See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/46543602

Aggregate Earnings, Firm-Level Earnings, and Expected Stock Returns

Article  in  Journal of Financial and Quantitative Analysis · September 2008

DOI: 10.1017/S0022109000004245 · Source: RePEc

CITATIONS READS
46 318

3 authors, including:

K. Ozgur Demirtas Hassan Tehranian

Sabanci University Boston College, USA
61 PUBLICATIONS   495 CITATIONS    91 PUBLICATIONS   5,641 CITATIONS   

SEE PROFILE SEE PROFILE

All content following this page was uploaded by K. Ozgur Demirtas on 29 September 2014.

The user has requested enhancement of the downloaded file.

 Aggregate Earnings, Firm-Level Earnings and Expected Stock Returns‡

Turan G. Bali, K. Ozgur Demirtas and Hassan Tehranian∗

Abstract

This paper provides an analysis of the predictability of stock returns using market, industry, and firm-level
earnings. Contrary to Lamont (1998), we find that neither dividend payout ratio nor the level of aggregate
earnings can forecast the excess market return. We show that these variables do not have robust predictive
power across different stock portfolios and sample periods. In contrast to the aggregate-level findings,
earnings yield has significant explanatory power for the time-series and cross-sectional variation in firm-
level stock returns and 48-industry portfolio returns. It is the mean-reversion of stock prices as well as the
earnings’ correlation with expected stock returns that are responsible for the forecasting power of earnings
yield. These results are robust after controlling for book-to-market, size, price momentum and post-earnings
announcement drift. At the aggregate-level, the information content of firm-level earnings about future cash
flows is diversified away and higher aggregate earnings do not forecast higher returns.

JEL # G10, G12, G14

Keywords: earnings, dividends, stock returns, market returns, predictability, business cycle
‡ We thank Hendrik Bessembinder (the editor) and two anonymous referees for their extremely helpful comments and

suggestions. We owe a great debt to Wayne Ferson, Pierluigi Balduzzi and Jeff Pontiff. We benefited from discussions
with George Aragon, Armen Hovakimian, Jonathan Lewellen, Salih Neftci, Lin Peng, Robert Schwartz, Jonathan
Wang, and Liuren Wu. We also thank the seminar participants at Baruch College, Boston College, City University of
New York, Drexel University, Koc University, and Rice University. Turan G. Bali and K. Ozgur Demirtas gratefully
acknowledge the financial support from the PSC-CUNY Research Foundation of the City University of New York.
Any remaining errors are our own.
∗ Turan G. Bali is professor at the Department of Economics and Finance, Zicklin School of Business, Baruch College,

City University of New York, One Bernard Baruch Way, Box 10-225, New York, New York 10010. Phone: (646)
312-3506, Fax: (646) 312-3451, E-mail: [email protected]; K. Ozgur Demirtas is an assistant professor
at the Department of Economics and Finance, Zicklin School of Business, Baruch College, City University of New
York, One Bernard Baruch Way, Box 10-225, New York, New York 10010. Phone: (646) 312-3484, Fax: (646) 312-
3451, E-mail: [email protected]; Hassan Tehranian is Griffith Millenium Chair at the Department
of Finance, Carroll School of Management, Boston College, Fulton Hall 324C, Chestnut Hill, MA 02167. Phone:
(617) 552-3944, Fax: (617) 552-043, E-mail: [email protected]
 Aggregate Earnings, Firm-Level Earnings and Expected Stock Returns

Abstract

This paper provides an analysis of the predictability of stock returns using market, industry, and firm-level
earnings. Contrary to Lamont (1998), we find that neither dividend payout ratio nor the level of aggregate
earnings can forecast the excess market return. We show that these variables do not have robust predictive
power across different stock portfolios and sample periods. In contrast to the aggregate-level findings,
earnings yield has significant explanatory power for the time-series and cross-sectional variation in firm-
level stock returns and 48-industry portfolio returns. It is the mean-reversion of stock prices as well as the
earnings’ correlation with expected stock returns that are responsible for the forecasting power of earnings
yield. These results are robust after controlling for book-to-market, size, price momentum and post-earnings
announcement drift. At the aggregate-level, the information content of firm-level earnings about future cash
flows is diversified away and higher aggregate earnings do not forecast higher returns.

1
I Introduction

Both academics and practitioners have sought to identify variables that forecast stock returns.
Many variables in the form of value-to-price ratios have been shown to predict time-series varia-
tion in aggregate stock returns. Shiller (1984) and Fama and French (1988a) estimate predictive
regressions of aggregate stock returns on dividend and earnings yield, and find that earnings yield
(E/P) has less predictive power than dividend yield (D/P).1 Fama and French (1988a) argue that
if higher variability of earnings is unrelated to the variation in expected returns, E/P is a noisier
measure of expected returns than (D/P).2

As a response to the interpretation of Fama and French (1988a) about the predictive power
of E/P, Lamont (1998) argues that aggregate earnings are negatively correlated with expected
returns, and that the variability of aggregate earnings is not noise but indeed related to future
returns. He also indicates that the aggregate dividend payout ratio forecasts excess returns on the
S&P Composite Index because high dividends (earnings) forecast high (low) returns. He explains
his finding of the negative and significant relation between aggregate earnings and expected returns
as follows: “The level of earnings is a good measure of current business conditions. Risk premia
on stocks covary negatively with current economic activity: investors require high expected returns
in recessions, and lower expected returns in booms. Since earnings vary with economic activity,
current earnings predict future returns.” Moreover, P is negatively correlated with future returns
due to mean-reversion in stock prices. Since both earnings (E) and price (P) covary negatively
with expected returns and their forecasting powers are offsetting, Lamont (1998) concludes that
earnings yield (E/P) fails to forecast aggregate stock returns.

We find that Lamont’s (1998) empirical results are not robust across different sample periods.
Contrary to the findings of Lamont (1998), we do not reject the hypothesis that the relations
between future returns and aggregate earnings, and future returns and dividend payout ratio are
flat. In other words, neither dividend payout ratio nor the level of aggregate earnings are correlated
with the excess market return.
1
In a recent paper, Lewellen (2004) also finds that D/P has more explanatory power than E/P for the excess
market return.
2
We should note that Fama and French (1988a) do not explicitly test whether earnings are related to expected
returns. Upon observing little forecasting power of earnings yield, they hypothesize that higher variability in earnings
may just be noise, unrelated to the variation in expected returns.

2
 Lamont (1998) examines the forecasting ability of aggregate earnings and dividend payout
ratio from 1947:Q1 to 1994:Q4 for the S&P Composite Index. We replicate his results for the
sample period of 1947:Q1 to 1994:Q4, but show that the predictive power of aggregate earnings
and dividend payout ratio does not exist for the extended sample periods of 1947:Q1-2002:Q4
and 1940:Q1-2002:Q4. In fact, once his sample is extended to 1995, the estimated coefficients on
earnings become marginally significant (only at the 10% level), and extending his data set to 1996
is enough for the disappearance of the results. Thus, it is the specific time period which drive the
results of Lamont (1998).

Since we do not find any evidence for a significant link between aggregate earnings and excess
market return, the potential question is: Can the aggregation of firm-level earnings diversify away
the information content of earnings about expected future cash flows and expected returns? To
answer this question, we investigate the predictability of firm-level stock returns and show that,
in contrast to the aggregate-level findings, earnings yield forecasts the firm-level stock returns.
We argue that firm-level earnings consist of two components: The portion that is explained by
aggregate earnings (systematic earnings) and the portion that is orthogonal to aggregate earnings
(unsystematic earnings). When firm-level earnings are aggregated to generate the market-level
earnings used in previous studies, the unsystematic earnings component diversifies away. Thus,
the aggregate-level earnings do not have any explanatory power for future returns. In contrast,
firm-level earnings contain significant information about future stock returns. Furthermore, at the
firm-level, prices and earnings have an opposite relationship with future returns, and as a result,
the informativeness of earnings and prices about expected returns is not offsetting.

Since there is a significantly positive relation between earnings and expected returns at the firm
level, but the relation is flat at the market level, a natural question to ask is whether there is any
predictability at the industry level. First, we use the 17-industry portfolios of Fama and French
(1997) to test whether industry-level earnings can predict excess returns on industry portfolios or
whether industry-level diversification is sufficiently large to eliminate predictability. The results
indicate that the aggregation of firm-level earnings to 17-industry portfolios diversifies away the
information content of firm-level earnings about future cash flows and hence there is no significant
relation between industry-level earnings and expected returns. Second, we conjecture that earnings
for a finer industry partition may contain significant information about future cash flows and this

3
information is partially diversified away, not completely as in the case of 17-industry portfolios.
To test our conjecture, we use the 48-industry portfolios of Fama and French (1997) and find
a positive and highly significant relation between earnings and expected returns for 48-industry
partition. Hence, we conclude that the effect of aggregation of firm-level earnings on the relation
between industry earnings and expected returns is considerably different for 17 and 48-industry
portfolios.

Regardless of the reason for why stock prices are mean-reverting, the level of stock price is an
important variable in predicting future returns. The predictive power of cash-flow-to-price ratios,
such as E/P, does not necessarily indicate that the cash flow proxy itself contains information
about future returns. Although our motivation for exploring the forecasting power of firm-level
earnings yield stems from our belief that firm-specific earnings are informative about the expected
future cash flows, it does not eliminate the possibility that only the price itself is responsible for the
correlation between earnings yield and future stock returns. Thus, one of our objectives is to use
realized firm-level earnings not as normalizing variables for stock price but as predictive variables
in their own right.

We show that normalized earnings themselves covary positively with future stock returns.
Hence, it is the mean-reversion of stock prices as well as the earnings’ correlation with expected
stock returns that are responsible for the forecasting power of earnings yield. Moreover, only the
unsystematic portion of earnings is correlated with future stock returns, while the variability in
systematic earnings is unrelated to expected returns.

In addition to testing the significance of a time-series relation between firm-level earnings and ex-
pected returns, we provide a comprehensive study on the cross-sectional relation between firm-level
earnings and expected stock returns. The average slope coefficients from Fama-MacBeth (1973)
regressions indicate a positive and highly significant relation between one quarter ahead returns
and firm-level earnings yield and normalized earnings. These results hold even after controlling for
size, book-to-market, momentum, and post-earnings announcement drift.

The paper is organized as follows. Section II presents the framework that forms the relation
between expected returns and earnings, dividends, and price. Section III explains the data con-
struction process and provides summary statistics. Section IV tests whether aggregate earnings can
explain the time-series variation in excess market returns. Section V presents empirical results on

4
the time-series and cross-sectional relation between firm-level earnings and expected stock returns.
Section VI examines the relation between industry-level earnings and expected returns. Section
VII concludes the paper.

II Framework

Market multiples such as dividend yield (D/P), book-to-market ratio (B/P), and earnings yield
(E/P) have been identified as important determinants of the time-series of expected stock returns
at the aggregate-level. Examples include Rozeff (1984), Shiller (1984), Fama and French (1988a,
1989), Campbell and Shiller (1988, 1989), Hodrick (1992), Kothari and Shanken (1997), Pontiff
and Schall (1998), Lamont (1998), Lewellen (2004), and Kothari, Lewellen, and Warner (2006).

