The Journal of Financial Research • Vol. XXX, No.

1 • Pages 111–127 • Spring 2007

APPLICABILITY OF THE FAMA-FRENCH THREE-FACTOR MODEL

IN FORECASTING PORTFOLIO RETURNS

Ou Hu
Youngstown State University

Abstract

For the model-based estimation of the equity cost of capital, evidence shows
that the common practice of using the average historical factor premiums as the
estimates of the next-period factor premiums generates inaccurate estimates. I
propose an alternative way to estimate factor premiums by using the structural
variables that are important predictors of future asset returns. Based on the out-
of-sample results from a trading strategy with four in-sample model-selection
criteria, I find that my estimation procedure performs better than the common
practice even when transaction costs are considered.

JEL Classification: G12, G31, C13, C22

I. Introduction

Since its origination in the 1960s, the capital asset pricing model (CAPM) has been
used by most project managers to estimate the equity cost of capital. In reality,
practitioners estimate the future market risk premium by averaging the long-term
historical market risk premium. However, many studies show that the historical
average market excess return is higher than the actual market risk premium. Thus,
the estimate of equity returns overstates what rational investors would have expected
to earn. The poor performance of this common practice has cast doubt on the
application of the CAPM.
As Fama and French (1997) demonstrate, the imprecise use of the CAPM
to estimate equity returns has two sources: the estimation error of the risk loadings
and the estimation error of the risk premiums. They conclude that the uncertainty
about risk premiums is responsible for a larger part of the problem in estimating
the equity cost of capital.
Ferson and Locke (1998), who analyze the sources of errors in CAPM-based
estimates of expected returns on industry portfolios, reach a similar conclusion:
that errors in estimating betas (factor loadings) probably do not matter as much as

The author is grateful for the comments from Ronald J. Balvers, Joseph Palardy, and especially William
T. Moore (the former editor) and an anonymous referee.

111
112 The Journal of Financial Research

errors in estimating market premium. Pastor and Stambaugh (1999) use a Bayesian
approach to examine the estimation of costs of equity for individual firms and
to compare estimates of three factor-based pricing models. They find that model
uncertainty is less important than parameter uncertainty within any given model,
and without mispricing uncertainty, estimation errors of factor premiums probably
account for more of the uncertainty about the cost of capital.
In this article, I examine the effectiveness of the Fama and French (1993)
three-factor model in predicting future returns. Following Elton (1999),1 I propose
an alternative way to estimate factor premiums. I do so by using several structural
variables, such as term premium, default risk premium, and dividend yield. Instead
of imposing a fixed model in forecasting risk premiums, I assume that investors do
not know the model specification but search for the optimal specification according
to some model selection criteria. To see whether the predictability of portfolio excess
returns can be exploited successfully, I use a trading strategy: to hold the portfolio(s)
with the highest expected excess return and to sell the portfolio(s) with the lowest
expected excess return. Then, I calculate the realized excess returns for long and
short positions, and the excess profits for a zero-investment strategy (long minus
short).
Based on the empirical results, my estimation procedure performs bet-
ter than the common practice. Without consideration of transaction costs, almost
all the trading strategies generate significant positive zero-investment profits. Al-
though the transaction costs can reduce the profit, by imposing a transaction cost
filter, investors can still earn significant positive profits from most of their trading
strategies.
Moreover, a risk analysis study shows that the excess returns from the
trading strategies cannot be explained by risk factors. The abnormal return from
the long position is significantly higher than that from the buy-and-hold benchmark,
the market portfolio. Finally, Bossaerts and Hillion’s (1999) forecast betas provide
some support for the Fama-French (1993) model in predicting asset returns.
The method I apply is fundamentally different from that used in previous
studies. For example, Ferson and Harvey (1999) allow factor loadings to be condi-
tional on some predetermined variables to test the Fama-French (1993) three-factor
model for the cross-section of stock returns. Cooper, Gulen, and Vassalou (2001)
use business-cycle variables and macroeconomic variables to predict asset future
returns. These authors find some evidence that the size factor and book-to-market
(B/M) factor are representatives of fundamental economic risks. However, those
studies do not examine the applicability of any particular asset pricing model.

1
While arguing that the average realized returns are poor proxies for expected returns, Elton (1999)
proposes some alternative ways to estimate expected returns.
 Forecasting Portfolio Returns 113

II. Endogenization of Model Selection

The Fama-French (1993) three-factor model is as follows:

rit − rft = bi (rmt − rft ) + si rSMB,t + h i rHML,t + eit , (1)

where rit is the return on asset i; rft is the return on the risk-free asset; rmt is the
return on the overall market; the size factor, r SMB,t , is the return on a zero-investment
portfolio that is long on small stocks and short on big stocks; the B/M factor, r HML,t ,
is the return on a zero-investment portfolio that is long on high-B/M stocks and
short on low-B/M stocks;2 bi , si , and hi are factor loadings; and eit is a mean-zero
regression disturbance. Fama and French (1993) generate 25 portfolios sorted by
size and B/M factors. For each of the 25 regressions in the form of model (1), the
typical R2 is above 0.9. Fama and French (1997) use the same regression on 48
industries and find that the average R2 (0.68) is slightly higher than that (0.63) of
the CAPM. This and other evidence support the power of the Fama-French (1993)
three-factor model in explaining asset returns.
To investigate its forecasting ability, I rewrite model (1) in a conditional
form:

rit − rft = b̂i E t−1 (rmt − rft ) + ŝ i E t−1rSMB,t + ĥ i E t−1rHML,t + εit . (2)

