Campbell 2007 Equity Returns

Predicting Excess Stock Returns Out of
Sample: Can Anything Beat the Historical

Average?
John Y. Campbell
Department of Economics, Harvard University and NBER
Goyal and Welch (2007) argue that the historical average excess stock return forecasts
future excess stock returns better than regressions of excess returns on predictor variables.
In this article, we show that many predictive regressions beat the historical average return,
once weak restrictions are imposed on the signs of coefficients and return forecasts. The
out-of-sample explanatory power is small, but nonetheless is economically meaningful for
mean-variance investors. Even better results can be obtained by imposing the restrictions
of steady-state valuation models, thereby removing the need to estimate the average from
a short sample of volatile stock returns. (JEL G10, G11)
Towards the end of the last century, academic finance economists came to
take seriously the view that aggregate stock returns are predictable. During the
1980s, a number of papers studied valuation ratios, such as the dividend-price
ratio, earnings-price ratio, or smoothed earnings-price ratio. Value-oriented
investors in the tradition of Graham and Dodd (1934) had always asserted
that high valuation ratios are an indication of an undervalued stock market
and should predict high subsequent returns, but these ideas did not carry much
weight in the academic literature until authors such as Rozeff (1984), Fama and
French (1988), and Campbell and Shiller (1988a, 1988b) found that valuation
ratios are positively correlated with subsequent returns and that the implied
predictability of returns is substantial at longer horizons. Around the same
time, several papers pointed out that yields on short- and long-term treasury
and corporate bonds are correlated with subsequent stock returns (Fama and
Schwert, 1977; Keim and Stambaugh, 1986; Campbell, 1987; Fama and French,
1989).
The authors are grateful to Jan Szilagyi for able research assistance, to Amit Goyal and Ivo Welch for sharing
their data, and to Malcolm Baker, Lutz Kilian, Martin Lettau, Sydney Ludvigson, Rossen Valkanov, the editor,
and two anonymous referees for helpful comments on an earlier draft entitled Predicting the Equity Premium
Out of Sample: Can Anything Beat the Historical Average? This material is based upon the work supported
by the National Science Foundation under Grant No. 0214061 to Campbell. Address correspondence to John Y.
Campbell, Department of Economics, Littauer Center, Harvard University, Cambridge, MA 02138, telephone:
(617) 496-6448, e-mail: john [email protected].

C The Author 2007. Published by Oxford University Press on behalf of The Society for Financial Studies. All
rights reserved. For Permissions, please email: [email protected].
doi:10.1093/rfs/hhm055
Advance Access publication November 20, 2007
Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015
Samuel B. Thompson
Arrowstreet Capital, LP
The Review of Financial Studies / v 21 n 4 2008
1510
During the 1990s and early 2000s, research continued on the prediction of
stock returns from valuation ratios (Kothari and Shanken, 1997; Pontiff and
Schall, 1998) and interest rates (Hodrick, 1992). Several papers suggested new
predictor variables exploiting information in corporate payout and financing
activity (Lamont, 1998; Baker and Wurgler, 2000), the level of consumption
in relation to wealth (Lettau and Ludvigson 2001), and the relative valuations
of high- and low-beta stocks (Polk et al., 2006). At the same time, several
authors expressed concern that the apparent predictability of stock returns
might be spurious. Many of the predictor variables in the literature are highly
persistent: Nelson and Kim (1993) and Stambaugh (1999) pointed out that
persistence leads to biased coefficients in predictive regressions if innovations
in the predictor variable are correlated with returns (as is strongly the case for
valuation ratios, although not for interest rates). Under the same conditions,
the standard t-test for predictability has incorrect size (Cavanagh et al., 1995).
These problems are exacerbated if researchers are data mining, considering
large numbers of variables, and reporting only those results that are apparently
statistically significant (Foster et al., 1997; Ferson et al., 2003). An active recent
literature discusses alternative econometric methods for correcting the bias and
conducting valid inference (Cavanagh et al., 1995; Mark, 1995; Kilian, 1999;
Lewellen, 2004; Torous et al., 2004; Campbell and Yogo, 2006; Jansson and
Moreira, 2006; Polk et al., 2006; Ang and Bekaert, 2007).
A somewhat different critique emphasizes that predictive regressions have
often performed poorly out-of-sample (Goyal and Welch 2003, 2007; Butler et
al., 2005). This critique had a particular force during the bull market of the late
1990s, when low valuation ratios predicted extraordinarily low stock returns
that did not materialize until the early 2000s (Campbell and Shiller, 1998).
Goyal and Welch (2007) argue that the poor out-of-sample performance of
predictive regressions is a systemic problem, not confined to any one decade.
They compare predictive regressions with historical average returns and find
that historical average returns almost always generate superior return forecasts.
They conclude that the profession has yet to find some variable that has
meaningful and robust empirical equity premium forecasting power.
While it is not clear how much weight should be placed on out-of-sample
statistics in judging the predictability of stock returns (Inoue and Kilian, 2004),
in this article we take up Goyal and Welchs (2007) challenge and ask whether
standard variables could have been used in real time to forecast twentiethand early twenty-first-century stock returns. We show that simple restrictions
on predictive regressions, suggested by investment theory, improve the out-ofsample performance of key forecasting variables and imply that investors could
have profited by using market timing strategies.
We begin in Section 2 by comparing the in-sample and out-of-sample
forecasting power of standard predictor variables through the end of 2005. We
use at least 20 years of data to obtain initial coefficient estimates and restrict
the forecast evaluation period to the period since 1927 when high-quality
Predicting Excess Stock Returns Out of Sample
1. Theoretical Restrictions on Predictive Regressions

In this section, we conduct an out-of-sample forecasting exercise inspired by
Goyal and Welch (2007), with modifications that reveal the effectiveness of
theoretically motivated restrictions. We use monthly data and predict simple
monthly or annual stock returns on the S&P 500 Index. This immediately creates a tradeoff between the length of the data sample and the quality of the
available data. High-quality total return data are available monthly from CRSP
since 1927, while total monthly returns before that time are constructed by
interpolation of lower-frequency dividend payments and therefore may be suspect. Accordingly, we use the CRSP data period as our out-of-sample forecast
evaluation period, but use earlier data to estimate an initial regression.
Table 1, whose format is based on the tables in Goyal and Welch (2007),
reports the results. We begin by discussing monthly return forecasts reported
in Panel A, and then discuss annual forecasts reported in Panel B. Each
row of the table considers a different forecasting variable. The first four
rows consider valuation ratios: the dividend-price ratio, earnings-price ratio,
smoothed earnings-price ratio, and book-to-market ratio. Each of these ratios
1511
total returns data are available from CRSP. We calculate an out-of-sample

