A Practitioner's Guide To Factor Models

A Practitioner's Guide to
Factor Models
The Research Foundation of

The Institute of Chartered Financial Analysts
978-0-943205-24-3
The Research Foundation of the Institute of Chartered Financial Analysts
Mission
The mission of the Research Foundation is to

identify, fund, and publish research material that:
expands the body of relevant and useful
knowledge available to practitioners;
assists practitioners in understanding and
applying this knowledge; and
enhances the investment management com-
munity's effectiveness in serving clients.
The Frontiers of
Investment Knowledge
Evolving
Concepts/Techniques
Body of
Investment
Gaining Validity
Ideas Whose Time
The Research Foundation of

The Institute of Chartered Financial Analysts
P . 0. Box 3668
Charlottesville, Virginia22903
U.S.A.
Telqbhone: 8041977-6600
F a : 8041977-1103
Table of Contents
Foreword ..................................... ix
A Practitioner's Guide to Arbitrage Pricing

Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
By Edwin Burmeister, Richard Roll, and Stephen A. Ross
AppendixA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
AppendixB . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
Multi-Index Models Using Simultaneous Estimation of
All Parameters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
By Edwin J. Elton and Martin J. Gruber
AppendixA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
Multiple-Factor Models for Portfolio Risk . . . . . . . . . . . . 59
By Richard Grinold and Ronald N. Kahn
AppendixA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
Bibliography ......... . . .......... . . .. ....... . . . 81
vii
Foreword
Ever since the seminal work of Harry M. Markowitz (1952), the identification
and measurement of investment risk have been hotly debated. Markowitz
constructed a mean-variance model to demonstrate how to quantify both the
risk and return of an asset or a portfolio of assets. The Markowitz model reveals
that, in an efficient marketplace, higher returns can be accomplished only by
accepting greater risks. Consequently, one of the most widely accepted
financial principles is the tradeoff between risk and return.
Although the concept of investment risk is universally recognized, the
appropriate measure of risk remains controversial. Financial researchers
generally agree that specific (nonsystematic) risks, such as those pertaining
only to individual companies, tend to cancel out in well-diversified portfolios.
Because systematic risk is nondiversifiable, however, it cannot be eliminated.
The first theory to measure systematic risk was the capital asset pricing model
(CAPM) for which William F. Sharpe (1964) shared the 1990 Nobel Memorial
Prize in Economic Sciences. The CAPM postulates that a single type of risk,
known as market risk, affects expected security returns. Only by exposing a
well-diversified portfolio to higher market risk can an investor expect to achieve
a higher rate of return.
Under the CAPM, market risk is defined as the variability of an asset's rate
of return relative to that of the overall market as measured by some market
index such as the S&P 500. Beta, the coefficient of the independent variable
(the market's rate of return) in an ordinary least squares regression equation to
explain the dependent variable (a security's rate of return), measures a
security's relative amount of systematic (market) risk. A beta equal to 1.0
indicates risk equivalent to that of the overall market, whereas a beta less than
1.0 denotes lower-than-market risk and a beta exceeding 1.0 indicates greater-
than-market risk.
The arbitrage pricing theory (APT), first presented by Stephen A. Ross
(1976), was the next major asset pricing model to appear. Also focusing on
A Practitioner's Guide to Factor Models
systematic risk, the APT recognizes that several different broad risk sources
may combine to influence security returns. The intuitive appeal of the APT
results from its recognition that the interaction of several macroeconomic
factors such as inflation, interest rates, and business activity affects rates of
return. The statistical process of factor analysis is employed to quantify the
broad risk factors and to estimate individual securities' degree of exposure to
these factors. A security effectively has a sensitivity to each systematic risk
factor. A series of beta coefficients are estimated to measure the sensitivity to
the respective factor risks for a particular security. Unlike the CAPM,
however, the individual factors, although precisely quantified, are not specifi-
cally associated with readily identifiable variables. Although considerable dis-
cussion continues about the number and the identification of these broad
factors, the APT nevertheless provides investment managers with a valuable
risk-management tool.
Factor models, the focus of this monograph, have existed for many years.
Even before the introduction of the popular CAPM and APT, Markowitz (1959)
proposed the use of a single-factor model to explain security returns. Some-
times referred to as index models, factor models often rely on the use of factor
analysis to identlfy factors that influence security returns.
A good portfolio manager, whether explicitly or implicitly, evaluates the
impact of a series of broad factors on the performances of various securities. In
this sense, a reliable factor model provides a valuable tool to assist portfolio
managers with the identification of pervasive factors that affect large members
of securities. According to a factor model, the return-generating process for a
security is driven by the presence of the various common factors and the
security's unique sensitivities to each factor (factor loadings). The common
factors may be readily identifiable fundamental factors such as price-earnings
ratio, size, yield, and growth. Factor models can be used to decompose
portfolio risk according to common factor exposure and to evaluate how much
of a portfolio's return was attributable to each common factor exposure.
Consequently, factor models offer a useful extension of the CAPM and the APT
because they advance our understanding about how key factors influence
portfolio risk and return.
The CAPM is clear about the source of risk (the market) but suffers because
no practical measure of the market exists. The APT causes difficulties because
it does not identlfy the number of important factors or define them. Financial
researchers and investment managers undoubtedly agree that only a few
important factors explain an overwhelming degree of investment risk and
return. Therefore, the appeal of factor models that define these factors
becomes apparent.
Foreword
Sharpe (1984) effectively summarizes why investment professionals should

pay heed to factor models:
While the relative importance of various actors changes over time,
as do the preferences of investors, we need not completely
abandon a valuable framework within which we can approach
investment decisions methodically. We have developed a useful
set of tools and should certainly continue to develop them.
Meanwhile, we can use the tools we have, as long as we use them
intelligently, cautiously, and humbly.
In this monograph, prominent academic and professional researchers of
factor models unite to present a practitioner's guide to factor models. In the first
article, Edwin Burmeister (Duke University), Richard Roll (University of
California, Los Angeles), and Stephen Ross (Yale University) offer "A Practi-
tioner's Guide to Arbitrage Pricing Theory." The authors not only explain the
basics and the equations of APT but also discuss the macroeconomic forces that
are the underlying sources of risk and how these factors combine to influence
rates of return. Importantly, they present several ways for investment
professionals to use APT effectively.
In the second article, Edwin J. Elton and Martin J. Gruber, both of New York
University, present "Multi-Index Models Using Simultaneous Estimation of All
Parameters." Beginning with an explanation of a single-factor model, the
authors next discuss multi-index models. They point out that the familiar
single-factor model provides a useful framework within which to compare
multi-index models. Their article examines methodology for simultaneously
estimating the indexes and sensitivities in a multi-index model. In addition, the
authors carefully test factor models, thus providing guidance with respect to the
reliability and usefulness of these models.
In the third article, Richard C. Grinold and Ronald N. Kahn, both of BARRA,
address "Multiple-Factor Models for Portfolio Risk." They present a practical
application of factor models to predict and control investment risk. Using a
widely recognized multiple-factor risk model developed at BARRA, Grinold and
Kahn emphasize the importance of identifying key fundamental factors that are
relatively easy for investment professionals to use. They stress the use of
factors that represent the recognized key investment attributes-volatility,
momentum, size, liquidity, growth, value, earnings volatility, and financial
leverage-and present specific measures of each. Thus, they advance the
factor model literature by moving from quantitative, but unspecified, factors to
readily identifiable fundamental characteristics.
As a result of the availability of substantial amounts of data and the increasing
A Practitioneh Guide to Factor Models
complexity of the investment decision-making process, the use of quantitative

models now has heightened prominence in investment management. Factor
models provide a useful extension of existing asset pricing models and allow
managers to decompose portfolios into sensitivities to broad factors. This
approach provides invaluable information about the attributes that influence
portfolio returns, thus allowing managers to explain these returns.
Although the mathematics of factor models may be complex, the underlying
concept is intuitively apparent. This series of articles explains factor models in
ways that investment practitioners can understand and use. The Research
Foundation of the Institute of Chartered Financial Analysts is pleased to sponsor
"A Practitioner's Guide to Factor Models." It is our hope and belief that these
articles will further the understanding and use of these valuable concepts.
John W. Peavy 111, CFA
xii
A Practitioner's Guide to
Arbitrage Pricing Theory
Edwin Burmeister
Duke University
Richard Roll
University of California, Los Angeles
Stephen A. Ross
Yale University
A fundamental principle of finance is the trade-off between risk and return.

Unless a portfolio manager possesses special information, one portfolio can be
expected to outperform another only if it is riskier in some appropriate sense.
The crucial question is: "What is the appropriate measure of risk?"
Many attributes might be related to an asset's risk, including market
capitalization (size), dividend yield, growth, price-earnings ratio (PIE), and so
on. Use of these traditional descriptors, however, presents at least three
problems:
1. Most are based on accounting data, and such data are generated by rules
that may differ sigmficantly across firms.
2. Even if all firms used the same accounting rules, reporting dates differ, so
constructing time-synchronized interfirm comparisons is difficult.
3. Most importantly, no rigorous theory tells us how traditional accounting
variables should be related to an appropriate measure of risk for computing the
risk-return trade-off. Even if historical empirical relationships can be uncov-
ered, without the foundation of a rigorous theory, one must be concerned that
any historical correlation might be spurious and subject to sudden and material
change.
Currently, only two theories provide a rigorous foundation for computing the
trade-off between risk and return: the capital asset pricing model (CAPM) and
the arbitrage pricing theory (APT).
The CAPM, for which William F. Sharpe shared the 1990 Nobel Memorial
Prize in Economic Sciences, predicts that only one type of nondiversifiable risk
influences expected security returns, and that single type of risk is "market
risk."l In 1976, a little more than a decade after the CAPM was proposed,
Stephen A. Ross invented the APT. The APT is more general than the CAPM
in accepting a variety of different risk sources. This accords with the intuition
that, for example, interest rates, inflation, and business activity have important
impacts on stock return volatility.
Although some theoretical formulations of the APT can be more intellectu-
ally demanding than the CAPM, the intuitively appealing basics behind the APT
are easy to understand. Moreover, the APT provides a portfolio manager with
a variety of new and easily implemented tools to control risks and to enhance
portfolio performance.
In the remainder of this paper, we will explain APT basics and the equations
of the APT. We will also discuss macroeconomic forces that are the underlying
sources of risk. We will then illustrate some risk exposure profiles and the
resulting APT-based risk-return trade-offs, and we will show how these
fundamental risks contribute to the expected and unexpected components of
realized return. Finally, we will discuss several uses of the APT that every
practitioner could easily apply.
The APT Basics

The CAPM and the APT agree that, although many different firm-specific
forces can influence the return on any individual stock, these idiosyncratic
effects tend to cancel out in large and well-diversified portfolios. This cancella-
tion is called the principle of diversification, and it has a long history in the field
of insurance. An insurance company has no way of knowing whether any
particular individual will become sick or will be involved in an accident, but the
company is able to predict its losses accurately on a large pool of such risks.
More precisely, if rm(t) is the return (in time period t) on a market index, such as the
S&P 500, the CAPM measure of the riskiness for asset i with return r,(t) is equal to that asset's
CAPM beta defined by Pi = cov[r,(t), rm(t)]lvar[rm(t)I.
The CAPM is equivalent to the statement that the market index is itself mean-variance efficient
in the sense of providing maximum average return for a given level of volatility. The index used
to implement the CAPM is implicitly assumed to be an effective proxy for the entire market of
assets.
An insurance company is not entirely free of risk, however, simply because
it insures a large number of individuals. For example, natural disasters or
changes in health care can have major influences on insurance losses by
simultaneously affecting many claimants. Similarly, large, well-diversified port-
folios are not risk free, because common economic forces pervasively influence
all stock returns and are not eliminated by diversification. In the APT, these
common forces are called systematic or pervasive risks.
According to the CAPM, systematic risk depends only upon exposure to the
overall market, usually proxied by a broad stock market index, such as the S&P
500. This exposure is measured by the CAPM beta, as defined in Footnote 1.
Other things equal, a beta greater (less) than 1.0 indicates greater (less) risk
relative to swings in the market index.2
The APT takes the view that systematic risk need not be measured in only
one way. Although the APT is completely general and does not specify exactly
what the systematic risks are, or even how many such risks exist, academic and
commercial research suggests that several primary sources of risk consistently
impact stock returns. These risks arise from unanticipated changes in investor
confidence, interest rates, inflation, real business activity, and a market index.
Every stock and portfolio has exposures (or betas) with respect to each of
these systematic risks. The pattern of economic betas for a stock or portfolio
is called its risk exposure profile. Risk exposures are rewarded in the market
with additional expected return, and thus the risk exposure profile determines
the volatility and performance of a well-diversified portfolio. The profile also
indicates how a stock or portfolio will perform under different economic
conditions. For example, if real business activity is greater than anticipated,
stocks with a high exposure to business activity, such as retail stores, will do
relatively better than those with low exposures to business activity, such as
utility companies.
Most importantly, an investment manager can control the risk exposure
profile of a managed portfolio. Managers with different traditional styles, such as
small-capitalization growth managers and large-capitalization value managers,
have differing inherent risk exposure profiles. For this reason, a traditional
manager's risk exposure profile is congruent to a particular APT style.
Given any particular APT style (or risk exposure profile), the difference
between a manager's expected return and his or her actual performance is
attributable to the selection of individual stocks that perform better or worse
Of course, "other things equal" can only be expected to hold on average over many time
periods.
A Practitionefs Guide to Factor Models
than a priori expectations. This extraordinary performance defines ex post APT

selection.
APT Equations
The APT follows from two basic postulates:
Postulate 1. In every time period, the difference between the actual

(realized) return and the expected return for any asset is equal to the sum, over
all risk factors, of the risk exposure (the beta for that risk factor) multiplied by
the realization (the actual end-of-period value) for that risk factor, plus an
asset-specific (idiosyncratic) error term.
This postulate is expressed by equation (1):
where
ri(t) the total return on asset i (capital gains plus dividends) realized at the
=
end of period t,
E[ri(t)] = the expected return, at the beginning of period t,
p,j = the risk exposure or beta of asset i to risk factor j for j = 1, . . . , K,
fi(t) = the value of the end-of-period realization for the jth risk factor, j =
1, .. .
, K, and
~ ~ (=t )the value of the end-of-period asset-specific (idiosyncratic) shock.
It is assumed that the expectations, at the beginning of the period, for all of the
factor realizations and for the asset-specific shock are zero; that is,
It is also assumed that the asset-specific shock is uncorrelated with the factor
realizations; that is,
cov[ei(t), fi(t)] = 0 for allj = 1, . . . , K.
Finally, all of the factor realizations and the asset-specific shocks are assumed
to be uncorrelated across time:
cov[fi(t), fi(tl)] = COV[E;(~),E i(t')] = 0
for all j = 1, . .. , K and for all t # t'.