As model (2) indicates, I must estimate all the three-factor premiums at time t by
means of historically available information up to time t–1. But instead of taking the
average of historical risk premiums, I propose using structural variables to estimate
the three-factor premiums. Moreover, I assume that investors have no knowledge
of what specific model will be applied to predict risk premiums but that they can
select an optimal model specification based on some criteria.
Following Pesaran and Timmermann (1995), I apply four model-selection
criteria: adjusted R2 , Akaike’s information criterion (AIC), Schwarz’s Bayesian
information criterion (BIC), and the sign criterion. Once investors choose a model,
they can make one-period-ahead predictions of risk premiums. As time progresses
and more historical information is available, investors with no assumption on the
model specification must reevaluate their model selection. Consequently, investors
might not choose the same model at time t+1 as they did at time t.
Suppose that at period t an investor searches over a set of k variables to
make one-period-ahead predictions of risk premiums (r mt − r ft ), rSMB,t , and rHML,t .
The total number of all the possible combinations of k variables [x 1 , x 2 , . . . , x k ]

2
Details on how to construct the SMB and HML factors are available on Kenneth French’s Web site
(http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/).
114 The Journal of Financial Research

is 2k , where k i denotes the ith combination of k variables. For each of the 2k

combinations, I conduct a linear regression as follows:

rps = X i,s−1 βi + ωi,s s = 1, 2, . . . , t, (3)

where rps is a risk premium at time s, and X i ,s−1 is a row vector that contains a
combination of k variables − k i . To predict r p t+1 , I obtain an ordinary least squares
(OLS) estimate β̂i for each of the 2k regressions. Next, I choose an optimal β̂i
according to a model-selection criterion and calculate an estimated rpt+1 by multi-
plying the optimal β̂i with its corresponding X i,t .

III. Data, Method, and In-Sample Estimation

All variables are measured at monthly frequencies from 1954:04 to 2001:10. The
start and end dates are determined by data availability. I examine the 17 equally
weighted industry portfolios3 defined by Fama and French (1993). The Fama-
French three factors—the market excess return (MKTEXRT ), the size factor (SMB),
and the B/M factor (HML)—are obtained from Kenneth French’s Web site (http://
mba.tuck.dartmouth.edu/pages/faculty/ken.french/).
To predict the three factor premiums, I use the following variables: divi-
dend yield (DIV, the difference between market return with dividend and market
return without dividend), the nominal one-month Treasury bill rate (TB), the in-
dustrial production growth rate (IP), the term premium (TERM, the difference
between 10-year and three-month Treasury yields), and the default risk premium
(DEFP, the spread between Moody’s Bbb and Aaa corporate bond yields). I obtain
DIV from the Center for Research in Security Prices (CRSP); TB from Ibbot-
son and Associates; and IP, TERM, and DEFP from the Federal Reserve Bank of
St. Louis.
Table 1 reports descriptive statistics of the 17 equally weighted industry
portfolios and predictive variables. No obvious patterns are evident across the means
and standard deviations of the 17 industries. Unlike the portfolios formed on size
or B/M ratio, the 17 industries were formed according to the characteristics and
activities of individual firms. Therefore, the advantage of examining 17 industries
is that I can avoid the inherent relation between the two factors (SMB and HML)
and the size and B/M portfolios.

3
Besides the 17 equally weighted industry portfolios, I apply the same method to 17 value-weighted
industry portfolios, and 30, 38, and 48 value-weighted and equally weighted industry portfolios. The results
are consistent over all portfolios grouped by different definitions, although the results of value-weighted
portfolios are not as significant as those of equally weighted portfolios.
 Forecasting Portfolio Returns 115

TABLE 1. Summary Statistics of 17 Equally Weighted Industry Portfolios and Other Predictive
Variables: Monthly Data from 1954:04 to 2001:10.
Panel A. Industry Portfolios
Industry
Portfolio Mean Std. Dev. ρ1 ρ6 ρ 12 ρ 36

Food 1.14 4.40 0.20 (.000) −0.027 (.512) 0.12 (.006) 0.07 (.104)
Mines 1.22 6.58 0.13 (.002) 0.012 (.777) 0.08 (.062) 0.18 (.000)
Oil 1.32 6.60 0.14 (.001) 0.022 (.596) 0.09 (.032) 0.08 (.091)
Clths 1.00 6.03 0.25 (.000) 0.014 (.747) 0.13 (.003) 0.11 (.016)
Durbl 1.14 6.17 0.22 (.000) −0.038 (.361) 0.11 (.009) 0.07 (.143)
Chems 1.23 5.27 0.15 (.000) −0.040 (.341) 0.08 (.072) 0.09 (.051)
Cnsum 1.57 6.57 0.18 (.000) −0.016 (.702) 0.02 (.585) 0.05 (.345)
Cnstr 1.21 5.91 0.22 (.000) −0.025 (.553) 0.13 (.003) 0.09 (.042)
Steel 1.08 6.16 0.14 (.001) −0.013 (.759) 0.04 (.409) 0.08 (.091)
FabPr 1.21 5.62 0.19 (.000) −0.006 (.894) 0.12 (.006) 0.09 (.043)
Machn 1.39 6.91 0.21 (.000) −0.004 (.930) 0.05 (.303) 0.07 (.133)
Cars 1.15 6.06 0.21 (.000) 0.018 (.666) 0.10 (.018) 0.07 (.090)
Trans 1.16 5.80 0.20 (.000) 0.040 (.342) 0.10 (.019) 0.07 (.089)
Utils 1.07 3.36 0.09 (.040) −0.015 (.719) 0.04 (.322) 0.04 (.320)
Rtail 1.09 5.71 0.25 (.000) −0.014 (.740) 0.09 (.044) 0.08 (.074)
Finan 1.25 4.87 0.25 (.000) 0.036 (.389) 0.13 (.002) 0.05 (.239)
Other 1.31 6.43 0.21 (.000) −0.013 (.755) 0.06 (.205) 0.07 (.119)
Panel B. Explanatory Variables
Explanatory
Variable Mean Std. Dev. ρ1 ρ6 ρ 12 ρ 36
MKTEXRT 0.57 4.31 0.06 (.140) −0.03 (.503) 0.03 (.525) −0.01 (.749)
SMB 0.13 3.02 0.08 (.065) 0.06 (.146) 0.15 (.001) 0.06 (.136)
HML 0.38 2.80 0.14 (.001) 0.08 (.063) 0.04 (.372) 0.04 (.450)
TB 0.44 0.23 0.96 (.000) 0.86 (.000) 0.76 (.000) 0.46 (.000)
IP 0.25 0.92 0.42 (.000) 0.03 (.536) −0.15 (.000) −0.04 (.282)
TERM 0.11 0.10 0.95 (.000) 0.65 (.000) 0.43 (.000) −0.09 (.032)
DEFP 0.08 0.04 0.97 (.000) 0.83 (.000) 0.69 (.000) 0.35 (.000)
DIV 0.29 0.18 −0.13 (.001) 0.92 (.000) 0.92 (.000) 0.84 (.000)