In a time-series setting the log-linear approximation of stock returns developed by Campbell

and Shiller (1988, 1989) provides a convenient framework for examining predictive relations. The
approximation starts with the definition of log stock return r i,t+1 ,

(1) ri,t+1 = log(Pi,t+1 + Di,t+1 ) − log(Pi,t ),

where P i,t is the stock price at the end of period t and D i,t+1 denotes dividends for firm i paid
during period t + 1. Presenting log variables by lowercase letters and using a first-order Taylor
series expansion, equation (1) becomes
⎛ ⎞ ⎛ ⎞
X∞ X∞
(2) pi,t = Ci + Et ⎝ ρji (1 − ρi )di,t+1+j ⎠ − Et ⎝ ρji ri,t+1+j ⎠ ,
j=0 j=0

where C i is a linearization constant and ρi is a constant discount factor close to one for each firm
i. Given equation (2), Lamont (1998) shows that the time t expectation of time t + 1 returns can
be written as,
⎛ ⎞ ⎛ ⎞
X∞ X∞
(3) Et (ri,t+1 ) = −pi,t + Et ⎝ ρji (1 − ρi )di,t+1+j ⎠ − Et ⎝ ρji ri,t+1+j ⎠ + Ci .
j=0 j=1

It is clear from the right-hand side of equation (3) that the level of stock price (or the index level)
is an important explanatory variable. Higher (lower) stock price implies lower (higher) expected
returns. The second term on the right side of equation (3) is the expected sum of future discounted
log dividends, denoted by p cf . The third term is the discounted sum of future returns starting next

5
period, denoted by p r . Equation (3) indicates that after controlling for stock price, any variable
known as of time t predicts time t + 1 returns only if it proxies for p cf and/or p r .3,4

When, for example, log of earnings yield (E/P) is used to forecast the aggregate returns, earnings
in the numerator can be viewed as a proxy for p cf , whereas price in the denominator controls for
the mean-reversion in the market index level. Earlier research has interpreted the poor forecasting
ability of earnings yield (E/P) as the noise in earnings which is not related to the expected returns.
However, Lamont (1998) presents equation (3) as:

(4) Et (ri,t+1 ) = −(pi,t − ni,t ) + (pcf

i,t
− ni,t ) − pri,t + Ci ,

where ni,t is a normalization factor used to create stationary right-hand side variables. Equation
(4) gives a helpful hint in disentangling different forecasting powers of the numerator and the
denominator of a value-to-price ratio, by using the normalized stock price and the normalized
cash flow proxy as the right-hand side variables. In this representation, the coefficients of the
stock price and the cash flow proxy are not restricted to be the same. Lamont (1998) argues that
aggregate dividends proxy for p cf and aggregate earnings are positively correlated with business
cycle fluctuations. He finds that normalized earnings of the S&P Composite Index as well as the
normalized price (i.e., normalized index level) have negative coefficients in predictive regressions,
whereas normalized dividends have a positive coefficient. Thus, subtracting price from earnings
(i.e., using earnings yield as a predictive variable) masks the relation between earnings yield and
expected returns. Furthermore, higher dividend payout ratio forecasts higher future returns on the
S&P Composite Index because subtracting earnings from dividends does not have such offsetting
effect.

In this paper, we show that Lamont’s (1998) findings are driven by the specific sample period and
he uses. To compare with the aggregate-level findings, we test whether firm-level earnings yield is
an important predictor of the time-series of stock returns, because firm-level earnings may contain
relevant information about p cf . We argue that when earnings are aggregated, any information
3
Stock prices themselves predict future returns and we call this “the price effect”, because it is most compelling
when the cash flow proxy in the numerator of the ratio does not covary with expected returns.
4
We should note, however, that the results of the paper neither support nor reject market efficiency, since our
analyses hold regardless of whether discount rates are affected by irrational traders. Therefore, our analyses cannot
be used to determine whether market rationally prices assets.

6
about firm-level cash flows are diversified and what is left is the systematic portion of earnings
which does not have any explanatory power for the expected returns.5

III Data

A Construction of Variables

We obtain the monthly returns, prices, dividends and adjustment factors from the Center for Re-
search in Securities Prices (CRSP) NYSE/AMEX/Nasdaq monthly file. Quarterly realized earn-
ings, total assets, total liabilities, book value, sales and shares outstanding data come from the
quarterly CRSP-COMPUSTAT merged database. We obtain the average yields to maturity on
Aaa and Baa-rated bonds and the yields on the 3-month Treasury bill and 10-year Treasury bond
from the Federal Reserve Statistical Release. We obtain the one-month Treasury bill returns from
the Ibbotson Associates.

All variables are formed on a quarterly basis. Quarterly returns are constructed by compounding
monthly returns. Log excess return (ri,t ) is defined as the difference between the log quarterly return
and the log risk-free rate. RRELt is the relative Treasury bill rate defined as the end of quarter
3-month T-bill rate minus its twelve-month backward moving average. DEFt is the default spread
calculated as the difference between the yields on Baa and Aaa-rated corporate bonds. TERMt is
the term spread calculated as the difference between the yields on the 10-year Treasury bond and
the 3-month Treasury bill.

For every quarter t, we compute dividends per share (Dt ) as the sum of the past four quarters of
dividends per share. To ensure that earnings per share and other balance sheet variables are known
to the market as of quarter t, we extract the report dates of quarterly earnings and take the latest
reported values.6 We add deferred taxes and investment tax credit (Balance Sheet) to common
5
This argument is also consistent with previous research which shows that, contrary to firm-level returns, aggregate
returns are driven by expected return news (i.e., change in expected future returns). For example, Campbell (1991)
finds that most of the variation in aggregate returns comes from variation in changes of expected returns. Furthermore,
there is a long accounting literature which argues that changes in firm-level earnings are positively correlated with
cash flow news (see, e.g., Easton and Harris (1991)).
6
To ensure that realized earnings are not stale, as early as three quarters prior earnings are assigned to quarter t.
If the report date of quarter t earnings is missing, we assign earnings of quarter t-1 to quarter t. We note that the
empirical results are not sensitive to these choices.

7
equity to compute the book value. We treat negative or zero sales, total assets, total liabilities and
book value as missing.

For each stock, we construct a normalization variable (Ni,t ) as the total assets per share of stock
i at the end of quarter t. The logs of the ratios of prices, dividends, and earnings to normalization
variable are used to generate normalized price, dividends, and earnings. Normalization variable is
used to obtain stationary right-hand side variables. Unfortunately, there is no sound methodology
that will guide the search for the most appropriate normalization variable. Thus, as a robustness
check we use several other normalization variables and show in Section V.D that our results are
insensitive to the choice of the normalization variable.

In all estimations, we use log returns and log ratios following previous studies including Lamont
(1998). In order to work with the log data, we consider non-negative earnings. Specifically, we
E
define normalized earnings (ei,t −ni,t ) as log( Ni,t
i,t
). All other log ratios are defined similarly. To
avoid giving extreme observations heavy weight in our analysis, following Fama and French (1992),
the smallest and largest 0.5% of the observations on normalized variables and value-to-price ratios
are set equal to the next largest and smallest values of the ratio (the 0.005 and 0.995 fractiles).

To be included in our sample, a firm in quarter t must have time t and t + 1 log excess return,
normalized price, normalized earnings, size and book value available. We also consider another
sample which requires each stock to have normalized dividends as of quarter t. Furthermore, in
firm-level time-series panel regresions, we require each stock to have at least four years (16 quarters)
of data available.7

To replicate and check the robustness of Lamont’s (1998) results, we obtain the aggregate-level
earnings, dividends, and returns from the data source of Lamont (1998). Specifically, aggregate-
level earnings and dividends data are obtained from the Security Price Index Record published by
Standard & Poor’s Statistical Service.8 Quarterly return on the S&P Composite Index is defined
³ ´
as ln CSTIND t+1
CSTIND t , where CSTIND is the index of total return (including the reinvested dividends)
on the S&P Composite Index. At the aggregate-level, following Lamont (1998), we use log of the
7
However, to control for survivorship bias, we do not impose this restriction in our firm-level cross-sectional
regressions.
8
We used the latest available Security Price Index Record (published in 2002), which reports data until the last
quarter of 2001. We acquire the data for the last quarter of 2001 and the four quarters of 2002 from the Standard &
Poor’s electronically.

8
average of the past five years of annual earnings per share, calculated as the sum of the past 20
observations of quarterly earnings per share divided by five as the normalization variable.

B Summary Statistics

The overall sample consists of 289,958 quarterly observations. We also use an alternative sample
that requires dividends to exist and it consists of 167,699 observations. We use the second sample,
whenever dividend yield or normalized dividends are used in our estimations. Although the data
range from 1966:Q2 to 2002:Q4, quarterly Compustat database provides very small number of
observations prior to the early 1970s (there are only 223 observations prior to 1972:Q3). Therefore,
when we conduct our firm-level analysis, we require each quarter to have at least 30 observations.
This requirement led us to use data starting from 1972:Q3.

Table 1 shows the summary statistics of the firm-level data. Mean, median, standard deviation,
and correlation statistics are computed for the time-series of each stock and averages of these
statistics across firms are reported. Panel A shows that earnings are more volatile than dividends.
Earnings have an average standard deviation of 0.64, whereas dividends have an average standard
deviation of 0.31. However, whether this higher variability in earnings is related to expected stock
returns remains to be seen. Panel B shows the correlation statistics. As expected, both dividend
yield and earnings yield are negatively correlated with the scaled stock price. Higher scaled stock
prices imply lower yields. Similarly, higher scaled earnings and dividends are positively correlated
with earnings yield and dividend yield, respectively. However, relatively low correlation between
normalized earnings, (ei,t −ni,t ), and earnings yield, (ei,t −pi,t ), is due to the positive correlation
between contemporaneous normalized stock price and normalized earnings.

IV Aggregate Earnings and Expected Returns

To test whether earnings yield, dividend yield, and dividend payout ratio can predict one-quarter-
ahead excess market return, we run the following OLS regression:

(5) rm,t+1 = α + βXm,t + εm,t+1 ,

9
where rm,t+1 is the log excess return on the market portfolio at time t + 1, and Xm,t is a vector of
predictive variables which are in the information set as of time t.9 In order to correct for the small-
sample bias in parameter estimates, we use the randomization tecnique of Nelson and Kim (1993).
To further account for the effect of autocorrelation and heteroscedasticity in the error terms, the
Nelson and Kim methodology is implemented based on the Newey-West (1987) standard errors. For
each regression, we report the bias adjusted slope coefficients, t-statistics, bias adjusted one-sided
p-values and bias adjusted R2 values.10 In equation (5), the slope coefficient (β) along with its bias
adjusted p-value determines whether there is a significantly positive or negative relation between
earnings and expected returns at the market level.