The above conditions are summarized by saying that asset returns are
generated by a linear factor model. Note that the risk factors themselves may
be correlated (iation and interest rates, for example), as may the asset-
specific shocks for different stocks (as would be the case, for example, if some
unusual event influenced all of the firms in a particular industry).
Postulate 2. Pure arbitrage profits are impossible. Because of compe-

tition in financial markets, an investor cannot earn a positive expected rate of
return on any combination of assets without undertaking some risk and without
making some net investment of funds.
Postulate 2 is, in fact, an appealing equilibrium concept that has far-ranging
implications for broad areas of financial economics well beyond the determina-
tion of asset prices. It is hard to imagine any model of financial behavior that fails
to conclude that pure arbitrage profits tend to zero. This generality brings many
advantages. The APT is free of restrictive assumptions on preferences or
probability distributions, and it provides a rigorous logical foundation for the
trade-off between expected returns and risks.
Given Postulates 1 and 2, the main APT theorem is that there exist K + 1
..
numbers Po, P,, . , P,, not all zero, such that the expected return on the ith
asset is approximately equal to Poplus the sum over j of Pii times P,; that is,
Although equation (2) holds only approximately, with additional assumptions,

it can be proved that it holds exactly (see, e.g., Chen and Ingersoll1983). More
importantly, even without any additional assumptions, it has been proved that
the approximation in equation (2) is sufficiently accurate that any error can be
ignored in practical applications (see, e. g., Dybvig 1983). Thus the approxima-
tion symbol, =, can be replaced by an equal sign:
Here, Pj is the price of risk, or the risk premium for the jth risk factor. Via
equation (3), these Pi's determine the risk-return trade-off.3
Imagine a portfolio that is perfectly diversified (i.e., one for which ep(t) = 0)
..
and with no factor exposures (Ppi = 0 for all j = 1, . , K); such a portfolio has
zero risk, and from equation (3) its expected return is Po.Thus, Pomust be the
risk-free rate of return. Reasoning similarly, the risk premium for the jth risk
An equivalent interpretation of equation (3) uses an analogy to the familiar relationship that
"quantity x price = value." Thus, if we think of Pii as the quantity of type-j risk in the ith asset
and Pj as the price of type-j risk, then the product is the value of the contribution of type-j
risk to the expected return of the ith asset. If we let V,? denote this value, then it follows from
equation (3) that the sum of all the values is equal to the expected excess return (the expected
.
return in excess of the risk-free rate) for the ith asset; that is, E[r,(t)]- Po = V,, + . . + V.,
A PractitionerS Guide to Factor Models
factor, Pi, is the return, in excess of the risk-free rate, earned on an asset that
has one unit of risk exposure to the jth risk factor (Pii = 1) and zero risk
exposures to all of the other factors (Pi, = 0 for all h Z j).
The full APT is obtained by substituting equation (3) into equation (I),which
after rearranging terms yields:
It is at this level of the determination of expected returns that the CAPM and
the APT differ. In the CAPM, the expected excess return for an asset is equal
to that asset's CAPM beta times the expected excess return on a market index,
even for multifactor versions of the standard CAPM. For such a multifactor
CAPM to be true, the APT risk premiums-the Pis-must satisfy certain
restrictions. In statistical tests, these CAPM restrictions have repeatedly been
rejected in favor of the APT.
A portfolio manager controls a portfolio's betas-the portfolio's risk expo-
sure profile-by stock selection. Note that as the risk exposure to a particular
factor is, for example, increased, the expected return for that portfolio is also
increased (assuming that this risk factor commands a positive risk premium).
Thus, risk exposures and hence the implied expected return for a portfolio are
determined by a manager's stock selection.
In many applications, data are observed monthly, and the 30-day Treasury
bill rate is taken as a proxy for risk-free rate; that is, Po in equation (4) is
replaced by TB(t), the 30-day Treasury bill rate known to investors at the
beginning of month t. Then, for a model with N assets (i = 1, . . . , N ) and a
sample period of T time periods (t = 1, . . . , T), the data are the asset returns,
ri(t), the Treasury bill rates, TB(t), and the factor realizations, fi(t). From these
data, the statistical estimation problem is to obtain numerical values for the N
Pis and the (N x K)Pii)s. Discussion of this econometric problem is beyond the
scope of this paper, but the bibliography lists further readings that cover the
topic in detail.4
Macroeconomic Forces Impacting Stock Returns

Taking the time period to be one month and using the 30-day Treasury bill
rate as a proxy for the risk-free rate of return, the APT model, equation (4),
becomes:
See, for example, Brown and Weinstein (1983); McElroy, Burmeister, and Wall (1985);
Chen, Roll, and Ross (1986); Burmeister and McElroy (1988); and McElroy and Burmeister
(1988).
From h s point, there are three alternative approaches to estimating an APT
model:
..
1. The risk factors f,(t), f,(t), . , fK(t) can be computed using statistical
techniques such as factor analysis or principal components.
2. K different well-diversified portfolios can substitute for the factors (see
Appendix B).
3. Economic theory and knowledge of financial markets can be used to
specify K risk factors that can be measured from available macroeconomic and
financial data.
Each of these approaches has its merits and is appropriate for certain types
of analysis. In particular, the first approach is useful for determining the number
of relevant risk factors, or the numerical value of K. Many empirical studies
have indicated that K = 5 is adequate for explaining stock returns.
The estimates extracted using factor analysis or principal components have
an undesirable property, however, that renders them difficult to interpret; this
problem arises because, by the nature of the technique, the estimated factors
are nonunique linear combinations of more fundamental underlying economic
forces. Even when these linear combinations can be given an economic
interpretation, they change over time so that, for example, Factor 3 for one
sample period is not necessarily the same combination-in fact, it is almost
certainly different-as the combination that was Factor 3 in a different sample
period.
The second approach can lead to insights, especially if the portfolios
represent different strategies that are feasible for an investor to pursue at low
cost. For example, if K were equal to 2, one might use small- and large-
capitalization portfolios to substitute for the factors.
The advantage of the third approach is that it provides an intuitively
appealing set of factors that admit economic interpretation of the risk exposures
(the p i s ) and the risk premiums (the Pis). From a purely statistical view, this
approach also has the advantage of using economic information in addition to
stock returns, whereas the first two approaches use "stock returns to explain
stock returns." This additional information (about inflation, for example) will, in
general, lead to statistical estimates with better properties, but of course,
insofar as the economic variables are measured with errors, these advantages
are diminished.
Selecting an appropriate set of macroeconomic factors involves almost as
much art as it does science, and by now, it is a highly developed art. The
practitioner requires factors that are easy to interpret, are robust over time,
and explain as much as possible of the variation in stock returns. Extensive

research work has established that one set of five factors meeting these criteria
is the following:
f,(t): Confidence risk. Confidence risk is the unanticipated changes in
investors' willingness to undertake relatively risky investments. It is measured
as the difference between the rate of return on relatively risky corporate bonds
and the rate of return on government bonds, both with 20-year maturities,
adjusted so that the mean of the difference is zero over a long historical sample
period. In any month when the return on corporate bonds exceeds the return
on government bonds by more than the long-run average, this measure of
confidence risk is positive Cf, > 0). The intuition is that a positive return
difference reflects increased investor confidence because the required yield on
risky corporate bonds has fallen relative to safe government bonds. Stocks that
are positively exposed to this risk (P, > 0) then will rise in price. Most equities
do have a positive exposure to confidence risk, and small stocks generally have
greater exposure than large stocks.
f2(t): Time horizon risk. T i e horizon risk is the unanticipated changes
in investors' desired time to payouts. It is measured as the difference between
the return on 20-year government bonds and 30-day Treasury bills, again
adjusted to be mean-zero over a long historical sample period. A positive
realization of time horizon risk V2 > 0) means that the price of long-term bonds
has risen relative to the 30-day Treasury bill price. This is a signal that investors
require less compensation for holding investments with relatively longer times
to payouts. The price of stocks that are positively exposed to time horizon risk
(pi2 > 0) will rise to appropriately decrease their yields. Growth stocks benefit
more than income stocks when this occurs.
f3(t): Injation risk. Inflation risk is a combination of the unexpected
components of short- and long-run inflation rates. Expected future inflation
rates are computed at the beginning of each period from available information:
historical inflation rates, interest rates, and other economic variables that
influence inflation. For any month, inflation risk is the unexpected surprise that
is computed at the end of the month-the difference between the actual inflation
for that month and what had been expected at the beginning of the month.
Because most stocks have negative exposures to inflation risk (pis < 0), a
positive inflation surprise Cf, > 0) causes a negative contribution to return,
whereas a negative inflation surprise Cf, < 0), a deflation shock, contributes
positively toward return.
Luxury-product industries are most sensitive to inflation risk. Consumer
demand for luxury goods plummets when real income is eroded through
inflation, thus depressing profits for industries such as retailing, services, eating
places, hotels and motels, and toys. In contrast, industries least sensitive to
inflation risk tend to sell necessities, the demands for which are relatively
insensitive to declines in real income. Examples include foods, cosmetics, tires
and rubber goods, and shoes. Also, companies that have large asset holdings
such as real estate or oil reserves may benefit from increased inflation.
f4(t): Business cycle risk. Business cycle risk represents unanticipated
changes in the level of real business activity. The expected values of a business
activity index are computed both at the beginning and end of the month, using
only information available at those times. Then, business cycle risk is calculated
as the difference between the end-of-month value and the beginning-of-month
value. A positive realization of business cycle risk (f, > 0) indicates that the
expected growth rate of the economy, measured in constant dollars, has
increased. Under such circumstances, firms that are more positively exposed to
business cycle risk-for example, firms such as retail stores, which do well
when business activity increases as the economy recovers from a recession-
will outperform those such as utility companies that respond only weakly to
increased levels in business activity.
f,(t): Market-timing risk. Market-timing risk is computed as that part of
the S&P 500 total return that is not explained by the first four macroeconomic
risks and an intercept term. Many people find it useful to think of the APT as
a generalization of the CAPM, and by including this market-timing factor, the
CAPM becomes a special case. If the risk exposures to all of the first four
.
macroeconomic factors were exactly zero (if Pil = . . = Pi4 = O), then
market-timing risk would be proportional to the S&P 500 total return. Under
those extremely unlikely conditions, a stock's exposure to market-timing risk
would be equal to its CAPM beta. Almost all stocks have a positive exposure to
market timing risk (pis > O), and hence positive market-timing surprises Cf, >
0) increase returns, and vice versa.
A natural question, then, is whether confidence risk, time horizon risk,
inflation risk, and business cycle risk help to explain stock returns better than
the S&P 500 alone. This question has been answered using rigorous statistical
tests, and the answer is very clearly that they do.6
Market-timing risk is not required in an APT model that includes all the relevant
macroeconomic factors. As a practical matter, some relevant macroeconomic factor may be
difficult to measure or may not even be observable. Market-timing risk will capture the effects of
any such unobserved macroeconomic factor.
The probability that the first four macroeconomic factors do not add any information that is
useful for explaining stock returns is less than the probability that a standard normal variable (a
Risk Exposure Profiles and Risks--Return Trade-off

The risk exposure profile for the S&P 500 and the corresponding prices of
risk (the risk premiums) are shown in Table 1.7 For each risk factor, the
contribution to expected return is the product of the risk exposure (Column 1)
and the corresponding price of risk (Column 2), and the sum of these products
is equal to the expected return in excess of the 30-day Treasury bill rate
(Column 3). Thus, if the 30-day Treasury bill rate were, say, 5.00 percent, the
forecasted return for the S&P 500 would be 5.00 + 8.09 = 13.09 percent a year.
TABLE 1. Calculation of Expected Excess Return for the

S&P 500
Contribution of
Price of Risk Factor to
Risk Expected Return
Risk Factor Exposure x (%/year) = (%/year)
Confidence risk 0.27 2.59% 0.70%

T i e horizon risk 0.56 -0.66 -0.37
Inflation risk -0.37 -4.32 1.60
Business cycle risk 1.71 1.49 2.55
Market-timing risk 1.00 3.61 3.61
Expected excess return 8.09%
In general, then, for any asset, i, the APT risks-return trade-off defined by
equation (3) is:
random variable that is normally distributed with a mean of zero and standard deviation of 1)
exceeds 20 in value; that is, it is virtually zero. See McElroy and Burmeister (1988).
The model presented in this section uses parameters estimated by the BIRR@Risks and
Returns AnalyzeF ("BIRR is an acronym for Burmeister, Ibbotson, Roll, and Ross). The model
is re-estimated every month, and the examples here and in the next sections use numbers taken
from the April 1992 release, which is based on monthly data through the end of March 1992.
The Risks and Returns Analyze@ is a PC-based software package for doing APT-based risk
analysis with a model of the sort described in this paper. Although econometric estimation of APT
parameters (the risk exposures, P i s , and the risk premiums or prices, P's) is beyond the scope
of this paper, complete discussions of the more technical statistical issues involved in parameter
estimation can be found in Brown and Weinstein (1983); McElroy, Burmeister, and Wall (1985);
Chen, Roll, and Ross (1986); Burmeister and McElroy (1988); and McElroy and Burmeister
(1988).
where TB is the 30-day Treasury bill rate. The following four observations will
help clarify this risks-return trade-off:
1. The price of each risk factor determines how much expected return will
change because of an increase or decrease in the portfolio's exposure to that
type of risk. Suppose, for example, a well-diversified portfolio, p, has a risk
exposure profile identical to that of the S&P 500, except that it has an exposure
to confidence risk, P*,, of 1.27 instead of 0.27 (= P,,,,). Because the price
of confidence risk (from Table 1) is 2.59 percent a year, the reward for
undertaking this additional risk is 1.00 times 2.59-that is, the portfolio will
have an expected return that is 2.59 percent a year higher than the expected
return for the S&P 500.
2. APT risk prices can be negative, and they are for both time horizon risk
and inflation risk (P, < 0 and P, < 0). Consider first inflation risk. Almost all
stocks have negative exposures to inflation risk because their returns decrease
with unanticipated increases in inflation. Thus, the inflation risk contribution to
expected return is usually positive (the negative risk exposure times the
negative price for inflation risk equals a positive contribution to expected
return). That is, for most i, Pa < 0, and because P, < 0, P, x P, > 0 for
most i.
3. Many stocks have a positive exposure to time horizon risk (P, > O),
however, and thus, when the price of long-term government bonds rises
relative to the price of 30-day Treasury bills, their return increases. Because
the reward for time horizon risk is negative (P, < 0), time horizon's contribu-
tion to the expected return for such stocks is negative; for stocks with a
negative exposure to time horizon risk, its contribution is positive.
Why should this be the case? The answer is that, just as you pay for an
insurance policy that pays off when your house bums down, investors desire to
hold stocks with returns that increase when the relative price of long-term
government bonds rises. The fact that investors want to hold stocks having this
characteristic means that the prices of those stocks have been driven higher
than they otherwise would have been, and therefore, their expected returns are
lower. Thus, the negative price for time horizon risk produces the desired
result: stocks with larger (positive) exposures to time horizon risk also have
lower expected returns.
New York Stock Exchange is indicated by the vertical dashed line. Again, note
that the distribution is not normal and appears to be skewed to the right.
As is evident from Figure 1, the business cycle risk for Reebok is much
larger than for the S&P 500. These risk exposure profiles are shown below.
Exposure for Reebok Exposure for S&P 500
Confidence Risk 0.73

Time Horizon Risk 0.77
Inflation Risk -0.48
Business Cycle Risk 4.59
M a r k e t - T i g Risk 1.50
These exposures give rise to an expected excess rate of return for Reebok
equal to 15.71 percent a year, compared with the 8.09 percent a year computed
for the S&P 500. Figure 3 compares the risk exposure profiles for Reebok and
the S&P 500.9
In general, the risk exposure profiles of individual stocks and of portfolios
can differ sigdicantly. For example, Figures 4, 5, 6, and 7 compare the
respective risk exposure profiles for portfolios of low-capitalization versus
high-capitalization stocks, growth stocks versus the S&P 500, a value portfolio
versus the BIRR stock database, and a growth versus high-yield portfolio.
These risk exposure profiles define APT styles, and they enable us to view
traditional portfolio management styles from a new perspective that reveals
their inherent macroeconomic risks.
The usefulness to practitioners of risk exposure profiles and the risk-return
trade-off is an empirical issue. Abundant evidence shows that market indexes
are not mean-variance efficient; if so, the usual implementations of the CAPM
using some market index as a proxy are invalid. More importantly, recent
empirical evidence demonstrates that CAPM betas do not accurately explain
returns.
The multifactor APT approach has far greater explanatory power than the
CAPM. Many econometric studies have verified the superior performance of
models that include multiple factors (Postulate 1of the APT) to explain returns
and that use multiple factor premiums (Postulate 2 of the APT) to explain
expected returns. These results are discussed in some of the papers listed in
The BIRR Risk Index plotted in this and the following graphs is a single number that gives an
approximate answer to the question, "Does A have more systematic risk, relative to the market,
than B?"
How to Use the APT: Some Examples

A primary concern for practitioners is not only to acquife an understanding
of the APT but also to learn how to use it to enhance their investment
performance. So far, we have concentrated on explaining the APT; now, we will
briefly discuss several uses of the APT that every practitioner could easily
apply. The following list is chosen to be exemplary of some widely used APT
techniques, but it is by no means exhaustive.
Evaluationof Macroeconomic Risk Exposuresand Attribution of

Return. Risk exposure profiles can vary widely for stocks and portfolios.
They are determined by the risks a manager undertakes through stock
selection, and in turn, determine a manager's APT style. A basic first task,
then, is to identify the risk exposure profiles for portfolios. Usually, managers
will want to compare their risk exposure profiles with those for appropriate
benchmarks. A small-cap manager, for example, should know whether his or
her portfolio differs in its exposure to macroeconomic risks from an appropriate
index of small-cap firms. Any differences will account for performance differen-
tials from the index. Only if the risk exposures are the same as the index can
ex post superior performance be attributed to APT selection-selection of
individual stocks that returned more than would be expected on the basis of the
risks undertaken.
Whatever the manager's risk exposure profile, the APT should be used to
divide the mean ex post actual return into: (1) expected return, which is the
reward for the risks taken, (2) unexpected macroeconomic factor return, which
arises from factor bets and factor surprises, and (3) a , which arises from stock
selection. Moreover, expected and unexpected factor return can be attributed
to the manager's risk exposure profile. Thus, APT analysis will provide a better
understanding of the true sources of actual portfolio performance.
Index Portfolios. A closely related use of the APT is in the formation of

index portfolios designed to track particular well-diversified benchmarks. The
APT provides powerful tools for tracking any such benchmark portfolio. A
tracking portfolio can be constructed simply by forming a portfolio with a
matching risk exposure profile. The expost APT a can be made small by making
the tracking portfolio well diversified so that the portfolio-specific return, call it
E*, is near zero.
Tracking a benchmark that itself is not well diversified, in the sense that its
expost a usually is not near zero, is more difficult. In this case, not only the risk
exposure profiles but also the benchmark's a must be matched. One way to do
this is to form the tracking portfolio by random sampling from the stocks that
constitute the benchmark.
Tilting, or Making a Factor Bet. Good managers may possess

superior knowledge about the economy. Suppose, for example, a manager
believes that the economy is going to recover from a recession faster than most
market participants do. If the manager is correct in this belief, the realizations
of business cycle risk will be positive Cf4 > 0), and stocks that have greater risk
exposures to business cycle risk (stocks for which Pi, is larger) will, ceteris
paribus, outperform.
To take advantage of this superior knowledge, the manager will want to
make a factor bet on (or tilt toward) business cycle risk-alter the existing
portfolio to increase its business cycle risk exposure without changing any other
macroeconomic risks. Conversely, if a manager has special knowledge that the
economy is going to slide into a recession, he or she will want to lower the
portfolio's exposure to business cycle risk.
Multimanager Fund Performance. Most sponsors employ more

than one manager. Even though each may perform well when compared with a
particular style benchmark, that is not the issue of most importance to a
sponsor. A sponsor wants to evaluate the risks and performance of the overall
fund.
The sponsor should combine the portfolios of individual managers into one
overall fund portfolio and then use the APT to examine the risk exposure profile
and performance of the fund portfolio. Often, the combination of managers leads
to risk exposures that the sponsor finds uncomfortable. If so, funds should be
reallocated among the managers to achieve the desired fund risk exposure
profile.
The sponsor must also examine whether or not the overall fund return
exceeds the benchmark and determine the sources of differences.
Optimized Risk Control with Manager-Supplied Rankings.