Note: Data and definitions of 17 industries are available on Kenneth French’s Web site (http://mba.
tuck.dartmouth.edu/pages/faculty/ken.french/). MKTEXRT is the excess return of market portfolio
over the risk-free rate. SMB is the excess return from a zero-investment strategy of buying small
portfolios and selling large portfolios. HML is the excess return from a zero-investment strategy
of buying value portfolios (high book-to-market (B/M) ratio) and selling growth portfolios (low
B/M ratio). TB is the nominal one-month Treasury bill rate. IP is industrial production growth rate.
TERM is the term premium, defined as the difference between 10-year and three-month Treasury
yields. DEFP is the spread between Moody’s Bbb and Aaa corporate bond yields (i.e., default risk
premium). DIV is the dividend yield (i.e., the difference between market return with dividend and
market return without dividend). TB, IP, TERM, and DEFP are all obtained from Federal Reserve
Bank of St. Louis. DIV is obtained from the Center for Research in Security Prices (CRSP) tapes.
The variable ρ i represents the autocorrelation coefficient over i month(s). The p-values are given in
parentheses.
116 The Journal of Financial Research

I find significant positive first-order autocorrelation over the 17 indus-

tries, with all p-values less than .05. This result is not surprising, based on the
finding in Moskowitz and Grinblatt (1999) that there is a strong and prevalent mo-
mentum in industries. Autocorrelations of MKTEXRT are weak. In contrast, SMB
and HML both exhibit strong first-order autocorrelation with p-values of .065 and
.001, respectively. Therefore, I use lags of SMB and HML as predictors for their
next period’s values. The other explanatory variables—TB, IP, TERM, DEFP, and
DIV—exhibit strikingly significant autocorrelations over almost all time intervals
(from 1 month to 36 months), and the magnitude of autocorrelations tends to de-
crease as the time interval increases.
I evaluate the predictability of portfolio returns by using a trading strategy
that includes in-sample estimation and out-of-sample evaluation. For the in-sample
estimation, I apply 5- and 10-year rolling windows as well as a 5-year expanding
window. For instance, with the 5-year expanding window, the data from 1959:04
to 1964:03 are used to estimate excess returns of the 17 industry portfolios for
1964:04. First, I obtain the estimates of the three risk premiums: E t−1 (r mt − r ft ),
E t−1r SMB,t , and E t−1r HML,t . Next, I develop the risk loadings—b̂i , ŝ i , and ĥ i —for
each industry portfolio by estimating model (1). Then, I calculate the estimated
excess returns of the 17 industry portfolios at month 1964:04 by

E t−1 (rit − rft ) = b̂i E t−1 (rmt − rft ) + ŝ i E t−1rSMB,t + ĥ i E t−1rHML,t . (4)

There is no intercept in this equation. If I assume that the Fama and French (1993)
three-factor model correctly prices assets, the intercept that measures mispricing
should be zero.
Based on the estimated excess returns of the 17 industry portfolios, the
trading strategy is to hold the industry portfolio with the highest expected excess
return and sell the industry portfolio with the lowest expected return. The next step
is to record the realized returns of the long position, the short position, and the
zero-investment strategy (the difference between the long position and the short
position) for 1964:04. With the in-sample window moving one month forward, the
same steps are repeated 451 times until the end of the entire period, 2001:10. Finally,
the average realized excess returns are calculated. Using the same starting point,
1964:03, the same procedure is applied to the 5- and 10-year rolling windows.
Table 2 presents the monthly in-sample estimation of the Fama-French
(1993) three factors over the entire period 1954:04 to 2001:10. As indicated by the
p-values, among all the predictive variables, the 12-month lagged value of TERM,
the 1- and 12-month lagged values of DEFP, the 12-month lagged value of DIV,
and the 1-month lagged value of TB are important in predicting market excess
returns. In contrast, only the 1-month lagged value of MKTEXRT and the 12-month
lagged value of TERM play a significant role in predicting the size factor, SMB. In
TABLE 2. In-Sample Estimation of the Fama-French (1993) Three Factors.

