1 s2.0 S0304405X06001164 Main 2 PDF

ARTICLE IN PRESS
Journal of Financial Economics 82 (2006) 491518

www.elsevier.com/locate/jfec
Protability, investment and average returns$

Eugene F. Famaa, Kenneth R. Frenchb,
a
Graduate School of Business, University of Chicago, Chicago, IL, 60637, USA
b
Amos Tuck School of Business, Dartmouth College, Hanover, NH, 03755, USA
Received 15 January 2005; received in revised form 9 June 2005; accepted 27 September 2005
Available online 28 July 2006
Abstract
Valuation theory says that expected stock returns are related to three variables: the book-to-
market equity ratio (Bt/Mt), expected protability, and expected investment. Given Bt/Mt and
expected protability, higher expected rates of investment imply lower expected returns. But
controlling for the other two variables, more protable rms have higher expected returns, as do
rms with higher Bt/Mt. These predictions are conrmed in our tests.
r 2006 Elsevier B.V. All rights reserved.
JEL classification: G12
Keywords: Average returns; Protability; Investment; Book-to-market equity
1. Introduction
In the dividend discount model, the market value of a share of a rms stock is the
present value of expected dividends,
X
1
Mt EDtt =1 rt , (1)
t1
where Mt is the price at time t, E(Dt+t) is the expected dividend in period t+t, and r is
(approximately) the long-term average expected stock return or, more precisely, the
$
We acknowledge the helpful comments of Richard Roll, Richard Sansing, Clyde Stickney, and two
anonymous referees.
Corresponding author.
E-mail address: [email protected] (K.R. French).
0304-405X/$ - see front matter r 2006 Elsevier B.V. All rights reserved.
doi:10.1016/j.jneco.2005.09.009
ARTICLE IN PRESS
492 E.F. Fama, K.R. French / Journal of Financial Economics 82 (2006) 491518
internal rate of return on expected dividends. With clean surplus accounting, the time t
dividend, Dt, is equity earnings per share, Yt, minus the change in book equity per share,
dBt BtBt1. The dividend discount model then becomes
X
1
Mt EY tt dBtt =1 rt (2)
t1
or, dividing by time t book equity,
P
1
EY tt dBtt =1 rt
M t t1
. (3)
Bt Bt
Eq. (3) makes three predictions about expected stock returns. (1) Controlling for
expected earnings and expected changes in book equity (both measured relative to current
book equity), a higher book-to-market equity ratio, Bt/Mt, implies a higher expected stock
return, r. This is the motivation for using the book-to-market ratio as a proxy for expected
returns. (2) Controlling for Bt/Mt and expected growth in book equity due to reinvestment
of earnings, more protable rmsspecically, rms with higher expected earnings
relative to current book equityhave higher expected returns. (3) Given Bt/Mt and
expected earnings relative to book equity, rms with higher expected growth in book
equity due to reinvestment of earnings have lower expected stock returns.
We test for the book-to-market, protability, and investment effects in expected returns
predicted by the valuation equation (3). This is not virgin territory. Though our methods
are different, our work can be viewed as providing a unifying perspective on many papers
that link average stock returns to book-to-market equity and proxies for expected
protability and investment.
For example, there is much evidence that rms with higher book-to-market ratios have
higher average stock returns (Rosenberg, Reid, and Lanstein, 1985; Chan, Hamao, and
Lakonishok, 1991; Fama and French, 1992; Capaul, Rowley, and Sharpe, 1993;
Lakonishok, Shleifer, and Vishny, 1994). Haugen and Baker (1996) and Cohen, Gompers,
and Vuolteenaho (2002) nd that, controlling for book-to-market equity, average returns
are positively related to protability. Faireld, Whisenant, and Yohn (2003), Richardson
and Sloan (2003), and Titman, Wei, and Xie (2004) show a negative relation between
average returns and investment. An extensive literature initiated by Sloan (1996) shows
that accruals are negatively related to future protability and that higher accruals predict
lower stock returns. (See Xie, 2001; Faireld, Whisenant, and Yohn, 2003; Richardson,
Sloan, Soliman, and Tuna, 2004, 2005; Chan, Chan, Jegadeesh, and Lakonishok, 2006.)
Working within the connes of a valuation equation like Eq. (2), Abarbanell and Bushee
(1998), Frankel and Lee (1998), Dechow, Hutton, and Sloan (2000), and Lee, Ng, and
Swaminathan (2004) combine analyst forecasts of earnings with assumptions about future
investment to estimate expected stock returns. The general result is that higher expected
net cash ows (expected protability minus expected investment) relative to current market
value forecast higher stock returns. Finally, Piotroski (2000) and Grifn and Lemmon
(2002) show that composite measures of rm strength, which are proxies for expected net
cash ows, are positively related to future stock returns. All these results are in line with
Eq. (3).
ARTICLE IN PRESS
E.F. Fama, K.R. French / Journal of Financial Economics 82 (2006) 491518 493
In this earlier work, evidence that the book-to-market ratio, expected protability, and
expected investment are related to future stock returns is typically attributed to mispricing.
As usual, irrational pricing is not the only possibility. With rational pricing, the book-to-
market, protability, and investment effects in expected returns implied by the valuation
equations are due to differences in risk: Controlling for other variables, more protable
rms and rms with higher book-to-market ratios are more risky, and faster-growing rms
are less risky. We take no stance on whether the patterns in average returns observed here
are rational or irrational. Indeed, one of our themes is that tests based (explicitly or
implicitly) on the valuation equations are generally powerless to determine whether
observed relations between average returns and Bt/Mt, protability, and investment are
due to rational or irrational pricing.
What do we add on the empirical side? Most existing papers look for book-to-market,
protability, or investment effects in average returns and treat them as isolated anomalies.
Our setup says that all this evidence is consistent with the predictions of valuation theory.
Working within the connes of valuation theory makes it clear, however, that cleanly
identifying book-to-market, protability, or investment effects in expected returns requires
controls for the other two variables, which are often missing in earlier tests. Our goal is to
provide an overall perspective on how the three combine to explain the cross section of
average stock returns.
The paper proceeds as follows. Section 2 discusses what tests based on the valuation
equation (3) can and cannot reveal about expected returns and the rationality of asset
prices. Section 3 uses cross-section regressions to develop proxies for expected protability
and investment. We nd that lagged values of many variables, including size, accounting
fundamentals, stock returns, analyst earnings forecasts, and two measures of rm strength
forecast protability and investment. Section 4 uses cross-section return regressions to
examine whether the book-to-market ratio and various proxies for expected protability
and investment (including the tted values from the regressions of Section 3) help explain
average returns in the manner predicted by Eq. (3). These cross-section return regressions
identify book-to-market, protability, and investment effects in average stock returns, but
they do not give a clean picture of their economic importance. Section 5 presents portfolio
tests that address this issue. The concluding Section 6 summarizes our evidence and
inferences.
2. Tests of valuation equations: strengths and weaknesses
Campbell and Shiller (1988) emphasize that the valuation equation (1) is a tautology
that denes the internal rate of return, r. Given the stock price and estimates of expected
dividends, there is a discount rate r that solves Eq. (1). With clean surplus accounting,
Eq. (2) is equivalent to Eq. (1), so Eq. (2) is a tautology. Eq. (3) is obtained by dividing
Eq. (2) by book equity, so Eq. (3) is also a tautology.
Tautology, however, does not mean Eq. (3) lacks content. In fact, the tautology
conclusion confers some robustness on tests that infer the discount rate, r, from Eq. (3).
For example, as long as rms are expected to follow clean surplus accounting in the future,
the past accounting rules that generate book equity, Bt, do not affect inferences about r.
Suppose two all-equity rms have identical current market values and identical expected
future earnings and investments. With clean surplus accounting, we can use Eq. (2) to infer
that the rms must have the same expected return, r. And because we derive Eq. (3) from
ARTICLE IN PRESS
Eq. (2) simply by dividing both sides by current book equity, Eq. (3) also implies they have
the same r even if the two rms assets are carried at different book values. The fact that
they have different Bt cancels out in Eq. (3), leaving the discount rate r unaffected. The
important implication is that if rms are expected to use clean surplus accounting, then our
cross-section tests to estimate how expected returns vary with Bt/Mt, expected
protability, and expected investment are valid, as long the tests control for all three
variables. And this serves to emphasize the importance of joint controls for the three
variables, which are typically missing in earlier work.
Deviations from clean surplus accounting are a potential problem. But there are reasons
to expect that actual deviations are not fatal. First, the transition from Eq. (1) to Eq. (2)
requires clean surplus only in expectation. Firms can deviate from clean surplus as long as
the expected value of future deviations is zero. Second, the intuition behind Eq. (2) is that
if two rms have the same stock price and the same expected growth in book equity, but
one has higher expected earnings, it must have a higher expected stock return (cost of
equity capital). Likewise, if two rms have the same stock price and expected earnings but
one requires more expected equity investment to generate the earnings, it must have a
lower expected stock return. We judge that accounting problems must be severe to obscure
all traces of these predictions. There is evidence that this is not the case. Thus, despite the
vagaries of accounting, the existing literature identies differences in average stock returns
associated with Bt/Mt, expected protability, and expected investment, even without
simultaneous controls for all three.
Now comes perhaps the most important point. Even with clean surplus accounting, tests
of Eq. (3) face a timeworn problem: We cannot tell whether the book-to-market,
protability, and investment effects in average stock returns are due to rational or
irrational pricing. To see the point, note rst that Eqs. (1) to (3) hold (they are tautologies)
whether the expected values of protability and investment in the equations are rational or
irrational. The implied discount rate, r, does vary with the expectations that are used.
When the expected values are rational, r is the discount rate (roughly the true expected
stock return) implied by rational beliefs. When the expected values are irrational, r is the
expected return implied by these irrational beliefs (and it is not the true expected return).