Many managers have their own proprietary methods for evaluating stock return
performance, yet lack adequate methods for estimating their accompanying
risks. The APT, or more accurately, part of the APT, is a perfect tool for such
managers.
To keep matters simple, suppose a manager has a personal ranking system
that scores every stock on a scale from 1to 10, where 10 is the score given to
the stocks in the best expected return category. The objective is to emulate the
volatility of the S&P 500 but achieve a higher return. How could the manager
use the APT?
Let si be the score from 1 to 10 assigned to the ith stock (i = 1, . . . , N).
The formal problem is to find portfolio weights, w,, w,, ..
. , wN, for the N
stocks in the selection universe such that the portfolio score is maximized but
the risk exposure profile is similar to that of the S&P 500. More formally, the
weights should result in the highest possible value for
subject to the constraint that the portfolio betas,
..
for j = 1, . , K, are close to the betas for the S&P 500. That is, the weights
should make the risk exposure profile for the portfolio close to the risk exposure
profile for the S&P 500 while maximizing the value of the portfolio's ranking
score. If the ranking system works, the return will be superior to the S&P 500.
If the resulting portfolio is well-diversified, it and the S&P 500 will have
approximately equal volatilities. The proper diversification can be achieved by
making N sufficiently large and by imposing a maximum value for the weights so
that the portfolio contains a large number of stocks. This optimization problem
is easily solved using linear programming.
Long-Short Investment Strategies. Long-short, or market-neu-

tral, investment strategies are receiving increased attention. The pure APT
view of such strategies will be discussed first; then, it will be shown how
managers with superior knowledge can use the APT to implement those
strategies effectively.
Suppose a manager holds a long portfolio with return rL(t) and a short
portfolio with return rs(t); both have equal dollar values. Let the risk exposures
.
for these portfolios be denoted by pLjand psi, j = 1, . . ,K. Assuming that the
short position earns the 30-day Treasury bill rate, the manager's total return is
Now, let the risk exposure profile on the long portfolio exactly match the risk
exposure profile on the short position. Then, using equation (3), the expected
returns on the long and short portfolios are equal, the expected return to the
long-short strategy is simply TB(t), and the variance of the realized return is
Because no stock is held in both the long and short portfolios, this variance is
approximately
The position has greater volatility than 30-day Treasury bills but no greater
mean return. Therefore, it is not a very attractive strategy, particularly after
trading costs.
This strategy could become attractive if the APT alphas on the long position
were sigdicantly larger than the APT alphas on the short position; that is, it is
an attractive strategy for a manager with superior APT selection. Consider an
exceptional manager who can pick two well-diversified portfolios of stocks, with
no stocks in common, such that ol, > 0 for the long portfolio and a, < 0 for the
short portfolio. If the manager also can match the risk exposure profiles of the
long and short positions, the return would be oc, - as + TB(t)with a volatility
approximately equal to that of 30-day Treasury bills.
The APT can play a crucial role for such a manager: It provides an easy and
quick way to match the risk exposure profiles of the long and short positions. As
an example of this role of the APT, we constructed a long portfolio consisting
of approximately 50 NYSE-listed stocks with the largest ex Post alphas over a
sample period of 72 months (April 1986 to March 1992). We then computed the
risk exposure profile for this long portfolio. A short portfolio of approximately 50
NYSE-listed stocks, not in the long portfolio, was also selected. An optimization
problem was solved to find portfolio weights for the short position that matched
its risk exposure profile to that of the long position. The resulting risk exposure
profile for the overall long-short strategy is illustrated in Figure 8; it has
essentially zero systematic risk. The sole source of volatility (beyond the
volatility of 30-day Treasury bills) for this long-short strategy comes from the
E'S for the long and short positions. By having portfolios of 50 stocks or more,
this volatility can be kept small.
The performance of this long-short, or market-neutral, strategy for the
most recent 12 months of the sample period (April 1991 to March 1992) is
illustrated in Figure 9. The mean realized return was 30.04 percent a year,
compared with 11.57 percent for the S&P 500, and the standard deviation of
this realized return was only 6.26 percent a year, compared with 18.08 percent
for the S&P 500.
Mean--Variance Efficiency. The standard optimization problem of

finding the portfolio with the highest expected rate of return for a given variance
is easily solved within an APT framework. For this problem, the expected
return could either be given by the APT equation, equation (3), or it could come
factors do not contribute to expected return, they do help to explain

volatility, and they provide managers with a tool to evaluate the
diversification of their portfolios.
Other enhancements are being invented every day. As more and more tools
become available and as understanding of the APT spreads, so does its
application to portfolio management problems.
Appendix A
To derive the restrictions that a multifactor CAPM must obey, suppose that the
CAPM were true for some market index of N assets. This index has a return
denoted by rm(t)and has weights wml, w, .. ., wd summing to 1. Suppose also
that Postulate 1of the APT holds, that is, that the N asset returns are generated
by the linear factor model (LFM) given in equation (1). We will then show that the
APT is valid and find the CAPM restrictions that the APT risk prices must satisfy.
This problem is solved by recognizing that the CAPM beta for any asset can
be computed as a linear function of the LFM risk exposures; that is, the CAPM
beta is equal to a linear function of the APT Pis.
First note that the return on the market index is
and hence is generated by a LFM with

Pmj = Wml X Plj + . . + w m X~PNj for j = 1, . . . , K.
Using Footnote 1, the CAPM beta for the ith asset is
The latter can be computed from the LFM generating the return for the ith asset:
Because by Postulate 1cov[e,(t), fi(t)] = 0, it follows that cov[a,(t), rm(t)] =

cov[ei(t), ~,(t)]. Thus, under the usual assumption that the market index is well
diversified and am(t)is approximately zero, we may set the last covariance term
in the above expression for pi equal to zero.
Under the CAPM, E[ri(t) - TB(t)] = Pi x E[rm(t) - TB(t)]. The APT is
true when there exist numbers P I , . . . , PK such that
It then follows immediately that the APT holds provided that
for all j = 1, . . . , K. Conversely, if the APT is true and the above K CAPM
restrictions on the Pis hold, then the CAPM is also true. Given an LFM for
asset returns, these are the CAPM restrictions that are rejected in favor of the
APT in statistical tests.
Appendix B
We will show that K well-diversified portfolios can substitute for the factors in
an APT model. To simplify the computations, we assume that K = 2; the
general case is easily handled using matrix algebra. Thus, suppose that two
different well-diversified portfolios have returns given by
and
Also assume that the risk exposure profiles for the two portfolios are not
proportional. We will show below that
(a) The APT equation for the return on the ith asset, ri(t), given by equation
(5), can be rewritten in terms of the portfolios with returns Rl(t) and
R2(t).
(b) Given the answer to (a), E[ri(t)] can be expressed in terms of the
expected returns for the two portfolios.
To prove (a) and (b), we introduce the following simphfymg notation:
In this notation, the APT equations for the two portfolios are
and
Takingy,(t) andy2(t) as given, the latter are two equations in two unknown z's,
and they may be solved for
and
where
and
Note that as long as the risk exposure profiles for the two portfolios are not
proportional, 6 # 0 and the solution given above exists.
Given these results, with straightforward algebraic manipulation, equation
(5) may be rewritten as
where
and
This exercise establishes (a) above.

Finally, taking expectations of the latter equation gives
This formulation proves (b) above.

Multi-Index Models Using
Simultaneous Estimation of all
Parameters
Edwin J. Elton
Martin J. Gruber
Leonard N. Stern School of Business
New York University
The concept of index models and their role in explaining and understanding the
pattern of security returns, what affects individual security returns, the
selection of optimal portfolios, and the level of relative long-run (equilibrium)
returns have been widely discussed. We will review these techniques as an
introduction to the concepts of estimating multi-index models.
Single- vs. Multi-Index Models

Most of our analysis will be in the context of a generic multi-index model and
a generic single-index model. The single-index model is included because it is
familiar and because it serves as a useful benchmark against which to judge
multi-index models. In addition, many of the problems with the multi-index
model, and the economic intuitions behind the solution to those problems, can
be illustrated for the case of a single-index model.
The single-index model is simply a way of decomposing return on an asset
into two parts. The first part, the systematic part, is the portion of return
affected by influences common to all assets. The second part is the unsystem-
atic part, which is assumed to be unique to the asset. Thus,
Asset
where
Rit = is the return on security i in time t,
bil = is the sensitivity of security i to returns on Index 1,
f1, = is the return on Index 1 in period t,
ai = is the expected level of nonindex-related return for security i, and
eit = is a random variable with mean of zero and variance 0:;;
Note that the unique return is also split into two parts: its mean level, a , and
its variability, ei, For the single-index model to be a reasonable description of
reality, the unique part should be truly unique to the security in question and not
related to another influence. Technically, this means that the value of the unique
return for security i in period t is unrelated to the value for security j in period
t. Because the a's are constants, this condition means
E(eigjt)= 0 for all i and j, where i # j.
Likewise, for the researcher to have correctly divided the return into its
systematic and unique parts, the unique return must be unrelated to the index
return so that
E(eiJlt)= 0 for all securities.
The single-index model describes return in terms of one common influence,
and the multi-index model describes returns in terms of more than one common
influence. Its structure is exactly the same as that of the single-index model
except for the inclusion of additional indexes. Thus, the model can be written as
where
6, = the return on the jth index affecting stock return, and
bC = the sensitivity of security i to the jth index.
As in the single-index case, unique return is assumed to be uncorrelated
across all securities; that is,
E(eigjt)= 0 for all i and all j, where i # j,
and systematic influences are assumed to be independent of unique influences: l
E(eitfit)= 0 for all i and all j.

The major reason for going to a multi-index model is the belief that influences
beyond one index cause securities to be correlated with each other (move
together). In fact, additional indexes are introduced in an attempt to have the e,
be the unique influence so the covariance between the residuals in equation (2)
will be approximately zero. This attempt to find a set of indexes that captures
all sigmficant influences that affect multiple securities results in an approximate
zero covariance between residuals that will be the key to designing a multi-index
model.
To this point, we have intentionally defined our analysis in the terminology
of multi-index models. Historically, starting with Sharpe's single-index model,
this terminology was commonly used both in the literature and practice of
financial analysts. With Ross's (1976) description of the arbitrage pricing theory
(APT) and the initial tests of Roll and Ross (1980) of the APT methodology, a
different terminology came into existence. Expressions like equation (2)
became known as a multi-factor return-generating process. The fi, which we
denoted as returns on indexes, became known as factors, and the bii became
factor loadings.
Estimating the Multi-Index Model

To use a multi-factor (multi-index) model, we need to estimate the
right-hand side of equation (2). We need estimates of both thefi factors and the
factor loadings, bii. Either the fi or the bG can be asserted on a priori grounds
with the other identified empirically, or both can be identified empirically.
Clearly, estimating one set of parameters should be easier than estimating
both sets simultaneously, but we do not do so because we do not know what the
correct set is. We can illustrate this problem with the case of the single-index
model. Usually, when estimating a single-index model, the index is taken to be
the return or excess return (over the riskless rate) on some widely diversified
portfolio. The most commonly used portfolios are the S&P 500 Index or the
CRSP Index. One of these is simply asserted as being the relevant factor.
If the Sharpe-Lintner version of the CAPM is correct, the right index to use
An additional assumption that is frequently made is that the indexes are uncorrelated. This
assumption does not create problems, because a set of correlated indexes can always be
converted to a set of uncorrelated indexes.
for explaining equilibrium returns is an index of the return on all risky assets in
which the weight on each return is the relevant market proportion of that asset.
If the model is not correct, there is no theoretical way of identdymg an
appropriate index and the definition of the relevant index is not clear-cut.
Furthermore, even if the standard CAPM is a relevant equilibrium model, no
one would be able to calculate the return on a market-weighted index of all
assets. Although a market-weighted index of equities is readily available,
market-weighted indexes of other assets, such as real estate, are not.
Plausibly, a non-market-weighted index of equities (one that places greater
weight on equities correlated with excluded assets such as real estate) could be
a better representation of the true "market" index than is a market-weighted
index of equities alone.
If we feel a multi-index model better represents the return structure, the
problems become more severe. No theoretical multi-index equilibrium model is
generally accepted. Although alternative theory suggests certain broad influ-
ences that might affect equilibrium returns, these influences are not easily
translated into empirically measurable influences.
An alternative to prespecdymg indexes is to try to have the historical return
series itself suggest what portfolios of securities would best serve as indexes.
Equation (1) and its explanation suggest the characteristics these indexes
should possess. In particular, they should separate the common influences in
returns from the unique influences in returns. After we have specified the
indexes, the unique returns on securities should be uncorrelated with each
other. In addition, the structure should be parsimonious; that is, returns can be
described in terms of a limited number of indexes. Finally, having the indexes
represent separate influences would be desirable.
Two statistical techniques accomplish these goals: factor analysis and
principal components analysi~.~ The most common technique is factor analysis.
Factor analysis was devised to define a set of indexes mathematically so that the
covariance between security returns is minimized after the indexes have been
removed. This assures that cov(e,ej) is as close to zero as it can be.
More precisely, once the user sets the number of indexes desired, factor
analysis wik
Define the optimal composition of each index (the weight on each

security in each index),
See the appendix to this chapter for a more detailed discussion of factor analysis and principal
components analysis, as well as the differences between them.
Speclfy the return of each index at each point in time (this is the same
concept as the return of the S&P Index at each point in time),
Calculate the bG, or sensitivity of each stock to each index, and
Measure the average explanatory power of the model for each stock.
One can then repeat the analysis for a different number of factors (indexes)
and determine the probability of needing to add another factor to the model.
Design Issues
Although the idea of letting the data design the model has a lot of appeal, in
statistical methodology, as in economics, there are few free lunches. The
techniques come with their own problems and their own set of choices. We will
discuss four of these: the effect of the choice of data, the number of indexes to
use, indeterminacy of the model, and computational diaculties.
The Choice of Data. The input to factor or principal component analysis

is a sample of security returns. In preparing the return data, the researcher
must select both the time period of returns and the sample of stocks (or
portfolios of stocks) to use to estimate a factor structure. Ideally, the structure
will hold for time periods and securities beyond those used in the estimation
sample.
Obviously, the data cannot suggest a factor if that factor was not present
during the time period chosen. For example, assume that changes in oil prices
affect equity returns. Using returns from a period with minimal changes in oil
prices will probably mean that changes in oil prices had a very small influence on
security returns in the period and that this factor will not be recovered by factor
analysis. Similarly, if the researcher selected a sample of stocks that happened
to have only a few stocks that are sensitive to changes in oil prices, the influence
of oil prices would not show up as a factor.
Researchers frequently attempt to mitigate computational problems by
factor analyzing the returns on portfolios of securities rather than returns on
individual securities. This technique can introduce other problems into the
analysis. For example, when factor analyzing a time series of returns, generally
the first index is a portfolio of most of the assets whose returns are being factor
analyzed and the sensitivity of all assets to the index is positive. This is not true
with subsequent indexes, however. Additional indexes generally measure
nonmarket influences. For example, it is not uncommon when factor analyzing
equity returns that the second index is highly correlated with firm capitalization
(captures the small-stock phenomenon) with the effect of the market removed.
Some securities will have positive sensitivities to the influence of size and
others negative. The sensitivities on a portfolio of securities is a weighted

average of the sensitivities of the securities that compose it. If some sensitiv-
ities are positive and some negative, the portfolio's sensitivity could be zero or
close to zero. Factor analyzing returns at the portfolio level could mean that
important influences would not be detected.
The Number of Indexes. A second design issue that affects results is