Table 2 presents the results of predictive regressions using lagged return, dividend yield, earnings
yield, and dividend payout ratio.11 Panel A of Table 2 replicates the results given in Table IV of
Lamont (1998, p.1572) using the data from 1947:Q1 to 1994:Q4. Lamont (1998) reports OLS
coefficent estimates with OLS standard errors. For ease of comparison, we report in Panel A, both
OLS and small-sample bias adjusted coefficients. However, in all other panels, we only present bias
adjusted coefficients. Based on the bias adjusted p-values, in the first two regressions, dividend yield
is highly significant and has a positive coefficient, whereas earnings yield is not significant.12 Third
regression shows that conditional on the dividend yield, higher earnings yield forecasts lower returns
for the S&P Composite Index. Fourth and fifth regressions show that with or without dividend
yield, dividend payout ratio has significant forecasting power and higher dividend payout ratio
forecasts higher future returns. Last row of Panel A presents results from a multivariate regression,
which consists of lagged excess return, relative T-bill rate, dividend yield and dividend payout ratio.
Both dividend yield and payout ratio have positive and significant coefficients. Panel A reiterates
the finding of Lamont (1998) that dividend payout ratio has significant power in forecasting future
returns for the period of 1947:Q1-1994:Q4.
9
In addition to earnings yield, dividend yield, and dividend payout ratio, following Lamont (1998), we include the
lagged excess market return, rm,t , and RRELt in the information set.
10
See Section V and VI for a detailed discussion and implementation of the Nelson and Kim (1993) methodology.
11
Ferson, Sarkissian and Simin (2003, p.1409) state that “If the expected returns are persistent, there is a risk of
finding a spurious regression relation between the return and an independent, highly autocorrelated lagged variable”.
We address this issue by adding the lag of the dependent variable in predictive regressions.
12
Lewellen (2004) finds no evidence of a significant link between earnings yield and excess return on the value-
weighted and equal-weighted NYSE index for the sample period of 1963 to 1994. When the data for 1995-2000 are
included, earnings yield forecasts only the equal-weighted index. However, the statistical significance is only marginal
with a p-value of 0.09 (see Table 6 of Lewellen (2004)).

10
 In Panel B of Table 2, we run the first set of robustness checks by extending the sample period
from 1994 to 2002. First row shows dividend yield has a smaller coefficient than in Panel A and it
loses statistical significance. Second row indicates that the coefficient of earnings yield is larger than
that in the original period, but it is insignificant. Third row shows that conditional on dividend
yield, earnings yield has no forecasting power, which contradicts with Lamont’s finding for the
shorter sample period. Fourth and fifth rows present the most striking feature of Table 2 that with
or without dividend yield, dividend payout ratio has no forecasting power for the excess market
return. Furthermore, in a multivariate regression, relative T-bill rate is the only significant variable.
The coefficient of dividend payout ratio in row 6 is close to zero (0.005) with a one-sided p-value of
0.43. Thus, the findings of Lamont (1998) regarding earnings yield and dividend payout ratio do
not hold for the extended sample period.13

Panels A and B use the postwar data. However, we have available data for the S&P 500
Composite Index starting from 1940. In Panel C, we use 7 years of extra data and include the
prewar data as well. The results are similar to those in Panel B such that dividend payout ratio
has no forecasting power which contradicts with Lamont (1998). The failure of dividend payout
ratio in forecasting the future returns can be explained by the relative weakness of dividends in
explaining the future returns and/or by the fact that the variability in aggregate earnings is not
related to the variability in expected returns.

To further investigate the weakness of dividend payout ratio in predictive regressions, we con-
sider aggregate earnings and dividends not as normalizing variables for price but as predictive
variables in their own right. Table 3 shows the bias adjusted parameter estimates from the pre-
dictive regressions using scaled prices, dividends, and earnings.14 Panel A of Table 3 replicates
the results given in the first row of Table V of Lamont (1998, p.1575). Observe that the scaled
price has a negative and significant coefficient, indicating the presence of mean-reversion. Scaled
dividend has a positive and significant coefficient implying that it proxies for expected future cash
flows (pcf ) in equation (4). Moreover, the coefficient on scaled earnings is negative and significant.
Lamont’s (1998) interpretation of the negative coefficient on earnings is that earnings measure
13
Similar results are also obtained by Lettau and Ludvigson (2001), Goyal and Welch (2005), and Ang and Bekaert
(2006) for different sample periods.
14
Note that the regressions in Table 2 are similar to the regressions run by Kothari and Shanken (1997) and Pontiff
and Schall (1998) with D/P and B/P. They argue that (B/P)’s predictive ability is related to the ability of book
value to forecast future cash flows.

11
macroeconomic activity associated with business cycle fluctuations. Since both prices and earnings
have a negative relationship with future returns, earnings yield has no predictive power as shown
in Panel A of Table 2.

Panel B of Table 3 uses an extended sample period from 1947:Q1 to 2002:Q4. Observe that
scaled dividend has a positive but insignificant coefficient. More importantly, scaled earnings cannot
predict the excess return on the market. Extending the data to 2002 decreases the coefficient
estimate to zero. In Panel C, we use the prewar data and find that earnings’ coefficient estimate is
almost zero and insignificant. Thus, the forecasting power of aggregate earnings is limited to the
sample period considered by Lamont (1998).

Tables 2 and 3 show that when we extend the S&P sample to 2002, both dividend payout
ratio and normalized earnings become insignificant. However, these results do not answer the
question whether the part of the data analyzed by Lamont (1998) is special and hence his results
are sample-specific, or the additional data is special and hence our results are sample-specific (i.e.,
the lack of significance may be driven entirely by the inclusion of 1995-2002 data).15 In order to
clarify these points, in Table 4, we extended the data only backwards rather than forwards and
consider Lamont’s sample with additional 2 years of data. Panel A of Table 4 shows that even
when the original sample is only extended backwards and the data from 1940 to 1994 are used,
dividend payout ratio becomes insignificant. Furthermore, Panel B shows that when we extend the
sample by only two years and consider 1947 to 1996 period, dividend payout ratio still becomes
insignificant. In Panel C, we consider the prewar period and show that normalized earnings is
marginally significant only at the 8% level. Finally, Panel D shows that inclusion of only two
years of data is enough to wipe out the relation between earnings and expected returns. Thus, we
conclude that the evidence that Lamont’s (1998) results are sample-specific, is not entirely driven
by the inclusion of the bubble period.
15
For example, the additional period that we consider includes the bubble period in which the US market exhibited
unusually high returns. At the same time, the dividend yield was at a historically low level.

12
V Firm-Level Earnings and Expected Stock Returns

A Value-to-Price Ratios

We use fixed-effects panel data regressions to examine the predictability at the firm-level. A pre-
dictive regression for the stock returns in a fixed-effects setting can be demonstrated as:

(6) ri,t+1 = αi + βXi,t + εi,t+1 ,

where ri,t+1 is the log excess stock return for firm i at time t + 1, and Xi,t is a vector of predictive
variables which are in the information set as of time t. The important point in equation (6) is
that the intercepts (αi ) are estimated separately for each firm i, which distinguishes fixed effects
panel data regressions from pooled panel data regressions where the intercept is the same for each
stock. Therefore, pooled panel data regressions can be viewed as stacked time-series regressions as
well as stacked cross-sectional regressions. However, estimating intercepts separately is equivalent
to demeaning each stock level data and ensures that each stock’s error term is orthogonal to the
explanatory variables for that stock. Thus, fixed-effects panel data regression in equation (6) is
equivalent to a stacked time-series regression.

Table 5 presents the results of predictive fixed-effects panel data regressions using value-to-price
ratios. Earnings yield, dividend yield, and dividend payout ratio are used as independent variables.
These are firm-specific variables. Furthermore, we use the relative Treasury bill rate RRELt , term
spread TERMt , and default spread DEFt . At each quarter t, these macroeconomic variables are
the same for each stock. RRELt , TERMt and DEFt have been used to control for pr in equation
(4).16

It is well known that OLS standard errors are not appropriate to draw conclusions on the
statistical significance of the estimated coefficients in a panel-data setting due to cross-sectional
correlation in residuals. Therefore, we use two different methodologies to compute the standard
errors that are adjusted for contemporaneous cross-sectional correlation. The first methodology is
the jackknife method of Shao and Rao (1993). The other methodology follows Rogers’ (1983, 1993)
method for computing standard errors in the existence of heteroscedasticity and contemporaneous
cross-correlations. At an earlier stage of the study, we find that the results from the jackknife and
16
See, for example, Campbell and Shiller (1988), Fama and French (1988a, 1989), Campbell (1991), Ferson and
Harvey (1991), Hodrick (1992), and Lamont (1998).

13
Rogers’ standard errors are very similar. In Table 5 we report the Rogers’ t-statistics (known as
the clustered t-statistics) in parentheses.

As mentioned earlier, a fixed-effects panel data regression is essentially equivalent to estimating

a collection of time-series regressions at the firm-level, constraining the slopes to be the same
but allowing for different intercepts. In other words, a fixed-effects regression is like running an
individual, firm-level, time-series regression and thus the effective number of observations per firm
can be quite small. To control for the effect of small sample bias on our estimated slope coefficients,
we use the Nelson and Kim (1993) methodology in our predictive panel data regressions. We obtain
the simulated independent variables and simulated return variable by randomizing the error terms
of each stock in the panel. Hence, when we compute the bias in estimated coefficients, we assure
that no predictability is assumed for each stock in our sample. Finally, to obtain the p-values,
we compare the jackknife t-statistics of the overall sample with the distribution of the simulated
jackknife t-statistics.

For each regression in Table 5, the first row reports the small sample bias-adjusted slope coef-
ficients, the second row presents the clustered t-statistics in parentheses, and the third row gives
the small sample bias-adjusted p-values in square brackets. The last three columns report the bias-
adjusted R2 values, the number of observations, and the number of stocks used in panel regressions.
The first and second rows in Table 5 show the univariate regressions of the one-quarter-ahead firm-
level excess stock returns on earnings and dividend yield. Both earnings and dividend yield are
positively correlated with the time-series of future stock returns. The third row shows that earn-
ings yield is a significant predictor of firm-level stock returns, even in the presence of dividend
yield. Row 4 confirms this finding by using dividend payout ratio together with dividend yield.
Since conditional on dividend yield, earnings yield is positively correlated with future returns, after
controlling for dividend yield, a higher dividend payout ratio forecasts lower future returns.

In the last regression, we introduce RRELt , TERMt and DEFt as independent variables along
with earnings and dividend yield. After controlling for macroeconomic variables, there is still
a positive and significant relation between earnings yield and expected returns at the firm level.
However, the slope coefficient on dividend yield is not significant. The bias-adjusted slope coefficient
on firm-level earnings yield is 0.022 with the t-statistic of 5.26 and the bias-adjusted p-value of zero.

14
 We find that earnings yield has a significant predictive power for expected stock returns. Either
the mean-reversion in stock prices alone drives the forecasting power of earnings yield and/or firm-
level earnings are positively correlated with expected stock returns. Next, we examine the source
of this finding by using normalized variables.