Next consider what we measure. Our estimates of expected protability and investment
(for example, from regressions of future protability and investment on lagged predictors)
are estimates of rational (actual or true) conditional expected values. And our return tests
provide estimates of how rationally assessed (actual or true) expected returns (proxied by
observed average returns) vary with the book-to-market ratio and rational assessments of
expected protability and investment. If the estimates of expected protability and
investment implicit in the pricing of stocks are also rational, then, up to sampling error, the
variation in expected returns we measure corresponds to that predicted by investors.
Suppose, however, that stock prices are based on irrational protability and investment
forecasts, so the book-to-market ratio Bt/Mt contains an irrational price. Eq. (3) still
implies that, as long as we use rational assessments of expected protability and growth,
our tests provide estimates of how true expected returns vary with rational assessments of
expected protability and investment and a book-to-market ratio that contains an
irrational price. In other words, the true expected returns we measure vary in the same way
with rational assessments of expected protability and growth whether or not the price in
Bt/Mt is based on these rational assessments. Irrational beliefs about expected protability
and investment do affect our estimates of true expected returns through their effects on the
ARTICLE IN PRESS
price Mt in Bt/Mt. And here we face the usual conundrum: Denitive statements about
how variation across rms in Bt/Mt in Eq. (3) splits between differences in rational risks
and irrational beliefs are (in our view) impossible. In short, despite common claims to the
contrary in the literature, tests of Eq. (3) cannot in themselves tell us whether the investor
forecasts of protability and investment that determine Mt are rational or irrational.
We revisit this issue throughout the paper.
3. Expected protability and investment
The rst step in our tests of the valuation equation (3) is to develop proxies for expected
protability and investment. The more complicated proxies are tted values from cross-
section regressions to predict protability, Yt+t/Bt, and the growth of assets, dAt+t/
At (At+tAt)/At, one, two, and three years ahead (t 1, 2, 3). The explanatory variables,
measured at the end of scal year t, are accounting fundamentals, the rms stock return for
scal year t and its combined return for years t-1 and t-2, analyst earnings forecasts for t+1,
and the composite measures of rm strength of Piotroski (2000) and Ohlson (1980). We use
the expected protability and asset growth estimates given by the tted values from these
rst-stage regressions as explanatory variables in second-stage cross-section return
regressions that test for protability and investment effects in average returns (Section 4).
The accounting fundamentals used as explanatory variables in the proxies for expected
protability and investment include lagged values of Bt/Mt, a dummy variable for negative
earnings, protability (Yt/Bt) for rms with positive earnings, accruals relative to book
equity for rms with positive (+ACt/Bt) and negative (ACt/Bt) accruals, investment
(dAt/At1), a dummy variable for rms that do not pay dividends (No Dt), and the ratio of
dividends to book equity (Dt/Bt). The book-to-market ratio is known to be negatively
related to protability and investment (rms with lower Bt/Mt tend to be more protable
and to invest more), and protability and investment are known to be persistent (Penman,
1991; Lakonishok, Shleifer, and Vishny, 1994; Fama and French, 1995). It also seems
reasonable that current protability is related to future investment and that current
investment is related to future protability. There is evidence that accruals forecast
protability (Sloan, 1996; Faireld, Whisenant, and Yohn, 2002, 2003; Richardson, Sloan,
Soliman, and Tuna, 2004, 2005). Previous work also shows that dividend-paying rms tend
to be more protable but to grow more slowly (Fama and French, 2001). We include rm
size (the log of total market cap, ln MCt) among the fundamental variables because smaller
rms tend to be less protable (Fama and French, 1995). The precise denitions of the
variables are in the Appendix.
Consistent with the logic of the valuation equations, all accounting variables are on a
per share basis. Throughout the paper, the dating convention is that year t includes the
accounting data for scal yearends in calendar year t. For consistency, the lagged returns
and market cap used in the protability and growth regressions are also measured at the
end of a rms scal year. Finally, the valuation equation (3) calls for equity investment,
dBt+t/Bt, but we measure investment as asset growth, dAt+t/At, which we judge gives a
better picture of investment. And we call Yt+t/Bt protability, but for t41, it clearly is a
mix of current protability and future earnings growth.
The explanatory variables used in the rst-stage regressions to develop proxies for expected
protability and asset growth also include It/Bt, the I/B/E/S consensus forecast of earnings
per share one year ahead (as available at the end of a rms scal year) divided by book equity
ARTICLE IN PRESS
per share at t; PTt, the composite measure of rm strength used by Piotroski (2000) to predict
stock returns; and OHt, the probability of debt default developed by Ohlson (1980) and used
by Grifn and Lemmon (2002) to forecast stock returns. Piotroski (2000) assigns rms binary
scores, 0 (bad) and 1 (good) each year on nine accounting fundamentals (including measures
of protability and past earnings growth). PTt is the sum of a rms scores on the nine
variables at the end of scal year t, with higher values indicating stronger past performance.
OHt is the tted value from Ohlsons (1980) cross-section logit regression (Model 1) that uses
accounting fundamentals for year t to assess the probability of default on debt, with higher
values implying weaker rms. From the construction of PTt and OHt (see Appendix), it is
clear that the two variables are proxies for expected net cash ows (the spread of expected
earnings over investment) in Eq. (3). Finally, I/B/E/S earnings forecasts begin in 1976, and
PTt requires data from cash ow statements, which are not available on Compustat until
1971. The period for most of our tests is 19632003, but tests that use I/B/E/S forecasts or
PTt are limited to periods of data availability.
Tables 1 and 2 show average slopes and their t-statistics for year-by-year cross-section
protability and asset growth regressions, estimated in the manner of Fama and MacBeth
(1973). The tables show results only for the full sample period for each regression, but we
can report that average slopes for the rst and second halves of the sample period support
inferences about the marginal explanatory power of different variables much like those
from the full-period tests.
We drop rms from the tests for several reasons. First, we exclude nancial rms
(Standard Industrial Classication codes between 6000 and 6999). In addition, to be
included in the sample for calendar year t (predicting protability and asset growth for
t+1, t+2, and t+3 in Tables 1 and 2, and predicting returns for July of t+1 to June of t+2 in
Tables 3 and 4), a rm must have Compustat data for year t on book equity, earnings
before extraordinary items, dividends, shares outstanding, and accruals, as well as data for
assets for t and t-1. A rm must also have market cap (price times shares outstanding)
available in the Center for Research in Security Prices (CRSP) database for its (last) scal
yearend in t, December of t, and June of t+1. We exclude rms with negative book equity
in year t. Firms are also deleted from specic regressions if they do not have other data,
such as PTt, OHt, and It/Bt, required for that regression. To avoid inuential observation
problems, we delete a rm from the protability and growth regressions if an explanatory
variable in the regression is outside the 0.5 or 99.5 percentile for that variable in year t. (We
consider only the upper or lower bound for one-sided variables, such as +ACt/Bt, ACt/
Bt, and Dt/Bt.) To avoid undue inuence of small rms, those with total assets less than
$25 million or book equity less than $12.5 million in year t are also excluded. (Using $5
million and $2.5 million as the cutoffs produces similar results.)
When the forecast horizon is more than a year ahead, there is overlap in the dependent
variables in the year-by-year protability and growth regressions. This can produce
autocorrelation of the slopes that affects the standard errors of the average slopes.
Inspection of the autocorrelations (not shown) in the asset growth regressions suggests no
evidence of a problem for any forecast horizon. The autocorrelations of the slopes in the
multiyear protability regressions are more often positive, but they are not systematically
large. Given the large standard errors of the autocorrelations, we are reluctant to impose
corrections that may not be warranted. Moreover, there is no overlap in the year-by-year
regressions that forecast protability and growth one year ahead, and the one-year and
multiyear regression results are always generically similar.
Table 1
Regressions to predict protability and asset growth
The table shows average slopes and their Fama-MacBeth t-statistics from annual cross-section regressions to predict protability, Yt+t/Bt, and asset growth, dAt+t/
At (At+tAt)/At, one, two, and three years ahead (t 1,2,3). Yt, Dt, and ACt are earnings, dividends, and accruals per share for the scal year ending in calendar
year t. ACt is accruals for rms with negative accruals (zero otherwise) and +ACt is accruals for rms with positive accruals. Bt, At, and Mt are book equity, total
assets, and stock price per share at the end of scal year t. MCt is market capitalization (price times shares outstanding) at the end of scal year t. It is the I/B/E/S
consensus forecast of earnings for the coming year, sampled at the end of scal year t. 1Yrt is the stock return for the year up to the end of scal year t, and 23Yrt is
the two-year return for the years up to the end of scal year t1. OHt is the probability of default on debt, estimated at the end of scal year t, from the logit
regression model of Ohlson (1980). PTt is the Piotroski (2000) composite index of rm strength. Neg Yt is a dummy variable that is one for rms that have negative
earnings for scal year t (zero otherwise), and No Dt is a dummy variable that is one for rms that pay no dividends during scal year t. The regressions are univariate
(dAt/At and It/Bt) or they use natural subsets of explanatory variables (ln Bt/Mt and ln Mt, or Neg Yt, Yt/Bt, ACt/Bt, and +ACt/Bt, or No Dt and Dt/Bt, or 1Yrt
and 2 3Yrt, or OHt and PTt). The time period for the dependent variable in the regressions that forecast protability and growth one year ahead is 19632004, except
for the regressions that use It/Bt, where the period is 19772003, and the regressions that use PTt, where the time period is 19722004. Overhead lines indicate the
explanatory variables used in a regression.
t ln Bt =M t ln MC t Neg Y t Y t =Bt AC t =Bt AC t =Bt No Dt Dt =Bt dAt =At 1 Yrt 2 3Yrt OH t PT t I t =Bt
Regressions to predict asset growth, dAt+t/At