MKTEXRT HML SMB IP IP TERM TERM DEFP DEFP DIV DIV TB

Factor Intercept (–1) (–1) (–1) (–1) (–12) (–1) (–12) (–1) (–12) (–1) (–12) (–1) R2

MKTEXRT −0.202 −0.023 −0.055 0.065 0.030 −0.219 1.755 5.196 37.007 −28.246 0.134 1.677 −2.557 0.038
(.753) (.631) (.440) (.296) (.886) (.288) (.542) (.036) (.000) (.002) (.900) (.106) (.058)
SMB −0.001 0.159 0.078 0.027 −0.026 0.030 0.439 −3.171 4.103 2.508 0.030 −0.083 −0.419 0.040
(.999) (.000) (.122) (.545) (.861) (.837) (.828) (.068) (.563) (.698) (.968) (.910) (.658)
HML 0.580 0.051 0.149 −0.002 −0.008 −0.055 −3.256 0.524 −10.879 20.356 −0.535 −1.706 −0.107 0.039
(.166) (.093) (.001) (.957) (.957) (.679) (.082) (.744) (.098) (.001) (.440) (.012) (.903)

Note: Monthly returns of the Fama-French (1993) three factors are regressed for the entire period, 1954:04 to 2001:10, on the following explanatory
variables: one-month lags of the three factors: market excess return (MKTEXRT), size factor (SMB), and book-to-market factor (HML); one-month lags of
the business-cycle-related variables: growth rate of industrial production (IP), the difference between 10-year and three-month Treasury yields (TERM), the
spread between Moody’s Bbb and Aaa corporate bond yields (DEFP), dividend yield, that is, the difference between market return with dividend and market
Forecasting Portfolio Returns

return without dividend (DIV ), and the nominal one-month Treasury bill rate (TB); and 12-month lags of IP, TERM, DEFP, and DIV . The p-values are given in
parentheses. The R2 s are corrected for degrees of freedom.
117
118 The Journal of Financial Research

estimating the future value of HML, the important variables include the 1-month
lagged values of MKTEXRT, HML, TERM, and DEFP, and the 12-month lagged
values of DEFP and DIV . The adjusted R2 s for the three regressions have about
the same value of 0.04, which indicates that over the entire period, the overall
predictability of the three factors is about the same.

IV. Out-of-Sample Performance of a Trading Strategy

Performance of the Common Practice

To report out-of-sample performance, Max1 is the return from the long position, that
is, to hold the portfolio with the highest estimated excess return; Min1 is the return
from the short position, that is, to sell short the portfolio with the lowest estimated
excess return; and Max1–Min1 is the profit from a zero-investment trading strategy,
that is, to hold (long) the portfolio with the highest expected excess return and sell
(short) the portfolio with the lowest expected excess return. Max3–Min3 is the profit
from a zero-investment strategy of holding the three portfolios with the highest
expected excess returns and selling the three portfolios with the lowest expected
excess returns.
Following the common practice, I use the historical averages of the Fama-
French (1993) three-factor premiums as the estimates of their future values. The
factor loadings b̂i , ŝ i , and ĥ i are estimated by model (1). Then, using equa-
tion (4), I calculate the expected excess returns. Table 3 shows that the out-of-
sample performance of the common practice is poor indeed. For Max1–Min1 with
the five-year expanding window, the difference (Max1–Min1 = 1.506%) between
the average realized annualized return (9.115%) of the long position (Max1) and
the average annualized return (7.506%) of the short position (Min1) is not signifi-
cantly different from zero (t = 0.571), which contravenes the spirit of the trading
strategy in which Max1 is expected to be significantly greater than Min1. More-
over, if the market excess return is treated as a benchmark, Max1 is expected to
be significantly greater than MKTEXRT, and Min1 is expected to be significantly
lower than MKTEXRT. However, based on the common practice, not only are Max1
and Min1 similar over all the in-sample windows, but they also are not statistically
different from annualized market excess return (5.352%). A similar result holds
for Max3–Min3. Therefore, I conclude that this trading strategy provides evidence
against the common practice.

Performance of the New Approach

In contrast, Tables 4 and 5 present different results for the performance of the Fama-
French (1993) three-factor model based on the estimation procedures described in
section III. Both the fixed model and the model with the adjusted R2 demonstrate
TABLE 3. Out-of-Sample Trading Strategy Performance of Fama-French (1993) Three-Factor Model Based on the Common Practice: Monthly Results from 1954:04 to 2001:10.

5-Year Expanding Window 5-Year Rolling Window 10-Year Rolling Window

Fama- Fama- Fama-

Trading Mean Terminal Sharpe French Mean Terminal Sharpe French Mean Terminal Sharpe French
Strategy Return t-stat Wealth Ratio α t-stat Return t-stat Wealth Ratio α t-stat Return t-stat Wealth Ratio α t-stat
Panel A. Max1–Min1
Max1 9.115 2.707 1,263 0.127 −0.063 −0.542 8.438 2.400 930 0.113 0.069 0.441 8.783 2.386 986 0.112 −0.160 −1.071
Min1 7.506 3.042 1,029 0.143 0.094 0.672 2.274 0.697 112 0.033 −0.533 −2.844 5.360 1.696 364 0.080 −0.188 −0.980
Max1–Min1 1.506 0.571 108 0.027 −0.158 −0.862 6.039 1.789 425 0.084 0.602 2.245 3.263 1.068 177 0.050 0.028 0.117
Max1–mkt 3.587 1.750 284 0.082 −0.063 −0.542 2.942 1.275 209 0.060 0.069 0.441 3.271 1.360 229 0.064 −0.160 −1.071
Min1–mkt 2.053 1.022 163 0.048 0.094 0.672 −2.935 −1.197 21 −0.056 −0.533 −2.844 0.007 0.003 65 0.000 −0.188 −0.980
Panel B. Max3–Min3
Max3 7.408 2.122 651 0.100 −0.242 −2.265 10.354 2.957 1,848 0.139 0.172 1.508 9.071 2.503 1,114 0.118 −0.116 −1.199
Min3 7.817 2.770 1,006 0.130 0.037 0.385 7.921 2.475 903 0.117 −0.094 −0.738 8.278 2.703 1,075 0.127 0.001 0.005
Max3–Min3 −0.382 −0.214 69 −0.010 −0.279 −1.977 2.269 1.005 165 0.047 0.266 1.457 0.737 0.362 99 0.017 −0.116 −0.748
Max3–mkt 1.959 0.970 157 0.046 −0.242 −2.265 4.769 2.280 434 0.107 0.172 1.508 3.545 1.624 271 0.076 −0.116 −1.199
Min3–mkt 2.350 1.476 202 0.070 0.037 0.385 2.449 1.241 191 0.058 −0.094 −0.738 2.789 1.460 220 0.069 0.001 0.005
Note: Following the common practice, I use the historical averages of the Fama-French (1993) three-factor premiums as the estimates of their future values. The factor loadings, b̂i , ŝ i , and ĥ i are
estimated using model (1) and the expected excess returns are calculated using equation (4). The sample comprises 17 equally weighted industry portfolios with three in-sample windows: the
Forecasting Portfolio Returns