B Normalized Earnings, Dividends and Stock Price

Panel A of Table 6 presents the bias-adjusted slope coefficients, clustered t-statistics and the bias-
adjusted p-values from the regressions of one-quarter-ahead excess returns on normalized earnings,
dividends, and stock price. The results in the first regression indicate that when stock price and
earnings are used as independent variables, scaled stock price has a negative coefficient while
earnings has a positive coefficient. The positive relation between earnings and future stock re-
turns implies that current earnings proxy for expected future cash flows (i.e., pcf in equation (4)).
Normalized price and normalized earnings have the coefficient estimates of -0.031 and 0.019, re-
spectively. Based on the clustered t-statistics and the bias-adjusted p-values, both normalized price
and earnings are highly significant.

The second regression of Panel A indicates that conditional on dividends and price, earnings
are positively correlated with future stock returns. This suggests that firm-level earnings contain
information about expected future cash flows beyond dividends. In fact, the bias-adjusted slope
coefficient on normalized dividend is found to be statistically insignificant with the t-statistic of
0.57 and the bias-adjusted p-value of 0.31. The last regression in Panel A of Table 6 shows that
after controlling for RRELt , TERMt , and DEFt , normalized earnings and normalized price remain
highly significant, whereas normalized dividend is statistically insignificant.

Firm-level earnings are affected by firm-specific and macroeconomic events. Therefore, in order
to investigate the importance of firm-specific information in forecasting the time-series of stock
returns, we decompose earnings into a firm-specific and systematic part by regressing firm-level
normalized earnings on aggregate normalized earnings:

(7) ei,t − ni,t = υ i + γ(eagg agg

i,t − ni,t ) + εi,t ,

where ei,t is the log of realized earnings of firm i known as of time t and ni,t is the log of normalization
variable for firm i, eagg agg
i,t and ni,t are aggregate log earnings and aggregate normalization variable

15
respectively, and eagg
i,t is defined as the log of sum of all the available firm-level earnings as of

time t. In equation (7), γ(eagg agg

i,t −ni,t ) is the portion of normalized firm-level earnings explained by

normalized aggregate-level earnings (systematic earnings), denoted by eexp exp

i,t −ni,t , whereas (υ i +εi,t )

is the portion of normalized firm-level earnings orthogonal to normalized aggregate-level earnings

(unsystematic earnings), denoted by eorth orth
i,t −ni,t .

By simultaneously decomposing earnings as suggested in equation (7) and using eexp exp
i,t −ni,t and

eorth orth
i,t −ni,t as explanatory variables in predictive regressions of the form given in equation (6), we

investigate which portion of firm-level earnings is informative about expected stock returns. To
the extent that unsystematic earnings are informative about pcf , we expect eorth orth
i,t −ni,t to have a

positive and significant coefficient in predictive regressions.

Panel B of Table 6 reports the parameter estimates and the related statistics from the panel
data regressions with normalized price, systematic earnings, eexp exp
i,t −ni,t , and unsystematic earnings,

eorth orth
i,t −ni,t , as explanatory variables. The results suggest that it is the unsystematic portion of

earnings which forecasts stock returns, while the coefficient estimate on systematic earnings is
insignificant. Similar to our earlier findings, the bias-adjusted coefficient on normalized price is
negative and significant. The bias-adjusted slope coefficient on (eorth orth
i,t −ni,t ) is 0.020 with the t-

statistic of 6.33 and the bias-adjusted p-value of zero. The bias-adjusted slope on (eexp exp
i,t −ni,t ) is

0.004 and statistically insignificant.

C Firm-Level Cross-Sectional Regressions

We have so far tested the significance of a relation between expected returns and firm-level, and
market-level earnings using time-series regressions. We will now examine the cross-sectional rela-
tion between firm-level earnings and expected stock returns using the Fama and MacBeth (1973)
regressions. There are many advantages of using the firm-level cross-sectional regressions in our
context. First, the Fama-MacBeth regressions will pick up time-series predictability if it really
exists. Hence, this gives us an opportunity to check the robustness of our earlier findings based
on the predictive panel data regressions. Second, there is no small sample bias in cross-sectional
regressions since bias afflicts only time-series regressions. Third, one may be concerned about the
impact of survivorship bias on our time-series tests, arising from the requirement that firms have at

16
least 16 quarters of data. This issue can be resolved using the Fama-MacBeth regressions because
a firm can be included in a cross-sectional regression as long as it has a valid data point.

In this section, we present the time-series averages of the slope coefficients from the cross-
sectional regressions of average stock returns on the lagged earnings yield, dividend yield, dividend
payout ratio, normalized earnings, normalized dividends, and normalized price. The average slopes
provide standard Fama-MacBeth tests for determining which explanatory variables have a signif-
icant predictive power for the cross-section of expected stock returns. Quarterly cross-sectional
regressions are run for the following econometric specification:

(8) ri,t+1 = αt + β t Xi,t + εi,t+1 ,

where ri,t+1 is the log excess return for stock i at time t + 1, and Xi,t is the log value-to-price ratio
or normalized earnings, dividend, and price for stock i at time t. In equation (8), the intercepts
(αt ) and slope coefficients (β t ) are estimated for each quarter using OLS. The time-series average
_
of the slope coefficients, β t , along with its standard error determines whether there is a positive
and significant relation between earnings and the cross-section of expected returns at the firm level.
We use the Newey and West (1987) adjusted standard errors to compute the statistical significance
_ P
of β t = T1 Tt=1 β t , where T is the number of quarters in our sample. This procedure calculates
the standard error of the time-series average of the estimated slope coefficients in equation (8) by
taking into account heteroscedasticity and autocorrelation in the time-series of the slopes.

In Panel A of Table 7, the first row reports the time series averages of the slope coefficients,
the second row presents the Newey-West adjusted t-statistics in parentheses, and the third row
shows the corresponding p-values in square brackets. The univariate regression results indicate a
positive and significant relation between firm-level earnings yield and the cross-section of expected
stock returns since the average slope from the quarterly regressions of realized returns on earnings
yield is about 0.018 with the Newey-West t-statistic of 5.67. It is important to note that the
magnitude and statistical significance of average slope coefficients are similar to our earlier findings
from the predictive panel data regressions. As shown in Table 5, the small sample bias-adjusted
slope coefficient is 0.030 with the t-statistic of 7.16. After controlling for other variables, the
bias-adjusted slope coefficient becomes 0.022 with the t-statistic of 5.26.

17
 The second regression in Panel A compares the forecasting powers of earnings and dividend
yield. The average slope coefficient on dividend yield is insignificant with a t-statistic of 1.08,
whereas the average slope on earnings yield is positive and highly significant with the t-statistic of
7.13. The third regression in Panel A shows that both dividend yield and dividend payout ratio
have significant predictive power for the cross-section of expected returns. However, as will be
presented in Panel B of Table 7, the statistical significance of dividend yield and dividend payout
ratio stems from the strong predictive power of firm-level earnings and price.

The existing literature provides evidence for a significantly positive relation between earnings
yield and the cross-section of expected returns (see, e.g., Basu (1977, 1983), Fama and French
(1992), and Lakonishok, Shleifer and Vishny (1994)).17 Fama and French (1992) find the coeffi-
cient on earnings yield to be significant (with t-stat=4.57) in univariate cross sectional regressions.
However, in multivariate regressions, earnings yield loses its significance. Contrary to Fama and
French (1992), Lakonishok, Shleifer and Vishny (1994) find the coefficient on earnings yield to be
positive and significant in both univariate and multivariate cross-sectional regressions.18 Note that
these studies use annual earnings and in most cases match the realized earnings to future returns
long after the earnings are announced. Hence, they practically analyze long term predictability.
In this paper, we use the latest reported quarterly earnings, and hence we have more time-series
variation in earnings. Moreover, earnings and future returns are matched in a much more timely
manner and one quarter ahead returns are forecasted. Furthermore, our sample period is longer as
compared to earlier studies.

Although, our focus is different from the existing literature on many dimensions, we should
control for the commonly used variables in the literature. Both Fama and French (1992) and
Lakonishok Shleifer and Vishny (1994) use book-to-price and size in their regressions. In the fourth
regression of Panel A, we use log book-to-price and log size as independent variables. Observe that
both variables have statistically significant coefficients. We conclude that book-to-price ratio has
some forecasting power even for one quarter ahead returns.19 In the fifth regression of Panel A,
17
Also see Haugen and Baker (1996) and Fama and French (2006).
18
Fama and French (1992) use monthly returns in their cross-sectional regressions from July 1963 to December
1990, where the returns are matched with annual earnings from the previous calendar year. Lakonishok, Shleifer and
Vishny (1994) use annual returns in their cross-sectional regressions from 1968 to 1989, where the returns are again
matched with annual earnings from the previous calendar year.
19
The existing literature use book-to-market ratio to forecast one year ahead returns starting from at least six
months after the measurement of the book value to one and a half year after. Furthermore, these studies document

18
we use book-to-price and size as control variables and show that earnings yield is still significant
in the presence of these control variables, whereas book-to-price ratio loses significance. The sixth
regression of Panel A shows similar evidence in our smaller dividend sample. Earnings yield and
size variables are the only significant variables.

In the firm level tests, our main variable is the normalized quarterly earnings. This raises the
concern that the tests might be picking up the well-known post-earnings announcement drift in
returns. Therefore, we need to control for the earnings momentum. For the same reason, we also
should control for the price momentum. Thus, we use standardized unexpected earnings and lagged
compounded returns as additional control variables.

There is a long literature on the post-earnings announcement drift (see, e.g., Ball and Brown
(1968), Bernard and Thomas (1989,1990)). It is defined as the tendency for a stock’s price to
drift in the direction of an earnings surprise in the months following an earnings announcement.
To control for the post-earnings announcement drift, we utilize the most commonly used earnings
surprise variable: Standardized Unexpected Earnings (SUE). We follow a wide array of papers and
compute the SUE as unexpected earnings scaled by the standard deviation. Specifically, we define
SUE= ( eit −e it−4
σ it ), where eit is the latest reported quarterly earnings of stock i at quarter t, eit−4
is its value four quarters ago, and σ it is the standard deviation of unexpected earnings over the
past eight quarters. Furthermore, in results that are not reported here to save space, we find that
i) scaling the unexpected earnings by price instead of the standard deviation, ii) substracting a
drift term from the unexpected earnings, and iii) calculating the standard deviation of unexpected
earnings over the past four quarters instead of eight quarters do not change our qualitative findings.

To control for the price momentum, we compute cumulative returns over the prior six months
excluding the current month. Specifically, at the end of month t, we define MOM as the cumulative
six month return between months t − 6 to t − 1. Similarly, in results that are not reported here,
we find that i) using the cumulative returns over the prior three months and ii) not skipping the
current month do not change our qualitative findings about the normalized earnings.20 Further-
more, following Bernard and Thomas (1990), Chan, Jegadeesh and Lakonishok (1996), Mendenhall
(2004) and others, we first express SUE and MOM variables in terms of their ordinal ranking and
that the forecasting power of book-to-price ratio is stronger for two to three-year ahead returns. Hence, using
book-to-price ratio in quarterly regressions yields different results in cross-sectional analysis.
20
The results from alternative specifications of price momentum and SUE are available upon request.

19
scale them to lie between zero and one. As widely discussed in the literature, this has the advantage
of accounting for possible nonlinearities in the relation.