Average slopes
1 0.11 0.70 0.06 0.58 0.00 0.01 0.02 0.25 0.16 0.10 0.05 2.61 1.23 0.62
2 0.22 1.71 0.06 1.23 0.11 0.01 0.05 0.48 0.30 0.20 0.09 4.67 2.24 1.23
3 0.32 2.87 0.03 1.93 0.08 0.02 0.08 0.78 0.43 0.28 0.12 6.78 3.03 1.88
t-statistics
1 16.79 4.27 8.52 18.97 0.05 0.62 2.50 4.56 12.42 17.72 14.91 17.79 8.42 9.12
2 18.91 6.19 3.74 20.07 1.40 0.18 4.16 4.23 10.63 17.65 12.95 18.77 9.74 9.94
ARTICLE IN PRESS
3 17.97 6.69 1.21 19.27 0.97 0.48 5.55 4.06 10.02 14.76 13.63 20.02 7.61 10.59
Regressions to predict profitability, Yt+t/Bt

Average slopes
1 0.08 1.05 0.08 1.01 0.02 0.08 0.02 1.06 0.10 0.10 0.04 2.37 2.47 0.97
2 0.07 0.84 0.03 0.98 0.01 0.09 0.00 1.14 0.08 0.08 0.03 1.71 1.89 0.89
3 0.07 0.77 0.01 1.01 0.05 0.07 0.00 1.17 0.07 0.06 0.02 1.51 1.53 0.91
t-statistics
1 12.19 4.72 6.23 58.31 1.10 8.44 2.30 27.19 6.68 11.89 11.22 18.26 13.05 34.56
E.F. Fama, K.R. French / Journal of Financial Economics 82 (2006) 491518
2 8.03 3.56 1.97 40.67 0.49 6.76 0.50 24.24 5.02 9.62 9.33 17.15 12.10 27.54
3 6.52 2.78 0.58 33.06 3.25 4.33 0.35 21.31 3.99 8.63 4.94 10.22 11.08 20.42
497
498
Table 2
Multiple regressions to predict protability and asset growth
The table shows average slopes and their Fama-MacBeth t-statistics from annual cross-section regressions to predict protability, Yt+t/Bt, and asset growth, dAt+t/
At (At+tAt)/At, one, two, and three years ahead (t 1,2,3). Yt, Dt, and ACt are earnings, dividends, and accruals per share for the scal year ending in calendar
year t. ACt is accruals for rms with negative accruals (zero otherwise) and +ACt is accruals for rms with positive accruals. Bt, At, and Mt are book equity, total
assets, and stock price per share at the end of scal year t. MCt is market capitalization (price times shares outstanding) at the end of scal year t. It is the I/B/E/S
consensus forecast of earnings for the coming year, sampled at the end of scal year t. 1Yrt is the stock return for the year up to the end of scal year t, and and 2 3Yrt
is the two-year return for the years up to the end of scal year t1. OHt is the probability of default on debt, estimated at the end of scal year t, from the probit
regression model of Ohlson (1980). PTt is the Piotroski (2000) composite index of rm strength. Neg Yt is a dummy variable that is one for rms that have negative
earnings for scal year t (zero otherwise), and No Dt is a dummy variable that is one for rms that pay no dividends during scal year t. Firms is the average number of
rms in the regressions. The time period for the dependent variable in the regressions that forecast protability and asset growth one year ahead is 19632004, except
for the regressions that use It/Bt, where the period is 19772003, and the regressions that use PTt, where the time period is 1972-2004. Int is the regression intercept
and the regression R2 is adjusted for degrees of freedom.
t Firms Int ln Bt/Mt ln MCt Neg Yt Yt/Bt ACt/Bt +ACt/Bt dAt/At No Dt Dt/Bt 1Yrt 2 3Yrt OHt PTt It/Bt R2
Regressions to predict asset growth, dAt+t/At

Average slopes
ARTICLE IN PRESS
1 1,953 1.13 0.10 0.53 0.09 0.19 0.03 0.09 0.05 0.01 1.13 0.12
2 1,810 1.25 0.19 1.23 0.14 0.46 0.04 0.21 0.10 0.02 2.16 0.15
3 1,675 1.40 0.28 2.03 0.15 0.79 0.04 0.36 0.11 0.03 3.26 0.16
t-statistics
1 131.46 12.93 3.25 13.55 4.51 1.02 5.73 6.82 2.56 14.26
2 69.97 15.98 4.42 6.49 5.66 0.46 8.58 5.80 2.62 15.09
3 47.87 16.46 5.23 4.14 6.91 0.35 9.42 5.46 2.76 14.25
Average slopes
1 1,458 1.09 0.08 0.31 0.06 0.10 0.06 0.10 0.00 0.02 1.07 0.04 0.02 1.11 0.03 0.09 0.16
2 1,353 1.16 0.17 0.87 0.09 0.25 0.03 0.17 0.00 0.04 2.02 0.08 0.04 2.28 0.26 0.10 0.20
3 1,259 1.25 0.27 1.60 0.10 0.43 0.10 0.28 0.02 0.05 3.17 0.12 0.05 3.12 0.81 0.12 0.20
t-statistics
1 57.52 10.72 0.89 11.22 3.25 2.64 3.40 0.04 2.89 7.82 3.53 6.66 11.14 0.24 2.28
2 30.42 12.11 1.60 8.37 3.64 0.82 3.93 0.10 2.89 10.90 7.80 5.25 13.12 1.07 1.18
3 18.28 13.33 2.25 3.74 2.50 1.40 4.16 0.51 2.63 10.44 5.58 4.68 10.10 1.45 1.22
Regressions to predict profitability, Yt+t/Bt
Average slopes
1 1,953 0.03 0.04 0.00 0.07 0.78 0.03 0.07 0.00 0.03 0.02 0.39
2 1,810 0.04 0.04 0.03 0.02 0.71 0.03 0.07 0.02 0.02 0.13 0.26
3 1,675 0.15 0.04 0.02 0.01 0.70 0.07 0.05 0.02 0.02 0.14 0.20
t-statistics
1 3.95 8.05 0.00 7.50 35.16 2.21 7.89 0.81 10.21 0.51
2 3.67 5.32 0.27 2.84 21.04 1.89 5.35 3.15 6.42 2.15
3 3.11 3.80 0.15 1.13 14.82 4.18 2.82 3.23 3.86 1.56
Average slopes
1 1,458 0.04 0.04 0.21 0.05 0.47 0.02 0.11 0.02 0.03 0.04 0.04 0.00 0.25 0.43 0.37 0.39
2 1,353 0.03 0.04 0.36 0.01 0.41 0.02 0.09 0.03 0.03 0.07 0.02 0.00 0.09 0.42 0.35 0.24
3 1,259 0.04 0.04 0.57 0.01 0.45 0.06 0.07 0.02 0.03 0.03 0.00 0.02 0.11 0.30 0.45 0.18
t-statistics
1 3.86 9.37 2.18 8.83 10.94 0.83 10.30 3.26 9.01 0.78 6.16 0.65 3.23 5.12 7.19
2 1.99 7.41 2.56 1.83 9.12 0.78 3.71 4.31 8.22 1.11 2.48 1.64 0.93 3.16 8.90
3 1.70 3.99 3.19 0.68 6.63 2.29 2.45 1.92 4.03 0.29 0.11 1.79 0.45 2.33 5.96
ARTICLE IN PRESS
499
ARTICLE IN PRESS
The multiple regressions to forecast protability and asset growth that provide our
proxies for expected protability and investment are in Table 2. Table 1 is background. It
summarizes preliminary regressions to show that, used alone or in small natural
subgroups, all variables in the multiple regressions of Table 2 forecast protability or
asset growth, and typically both. Our discussion largely focuses on the evidence about
marginal explanatory power from the multiple regressions of Table 2. There are two sets of
regressions in Table 2. The rst uses only lagged size and accounting fundamentals to
forecast protability and growth. The second set adds lagged returns, analyst earnings
forecasts, and the two general measures of rm strength, PTt and OHt, to the explanatory
variables.
3.1. Asset growth
Consider the Table 2 regressions to forecast asset growth. In the rst set, all accounting
fundamentals are related to future asset growth in plausible ways. Smaller rms and more
protable rms tend to grow faster, but rms that pay more dividends grow more slowly.
Firms with higher book-to-market ratios (so-called value rms) grow less rapidly than low
Bt/Mt rms (growth rms). Among rms with positive accruals (reported earnings exceed
cash earnings from operations), larger accruals are associated with slower future asset
growth. The relation between accruals and growth is not discernible when accruals are
negative. In terms of t-statistics, Bt/Mt and Dt/Bt have the strongest explanatory power,
with average slopes more than 12 standard errors from zero. Lagged asset growth also
helps predict future growth, but in economic terms the effects are small. Without showing
the details, we can report that adding more lags of growth to the multiple regressions in
Table 2, or replacing the rst lag of growth with a three-year average, does not produce
stronger evidence for the importance of lagged asset growth in predicting future growth.
This is in contrast to the univariate regressions in Table 1, in which lagged growth shows
strong power to forecast asset growth up to three years ahead.
Adding lagged returns, I/B/E/S protability forecasts, PTt, and OHt to the asset growth
regressions tends to reduce the size and precision of other slopes, which nevertheless
continue to have explanatory power, with two exceptions. The average slopes on size are
still negative, but they are now less than two standard errors from zero for forecasts one
and two years ahead. More interesting, lagged asset growth loses its power to forecast
future growth, a result of some import in interpreting the return regressions later. Lagged
returns and OHt have marginal forecast power in the full regressions, and I/B/E/S
protability forecasts may have explanatory power, at least for forecasts one year ahead.
Not surprisingly, rms with higher past returns and higher forecasted protability tend to
invest more, while rms with higher probability of default (OHt) grow less rapidly. Used
alone, there is a strong positive relation between the Piotroski measure of rm strength and
future asset growth (Table 1), but in the full regressions, PTt does not have reliable forecast
power.
3.2. Profitability
When size and the accounting fundamentals are used to forecast protability, Yt+t/Bt,
one, two, and three years ahead (t 1, 2, 3), lagged protability has by far the strongest
forecast power. For example, the average slope on Yt/Bt for forecasts one year ahead, 0.78,
ARTICLE IN PRESS
is 35.16 standard errors from zero. Thus, there is considerable persistence in protability.
But protability is mean reverting; the one-year slope on lagged protability is about ten
standard errors below 1.0, and the slope decays to 0.70 for forecasts three years ahead.
Without showing the details, we can report that adding more lags of protability, or
replacing the rst lag with a three-year average, does not produce stronger evidence for the
importance of lagged protability in predicting future protability than the rst lag of
protability alone.
As expected, the book-to-market ratio helps predict protability; rms with higher
Bt/Mt (value rms) tend to be less protable. The forecast power of the ratio of dividends to
book equity, which shows up clearly in the regressions of Table 1, largely disappears when
placed in competition with other fundamentals in Table 2. But the multiple regressions of
Table 2 produce stronger evidence that rms that do not pay dividends are less protable.
The link between lagged asset growth and future protability merits discussion. In
univariate regressions (Table 1), lagged growth is positively related to future protability,
but the slope turns negative in the multivariate regressions of Table 2. Thus, with controls
for size and other fundamentals (especially past protability), higher asset growth is
associated with lower future protability and growth in earnings. (We return to this nding
later.)
When lagged returns, I/B/E/S earnings forecasts, and the Piotroski and Ohlson
measures of rm strength (PTt and OHt) are added to the protability regressions, not
much happens to the average slopes for Bt/Mt, the ratios of negative and positive accruals
to book equity, and the ratio of dividends to book equity. But the slopes on lagged
protability are smaller, and the slopes on lagged asset growth tend to be more reliably
negative. When the two lagged returns are used alone to forecast protability, their
average slopes are strongly positive (Table 1), but in competition with other variables
(Table 2), the slopes on lagged returns decline and only the rst lagged return (year t1)
shows reliable forecast power. OHt produces strong negative average slopes when used
alone to forecast protability; higher probability of default is (not surprisingly) associated
with lower future protability. But in the multiple regressions, OHt loses most of its
explanatory power, at least for forecasts more than a year ahead. In contrast, though the
positive average slopes on the PTt measure of rm strength are smaller when other
variables are in the protability regressions, they remain more than 2.3 standard errors
from zero.