5-year expanding window, and the 5-year and 10-year rolling windows. Reported in this table are the average annualized returns, t-statistics, terminal wealth (assuming investors start with $100
in 1954:04 and reinvest portfolio income every month until the end of the sample), Sharpe ratio, and Fama-French three-factor model alphas. The notations are as follows: Max1 is the realized
return from holding the portfolio with the highest expected return, Min1 is the realized return from selling the portfolio with the lowest expected return, Max1-Min1 is the realized return from the
zero-investment portfolio (difference between long and short positions), Max1–mkt is the realized excess return of long position over market excess return, Min1–mkt is the realized excess return
of short position over market excess return, Max3 is the realized return from holding three portfolios with the highest expected returns, Min3 is the realized return from selling three portfolios
with the lowest expected returns, and Max3–Min3, Max3–mkt, and Min3–mkt are defined the same way as Max1–Min1, Max1–mkt, and Min1–mkt, respectively.
119
 120

TABLE 4. Out-of-Sample Trading Strategy Performance of the Fama-French (1993) Three-Factor Model with a Fixed Model in Predicting Risk Premiums: Monthly Results from 1954:04
to 2001:10.

5-Year Expanding Window 5-Year Rolling Window 10-Year Rolling Window

Fama- Fama- Fama-
Trading Mean Terminal Sharpe French Mean Terminal Sharpe French Mean Terminal Sharpe French
Strategy Return t-stat Wealth Ratio α t-stat Return t-stat Wealth Ratio α t-stat Return t-stat Wealth Ratio α t-stat
Panel A. Max1–Min1
Max1 13.947 3.951 6,259 0.186 0.372 2.248 15.081 3.842 7,683 0.181 0.485 2.498 13.957 3.93 6,209 0.185 0.470 2.665
Min1 1.486 0.419 72 0.02 −0.5 −2.71 4.1 1.121 182 0.053 −0.28 −1.31 3.061 0.845 125 0.04 −0.432 −2.148
Max1–Min1 12.295 3.387 3,359 0.159 0.876 2.988 10.586 2.634 1,557 0.124 0.764 2.342 10.6 2.821 1,775 0.133 0.902 2.938
Max1–mkt 8.195 3.423 1,340 0.161 0.372 2.248 9.276 3.304 1,709 0.156 0.485 2.498 8.204 3.277 1,300 0.154 0.470 2.665
Min1–mkt −3.686 −1.479 16 −0.07 −0.5 −2.71 −1.194 −0.447 39 −0.021 −0.28 −1.31 −2.184 −0.834 27 −0.039 −0.432 −2.148
Panel B. Max3–Min3
Max3 13.308 3.844 5,208 0.181 0.317 2.608 13.361 3.689 4,921 0.174 0.326 2.575 10.909 3.126 2,248 0.147 0.174 1.337
Min3 5.807 1.626 357 0.077 −0.18 −1.33 4.91 1.436 275 0.068 −0.26 −1.78 4.689 1.395 259 0.066 −0.313 −2.271
Max3–Min3 7.124 2.807 852 0.132 0.502 2.413 8.089 2.94 1,139 0.138 0.588 2.607 5.965 2.395 581 0.113 0.487 2.346
Max3–mkt 7.585 3.733 1,197 0.176 0.317 2.608 7.636 3.399 1,157 0.16 0.326 2.575 5.298 2.544 524 0.12 0.174 1.337
Min3–mkt 0.433 0.193 84 0.009 −0.18 −1.33 −0.422 −0.199 63 −0.009 −0.26 −1.78 −0.632 −0.301 58 −0.014 −0.313 −2.271
Note: To estimate the Fama-French (1993) three risk factors—market market excess return (MKTEXRT), size factor (SMB), and book-to-market factor (HM)—I apply a fixed model with lags of the
risk premiums. I also use all the structural variables: industrial production rate, term premium, default risk premium, dividend yield, and nominal one-month Treasury bill rate. The sample comprises
17 equally weighted industry portfolios with three in-sample windows: the 5-year expanding window, and the 5-year and 10-year rolling windows. Reported in this table are the average annualized
The Journal of Financial Research

returns, t-statistics, terminal wealth (assuming investors start with $100 in 1954:04 and reinvest portfolio income every month until the end of the sample), Sharpe ratio, and Fama-French three-factor
model alphas. The notations are as follows: Max1 is the realized return from holding the portfolio with the highest expected return, Min1 is the realized return from selling the portfolio with the lowest
expected return, Max1-Min1 is the realized return from the zero-investment portfolio (difference between long and short positions), Max1–mkt is the realized excess return of long position over market
excess return, Min1–mkt is the realized excess return of short position over market excess return, Max3 is the realized return from holding three portfolios with the highest expected returns, Min3 is the
realized return from selling three portfolios with the lowest expected returns, and Max3–Min3, Max3–mkt, and Min3–mkt are defined the same way as Max1–Min1, Max1–mkt, and Min1–mkt, respectively.
TABLE 5. Out-of-Sample Trading Strategy Performance of the Fama-French (1993) Three-Factor Model with Adjusted R2 as the Model-Selection Criterion: Monthly Results: 1954:04
to 2001:10.