R 2 statistic that can be compared with the usual in-sample R 2 statistic. Like
Goyal and Welch (2007), we find unimpressive in-sample results and poor
out-of-sample performance for many of the usual linear regressions.
Our contribution is to show that restricted regressions perform considerably
better than these unrestricted regressions. We begin by considering two alternative restrictions that can be imposed on any theoretically motivated forecasting
regression: first, the regression coefficient has the theoretically expected sign;
and second, the fitted value of the equity premium is positive. We impose these
restrictions sequentially and then together, and find that they substantially improve the out-of-sample performance of predictive regressions.
Next, we look more closely at valuation ratios, for which theory restricts
not only the sign but also the magnitude of the regression coefficients. We find
that out-of-sample predictability can be improved even more by restricting the
magnitude of the coefficients to the values implied by a simple steady-state
model, such as the Gordon (1962) growth model in the case of the dividendprice ratio.
Section 1 shows that several commonly used forecasting variables do have
some ability to predict stock returns out-of-sample. The out-of-sample R 2
statistics are positive, but small. This raises the question of whether the predictive power is economically meaningful. In Section 2, we show that even very
small R 2 statistics are relevant for investors because they can generate large
improvements in portfolio performance. In a related exercise, we calculate the
fees that investors would be willing to pay to exploit the information in each of
our forecasting variables. Section 3 briefly concludes.
1872m2
1872m2
1881m2
1926m6
1936m6
1920m1
1870m1
1920m1
1919m1
1871m5
1927m12
1951m12
Dividend-price ratio
Earnings-price ratio
Smooth earnings-price ratio
Book-to-market
ROE
T-Bill rate
Long-term yield
Term spread
Default spread
Inflation
Net equity issuance
Consumption-wealth ratio
1927m1
1927m1
1927m1
1946m6
1956m6
1940m1
1927m1
1940m1
1939m1
1927m1
1947m12
1971m12
1927m1
1927m1
1927m1
1946m6
1956m6
1940m1
1927m1
1940m1
1939m1
1927m1
1947m12
1971m12
2.69
2.84
3.01
1.98
0.35
1.77
0.91
1.72
0.07
0.17
0.54
3.76
1.25
2.29
1.85
1.96
0.36
2.44
1.46
2.16
0.74
0.39
1.74
4.57
In-Sample
t-statistic
0.65%
0.12
0.33
0.43
0.93
0.52
0.19
0.46
0.19
0.22
0.34
1.36
5.53
4.93
7.89
3.38
8.60
5.54
0.15
4.79
3.81
0.71
4.27
7.75
B: Annual Returns
10.8
6.78
13.57
8.26
0.32
4.26
0.77
3.10
0.01
0.07
0.35
19.87
Unconstrained
A: Monthly Returns
1.13%
0.71
1.36
0.61
0.02
0.86
0.19
0.65
0.10
0.06
0.48
2.60
In-Sample
R-squared
5.53
4.93
7.89
3.38
0.03
5.54
0.15
4.79
3.81
0.71
4.27
7.75
0.05%
0.18
0.42
0.43
0.06
0.51
0.19
0.47
0.19
0.21
0.34
1.36
Positive Slope
Both
0.08%
0.18
0.43
0.00
0.06
0.55
0.20
0.46
0.19
0.17
0.50
0.27
5.63
4.94
7.85
1.39
0.03
7.47
2.26
4.74
3.81
0.71
2.38
1.48
Pos. Forecast
0.07%
0.14
0.38
0.00
0.93
0.57
0.20
0.45
0.19
0.18
0.50
0.27
5.63
4.94
7.85
1.39
8.35
7.47
2.26
4.74
3.81
0.71
2.38
1.48
Out-of-Sample R-squared with Different Constraints
This table presents statistics on forecast errors for stock returns. We use S&P 500 total returns (dividend included) where data prior to January 1927 was obtained from Robert Shillers Web site. Sample Begin denotes when the predictor
was first available. All statistics are for the period that starts at Forecast Begin and ends on December 31, 2005 (for monthly forecasts), or December 31, 2004 (for annual forecasts). The Unconstrained out-of-sample R-squared compares
the forecast error of the historical mean versus the forecast from unconstrained ordinary least squares. Positive Slope introduces the restriction that the coefficient on the predictor must be of the correct sign, otherwise the historical mean
is used as a predictor instead. Pos. Forecast requires that the prediction be positive, otherwise we use zero as the forecast. Both indicates that we impose both restrictions. The in-sample t-statistic and R-squared both come from the last
regression (i.e., over the whole sample from the Sample Begin date to December 2005 for monthly data and December 2004 for annual data). The In-Sample t-statistic is heteroskedasticity robust. The Annual forecasts in Panel B are
based on 12 overlapping annual returns per year. The t-statistics in Panel B are corrected to take into account correlation induced by the overlapping nature of the dependent variable. For the consumption-wealth ratio, we report the robust
F-statistic that all the slope coefficients are zero.
1872m2
1872m2
1881m2
1926m6
1936m6
1920m1
1870m1
1920m1
1919m1
1871m5
1927m12
1951m12
Dividend-price ratio
Earnings-price ratio
Smooth earnings-price ratio
Book-to-market
ROE
T-Bill rate
Long-term yield
Term spread
Default spread
Inflation
Net equity issuance
Consumption-wealth ratio
Forecast
Begin
1512
Sample
Begin
Table 1
Excess return prediction with regression constraints
Goyal and Welch (2007) consider these variables, along with the ratio of lagged dividends to lagged prices (the
dividend yield in Goyal and Welchs terminology). We drop this variable as there is no reason to believe that
it should be a better predictor than the ratio of lagged dividends to current prices (the dividendprice ratio).
If earnings obey the clean-surplus relation, the growth rate of real book equity equals real ROE minus the
payout ratio. ROE can alternatively be measured from the growth rate of book equity rather than from reported
accounting earnings, as in Polk et al. (2006).
In 2003 the BEA revised the definitions of several variables used by Lettau and Ludvigson (2001) to construct
their series. We use the updated data available on Martin Lettaus Web site, not the original series used in Lettau
and Ludvigson (2001). Data revisions raise the deeper problem that the series may not have been available to
investors in real time, but we do not try to deal with this issue here.
1513
is measured in levels, rather than logs, with some accounting measure of