Used alone to forecast protability (Table 1), lagged protability and analyst earnings
forecasts have average slopes close to 1.0. Thus differences in lagged protability or in
analyst forecasts show up roughly one for one in future protability. But in the multiple
regressions that use the full set of variables to forecast protability (Table 2), the slopes on
lagged protability and analyst forecasts typically fall to less than half the values observed
in Table 1, and the sum of the slopes is now less than 1.0. The average slopes for both
variables are more than 5 standard errors from zero. Thus, in the multiple regressions, the
two variables (correlated 0.35) split the information they share about future protability.
Moreover, many variables help forecast protability in the full regressions. This result
conrms earlier evidence that analysts overlook information when making earnings
forecasts. (See, for example, Ali, Klein, and Rosenfeld, 1992; Abarbanell and Bernard,
1992; Easterwood and Nutt, 1999; Ahmed, Nainar, and Zhang, 2003.)
Spawned by Sloan (1996), a large literature shows that accruals result in transitory
variation in earnings. The negative slopes on accruals in the protability regressions
ARTICLE IN PRESS
conrm this result. Note, however, that the average slopes for +ACt/Bt in the one- and
two-year protability regressions of Table 2 are more negative than for ACt/Bt, but the
slope for +ACt/Bt tends to become less negative for longer horizons, and the coefcient
for ACt/Bt becomes more negative. As a result, the slopes for positive and negative
accruals are about equal in the three-year regressions, around 0.06. The behavior of the
slopes suggests that the reversal of positive accruals in reported earnings occurs faster, but
positive and negative accruals have comparable long-run transitory effects on earnings.
Accruals, however, do not mean revert much faster than the cash component of
earnings. In the Table 2 regressions that use only size and lagged accounting fundamentals
to forecast protability, the mean reversion of protability (which includes cash earnings
and accruals) is picked up by lagged protability and accruals, with accruals measuring
marginal mean reversion beyond that captured by lagged protability. (Other explanatory
variables in the regressions largely just allow for differences in long-term average
protability across rms.) The point estimates of the accrual slopes, around minus 6% at
the three-year horizon, suggest that the long-term marginal mean reversion of protability
associated with accruals is small.
Sloans (1996) hypothesis, adopted near uniformly in the literature on stock returns and
accruals, is that investors do not understand the faster mean reversion of the accruals part
of earnings. This leads to a negative relation between current accruals and future stock
returns observed when the mean reversion of accruals hits measured earnings. But the fact
that accruals do not mean revert much faster than the cash component of earnings suggests
that Sloans story cannot in itself explain large spreads in average returns associated with
accruals. And our estimates of the marginal mean reversion of protability due to accruals
are similar to those of Sloan (1996) and others.
The protability regressions in Table 2 that include all explanatory variables produce the
same or slightly lower R2 than the regressions that use only size and accounting
fundamentals to forecast protability. Thus, though lagged returns, PTt, and analyst
forecasts have marginal explanatory power in the protability regressions, it comes at the
expense of other variables, primarily lagged protability. Lagged returns, PTt, and analyst
forecasts do not add to the overall power of the protability forecasts provided by size and
the accounting fundamentals. Without showing the details, we can also report that lagged
protability alone produces protability forecasts near as powerful (in terms of R2) as
those from the expanded regressions in Table 2. These comments may be pertinent when
we nd next that a small set of explanatory variables (including lagged protability) seem
to provide a simple proxy for expected protability that shows more power to forecast
stock returns than the tted values from the protability regressions in Table 2.
4. Expected returns: cross-section regressions
We test for the protability and investment effects in expected returns predicted by the
valuation equation (3) in three steps. We rst present cross-section regressions that explain
average stock returns with lagged values of size, Bt/Mt, asset growth, protability,
accruals, and the PTt and OHt measures of rm strength. The goal is to examine whether
simple proxies for expected protability and asset growth add to the explanation of
average returns provided by size and Bt/Mt. We then use more complicated proxies for
expected protability and asset growththe tted values from the regressions of Table 2
to test for protability and investment effects in average returns. The nal tests use
ARTICLE IN PRESS
portfolios to examine whether the protability and investment effects identied in the
cross-section regressions are large and pervasive in the sample as a whole and within
portfolios formed on size and Bt/Mt.
We estimate cross-section return regressions monthly, starting in July 1963, with the
explanatory variables updated annually, at the end of June. To ensure that the explanatory
variables are known at the beginning of the month of the dependent returns, the
accounting variables in the regressions are for scal years that end in the calendar year
preceding the July when they are rst used. Thus we use data from scal yearends between
January and December of year t to forecast monthly returns from July of t+1 to June of
t+2. As in Fama and French (1992), market equity for the size variable is measured at the
end of June of t+1, and market equity in the book-to-market ratio is for the end of
December of t. To reduce the impact of outliers, we winsorize the independent variables in
the return regression at the 0.5% level. Thus extreme values are shrunk to the 0.5 and 99.5
percentiles for year t. (As in Tables 1 and 2, we consider only the upper or lower bound for
one-sided variables.)
4.1. Baseline tests
Conrming previous evidence, Table 3 shows that when size and the book-to-market
ratio are used alone to explain returns, there is a strong positive relation between average
return and Bt/Mt. The t-statistic for the average Bt/Mt slope is near three standard errors
from zero. Thus high book-to-market (value) rms have higher average returns than low
book-to-market (growth) rms. As in previous work, small (low market cap) rms have
higher average returns than big rms, but the negative average size slope is only 1.20
standard errors from zero.
More interesting, simple proxies for expected protability and asset growth seem to
conrm the positive protability and negative growth effects in average returns predicted
by the valuation equation (3). When lagged protability and asset growth are added to the
return regressions that include size and Bt/Mt, there is a strong positive relation between
protability and average return (t 2:55) and a stronger negative relation between average
return and asset growth (t 3.87). Moreover, adding lagged protability and asset
growth to the return regressions has almost no effect on the average slope for Bt/Mt and
enhances the average slope for size, which is now 1.83 standard errors from zero. We can
also report that adding lags of protability and growth or replacing the rst lags with
averages of three years of past values does not produce reliable improvements in
explanatory power.
Because accruals are negatively related to future protability (Table 2), the valuation
equation (3) predicts a negative relation between accruals and future returns. The average
slope for positive accruals, +ACt/Bt, in the return regressions of Table 3 is reliably
negative (t 6.82). This is consistent with earlier evidence (Sloan, 1996; Collins and
Hribar, 2000; Chan, Chan, Jegadeesh, and Lakonishok, 2006) that accruals predict
returns. The average slope on negative accruals, ACt/Bt, is also negative but less than one
standard error from zero. Thus, with controls for other variables, negative accruals do not
reliably predict higher future returns. This result does not seem to have a precedent in the
literature.
Some of the information in positive accruals about future returns is related to the
information in lagged growth. The average correlation between +ACt/Bt and dAt/At is
504
Table 3
Monthly cross-section return regressions
The table shows average slopes and their Fama-MacBeth t-statistics from monthly cross-section regressions to predict stock returns. Yt, Dt, and ACt are earnings,
dividends, and accruals per share for the scal year ending in calendar year t. ACt is accruals for rms with negative accruals (zero otherwise) and +ACt is accruals
for rms with positive accruals. Bt, At, and Mt are book equity, total assets, and stock price per share at the end of scal year t. MCt is market capitalization (price
times shares outstanding) at the end of June of year t+1. It is the I/B/E/S consensus forecast of earnings for the coming year, sampled at the end of scal year t. 1Yrt is
the stock return for the year up to the end of scal year t, and 23Yrt is the two-year return for the years up to the end of scal year t1. OHt is the probability of
default on debt, estimated at the end of scal year t, from the logit regression model of Ohlson (1980). PTt is the Piotroski (2000) composite index of rm strength.
Neg Yt is a dummy variable that is one for rms that have negative earnings for scal year t (zero otherwise), and No Dt is a dummy variable that is one for rms that
pay no dividends during scal year t. F(Yt+t/Bt) and F(dAt+t/At), expected protability and asset growth, are tted values from the rst pass regressions in Table 2
that include lagged fundamentals (ln Bt/Mt, ln MCt, Neg Yt, Yt/Bt, ACt/Bt, +ACt/Bt, dAt/At, No Dt, and Dt/Bt) and lagged fundamentals, lagged returns, It/Bt,
OHt, and PTt. Firms is the average number of rms in the regressions. The regressions are estimated monthly, beginning in July of 1963, using explanatory variables
that are updated at the end of each June. The accounting explanatory variables in the regression for July of year t+1 are for scal years ending in calendar year t. The
size variable, ln MCt, is measured at the end of June of year t+1, but in Bt/Mt, Mt is measured at the end of December of year t. The time period for the dependent
returns in the regressions is July 1963 to December 2004, except for the regressions that require It/Bt, where the period is July 1977 to December 2003, and the
regressions that require PTt, where the period is July 1972 to December 2004. Int is the regression intercept and the regression R2 is adjusted for degrees of freedom.
Firms Int ln Bt/Mt ln MCt Neg Yt Yt/Bt ACt/Bt +ACt/Bt dAt/At OHt PTt R2
Part A: Regressions use lagged profitability, asset growth, accruals, OHt, and PTt
ARTICLE IN PRESS
Average 2,058 1.66 0.28 0.06 0.02

