Professional Documents
Culture Documents
North-Holland
S.P. Kothari
Unioersily qf Rochester. Rochester. NY 14627, USA
Richard G. Sloan
Uniuersily of Pennsylvania. Philadelphia. PA 19104-6365, USA
Stock return over a period reflects the market’s revision in expectation of future earnings. Account-
ing earnings over the same period, however, have limited ability to reflect such revised expectations.
Therefore, returns anticipate earnings changes and the earnings response coefficient from a regres-
sion of returns on contemporaneous earnings changes is biased toward zero. We reduce this bias by
including leading-period returns in pricexarnings regressions. The resulting estimated earnings
response coefficient magnitudes suggest that the capital market, on average, views earnings changes
to be largely permanent. This is consistent with the random walk time series property of annual
earnings.
1. Introduction
Recently the relation between time series properties of earnings and earnings
response coefficients [formally defined below in eq. (l)] has been a topic of
considerable interest among accounting researchers [see, e.g., Kormendi and
Lipe (1987), Collins and Kothari (1989), Easton and Zmijewski (1989)J Since
*We acknowledge Celal Aksu, Andrew Alford, Vie Bernard, Fischer Black, Carol Frost, Dan
Givoly, Paul Healy, Pat Hughes, Bob Lipe, Tom Lys, Jim Manegold, Krish Palepu, K. Ramesh,
Doug Skinner, Charlie Wasley, Ross Watts, Peter Wilson, Jerry Zimmerman, Mark Zmijewski,
workshop participants at the University of Arizona, Harvard University, Northwestern University,
Ohio State University, University of Southern California, Temple University, Washington Univer-
sity at St. Louis, and especially an anonymous referee and Ray Ball (the editor) for many useful
comments. S.P. Kothari acknowledges financial support from the Bradley Policy Research Center at
the Simon School and the John M. Olin Foundation.
time series properties of annual earnings suggest that firms’ earnings changes are
largely permanent, the earnings response coefficient is predicted to be about
(1 + l/ri), where ri is the expected rate of return on firm i’s equity [Kormendi
and Lipe (1987) and Collins and Kothari (1989)l.l Earnings response coeffi-
cients reported in the literature, however, are considerably smaller than implied
by the time series properties of earnings. For example, Penman’s (1990, table 2)
estimate using annual returns/earnings data is 0.894, whereas Kormendi and
Lipe (1987, table 1) report a median coefficient of 2.5. Ali and Zarowin (1992),
who control for the effect of serial correlation in earnings, report a median
earnings response coefficient of 1.59. Use of analysts’ earnings forecasts as better
proxies for the market’s expectation also yields coefficients of similar mag-
nitudes [see, for example, Easton and Zmijewski (1989a,b) and Brown, Griffin,
Hagerman, and Zmijewski (1987) who use Value Line’s quarterly earnings
forecasts].
In this paper we resolve the apparent inconsistency between the time series
behavior of annual earnings and the market’s valuation of earnings changes, as
reflected in an earnings response coefficient estimate, by using the information in
stock prices about future earnings. By exploiting the earnings forecasts em-
bedded in security prices, we obtain coefficients that are comparable to those
implied by the time series properties of earnings.
One source of stock prices’ predictive ability with respect to future earnings
changes is that the information set reflected in the market’s expectations, and
thus in prices, is richer than that in the past time series of earnings. Price change
over a period reflects revision in the market’s expectation of future earnings as
well as realized earnings over the period. In comparison, accounting earnings
over a period primarily summarize: (i) the effect of sales transactions during
a fiscal period that have generated cash or almost certainly will generate cash
(e.g., cash sales and net change in receivables), (ii) the effect of past periods’
activities (e.g., via depreciation expense and cost of goods sold), and (iii) cash
outlays for investments generating uncertain future benefits (e.g., research and
development and advertising expenses). Compared to stock prices, earnings thus
have a limited ability to contemporaneously reflect the market’s revised expec-
tations of future cash flows. Such limited ability is attributable to the conserva-
tism, objectivity, verifiability, and other conventions that underlie Generally
Accepted Accounting Principles (GAAP). For example, earnings for a period are
unlikely to reflect the expected net cash flow from a new growth opportunity
facing the firm. Since, on average, the market’s expectations eventually get
reflected in earnings, price changes anticipate earnings changes, or prices lead
‘Annual earnings changes exhibit a small degree of negative serial correlation [see, for example,
Ball and Watts (1972, tables 3.4). Brooks and Buckmaster (I 980), and Ali and Zarowin (1992)]. For
a typical firm, we therefore expect its earnings response coefficient to be somewhat smaller than
(1 + l/r,).