corporate value in the numerator, and market value in the denominator. The
smoothed earnings-price ratio, proposed by Campbell and Shiller (1988b,
1998) is the ratio of a 10-year moving average of real earnings to current real
prices. Campbell and Shiller argue that this ratio should have better forecasting
power than the current earnings-price ratio because aggregate corporate
earnings display short-run cyclical noise; in particular, earnings drop close to
zero in recession years, such as 1934 and 1992, creating spikes in the current
earnings-price ratio that have nothing to do with stock market valuation levels.1
The next row of Table 1 uses a smoothed measure of accounting real return
on equity (ROE) as the predictor variable. Real ROE is measured as the level of
real earnings divided by lagged real book equity and the gross one-year inflation
rate. This gives a measure of the total resources available to be divided between
real payouts and real growth in book equity.2 We smooth real ROE over 10
years to remove short-run cyclical noise in corporate profitability. We include
this variable in Table 1 not because it has been claimed to predict aggregate
stock returns by itself, but because we will use it in combination with valuation
ratios in Table 2.
The next five rows of Table 1 consider nominal interest rates and inflation:
the short-term interest rate, the long-term bond yield, the term spread between
long- and short-term treasury yields, the default spread between corporate and
treasury bond yields, and the lagged rate of inflation.
The last two rows of Table 1 evaluate forecasting variables that have been
proposed more recently: the equity share of new issues proposed by Baker
and Wurgler (2000) and the consumption-wealth ratio of Lettau and Ludvigson
(2001). Lettau and Ludvigsons variable is based on a cointegrating relationship
between consumption, aggregate labor income, and aggregate financial wealth.
Rather than estimate a separate cointegrating regression, we simply include the
three variables directly in the forecasting equation for stock returns.3
The first column of Table 1 reports the first date for which we have data
on each forecasting variable. For dividends, earnings, stock returns, long-term
bond yields, and inflation we have data, originally assembled by Robert Shiller,
back to the early 1870s. Other data series typically begin shortly after the end of
World War I, although book equity does not become available until 1926, which
means that the book-to-market ratio begins then and the 10-year smoothed ROE
1872m2
1872m2
1881m2
1891m2
1892m2
1892m2
1936m6
1891m5
1892m2
1892m2
1936m6
Dividend/price
Earnings/price
Smooth earnings/price
Dividend/price + growth
Earnings/price + growth
Smooth earnings/price + growth
Book-to-market + growth
Dividend/price + growth real rate
Earnings/price + growth real rate
Smooth earnings/price + growth real rate
Book-to-market + growth real rate
1927m1
1927m1
1927m1
1927m1
1927m1
1927m1
1956m6
1927m1
1927m1
1927m1
1956m6
1927m1
1927m1
1927m1
1927m1
1927m1
1927m1
1956m6
1927m1
1927m1
1927m1
1956m6
In-Sample
R-squared
2.69
2.84
3.01
1.77
1.42
1.75
1.97
1.46
1.13
1.53
2.03
B: Annual Returns
10.89
6.78
13.57
9.30
4.44
10.45
5.45
7.69
3.27
7.90
5.77
A: Monthly Returns
1.25
1.12%
2.28
0.71
1.85
1.35
1.40
1.03
1.82
0.49
2.00
1.10
1.61
0.33
1.47
0.86
1.53
0.36
1.97
0.84
1.68
0.36
In-Sample
t-statistic
5.53
4.93
7.89
2.49
1.69
3.16
3.53
2.87
2.01
3.35
1.73
0.66%
0.12
0.32
0.05
0.05
0.11
0.35
0.02
0.00
0.15
0.45
Unconstrained
5.63
4.94
7.85
2.99
2.11
3.33
0.64
3.24
2.05
3.35
1.12
0.08%
0.18
0.43
0.20
0.08
0.25
0.34
0.21
0.12
0.26
0.45
Positive Slope,
Pos. Forecast
Fixed Coefs
0.42%
0.76
0.97
0.63
0.57
0.72
0.33
0.41
0.39
0.52
0.24
2.20
5.87
7.99
4.35
3.89
5.39
3.63
1.89
1.85
3.22
2.33
Pos. Intercept,
Bounded Slope
0.19%
0.25
0.43
0.17
0.07
0.21
0.34
0.18
0.09
0.23
0.42
3.76
4.34
6.44
2.67
1.80
3.23
2.39
2.95
2.04
3.38
1.82
Out-of-Sample R-squared with Different Constraints
The table presents forecast statistics for value predictors under various constraints. The predictor label + growth indicates that we add an earnings growth forecast to the
predictor. The predictor label real rate indicates that we subtract a forecast of the real risk-free rate from the predictor. See the text for details. The In-Sample statistics are
defined as in Table 1. The Unconstrained and Positive Slope, Pos. Forecast columns are described in Table 1. Pos. Intercept, Bounded Slope indicates that we constrain
the intercept to be positive and the slope to be between zero and one. Fixed Coefs indicates that we fix the intercept at zero and the slope at one.
1872m2
1872m2
1881m2
1891m2
1892m2
1892m2
1936m6
1891m5
1892m2
1892m2
1936m6
Dividend/price
Earnings/price
Forecast
Begin
1514
Sample
Begin
Table 2
Excess return prediction with valuation constraints
where
rt is the fitted value from a predictive regression estimated through period
t 1, and r t is the historical average return estimated through period t 1.
If the out-of-sample R 2 is positive, then the predictive regression has lower
average mean-squared prediction error than the historical average return.5 We
use the entire available history of stock returns, back to 1871, to estimate the
historical average return. This gives the historical mean an advantage over
predictive regressions with variables that have become available more recently,
because more data are available to estimate the historical mean than to estimate
such predictive regressions. However, this is a real-world advantage of the
historical mean that should be taken into account in our tests.
The out-of-sample performance of the predictor variables is mixed. The
fifth column of Table 1 shows that only two out of the four valuation ratios
(the earnings yield and smoothed earnings yield) and two out of the five
interest-rate variables (the treasury bill rate and term spread) deliver positive
out-of-sample R 2 statistics. The interest rate results are consistent with the
4
The adjustment of the R 2 statistic for degrees of freedom makes only a very small difference in samples of the
size used here. The adjustment is about 5 basis points for a regression starting in 1871, and about 10 basis points
for a regression starting in 1927.
Clark and West (2005) point out that if the return series is truly unpredictable, then in a finite sample the
predictive regression will on average have a higher mean squared prediction error because it must estimate an
additional coefficient. Thus, the expected out-of-sample R 2 under the null of unpredictability is negative, and
a zero out-of-sample R 2 can be interpreted as weak evidence for predictability. We do not pursue this point
here because, like Goyal and Welch (2007), we ask whether predictive regressions or historical average return
forecasts have delivered better out-of-sample forecasts, not whether stock returns are truly predictable.
1515
begins in 1936. All data series continue to the end of 2005. The second column
reports the date at which we begin the out-of-sample forecast evaluation. This
is the beginning of 1927, when accurate data on total monthly stock returns
become available from CRSP, or 20 years after the date in column 1, whichever
comes later.
The third and fourth columns of Table 1 report the full-sample t-statistic for
the significance of each variable in forecasting stock returns, and the adjusted
R 2 statistic of the full-sample regression.4 It is immediately obvious from
the column of t-statistics that many of the valuation ratios and interest-rate
variables are statistically insignificant in predicting stock returns over the long
sample periods considered here. The most successful variables are the earnings
yield, the treasury bill rate, and the term spread. Of the two recently proposed
variables, net equity issuance is modestly successful and the consumptionwealth ratio is strikingly successful in-sample.
The remaining columns of Table 1 evaluate the out-of-sample performance
of these forecasts, using an out-of-sample R 2 statistic that can be compared
with the in-sample R 2 statistic. This is computed as
T
(rt
r t )2
2
ROS = 1 Tt=1
,
(1)
2
t=1 (r t r t )
1.1 Sign restrictions on slope coefficients and forecasts