5-Year Expanding Window 5-Year Rolling Window 10-Year Rolling Window

Fama- Fama- Fama-
Trading Mean Terminal Sharpe French Mean Terminal Sharpe French Mean Terminal Sharpe French
Strategy Return t-stat Wealth Ratio α t-stat Return t-stat Wealth Ratio α t-stat Return t-stat Wealth Ratio α t-stat
Panel A. Max1–Min1
Max1 13.945 3.928 6,271 0.185 0.424 2.351 13.125 3.514 4,385 0.165 0.435 2.236 16.337 4.129 11,735 0.194 0.706 3.867
Min1 2.102 0.593 92 0.028 −0.46 −2.47 5.35 1.474 294 0.069 −0.14 −0.65 0.835 0.246 61 0.012 −0.626 −3.248
Max1–Min1 11.621 3.073 2,499 0.145 0.881 2.868 7.413 1.852 514 0.087 0.575 1.736 15.385 3.881 8,535 0.183 1.332 4.392
Max1–mkt 8.193 3.165 1,259 0.149 0.424 2.351 7.411 2.708 910 0.128 0.435 2.236 10.474 3.597 2,553 0.169 0.706 3.867
Min1–mkt −3.099 −1.24 20 −0.058 −0.46 −2.47 −0.002 −0.001 60 0 −0.14 −0.65 −4.306 −1.743 12 −0.082 −0.626 −3.248
Panel B. Max3–Min3
Max3 13.031 3.85 4,892 0.181 0.318 2.632 11.117 3.12 2,358 0.147 0.172 1.353 12.077 3.431 3,314 0.162 0.317 2.572
Min3 6.647 1.87 487 0.088 −0.14 −0.99 7.312 2.117 642 0.1 −0.07 −0.47 5.185 1.545 315 0.073 −0.317 −2.292
Max3–Min3 6.019 2.438 588 0.115 0.455 2.228 3.568 1.354 232 0.064 0.24 1.085 6.58 2.565 709 0.121 0.634 3.035
Max3–mkt 7.322 3.675 1,102 0.173 0.318 2.632 5.497 2.482 545 0.117 0.172 1.353 6.412 2.953 762 0.139 0.317 2.572
Min3–mkt 1.234 0.554 114 0.026 −0.14 −0.99 1.868 0.86 146 0.04 −0.07 −0.47 −0.159 −0.076 70 −0.004 −0.317 −2.292
Note: To estimate the Fama-French (1993) three risk factors— excess return (MKTEXRT), size factor (SMB), and book-to-market factor (HM)—I use the optimal model based on the value of
adjusted R2 . The sample comprises 17 equally weighted industry portfolios with three in-sample windows: the 5-year expanding window, and the 5-year and 10-year rolling windows. Reported in
Forecasting Portfolio Returns

this table are the average annualized returns, t-statistics, terminal wealth (assuming investors start with $100 in 1954:04 and reinvest portfolio income every month until the end of the sample),
Sharpe ratio, and Fama-French three-factor model alphas. The notations are as follows: Max1 is the realized return from holding the portfolio with the highest expected return, Min1 is the realized
return from selling the portfolio with the lowest expected return, Max1-Min1 is the realized return from the zero-investment portfolio (difference between long and short positions), Max1–mkt is the
realized excess return of long position over market excess return, Min1–mkt is the realized excess return of short position over market excess return, Max3 is the realized return from holding three
portfolios with the highest expected returns, Min3 is the realized return from selling three portfolios with the lowest expected returns, and Max3–Min3, Max3–mkt, and Min3–mkt are defined the
same way as Max1–Min1, Max1–mkt, and Min1–mkt, respectively.
121
122 The Journal of Financial Research

a consistently significant performance of the Fama-French three-factor model in

predicting the excess returns of the 17 industry portfolios.
For the strategy of Max1–Min1 (Panel A in both Tables 4 and 5), the av-
erage annualized return of Max1 for the fixed model and the adjusted R2 model
over all three in-sample windows is about 13%, which is positive and significant,
and the zero-investment profit (Max1–Min1) is positive and significant with an
average 10% annualized return. The out-of-sample result shows that with the help
of some structural variables, the Fama-French (1993) three-factor model can pre-
dict asset returns and correctly distinguish the portfolios with different expected
returns. Moreover, Max1–mkt is positive and significant. The average annualized
return of Max1–mkt is about 8%, and both the terminal wealth and the Sharpe ratio
indicate that holding the portfolio with the highest expected returns outperforms
the benchmark of buying and holding the market portfolio.
As a robustness check, Panel B in both Tables 4 and 5 presents the result for
the strategy of Max3–Min3. Although the returns of Max3–Min3 are not as high as
those of Max1–Min1, they are positive and significant, with an average annualized
return of 7%, and the returns of Max3–mkt from all the strategies are positive and
significant. Therefore, the results from Max3–Min3 provide additional support for
the performance of the Fama-French (1993) three-factor model with risk premiums
estimated by structural variables.
A comparison between Table 3 and Tables 4 and 5 shows that using fore-
casted premiums performs better than using realized values of the premiums. In
Table 3, Max1–mkt and Max3–mkt are not different from zero; neither are Min1–
mkt and Min3–mkt. In contrast, in Tables 4 and 5, Max1–mkt and Max3–mkt are
positive and significant, and Min1–mkt and Min3–mkt are not statistically different
from zero.
In addition, I conduct formal statistical tests to compare the common prac-
tice with the proposed estimation procedures. I directly compare the out-of-sample
realized returns (Max1–Min1 and Max3–Min3) from the common practice with
those from the five estimation procedures: the fixed model and the four models
with different criteria. For the 5-year expanding window and the 10-year rolling
window, all five estimation procedures outperform the common practice. For in-
stance, from the 5-year expanding window, the difference of the realized annual
returns of Max1–Min1 between the fixed model and the common practice is 10.64%
(t = 2.54). I generalize that the difference in annualized returns is statistically
significant, with about 10% for Max1–Min1 and about 6% for Max3–Min3. Al-
though the results for the 5-year rolling window are not as significant as both
the 5-year expanding window and the 10-year rolling window, the difference re-
turns is still positive and the fixed model still performs better than the common
practice.
This analysis shows that by using structural variables, the estimated risk
premiums—MKTEXRT, SMB, and HML—do a better job in forecasting
 Forecasting Portfolio Returns 123