the number of indexes in the return-generating process. The factor analytic
techniques speclfy, for a fixed number of indexes, the indexes that best
separate out common influences from unique influences. Although statistical
techniques can determine whether adding another influence "statistically"
improves the explanatory power of the model, common sense and economic
significance play a major role in deciding on the number of factors to analyze.
For example, does the composition of the last index make intuitive sense in
terms of capturing an influence the analyst feels might affect security prices. For
testing economic sigmficance, the analyst will examine such issues as whether
the index improves the ability to construct portfolios that match a market index
and whether the index helps explain the time series of security returns in a
period other than when the model was fit.
The Nonuniqueness of Factors. A third design issue is concerned

with the indeterminacy of the structure of the multi-index model.3 An infinite
number of models can separate returns into systematic and nonsystematic
components. For example, consider the single-index model, and assume the
index is the return on the S&P Index. In a second model, the index is half the
return of the S&P Index and all the sensitivities are doubled, but the result is
the same separation of return into systematic and unsystematic parts.
Multi-index models become even more complex. An infinite number of
specifications will result in the same separation of systematic and unsystematic
returns.4 When the indexes or the sensitivities are prespecified, generally a
natural scaling occurs and a set of indexes suggests itself. When a researcher
uses techniques that produce a model that best captures the past return
structure, he or she must realize that the resulting structure is not unique.
Some researchers will examine alternative structures in an attempt to under-
Principal components analysis does provide one particular determinate solution to the factor
solution. See the appendix to this paper.
In technical terms, solutions are determined only up to a linear transformation of the factor
structure.
stand what influences are affecting security returns and to convince themselves
that the overall separation makes intuitive sense.
Nonuniqueness is a concern in certain applications and not in others. Any
solution is correct in the sense that it explains (and predicts) returns as well as
any other solution. Some solutions, however, are easier to interpret econom-
ically than others. Also, two researchers using slightly different solution
algorithms or slightly different samples can come up with solutions that are
simply transformations of each other but that appear to be very different.
Computational Problems. A fourth problem with factor analysis is the

difficulty of factor analyzing returns for a large number of securities. The
primary problem is computational. Analyzing a large number of securities is
costly and impossible to do with most standard statistical packages. To get a
meaningful factor analysis, however, requires a longer data series than the
number of firms being analyzed. With 20 years of monthly data, one is restricted
to fewer than 240 firms.
Two procedures have been introduced to deal with difficulties in factor
analyzing a large number of securities. The first involves performing factor
analysis on subsets or groups of securities. This multisample approach was
introduced by Roll and Ross (1980). The second is the portfolio approach
introduced by Chen (1983).
In the group, or multisample, approach, securities are divided into samples
and maximum likelihood factor analysis is used to extract factors from each
sample. The difficulty with this approach is that, because of indeterminacy, the
factors extracted in each sample need not be extracted in the same order or
even with the same sign. How then can one determine whether the same
factors or different factors (unique to each group) have been extracted in going
from group to group? The solution to this problem represents an opportunity to
improve our identification of the return-generating process. By analyzing the
factors produced from different groups and identifying those influences that are
common across two or more groups, we can differentiate between factors
unique to one group and general factors that apply to all groups.
Chen's approach rests on examining the return behavior of portfolios of
securities rather than individual issues. This approach has two disadvantages.
F i s t , no matter how large the sample used to form portfolios, one cannot tell
whether the factors are unique to that sample or common to all securities.
Two papers, Cho, Elton, and Gruber (1984) and Brown and Weinstein (1983), discuss
techniques for identifying the common factors across groups.
Second, the results obtained by this methodology are very sensitive to the
portfolio formatioli technique that was used.
Applications of Multi-Index Models

In this section, we will discuss one simple example to illustrate the principles
discussed earlier, and then we will describe two applications: one to the
Japanese equity market and the second to the U. S. bond market. The purposes
of these examples are to illustrate the basic ideas discussed earlier and to
convey to the reader the kinds of tests and verification that are needed to be
confident that the models are useful.
A Simple Example. To illustrate the principles discussed earlier, we

ran a principal components analysis on historical return data for four market
indexes. The indexes were the Morgan Stanley Capital International return
indexes for common stocks in Canada, the United States, France, and Belgium.
The data were monthly return data for the decade ending December 1988. The
Morgan Stanley indexes are market-weighted indexes of the major stocks in
each market. Table 1shows the variance-covariance matrix and the correlation
matrix. The variance entries are on the diagonal of the matrix, the covariance
entries are above the diagonal, and the correlations are below the diagonal. The
returns are positi;ely correlated between all pairs of countries, and the highest
correlation is between the two North American countries and the two European
countries. Thus, two indexes will be needed: a general world index and an index
that reflects whether the country is North American or European.
TABLE 1. Variance--Covariance Matrix and Correlations
Country Belgium Canada France United States
Belgium 50.34 17.55 34.22 13.79

Canada 0.38 42.41 19.64 22.39
France 0.65 0.41 55.11 15.64
United States 0.41 0.72 0.43 23.03
We ran a principal components analysis on the variance-covariance matrix

and obtained two explanatory factors. The return on one factor was given by
where
R = return,
f = factor value,
B = Belgium,
C = Canada,
U = United States,
F = France, and
t = time period.
Note from equation (3) that the first index is very close to an equally weighted
index of the four country indexes. Thus, we might label the first factor a "world
factor.
The second factor is
This index is long in North America and short in Europe. It could be

characterized as measuring the performance of North American stocks relative
to Europe. The returns for each country and on each of the two factors for the
last four months of our sample period are shown in Table 2.
TABLE 2. Returns in Countries and on Factors
Period Belgium Canada France U.S. Index 1 Index 2
1 2.08% 14.64% 7.25% 6.08% 10.99% 9.85%

2 -4.95 13.41 -0.89 -1.03 -0.84 11.83
3 -16.57 -22.03 -15.95 -8.82 -34.85 -7.16
4 18.72 3.36 11.00 4.33 17.24 -7.13
Mean for 120
months 2.16 1.02 1.54 1.33
Note: To get the return on the index, subtract the mean return shown at the bottom of the table from each
month's return.
The b's (the sensitivities, or factor loadings) could be estimated by

regressing the country returns on the two factors. Alternatively, the sensitiv-
ities are an output of the analysis. As one would expect, the b's are positive for
the North American countries and negative for the European countries on the
second index. The R2s (explanatory power) of the two-index model for each
country were 0.81 for Belgium, 0.95 for Canada, 0.84 for France, and 0.74 for
the United States. The two-factor model does an excellent job of explaining
returns.
A Multi-Index Model of the Japanese Stock Market. We selected

Japanese data to demonstrate factor analytical solutions because of general
interest in the Japanese markets and because Japan is an economy for which the
multi-index models have an especially clear-cut advantage over the single-index
model.6 The sample of stocks originally selected for study was the 400 stocks
that make up the Nomura Research Institute (NRI) 400 stock index. The
Nomura Index, like the S&P Index, is market weighted. The 400 stocks in the
index are among the largest in Japan, and attention is given to industry balance
in selecting them. The 400 stocks in this index represent more than 60 percent
of the total market value of all stocks listed on the Tokyo stock exchange. Seven
of the stocks had incomplete data and were dropped from our sample.
We divided the remaining 393 securities into four approximately equal-sized
samples. We then performed maximum likelihood factor analysis on the
variance-covariance matrixes for each of these groups. The input data were the
returns for each security for 180 months.
How many factors? The first problem in doing a factor analysis is
deciding on the appropriate number of factors to include in the model. Tests of
the appropriate number of factors can be separated into those that look at one
group at a time and those that take advantage of the fact that solutions for more
than one group have been obtained. Although we will emphasize the latter
because it focuses on the commonality of the factors across groups, we will
briefly examine the standard single-group tests.
Single-group tests. To obtain an initial idea of how many factors might be
present in the return-generating process, ten factor solutions (using maximum
likelihood factor analysis) were performed on each of our four samples. That is,
for each sample, a maximum likelihood solution involving first one factor, then
two factors, then three factors, and continuing through ten factors was
calculated. Three techniques in the literature on factor analysis are commonly
used to decide on the correct number of factors needed to explain the
covariance matrix. These are chi square (Lawley & Maxwell 1971), information
This section is based on an article by Elton and Gruber (1990). Other authors have
documented the success of alternative forms of multi-index models in Japan. See Brown (1990)
and Harnao (1990). The reason the multi-index model works especially well in Japan is the higher
residual covariance between security returns after the iduence of the market has been removed.
This result may be attributable to some structural difference between the U.S. and Japanese
economies or it may be attributable to the high percentage of corporate cross-ownership in Japan.
criteria (Akaike 1974b), and Baysian criteria (Schwartz 1978). Of the three,
Schwartz's technique is the most conservative in estimating the number of
sigmficant factors. The chi square test and Akaike's information criteria tend to
include factors that, although statistically sigmficant, have little economic
importance. When we used either of those two tests, the results showed that,
for each sample, at least ten factors were present in the return-generating
process. We did not test for the optimum number because we had not
performed factor solutions involving the extraction of more than ten factors.
Also, alternative tests, as well as previous attempts by others to identify the
return-generation processes for other samples of securities, indicated the
presence of fewer than ten factor^.^
Schwartz's Baysian criteria provided a much more parsimonious description
of the return-generating process. Schwartz's method produces a statistic that
reaches its minimum at the "correct" number of factors extracted. The value of
Schwartz's statistic for each group for alternative numbers of factors (one
through ten) is shown in Figure 1. Schwartz's criteria identified three factors as
'
sigmficant in the return-generating process for Sample 1 and four factors for
Samples 2, 3, and 4.
The conclusion from examining the number of factors present in the
return-generating process of each sample separately is somewhat ambiguous. It
rests on the choice of the test used to determine sigmficance. The answer
would seem to be either four, ten, or more factors. This ambiguity illustrates
how cautious one should be about placing too much reliance on statistical
sigmficance in deciding on the number of factors. The next step is to use
information from more than one group to decide on the number of factors.
Mult$le-group tests. The intent of our analysis was to estimate a return-
generating process that describes the return on all stocks that are comparable
to the stocks in the NRI 400 Index. For any one of the four samples, as more
factors were added to the solution, the probability increased that the added
factors are idiosyncratic to the stocks in that sample or a subset of those stocks
rather than factors that explain the covariance structure of returns among large
groups of securities. If, in fact, the factor solution from a sample captured
general influences, then the solution from a second group should reflect the
same general influences. We could not, however, simply compare the first
factor from a sample with the first factor from another sample and the second
with the second, and so forth. Factor solutions are only unique up to a linear
transformation. Therefore, the first factor from one sample may be the second
See, for example, Roll and Ross (1980), Hamao (1990).

FIGURE 1. Schwartz's Baysian Criterion
/
/
/ '
\ /
/.
...
\ /.
'. --. ./ .
5.0 - .
'. -_ -.-...
0
0.
- .
.
I I I I I I I I
4.8
1 2 3 4 5 6 7 8 9 10
Number of Factors
Group 1
- - - Group 2
. .. . . . . Group 3
Group 4
factor from a second sample or even a linear combination,of the first, second,
third, and fourth factors from the second sample. Although some attributes of
maximum likelihood factor analysis will tend to extract factors in a parallel
manner and to some extent lessen this problem, they do not eliminate it.*
This problem can be corrected by canonical correlation. For an n-factor
solution, find that linear combination of the n factors from one sample that is
most highly correlated with the best linear combination of the n factors from a
second sample. After removing this correlation, find the second linear combi-
nation of the n factors from the first sample that is most highly correlated with
the linear combination of the n factors from the second sample. This process is
repeated n times. By finding best-fit linear combinations, canonical correlation
removes the problem of factors from one sample being linear transformations of
See Roll and Ross (1980) and Dhrymes et al. (1984) for a debate on this issue.
factors from a second sample. If, in fact, one has estimated too many factors,
then after removing the common factors, the remaining canonical variates
should be uncorrelated.
Table 3 presents the average (across four samples) squared canonical
correlation for the first canonical variate out of the one-factor solution, the
second canonical variate out of the two-factor solution, proceeding through the
seventh canonical variate out of the seven-factor solution. The results indicate
that the likely solution is either four or five factors. The canonical R2 of the
fourth linear combination from a four-factor solution is almost 60 percent; for
the fifth linear combination from a five-factor solution, it is slightly more than 20
percent; but for the sixth linear combination of the six-factor solution, it is less
than 5 percent.
TABLE 3. Canonical R, ith Canonical R* from the i Factor

Solution
The evidence so far would seem to support a four-factor solution with the
fifth factor worth considering. Although we have argued that canonical corre-
lation is the correct way to determine whether factor structures from one group
are the same as those from a second group, the simple correlation pattern
between factors is also worth examining in order to see the type of orthogonal
transformation that can take place. Table 4 presents the simple correlation
between the factors extracted from Samples 1 and 2 for the four-factor and
five-factor solutions. Note that in the four-factor solution, the only correlations
above 0.10 occur for the first factor from Sample 1 with the first factor from
Sample 2, the second with the second, and so forth. For the five-factor solution,
however, this clear pattern fractures. In addition, some factors from one group
do not seem to be associated at all with factors from the other group. For
example, Factor 5 from Sample 1and Factor 4 from Sample 2 have only minimal
correlation with any of the factors from the other group. This result also
supports the four-factor solution.
TABLE 4. An Example of a Factor Pattern: Correlation of

Factor between Samples 1 and 2
Sample 1
Sample 2 F1 F2 F3 F4 F5
Four factors
F1
F2
F3
F4
Five factors
F1
F2
F3
F4
F5
Note: Dash = less than 0.1.
With one four-factor solution for each of the four groups, the problem could
be which solution to accept, but this is not really a problem. Each four-factor
solution is close to a linear transformation of any other four-factor solution.
Therefore, they should, for all practical purposes, work about equally well in
explaining returns. In fact, they all have about the same explanatory power
across securities and portfolios of securities. For example, the R2 of the Tokyo
Stock Exchange (TSE) varied between 0.902 to 0.928 across the four different
factor solutions.
The model's explanatory power. Having determined that returns are
related to four factors and having produced a particular four-factor solution, the
next step is to examine how much of the total return these four factors explain
and to compare this result to the amount explained by the more conventional
single-index model.
To examine this question, we used returns on 20 groups obtained by ranking
the securities in the NRI 400 by size (total equity asset capitalization). Grouping
by size will, of course, increase the amount explained by almost any model. At
the same time, it creates a manageable set of data that allows examination of
average explanatory power and explanatory power across sets of stocks.
Table 5 shows the sensitivities and adjusted R2s when the returns on each
of the 20 portfolios are regressed against the four factors for the 15-year period
April 1971 to March 1986. Across the 20 portfolios, the average adjusted R2 is
78 percent. Of the 80 different sensitivity estimates, all but 18 are significant at
the 5 percent level.
TABLE 5. Sensitivities and Explanatory Power for the

Four-Factor Model
Beta Coefficients
Portfolio F1 F2 F3 F4 Adiusted R2
1 0.0428 0.0092
2 0.0442 0.0048
3 0.0427 0.0034
4 0.0417 -0.0003"
5 0.0421 0.0041
6 0.0396 0.0022"
7 0.0388 0.0077
8 0.0380 0.0078
9 0.0417 0.0059
10 0.0356 0.0025"
11 0.0375 0.0023"
12 0.0384 -0.0007"
13 0.0347 -0.0037
14 0.0374 -0.0029"
15 0.0368 -0.0037
16 0.0384 -0.0059
17 0.0330 -0.0087
18 0.0385 -0.0004"
19 0.0364 -0.0022"
20 0.0364 -0.0065
Average
"Insignficant at the 5 percent level.
To compare these results with those for the single-index model, returns on
the 20 portfolios were regressed against the NRI 400 index. Because the NRI
400 index is value weighted and because it is made up of the same 400 stocks
used to form the 20 groups, the relationship between the 20 portfolios and the
index is likely to be higher than if we had chosen another market index. Table
6 shows the results. As one might expect given the construction of the index,
TABLE 6. Sensitivity and Explanatory Power for One-Index