Despite the mixed performance of unrestricted predictive regressions out of
sample, it is premature to conclude with Goyal and Welch (2007) that predictive
regressions cannot profitably be used by investors in real time. A regression
estimated over a short sample period can easily generate perverse results, such
as a negative coefficient when theory suggests that the coefficient should be
positive. Since out-of-sample forecast evaluation begins as little as 20 years
after a predictor variable first becomes available, while the historical mean
return is estimated from the beginning of our data set, this can be an important
problem in practice. For example, in the early 1930s the earnings-price ratio
was very high, but the coefficient on the predictor was estimated to be negative.
This led to a negative forecast of the equity premium in the early 1930s and a
subsequent poor forecast performance. In practice, an investor would not use
a perverse coefficient but would likely conclude that the coefficient is zero, in
effect imposing prior knowledge on the output of the regression.
In the remaining columns of Table 1, we explore the impact of imposing
sensible restrictions on the out-of-sample forecasting exercise. In the sixth
column, we set the regression coefficient to zero whenever it has the wrong
sign (different from the theoretically expected sign estimated over the full
sample). In the seventh column, we assume that investors rule out a negative
equity premium, and set the forecast to zero whenever it is negative. In the
eighth and final columns, we impose first the sign restriction on the coefficient,
and then the sign restriction on the forecast.
These restrictions never worsen and almost always improve the out-ofsample performance of our predictive regressions. With no restrictions, as
6
Earlier drafts of this article reported better out-of-sample performance for this series. BEA data revisions in 2003
and the addition of recent years to the forecast evaluation period are responsible for the change in results.
1516
conclusion of Ang and Bekaert (2007) that the treasury bill rate and term
spread are robust return predictors. The performance of these variables would
be stronger if we started the sample period later, because the interest rate
process changed dramatically at the time of the Federal Reserve-Treasury
Accord in 1951. Of the two recently proposed variables, net issuance performs
reasonably well but the consumption-wealth ratio does not. The difficulty of
estimating coefficients in a short sample is particularly severe for this series
because it includes three separately estimated components.6
All the regressions we have reported predict simple stock returns rather
than log stock returns. The use of simple returns makes little difference to the
comparison of predictive regressions with historical mean forecasts, but all
forecasts tend to underpredict returns when log returns are used. The reason is
that high stock market volatility in the 1920s and 1930s depressed log returns
relative to simple returns in this period. Thus, the gap between average stock
returns in the late twentieth century and the early twentieth century is greater
in logs than in levels.
1.2 Quantitative coefficient restrictions for valuation ratios

So far we have used theory only to impose weak sign restrictions on coefficients
and excess return forecasts. We now go further and impose the restrictions of a
simple steady-state model on the coefficients for valuation ratios. This approach
can be interpreted as a conditional version of the procedure advocated by
Fama and French (2002) to estimate the unconditional average mean return.
It can also be interpreted as a way to combine the information from multiple
data sources, for example, on market valuations, corporate profitability, and
payout policy, without increasing the number of free coefficients that must be
estimated.
We start from the Gordon (1962) growth formula,
D
= R G,
P
(2)
1517
we noted above, only earnings-based valuation ratios have a positive out-ofsample R 2 , but the slope restriction delivers a positive out-of-sample R 2 for the
dividend yield and the sign restriction brings the out-of-sample R 2 close to zero
for the book-to-market ratio. The sign restriction also delivers a positive outof-sample R 2 for the long-term bond yield and the consumption-wealth ratio.
Figure 1 illustrates the effect of the restrictions for the dividend-price ratio.
The top panel shows annualized excess return forecasts based on the fullsample OLS regression coefficient, the rolling (out-of-sample) OLS regression coefficient without restrictions, and the out-of-sample coefficient with both
coefficient and forecast sign restrictions. The bottom panel shows the cumulative out-of-sample R 2 for these three forecasts. The coefficient sign restriction
significantly improves the forecasts in the 1930s, when the coefficient was estimated to be negative. The forecast restrictions bind periodically during the
1960s and 1990s. Valuation ratios were unusually low during these periods,
leading to unprecedentedly low forecasts. Campbell and Shiller (2001) also
noted the unusually low valuation ratios of the 1990s, and wrote We do not
find this extreme forecast credible; when the independent variable has moved
so far from the historically observed range, we cannot trust a linear regression line. Our forecast restrictions are a simple way to avoid such incredible
forecasts.
Panel B of Table 1 reports comparable results for annual regressions, estimated using overlapping monthly data. In-sample t-statistics are corrected for
serial correlation and are even lower than those reported in Panel A for monthly
data. Despite this weak in-sample predictive power, these regressions perform
quite well out-of-sample. When both our theoretical restrictions are imposed,
all four valuation ratios and the three variables based on the treasury yield curve
have out-of-sample R 2 statistics of at least 2%. Net equity issuance and the
consumption-wealth ratio also benefit from our restrictions but do not beat the
historical mean return at the annual frequency.
1518
Figure 1
Forecasting excess returns with the dividend yield
The top panel shows the historical annualized excess return forecasts for the historical mean and three different regression models, as labeled in the figure, where oos refers to out-ofsample and ols refers to ordinary least squares. The bottom panel shows the cumulative out-of-sample R-squared up to each point in the historical sample, for the three regression models
relative to the historical mean.
which describes the dividend-price ratio in a steady state with a constant discount rate and dividend growth rate. We combine this formula with the steadystate relation between growth and accounting return on equity,

D
ROE,
(3)
G = 1
E
The same approach can be used to obtain growth forecasts from the earnings
yield. Using the fact that D/P = (D/E)(E/P), we have

D
D E

+ 1
ROE,
(5)
RE P =
E P
E
a payout-ratio-weighted average of the earnings yield and the accounting return
on equity. When return on equity equals the expected return, as might be the
case in long-run equilibrium, then this implies that
R E P = E/P.
Finally, since E/P = (B/M)ROE, we have

D B

1 .
(6)
R B M = ROE 1 +
E M
To use these formulas in practice, one must decide how to combine historical
and contemporaneous data on the right-hand side variables. We follow Fama and
French (2002) by using historical average data on payouts and profitability, but
differ from them by using current rather than historical average data on valuation
ratios to obtain a return forecast conditional on the markets current valuation
level. This procedure assumes that movements in valuation ratios, relative to
historical cash flows, are explained by permanent changes in expected returns.
It is a compromise between the view that valuation ratios are driven by changing
forecasts of profitability, in which case the implied movements in returns would
be smaller, and the view that valuation ratios are driven by temporary changes
in discount rates, in which case the implied return movements would be larger
(Campbell and Shiller, 1988a).7
Table 2 reports regression forecasts using both the unadjusted valuation ratios
from Table 1, and two variants of growth-adjusted ratios. Our first set of growthadjusted ratios uses Equations (4)(6) with current data on the valuation ratios
and historical data on the payout ratio (an average from the beginning of the
7
Cochrane (2007) also emphasizes the importance of the fact that valuation ratios do not forecast growth rates
of cash flows. The ability of the theoretical models in this section to predict stock returns is consistent with
Cochranes results.
1519
where D/E is the payout ratio, to obtain a growth-adjusted return forecast:

D
D

RD P =
+ 1
ROE.
(4)
P
E
Before 1926, we do not have data on ROE and thus we cannot calculate 10-year smoothed ROE until 1936.
Before that date, we use real earnings growth to estimate G.
One could consider forecasting the short-term real interest rate using recent data on inflation and nominal interest
rates. There are two reasons not to pursue this approach. First, the volatility of inflation makes it difficult in
practice. Second, the steady-state growth model delivers a long-run forecast of stock returns that should be
matched to a long-run forecast of real bond returns, such as the TIPS yield or (since this is not available before
1997) the long-run historical average real interest rate. If the short-term real interest rate exceeds such a long-run
real rate forecast, then it is quite possible that the short-term real stock return shifts upwards in parallel, leaving
the excess stock return forecast unchanged.
10
A subtle but important issue is that we use the steady-state model to forecast the arithmetic average stock
return, and take arithmetic averages of ROE and other historical data. Some authors, such as Siegel (1994), take
geometric averages of historical data and forecast the geometric average stock return. These two approaches
are equivalent if the volatility of dividend growth and stock returns are the same, as implied by the steady-state
model, but are different in the data because stock returns are much more volatile than dividend growth and ROE.
A full exploration of the two approaches is beyond the scope of this article, but we note that the Siegel approach
would generate higher forecasts of arithmetic average stock returns, and thus would perform even better in the
late twentieth century than the approach we use here.
1520
sample) and the return on equity (a 10-year moving average as in Table 1).8 Our
second set of growth-adjusted ratios in addition subtracts the historical average
real interest rate from the beginning of the sample period in order to convert
a theoretical real return forecast into a theoretical excess return forecast. The
historical average real interest rate is extremely stable, so this final step is close
to an intercept adjustment.9
We use these ratios in four different ways. First, we report unrestricted regressions of returns on the ratios. Second, we report regressions restricted as in
Table 1 to have positive slope coefficients and positive return forecasts. Third,
given that valuation ratios are positive, we can obtain reasonable return forecasts merely by bounding the intercept above zero and the slope coefficient
to lie between zero and one. This is similar to the second approach but can
be implemented by coefficient restrictions rather than by restricting forecasts
directly. Finally, we impose the restrictions of the steady-state theory by restricting the intercept of the regression to be zero and the slope coefficient to be
one. That is, we estimate the return forecast directly from the data without using any historical information on the covariance between returns and valuation
ratios.10
Table 2 shows that the last and most restrictive approach delivers the best
out-of-sample performance in monthly data. The out-of-sample R 2 statistics
range from 0.66% to 0.32% when no restrictions are imposed, from 0.45
to 0.43% when the restrictions of Table 1 are imposed, and from 0.24% to
0.97% when the zero-intercept and unit-slope restrictions are imposed. These
restrictions improve the out-of-sample monthly forecasting power of every predictive regression we consider. The out-of-sample R 2 statistics are also reliably
positive in annual regressions that impose zero-intercept and unit-slope restrictions, ranging from 1.85% to 7.99%, but here the theoretical restrictions worsen
out-of-sample predictive power in a few cases. The most striking example is
the dividend-price ratio with no growth adjustment, where the zero-intercept
restriction is the least theoretically appealing as it effectively assumes zero real
growth in dividends.
2. How Large an R2 Should We Expect?

In the previous section, we showed that many of the forecasting variables that
have been discussed in the literature do have positive out-of-sample predictive power for aggregate U.S. stock returns, when reasonable restrictions are
1521
Goyal and Welch (2007) emphasize that out-of-sample performance has

been particularly poor for predictive regressions in the last few decades. To
investigate this issue, Table 3 repeats Table 2 for three subsamples: 1927
1956, 19561980, and 19802005. We choose the first subsample to end
20 years after we are first able to calculate 10-year smoothed ROE, so that
growth adjustments in the second and third subsamples are entirely based
on ROE and not on historical earnings growth. We omit the book-to-market
ratio in the first subsample, as the series does not become available until
1926 and thus we cannot estimate coefficients before the first subsample
begins.
As Goyal and Welch (2007) have noted, the ability of valuation ratios to
forecast stock returns is strongest in the first subsample, which includes the
Great Depression and World War II, slightly weaker in the second subsample,
which includes the postwar boom and the oil shocks of the 1970s, and weakest
in the third subsample, which includes the great equity bull market at the end
of the twentieth century. Even in the third subsample, however, earnings-based
valuation ratios outperform the historical mean return in forecasting future
monthly stock returns. The outperformance is particularly convincing when
the ratios are adjusted for growth, in which case it can be achieved either by
imposing a zero intercept and a unit slope or by merely imposing a positive
intercept and a slope between zero and one. Results are more mixed in annual
data during the third subsample, but here too earnings-based valuation ratios
perform reasonably well. The relatively weak performance of the dividendprice ratio in this period is consistent with the observation of Boudoukh et al.
(2007), who find that in recent years share repurchases have played a more
important role in total payouts to shareholders.
Figure 2 compares the historical mean stock return with four forecasts based
on the smoothed earnings yield: the unrestricted rolling regression forecast,
the zero-intercept unit-slope forecast based on the smoothed earnings yield
with no adjustments, and the zero-intercept unit-slope forecasts based on
the two growth-adjusted versions of the smoothed earnings yield. The theoretically restricted forecasts are more stable over time and have superior
out-of-sample R 2 statistics as illustrated in the bottom panel of the figure.
Although there is a very gradual decline over time in the out-of-sample R 2
statistics of these forecasts, this does not mean that the forecasts are not
working in the later years of the sample; it simply means that they are
not beating the historical mean forecast as decisively as they did in earlier
years.
4.53
5.34
8.22
3.05
2.38
3.87
4.51
3.49
5.37
9.95
7.45
12.51
2.77
2.21
3.73
4.40
3.44
5.34
0.01
0.06
0.27
1.67
1.25
3.19
3.67
7.58
10.49
4.83
4.37
6.38
0.45
0.41
0.60
9.46
5.08
4.93
1.76
0.85
1.27
7.09
3.57
4.68
4.91
3.36
0.67%
0.30
0.51
0.59
0.33
0.47
0.73
0.76
0.66
0.74
0.89
B: Annual Returns
5.99
6.88
3.25
2.56
3.71
3.71
1.74
6.61
0.85
3.97
1.19
4.65
5.16
10.34
3.56
8.28
4.68
7.16
4.84
7.32
4.22
11.85
A: Monthly Returns
0.88%
0.57%
0.56
0.45
0.80
0.48
0.18
0.18
0.12
0.12
0.19
0.19
0.62
0.73
0.24
0.24
0.34
0.34
0.28
0.28
0.82
0.91
Fixed
Coefs
16.19
6.06
8.86
1.87
1.63
2.30
0.24
2.19
1.88
2.36
0.25
1.30%
0.53
1.06
0.11
0.05
0.06
0.12
0.11
0.06
0.04
0.14
Unconstrained
Fixed
Coefs
0.54%
0.07
0.01
0.14
0.16
0.16
0.00
0.08
0.03
0.02
0.27
7.98
1.47
1.33
0.28
1.60
1.81
2.43
2.95
0.64
0.47
6.20
Pos.
Intercept,
Bounded
Slope
0.21%
0.09
0.06
0.11
0.05
0.06
0.02
0.11
0.06
0.04
0.02
1.38
0.88
1.33
1.82
1.63
2.23
0.14
2.19
1.88
2.36
0.35
Sample: 19802005
The table provides out-of-sample R-squared statistics from predicting the equity premium with a forecasting variable versus the historical mean. The subsamples
roughly but not exactly divide the data into thirds. The column labels Unconstrained, Pos. Intercept, Bounded Slope, and Fixed Coefs are all defined in Tables
1 and 2. We do not provide results for the Book-to-market + Growth predictor in the 19271956 subsample because we do not have 20 years of data until 1956.
Dividend/price
Earnings/price
0.30
0.20
0.39
0.86%
0.16
0.56
0.15
0.06
0.09
0.63%
1.04
1.33
0.78
0.73
0.93
0.21%
0.28
0.53
0.18
0.12
0.25
Unconstrained
Unconstrained
Pos.
Intercept,
Bounded
Slope
Pos.
Intercept,
Bounded
Slope
Fixed
Coefs
Sample: 19561980
Sample: 19271956
1522
Dividend/price
Earnings/price
Table 3
Subsample stability
1523
Figure 2
Forecasting excess returns with the smoothed earnings yield
The top panel shows the historical annualized excess return forecasts for the historical mean, an unrestricted regression on the smoothed earnings yield, and three theoretically restricted
forecasts from the smoothed earnings yield, as labeled in the figure. The bottom panel shows the cumulative out-of-sample R-squared up to each point in the historical sample, for the four
models that use the earnings yield relative to the historical mean.
imposed on the predictive regression. However, the R 2 statistics are small in

magnitude. This raises the important question of whether they are economically
meaningful.
To explore this issue, consider the following example:
rt+1 = + xt + t+1 ,
(7)
and earns an average excess return of