t-statistics 3.85 2.97 1.20
Average 2,058 1.69 0.28 0.08 0.00 1.10 0.40 0.03
t-statistics 4.36 2.74 1.83 0.02 2.55 3.87
Average 2,058 1.83 0.26 0.10 0.00 1.38 0.24 1.42 0.19 0.04
t-statistics 4.93 2.61 2.37 0.03 3.21 0.80 6.82 1.99
Average 2,253 1.35 0.34 0.07 0.04 0.06 0.03

t-statistics 2.43 3.02 1.33 2.25 2.58
Average 2,253 1.51 0.34 0.09 0.11 1.51 0.09 1.32 0.25 0.03 0.04 0.03
t-statistics 3.44 2.89 1.76 0.77 3.40 0.41 5.42 2.44 1.55 2.55
t Firms Int ln Bt/Mt ln MCt F(dA1+t/At) F(Yt+t/Bt) R2
Part B: Regressions use expected profitability, F(Yt+t/Bt), and asset growth, F(dAt+t/At), from first-stage regressions
Expected profitability and growth estimated with lagged fundamentals

Average slopes
1 2,058 1.61 0.37 0.08 0.04 1.58 0.03
2 2,058 1.11 0.48 0.08 0.39 2.05 0.03
3 2,058 0.97 0.53 0.07 0.42 2.04 0.03
t-statistics
1 1.86 3.42 1.92 0.05 2.03
2 1.92 3.87 1.76 0.87 2.37
3 1.89 4.14 1.56 1.34 2.40
Expected profitability and growth estimated with lagged fundamentals, lagged returns, It/Bt, OHt, and PTt
Average slopes
1 1,530 2.06 0.20 0.09 0.20 1.27 0.03
2 1,530 1.75 0.25 0.10 0.08 1.58 0.03
3 1,530 1.57 0.28 0.09 0.21 1.43 0.03
t-statistics
1 1.97 1.49 1.75 0.18 1.28
2 2.75 1.67 1.80 0.16 1.49
3 2.71 1.92 1.71 0.66 1.64
ARTICLE IN PRESS
505
ARTICLE IN PRESS
0.29, and adding accruals to the return regressions cuts the average slope on asset growth
in half, from 0.40 (t 3.87) to 0.19 (t 1.99). This is in line with previous evidence
that the accruals may in part pick up a growth effect in average returns (Faireld,
Whisenant, and Yohn, 2003). Another interpretation, however, is that asset growth
predicts returns because it helps predict protability, and accruals absorb some of the
protability information in asset growth. This is consistent with the evidence (Table 2) that
with controls for other variables (primarily lagged protability) the marginal relation
between lagged asset growth and future protability is negative. In contrast, adding
accruals to the return regressions increases the slope on lagged protability, from 1.10
(t 2.55) to 1.38 (t 3.21). This is in line with previous evidence that adding accruals
helps clean up the information in lagged protability about future protability (Sloan,
1996). The important point, however, is that all these results on how Bt/Mt, protability,
growth, and accruals predict returns are consistent with the valuation equation (3).
Similarly, the PTt and OHt measures of rm strength are proxies for expected net cash
ows. The valuation equation (3) thus implies that they are candidates for identifying variation
in average returns missed by size and Bt/Mt. Conrming Piotroski (2000) and Grifn and
Lemmon (2002), Table 3 shows that PTt and OHt have explanatory power (average slopes
more than 2.2 standard errors from zero) when added to return regressions that include size
and Bt/Mt. Controlling for size and Bt/Mt, stronger rms (higher PTt) have higher average
returns, and rms with higher default probabilities (OHt) have lower average returns.
Adding lagged protability, asset growth, and accruals to the return regressions
dampens the average slopes for PTt and OHt, from 0.06 to 0.04 (t 2.55) for PTt and from
0.04 to 0.03 (t 1.55) for OHt (Table 3). Collinearity thus takes its toll, but each of
these variables (lagged protability, growth, accruals, PTt, and OHt) seems to capture
information about average returns missed by the others. (Without showing the details, we
can report that adding the two dividend variables, No Dt and Dt/Bt, and the I/B/E/S
earnings forecast variable, It/Bt, does not enhance the explanatory power of the return
regressions.)
Finally, the previous literature typically interprets observed relations between returns
and lagged protability, investment, accruals, PTt, and OHt as evidence of mispricing. But
as emphasized in Section 2, the protability, investment, and net cash ow effects in
average returns captured by these variables are consistent with the valuation equation (3)
whether or not pricing is rational. And tests of Eq. (3) cannot in themselves distinguish
rational from irrational pricing.
4.2. Better proxies for expected profitability and investment
The valuation equation (3) suggests that protability, asset growth, accruals, PTt, and
OHt predict returns because they have information about expected protability and asset
growth. If so, it seems reasonable that the tted values from the rst-stage protability and
growth regressions in Table 2, which aggregate the information in these and other
variables about expected protability and growth, should forecast returns at least as well.
The monthly return regressions that use the tted values, in Panel B of Table 3, do not
support this conclusion.
When lagged accounting fundamentals (including protability, asset growth, and
accruals) are used along with size and Bt/Mt to construct proxies for expected protability
and growth, there is a reliable positive relation between expected protability and average
ARTICLE IN PRESS
return. The t-statistics for the average slopes on expected protability are 2.03 in the return
regressions that use expected protability and expected growth one year ahead and more
than 2.3 in the regressions that use forecasts two and three years ahead. Contrary to the
predictions of the valuation equation (3), however, the return regressions of Table 3
produce positive average slopes on the Table 2 regression proxies for expected asset
growth, but they are not reliably different from zero.
Table 2 says that lagged returns, analyst earnings forecasts, PTt, and OHt have
explanatory power in the rst-stage protability and asset growth regressions that also
control for size, Bt/Mt, and lagged accounting fundamentals. But Table 3 says that the
tted values from these full rst-stage protability and growth regressions produce weaker
evidence of protability effects in average returns (the largest t-statistic is 1.64), and there
is still no evidence of asset growth effects.
4.3. Discussion
Why do the simple proxies for expected protability and investment provided by lagged
protability, asset growth, accruals, PTt, and OHt produce better descriptions of average
returns than the more complicated proxies from the rst-stage protability and asset
growth regressions that summarize the information in these and other variables? We offer
some possibilities.
There are two potential measurement error problems in the way we use the rst-stage
protability and asset growth regressions in the second-stage return regressions. First,
though the protability and asset growth regressions identify many variables that have
forecast power, the average slopes have measurement error, so there is a measurement
error problem when the regression tted values are used as explanatory variables for
returns. Second, the tted values from the rst-stage regressions used in the second-stage
return regressions are computed with full-period average slopes from the year-by-year rst-
stage regressions. The implicit assumption is that the true rst-stage slopes are constant. If
the slopes are not constant, full-period average slopes produce noisy period-by-period
estimates of expected protability and growth.
To explore whether variation in the true rst-stage slopes affects our results, we estimate
the second-stage return regressions using tted values from rolling rst-stage protability
and growth regressions. Specically, we estimate each years tted values with the average
slopes from the rst-stage regressions for the most recent ten years. If the true rst-stage
slopes are constant, tted values constructed from a rolling ten years of slopes should not
work as well in the second-stage return regressions as tted values that use full-period
average slopes. But without showing the details, we can report that the forecasts of
protability and growth based on rolling ten-year average slopes work about as well as (no
better or no worse than) forecasts that use the full-period average slopes from the rst-
stage regressions. This suggests that there is enough variation in the true slopes to offset
the larger estimation error of ten-year average rst pass regression slopes.
More positively, we suggest that entering lagged size, Bt/Mt, protability, asset growth,
accruals, PTt, and OHt directly as explanatory variables in the return regressions provides
a exible solution to these measurement error problems. Specically, entering the
explanatory variables for expected protability and asset growth into the monthly return
regressions in an unrestricted way implicitly allows them to pick up whatever rst-stage
slopes are currently relevant for predicting protability and growth.
ARTICLE IN PRESS
The failure of the tted values from the protability and asset growth regressions in
the second-stage return regressions may in part be due to collinearity. We use the
same explanatory variables in the protability and growth regressions. Many variables
affect the two tted values in similar ways. The coefcients on Bt/Mt, Neg Yt, +ACt/Bt,
and No Dt are negative in both sets of regressions (Table 2), and the coefcients on Yt/Bt
are positive. As a result, the tted values from the rst-stage regressions are highly
correlated. For example, the average of the annual correlations between the one-year-
ahead tted values from the comprehensive rst pass protability and growth regressions
is 0.76.
The tted values from the rst-stage protability and asset growth regressions are also
correlated with the size and Bt/Mt variables in the second-stage return regressions. Book-
to-market is a powerful explanatory variable in the rst-stage growth regressions; the
correlation between the estimates of expected growth and Bt/Mt is typically about 0.8.
Because Bt/Mt is also an explanatory variable in the second-stage return regressions, this
collinearity may obscure the growth effects in average returns.
Both size and Bt/Mt have strong slopes in the rst-stage protability regressions and, as
a result, they are correlated with the regression tted values. The correlation of the
estimates of expected protability with size is around 0.4, and the correlation with Bt/Mt is
about 0.7. These links are not as tight as those between Bt/Mt and the estimates
of expected asset growth, but they do make it more difcult to identify the marginal
relation between expected protability and expected return in the second-stage return
regressions.
This is a good place to note that the valuation equation (3) does not imply that there
must be variation in expected returns independent of size and Bt/Mt. Suppose differences
in expected returns are perfectly explained by size and Bt/Mt. Then the best possible
forecasts of expected net cash ows must be perfectly correlated with linear combinations
of size and Bt/Mt, so there are no protability and investment effects in expected returns
left unexplained by size and Bt/Mt. Because the proxies for expected growth from the asset
growth regressions are highly correlated with Bt/Mt, this story may explain why the
proxies do not identify growth effects in average returns. The proxies for expected
protability from the cross-section protability regressions are less correlated with Bt/Mt,
which may explain why they show up more strongly in the return tests.
Finally, lagged asset growth may show up in the return regressions with a negative
average slope because of the information in asset growth about future protability, not
because of its information about expected growth. Lagged asset growth is not important in
predicting future growth in the multiple regressions of Table 2. Moreover, in the
regressions to forecast protability (where accruals and especially lagged protability have
powerful roles), higher growth is associated with lower future protability. This is
consistent with lower expected returns for faster growing rms, especially when the return
regressions control for lagged protability and accruals.
5. Expected returns: portfolio tests
Cross-section return regressions can identify variables that help describe average stock
returns, but the economic signicance of the average slopes is not always easy to judge.
Moreover, the average slopes from the return regressions cannot tell us whether the
regressions are well-specied. For example, do the protability and asset growth effects in
ARTICLE IN PRESS
average returns identied by the regressions show up in a general way among stocks in
different size and book-to-market groups? This section uses portfolio tests to address these
issues.
5.1. Economic significance
Table 4 shows predicted and actual returns on portfolios formed using predicted values
from the cross-section return regressions of Table 3. The explanatory variables in the
monthly return regressions change once a year, at the end of June. Thus, at the end of
each June, we compute predicted monthly returns on individual stocks for the following
year by combining the current values of the explanatory variables in the return regres-
sions of Table 3 with the average monthly regression slopes for the full sample period.
(Because the goal is to develop perspective on the results from the return regressions, the
look-ahead bias suggested by our use of full-period average regression slopes is not an
issue.) We then allocate stocks to high and low expected return portfolios based on
whether their predicted monthly returns for the next year are above or below the sample
median for the year. For each return regression in Table 3, Table 4 shows the average
difference between predicted high and low returns and the average difference between
actual returns. We report both equal-weight (EW) returns, which give heavy weight to the
many small rms in the sample, and value-weight (VW) returns, which give heavy weight
to large rms.
The rst-stage protability and asset growth regressions of Table 2 examine forecast
horizons of one, two, and three years. Table 3 uses these forecasts in three separate sets of
regressions to explain the cross section of average returns. We also estimate the high minus
low return spreads in Table 4 using the Table 3 regressions for each of the three forecast
horizons. The results for different forecast horizons are near identical. To save space,
Table 4 shows only the predicted and actual return spreads based on the Table 3 return
regressions that use forecasts of protability and growth one year ahead.
Predicted average spreads are fairly similar for equal-weight and value-weight returns.
This suggests that the variation across stocks in the underlying regression explanatory
variables is roughly similar for small and big rms. When we equal weight returns, the
average actual return spread for every regression is higher than the predicted spread, but
with value weighting, actual spreads are below predicted spreads. We infer that the average
return effects measured by the Table 3 regressions are stronger among smaller rms. Still,
the ordering of the average spreads in actual returns produced by successive regressions in
Table 4 is the same for equal-weight and value-weight returns, so we also infer (and Table 5
conrms) that, though the magnitudes differ, the average return effects observed in the
regressions of Table 3 are common to small and big rms.
Table 4 conrms existing evidence that differences in size and book-to-market equity are
associated with large spreads in average returns. The average spreads in high minus low
returns predicted by the cross-section regressions that use just size and Bt/Mt to explain
returns are 0.42% (EW) and 0.49% (VW) per month; the average actual spreads in returns
are 0.52% (EW) and 0.43% (VW), and they are 4.66 and 3.42 standard errors from zero.
We know from Table 3 and previous work that the lions share of these spreads is due to
the value premium identied by Bt/Mt.
The regressions in Table 3 say that lagged protability, asset growth, and accruals have
statistically reliable power to forecast returns when added to regressions that also include
ARTICLE IN PRESS
Table 4
Equal-weight (EW) and value-weight (VW) predicted and actual average high minus low returns
Each month the tted values, computed using the average monthly slopes for the full sample period from the
return regressions in Table 3, are used to allocate stocks to high and low predicted return portfolios based on
whether their regression tted values for the month are above or below the sample median for the month. For
each return regression in Table 3, we compute the average monthly predicted and actual spreads between the
equal-weight and value-weight average high and low returns. The t-statistics, t( ), are the ratios of the average
actual spreads to their time series standard errors.
The explanatory variables in the return regressions (dened in Table 3) used to allocate rms to high and low
predicted return portfolios are:
1. ln Bt/Mt, ln MCt,
2. ln Bt/Mt, ln MCt, Neg Yt, Yt/Bt, dAt/At,
3. ln Bt/Mt, ln MCt, Neg Yt, Yt/Bt, ACt/Bt, +ACt/Bt, dAt/At,
4. ln Bt/Mt, ln MCt, OHt, PTt,
5. ln Bt/Mt, ln MCt, Neg Yt, Yt/Bt, ACt/Bt, +ACt/Bt, dAt/At, OHt, PTt,
6. ln Bt/Mt, ln MCt, F(Yt+1/Bt), F(dAt+1/At),
7. ln Bt/Mt, ln MCt, F(Yt+1/Bt), F(dAt+1/At).
Regressions 15 (Reg) use lagged protability, asset growth, accruals, OHt, and PTt as proxies for expected
protability and asset growth. Regressions 6 and 7 use F(Yt+1/Bt) and F(dAt+1/At), the tted values from the
protability and asset growth regressions of Table 2 for forecasts one year ahead, as proxies for expected
protability and asset growth. In Regression 6, F(Yt+1/Bt) and F(dAt+1/At) use ln Bt/Mt, ln MCt, Neg Yt, Yt/Bt,
ACt/Bt, +ACt/Bt, dAt/At, No Dt, and Dt/Bt as explanatory variables. Regression 7 adds 1Yrt, 2-3Yrt, OHt, PTt,
and It/Bt to the variables used to construct F(Yt+1/Bt) and F(dAt+1/At). Return spreads computed using
Regressions 13 and 6 are for July 1963 to December 2004, the spreads computed using Regressions 4 and 5 start
in July 1972, and those computed using Regression 7 start in July 1977.
Average predicted spread Average actual spread t(Average actual spread)
Reg EW VW EW VW EW VW
1 0.42 0.49 0.52 0.43 4.66 3.42