Model
Portfolio Beta Coefficient Adjusted R2 Alpha Average Return
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
Average
the R2 declines with size. The extent of the decline is dramatic. The adjusted
R2 is less than 50 percent for the smaller eight portfolios and less than 15
percent for the portfolio of smallest stocks. The four-factor model explains
considerably more of the time series of security returns than does the
single-factor model. The average adjusted R2 is 55 percent compared with 78
percent for the four-factor model. 9
Comparing the explanatory power of the first factor in the four-factor model
with the explanatory power when all four factors are included shows that the
added three factors explain a significant proportion of the variability of returns.
The sensitivity of portfolio returns to the NRI 400 index (beta) also declines
That the four-factor model has a higher explanatory power than a single-factor market model
in the fit period is to be expected. The size of the difference, however, is not the same as for U.S.
studies, nor is the deterioration with size.
with size, which was not at all expected, because beta is usually considered a
measure of risk. For U. S. data, the beta coefficient increases as size decreases,
so smaller firms are viewed as having greater risk. For Japanese data, the
reverse is true. This result must be interpreted with some caution, however.
The firms in the sample are all fairly large. The 400 companies that compose the
NRI 400 are selected from among the largest firms on the TSE, which lists
1,100 h s in its first section. Thus, the relationship between size and beta is
found in the larger firms of the first section of the TSE.
Also evident from Table 6 is that return is strongly related to size. The
difference between the average return on the small and large firms is more than
1 percent a month. Furthermore, the relationship is almost monotonic. These
results mean that the smaller firms provide a higher return as well as lower
beta. If beta is a risk measure, this evidence strongly favors the purchase of
small stocks. Alternatively, perhaps beta is not a sufficient metric for risk. The
relationship between return and size is at least partially captured by the
four-factor model. For example, the sensitivities shown to Factor 4 are ranked
by size. A similar pattern, although less pronounced, is seen in Factors 2 and 3.
Thus, part of what the four-factor model is picking up relative to the one-factor
model is a size effect.
Factor sensitivity stationarity. Another interesting question is the
stability of the sensitivity coefficient, b$ We concentrated on Factor 4 because
it generally has the least stable sensitivity of the four factors. Table 7 shows the
sensitivity coefficients for Factor 4 for the 15-year period and three nonover-
lapping 5-year periods. Although not identical, clearly the sensitivities have the
same pattern across the 20 groups. The correlation between the sensitivities
for the 1971-76 and 1976-81 periods is 0.97; the correlation between the
1976-81 and 1981-86 periods is 0.95.
The average absolute difference in sensitivity between the 1971-76 and
1976-81 periods was 0.0024 when the average absolute value of the sensitivity
in the 1971-76 period was 0.0105. On average, the change was less than 23
percent. Likewise, the average absolute difference in sensitivity between the
1976-81 and 1981-86 periods was 0.0039, with an average absolute value of
the sensitivity in 1976 to 1981 of 0.0139. Thus, the average change was about
28 percent. The sensitivity measures between nonoverlapping periods are very
stable for Factor 4. The stability is even more pronounced for the other factors.
In this section, we have shown that a four-factor model explains returns
better than a one-factor model and exhibits stable sensitivities over time.
Although some increase in explanatory power is guaranteed as we move from
a one-factor to a four-factor model, the magnitude of the increase, particularly
for low-capitalization portfolios was unusually large. A much more powerful test
TABLE 7. Sensitivity to Factor 4
April 1976- April 1981- April 1976- April 1971-

Portfolio March 1986 March 1986 March 1981 March 1976
'Insimcant at 5 percent level.
of whether a four-factor model is superior to the one-index model is to make

comparisons in a forecast mode. We will now examine one such comparison.
Index Matching. In index matching, for each model (the one-index and
four-index models), a portfolio is constructed that has the same sensitivity
(beta) or sensitivities as the index being matched, with minimum residual risk
and (approximately) a fixed number of securities. The second step is to examine
the ability of these portfolios to match the index over a period of time
subsequent to when they are formed.
The index-matching test is a joint test of a number of hypotheses. One
aspect affecting performance is whether the market has one or four factors.
Even with four factors, the one-factor model could still perform better than the
four-factor model if the historically estimated sensitivities for the four-factor
model were poor forecasts of future sensitivities and the historical sensitivities
for the single-index model were a good predictor of future sensitivities. Finally,
if the sensitivity of the market to the four factors is very stable over time, the
performance of the two models might be quite similar, even if the four-factor
model were a superior description of reality. In this case, matching market
betas is equivalent to matching factor sensitivities. Thus, the index-matching
tests are joint tests of the model, the stability of the sensitivities, and the
stability of the relationship of the market to the factors.
We attempted to match the Nikkei 225 Index over the five-year period
January 1, 1981, to December 31, 1986. The first step was estimating
sensitivities for each security. Sensitivities were estimated on a quarterly basis
for both the single-factor model and the four-factor model. For the market
model, we simply ran a regression of the return on each of the 393 stocks in our
study against the TSE Index at the beginning of each quarter using the previous
five years of monthly data. For the four-factor model, a factor analysis was run
at the beginning of each quarter, using the prior 11 years of return data to
determine the composition of the factors. Then, at the beginning of each
quarter, we regressed the prior five years of monthly returns for each of the
393 securities against returns on the four factors. The regression coefficients
were the sensitivities or betas used in the next step. To match the Nikkei, we
need to know the sensitivity of that index to our model. To estimate the
sensitivity of the Nikkei 225, we regressed it against both the TSE Index for the
market model and the four indexes for the four-factor model.
Once the sensitivities were determined, we calculated the composition of
the portfolio with minimum tracking error. Matching portfolios were determined
for portfolios of about 25, 50, and 100 securities. For example, for the market
model with 25 securities, quadratic programming was used to produce a
portfolio with approximately 25 securities that (1) had the same beta as the
Nikkei index and minimal residual risk from the market model, (2) involved full
investment, and (3) had appropriate upper and lower limits on investment in
each security. For the four-factor model, the portfolio is constrained to match
the sensitivity of the Nikkei index to each of the four factors. This procedure is
repeated for 25, 50, and 100 securities and for each quarter, or 20 times for the
five-year period.
Having determined the portfolio composition at the beginning of each
quarter, we then calculated the return on the Nikkei 225 adjusted for dividends
each month, as well as the return on the matching portfolio, using the market
lo Getting the quadratic programming solution to contain approximately 25 securities involved

iterating the solution for different limits on the fraction invested in each security.
four-factor model cut the average squared forecasting error, in some cases, by
more than half.
The index-matching test involves forecasting the model and parameters.
Thus, the difference in number of factors during the fit period does not bias the
results. The ability of the four-factor model to allow the construction of a
portfolio that tracks a second portfolio more closely is powerful evidence that
the four-factor model describes reality better than the single-factor model. l2
A Multi-Index Model of Bond Returns. As a second example of the

estimation of a multi-index model from historical returns, we will discuss an
application to bonds. In this case, we were attempting to find a return-
generating model that would be useful in protecting against shifts in the term
structure of interest rates.
Estimation of a multi-index model using bond returns has a basic problem.
An assumption underlying most estimation techniques is that the structure of
the multi-index model remains stable over time. For bond returns, however,
this is unlikely to be true. The factor determining how bond returns move
together is a shift in the term structure, which affects long bonds more than
short. Further, examining the return series for a single bond would mean that
the return series is for a bond with changing maturity and changing sensitivity
to a shift in the term structure over time. To mitigate this influence, each of the
portfolios of bonds for which a time series of returns was constructed had a
constant maturity over time and hence roughly a constant sensitivity to the
factors.
Using these time series, we constructed the correlation matrix by standard-
izing returns to have a mean of zero and a variance of 1. We used the correlation
matrix because of evidence of Gibbons (1982) that factor analysis of the
correlation matrix for U.S. bond data is more stable than factor analysis of the
variance-covariance matrix. We then performed a maximum likelihood factor
analysis on the correlation matrix. To obtain indexes over a large sample, we
constructed portfolios of bonds that were perfectly correlated with each factor
from the factor analysis and that were uncorrelated with other indexes. We then
estimated the sensitivities by using least squares regression of the constant-
maturity portfolios on the indexes.
The question now is how to decide if this multi-index model is useful. We
l2 AS we discussed earlier, this is a joint test. It is not very plausible, however, that the
four-factor model is more stable than the one-factor model. Its superior performance is likely to
be because of the presence of multiple factors rather than greater stability of sensitivities.
examined its usefulness on a relative basis: Did it perform better than

alternatives others had suggested? We compared this multi-index model with
other models in which either the indexes or the sensitivities (durations) were
prespecified. We examined both a single-index and a two-index model. We
stopped at two indexes because, after removing two indexes, the covariance
between unique return, E(eiej), for our portfolios was so close to zero that
extracting additional indexes was not feasible.
We used three different sets of tests.13 First, we examined which model
serves as a better forecaster of future returns. Within the period that the
factors are extracted, the two-index model would explain more than the one
index. If the second index is spurious, however, or if sensitivity to the second
index is unstable over time, then the single-index model might well work best.
The models were compared by examining the average absolute difference
between forecasted return and actual return.
The second tests were correlation tests. If a multi-index model is the correct
one, then the covariance of unique return across portfolios, E(eiej), should be
close to zero. Once again, this idea was tested for forecasted returns. In this
case, we compared the average absolute value of the covariance between
residuals.
The third test has its origins in immunization. The cash flows from an
arbitrary portfolio of bonds was used to represent a set of liabilities. Because
the value of the portfolio of liabilities at each point is known, the change in value
of the liabilities (the portfolio's return) can be calculated. We tried to immunize
these liabilities by constructing a second portfolio of bonds that moved in unison
with (had the same return as) the portfolio of liabilities. If we have the correct
return-generating process, then a portfolio of assets with the same sensitivities
to the factors as the liabilities should be immunized and differences in return
should be random.
In this study, the two-index model outperformed the single-index models;
for bonds, two factors seemed to be necessary. Furthermore, prespecifying the
sensitivities as durations rather than estimating them from the data produced
the poorest results. In this case, however, prespeclfylng the two indexes as
returns on a widely diversified portfolio of short bonds and a widely diversified
portfolio of long bonds outperformed the model for which the two indexes were
estimated empirically by factor analysis. The factor structure was sufficiently
unstable over time that it was better to select two portfolios widely separated
l3 See Elton, Gruber, and Nabar (1988) for a more detailed description of the methodology and
results described in this section.
in maturity and broadly diversified and use them as the portfolios (indexes) of
interest over time.
Conclusions
In this paper, we have reviewed some methodology for simultaneously
estimating the indexes and sensitivities in a multi-index model. We have
discussed some of the problems with this methodology and described one
example, the Japanese stock market, for which the method worked extremely
well, and one example, U.S. bonds, for which it worked less well.
We also described some tests of how well a factor model works. Two points
are particularly important: Whether a model works well or poorly can only be
judged in the framework of a particular application (and then it should be judged
in a forecast mode), and the performance of the model is only good or bad
relative to alternative models. Good performance is not an absolute concept.
We should close with a brief discussion of the methodology we have
discussed in this chapter versus the methodology of prespecifying sensitivities
or indexes. Estimating both parameters of a multi-index model has a charac-
teristic that is both its fundamental strength and its major weakness. Its
strength is that it requires no a priori specification of the influences that affect
returns. This makes such models an ideal tool for explaining new types of data.
This analysis can provide real insight into the influences that affect returns.
Furthermore, an n-index model derived from a set of data via factor analysis
explains those data better than any other n-index model. The weakness of the
factor solution is that it may or may not perform better in any application
involving forecasts. Moreover, the factor solution lacks the intuition contained
in a model constructed on the basis of economic logic, and it is usually more
difficult to explain to clients.
A researcher might well wish to use combinations of the models discussed
in the various sections of this monograph. If an analyst feels confident that he
or she can identify one or more economic variables that affect security returns
but that other, unknown factors might be important, factor analysis can help
identify these unknown factors. For example, a researcher might believe that all
equity returns are affected by the market but that undefined sector influences
are also instrumental. One solution is to estimate the single-index model using
the market index and then factor analyze the residuals of the single-index model
to derive other factors.
Appendix A
Several statistical techniques can be used to produce a set of indexes that

explain the interrelationship among security returns. The most commonly used
techniques for identifying the influences (factors) affecting security returns and
sensitivities (factor loadings) simultaneously are principal component analysis
and factor analysis.
Principal Component Analysis

Principal component analysis is perhaps easier to understand than factor
analysis. For returns on a defined set of stocks over a predetermined time span,
compute an index (weighted average of that set of returns) that explains the
maximum amount of the variation in the variance-covariance matrix of security
returns. l4 Usually, the first principal component looks somewhat like a market
index with all stocks entering with positive weights. Then, search for the index,
constrained to be orthogonal (uncorrelated with the first index), that explains as
much as possible of the unexplained portion of the variance-covariance matrix.
Additional principal components are then extracted until the user decides that
they are picking up random influences in the data rather than real information.
Of course, a prior estimate of the number of relevant influences will narrow the
choice of how many principal components to extract.
To obtain a multi-index model from a principal component solution some
adjustments are usually performed. First, the indexes obtained from a principal
components solution have decreasing standard deviations as additional influ-
l4 Many researchers choose to perform principal component or factor analysis on the

correlation matrix rather than the variance-covariance matrix. This choice is made because the
solution is sensitive to the scale of the data. Although historical returns on securities have a
natural scaling, an analyst may not want the solution to be affected by the difference in volatility
across securities.
ences are extracted. Normally, the indexes are adjusted so the standard
deviations of all indexes are unity.15
As additional influences are extracted using principal components analysis,
the proportion of random noise in the yet-unexplained variance-covariance
matrix increases. Therefore, each successive principal component is more
likely to be measuring random influences.
Principal component analysis has several advantages over other methods of
extracting factors:
There is a unique solution.

The meaning of successive factors is easy to understand.
The solution technique is simple compared with most other factor
solution techniques, and a large sample can be used.
When residual variance across securities is unrelated to factor loadings
and sample size goes to infinity, the principal component solution is a
transformation of the factor solution while being solvable (see Connor
and Korayczk 1986).
The major weakness of principal component analysis is that it is variance

oriented rather than covariance oriented. The assumption is that residual
variance is not related to factor 1oadings.lG
Factor Analysis
Factor analysis operates directly on the covariance matrix and produces a
result that is intuitively appealing, given the nature of multifactor models. A
number of alternative estimation procedures are available, including maximum
likelihood, generalized least squares, and unweighted least squares. Most
analysts use maximum likelihood methods. For any hypothesized number of
factors, factor analysis finds the indexes and the loadings on each index for each
security to make the covariance between the unique returns as small as
possible.
Although the indexes produced by factor analysis need not be orthogonal to
l5 In addition, researchers often want to construct the indexes from a larger sample of stocks
or to have other properties such as being widely diversified. See, for example, Lehrnann and
Modest (1988).
l6 For example, in a single-index model, beta has been shown to be positively related to
residual variance. Thus, a principal components estimation of the single-index model can lead to
false inferences about beta.
each other, researchers in this area have generally constructed their solutions
so that they are orthogonal. Even with the orthogonal constraint, interpreting
the solution is quite complex. For any given data set, an infinite number of
equivalent factor solutions are possible, which makes interpreting any particular
factor very difficult. The information spanned by the multiple-factor solution
from one sample should be the same as the information spanned by the
multiple-factor solution from the second sample (except for sampling error). As
a practical matter, we know what the set of factors from a factor solution
represents, but we are less sure about the meaning of any individual factor.
A second drawback to factor analysis is that factor solutions are very difficult
to estimate and the sample size that can be factor analyzed is limited by the
length of the return series. Maximum likelihood factor analysis involves a
complex nonlinear optimization problem. It is sufficiently complex that many
analysts have resorted to small sample sizes to estimate factors and factor
loadings.
Multiple-Factor Models for
Portfolio Risk
Richard Grinold
Ronald N. Kahn
BARRA
Multiple-factor models have been applied to forecasting expected returns, as

well as forecasting risk. This paper will focus on the use of multiple-factor
models to predict and control portfolio risk rather than on their use to identlfy
sources of expected return. Portfolio risk depends both on asset risks and on
correlations; factor models accurately capture both of these characteristics.
The goals of risk-forecasting models are to help portfolio managers analyze the
sources of risk in their portfolios; to determine (ex post) if the risks were
justified; and, with a forecast of expected returns, to build new portfolios that
have desirable characteristics.
The emphasis in this paper is on factor models that are relatively easy for
skilled investment practitioners to interpret and use. Thus, we will stress the
use of factors that capture familiar investment themes such as value, growth,
momentum, volatility, and size. We will also stress the use of models that
quickly adapt to the changing nature of the marketplace, whether the change is
the merger of U.S. Steel and Marathon Oil, the birth of the Baby Bells, the
emergence of an electrical utilities industry in the United Kingdom, or a drastic
reassessment of the earnings prospects for IBM.
A Brief History of Multiple-Factor Models

Modem portfolio theory started in the 1950s with Markowitz's (1959)
statement of the portfolio management problem as one of balancing risk against
expected return. He also stressed the role of diversification in reducing portfolio

risk. In the 1960s, Lintner (1965a, 1965b), Mossin (1966), Sharpe (1964), and
Treynor (1961) went one step further and proposed a single-factor model for
the explanation of expected return. In the 1970s, Merton (1972), Rosenberg
and Marathe (1975), Ross (1976), Sharpe (1977), and others proposed
multiple-factor models for the explanation of expected returns. In the 1980s,
academic debate still centered around the issue of expected return-in
particular, the correct way to explain expected return on financial assets (asset
pricing). The question was whether current statistical procedures are up to the
task of recognizing the correct answer if we have it in hand.
Rosenberg (1974) noted early on that the same procedures used in the
search for expected returns could also be used to explain portfolio risk.
Moreover, these efforts had the benefits of being independent of any particular
theory of expected returns and having practical value. They helped portfolio
managers deal in a quantitative and effective way with the issue of risk control.
With the help of a model to handle risk, the managers could then concentrate
their efforts on the difficult task of finding assets that are potential over- or
underachievers.
Multiple-factor risk models of equity returns are mentioned in a host of
academic papers. They are usually not the focus of the paper, merely a
technique or an abstract structure made to demonstrate the point. The
literature on arbitrage pricing theory has made the multiple-factor model a
standard part of the academic landscape. As is well known, the arbitrage pricing
theory posits the existence of an unknown number of unidentified factors that
can be used to explain expected returns. This vague specification has caused
some scoffing, but it has also opened the door to a tremendous amount of
creativity in building linear models. The rule seems to be "anything that is not
forbidden is allowed." Given the wide availability of data, the low cost of
computation, the human imagination, and the need to publish, attempts to build
such models have taken widely different approaches. Witness, for example, the
efforts of Roll and Ross (1980), Chen, Roll, and Ross (1986), Conner and
Korajczyk (1988), and Lehrnann and Modest (1988).
The Structure of Multiple-Factor Risk Models