The last regression in Panel A of Table 7 shows that even after controlling for standardized unex-
pected earnings and momentum variables, earnings yield has a positive and statistically significant
coefficient.

As discussed earlier and shown in Panel A of Table 6, when normalized price and normalized
earnings are used as independent variables in the predictive panel regressions, scaled stock price
has a significantly negative coefficient while earnings has a significantly positive coefficient. Also,
conditional on dividends and price, earnings is positively correlated with future stock returns. To
check whether the magnitude and statistical significance of the slope coefficients from time-series
regressions are robust, we run the firm-level cross-sectional regressions with normalized variables.

Panel B of Table 7 shows that when normalized price and normalized earnings are used as
independent variables in the cross-sectional Fama-MacBeth regressions, the average slope coefficient
on normalized earnings is 0.015 with the Newey-West t-statistic of 6.39 and the average slope
coefficient on normalized price is -0.018 with the t-statistic of -4.84. This result is very similar to
our estimates from the predictive panel data regressions. As shown in Panel A of Table 6, the small
sample bias-adjusted slope coefficient on normalized earnings is 0.019 with the t-statistic of 5.80.
After controlling for other variables, the bias-adjusted slope coefficient becomes 0.016 with the
t-statistic of 4.14. The second regression in Panel B indicates that the average slope coefficient on
normalized dividend is insignificant, whereas the average slope coefficients on normalized earnings
and normalized price are highly significant. This result is also very similar to what we obtain
from the predictive panel data regressions and provides evidence that firm-level earnings have
much better predictive power for future returns than firm-level dividends. The third and fourth
regressions in Panel B use book-to-price ratio and size as control variables. After controlling for
these variables, firm level earnings remains to be highly significant. Finally in the last regression of
Panel B, we control for post earnings announcement drift and momentum, and show that normalized
earnings has a positive and highly significant coefficient after controlling for these additional control
variables.21
21
In additional analysis, we also run estimations where each explanatory variable is expressed in terms of its ordinal
ranking. This has the additional benefit of expressing all the explanatory variables on a common scale, so that their
coefficients can be directly compared. We find that, in a multivariate setting, the coefficient on SUE is 0.027 and the

20
D Robustness Checks

In this section, we test the sensitivity of our results to the choice of a normalization variable.
Panel A of Appendix repeats our predictive panel data regressions using sales, book value, and
total liabilities as alternative normalization variables. As shown in Panel A, for all normalization
variables, scaled stock price has a negative and significant coefficient, and normalized earning
has positive and significant coefficient. Therefore, the results are not sensitive to the choice of
the normalization variable. Panel B of Appendix provides very similar results from the cross-
sectional Fama-MacBeth regressions. For all normalization variables, the average slope coefficients
on normalized earnings and normalized price turn out to be significantly positive and significantly
negative, respectively. The most striking observation in Appendix is the affinity of results from
the cross-sectional Fama-MacBeth and time-series predictive panel regressions. For example, when
sales is used as a normalization variable, the average slope coefficient on earnings is 0.016 with
the t-statistic of 7.08, which is almost identical to the bias-adjusted slope coefficient from the
predictive panel regressions: 0.017 with the t-statistic of 5.30. The cross-sectional and predictive
panel regression results are very similar for book value and total liabilities as well.

While panel data regressions result in precise estimations due to the availability of large number
of observations, the coefficient estimates are restricted to be the same across firms. In the present
context, this implies that all types of stocks have the same magnitude of mean-reversion and the
earnings’ correlations with expected stock returns are also the same for all stocks. At an earlier stage
of the study, we relax this assumption by using portfolio level fixed-effects panel data regressions.
We sort stocks into ten size (market capitalization) deciles at the end of each quarter. We also sort
stocks into ten book-to-market deciles. Then, we run the fixed-effects panel data regressions for
each portfolio. This procedure does not restrict coefficients to be the same for stocks in different
portfolios. For each size and book-to-market portfolio, we find a positive and significant relation
between normalized earnings and expected stock returns. Consistent with our earlier results, the
coefficient on normalized price is found to be negative and significant.22
cofficient on normalized earnings is 0.40 and both are highly significant. Thus, the qualilative results are similar to
those reported in the paper. We do not show these results to save space, they are available upon request.
22
To save space, we do not present the parameter estimates and related statistics from the panel data regressions
within each size and book-to-market deciles. They are available upon request.

21
 In addition, we run a battery of further robustness checks. First, we repeat all of our cross-
sectional analyses using raw returns and conclude that if raw returns are used instead of log returns,
our qualitative results and conclusions do not change. Second, we address the survival bias/backfill
bias on Compustat by running the cross-sectional regressions based on a new sample obtained by
eliminating the first three years of each stock (i.e., the first 12 quarters). We repeat our analysis in
Table 7 with this new sample and find a positive and significant relation between firm-level earnings
and expected returns after controlling for size, book-to-market, momentum, and post-earnings
announcement drift. Third, to be consistent with Lamont (1998), we use quarterly earnings but
annual dividends. As a robustness check, we generate a new sample that uses quarterly dividends
instead of annual dividends and rerun our estimations in Table 7. The results indicate that when
quarterly dividends are used, dividend payout ratio, dividend yield and normalized dividends have
very similar coefficient estimates and significance levels. Finally, as mentioned earlier, we winsorize
the top and bottom 0.5% of the variables by pooling all of the data. As a further robustness check,
we winsorize the top and bottom 0.5% of the variables every quarter and repeat our estimations
in Table 7 and find that none of the results changes quantitatively. Because of space constraints,
we cannot report the findings related to i) the survival bias/backfill bias; (ii) raw returns vs. log
returns; (iii) quarterly dividends vs. annual dividends; and (iv) winsorizing the data by quarter.
They are available upon request.

VI Industry-Level Earnings and Expected Returns

We have so far provided evidence that there is no significant relation between aggregate earnings
and expected returns at the market level. Whereas, the relation between earnings and expected
returns is positive and highly significant at the firm-level. Then, a natural question is whether
there is any predictability at the industry-level. In this section, we investigate whether industry-
level earnings can forecast industry returns or whether industry-level diversification is sufficient to
eliminate predictability.

We use the 17 and 48 industry portfolios of Fama and French (1997) and follow the econometric
methodology of Jones, Kaul and Lipson (1994) to examine the time-series relation between earnings

22
and expected returns at the industry level. The following predictive regression is run for each
industry portfolio:

(9) rind ind

i,t+1 = αi + β i Xi,t + εi,t+1 ,

where rind ind

i,t+1 is the log excess return on industry i at time t + 1, and Xi,t is a vector of predictive

variables which are in the information set as of time t. In equation (9), the intercepts (αi ) and slope
coefficients (β i ) are estimated separately for each industry i using OLS. The average slope coefficient
along with its standard error determines whether there is a positive and significant relation between
earnings and expected returns at the industry level. We use the estimation procedure introduced
_ P
by Jones, Kaul, and Lipson (JKL) to compute the standard error of β i = n1 ni=1 β i , where n
equals 17 or 48. This procedure calculates the standard error of the cross-sectional mean of the
estimated slope coefficients in equation (9) by taking into account any cross-sectional correlation
in the individual industry level slope coefficients.23

One disadvantage of using individual time-series regressions is the issue of small sample bias.
As argued by Stambaugh (1999), there exists a small sample bias in predictive regressions of the
sort given in equation (9) because the regression disturbances are correlated with the regressors’
innovations, hence the expectation of the regression disturbance conditional on the future values
of regressors no longer equals zero (also see Lewellen (2004) and Amihud and Hurvich (2004)).24
Therefore, we consider the randomization technique of Nelson and Kim (1993) to correct for the
small sample bias.

We first run the predictive regression and estimate a first-order autoregression for the inde-
pendent variables. Then, we record all the residuals. Finally, we randomize the residuals of the
first-order autoregression to create pseudo-independent variables and returns that have similar
time-series properties as the actual series but they have been generated under the null of no pre-
dictability. The pseudo stock return is generated as the unconditional mean plus the randomized
error term. In each simulation, residuals from the predictive regression and the autoregressions
for the independent variables are randomized simultaneously, hence the correlation that drives the
23
For the calculation of standard errors of the average slope coefficients, see Jones, Kaul, and Lipson (1994, pp.646-
650).
24
The small sample bias indicated by Stambaugh (1999) is a function of the bias of the autoregressive coefficients
of the independent variables, the correlation between the error terms, and the sample size. The sign of the bias
depends on the sign of the correlation between the error terms. If the regression disturbance is positively (negatively)
correlated with the regressor’s innovation, there is a negative (positive) bias.

23
Stambaugh bias is preserved. We repeat this randomization procedure 1000 times and use the JKL
methodology in each cycle and record 1000 average JKL slopes and JKL t-statistics. Consequently,
the mean of these 1000 slope vectors gives the bias in coefficients.

Panels A and B of Table 8 present the average slope coefficients and their statistical significance
for the 17-industry portfolios. In each regression, the first row reports the small sample bias-
adjusted average slope coefficients, the second row presents the JKL t-statistics in parentheses,
and the third row gives the small sample bias-adjusted p-values in square brackets.25 The last
two columns report the bias-adjusted R2 values and the number of observations in industry-level
regressions. In Panel A, we investigate the relation between value-to-price ratios and expected
returns. The qualitative results from the JKL t-statistics and the bias-adjusted p-values are similar
and point out the same conclusion that there is no significant relation between earnings yield and
expected returns on 17-industry portfolios. We consider two regressions, one that uses dividend
payout ratio and the other using earnings yield in a multivariate setting. Indeed, these regressions
are identical in the sense that the coefficient estimate of dividend payout ratio is equal to -1 times
the coefficient estimate of earnings yield.26 The results in Panel A also indicate that, for the 17-
industry portfolios, neither dividend yield nor dividend payout ratio has a significant predictive
power for expected returns.

Panel B investigates whether normalized earnings, normalized dividends and normalized price
have any predictive power for expected returns after controlling for lagged return and RREL. Note
that the average slope coefficients on normalized dividend and normalized price are statistically
insignificant, whereas, the average slope coefficient on normalized earnings is marginally significant
at best with the t-statistic of 1.69 and p-value of 0.06.

Overall, the results in Panel A show that the aggregation of firm level earnings to 17-industry
portfolios diversifies away the information content of firm-level earnings about future cash flows
and hence there is no significant relation between industry-level earnings and expected returns. To
further examine whether industry earnings can forecast future returns, we repeat similar estimations
25
Small sample bias-adjusted p-values are computed as the percentage of times the simulated t-statistics are higher
than the sample t-statistics. In other words, a p-value of 0.99 (0.01) shows that the coefficient is negative (positive)
and significant at the 2% level. However, in our tables, for negative coefficients we report the bias-adjusted p-value
as 0.01 instead of 0.99 to be consistent with the results based on the cross-sectional regressions.
26
Small differences in the estimates are due the fact that we run a separate simulation for each one of our specifica-
tions. Note that the differences in coefficient estimates are very small, indicating the robustness of our randomization
procedure.

