in the market, the local time component evaluates the effect on
the relative return of the portfolio of stocks entering or leaving the index because
of rank changes. It also provides a mathematically rigorous explanation of the size
effect, the tendency of smaller stocks to outperform larger stocks over the long term.
However, the most important application of the idea of ranked market weights is
not to portfolio construction, but to provide a basis for a structural model for equity
markets which have a stable distribution of capital. This concept of stability is made
precise and developed in the next chapter.
Chapter 5, “Stable models for the distribution of capital,” presents a structural
theory for equity markets in which the capital distribution exhibits stability over
time. The capital distribution curve refers to the log-log plot of the market weights
arranged in descending order, i.e., the logarithms of the market weights versus the
logarithms of their respective ranks. If this curve is roughly a straight line, then the
capital distribution follows a Pareto distribution. Much is known about the Pareto
distribution, and attempts have been made to apply it to the capital distribution of
the market. However, empirically the capital distribution curve does not resemble a

3
straight line, but instead is notably concave (dome-shaped).
An asymptotically stable market is a market in which the frequency of changes
in rank among the stocks is stable over the long term. For asymptotically stable
markets, it is shown that there are certain characteristic parameters that determine
the shape of the capital distribution curve. The U.S. equity market is analyzed and
appears to be asymptotically stable, and the values of the characteristic parameters
for the U.S. market are estimated.
Finally, a first-order model is constructed for an asymptotically stable market, and
it is shown that the first-order model has (almost) the same capital distribution curve
as the market from which it is derived. For the U.S. market, the capital distribution
curve for the first-order model is stable, and does not vary as much over time as the
actual capital distribution curve of the market.
In Chapter 6, “Performance of functionally generated portfolios,” a number of
examples of portfolios of stocks in the U.S. equity market are studied. Market en-
tropy and a diversity-weighted portfolio generated by the generating function Dp are
analyzed. Trading turnover is analyzed for a Dp -generated portfolio, and it is shown
that in the U.S. market a Dp -generated portfolio will have about 7% turnover a year if
the stocks are traded when their actual weights differ from their theoretical portfolio
weights by a factor of .1 or greater.
In this chapter, the explanation of the size effect given in Chapter 4 is tested in
the U.S. market. It is shown that the “crossovers” of stocks from a large-stock index
to a small-stock index generate about 1% a year superior logarithmic return for the
small-stock index. The portfolio containing only the largest stock in the market is
analyzed, and this portfolio is shown to underperform the market by about .19% a
year.
Chapter 7, “Applications of stochastic portfolio theory,” provides examples of
various types of applications for the methodology. Optimization models are con-
structed using the first-order model, and these models enjoy the property that all the
necessary parameters, i.e., the growth rates and variances, can be easily calculated.
Estimation of these parameters is a serious weakness in classical portfolio theory, and
with the first-order model, such estimation is unnecessary. Also discussed here is
diversity-weighted indexing, the application of the Dp generating function to popular
large-stock indices such as the S&P 500 or the Russell 1000.
It turns out that change in diversity is a significant factor in manager performance,
and this factor explained more than half of the variation of manager performance
relative to the S&P 500 over the period from 1971 to 1998. The relationship between
manager performance and change in diversity is discussed here, as well as a method of
directly calculating the effect that changes in the capital distribution have on portfolio
return.