The multiple-factor risk models described in this paper have a simple linear
structure composed of four components: a stock's exposures to the factors, its
excess returns, the attributed factor returns, and the specific returns. For-
mally,
where
X,,,(t) = the exposure of asset n to factor k (known at time t),
rn(t) = the excess return (return above the risk-free return) on stock n
during the period from time t to time t + 1,
fk(t) = the factor return to factor k during the period from time t to time
t + 1, and
u,(t) = stock n's specific return during the period from time t to time t +
1.' This is the return that cannot be explained by the factors.
The exposures, X,,,, are frequently called factor loadings. For industry
factors, the exposures are either 1 or 0, indicating whether or not the stock
belongs to a given industry. For the other common factors, the exposures are
standardized so that the average exposure over all stocks is 0 and the standard
deviation across stocks is 1.
The specific return, u,, is sometimes called the idiosyncratic return to the
stock. It is the return the model does not explain. The risk model will account
for specific risk, however, so the risk predictions will explicitly consider the risk
of U,.
Equation (1) is not meant to convey any sense of causality. The factors may
or may not be the basic driving forces for security returns. In our view, they are
merely dimensions along which to analyze risk.
Now, assume that the specific returns are not correlated with the factor
returns and the specific returns are not correlated with each other. With these
assumptions and the return structure of equation (I), the risk structure is:
where
Vn,, the covariance of asset n with asset m (if n
= = m, this gives the
variance of asset n),
Xn,,, = the exposure of asset n to factor k1,
Although the model's time structure is defined in equation (I), in the rest of this paper, the
explicit time variables will be suppressed.
F,,,, = the covariance of factor k l with factor k2 (if k l = k2, this gives the
variance of factor kl), and
A , , = the specific covariance of asset n with asset m. By assumption, all
specific risk correlations are zero, so this term is zero unless n =
m. In that case, this term gives the specific variance of asset n.
Building the Model

The process of building a multiple-factor risk model consists of three phases:
choosing the factors, estimating factor returns, and forecasting risk. We shall
examine these in detail.
Choosing the Factors. The art of building a multiple-factor risk model

is to choose appropriate factors. This search for factors is limited by one key
constraint: All factors must be a priori factors. That is, even though the factor
returns are uncertain, the factor exposures must be known at the beginning of
the period.
Within the constraint of a priori factors, a wide variety of factors are
possible. To classlfy the factors, they are first divided into three categories:
responses to extemal influences, cross-sectional comparisons of asset at-
tributes, and purely internal or statistical factors.
Responses to external influences. One of the prevalent themes in the
academic literature of financial economics is that a demonstrable link should
exist between outside economic forces and the equity markets. The response
factors are an attempt to capture that link. These factors include responses to
return in the bond market (sometimes called bond beta), unexpected changes in
inflation ( i a t i o n surprise), changes in oil prices, changes in exchange rates,
and changes in industrial production. These factors are sometimes called
macrofactors. BARRA models use this type of factor, in particular in response
to interest rate changes and, in some countries, response to changes in
exchange rates.
These measures suffer from two serious defects. The first is that the
response coefficient has to be estimated through regression analysis or some
similar technique. This requirement leads to errors in the estimates, commonly
called the error-in-variables problem. The second drawback is that the esti-
mates are based on behavior during a past period, generally five years. Even if
these past estimates are accurate in the statistical sense of capturing the true
situation in the past, they may not accurately describe the present. In short,
these response coefficients can be nonstationary.
Cross-sectional comparisons. These factors, which have no link to the
remainder of the economy, compare attributes of the stocks. Cross-sectional
attributes can themselves be classified in two groups: fundamental and market.
Fundamental attributes include such ratios as dividend yield and earnings yield,
plus analysts' forecasts of future earnings per share. Market attributes include
volatility during a past period, return during a past period, option-implied
volatility, share turnover, and so forth.
To some extent, market attributes such as volatility and momentum may
have the same difficulties (errors in variables, nonstationarity) as the external
response factors. Here, however, the factor interpretation is somewhat
different. For example, a momentum factor, taken to be a measure of the price
performance of the stock for the past 12 months, is not intended as a forecast
of continued success or of mean reversion. It is merely a recognition that stocks
that have been relatively successful (unsuccessful) during the past year will
quite frequently behave in a common fashion. Sometimes the momentum will be
reinforced, in other times it will be reversed, and in yet other times it will be
irrelevant. We are accounting for the fact that in five or six months of the year,
controlling for other attributes, previously successful stocks behave in a much
different manner than previously unsuccessful stocks. The same is true for
stocks with high historical volatility and so forth.
Statisticalfactors. It is possible to amass returns data on a large number
of stocks, turn the crank of a statistical meat grinder, and admire the factors the
machine produces: factor ex machina. This procedure can be accomplished in an
amazingly large number of ways, including principal component analysis,
maximum likelihood analysis, and expectations maximization analysis. One can
use a two-step approach-first get the factors and then the exposures-or
simultaneously estimate both factors and exposures, or turn the problem on its
head in the imaginative approach taken by Connor and Korajczyk (1988).
BARRA models do not usually include statistical factors because they are very
difficult to interpret and because the statistical estimation procedure is prone to
discovering spurious correlations.
Among the many possible factors, those chosen should satisfy three criteria:
They should be incisive, intuitive, and interesting. Incisive factors differentiate
returns. For example, low-volatility stocks perform differently from high-
volatility stocks at least three times a year. If the overall volatility exposure is
not monitored, then returns can be upset with disturbing frequency.
Intuitive factors relate to interpretable and recognizable dimensions of the
market. Credible stories define these factors. For example, size has the big
companies at one end and the small companies at the other. Momentum has the
firms that have performed well separated from the firms that have done
relatively poorly. Intuitive factors arise from recognizable investment themes.
A Practztioneh Guide to Factor Models
Factors in the U.S. equity market include industries, size, yield, value, success,
volatility, growth, leverage, liquidity, foreign income, and labor sensitivity.
Interesting factors explain some part of performance. We can attribute a
certain amount of return to each factor in each period. That factor might help
explain exceptional return or beta or volatility. For example, large stocks did
well over a particular period, or high-volatility stocks are high-beta stocks.
Research leading to the appropriate factors, then, depends both on statistical
techniques and on investment intuition. statistical techniques can idenbfy the most
incisive and interesting factors. Investment intuition can help idenbfy intuitive
factors. Factors can have statistical signdicance or investment significance or both.
Model research must take both forms of significance into account.
Exposures. The factors typically chosen for use in a multiple-factor risk

model fall into two broad categories: industries and risk indexes. Industry
factors measure the differing behavior of stocks in different industries. Risk
indexes measure the differing behavior of stocks across other, nonindustry
dimensions.
Industly exposures. Industry groupings partition stocks into nonoverlap-
ping classes. Industry groupings should satisfy several criteria. They should
represent a reasonable number of companies in each industry, have a reason-
able fraction of capitalization in each industry, and be in reasonable accord with
the conventions and mind-set of investors in that market.
Industry exposures are usually 110 variables, because stocks are either in an
industry or they are not. The market itself has unit exposure in total to the
industries. Because large corporations can do business in several industries, the
industry factors must account for multiple industry memberships. For example,
for March 1992, BARRA's USE2 model classifies General Electric as 39
percent producer goods, 28 percent aerospace, 23 percent consumer products,
5 percent miscellaneous finance, and 5 percent media.
Risk index exposures. Industries are not the only sources of stock risk.
Risk indexes measure the movements of stocks exposed to common invest-
ment themes. Risk indexes we have identified in the United States and other
equity markets fall into these broad categories: .
Volatility distinguishes stocks by their volatility. Assets that rank high in
this dimension have been and are expected to be more volatile than
average.
Momentum distinguishes stocks by recent performance.
Size distinguishes large stocks from small stocks.
Liquidity distinguishes stocks by how often their shares trade.
Growth distinguishes stocks by past and anticipated earnings growth.
Value distinguishes stocks by their fundamentals, including ratios of
earnings, dividends, cash flows, book value, and sales to price; is the
stock cheap or expensive relative to fundamentals?
Earnings volatility distinguishes stocks by their earnings stability.
Financial leverage distinguishes firms by their debt-to-equity ratios and
exposure to interest rate risk.
Any particular equity market can contain fewer or more risk indexes, depending
on its own idiosyncrasies.
Each of the broad categories listed above can contain several specific
measurements, or descriptors, of the category. For instance, volatility measures
might include recent daily return volatility, option-implied volatility, recent price
range, and beta. Although the descriptors are typically correlated, each
captures one aspect of the risk index. We construct risk index exposures by
weighting exposures of the descriptors within the risk index. The weights are
chosen to maximize the model's explanatory power. Relying on several different
descriptors can improve model robustness.
The various categories of descriptors and risk indexes involve different sets
of natural units and ranges. To quantlfy them, all raw exposure data must be
rescaled:
where mean (X,,) is the raw exposure value mean and SDCX,,,) is the raw
exposure value standard deviation across the universe of assets. The result is
that each risk index exposure has a mean of zero and a standard deviation of 1.
This standardization also facilitates the handling of outliers.
As an example of how this procedure works, BARRA's USE2 model assigns
General Motors a size exposure of 1.30 for March 1992. Thus, on the size
dimension, General Motors lies sigdicantly above average. For the same date,
the model assigns Apple Computers a size exposure of -0.26. On this
dimension, Apple Computers lies somewhat below average.
Factor Returns. Given exposures to the industry and risk index factors,
the next step is to estimate returns via multiple regressions. This procedure
was developed in Fama-MacBeth (1973). The model is linear, and equation (1)
has the form of a multiple regression. Stock excess returns are regressed
against factor exposures, choosing factor returns that minimize the (possibly
weighted) sum of squared specific returns. For the United States, we use a
universe of 1,100 of the largest companies. The R2 statistic, which measures
the explanatory power of the model, tends to average between 30 percent and
40 percent for models of monthly equity returns with roughly 1,000 assets and
50 factors. Larger ' R statistics tend to occur in months with larger market
moves.
In this cross-sectional regression, which is performed every period, gener-
ally one month, the industry factors play the role of intercepts. The market as
a whole has an exposure of 1to the industries, and industry factor returns tend
to pick up the market return. They are the more volatile factors in the model.
The market has close to zero exposure to the risk indexes, and risk index factor
returns pick up extra-market returns. They are the less volatile factors in the
market.
To estimate factor returns efficiently, we run generalized least squares
regressions, weighting each observed return by the inverse of its specific
variance. In some models, we instead weight each observation by the square
root of its market capitalization, which acts as a proxy for the inverse of its
specific variance.
Although these cross-sectional regressions can involve many variables (the
USE2 model uses 68 factors), the models do not suffer from multicollinearity.
Most of the factors are industries (55 out of 68 in USEZ), which are orthogonal.
In addition, tests of variance inflation factors, which measure the inflation in
estimate errors attributable to multicollinearity, lie far below serious danger
levels.
Factor Portfolios. The regression approach to estimating factor returns

leads to an insightful interpretation of the factors. Weighted regression
gymnastics lead to the following matrix expression for the estimated factor
returns:
where X is the exposure matrix, W is the diagonal matrix of regression weights,

and r is the vector of excess returns. For each particular factor return, this
calculation is simply a weighted sum of excess returns:
Our research has shown that the square root is the appropriate power of market capitalization
to mimic inverse specific variance. Larger companies have lower specific variance, and as
company size doubles, market variance shrinks by a factor of 0.7.
In this form, each factor return, f,, can be interpreted as the return to a
portfolio with portfolio weights c, ,. So factor returns are the returns to factor
portfolios. The factor portfolio hoidings, which are known a priori, ensure that
the portfolio has unit exposure to the particular factor, zero exposure to every
other factor, and minimum risk, given those constraints.
These portfolios have two different interpretations. They are sometimes
interpreted as factor-mimicking pot@olios, because they mimic the behavior of
some underlying basic factor. We interpret them more simply as portfolios that
capture the specific effect we have defined through our exposures.
Factor portfolios typically contain both long and short positions. For
example, the factor portfolio for the earnings-to-price factor in the U.S. market
will have an earnings-to-price ratio that is one standard deviation above the
market while having zero exposure to all other factors. A zero exposure to an
industry implies that the portfolio will hold some industry stocks long and others
short, with longs and shorts balancing. This portfolio will contain every single
asset with some weight.
Factor Covariance and Specific-Risk Matrixes. Once the factor

returns each period are estimated, we can estimate a factor covariance
matrix-an estimate of all the factor variances and covariances. To operate
effectively as a risk model, this factor covariance matrix should constitute our
best forecast of future factor variances and covariances over the investor's time
horizon. 3
Generating an asset-by-asset covariance matrix requires both the factor
covariance matrix, F, and the specific risk matrix, A. By definition, a stock's
specific return, un, is that component of its return that the model cannot
explain. So the multiple-factor model can provide no insight into stock-specific
returns. For specific risk, we need to model specific return variance, u:
assuming that mean specific return is zero.
In general, the model for specific risk is
with
Forecasting covariance from a past history of factor returns is a subject worthy of a paper in
itself, and the details are beyond the scope of this effort. Basic techniques rely on weights over
the past history and Bayesian priors on covariance. More advanced techniques include forecasting
variance conditional on recent events, as first suggested by Engle (1982). Such techniques
assume that variance is only constant conditional on other variables. For a review of these ideas,
see Bollerslev et al. (1992).
and
S(t) measures the average specific variance across the universe of stocks, and
vn captures the cross-sectional variation in specific variance.
To forecast specific risk, we use a time series model for S(t) and a linear
multiple-factor model for vn(t). Models for vn(t) typically include some risk index
factors, plus factors measuring recent squared specific returns. The time
dependence in the model of v,(t) is captured by time variation in the exposures.
One pooled regression over assets and time periods, with outliers trimmed, is
used to estimate model coefficients.
Data Requirements. Multiple-factor risk models require data on stock