24
for 48-industry portfolios. We expect that earnings for a finer industry partition may contain
significant information about future cash flows and this information is partially diversified away,
but not completely like we observed for the 17-industry portfolios. In other words, we expect the
degree of aggregation of earnings to be significantly different for 17 and 48-industry portfolios.

Panel C of Table 8 indicates a positive and significant relation between earnings and expected
returns for 48-industry portfolios. Observe that the earnings yield has a coefficient of 0.040 with
a JKL t-statistic of 3.01 and p-value of 0.01. Furthermore, note that dividend yield’s coefficient
estimate decreases considerably from 0.041 to -0.001, when earnings yield is used in the multivariate
regression instead of dividend payout ratio. This finding indicates that the high coefficient estimate
of dividend yield in the first specification is mostly due to price and that the significant forecasting
power of dividend payout ratio is mostly driven by earning’s forecasting power. Indeed the results
in Panel D clarify that point. In Panel D, we investigate whether normalized earnings, normalized
dividends and normalized price have any predictive power for expected returns after controlling
for lagged return and RREL. We find that the average slope coefficient on normalized earnings is
highly significant (with a JKL t-statistic of 3.19 and p-value of 0.01), whereas the average slope on
normalized dividend is insignificant confirming our previous interpretation.

VII Conclusion

Previous research finds that aggregate earnings yield has little or no forecasting power for aggregate
stock returns. One earlier interpretation of this finding is that the higher variability of earnings
(compared to other cash flow proxies such as dividends) is noise, unrelated to expected returns.
To provide an alternative explanation, Lamont (1998) takes another look at the relation between
aggregate earnings and excess market return, and finds that earnings are positively correlated with
business conditions and hence are negatively correlated with expected returns. He indicates that
(E/P) fails to forecast aggregate stock returns because forecasting powers of earnings (E) and price
(P) are offsetting.

This paper provides evidence that contrary to the findings of Lamont (1998), aggregate earnings
and dividend payout ratio do not contain any information about future returns. We show that his

25
empirical results hold for the specific sample period he uses. His conclusions disappear when we
use the extended versions of his data set.

We investigate the time-series predictability of firm-level stock returns and show that earnings
yield can explain the time-series variation in individual stock returns. It is the mean reversion
in stock prices as well as the earnings’ correlation with expected stock returns that is responsible
for the forecasting power of earnings yield. We decompose firm-level earnings into systematic and
unsystematic components, and find that the correlation of unsystematic (firm-specific) earnings
with expected stock returns leads earnings to have a significant predictive power. However, the
systematic portion of firm-level earnings is not significantly correlated with expected returns. These
results indicate that when firm-level earnings are aggregated to generate the market-level earnings,
the information content of firm-level earnings about future cash flows diversifies away. Hence, the
aggregate-level earnings do not have any explanatory power for the excess market returns.

We also examine the cross-sectional relation between firm-level earnings and expected stock
returns using the Fama and MacBeth (1973) regressions to check the robustness of our findings
obtained from the stacked time-series regressions. The average slope coefficients on earnings indicate
a positive and significant relation between firm-level earnings and the cross-section of expected
returns. This result holds even after controlling for size, book-to-market, momentum, and post-
earnings announcement drift.

Since there is a significantly positive relation between earnings and expected returns at the
firm-level, but the relation is flat at the market-level, we test if there is any predictability at
the industry-level. The results indicate that the aggregation of firm-level earnings to 17-industry
portfolios diversifies away the unsystematic portion of firm-level earnings and hence there is no
significant relation between industry-level earnings and expected returns. However, when we extend
the industry partition from 17 to 48, the weak positive relation between earnings and expected
returns becomes strongly positive. Basically, the results from the market, industry, and firm level
regressions are consistent with each other. There is no predictability at the market-level and no
predictability when the market portfolio is decomposed into 17-industry portfolios. However, when
the market is decomposed into 48-industry, we find a significantly positive relation between earnings
and expected returns. The strong positive relation remains intact when the market portfolio is
decomposed into individual firms.

26
References

[1] Ang, A., and G. Bekaert. “Stock Return Predictability: Is It There?” Review of Financial Studies,
forthcoming (2006).

[2] Amihud, Y., and C. M. Hurvich. “Predictive Regressions: a Reduced-Bias Estimation Method.” Journal
of Financial and Quantitative Analysis, 39 (2004), 813-841.

[3] Ball, R., and P. Brown. “An Emprical Evaluation of Accounting Numbers.” Journal of Accounting
Research, 6 (1968), 159-178.

[4] Basu, S. “Investment Performance of Common Stocks in Relation to their Price-Earnings Ratios: A
Test of the Market Efficiency Hypothesis.” Journal of Finance, 32 (1977), 663-682.

[5] Basu, S. “The Relationship between Earnings’ Yield Market Value and Return for NYSE Common
Stocks.” Journal of Financial Economics, 12 (1983), 129-156.

[6] Berk, J. “A Critique of Size-Related Anomalies.” Review of Financial Studies, 8 (1995), 275-286.

[7] Bernard, V., and J. Thomas. “Post-Earnings Announcement Drift: Delayed Price Response or Risk
Premium?” Journal of Accounting Research, Suplement 27 (1989), 1-36.

[8] Bernard, V., and J. Thomas. “Evidence that Stock Prices do not Fully Reflect the Implications of
Current Earnings for Future Earnings” Journal of Accounting and Economics, 13 (1990), 305-340.

[9] Campbell, J. Y. “A Variance Decomposition for Stock Returns.” Economic Journal, 101 (1991), 157-
179.

[10] Campbell, J. Y., and R. J. Shiller. “Stock Prices Earnings, and Expected Dividends.” Journal of
Finance, 43 (1988), 661-676.

[11] Campbell, J. Y., and R. J. Shiller. “The Dividend-Price Ratio and Expectations of Future Dividends
and Discount Factors.” Review of Financial Studies, 1 (1989), 195-228.

[12] Chan, L., N. Jegadeesh., and J. Lakonishok. “Momentum Strategies.” Journal of Finance, 51 (1996),
1681-1713.

[13] Dechow, P., and R. Sloan. “Returns to Contrarian Investment Strategies: Test of Naive Expectations
Hypothesis.” Journal of Financial Economics, 43 (1997), 3-27.

[14] Easton, P. D., and T. S. Harris. “Earnings as an Explanatory Variable for Returns.” Journal of Ac-
counting Research, 29 (1991), 19-36.

[15] Fama, E. F., and K. R. French. “Dividend Yields and Expected Stock Returns.” Journal of Financial
Economics, 22 (1988a), 3-25.

27
[16] Fama, E. F., and K. R. French. “Permanent and Transitory Components of Stock Prices.” Journal of
Political Economy, 96 (1988b), 246-273.

[17] Fama, E. F., and K. R. French. “Business Conditions and Expected Returns on Stocks and Bonds.”
Journal of Financial Economics, 25 (1989), 23-49.

[18] Fama, E. F., and K. R. French. “Cross-Section of Expected Stock Returns.” Journal of Finance, 47
(1992), 427-465.

[19] Fama, E. F., and K. R. French. “Industry Costs of Equity.” Journal of Financial Economics, 43 (1997),
153-193.

[20] Fama, E. F., and K. R. French. “Profitability, Investment, and Average Returns.” Journal of Financial
Economics, forthcoming (2006).

[21] Fama, E. F., and J. MacBeth. “Risk and Return: Some Empirical Tests.” Journal of Political Economy,
81 (1973), 607-636.

[22] Ferson, W. E., and C. R. Harvey. “The Variation of Economic Risk Premiums.” Journal of Political
Economy, 99 (1991), 385-415.

[23] Ferson, W. E.; S. Sarkissian; and T. Simin. “Spurious Regressions in Financial Economics?” Journal
of Finance, 58 (2003), 1393-1414.

[24] Goyal, A., and I. Welch. “A Comprehensive Look at the Empirical Performance of Equity Premium
Prediction.” Review of Financial Studies, forthcoming (2005).

[25] Haugen, R. A., and N. L. Baker. “Commonality in the Determinants of Expected Stock Returns.”
Journal of Financial Economics, 41 (1996), 401-439.

[26] Hodrick, R. J. “Dividend Yields and Expected Stock Returns: Alternative Procedures for Inference and
Measurement.” Review of Financial Studies, 5 (1992), 357-386.

[27] Jones, C. M.; and G. Kaul; and M. L. Lipson. “Transactions, Volume, and Volatility.” Review of
Financial Studies, 7 (1994), 631-651.

[28] Keim, D. B., and R. F. Stambaugh. “Predicting Returns in the Stock and Bond Markets.” Journal of
Financial Economics, 17 (1986), 357-390

[29] Kothari, S. P.; J. Lewellen; and J. B. Warner. “Stock Returns, Aggregate Earnings Surprises, and
Behavioral Finance.” Journal of Financial Economics, 79 (2006), 537-568.

[30] Kothari, S. P., and J. Shanken. “Book-to-Market, Dividend Yield, and Expected Market Returns: A
Time-Series Analysis.” Journal of Financial Economics, 44 (1997), 169-203.

28
[31] La Porta, R. “Expectations and Cross-Section of Stock Returns.” Journal of Finance, 51 (1996), 1715-
1742.

[32] Lakonishok, J.; A. Shleifer; and R. W. Vishny. “Contrarian Investment, Extrapolation, and Risk.”
Journal of Finance, 49 (1994), 1541-1578.

[33] Lamont, O. “Earnings and Expected Returns.” Journal of Finance, 53 (1998), 1563-1587.

[34] Lettau, M., and S. Ludvigson. “Consumption, Aggregate Wealth, and Expected Stock Returns.” Journal
of Finance, 56 (2001), 815-849.

[35] Lewellen, J. “Predicting Returns with Financial Ratios.” Journal of Financial Economics, 74 (2004),
209-235.

[36] Mendenhall, R. “Arbitrage Risk and Post-Earnings-Announcement Drift.” Journal of Business, 77

(2004), 875-894.

[37] Nelson, C. R., and M. J. Kim. “Predictable Stock Returns: The Role of Small Sample Bias.” Journal
of Finance, 48 (1993), 641-661.

[38] Newey, W. K., and K. D West. “A Simple, Positive Semi-Definite, Heteroskedasticity and Autocorrela-
tion Consistent Covariance Matrix.” Econometrica, 55 (1987), 703-708.

[39] Pontiff, J., and L. D. Schall. “Book-to-Market Ratios as Predictors of Market Returns.” Journal of
Financial Economics, 49 (1998), 141-160.

[40] Rogers, W. H. “Analyzing Complex Survey Data.” Rand Corporation memorandum, Santa Monica,
CA (1983).

[41] Rogers, W. H. “Regression Standard Errors in Clustered Samples.” Stata Technical Bulletin Reprints
STB-13, STB-18 (1993), 88-94.

[42] Rozeff, M. “Dividend Yields are Equity Risk Premiums.” Journal of Portfolio Management, 11 (1984),
68-75.

[43] Shao, J., and J. N. K. Rao. “Jackknife Inference for Heteroscedastic Linear Regression Models.” Cana-
dian Journal of Statistics, 21 (1993), 377-385.