industry-based portfolios than do the realized values of those premiums, which

presumably contain some noisy information that dilutes their forecasting ability.

Transaction Costs
In reality, transaction costs are incurred in trading assets. Previous studies such as
those by Pesaran and Timmermann (1995), Carhart (1997), and Grundy and Martin
(2001) show that transaction costs can be substantial when portfolios are switched
monthly. Therefore, it is important to analyze the effect of transaction costs for a
more reasonable assessment of stock market predictability.
Following Pesaran and Timmermann (1995), I investigate transaction costs
of 0.5% and 1% per switch. A switch involves selling one portfolio and purchasing
another. Following Balvers and Wu (2004), for Max1–Min1, Max3, and Max3–
Min3, I count each switch of one of the two, three, or six portfolios as 1/2, 1/3, and
1/6 of a switch, respectively.
The higher the percentage of the switched portfolio is, the higher are the
transaction costs, and the lower are the realized profits. In fact, for Max1–Min1, the
average percentage of switch is 32% for all five estimation procedures across all
three in-sample windows. Formal out-of-sample tests show that with a transaction
cost of 0.5%, Max1–Min1 is no longer consistently positive and significant. For
example, with the adjusted R2 as the model-selection criterion, only the 5-year
expanding window and the 10-year rolling window generate positive returns of
Max1–Min1. Max3–Min3 is not significantly different from zero for all trading
strategies.
With a transaction cost of 1% per switch, Max1–Min1 and Max3–Min3
become even smaller. Max1–Min1 is not statistically positive, and Max3–Min3
becomes negative and significant for a few cases, such as the 5-year rolling window
using adjusted R2 as the criterion and the 5- and 10-year rolling windows using AIC
as the criterion. The results indicate that a reasonable amount of transaction costs
could wipe out a significant portion of the realized excess returns generated by the
trading strategies.
The results also imply that once transaction costs are taken into consid-
eration, it is not optimal to switch portfolios whenever the expected return of one
portfolio exceeds that of the currently held portfolio. Therefore, I impose a thresh-
old on the expected return differential; that is, if the portfolio with the highest
expected return is different from the currently held portfolio, a switch will happen
only when the expected return differential between these two portfolios reaches a
minimum level. For a 0.5% transaction cost, the expected return differential for a
switch is at least 0.5%, and for a 1% transaction cost, the differential is at least
1%.
The results show that for the 0.5% transaction cost per switch, the fixed
model and the model selected by adjusted R2 , AIC, and sign generate significant
124 The Journal of Financial Research

positive zero-investment profit (Max1–Min1) across all three in-sample windows.4

However, BIC does not perform as well. Only the five-year rolling window produces
a significant positive Max1–Min1 profit. For instance, with the five-year expanding
window, the fixed model produces an annualized return of 10.31% (t = 2.902) and
the adjusted R2 generates 7.64% (t = 2.264). These results are similar to those in
Peseran and Timmerman (1995).
For the 1% transaction cost per switch, although all four model-selection
criteria generate nonsignificant zero-investment profit (Max1–Min1), the fixed
model still demonstrates strong and statistically significant performance, showing
an average annualized return of 7%. For Max3–Min3 for both low and high trans-
action costs, none of the four criteria generates significant positive zero-investment
profits. However, the fixed model performs better than all the criteria. Only the
10-year rolling window fails to generate significant positive Max1–Min1 returns.
In addition, both the terminal wealth and Sharpe ratio show that the trading strategies
of Max1 and Max3 outperform the buy-and-hold strategy of the market portfolio.
I also perform formal tests to compare the common practice with the es-
timation procedures when transaction costs are present. The results show that the
estimation procedures outperform the common practice for most trading strategies,
although the results for high transaction costs are less significant than are those for
low transaction costs. In general, even after considering transaction costs, with a
reasonable prerequisite on the switch of portfolios, the Fama-French (1993) three-
factor model with forecast risk premiums still exhibit economic significance in
predicting asset returns.

Risk Analysis
The preceding analysis provides evidence on the predictability of asset returns.
Whether the predictability of asset returns indicates market inefficiency depends
on whether the excess returns can be explained by some risk factors. I use the
CAPM and the Fama-French (1993) three-factor model to identify whether the
excess returns are related to some standard risk factors. I treat the excess returns of
Max1–Min1 (Max3–Min3), Max1–mkt (Max3–mkt), and Min1–mkt (Min3–mkt) as
the dependent variables in the following two regressions:

erit = αi + bi (rmt − rft ) + eit , (5)

erit = αi + bi (rmt − rft ) + si rSMB,t + h i rHML,t + eit , (6)

where erit refers to the excess return of one trading strategy at time t. I test excess
returns from six trading strategies. These tests include Max1–Min1 (Max3–Min3),