S.P. Kothari and R.G. Sloan, Information in prices about firure earnings 145
earnings2 This phenomenon has long been recognized in the literature. For
evidence in the accounting literature, see Ball and Brown (1968), Beaver, Lam-
bert, and Morst (1980) Brown, Foster, and Noreen (1985), Collins, Kothari, and
Rayburn (1987), Freeman (1987) and Collins and Kothari (1989).
The effect of prices leading earnings on an earnings response coefficient, can
be seen from the following commonly estimated regression:
‘We hasten to add that by ‘prices lead earnings’ we do not imply there is any ‘error’ in historical
cost accounting earnings. While managers have incentives to ‘manipulate’ accruals [e.g., Healy
(1985)], it is neither necessary nor do we rely on such ‘manipulations’ or ‘error’ to argue that returns
anticipate earnings changes. Moreover, since accrual ‘manipulations’ do not have a first-order stock
price effect, returns are unlikely to be useful in forecasting future earnings changes stemming from
accrual ‘manipulations’. The latter are probably helpful in forecasting future earnings changes on the
basis of past earnings changes.
‘Collins and Kothari (1989) also include leading-period returns in estimating price-arnings
regressions. They measure return over a 15-month window that begins five (two) months earlier than
the fiscal period t for the large (small) firms. We include up to three leading years’ returns to more
fully exploit the information in prices with respect to future earnings.
146 S.P. Kothari and R.G. Sloan, Information in prices about future earnings
4The generally observed large price run-up in a target firm’s stock can also be viewed as a growth
opportunity in that the bidder firm is expected to utilize the target firm’s assets more efficiently and
generate greater amounts of cash flows, that will be reflected in future earnings.
1.A.E B
148 S.P. Kothari and R.G. Sloan, Information in prices about fulure earnings
P
t-2
P
I-3
Fig. 1. Earnings and returns measurement intervals in lead-lag price+arnings regressions. Two-
year buy-and-hold returns, inclusive of dividends, regressed on annual earnings deflated by price at
the beginning of the return measurement interval. Return measurement interval consists of the
contemporaneous and one leading year. Return observations are overlapping.
where Xi, is accounting earnings over the period t, Pit/Pit_, is one plus the
buy-and-hold return, inclusive of dividends, over the period from the end of
S.P. Kothari and R.G. Sloan, Information in prices about future earnings 149
‘We also considered a reverse regression of annual earnings on annual returns, in the spirit of
Beaver, Lambert, and Ryan (1987) to estimate earnings response coefficients. Since prices lead
earnings, the return variable contains information about the current year’s earnings (the dependent
variable in a reverse regression) as well as the future years’ earnings. The presence of the latter type of
information in the return variable means it measures the variable of interest (i.e., the component of
return that is attributable to the current period’s earnings information) with error. The estimated
slope is therefore expected to be biased, much the same way as in the direct regression. See Lipe
(1990, p. 64) and Beaver et al. (1987, pp. 150-151).
150 S.P. Kothari and R.G. Sloan, Information in prices about future earnings
3.1. Data
We use data from the Compustat Quarterly, Annual Industrial, and Annual
Res earth tapes, and the Center for Research in Security Prices (CRSP) monthly
tape. We use earnings data for 1979-88 from the quarterly tape for earnings
measurement intervals shorter than one year and data for 1950-88 from the
annual tapes when the earnings measurement interval is one year or longer.
We use monthly returns to calculate buy-and-hold returns including dividends
over holding periods ranging from one quarter to four years. Returns are meas-
ured over intervals that end in the earnings fiscal quarter- or year-end. The effect
of including earnings-announcement and post-earnings-announcement-period re-
turns on earnings response coefficients is discussed in sections 4.2 and 4.3.
We include all firms that have at least five consecutive earnings and return
observations available for the measurement interval of interest. We apply this
critierion separately for each measurement interval. Thus, if the measurement
inverval is one quarter, then earnings and return data for at least five consecu-
tive quarters must be available. This enables us to estimate a time series
regression separately for each firm.
Table 1 reports descriptive statistics for the earnings yield (X,_,,,/P,_,) and
price relative (P,/P,_,) variables. P,/P,_, is one plus the buy-and-hold return,
inclusive of dividends, over the period t - z to t. The descriptive statistics are
reported for measurement intervals, T, ranging from one to four years. For each
measurement interval, the average earnings yield and price relative are cal-
culated for each firm. Using these firm level data, sample mean, standard
deviation, median, minimum, and maximum values are obtained. For T = 1
year, the average earnings yield for a sample of 2,721 firms is 6.1% and the
median is 9.1%. The mean and median earnings yields for the four-year-interval
sample of 2,319 firms are 13.2% and 12.7%. The mean as well as median
earnings yields reported in table 1 increase monotonically with 7. One reason is
that the earnings per share is deflated by P,_, which, on average, declines in 5.