returns and sufficient data to calculate factor exposures. Calculating stock
returns requires not only stock price data but also data on stock dividends,
splits, and other adjustments. Factor exposures require industry identification,
including earnings, sales, and assets segmented by industry; historical returns;
associated option information; fundamental accounting data; and earnings
forecasts. 4
Model Validity
Considerable evidence supports the validity of multiple-factor risk models.
This evidence falls into three categories: in-sample tests, out-of-sample tests,
and empirical observations.
In-sample tests focus on the performance of the multiple-factor model
(equation 1)in explaining excess stock returns. Typically, these models will use
BARRA's USE2 model started in January 1973 and, for initial estimation, required data
covering the period from January 1968 through December 1972. This model relies mainly on
MARKET PLUS for market data, COMPUSTAT for fundamental accounting data, and IBES for
earnings forecasts, but it also requires data from many other sources, including Standard &
Poor's, the New York Stock Exchange, the American Stock Exchange, Value L i e , and
Interactive Data Corporation.
roughly 50 factors to explain the returns to roughly 1,000 assets each month.
Monthly R2 statistics for the models average about 30-40 percent, meaning
that the model "explains," on average, about 30-40 percent of the observed
cross-sectional variance of the universe of stock returns.
These R2statistics, averaged over many months, do not accurately convey
model performance, however. In fact, the R2 statistic can vary quite signifi-
cantly from month to month, depending in part on the overall market return.
Model R2statistics are highest when the market return differs very signdicantly
from zero. The R2statistic was very high in October 1987 because the market
return was so extreme. In months when the market return is near zero, the R2
statistic can be quite low, even if discrepancies between realized and modeled
returns are small.
Another measure of model performance is the root mean square error from
the regression. This averages 6 percent for monthly cross-sectional returns in
the United States and does not vary much from month to month. Because
monthly stock volatility in the United States averages 10 percent, the model
explains about 64 percent of individual stock variance, on average.
Because the goal of the model is to explain portfolio risk, a better way to
evaluate the model is by the fraction of portfolio risk it explains, and here is
evidence of the model's true power. For benchmark portfolios in the United
States, the multiple-factor risk model explains more than 98 percent of portfolio
variance.
Out-of-sample tests compare forecast risk with realized risk. One out-of-
sample test builds portfolios of randomly chosen assets and then compares the
forecast and realized active risk of those portfolios; active risk is defined as the
volatility of the active return, or the difference between the portfolio return and
a benchmark return. In tests in the United States involving 500 such portfolios
containing 100 assets each, we compared realized active risk for a 12-month
period with forecast active risk at the beginning of the 1988-91 period, using
the S&P 500 as the benchmark. At the 1 percent confidence level, we could
reject the hypothesis that forecast variance equaled realized variance only 2.8
percent of the time.
We have also examined risk forecasts cross-sectionally. With the same 500
portfolios, we examined standardized active returns: ratios of realized active
returns to forecast active risk. Pooled over four months, the standard deviation
of standardized active returns was 1.06, which according to X 2 tests, was
statistically consistent with the unbiased result of 1.0.
Finally using a variance-forecasting test suggested by Engle, Hong, and
Kane (1990), we have run options-based tests comparing multiple-factor risk
model forecasts with historical asset-by-asset risk. In these tests, we construct
30 random portfolios of 100, 150, 200, and 250 assets and generate the two
forecasts of active risk. We then use the Black-Scholes model to price
one-month at-the-money options on portfolio active value based on these two
forecasts. We create a synthetic market in these options, trading at the mean
price. At the end of the period, we calculate profit and loss. Over a 36-month
period from January 1988 through December 1990, the strategy using the
multiple-factor risk model forecasts, on average, returned 37 basis points per
option traded with the historical volatility trader, with a standard deviation of
134 basis points. In this zero-sum game, the strategy based on historical
volatility lost 37 basis points per option traded.
In both in-sample and out-of-sample tests of model validity, we occasionally
invoke standard distributional assumptions to interpret the statistical sigmfi-
cance of the results. Also, in both model building and testing, we make use of
Monte Carlo simulations to test statistical sigmficance while relaxing the
required assumptions. As to investment sigmficance, if some event occurs 3
times out of 12, an investor would want to know about it, even if a statistician
would not be sure of its importance at the 95 percent confidence level.
Empirical observations concerning model validity are more vague than
statistical tests, but they are still relevant. Simply put, these models success-
fully make use of intuitive factors to predict risk and understand return, and
they have been widely accepted by the investment community for those roles
for 18 years now.
How do multiple-factor risk models compare with their existing alternatives?
Historical asset-by-asset covariance matrixes consistently underperform multi-
ple-factor models in risk forecasting, and they suffer from severe estimation
problems. A covariance matrix for 1,000 assets contains 500,500 independent
entries, all estimated with errors. In addition, unless estimated over more than
1,000 time periods, the covariance matrix will not be full rank.
Simpler versions of the multiple-factor approach include a one-factor model
and a constant-correlation model. The one-factor model is a close relative of the
capital asset pricing model. This model includes only one common factor-the
market. The constant-correlation model assumes that all assets exhibit the
same correlation. Both models are simple and helpful for "quick and dirty"
applications but ignore linkages among stocks in specific industries and with
similar attributes.
Another approach to risk modeling uses statistical factor analysis. This
approach identifies factors based on past correlations between asset returns.
These factors are typically not intuitive or recognizable. This statistically driven
approach can lead to risk forecasts comparable in quality to multiple-factor risk
model forecasts but without any of the insight. Also, because they do not rely
on investor intuition, they can be less robust than multiple-factor models.
Overall, multiple-factor risk models outperform all alternative risk models in
providing incisive, intuitive, and interesting risk analysis.
Applications of Multiple-Factor Risk Models

The technical core of a multiple-factor risk model is the attribution of asset
returns to chosen common factor and specific returns, plus forecasts of the
variances and covariances of these common factor and specific returns. This
technical core supports three separate types of investment analysis, which
focus on the present, the future, and the past.
The Present: Current Portfolio Risk Analysis. The multiple-factor

risk model decomposes current, overall portfolio risk in several ways. This
decomposition of risk identifies the important sources of risk in the portfolio and
links those sources with aspirations for active return.
One way to divide the risk is to identlfy the market and the residual
components. An alternative is to look at risk relative to a benchmark and
identlfy the active risk. A third way to divide the risk is between the model risk
and the specific risk. The risk model can also perform marginal analysis,
identlfylng which assets, at the margin, are most and least diverslfylng in the
portfolio.
Risk analysis is important for both passive management and active manage-
ment. Passive managers attempt to match their portfolio returns to a particular
benchmark. They run index funds, but depending on the benchmark, the
passive managers' portfolios may not include all the stocks in the benchmark.
For example, a passive small-stock manager might face prohibitive transaction
costs for holding the thousands of assets in a broad small-stock benchmark.
Current portfolio risk analysis can tell a passive manager the active risk, or
tracking error, of a portfolio relative to its benchmark. The tracking error is the
volatility of the difference in return between the portfolio and the benchmark.
Passive managers want minimum tracking error.
The goal of active managers is not to track the benchmark as closely as
possible but to outperform the benchmark. Still, risk analysis is important in
active management, to focus active strategies. Active managers want to take on
risk only along those dimensions they believe they can outperform.
By suitably decomposing current portfolio risk, active managers can better
understand the positioning of their portfolios. Risk analysis can tell them not
only what their active risk is but also why and how to change it. Risk analysis
can classlfy active bets into inherent bets, intentional bets, and incidental bets:
Inherent bets. An active manager who is trying to outperform a bench-

mark (or the market) will have to bear the benchmark risk. This risk is a
constant part of the task, not under the portfolio manager's control.
Intentional bets. An active portfolio manager has identified stocks that he
believes will do well and stocks that he believes will do poorly. The manager
should expect that these stocks will appear as important marginal sources of
active risk. This is welcome news: It tells the portfolio manager that he has
taken active positions that are consistent with his beliefs.
Incidental bets. These are unintentional side effects of a manager's active
position. The manager has inadvertently created an active position on some
factor that is a significant contributor to marginal active risk. For example, a
manager who builds a portfolio by screening on yield will find a large incidental
bet on industries that have higher than average yields. Are these industry bets
intentional or incidental? Incidental bets often arise through incremental port-
folio management, where a sequence of stock-by-stock decisions, each plausible
in isolation, leads to accumulated incidental risk.
TABLE 1. Sample Portfolio
Stock Shares Weight
American Express
AT&T
Chevron
Coca Cola
Walt Disney Productions
Dow Chemicals
DuPont
Eastman Kodak
Exxon
General Electric
General Motors
IBM
International Paper
Johnson &Johnson
McDonalds
Merck
Minnesota Mining and Manufacturing
Philip Morris
Procter & Gamble
Sears
To understand portfolio risk characterization more concretely, consider the
following problem: Using as an investment portfolio the Major Market Index
(MMI), a price-weighted index of 20 of the largest U. S. stocks, analyze its risk
relative to the S&P 500 as of February 28, 1992. The portfolio's composition is
given in Table 1.
Comparing risk factor exposures against the benchmark, this portfolio
contains larger, less volatile stocks with higher leverage and foreign income and
lower earnings variability-what one might expect from a large-stock portfolio
versus a broader index. The portfolio also contains several industry bets.
The multiple-factor risk model forecasts 21.3 percent volatility for the
portfolio and 20.8 percent volatility for the index. The portfolio tracking error is
4.7 percent. Assuming that active returns are normally distributed, the portfolio
annual return will lie within 4.7 percent of the index annual return roughly
two-thirds of the time. The model also can forecast the portfolio's beta-its
exposure to movements of the index. Beta measures the portfolio's inherent
risk. The MMI portfolio beta is 1.02. This implies that if the S&P 500 exceeded
its expected return by 100 basis points, we would expect the portfolio return to
exceed its expected return by 102 basis points.
The mar@ contribution to tracking error-the increase in tracking error
from a 1percent increase in asset holding financed by a 1 percent decrease in
cash-can be used to determine the most and least diversify~ngassets. A more
detailed treatment of marginal contribution to tracking error is found in the
mathematical appendix to this paper.
In this example, increasing the holdings in American Express would do most
to reduce risk, and increasing holdings in Merck would do the most to
concentrate the portfolio. These are also the lowest and highest weighted
assets in the portfolio.
Looking Forward: Portfolio Construction. Given forecasts of ex-

pected returns, a multiple-factor risk model can help construct investment
portfolios that optimally implement bets on those returns. The idea is to
maximize utility, defined as risk-adjusted expected return:
Here, h, is the holding of asset n, r, is the expected return to asset n, and A

is a risk aversion parameter. The covariance, V,,,, comes from the multiple-
factor risk model. In typical examples, the holdings are active holdings relative
to a benchmark and the expected returns are exceptional returns (alphas)
relative to the benchmark. The quadratic optimization problem is solved to

determine the optimal portfolio weights. Of course, in real life, the problem
must account for transaction costs and add constraints and penalties.
Beyond providing risk forecasts, multiple-factor risk models can occasionally
help with return forecasts. Although this is not the main focus of this type of
model, the research needed to choose factors can identlfy those that generate
exceptional return; value factors, for example, often generate exceptional
return. Portfolios that implement such bets on factors are called tilt funds.
In portfolio construction, we assign forecasts of exceptional return to the
stocks in the MMI portfolio and then optimally weight them to maximize
risk-adjusted exceptional return relative to the S&P 500. In the portfolio shown
in Table 2, we arbitrarily assigned 2 percent exceptional return forecasts to
those stocks with ticker symbols that fell in the top half of the list alphabetically
and -2 percent forecasts to the rest. The optimization procedure constructs a
reweighted portfolio.
The alphas shown in Table 2 imply an initial portfolio alpha of -0.19 percent.
TABLE 2. Portfolio Construction Example
Stock Alpha Shares Weight
American Express
AT&T
Chevron
Coca Cola
Walt Disney Productions
Dow Chemicals
DuPont
Eastrnan Kodak
Exxon
General Electric
General Motors
IBM
International Paper
Johnson & Johnson
McDonalds
Merck
Minnesota Mining and Manufacturing
Philip Morris
Procter & Gamble
Sears
With this reweighting, the tracking error moves slightly, from 4.68 percent to
4.71 percent, and the alpha of the portfolio moves to 1.21 percent. For reasons
of risk control, the optimizer cannot eliminate the holdings of all the negative
alpha stocks, but it does reduce those holdings and eliminates three of them
from the portfolio.
The Past: Performance Analysis. Historical analysis of investment

performance is important for understanding realized investment performance
and for backtesting new investment strategies. Over any one period, the model
can attribute returns to the factors and to specific asset returns. Then, linking
many periods of attributed returns, it can analyze the series of returns to these
various bets. This helps measure investment skill and value added.
Within the factor structure of the model, past returns can be attributed to
bets on factors and bets on specific asset returns. For each historical period, we
know the exposures of the portfolio relative to its benchmark, as well as the
subsequent factor and specific returns. By examining many such periods, we
can aggregate returns attributed to each factor and returns attributed to specific
asset bets. With this time series of attributed returns, we can observe mean
achieved returns. This time series, as well as the model itself, leads to
estimates of the risk associated with those returns.
For example, we can compare the performance of the S&P 500 portfolio
with the BARRA ALL-US Index, a broad index of more than 5,000 stocks, for
the 65 months from September 30, 1986, through February 28, 1992. The S&P
500 outperformed the BARRA Index by 92 basis points a year, with an
annualized risk of 2.78 percent. Decomposition of this risk and return by source
shows that most of this active return arose from bets on the common factors:
Annual Return Annual risk
Active common factors

Specific asset selection
Market timing -
Given the large numbers of stocks in the portfolio, very little of the active return
arose from specific asset selection. The market-timing component measures
return contributions attributable to variation in portfolio beta over the time
period. The beta of the S&P 500 versus the BARRA ALL-US was 0.97 in
March 1992, but it does vary over time. The market-timing contribution of - 13
basis points arises because that beta tended to be above its mean value in
A Practitionefi Guide to Factor Models
months when the BARRA Index excess return was below its mean value, and
vice versa.
Among all the bets (policies) included in the S&P portfolio but in not the
BARRA ALL-US benchmark, the best performing was a positive bet on foreign
income, which gained 28 basis points a year; the worst performing was the size
bet, which lost 29 basis points a year during this period.
Two particular statistics can help assess the skill and value added of the S&P
portfolio. Letting Ram,,, represent annualized returns and M represent the
number of observation periods, the t-statistic for the mean return is:
This statistic measures whether the observed mean annualized return differs
sigmficantly from zero. It is one statistical measure of investment skill. If the
t-statistic exceeds 2.0 and returns are normally distributed, then the probability
that simple luck generated these returns is less than 5 percent.
Related to this distinction between skill and luck is the question of whether
the manager has added investment value. The utility defined in equation (9) can
be used to measure value added, or risk-adjusted active return. Detailed
analysis shows that value added rises in proportion to the square of the
manager's information ratio, IR, or the ratio of annual active return, a,, to
annual active risk, o:,
2
vAma = A(%) ,
4A '"annual
with
aannual
IR=-.
'"annual
Value added rises with the manager's information ratio, regardless of the level
of risk aversion.
If the M periods of observation of these returns correspond to T years, then
the information ratio is just the t-statistic divided by the square root of the
number of years of observation:
t-stat
I R = 6
Overall, the t-statistic measures the statistical sigtllficance of the return, but the
information ratio also captures the risk-reward trade-off of the strategy and the
manager's value added.5 An information ratio of 0.5 observed over five years
may be statistically more significant than an information ratio of 0.5 observed
over one year, but the value added will be equal. The distinction between the
t-statistic and the information ratio arises because the definition of value added
is based on risk over a particular horizon, in this case one year.
Using the results of single-period performance attribution over M periods,
this analysis of skill and value added can be applied factor by factor. This process
will identlfy not only whether the manager has overall skill and has added value
but also where the manager has skill and has added value. The result is a precise
analysis of the manager's style. For the example above, the information ratios
and t-statistics for each component of active return are as follows:
Information
Ratio t-statistic
Active common factors

Specific asset selection
Market timing
Other Uses of Multiple-Factor Models. Portfolio managers are not

the only users of multiple-factor risk models. Researchers, plan sponsors, and
traders also find them helpful. Investment researchers use multiple-factor risk
models to run controlled backtests of future investment strategies. For this,
their needs are similar to those of portfolio managers. They need to implement
strategies optimally on historical data and understand the subsequent perfor-
mance of those strategies. Researchers can use backtests to enhance their
strategies. They can also use performance analysis and portfolio risk charac-
terization to improve their understanding of the bets they are testing.
Pension plan sponsors can use multiple-factor risk models to coordinate their
multiple managers. Portfolio risk characterization allows them to understand
any gaps or overlaps among their managers or in their asset allocation mixes.
Plan sponsors also use performance analysis to assess their managers' value
added and to check on their managers' styles.
Traders can use multiple-factor models in at least two ways. The models can
aid in risk control during the course of trading. Multiple-factor models can also
For a more detailed discussion of the information ratio and its relationship to skill and value
added, see Grinold (1990).
aid in index arbitrage strategies through their use in constructing small baskets
of stocks to track index futures optimally.
Conclusions
Multiple-factor risk models perform well in predicting investment risk and
providing investment intuition. Across many asset classes and markets, these
models identlfy incisive, intuitive, and important common factors affecting risk
and return. They use intuitive, easy-to-understand factors to analyze invest-
ment risk and returns. They accurately forecast investment risk and help
explain past returns, but they do not forecast returns.
Multiple-factor risk models can be used to analyze current portfolio risk,
construct portfolios that optimally trade off risk with expected returns, and
analyze skill and value added associated with past returns. They are an
important tool for managing portfolios, conducting investment research, coor-
dinating multiple managers, and trading.
Portfolio managers use multiple-factor risk models to (1) analyze their
current risk and understand the size and location of their bets, (2) construct
portfolios that optimally trade off risk against expected returns, and (3) analyze
and provide insight into their past returns in order to understand their skill and
value added. Researchers use multiple-factor risk models in similar ways to
backtest and fine-tune strategies. Pension plan sponsors use multiple-factor risk
models to coordinate their multiple managers and to understand gaps and
overlaps in their asset allocation mixes. Traders use these models to control
investment risk over short horizons.
Multiple-factor risk models are central to structured investing and are also
extremely useful for traditional investment processes. Whether investors
structure their portfolios within a strict risk-return framework or whether they
simply pick stocks according to tradition, multiple-factor risk models help
control and understand risk and also help understand past performance.
Appendix A
The risk model in matrix notation is written as
where r is an N vector of a stock's excess returns, X is an N by K matrix of

stock factor exposures, f is a K vector of factor returns, and u is an N vector
of specific returns.
We assume:
the specific returns, u, are uncorrelated with the factor returns, f; that
is cov{u,, f,} = 0 for all n and k.
the covariance of stock n's specific return, u,, with stock m's specific
return, urn, is 0, if m # n; that is, cov{u,, urn}= 0 if m # n.
With these assumptions, we can express the N by N covariance matrix, V, of

stock returns as:
where F is the K by K covariance matrix of the factor returns and A is the N

by N diagonal matrix of specific variance.
A portfolio, P , is described by an N-element vector, hpl that gives the
portfolio's holdings in the N risky assets. The factor exposures of P are given
by:
The variance of P is given by:
A similar formula lets us calculate active variance. If h, is the benchmark

holdings vector, then we can define:
and
Notice that we have separated both total and active risk into common factor and
specific components. This method works because factor risks and specific risks
are uncorrelated.
We can also examine the marginal effects of any change in the portfolio. This
type of sensitivity analysis allows us to see what factors and assets have the
largest impact on risk. The marginal impact on risk is measured by the partial
derivative of the risk with respect to the asset holding.
We can compute these marginal contributions for total risk and active risk.
The N vector of marginal contributions to total risk is:
VhP
MCTR = -.
UP
The MCTR(n) is the partial derivative of upwith respect to hp(n). We can think
of it as the change in portfolio risk given a 1percent increase in the holding of
asset n, which was financed by decreasing the cash account by 1percent. The
cash holding, hp(0), is given by:
In a similar way, we can define the marginal contribution to active risk as:
VhA
MCAR = -.
VA
Bibliography
Akaike, H. 1974(a). "Markovian Representation of Stochastic Processes and its Application to the
Analysis of Autoregressive Moving Average Processes." Annuls of the Institute of Statistical
Mathematics 26: 363-87.
. 1974(b). "A New Look at the Statistical Identification Model." IEEE Transactions on
Automatic Control, 716-23.
Berry, Michael A. 1988. "A Practical Perspective on Evaluating Mutual Fund Risk." Investment
Management Review (MarcWApril):78-86.
Berry, Michael A., Edwin Burmeister, and Marjorie B. McElroy. 1988. "Sorting Out Risks Using
Known APT Factors." Financial Analysts Journal 44 (MarcWApril):29-42.
Bollerslev, T i , Ray Y. Chou, Narayanan Jayaraman, and Kenneth F. Kroner. 1992. "ARCH
Modeling in Finance: A Selective Review of the Theory and Empirical Evidence, with Suggestions
for Future Research." Journal of Econometrics 52 (April/May):5-59.
Brock, W i a m A. 1982. "Asset Prices in a Production Economy." In The Economics of