4
Because of limited space, I do not include the results from formal tests.
 Forecasting Portfolio Returns 125

Max1–mkt (Max3–mkt), and Min1–mkt (Min3–mkt). The nonzero value of α i indi-

cates whether an asset is overpriced or underpriced.
The results show that the α of Max1–Min1 is positive and significant except
for the 5-year rolling window using R2 and sign as selection criteria. The α of
Max1–mkt is also positive and significant except for a few cases in which the
Fama-French (1993) three-factor model α is not statistically different from zero,
such as the 10-year expanding window with model-selection criteria of R2 , AIC, and
sign.
In general, the results provide evidence that even with a risk adjustment,
the mean return of the long portfolio (Max1) is still significantly larger than that of
the buy-and-hold benchmark, the market portfolio. For some cases of Min1–mkt, the
value of α is negative and significant. However, for other cases, it is not statistically
different from zero. Therefore, the evidence that the short portfolio performs worse
than the benchmark is not as strong as the evidence that the long portfolio performs
better than the benchmark. Although showing less statistical significance, Max3–
Min3, Max3–mkt, and Min3–mkt exhibit a similar pattern in the CAPM α and the
Fama-French (1993) three-factor model α.

Bossaerts and Hillion’s (1999) Forecast Beta

The analysis of the forecasting ability of the Fama-French (1993) three-factor model
indicates that the structural variables capture the variations in risk premiums and
that using forecasted risk premiums results in better forecasts of asset returns
than does using average historical risk premiums. I next examine the compari-
son between realized and expected excess returns. This comparison provides addi-
tional information on how well the Fama-French three-factor model forecasts asset
returns.
The method I apply is the Bossaerts and Hillion’s (1999) forecast beta.
They investigate the forecasting ability (external validity) of a model by projecting
the actual excess returns onto the expected excess returns:

erit+1 = ai + bi z it + v t+1 , i = 1, 2, . . . , 17, (7)

where zit is the expected excess return of industry portfolio i at time t and is
calculated using the Fama-French (1993) three-factor model with risk premiums
forecasted by structural variables.
After examining all 17 industry portfolios, I find that for all four model-
selection criteria and the fixed model, at least 10 of 17 forecast betas are positive
and significant with the five-year expanding window, and that the actual ratio is
between 10/17 and 16/17. With the two rolling windows, the ratio is between 2/17
and 10/17, and in most cases, the ratio is about 7/17. As a result, I find that all five
estimation procedures perform well out of sample.
126 The Journal of Financial Research

V. Conclusion

The common practice in the model-based estimation of the equity cost of capital
uses the average historical factor premiums as estimates of future factor premiums.
However, much evidence shows that the common practice generates inaccurate
estimates.
I propose using structural variables to estimate risk premiums and imple-
ment trading strategies on 17 equally weighted industry portfolios to evaluate the
forecasting ability of the Fama-French (1993) three-factor model. Assuming that
investors do not possess foreknowledge of what model they should use to estimate
risk factors, I apply four model-selection criteria to choose an optimal model for
every point in the trading strategy.
The empirical results show that based on the common practice, all trading
strategies generate nonsignificant zero-investment profits, but with the proposed
estimation procedures, most trading strategies generate significant positive zero-
investment profits. When transaction costs are considered, the zero-investment
profits are materially reduced. However, an adjustment on trading strategies with a
transaction cost threshold can still prevent the profits from being completely eroded
by transaction costs, at least for low transaction costs.
A risk analysis shows that the returns generated by the trading strategies
cannot be explained by the CAPM or the Fama-French (1993) three-factor model.
Finally, to determine how well the Fama-French model forecasts asset returns, I
calculate Bossaerts and Hillion’s (1999) forecast betas. For the five-year expanding
window, the forecast betas are positive and significant for more than half of the
17 industry portfolios.
All results demonstrate that the Fama-French (1993) three-factor model
with factor premiums estimated from structural variables is more reliable than the
common practice in forecasting portfolio returns.

References

Balvers, R. J. and Y. R. Wu, 2004, Momentum and mean reversion across national equity markets, Working
paper.
Bossaerts, P. and P. Hillion, 1999, Implementing statistical criterion to select return forecasting models:
What do we learn? Review of Financial Studies 12, 405–28.
Carhart, M. M., 1997, On persistence in mutual fund performance, Journal of Finance 52, 57–82.
Cooper, M., H. Gulen, and M. Vassalou, 2001, Investing in size and book-to-market portfolio using infor-
mation about the macroeconomy: Some new trading rules, Working paper.
Elton, E. J., 1999, Expected return, realized return, and asset pricing tests, Journal of Finance 54, 1199–1220.
Fama, E. F. and K. R. French, 1993, Common risk factors in the returns on stocks and bonds, Journal of
Financial Economics 33, 3–56.
Fama, E. F. and K. R. French, 1997, Industry costs of equity, Journal of Financial Economics 43, 153–
93.
 Forecasting Portfolio Returns 127

Ferson, W. E. and C. R. Harvey, 1999, Conditioning variables and cross-section of stock returns, Journal
of Finance 54, 1325–60.
Ferson, W. E. and D. H. Locke, 1998, Cost of capital estimation without CAPM: Analysis of sources of
error, Management Science 44, 485–500.
Grundy, B. and J. S. Martin, 2001, Understanding the nature and the risks and the sources of the rewards to
momentum investing, Review of Financial Studies 14, 29–78.
Moskowitz, T. J. and M. Grinblatt, 1999, Do industries explain momentum? Journal of Finance, 1249–90.
Pastor, L. and R. F. Stambaugh, 1999, Costs of equity capital and model mispricing, Journal of Finance 54,
67–121.
Pesaran, M. H. and A. Timmermann, 1995, Predictability of stock returns: Robustness and economic
significance, Journal of Finance 50, 1201–28.