1
S2
2
=
,
2
2
x +
where S is the unconditional Sharpe ratio of the risky asset.

If the investor observes xt , the investor sets

+ xt
1
,
t =
(9)
(10)
where the denominator is now 2 rather than 2x + 2 because the variation in

the predictor variable xt is now expected and does not contribute to risk. The
investor earns an average excess return of
2

2
1
+ 2x
S + R2
1
=
,
(11)
1 R2
where
R2 =
2x
2x + 2
(12)
is the R 2 statistic for the regression of excess return on the predictor variable
xt .
11
Merton (1969) presents the analogous portfolio solution for the case where the investor has power utility with
relative risk aversion , asset returns are lognormally distributed, and the portfolio can be continuously rebalanced.
Campbell and Viceira (2002, chap. 2) use a discrete-time approximate version of Mertons solution. Sentana
(2005) also explores the relation between regression forecasts and optimal portfolio construction.
1524
where rt+1 is the excess simple return on a risky asset over the riskless interest
rate, is the unconditional average excess return, xt is a predictor variable with
mean zero and constant variance 2x , and t+1 is a random shock with mean
zero and constant variance 2 . For tractability, consider an investor with a
single-period horizon and mean-variance preferences. The investors objective
function is expected portfolio return less (/2) times portfolio variance, where
can be interpreted as the coefficient of relative risk aversion.11 If the investor
does not observe xt , the investor chooses a portfolio weight in the risky asset

1
(8)
t = =
2x + 2
The difference between the two expected returns is

1
R2
(1 + S 2 ),
1 R2
(13)
which is always larger than R 2 /, and is close to R 2 / when the time interval is
short and R 2 and S 2 are both small. The proportional increase in the expected
return from observing xt is
R2
1 R2
1 + S2
S2

,
(14)
which is always larger than R 2 /S 2 and is close to R 2 /S 2 when the time interval
is short and R 2 and S 2 are both small.
This analysis shows that the correct way to judge the magnitude of R 2 is
to compare it with the squared Sharpe ratio S 2 . If R 2 is large relative to S 2 ,
then an investor can use the information in the predictive regression to obtain
a large proportional increase in portfolio return. In our monthly data since
1871, the monthly Sharpe ratio for stocks is 0.108, corresponding to an annual
Sharpe ratio of 0.374. The squared monthly Sharpe ratio is S 2 = 0.012 = 1.2%.
This can be compared with the monthly out-of-sample R 2 statistic for, say, the
smoothed earnings-price ratio of 0.43% in the last column of Panel A of Table 1.
A mean-variance investor can use the smoothed earnings-price ratio to increase
the average monthly portfolio return by a proportional factor of 0.43/1.2 =
36%. The absolute increase in portfolio return depends on risk aversion, but is
about 43 basis points per month or 5.2% per year for an investor with unit risk
aversion, and about 1.7% per year for an investor with a risk aversion coefficient
of three. The calculation can also be done in the annual data shown in Panel B
of Table 1, comparing the squared annual Sharpe ratio of 11.8% to the annual
out-of-sample R 2 statistic for, say, the smoothed earnings-price ratio of 7.9%.
A mean-variance investor can use the smoothed earnings-price ratio in annual
data to increase the average annual portfolio return by a proportional factor of
7.9/11.8 = 67%. The absolute increase in portfolio return is 9.6% per year for
an investor with unit risk aversion, and 3.2% per year for an investor with a
risk aversion coefficient of three.
The investor who observes xt gets a higher portfolio return in part by taking
on greater risk. Thus, the increase in the average return is not a pure welfare
gain for a risk-averse investor. To take account of this, in Table 4 we calculate
the welfare benefits generated by optimally trading on each predictor variable
for an investor with a relative risk aversion coefficient of three. We impose
realistic portfolio constraints, preventing the investor from shorting stocks or
taking more than 50% leverage: that is, confining the portfolio weight on stocks
to lie between 0% and 150%. The investors optimal portfolio depends on an
estimate of stock return variance at each point in time, and we assume that the
1525
12
In an earlier draft, we reported similar results for the case where the investor estimates variance using all the
available historical data. If these variance estimates are incorrect, for example because the predictor variable
forecasts variance as well as return, this will reduce the utility generated by trading on the predictive variable.
1526
investor estimates variance using a rolling five-year window of monthly data.12

We report the change in utility, relative to investing with the historical mean
forecast of the equity premium, for each of the models considered in Table 3.
These utility differences have the units of expected annualized return, so they
can be interpreted as the transactions costs or portfolio management fees that
investors would be prepared to pay each year to exploit the information in the
predictor variable.
The utility gains in Table 4 are broadly consistent with the out-of-sample
R 2 statistics reported in Table 3. Utility gains are reliably positive using annual returns and coefficients fixed at theoretical levels: the only utility loss
occurs for the unadjusted dividend-price ratio in the 19802005 sample, when
share repurchases become important. The results are somewhat more mixed for
monthly returns, but here too are generally positive. The utility gains tend to
be greatest in the 19271956 subsample, but by contrast with Table 3, are not
always smaller in 19802005 than in 19561980.
The utility gains reported in Table 4 are limited by the leverage constraint,
together with the high average equity premium. Predictable variations in stock
returns do not generate portfolio gains when there is a binding upper limit on
equity investment. Utility gains would be larger if we relaxed the portfolio constraint or included additional assets, with higher average returns than treasury
bills, in the portfolio choice problem. On the other hand, Table 4 does not take
any account of transactions costs. Modest gains from market timing strategies
could be offset by the additional costs implied by those strategies. Optimal
trading strategies in the presence of transactions costs are complex, and so we
do not explore this issue further here. We note that even the baseline strategy
based on a historical-mean forecast incurs rebalancing costs and that utility
gains of 50 basis points per year, which are commonly achieved in Table 4, are
sufficient to cover substantial additional costs.
Since small R 2 statistics can generate large benefits for investors, we should
expect predictive regressions to have only modest explanatory power. Regressions with large R 2 statistics would be too profitable to believe. The saying
If youre so smart, why arent you rich? applies with great force here, and
should lead investors to suspect that highly successful predictive regressions
are spurious. Since the squared Sharpe ratio and average real interest rate increase in proportion with the investment horizon, however, larger R 2 statistics
are believable at longer horizons. Authors such as Fama and French (1988)
have found that R 2 statistics increase strongly with the horizon when the predictor variable is persistent, a finding that is analyzed in Campbell, Lo, and
MacKinlay (1997, chap. 7) and Campbell (2001). This behavior is completely
consistent with our analysis and empirical results.
0.02
0.60
0.39
1527
Sample: 19802005
4.74
1.44
3.06
0.19
0.04
0.41
0.10
0.06
0.02
0.24
0.17
3.69%
0.80
2.73
0.22
0.00
0.04
0.08
0.04
0.09
0.04
0.04
0.34
0.53
0.81
0.17
0.04
0.36
0.13
0.06
0.02
0.24
0.14
0.22%
0.07
0.13
0.22
0.00
0.06
0.02
0.04
0.09
0.04
0.10
Pos. Intercept,
Unconstrained Bounded Slope
0.95
0.68
0.81
0.34
0.24
0.50
0.26
0.23
0.15
0.36
0.26
1.32%
0.74
0.46
0.34
0.19
0.24
0.23
0.18
0.17
0.21
0.11
Fixed
Coefs
This table presents out-of-sample portfolio choice results. The numbers are the change in average utility from forecasting the market with the predictor instead of the historical
mean. All numbers are annualized, so we multiply the monthly numbers by 12. Unconstrained indicates that we use the unconstrained OLS predictor of the equity premium.
Pos. Intercept, Bounded Slope indicates that we use the forecast with the intercept bounded above zero and the slope bounded between zero and one. Fixed Coefs indicates
that we use the forecast that sets the intercept to zero and the slope to one. The utility function is E(Rp) (/2)Var(Rp), where Rp is the portfolio return and = 3. All utility
changes are annualized, so we multiply monthly utility changes by 12.
0.95
0.40
0.44
0.70
0.49
0.51
1.27
1.06
0.85
0.94
1.93
0.66
0.49
0.87
1.56
1.41
1.53
0.35
0.15
0.28
0.01
0.01
0.09
0.20
0.25
B: Annual Returns
0.59
0.41
0.44
0.35
0.15
0.25
0.14
0.01
0.09
0.18
0.13
0.34
0.78
1.42
1.04
1.52
1.34
0.76
1.30
1.91
0.34
0.28
0.53
Dividend/price
0.55
Earnings/price
1.29
1.81
0.18
0.14
0.05
0.44
0.34
0.68
0.74
0.79
0.20
0.96
0.61
1.14
0.42
0.31
0.36
1.46%
0.51
0.04
0.64
0.21
0.13
1.38
1.49
0.73
1.07
1.94
Fixed
Coefs
A: Monthly Returns
0.92%
0.12
0.05
0.46
0.08
0.04
0.15
0.01
0.17
0.38
0.05
0.11%
1.80
1.89
0.43
1.60
0.89
0.43%
1.42
1.14
0.64
0.36
0.77
0.03%
0.86
0.53
0.28
1.05
0.47
1.93%
0.28
0.75
0.46
0.08
0.04
0.43
0.01
0.17
0.38
0.38
Pos. Intercept,
Fixed
Coefs
Pos. Intercept,
Dividend/price
Earnings/price
Sample: 19561980
Sample: 19271956
Table 4
Portfolio choice
3. Conclusion
1528
A number of variables are correlated with subsequent returns on the aggregate