Information and Uncertainty, ed. J. McCall, 1-43. Chicago: The University of Chicago Press.
Brown, Stephen J. 1990. "Macroeconomic Factors and the Japanese Stock Market." In Jujanese
Capital Markets, eds. Edwin J . Elton and Martin J. Gruber, 175-92. New York: Harper and Row.
Brown, Stephen J., and Mark I. Weinstein. 1983. "A New Approach to Testing Asset Pricing
Models: The Bilinear Paradigm." Journal of Finance 38 Uune): 711-43.
Burmeister, Edwin, and Marjorie B. McElroy. 1988. "Joint Estimation of Factor Sensitivities and
Risk Premia for the Arbitrage Pricing Theory." Journal of Finance 43 uune):721-33.
Burmeister, Edwin, and Kent D. Wall. 1986. "The Arbitrage Pricing Theory and Macroeconomic
Factor Measures." The Financial Review 21 (February):1-20.
Burmeister, Edwin, Kent D. Wall, and James D. Hamilton. 1986. "Estimation of Unobserved
Expected Monthly Inflation Using Kalman Filtering." Journal of Business and Economic Statistics
4:147-60.
Chamberlain, G., and M. Rothschild. 1983. "Arbitrage, Factor Structure, and Mean-Variance
Analysis on Large Asset Markets." Econometrics 51: 1281-304.
Chen, Nai-fu. 1983. "Some Empirical Tests of the Theory of Arbitrage Pricing." Journal of
Finance 38 (December):1392-414.
Chen, Nai-fu, and Jonathan E. Ingersoll, Jr. 1983. "Exact Pricing in Linear Factor Models with
Finitely Many Assets: A Note." Journal of Finance 38 uune):985-88.
Chen, Nai-fu, Richard Roll, and Stephen A. Ross. 1986. "Economic Forces and the Stock
Market." Journal of Business 59 Uuly):383-403.
Cho, D. Chinhyung. 1984. "On Testing the Arbitrage Pricing Theory: Inter-Battery Factor
Analysis. " Journal of Finance 39 (December): 1485-1502.
Cho, D. Chinhyung, Edwin Elton, and Martin Gruber. 1984. "On the Robustness of the Roll and
Ross Arbitrage Pricing Theory. "Journal of Financial and Quanitative Analysis 19 (March):1-10.
Cho, D. Chinhyung., C. S. Em, and L. W. Senbet. 1986. "International Arbitrage Pricing

Theory: An Empirical Investigation." Journal of Finance 41 ume):313-30.
Cho, D. Chinhyung, and Wiarn Taylor. 1987. "The Seasonal Stability of the Factor Structure of
Stock Returns. " Journal of Finance 42 (December): 1195-21 1.
Connor, Gregory. 1984. "A Unified Beta Pricing Theory." Journal of Economic Theory 34: 13-31.
Connor, Gregory, and Robert A. Korajczyk. 1986. "Performance Measurement with the
Arbitrage Pricing Theory: A New Framework for Analysis." Journal of Financial Economics 15
UanuaryIFebruary):373-94.
. 1988. "Risk and Return in an Equilibrium APT: Application of a New Test Methodology."
Journal ofFinancia1 Economics 21 (September):255-89.
Copeland, Thomas E., and J. Fred Weston. 1988. Financial Theoly and Corporate Policy.
Reading, Mass.: Addison-Wesley Publishing Company.
Cox, John C., Jonathan E. Ingersoll, Jr., and Stephen A. Ross. 1985. "An Intertemporal General
Equilibrium Model of Asset Prices." Econometrics 53:363-84.
Dhrymes, Phoebus, Irwin Friend, and Bulent Gultekin. 1984. "A Critical Reexamination of the
Empiral Evidence on the Arbitrage Pricing Theory." Journal of Finance 39 Uune):323-46.
Dybvig, Philip. 1983. "An Explicit Bound on Deviations from APT Pricing in a Finite Economy."
Journal of Financial Economics 12 (December):483-96.
Dybvig, Philip H., and Stephen A. Ross. 1985. "Yes, the APT is Testable." Journal of Finance
40 (December): 1173-88.
Elton, Edwin J., and Martin J. Gruber. 1970. "Homogeneous Groups and the Testing of
Economic Hypotheses." Journal of Financial and QuantitativeAnalysis 5 Uanuary):581-602.
. 1971. "Improved Forecasting Through the Design of Homogenous Groups." Journal of

Business 44 (October):432-50.
. 1987. Modem Portfolio Theory and Investment Analysis, Third Edition. New York: John
Wiey & Sons.
. 1990. "A Multi-index Risk Model of the Japanese Stock Market." In Japanese Capital
Markets, eds. Edwin J. Elton and Martin J. Gruber, 127-54. New York: Harper and Row.
Elton, Edwin J., Martin J. Gruber, and Prafdla Nabar. 1988. "Bond Returns, Immunization, and
the Return Generating Process." Studies in Banking and Finance 5: 125-54.
Engle, Robert F. 1982. "Autoregressive Conditional Heteroskedasticity with Estimates of the
Variance of U. K. Inflation. " Econometrica 51: 987-1008.
Engle, Robert F, Che-Hsiung Hong, and Alex Kane. 1990. "Valuation of Variance Forecasts with
S i a t e d Option Markets." Sun Diego Economics Discussion Paper, University of California.
Fama, Eugene F., and Kenneth R. French. 1992. "The Cross-Section of Expected Stock
Returns." Journal of Finance 47 (June):427-65.
Fama, Eugene F., and James MacBeth. 1973. "Risk, Return, and Equilibrium: Empirical Tests. "
Journal of Political Economy (May/June):607-36.
Farrell, James. 1974. "Analyzing Covariation of Returns to Determine Homogeneous Stock
Groupings. " Journal of Business 47 (April): 186-207.
Francis, Jack Clark. 1991. Investments: Analysis and Management. New York: McGraw-Hill, Inc.
Gibbons, Michael R. 1982. "Multivariate Tests of Financial Models: A New Approach." Journal
of Financial Economics 10 (March):%27.
Grinold, Richard C. 1990. "The Fundamental Law of Active Management." In Managing
Institutional Assets, ed. Frank J. Fabozzi. New York: Harper & Row.
Hamao, Yasushi. 1990. "An Empirical Examination of the Arbitrage Pricing Theory: Using
Japanese Data." In Jeanese Capital Markets, eds. Edwin J. Elton and Martin J. Gruber, 155-74.
New York: Harper and Row.
Hansen, L. P., and S. F. Richard. 1987. "The Role of Conditioning Information in Deducing
Testable Restrictions Implied by Dynamic Asset Pricing Models." Econometrica 55587-613.
Hubennan, Gur. 1982. "A Simple Approach to Arbitrage Pricing Theory." Journal of Ecunomic
Theory 78: 183-91.
Huberman, Gur, and Robert Stambaugh. 1987. "Mimicking Portfolios and Exact Arbitrage
Pricing. " Journal of Finance 42 (March): 1-9.
Ingersoll, Jonathan E. Jr. 1984. "Some Results in the Theory of Arbitrage Pricing." Journal of
Finance 39 (September): 1021-39.
Ingersoll, Jonathan E. Jr. 1987. Theory of Financial Decision Making. Totowa, NR: Rowrnan &
Littlefield.
King, Benjamin F. 1966. "Market and Industry Factors in Stock Price Behavior." Journal of
Business 39 Uanuary): 139-40.
Lawley, D. N., and A. E. Maxwell. 1971. Factor Analysis as a Statistical Method. New York:
Macmillan.
Lehmann, Bruce N., and David M. Modest. 1987. "Mutual Fund Performance Evaluation: A
Comparison of Benchmarks and Benchmark Comparisons."Journal of Finance 42 dune): 233-65.
. 1988. "The Empirical Foundations of the Arbitrage Pricing Theory." Journal of Financial
Economics 20 (September):213-54.
Litner, John. 1965a. "The Valuation of Risk Assets and the Selection of Risky Investments in
Stock Portfolios and Capital Budgets." Review of Economics and Statistics (February): 13-37.
. 1965b. "Security Prices, Risk, and Maximal Gains from Diversification." Journal of
Finance 20 (December):587-615.
Litzenberger, Robert H., and Krishna Rarnaswamy. 1979. "The Effect of Personal Taxes and
Dividends on Capital Asset Prices: Theory and Empirical Evidence." Journal of Financial
Economics 72 Uune): 163-96.
Markowitz, Hany M. 1952. "Portfolio Selection," Journal of Finance 7 (March):77-91.
. 1959. Po&lw Selection: Ejicient Diversificationof Investments. New York: John Wiey
& Sons.
McElroy, Marjorie B., and Edwin Burmeister. 1988. "Arbitrage Pricing Theory as a Restricted
Nonlinear Multiple Regression Model: ITNLSUR Estimates." Journal ofBusiness and Economic
Statistics 6:29-42.
McElroy, Marjorie B., Edwin Burmeister, and Kent D. Wall. 1985. "Two Estimators for the APT
Model when Factors are Measured. " Economics Letters 19:271-75.
Merton, Robert C. 1972. "An Analytical Derivation of the Efficient Portfolio." Journal of
Financial and Quantitative Analysis 7 (September): 1851-72.
Mossin, Jan. 1966. "Equilibrium in a Capital Asset Market." Econometrica (October): 768-83.
Research Foundation of the Institute of Chartered Financial Analysts. 1991. The Founders of
Mo&rn Finance: Their Prize Winning ConcQts and 1990 Nobel Lectures. Charlottesville, Va.
Roll, Richard, and Stephen A. Ross. 1980. "An Empirical Investigation of the Arbitrage Pricing
Theory. " Journal of Finance 35 (December): 1073-103.
Rosenberg, Barr. 1974. "Extra-Market Components of Covariance in Security Returns-I."

Journal of Financial and Quantitative Analysis 9 (March):263-74.
Rosenberg, Barr, and Viay Marathe. 1975. "The Prediction of Investment Risk: Systematic and
Residual Risk." Proceedings of the Seminar on the Analysis of Security Prices, University of
Chicago (November): 85-224.
. 1979. "Test of Capital Asset Pricing Hypotheses." Research in Finance (annual):ll5-

224.
Ross, Stephen A. 1976. "The Arbitrage Theory of Capital Asset Pricing." Journal of Economic
Theory (December): 341-60.
. 1977. "Return, Risk, and Arbitrage." In Risk and Return in Finance, eds. I . Friend and
J. Bicksler, 189-219. Cambridge, Mass.: Ballinger.
. 1978. "Mutual Fund Separation in Financial Theory-The Separating Distribution."

Journal of Economic Theory 12254-86.
Ross, Stephen A., and Richard Roll. 1984. "The Arbitrage Pricing Theory Approach to Strategic
Portfolio Planning." Financial Analysts Journal 40 (Maydune): 14-26.
Ross, Stephen A., and Randolph W. Westerfield. 1988. Corporate Finance. St. Louis: T i e s
MirrorMosby College Publishing.
Rudd, Andrew, and Henry K. Clasing, Jr. 1988. Modern Portfolio Theory, Second Edition, Orinda,
California: Andrew Rudd.
Schwartz, G. 1978. "Estimating the Dimension of a Model." Annals of Statistics 6461-64.
Sharpe, Wiam F. 1964. "Capital Asset Prices: A Theory of Market Equilibrium Under
Conditions of Risk." Journal of Finance 19 (September):425-42.
. 1977. "The Capital Asset Pricing Model: A 'Multi-Beta' Interpretation." In Financial

Decision Making Under Uncertainty, eds. Haim Levy and Marshall Sarnat. New York: Academic
Press.
. 1984. "Factor Models, CAPMs, and the APT." Journal of Portfolio Management 11
(Fall):21-25.
Treynor, Jack L. 1961. "Toward a Theory of the Market Value of Risky Assets." Unpublished
manuscript.
Trzcinka, C. 1986. "On the Number of Factors in the Arbitrage Pricing Model." Journal of
Finance 41 (June):347-68.
Wei, K. C. John. 1988. "An Asset-Pricing Theory Unifying the CAPM and APT." Journal of
Finance 43 (September):881-92.

A Practitioner's Guide To Factor Models

Uploaded by

Copyright:

Available Formats

A Practitioner's Guide To Factor Models

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

A Practitioner's Guide To Factor Models

Uploaded by

Copyright:

Available Formats

A Practitioner's Guide to

The Research Foundation of

The mission of the Research Foundation is to

Ideas Whose Time

The Research Foundation of

A Practitioner's Guide to Arbitrage Pricing

Bibliography ......... . . .......... . . .. ....... . . . 81

Sharpe (1984) effectively summarizes why investment professionals should

complexity of the investment decision-making process, the use of quantitative

John W. Peavy 111, CFA

A fundamental principle of finance is the trade-off between risk and return.

The APT Basics

than a priori expectations. This extraordinary performance defines ex post APT

Postulate 1. In every time period, the difference between the actual

for all j = 1, . .. , K and for all t # t'.

Postulate 2. Pure arbitrage profits are impossible. Because of compe-

Although equation (2) holds only approximately, with additional assumptions,

Macroeconomic Forces Impacting Stock Returns

and explain as much as possible of the variation in stock returns. Extensive

Risk Exposure Profiles and Risks--Return Trade-off

TABLE 1. Calculation of Expected Excess Return for the

Confidence risk 0.27 2.59% 0.70%

Confidence Risk 0.73

How to Use the APT: Some Examples

Evaluationof Macroeconomic Risk Exposuresand Attribution of

Index Portfolios. A closely related use of the APT is in the formation of

Tilting, or Making a Factor Bet. Good managers may possess

Multimanager Fund Performance. Most sponsors employ more

Optimized Risk Control with Manager-Supplied Rankings.

subject to the constraint that the portfolio betas,

Long-Short Investment Strategies. Long-short, or market-neu-

Mean--Variance Efficiency. The standard optimization problem of

factors do not contribute to expected return, they do help to explain

and hence is generated by a LFM with

Because by Postulate 1cov[e,(t), fi(t)] = 0, it follows that cov[a,(t), rm(t)] =

It then follows immediately that the APT holds provided that

This exercise establishes (a) above.

This formulation proves (b) above.

Single- vs. Multi-Index Models

E(eitfit)= 0 for all i and all j.

Estimating the Multi-Index Model

Define the optimal composition of each index (the weight on each

The Choice of Data. The input to factor or principal component analysis

others negative. The sensitivities on a portfolio of securities is a weighted

The Number of Indexes. A second design issue that affects results is

The Nonuniqueness of Factors. A third design issue is concerned

Computational Problems. A fourth problem with factor analysis is the

Applications of Multi-Index Models

A Simple Example. To illustrate the principles discussed earlier, we

TABLE 1. Variance--Covariance Matrix and Correlations

Country Belgium Canada France United States

Belgium 50.34 17.55 34.22 13.79

We ran a principal components analysis on the variance-covariance matrix

This index is long in North America and short in Europe. It could be

TABLE 2. Returns in Countries and on Factors

Period Belgium Canada France U.S. Index 1 Index 2

1 2.08% 14.64% 7.25% 6.08% 10.99% 9.85%

The b's (the sensitivities, or factor loadings) could be estimated by