2 0.51 0.52 0.58 0.48 5.16 3.71
3 0.54 0.54 0.67 0.50 6.49 4.16
4 0.48 0.52 0.55 0.42 4.56 3.16
5 0.57 0.52 0.65 0.49 5.50 3.58
6 0.43 0.50 0.53 0.47 4.60 3.72
7 0.29 0.36 0.36 0.32 2.98 2.21
size and Bt/Mt as explanatory variables. But Table 4 says that the increments to average
returns produced by these variables are modest. Adding lagged protability and asset
growth to the return regressions in Table 3 (Regression 2) increases the predicted spreads
in Table 4 by 0.09% (EW) and 0.03% (VW) per month; the increments to average actual
spreads in returns are 0.06% (EW) and 0.05% (VW). Adding positive and negative
accruals further increases average predicted return spreads by just 0.03% (EW) and 0.02%
(VW) per month; the increases in average actual return spreads are 0.09% (EW) and
0.02% (VW).
If we use just PTt and OHt with size and Bt/Mt to forecast returns (Regression 4 in Table 4),
the average predicted and actual return spreads are below the spreads produced by
combining lagged protability, asset growth, and accruals with size and Bt/Mt. And the
full Regression 5 in Table 4 that uses lagged size, Bt/Mt, protability, growth, accruals,
PTt, and OHt to forecast returns produces average predicted and actual return spreads
close to those obtained without PTt and OHt. In short, the Piotroski (2000) and Ohlson
ARTICLE IN PRESS
Table 5
Predicted and actual average high minus low return spreads for six size-B/M groups
In June of each year, the NYSE, Amex, and Nasdaq rms in our sample are allocated to two size groups, small
(S) and big (B), according to whether their market cap is below or above the NYSE median. Firms are also
allocated to three book-to-market groups depending on whether their Bt/Mt is in the bottom 30% (L), middle
40% (M), or top 30% (H) of Bt/Mt for NYSE rms. Intersecting the size and Bt/Mt groups produces six
portfolios, SL, SM, SH, BL, BM, and BH. Each month the tted values, computed using the average monthly
slopes for the full sample period from the return regressions in Table 3, are used to allocate stocks in each of the
six size-Bt/Mt groups to high and low predicted return portfolios based on whether their regression tted values
for the month are above or below their groups median. For each return regression in Table 3 and for each of the
six size-Bt/Mt groups, the table shows the average predicted and actual differences between the equal-weight
average high and low returns. The table also shows time series averages of simple monthly averages (Ave) of the
six value-weight return spreads. The explanatory variables in the return regressions (dened in Table 3) used to
allocate rms to high and low predicted return portfolios are:
1. ln Bt/Mt, ln MCt,
2. ln Bt/Mt, ln MCt, Neg Yt, Yt/Bt, dAt/At,
3. ln Bt/Mt, ln MCt, Neg Yt, Yt/Bt, ACt/Bt, +ACt/Bt, dAt/At,
4. ln Bt/Mt, ln MCt, OHt, PTt,
5. ln Bt/Mt, ln MCt, Neg Yt, Yt/Bt, ACt/Bt, +ACt/Bt, dAt/At, OHt, PTt,
6. ln Bt/Mt, ln MCt, F(Yt+1/Bt), F(dAt+1/At),
7. ln Bt/Mt, ln MCt, F(Yt+1/Bt), F(dAt+1/At).
Regressions 15(Reg) use lagged protability, asset growth, accruals, OHt, and PTt as proxies for expected
protability and asset growth. Regressions 6 and 7 use F(Yt+1/Bt) and F(dAt+1/At), the tted values from the
protability and asset growth regressions of Table 2 for forecasts one year ahead, as proxies for expected
protability and asset growth. In Regression 6, F(Yt+1/Bt) and F(dAt+1/At) use ln Bt/Mt, ln MCt, Neg Yt, Yt/Bt,
ACt/Bt, +ACt/Bt, dAt/At, No Dt, and Dt/Bt as explanatory variables. Regression 7 adds 1Yrt, 2-3Yrt, OHt, PTt,
and It/Bt to the variables used to construct F(Yt+1/Bt) and F(dAt+1/At). Return spreads computed using
Regressions 13 and 6 are for July 1963 to December 2004, the spreads computed using Regressions 4 and 5 start
in July 1972, and those computed using Regression 7 start in July 1977.
Reg SL SM SH BL BM BH Ave
Average spread in expected returns

1 0.19 0.11 0.21 0.25 0.11 0.13 0.17
2 0.44 0.23 0.26 0.36 0.19 0.19 0.28
3 0.48 0.34 0.35 0.39 0.25 0.22 0.34
4 0.33 0.24 0.30 0.34 0.21 0.22 0.27
5 0.51 0.37 0.39 0.42 0.27 0.26 0.37
6 0.28 0.19 0.24 0.29 0.18 0.18 0.23
7 0.23 0.17 0.19 0.21 0.16 0.15 0.18
Average spread in actual returns
1 0.33 0.06 0.24 0.18 0.15 0.18 0.19
2 0.60 0.31 0.34 0.34 0.23 0.25 0.35
3 0.86 0.34 0.39 0.39 0.26 0.25 0.42
4 0.45 0.25 0.37 0.22 0.14 0.32 0.29
5 0.83 0.44 0.47 0.35 0.14 0.24 0.41
6 0.49 0.08 0.28 0.26 0.26 0.23 0.27
7 0.34 0.24 0.17 0.25 0.18 0.12 0.22
t-statistics for average spread in actual returns
1 3.41 0.66 2.50 1.72 2.11 1.94 3.93
2 4.50 4.04 3.75 2.47 3.19 3.03 5.89
3 6.57 4.41 5.05 3.09 3.88 2.86 7.30
4 3.37 3.04 4.71 2.00 1.86 3.00 5.27
5 5.82 4.94 5.82 2.42 1.85 2.19 6.12
6 3.15 0.84 3.42 1.84 3.27 2.41 3.63
7 1.83 2.05 1.43 1.94 1.82 0.99 2.64
ARTICLE IN PRESS
(1980) measures of rm strength, which aggregate the information in many accounting