A second reason is that there is a survivorship bias that results in higher
earnings yield samples as the measurement interval is increased. A survivorship
bias also reduces the sample size from 2,721 for 5 = one year to 2,319 for
T = four years.
Descriptive statistics for P,/P,_, for the various measurement interval sam-
ples are reported in the bottom panel of table 1. The last row reports annualized
mean values. The annualized mean price relatives measured over all intervals
are almost identical. They range from 1.168 to 1.170. This means the average
S.P. Kothari and R.G. Sloan, Information in prices about future earnings 151
Table 1
Descriptive statistics for earnings yield (X~_~,JP~_,) and price relative (Pr/Pr_,) variables measured
over intervals ranging from one to four years and data from 1950-88. Sample sizes are from 2,319 to
2,721 firms, l'b
Measurement interval
Earnings yield
Mean 0.061 0.091 0.112 0.132
Std. deviation 0.156 0.121 0.131 0.143
Median 0.091 0.104 0.116 0.127
Minimum - 2.874 - 2.072 - 2.534 - 2.020
Maximum 0.639 0.856 1.045 1.400
N 2,721 2,588 2,471 2,319
Price relatives
Mean 1.168 1.371 1.598 1.864
Std. deviation 0.139 0.310 0.543 0.788
Median 1.165 1.345 1.532 1.734
Minimum 0.527 0.288 0.102 0.221
Maximum 2.521 3.803 7.492 9.867
Annualized mean c 1.168 1.170 1.169 1.168
aSample selection: Any firm that had earnings data on the Compustat Annual Industrial or the
Compustat Annual Research tape and return data on the Center for Research in Security Prices
monthly tape for five consecutive measurement interval periods during 1950-88 is included. Data for
the longer than one-year measurement intervals contain overlapping periods because the measure-
ment window is moved forward only one year at a time.
~Earnings yield X , _ t , r / P , _ ~ primary annual earnings per share before extraordinary items and
=
discontinued operations for year t, divided by the price at the beginning of the return measurement
interval. All earnings per share numbers and prices are adjusted for stock splits and stock dividends.
Price relative Pf/Pr • = one plus the buy-and-hold return, inclusive of dividends, over the years
t-Ttot.
cAnnualized price relative is calculated as the geometric mean annual price relative for the
particular measurement interval.
a n n u a l r e t u r n o n t h e s a m p l e f i r m s is 1 6 . 8 % t o 1 7 . 0 % . T h e s a m p l e f i r m s ' r e t u r n s
t h u s d o n o t r e v e a l a n y o b v i o u s s u r v i v o r s h i p b i a s p r o b l e m as a f u n c t i o n o f t h e
return measurement interval. The mean and median raw price relatives nat-
urally increase with the measurement interval. The mean price relative increases
f r o m 1.168 f o r t h e o n e - y e a r - i n t e r v a l s a m p l e t o 1.864 for t h e f o u r - y e a r - i n t e r v a l
sample.
4. Regression results
Table 2
Selected fractiles from the distribution of estimated parameters of firm-specific regressions of price
relatives on earnings deflated by price. Earnings measurement interval is one year and price relatives
are for the contemporaneous one year and for periods that also include leading years. Annual
earnings data from 1950-88. Sample sizes from 2,319 to 2,721 firms?
Pit/Pit-r = ?io~ + 7iltXu/Pit-t + Eit b'c
aSampte selection: Any firm that has earnings data on the Compustat Annual Industrial or
Research tape for five consecutive years during 1950-88 and return data for the earnings measure-
ment years and the three leading years is included. The return observations contain overlapping
periods when one or more years of leading-period returns are included.
bEarnings yield Xt/Pt-~ = primary annual earnings per share before extraordinary items and
discontinued operations for year t, divided by the price at the beginning of the return measurement
interval. All earnings per share numbers and prices are adjusted for stock splits and stock dividends.
Price relative Pt/Pt-r one plus the buy-and-hold return, inclusive of dividends, over the years
=
t-~tot.
CThe model is estimated by modeling the residuals from the ordinary least squares regressions as
a first-order autoregressive process and then ordinary least squares parameters are reestimated
using transformed variables. The estimated slope coefficients are deflated by (1.039)'- 1 to control for
bias stemming from using dividend-reinvested returns.