U.S. stock market in the twentieth century. Some of these variables are stock
market valuation ratios; others reflect the levels of short- and long-term interest
rates, patterns in corporate finance or the cross-sectional pricing of individual
stocks, or the level of consumption in relation to wealth. Goyal and Welch
(2007) argue that in-sample correlations conceal a systematic failure of these
variables out of sample: none have been able to beat a simple forecast based on
the historical average stock return.
In this article, we have shown that most of these predictor variables perform
better out-of-sample than the historical average return forecast, once weak restrictions are imposed on the signs of coefficients and return forecasts. The
out-of-sample explanatory power is small, but nonetheless is economically
meaningful for investors. We have also explored theoretical restrictions on the
coefficients relating valuation ratios to future returns. Although these restrictions are based on steady-state models, which do not fully describe asset prices
in an economy with changing growth rates and discount rates, they improve
the out-of-sample performance of valuation forecasts by removing the need
to estimate coefficients from short samples of volatile stock returns. We find
that even in the period 19802005, where return prediction is most challenging, theoretically restricted valuation models often outperform return forecasts
based on the long-run historical mean of stock returns. These models also
generate meaningful utility gains for mean-variance investors. Importantly, the
performance of the models does not depend sensitively on the particular valuation ratio that is used or the manner in which it is adjusted for long-run
growth. We interpret this result as illustrating the well-known econometric
principle that even false theoretical restrictions can be helpful in forecasting if they reduce the variance of a predictor more than they increase its
bias.
We have deliberately kept our methods simple to avoid the charge that we
have tested a wide variety of methods and have only reported those that succeed. The results suggest that further improvements in performance can be
achieved by exploring alternative methods to combine theory with historical
data. One obvious possibility is a Bayesian approach with priors centered either on a model with a constant expected stock return, as in Wachter and
Warusawitharana (2006), or on a valuation model with constant or slowly
changing profitability and payout ratios of the sort we have explored in this
article. Of course, given the quantity of historical data that are now available, optimal forecasts of stock returns going forward may place greater
weight on the data, and less weight on theoretical restrictions, than those
methods that most successfully predicted stock returns during the twentieth
century.
References
Amihud, Y., and C. Hurvich. 2004. Predictive Regressions: A Reduced-Bias Estimation Method. Journal of
Financial and Quantitative Analysis 39:81341.
Ang, A., and G. Bekaert. 2007. Stock Return Predictability: Is It There? The Review of Financial Studies
20:651707.
Baker, M., and J. Wurgler. 2000. The Equity Share in New Issues and Aggregate Stock Returns. Journal of
Finance 55:221957.
Bollerslev, T. 1990. Modeling the Coherence in Short-Run Nominal Exchange Rates: A Multivariate Generalized
ARCH Model. Review of Economics and Statistics 72:498505.
Boudoukh, J., R. Michaely, M. Richardson, and M. Roberts. 2007. On the Importance of Measuring Payout
Yield: Implications for Empirical Asset Pricing. Journal of Finance 62:877916.
Butler, A. W., G. Grullon, and J. P. Weston. 2005. Can Managers Forecast Aggregate Market Returns? Journal
of Finance 60:96386.
Campbell, J. Y. 1987. Stock Returns and the Term Structure. Journal of Financial Economics 18:37399.
Campbell, J. Y. 2001. Why Long Horizons? A Study of Power Against Persistent Alternatives. Journal of
Empirical Finance 8:45991.
Campbell, J. Y., A. W. Lo, and A. C. MacKinlay. 1997. The Econometrics of Financial Markets. Princeton:
Princeton University Press.
Campbell, J. Y., and R. J. Shiller. 1988a. The Dividend-Price Ratio and Expectations of Future Dividends and
Discount Factors. The Review of Financial Studies 1:195228.
Campbell, J. Y., and R. J. Shiller. 1988b. Stock Prices, Earnings, and Expected Dividends. Journal of Finance
43:66176.
Campbell, J. Y., and R. J. Shiller. 1998. Valuation Ratios and the Long-Run Stock Market Outlook. Journal of
Portfolio Management 24(2):1126.
Campbell, J. Y., and R. J. Shiller. 2001. Valuation Ratios and the Long-Run Stock Market Outlook: An Update.
NBER Working Paper 8221.
Campbell, J. Y., and L. M. Viceira. 2002. Strategic Asset Allocation: Portfolio Choice for Long-Term Investors.
New York: Oxford University Press.
Campbell, J. Y., and M. Yogo. 2006. Efficient Tests of Stock Return Predictability. Journal of Financial
Economics 81:2760.
Cavanagh, C. L., G. Elliott, and J. H. Stock. 1995. Inference in Models with Nearly Integrated Regressors.
Econometric Theory 11:113147.
Clark, T. E., and K. D. West. 2005. Using Out-of-Sample Mean Squared Prediction Errors to Test the Martingale
Difference. NBER Technical Paper 305.
Cochrane, J. H. 2007. The Dog That Did Not Bark: A Defense of Return Predictability. The Review of Financial
Studies, published September 22, 2007, 10.1093/rfs/hhm046.
Fama, E. F., and K. R. French. 1988. Dividend Yields and Expected Stock Returns. Journal of Financial
Economics 22:325.
Fama, E. F., and K. R. French. 1989. Business Conditions and Expected Returns on Stocks and Bonds. Journal
of Financial Economics 25:2349.
Fama, E. F., and K. R. French. 2002. The Equity Premium. Journal of Finance 57:63759.
1529
Bollerslev, T., and J. Wooldridge. 1992. Quasi-Maximum Likelihood Estimation and Inference in Dynamic
Models with Time-Varying Covariances. Econometric Reviews 11:14372.
Fama, E. F., and G. W. Schwert. 1977. Asset Returns and Inflation. Journal of Financial Economics 5:11546.
Ferson, W. E., S. Sarkissian, and T. T. Simin. 2003. Spurious Regressions in Financial Economics? Journal of
Finance 58:1393413.
Foster, F. D., T. Smith, and R. E. Whaley. 1997. Assessing Goodness-of-Fit of Asset Pricing Models: The
Distribution of the Maximal R 2 . Journal of Finance 52:591607.
Gordon, M. 1962. The Investment, Financing, and Valuation of the Corporation. Homewood, IL: Irwin.
Goyal, A., and I. Welch. 2003. Predicting the Equity Premium with Dividend Ratios. Management Science
49:63954.
Graham, B., and D. L. Dodd. 1934. Security Analysis. First edition. New York: McGraw Hill.
Hodrick, R. J. 1992. Dividend Yields and Expected Stock Returns: Alternative Procedures for Inference and
Measurement. The Review of Financial Studies 5:25786.
Inoue, A., and L. Kilian. 2004. In-Sample or Out-of-Sample Tests of Predictability: Which One Should We Use?
Econometric Reviews 23:371402.
Jansson, M., and M. J. Moreira. 2006. Optimal Inference in Regression Models with Nearly Integrated Regressors.
Econometrica 74:681715.
Keim, D. B., and R. F. Stambaugh. 1986. Predicting Returns in the Stock and Bond Markets. Journal of Financial
Economics 17:35790.
Kilian, L. 1999. Exchange Rates and Monetary Fundamentals: What Do We Learn from Long-Horizon Regressions? Journal of Applied Econometrics 14:491510.
Kothari, S.P., and J. Shanken. 1997. Book-to-Market, Dividend Yield, and Expected Market Returns: A TimeSeries Analysis. Journal of Financial Economics 44:169203.
Lamont, O. 1998. Earnings and Expected Returns. Journal of Finance 53:156387.
Lettau, M., and S. Ludvigson. 2001. Consumption, Aggregate Wealth, and Expected Stock Returns. Journal of
Finance 56:81549.
Lewellen, J. 2004. Predicting Returns with Financial Ratios. Journal of Financial Economics 74:20935.
Litterman, R. 1986. Forecasting with Bayesian Vector Autoregressions: Five Years of Experience. Journal of
Business and Economic Statistics 4:2538.
Mark, N. C. 1995. Exchange Rates and Fundamentals: Evidence on Long-Horizon Predictability. American
Economic Review 85:20118.
Merton, R. C. 1969. Lifetime Portfolio Selection under Uncertainty: The Continuous Time Case. Review of
Economics and Statistics 51:24757.
Nelson, C., and M. Kim. 1993. Predictable Stock Returns: The Role of Small Sample Bias. Journal of Finance
48:64161.
Polk, C., S. Thompson, and T. Vuolteenaho. 2006. Cross-Sectional Forecasts of the Equity Premium. Journal of
Financial Economics 81:10141.
Pontiff, J., and L. D. Schall. 1998. Book-to-Market Ratios as Predictors of Market Returns. Journal of Financial
Economics 49:14160.
Rozeff, M. S. 1984. Dividend Yields are Equity Risk Premiums. Journal of Portfolio Management 11(1):6875.
Sentana, E. 2005. Least Squares Predictions and Mean-Variance Analysis. Journal of Financial Econometrics
3:5678.
1530
Goyal, A., and I. Welch. 2007. A Comprehensive Look at the Empirical Performance of Equity Premium
Prediction. The Review of Financial Studies, forthcoming.
Siegel, J. 1994. Stocks for the Long Run. New York: Norton.
Stambaugh, R. F. 1999. Predictive Regressions. Journal of Financial Economics 54:375421.
Torous, W., R. Valkanov, and S. Yan. 2004. On Predicting Stock Returns with Nearly Integrated Explanatory
Variables. Journal of Business 77:93766.
Wachter, J. A., and M. Warusawitharana. 2006. Predictable Returns and Asset Allocation: Should a Skeptical
Investor Time the Market? Unpublished Paper, University of Pennsylvania.
1531