variables, seem to have no economically important information about expected returns
beyond the information in lagged protability, growth, and accruals, which in turn seems
modest.
In the return regressions of Table 3, the tted values from the regressions to explain
protability and asset growth in Table 2 do not show consistent marginal explanatory
power. We speculated that one problem may be collinearity: The tted values from the
protability and growth regressions are highly correlated with each other and with Bt/Mt.
In the return forecasts of Table 4, however, we are interested in whether the tted
values for protability and growth together have information about average returns
beyond that in size and Bt/Mt. In this task, the high correlation of the tted values with
Bt/Mt remains a problem, but collinearity between the two tted values themselves is
not a problem. There is thus reason to hope that the tted values from the protability
and growth regressions identify substantial variation in average predicted and actual
returns.
The hope is not realized. Table 4 says that, used with size and Bt/Mt as explanatory
variables in the return regressions of Table 3, the tted values from the regressions in Table
2 that use size, Bt/Mt, and accounting fundamentals to predict protability and asset
growth produce average spreads in predicted and actual returns about as large as those
produced by size and Bt/Mt alone. Thus in economic terms these forecasts of protability
and investment add little or nothing to the prediction of returns provided by size and Bt/Mt.
Moreover, the protability and asset growth regressions that use the full set of apparently
important explanatory variables produce lower average predicted and actual return
spreads than the baseline return regressions that use just size and Bt/Mt as explanatory
variables. Thus these tted values partially obscure the return predictions provided by size
and Bt/Mt.
5.2. Pervasiveness and regression specification
Table 4 gives an overall picture of the economic signicance of the variation in average
returns uncovered in the return regressions of Table 3. The nal task is to examine whether
the return regressions are well-specied in the sense that the predictions they make show up
in the average returns of stocks in different size and book-to-market groups.
The size-Bt/Mt groups are the portfolios used in the construction of the SMB (small
minus big market cap) and HML (high minus low Bt/Mt) returns of the three-factor model
of Fama and French (1993). In June of each year beginning in 1963, NYSE, Amex, and
Nasdaq rms are allocated to two size groups, small (S) and big (B), according to whether
their market cap is below or above the NYSE median. Firms are also allocated to three
book-to-market groups depending on whether their Bt/Mt is in the bottom 30% (L),
middle 40% (M), or top 30% (H) of Bt/Mt for NYSE rms. Intersecting the size and
Bt/Mt sorts produces six portfolios, SL, SM, SH, BL, BM, and BH.
At the end of each June, we allocate the stocks in each of the six size-Bt/Mt groups to
high and low expected return portfolios based on whether their predicted monthly returns
(tted values) for the next year are above or below their groups median. We then compute
the predicted and actual returns on the high and low portfolios for the next 12 months. For
each return regression in Table 3 and for each of the six size-Bt/Mt groups, Table 5 shows
the average difference between equal-weight predicted high and low returns and the
ARTICLE IN PRESS
average difference between equal-weight actual high and low returns. (Value-weight
returns, omitted to save space, support the same conclusions.) Comparing average actual
return spreads with the spreads predicted for the six size-Bt/Mt groups gives perspective on
which groups deliver the variation in average returns predicted by the regressions. This, in
turn, provides information about whether the regressions are well-specied in different
size-Bt/Mt groups. Table 5 shows average predicted and actual return spreads for all the
return regressions in Table 3, but the discussion below focuses on the regressions (lagged
size, Bt/Mt, protability, asset growth, and accruals as explanatory variables) that produce
incremental return spreads in Table 4.
The baseline again is the return regression with only size and Bt/Mt as explanatory
variables. Within each of the six size-Bt/Mt groups there is variation in size and Bt/Mt
which produces rather large average spreads (from 0.11 to 0.25% per month) in predicted
high minus low returns. The average actual spreads in returns reproduce the predicted
spreads fairly well, with one notable exception. The average return spread for the small
growth group SL, 0.33% per month (t 3.41), is near 75% larger than the spread
predicted by the within-group variation in size and Bt/Mt, 0.19%.
Adding lagged protability and growth to return regressions that also include size and
Bt/Mt as explanatory variables increases the average predicted and actual return spreads
for all six size-Bt/Mt groups, so the relation between these variables and average returns is
general. We infer that the regressions that predict returns with size, Bt/Mt, protability,
and growth are well-specied; they identify patterns in average returns that show up within
all size-Bt/Mt groups. As in the overall return results in Table 4, however, the increments
to average predicted and actual return spreads obtained by adding protability and asset
growth to the return regressions are typically modest, except again for the small growth
group where the predicted average return spread rises by 0.25% per month (more than
twice the increase for any other group) and the actual rises by 0.27%. We infer that there is
wide variation in protability and asset growth among small growth stocks, and it shows
up as predicted in average returns.
Adding lagged accruals to the return regressions that also include lagged size, Bt/Mt,
protability, and asset growth as explanatory variables (Regression 3) increases the
average predicted high minus low return spreads for the six size-Bt/Mt groups by between
0.03% and 0.11% per month. The average spreads in actual returns also increase for all
groups except BH (big value stocks). The increases are modest, except (again) for the small
growth group, where adding accruals to the return regressions causes the average high
minus low return spread to rise from an already impressive 0.60% per month (t 4.50) to
0.86% (t 6.57). This is more than six times the predicted increase, from 0.44% to 0.48%.
In short, adding accruals to the explanatory variables in the return regressions produces
small increases in average high minus low returns for all groups except SL. We infer that
small growth stocks are inuential in the strong average slope for positive accruals in the
return regressions of Table 3.
Our modest incremental returns associated with accruals are in contrast to returns of
about 10% per year found by Sloan (1996) and others for strategies that buy the stocks of
low accruals rms and short rms with high accruals. Why are our results different? First,
earlier return tests do not simultaneously control for size, Bt/Mt, protability, and growth,
to isolate the marginal explanatory power of accruals. Second, the portfolio strategies
examined are typically extreme, buying and shorting equal-weight portfolios of the bottom
and top deciles of accruals. In contrast, we compare the top and bottom halves of
ARTICLE IN PRESS
predicted returns within the six size-B/M portfolios. Our results suggest that small growth
stocks are probably inuential in the large equal-weight returns observed for extreme
strategies.
Fama and French (1993) nd that small growth stocks are a problem for their three-
factor asset pricing model, and Mitchell and Stafford (2000) nd that small growth stocks
are inuential in many high prole event study anomalies. The evidence presented here
suggests that small growth stocks are also inuential in the accruals anomaly.
6. Conclusions
The valuation equation (3) says that controlling for expected protability and
investment, rms with higher book-to-market equity have higher expected stock returns;
given Bt/Mt and expected investment, higher expected protability also implies higher
expected returns; and given Bt/Mt and expected protability, higher expected rates of
investment are associated with lower expected returns.
Our evidence tends to conrm these predictions. Specically our cross-section
regressions say that lagged protability, asset growth, and accruals, used as simple
proxies for expected protability and investment, are related to average returns in the
manner predicted by Eq. (3). The Piotroski (2000) and Ohlson (1980) measures of rm
strength, which are proxies for expected net cash ows (earnings minus investment), are
also related to average returns in the manner predicted by Eq. (3).
A puzzle arises when the tted values from the cross-section regressions to forecast
protability and asset growth are used as proxies for expected protability and investment
in the cross-section return regressions. Many variables contribute to regression forecasts of
protability and asset growth. Thus there is information about expected protability and
asset growth beyond that in lagged protability and asset growth. Better proxies for
expected protability and investment should do a better job identifying the protability
and investment effects in average returns predicted by Eq. (3). But this is not what we
observe. We suggest that the problem is some combination of measurement error in the
tted values from the rst-stage protability and asset growth regressions, and collinearity
between these tted values and the book-to-market variable in the second-stage return
regressions. We argue that the measurement error problem is implicitly resolved by
entering important explanatory variables from the rst-stage protability and asset growth
regressions (lagged protability, growth, accruals, and the PTt and OHt measures of rm
strength) directly as explanatory variables in the second-stage return regressions.
Qualitatively, our results are in line with much existing evidence. It is not a surprise that
book-to-market equity is a powerful variable in describing the cross section of average
stock returns (for example, Fama and French, 1992). Existing evidence also says that more
protable rms have higher expected returns (for example, Haugen and Baker, 1996), and
rms that invest more have lower average returns (for example, Faireld, Whisenant, and
Yohn, 2003). Sloan (1996) and many subsequent papers show that higher current accruals
imply lower future protability and lower future stock returns. Piotroski (2000) and Grifn
and Lemmon (2002) nd that the PTt and OHt proxies for expected net cash ows are
related to average stock returns. At a minimum, our framing of the evidence emphasizes
that all these results are consistent with valuation theory, as summarized in Eq. (3).
Our evidence, however, provides more than perspective on existing results. Previous
work typically examines return effects one variable at a time. In contrast, we examine how
ARTICLE IN PRESS
lagged Bt/Mt, protability, asset growth, accruals, and the PTt and OHt proxies for
expected net cash ows contribute to the description of average returns in tests that
examine incremental effects. Specically, we examine the spreads in realized average
returns obtained when we allocate stocks to high and low expected return portfolios based
on the tted values from cross-section return regressions that successively add variables
identied by us and others as important. The spreads in realized average returns are large,
but the lions share is absorbed by the book-to-market ratio, with an assist from size. The
average high minus low portfolio returns from cross-section regressions that use size and
Bt/Mt to explain returns are 5% to 6% per year. Adding lagged protability and growth to
the regressions increases the average return spreads by less than 1% per year. When we add
accruals to these regressions, the incremental return is again less than 1% per year, and
most of this seems to be due to small growth stocks. Finally, adding PTt and OHt to
regressions that include lagged size, Bt/Mt, protability, asset growth, and accruals as
explanatory variables adds nothing to high minus low average returns.
We emphasize throughout that there is one important issue on which our results are silent:
whether the relations between average returns and Bt/Mt, protability, asset growth, accruals,
PTt, and OHt are due to rational or irrational pricing. To reiterate, the valuation equation (3)
says that expected returns vary with rational assessments of expected protability and growth
(like those we measure) whether pricing is rational or irrational. Irrational beliefs about
expected protability and investment do affect expected returns, through their effects on the
price Mt in Bt/Mt. Tests based implicitly or explicitly on the valuation equation, however,
cannot reveal how the relation between average returns and Bt/Mt splits between differences
in rational risks and the effects of irrational beliefs. In short, despite common claims to the
contrary, tests of Eq. (3) cannot by themselves tell us whether the investor forecasts of
protability and investment that determine Mt are rational or irrational.
7. Appendix
The base accounting variables, from Compustat, are At, total assets (Compustat data
item 6); Yt, income before extraordinary items (18); ACt, accruals [the change in current
assets (4), minus the change in cash and short term investments (1), minus the change in
current liabilities (5), plus the change in debt in current liabilities (34)]; Dt, total dividends
[dividends per share by ex date (26) times common shares outstanding (25)]; and Bt, book
equity [total assets (6), minus liabilities (181), plus balance sheet deferred taxes and
investment tax credit (35) if available, minus preferred stock liquidating value (10) if
available, or redemption value (56) if available, or carrying value (130)]. The accounting
variables for year t are measured at the scal yearend that falls in calendar year t. Market
capitalization MCt (price times shares outstanding) is from CRSP.
We compute the book-to-market ratio for year t, Bt/Mt, as book equity for the scal
yearend in calendar year t divided by market equity at the end of December of t. The
market cap variable, ln MCt, used to measure size in the protability and growth
regressions of Tables 1 and 2, is measured at the scal yearend. The market cap variable, ln
MCt+1, used to measure size in the Table 3 return regressions for July of t+1 to June of t+2
and when assigning rms to the six size-Bt/Mt portfolios at the end of June of t+1 in Table
4 is for the end of June of t+1.
We compute two summary measures of rm strength. The rst, OHt, is a mea-
sure of bankruptcy risk developed by Ohlson (1980). Ignoring the constant, OHt is
ARTICLE IN PRESS
dened as
OH t 4:07 ln At 6:03 Lt =At 1:43CAt CLt =At 0:0757 CLt =CAt