[44] Shiller, R. J. “Stock Prices and Social Dynamics.” Brooking Papers on Economic Activity, 2 (1984),
457-510.

[45] Stambaugh, R. F. “Predictive Regressions.” Journal of Financial Economics, 54 (1999), 375-421.

29
 Table 1
Summary Statistics

Panel A shows the means, medians, and standard deviations. ri,t is the quarterly log excess return computed as
the quarterly log return in excess of the one-month Treasury bill rate compounded within quarter t; (ei,t −pi,t )
E
is log( Pi,t ) where Ei,t is the most recently reported earnings per share and Pi,t is the stock price at the end of
i,t
Di,t
quarter t; (di,t −pi,t ) is log( P ) where Di,t is the sum of the past four quarters of dividend per share; Ni,t is
i,t
Ei,t
a normalization variable, defined as the latest reported total assets per share for stock i; (ei,t −ni,t ) is log( N );
i,t
Di,t Pi,t
(di,t −ni,t ) is log( N ) and (pi,t −ni,t ) is log( N ). Panel B shows the correlation matrix of these variables. Mean,
i,t i,t
median, standard deviation, and correlation statistics are computed for each stock i and averages of these time-series
statistics are reported.

Panel A
ri,t ei,t −pi,t di,t −pi,t pi,t −ni,t ei,t −ni,t di,t −ni,t

Mean 0.022 -4.029 -3.691 -0.496 -4.519 -4.460

Median 0.022 -3.967 -3.677 -0.491 -4.411 -4.433

Std Dev 0.206 0.620 0.398 0.444 0.639 0.307

Panel B
ri,t ei,t −pi,t di,t −pi,t pi,t −ni,t ei,t −ni,t di,t −ni,t

ri,t 1.00 -0.08 -0.11 0.10 0.05 -0.02

ei,t −pi,t -0.08 1.00 0.44 -0.49 0.27 -0.09

di,t −pi,t -0.11 0.44 1.00 -0.48 -0.19 0.37

pi,t −ni,t 0.10 -0.49 -0.48 1.00 0.70 0.64

ei,t −ni,t 0.05 0.27 -0.19 0.70 1.00 0.63

di,t −ni,t -0.02 -0.09 0.37 0.64 0.63 1.00

30
 Table 2
Forecasting S&P Composite Index Return with Value-to-Price Ratios

This table shows the parameter estimates and standard errors from predictive regressions using value-to-price ratios
for different sample periods. In each regression, dependent variable is the quarterly log excess return on the S&P
Composite Index, rm,t+1 . Independent variables are rm,t , (dt −pt ), (et −pt ), (dt −et ), and RRELt . RRELt is the
relative Treasury bill rate. (dt −pt ) is the log dividend yield, (et −pt ) is the log earnings yield, and (dt −et ) is the log
dividend payout ratio corresponding to the S&P Composite Index. In Panels B and C, the first row reports the small
sample bias-adjusted average slope coefficients, the second row presents the Newey-West t-statistics in parentheses,
and the third row gives the bias-adjusted p-values in square brackets. First row of Panel A also reports the OLS
coefficient estimates for Lamont’s (1998) original sample. The last columns report the bias-adjusted R2 values. a,b,c
indicate significance at the 1%, 5% and 10% level, respectively.

Panel A. (1947Q1-1994Q4)
Constant rm,t RRELt dt −pt dt −et et −pt R2

1 0.222 0.064 5.12 %

0.138a 0.043a 4.38 %
(2.53) (2.57)
[0.00] [0.00]
2 0.062 0.012 0.32 %
0.033 0.008 0.12 %
(0.61) (0.63)
[0.24] [0.25]
3 0.207 0.194 -0.111 12.57 %
0.123b 0.173a -0.110a 11.25 %
(2.14) (3.55) (-3.13)
[0.01] [0.00] [0.00]
4 0.207 0.083 0.111 12.57 %
0.130b 0.065a 0.111a 11.39 %
(2.28) (3.24) (3.15)
[0.01] [0.00] [0.00]
5 -0.042 0.083 4.40%
-0.058b 0.083b 3.84 %
(-2.37) (2.51)
[0.01] [0.01]
6 0.210 0.067 -0.861 0.078 0.086 14.87 %
0.132b 0.061 -0.873c 0.059a 0.085b 12.61 %
(2.45) (1.12) (-1.83) (3.27) (2.45)
[0.01] [0.15] [0.05] [0.00] [0.01]

31
 Panel B. (1947Q1-2002Q4)
Constant rm,t RRELt dt −pt dt −et et −pt R2

1 0.056c 0.017 2.68 %

(1.16) (1.14)
[0.05] [0.06]
2 0.072 0.018 0.92 %
(1.40) (1.41)
[0.08] [0.07]
3 0.018 0.045 -0.032 2.97 %
(0.34) (1.46) (-1.23)
[0.21] [0.11] [0.15]
4 0.020 0.014 0.033 3.06 %
(0.36) (0.87) (1.28)
[0.21] [0.11] [0.11]
5 -0.025 0.037 0.74 %
(-1.33) (1.44)
[0.07] [0.09]
6 0.048 0.054 -1.332b 0.016 0.005 6.05 %
(0.91) (1.00) (-2.51) (1.12) (0.21)
[0.09] [0.17] [0.01] [0.06] [0.43]

Panel C. (1940Q1-2002Q4)
Constant rm,t RRELt dt −pt dt −et et −pt R2

1 0.050 0.015 2.30 %

(1.08) (1.07)
[0.06] [0.07]
2 0.076 0.019 0.86 %
(1.48) (1.50)
[0.09] [0.08]
3 0.016 0.032 -0.022 2.24 %
(0.31) (1.26) (-0.96)
[0.24] [0.15] [0.20]
4 0.024 0.013 0.021 2.31 %
(0.45) (0.86) (0.95)
[0.25] [0.11] [0.20]
5 -0.023 0.033 0.53 %
(-1.35) (1.49)
[0.06] [0.07]
6 0.055 0.078 -1.314b 0.017 -0.001 5.38 %
(1.05) (1.52) (-2.49) (1.24) (-0.06)
[0.09] [0.07] [0.01] [0.06] [0.48]

32
 Table 3
Forecasting S&P Composite Index Return with Aggregate Normalized Variables

This table shows the parameter estimates and standard errors from predictive regressions using normalized prices,
dividends, and earnings for different sample periods. In each regression, dependent variable is the quarterly log
excess return on the S&P Composite Index, rm,t+1 . Independent variables are rm,t , (pt −nt ), (dt −nt ), (et −nt ), and
RRELt . RRELt is the relative Treasury bill rate. dt is the log of dividends per share paid out in the four quarters
including quarter t. et is the log of earnings per share in quarter t. pt is the log of the S&P price level. nt is the log
of the average of the past five years of annual earnings per share, calculated as the sum of the past 20 observations of
quarterly earnings per share divided by five. In Panels B and C, the first row reports the small sample bias-adjusted
average slope coefficients, the second row presents the Newey-West t-statistics in parentheses, and the third row gives
the bias-adjusted p-values in square brackets. First row of Panel A also reports the OLS coefficient estimates for
Lamont’s (1998) original sample. The last columns report the bias-adjusted R2 values. a,b,c indicate significance at

the 1%, 5% and 10% level, respectively.

Panel A. (1947Q1-1994Q4)
Constant rm,t RRELt pt −nt dt −nt et −nt R2

0.209 0.067 -0.859 -0.078 0.163 -0.086 14.87 %

0.109b 0.059 -0.844c -0.050a 0.154a -0.085b 11.81 %
(2.19) (1.10) (-1.78) (-2.61) (3.08) (-2.40)
[0.01] [0.14] [0.05] [0.00] [0.00] [0.02]

Panel B. (1947Q1-2002Q4)
Constant rm,t RRELt pt −nt dt −nt et −nt R2

0.046 0.043 -1.411a -0.003 0.062 -0.001 5.76 %

(0.91) (0.80) (-2.64) (-0.21) (1.70) (-0.03)
[0.07] [0.22] [0.00] [0.22] [0.08] [0.48]

Panel C. (1940Q1-2002Q4)
Constant rm,t RRELt pt −nt dt −nt et −nt R2

0.044 0.072 -1.339b -0.007 0.039 0.004 4.78 %

(0.85) (1.41) (-2.54) (-0.49) (1.08) (0.15)
[0.09] [0.08] [0.01] [0.14] [0.21] [0.42]

33
 Table 4
Forecasting S&P Composite Index Return (Alternative Periods)

This table shows the parameter estimates and standard errors from predictive regressions using value-to-price ratios
and normalized prices, dividends, and earnings for alternative sample periods. In each regression, dependent variable
is the quarterly log excess return on the S&P Composite Index, rm,t+1 . Independent variables are defined in Tables
2 and 3. In each regression, the first row reports the small sample bias-adjusted average slope coefficients, the second
row presents the Newey-West t-statistics in parentheses, and the third row gives the bias-adjusted p-values in square
brackets. The last columns report the bias-adjusted R2 values. a,b,c indicate significance at the 1%, 5% and 10%
level, respectively.

Panel A. (1940Q1-1994Q4)
Constant rm,t RRELt dt −pt dt −et R2

0.093c 0.090c -1.044c 0.036c 0.043 8.57 %

(1.61) (1.69) (-2.08) (2.09) (1.61)
[0.03] [0.05] [0.03] [0.03] [0.06]

Panel B. (1947Q1-1996Q4)
Constant rm,t RRELt dt −pt dt −et R2

0.057 0.081 -0.996c 0.031c 0.048 7.96 %

(1.04) (1.45) (-1.98) (1.63) (1.46)
[0.06] [0.07] [0.03] [0.05] [0.09]

Panel C. (1940Q1-1994Q4)
Constant rm,t RRELt pt −nt dt −nt et −nt R2

0.056b 0.096c -0.967c -0.034b 0.061 -0.051c 8.35 %

(0.91) (1.79) (1.96) (-1.97) (1.21) (-1.92)
[0.10] [0.04] [0.03] [0.02] [0.18] [0.04]

Panel D. (1947Q1-1996Q4)
Constant rm,t RRELt pt −nt dt −nt et −nt R2

0.056c 0.081 -1.018c -0.023 0.103c -0.045 7.32 %

(1.03) (1.16) (-1.70) (-1.42) (2.17) (-1.43)
[0.03] [0.08] [0.04] [0.07] [0.04] [0.12]

34
 Table 5
Firm-Level Earnings Yield and Expected Stock Returns

This table shows the parameter estimates and their statistical significance from the fixed-effects panel data regressions.
In each regression, the first row reports the small sample bias-adjusted slope coefficients, the second row presents the
clustered t-statistics in parentheses, and the third row gives the bias-adjusted p-values in square brackets. The last
three columns report the bias-adjusted R2 values, the number of observations (T), and the number of stocks (N). The
dependent variable is the quarterly log excess stock return, ri,t+1 . Independent variables are (ei,t −pi,t ), (di,t −pi,t ),
(di,t −ei,t ), RRELt , TERMt and DEFt . RRELt is the relative Treasury bill rate, DEFt is the default spread, and
TERMt is the term spread. a,b,c indicate significance at the 1%, 5% and 10% level, respectively.