Campbell 2007 Equity Returns

Uploaded by

Copyright:

Available Formats

Campbell 2007 Equity Returns

Uploaded by

Document Information

Original Description:

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

Campbell 2007 Equity Returns

Uploaded by

Copyright:

Available Formats

Predicting Excess Stock Returns Out of

Sample: Can Anything Beat the Historical

Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015

The Review of Financial Studies / v 21 n 4 2008

Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015

Predicting Excess Stock Returns Out of Sample

1. Theoretical Restrictions on Predictive Regressions

Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015

total returns data are available from CRSP. We calculate an out-of-sample

Out-of-Sample R-squared with Different Constraints

Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015

The Review of Financial Studies / v 21 n 4 2008

Predicting Excess Stock Returns Out of Sample

Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015

is measured in levels, rather than logs, with some accounting measure of

Out-of-Sample R-squared with Different Constraints

Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015

The Review of Financial Studies / v 21 n 4 2008

Predicting Excess Stock Returns Out of Sample

Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015

The Review of Financial Studies / v 21 n 4 2008

1.1 Sign restrictions on slope coefficients and forecasts

Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015

Predicting Excess Stock Returns Out of Sample

1.2 Quantitative coefficient restrictions for valuation ratios

Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015

Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015

The Review of Financial Studies / v 21 n 4 2008

Predicting Excess Stock Returns Out of Sample

Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015

where D/E is the payout ratio, to obtain a growth-adjusted return forecast:

The Review of Financial Studies / v 21 n 4 2008

Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015

Predicting Excess Stock Returns Out of Sample

2. How Large an R2 Should We Expect?

Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015

Goyal and Welch (2007) emphasize that out-of-sample performance has

Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015

The Review of Financial Studies / v 21 n 4 2008

Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015

Predicting Excess Stock Returns Out of Sample

The Review of Financial Studies / v 21 n 4 2008

imposed on the predictive regression. However, the R 2 statistics are small in

and earns an average excess return of

where S is the unconditional Sharpe ratio of the risky asset.

where the denominator is now 2 rather than 2x + 2 because the variation in

Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015

Predicting Excess Stock Returns Out of Sample

The difference between the two expected returns is

Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015

The Review of Financial Studies / v 21 n 4 2008

Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015

investor estimates variance using a rolling five-year window of monthly data.12

Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015

Predicting Excess Stock Returns Out of Sample

The Review of Financial Studies / v 21 n 4 2008

Downloaded from http://rfs.oxfordjournals.org/ at University of Tasmania Library on January 9, 2015

A number of variables are correlated with subsequent returns on the aggregate

Predicting Excess Stock Returns Out of Sample