2:37 NI t =At 0:285 Losst 1:72NegBookt 0:521DNI t 1:83 Opt =Lt ,
4
where ln At is the natural log of assets; Lt is liabilities (Compustat item 181); CAt is current
assets (4); CLt is current liabilities (5); NIt is net income (1 7 2); Losst is 1 if net income is
negative in t and t-1, and 0 otherwise; NegBookt is 1 if liabilities exceed assets and 0
otherwise; DNIt is the change in net income from t-1 to t divided by the sum of the absolute
values of net income in t-1 and t, (NIt NIt1)/(|NIt1|+|NIt|); and Opt, funds from
operations, is earnings before extraordinary items (18), plus income statement deferred
taxes (50), if available, plus equitys share of depreciation expense, which we dene as
MCt/(AtBt+MCt) times total depreciation expense (14).
The second composite measure of rm strength, PTt, is from Piotroski (2000). It
is the sum of nine binary variables, each equal to 1 if a given condition holds and 0
otherwise. The nine conditions are: (1) income before extraordinary items, Yt, is positive;
(2) cashow from operations, CFOt, is positive; (3) the change in the return on
assets, dened as income before extraordinary items at yearend divided by assets at
the beginning of the year, Yt/Bt, is positive; (4) cashow from operations exceeds
income before extraordinary items; (5) the change in leverage, dened as long-term
debt at scal yearend (Compustat items 9 and 44) divided by assets at yearend, is negative;
(6) the change in liquidity, dened as current assets divided by current liabilities, is
positive; (7) the change in the gross margin ratio, dened as one minus the ratio of the
cost of goods sold (41) to sales (12), is positive; (8) the change in turnover, dened
as sales divided by beginning of year assets, is positive; and (9) the company has a positive
cashow from the sale of common and preferred (108). The changes are measured
from year t-1 to t. If the Compustat format code for the statement of cashows (310)
indicates the company does report a statement of cashows (format code 7), cashow
from operations, CFOt, is net cash from operating activities (308). If the company
reports a statement of working capital (format code 1), CFOt is funds from operations,
Opt, minus other changes in working capital (236, if available). For other format codes,
CFOt is funds from operations, Opt, plus other changes in working capital (if available).
Because each binary variable is 0 if a condition does not hold, PTt increases with rm
strength.
Analyst earnings forecasts are from Thomson Financials I/B/E/S database. It is the
median forecast of earnings per share for scal year t+1 that is available at the end of scal
year t. (Using the forecasts for year t+2 or t+3 does not change the results materially.) It/Bt
is the median forecast at the end of scal year t times the I/B/E/S split factor for that
month (to reverse adjustments I/B/E/S makes for stock splits that occur after t) times the
shares outstanding at t reported by Compustat (data item 25) divided by book equity for
scal year t. We do not use an I/B/E/S forecast if the split adjusted version of the stock
price reported by I/B/E/S at the end of the scal year differs by more than 5% from the
price reported by CRSP.
All variables except ln MCt and ln At (in OHt) are on a per share basis. We use CRSPs
share factor (FACSHR) for stock splits and stock dividends (distribution codes 5510-5559)
to adjust nonsynchronous variables, such as Yt+1/Bt and dAt+1/At.
ARTICLE IN PRESS
To reduce the inuence of outliers, we delete a rm from the year t protability and
growth regressions of Tables 1 and 2 if an explanatory variable is outside the 0.5 or 99.5
percentile for that variable in year t. But we do not delete rms with extreme values when
computing the tted values from the protability and growth regressions for use as
explanatory variables in the return regressions of Table 3. Instead, we winsorize the
explanatory variables in the rst-pass regressions at the 0.5% level, shrinking extreme
values to the 0.5 and 99.5 percentiles for year t. Thus we delete rms with variables outside
the 0.5 and 99.5 percentiles when estimating the protability and growth regressions, but
we shrink extreme values when estimating expected protability and growth for the return
regressions. We consider only the upper or lower bound for one-sided variables, such as
+ACt/Bt and ACt/Bt. We also winsorize other independent variables in the return
regressions at the 0.5 and 99.5 percentiles.
References
Abarbanell, J.S., Bernard, V.L., 1992. Tests of analysts overreaction/underreaction to earnings information as an
explanation for anomalous stock price behavior. Journal of Finance 47, 11811207.
Abarbanell, J.S., Bushee, B.J., 1998. Abnormal returns to a fundamental analysis strategy. Accounting Review 73,
1945.
Ahmed, A.S., Nainar, S.M.K., Zhang, X.F., 2003. Further evidence on analyst and investor mis-weighting of
prior period cash ows and accruals. Unpublished working paper. University of Chicago, Chicago, IL.
Ali, A., Klein, A., Rosenfeld, J., 1992. Analysts use of information about permanent and transitory earnings
components in forecasting annual EPS. Accounting Review 67, 183198.
Campbell, J., Shiller, R.J., 1988. The dividend-price ratio and expectations for future dividends and discount
factors. Review of Financial Studies 1, 195228.
Capaul, C., Rowley, I., Sharpe, W.F., 1993. International value and growth stock returns. Financial Analysts
Journal, JanuaryFebruary, 2736.
Chan, K., Chan, L.K.C., Jegadeesh, N., Lakonishok, J., 2006. Earnings quality and stock returns. Journal of
Business 79, 10411082.
Chan, L.K.C., Hamao, Y., Lakonishok, J., 1991. Fundamentals and stock returns in Japan. Journal of Finance
46, 17391789.
Cohen, R.B., Gompers, P.A., Vuolteenaho, T., 2002. Who underreacts to cash-ow news? Evidence from trading
between individuals and institutions. Journal of Financial Economics 66, 409462.
Collins, D.W., Hribar, P., 2000. Earnings-based and accrual-based market anomalies: one effect or two? Journal
of Accounting and Economics 29, 101123.
Dechow, P.M., Hutton, A.P., Sloan, R.G., 2000. An empirical assessment of the residual income valuation model.
Journal of Accounting and Economics 26, 134.
Easterwood, J.C., Nutt, S.R., 1999. Inefciency in analysts earnings forecasts: Systematic misreaction or
systematic optimism? Journal of Finance 54, 17771797.
Faireld, P.M., Whisenant, S., Yohn, T.L., 2002. The differential persistence of accruals and cashows for future
operating income versus future return on assets. Unpublished working paper. Georgetown University,
Washington, DC.
Faireld, P.M., Whisenant, S., Yohn, T.L., 2003. Accrued earnings and growth: implications for future
protability and market mispricing. Accounting Review 78, 353371.
Fama, E.F., French, K.R., 1992. The cross-section of expected stock returns. Journal of Finance 47, 427465.
Fama, E.F., French, K.R., 1993. Common risk factors in the returns on stocks and bonds. Journal of Financial
Economics 33, 356.
Fama, E.F., French, K.R., 1995. Size and book-to-market factors in earnings and returns. Journal of Finance 50,
131155.
Fama, E.F., French, K.R., 2001. Disappearing dividends: changing rm characteristics or lower propensity to
pay. Journal of Financial Economics 60, 343.
Fama, E.F., MacBeth, J., 1973. Risk, return, and equilibrium: empirical tests. Journal of Political Economy 81,
607636.
ARTICLE IN PRESS
Frankel, R., Lee, C.M.C., 1998. Accounting valuation, market expectation, and cross-sectional stock returns.
Journal of Accounting and Economics 25, 283319.
Grifn, J.M., Lemmon, M.L., 2002. Does book-to-market equity proxy for distress risk or mispricing? Journal
of Finance 57, 23172336.
Haugen, R.A., Baker, N.L., 1996. Commonality in the determinants of expected stock returns. Journal of
Financial Economics 41, 401439.
Lakonishok, J., Shleifer, A., Vishny, R.W., 1994. Contrarian investment, extrapolation, and risk. Journal
of Finance 49, 15411578.
Lee, C.M.C., Ng, D., Swaminathan, B., 2004. International asset pricing: evidence from the cross section of
implied cost of capital. Unpublished working paper. Cornell University, Ithaca, NY.
Mitchell, M.L., Stafford, E., 2000. Managerial decisions and long-term stock price performance. Journal
of Business 73, 287329.
Ohlson, J.A., 1980. Financial ratios and the probabilistic prediction of bankruptcy. Journal of Accounting
Research 18, 109131.
Penman, S.H., 1991. An evaluation of accounting rate of return. Journal of Accounting, Auditing, and Finance 6,
233255.
Piotroski, J.D., 2000. Value investing: the use of historical nancial statement information to separate winners
from losers. Journal of Accounting Research 38 (Supplement), 142.
Richardson, S.A., Sloan, R.G., 2003. External nancing and future stock returns. Unpublished working paper.
University of Pennsylvania, Philadelphia, PA.
Richardson, S.A., Sloan, R.G., Soliman, M.T., Tuna, I., 2004. The implications of accounting distortions and
growth for accruals and protability. Unpublished working paper. University of Pennsylvania, Philadelphia,
PA.
Richardson, S.A., Sloan, R.G., Soliman, M.T., Tuna, I., 2005. Accrual reliability, earnings persistence, and stock
prices. Journal of Accounting and Economics 39, 437485.
Rosenberg, B., Reid, K., Lanstein, R., 1985. Persuasive evidence of market inefciency. Journal of Portfolio
Management 11, 917.
Sloan, R.G., 1996. Do stock prices fully reect information in accruals and cash ows about future earnings?
Accounting Review 71, 289315.
Titman, S., Wei, K.C.J., Xie, F., 2004. Capital investments and stock returns. Journal of Financial and
Quantitative Aanalysis 39, 677700.
Xie, H., 2001. The mispricing of abnormal accruals. Accounting Review 76, 357373.

1 s2.0 S0304405X06001164 Main 2 PDF

Uploaded by

Copyright:

Available Formats

1 s2.0 S0304405X06001164 Main 2 PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1 s2.0 S0304405X06001164 Main 2 PDF

Uploaded by

Copyright:

Available Formats

ARTICLE IN PRESS

Journal of Financial Economics 82 (2006) 491518

Protability, investment and average returns$

JEL classification: G12

Keywords: Average returns; Protability; Investment; Book-to-market equity

or, dividing by time t book equity,

2. Tests of valuation equations: strengths and weaknesses

3. Expected protability and investment

Regressions to predict asset growth, dAt+t/At

Regressions to predict profitability, Yt+t/Bt

Regressions to predict asset growth, dAt+t/At

3.1. Asset growth

4. Expected returns: cross-section regressions

4.1. Baseline tests

Average 2,058 1.66 0.28 0.06 0.02

Average 2,253 1.35 0.34 0.07 0.04 0.06 0.03

Expected profitability and growth estimated with lagged fundamentals

4.2. Better proxies for expected profitability and investment

5. Expected returns: portfolio tests

5.1. Economic significance

Average predicted spread Average actual spread t(Average actual spread)

1 0.42 0.49 0.52 0.43 4.66 3.42

Average spread in expected returns

(1980) measures of rm strength, which aggregate the information in many accounting

5.2. Pervasiveness and regression specification

OH t 4:07 ln At 6:03 Lt =At 1:43CAt CLt =At 0:0757 CLt =CAt

You might also like