ei,t −pi,t di,t −pi,t di,t −ei,t RRELt TERMt DEFt R2 T N

1 0.030a 1.25% 245,013 6,074

(7.16)
[0.00]

2 0.018a 1.12% 151,030 3,321

(2.25)
[0.00]

3 0.019a 0.009 1.57% 151,030 3,321

(4.98) (1.345)
[0.00] [0.10]

4 0.028a -0.019a 1.57% 151,030 3,321

(2.88) (-4.71)
[0.00] [0.00]

5 0.022a 0.010 -1.422 0.026 -0.082 3.11% 151,030 3,321

(5.26) (1.50) (1.64) (0.04) (-0.04)
[0.00] [0.08] [0.06] [0.18] [0.24]

35
 Table 6
Firm-Level Normalized Variables and Expected Stock Returns

This table shows the parameter estimates and their statistical significance from the fixed-effects panel data regressions.
In each regression, the first row reports the small sample bias-adjusted slope coefficients, the second row presents
the clustered t-statistics in parentheses, and the third row gives the bias-adjusted p-values in square brackets. The
last three columns report the bias-adjusted R2 values, the number of observations (T), and the number of stocks
(N). The dependent variable is the quarterly log excess stock return, ri,t+1 . Independent variables are (ei,t −ni,t ),
(di,t −ni,t ), (pi,t −ni,t ), RRELt , TERMt and DEFt . RRELt is the relative Treasury bill rate, DEFt is the default
exp exp
spread, and TERMt is the term spread. (ei,t −ni,t ) is the portion of normalized firm-level earnings explained
by normalized aggregate-level earnings (systematic earnings), (eorth orth
i,t −ni,t ) is the portion of normalized firm-level
earnings orthogonal to normalized aggregate-level earnings (unsystematic earnings) (see section V.B). a,b,c indicate
significance at the 1%, 5% and 10% level, respectively.

Panel A
pi,t −ni,t ei,t −ni,t di,t −ni,t RRELt TERMt DEFt R2 T N

1 -0.031a 0.019a 1.97% 245,013 6,074

(-3.84) (5.80)
[0.00] [0.00]

3 -0.029a 0.014a 0.003 1.83% 151,030 3,321

(-2.71) (3.79) (0.57)
[0.00] [0.00] [0.31]

7 -0.033a 0.016a 0.004 -1.432 0.017 -0.246 3.40% 151,030 3,321

(-3.00) (4.14) (0.72) (-1.66) (0.03) (-0.12)
[0.00] [0.00] [0.25] [0.06] [0.39] [0.35]

Panel B
pi,t −ni,t (eoi,trth −noi,trth ) (eexp exp
i,t −ni,t ) RRELt TERMt DEFt R2 T N

1 -0.032a 0.020a 0.004 1.99% 245,013 6,074

(3.84) (6.33) (0.03)
[0.00] [0.00] [0.35]

2 -0.033a 0.022a 0.123 -2.106b -0.152 -0.873 3.40% 245,013 6,074

(-4.11) (6.70) (0.85) (-2.23) (-0.21) (-0.40)
[0.00] [0.00] [0.21] [0.02] [0.41] [0.30]

36
 Table 7
Cross Sectional Regressions

This table shows the parameter estimates and their statistical significance from the firm level cross-sectional Fama-
MacBeth regressions. In each regression, the first row reports the time series averages of the slope coefficients,
the second row presents the Newey-West (1987) adjusted t-statistics in parentheses, and the third row shows the
corresponding p-values in square brackets. Panel A reports the statistics for the cross-sectional regressions with the
value-to-price ratios. Panel B reports the statistics for the cross-sectional regressions with the normalized variables.
SUE is the unexpected earnings (the change in the most recently reported quarterly earnings from its value four
quarters ago) scaled by the standard deviation of unexpected earnings over the past eight quarters. MOM is the
compound return over the prior six months excluding the current month. Both SUE and MOM are expressed in
terms of their decile rank and scaled to fall between zero and one. In both panels, log book-to-price ratio (bit −pit )
and log of market value (lmei,t ) are used as additional control variables. The last three columns report the average
R2 values, the number of observations (T), and the number of stocks (N). a,b,c indicate significance at the 1%, 5%
and 10% level, respectively.

Panel A. Value-to-Price Ratios

ei,t −pi,t di,t −pi,t di,t −ei,t bi,t −pi,t lmei,t SUE MOM R2 T N

1 0.018a 1.54% 289,958 12,824

(5.67)
[0.00]

2 0.014a 0.003 1.57% 167,699 5,635

(7.13) (1.08)
[0.00] [0.28]

3 0.017a -0.014a 3.33% 167,699 5,635

(5.63) (-7.41)
[0.00] [0.00]

4 0.012a -0.003b 4.0% 289,958 12,824

(3.16) (-2.39)
[0.00] [0.02]

5 0.013a 0.005 -0.003b 4.8% 289,958 12,824

(6.39) (1.55) (-2.41)
[0.00] [0.12] [0.02]

6 0.013a 0.003 0.000 -0.003b 6.0% 167,699 5,635

(7.34) (1.12) (0.03) (-2.47)
[0.00] [0.27] [0.98] [0.02]

7 0.011a 0.008a -0.003a 0.021a 0.012b 5.8% 258,616 11,400

(6.25) (2.91) (-2.58) (6.06) (2.37)
[0.00] [0.12] [0.01] [0.00] [0.02]

37
 Panel B. Normalized Variables
pi,t −ni,t ei,t −ni,t di,t −ni,t bi,t −pi,t lmei,t SUE MOM R2 T N

1 -0.018a 0.015a 2.97% 289,958 12,824

(-4.84) (6.39)
[0.00] [0.00]

2 -0.017a 0.014a 0.001 4.63% 167,699 5,635

(-5.21) (7.73) (0.59)
[0.00] [0.00] [0.56]

3 -0.014a 0.012a 0.003 -0.003b 5.90% 289,958 5,635

(-3.64) (6.72) (0.78) (-2.36)
[0.00] [0.00] [0.44] [0.02]

4 -0.016a 0.013a 0.002 -0.001 -0.003b 7.35% 167,699 5,635

(-4.01) (7.41) (0.85) (-0.37) (-2.31)
[0.00] [0.00] [0.40] [0.71] [0.02]

5 -0.011a 0.010a 0.007b -0.003b 0.022a 0.010c 6.80% 258,616 11,400

(-3.00) (6.85) (2.23) (-2.43) (6.58) (1.84)
[0.00] [0.00] [0.03] [0.02] [0.00] [0.07]

38
 Table 8
Industry-Level Earnings and Expected Returns

This table presents the average slope coefficients and their statistical significance based on the Jones, Kaul and Lipson
(JKL, 1994) methodology for the 17 and 48-industry portfolios. In each regression, the first row reports the small
sample bias-adjusted average slope coefficients, the second row presents the JKL t-statistics in parentheses, and the
third row gives the bias-adjusted p-values in square brackets. The last two columns report the bias-adjusted R2
values and the number of observations (T) in industry-level regressions. The dependent variable is the quarterly log
excess return on industry portfolios, rind,t+1 . Independent variables are rind,t , (dt −pt ), (et −pt ), (dt −et ), (pt −nt ),
(dt −nt ), (et −nt ) and RRELt . a,b,c indicate significance at the 1%, 5% and 10% level, respectively.

Panel A. 17-industry portfolios; value-to-price ratios

Constant rind,t RRELt dt −pt dt −et et −pt R2 T

0.110 0.005 -1.467a 0.030 -0.026 3.69 % 1,951

(1.83) (0.08) (-2.75) (1.94) (-1.45)
[0.31] [0.65] [0.01] [0.19] [0.12]

0.099 0.005 -1.466a 0.001 0.026 3.64% 1,951

(1.66) (0.08) (-2.75) (0.09) (1.49)
[0.30] [0.65] [0.01] [0.76] [0.14]

PanelB.17-industryportfolios;normalizedvariables
Constant rind,t RRELt pt -nt dt -nt et -nt R2 T

0.202 0.002 -1.482a -0.027 0.005 0.035 3.75 % 1,951

(1.73) (0.03) (-2.79) (-1.60) (0.31) (1.69)
[0.08] [0.59] [0.01] [0.33] [1.69] [0.06]

Panel C. 48-industry portfolios; value-to-price ratios

Constant rind,t RRELt dt −pt dt −et et −pt R2 T

0.154 -0.010 -1.509a 0.041 -0.039a 4.44 % 5,343

(2.74) (-0.19) (-2.79) (2.84) (-2.93)
[0.43] [0.66] [0.01] [0.16] [0.01]

0.148 -0.009 -1.509a -0.001 0.040a 4.45% 5,343

(2.63) (-0.18) (-2.79) (-0.10) (3.01)
[0.30] [0.67] [0.01] [0.93] [0.01]

Panel D. 48-industry portfolios; normalized variables

Constant rind,t RRELt pt −nt dt −nt et −nt R2 T

0.167 -0.006 -1.470a -0.047 0.001 0.044a 4.46 % 5,343

(2.52) (-0.12) (-2.74) (-3.16) (-0.10) (3.19)
[0.04] [0.65] [0.01] [0.13] [0.64] [0.01]

39
 Appendix
Normalization Variables

This table shows the parameter estimates and their statistical significance from the fixed-effects panel data regressions
and cross-sectional Fama-MacBeth regressions. For each regression in Panel A, the first row reports the small sample
bias-adjusted slope coefficients, the second row presents the clustered t-statistics in parentheses, and the third row
gives the small sample bias-adjusted p-values in square brackets. For each regression in Panel B, the first row reports
the time series averages of the slope coefficients, the second row presents the Newey-West (1987) adjusted t-statistics
in parentheses, and the third row shows the corresponding p-values in square brackets. In each panel, the results
are reported for three different normalization variables; sales per share, book value per share, and total liabilities per
share. a,b,c indicate significance at the 1%, 5% and 10% level, respectively.

Panel A. Predictive Panel Regressions

pi,t −ni,t ei,t −ni,t R2 T N
Ni,t ≡sales per share
1 -0.043a 0.017a 2.62% 269,176 6,380
(-4.55) (5.30)
[0.00] [0.00]

Ni,t ≡book value per share

2 -0.031a 0.015a 1.92% 268,378 6,393
(-2.82) (6.03)
[0.00] [0.00]

Ni,t ≡total liabilities per share

3 -0.033a 0.023a 1.87% 243,040 6,060
(-4.977) (6.272)
[0.00] [0.00]

Panel B. Cross-Sectional Regressions

pi,t −ni,t ei,t −ni,t R2 T N
Ni,t ≡sales per share
1 -0.019a 0.016a 2.87% 313,081 12,948
(-6.31) (7.08)
[0.00] [0.00]

Ni,t ≡bookvaluepershare
2 -0.021a 0.015a 2.65% 312,294 12,984
(-5.18) (8.33)
[0.00] [0.00]

Ni,t ≡total liabilities per share

3 -0.017 0.015 3.16% 287,845 12,790
(-5.39) (6.68)
[0.00] [0.00]

40

View publication stats