S.P. Korhari and R.G. Sloan, Information in prices about future earnings 153
While we provide an approximate correction for this bias, we also avoid adding
unnecessary leading-period returns.
When we include leading-period returns (i.e., r > l), we obtain overlapping
return observations, as shown in fig. 1, and the regression errors become serially
correlated (we discuss below an adjustment for serial correlation in errors).
Under these conditions, ordinary least squares estimators are unbiased, but less
efficient than generalized least squares estimates that take account of the
autocorrelation [Maddala (1977, p. 281)]. We therefore transform the price-
earnings variables by modelling residuals from the ordinary least squares
regressions as a first-order, autoregressive process and then reestimate the
priceearnings regression using the transformed variables. This two-step, full
transform procedure is similar to the Cochrane and Orcutt (1949) procedure.
We also estimate ordinary least squares regressions and regressions using
nonoverlapping return data. The former are discussed in the ‘diagnostic tests’
section and the latter are discussed later in this section.
We report regression results in table 2, comparing the average estimated
coefficient with its predicted value. By focusing on the average coefficient, we
draw inferences about the price-earnings relation of the average security in our
sample. The average coefficient is predicted to be (1 + l/r), where I is the average
security’s expected rate of return on equity.6 We use the average realized annual
return of 17.15% on the CRSP equal-weighted portfolio from 192688 as
a proxy for the expected rate of return. We consider this to be a reasonable
proxy for expected return on the sample firms because the average return on the
sample firms, as reported in table 1, is close to the CRSP equal-weighted return
over 192688 and the market model average beta for the sample using the CRSP
equal-weighted portfolio is 1.Ol. For r = 17.15%, the average earnings response
coefficient is predicted to be 6.83.’
As our hypothesis predicts, the average earnings response coefficient increases
as returns from leading years are included. The average earnings response
coefficient from the annual contemporaneous regressions for 2,721 sample firms
is 2.56. The median earnings response coefficient is 1.89, indicating that the
‘Since (I/N)x(l/r,) 2 l/r, where N is the number of securities, ri is the expected return on security
i, and r is the average expected return for the securities, (1 + l/r) underestimates the predicted
average coefficient. However, since expected return on equities is almost invariably greater than
8&10%, and for reasonable values of cross-sectional variance of the expected returns on equities, the
bias in using (1 + I/r) as the predicted coefficient can be shown to be relatively small. Throughout in
this paper we compare the estimated average earnings response coefficient against (1 + l/r).
7An alternative to using the CRSP realized rate of return as a proxy for r would be to use a proxy
for the expected real rate of return on common stocks. If inflation is assumed to have no real effect
on valuation, then the latter proxy would be consistent with the valuation model underlying the
price-earnings relation. We, however, do not adjust the average realized rate of return on the CRSP
equal-weighted portfolio for the average inflation rate of 3.2% over 192687 [Ibbotson and
Sinquefield (1989)] because it is not obvious how inflation affects the price-arnings multiple. If we
use a proxy for the expected real rate of return, then the predicted value of the earnings response
coefficient is 8.17.
154 S.P. Kothari and R.G. Sloan, It$wnation in prices ahour future earnings
distribution is skewed to the right. More than 90% of the coefficient estimates
are positive. As one, two, and three leading years’ returns are added to the
contemporaneous annual returns, the average earnings response coefficient
increases to 4.69, 5.08, and 5.45 (a 113% increase). Taken together, leading-
period returns are at least as important as contemporaneous returns in terms of
their sensitivity with annual earnings. The median earnings response coefficients
always are lower than the sample means, but they too increase substantially with
the inclusion of leading-period returns. For example, when three leading years’
return is included, the median earnings response coefficient is 4.42.
We next compare the estimated coefficients with their predicted values. The
sample mean’s standard error is affected by (i) cross-correlation among the
coefficient estimates, (ii) variation in the ‘true’ coefficient magnitudes across
firms, and (iii) cross-sectional variation in the degree to which stock prices lead
earnings. Since the price-earnings regressions’ errors are expected to be posit-
ively cross-correlated, the estimated standard errors will likely understate the
‘true’ standard error.8 Cross-sectional variation in the firms’ earnings response
coefficients due to variation in their expected returns will also affect the estim-
ated standard errors. Since we do not model individual firm’s earnings response
coefficients, this will tend to overstate standard errors. However, as seen from
the 10th and 90th percentiles of their distribution in table 2, many firms’
earnings response coefficient estimates appear to substantially deviate from their
predicted values. We, therefore believe sampling variation in the coefficient
estimates dominates the estimated standard errors. Cross-sectional variation in
the degree to which prices lead earnings will also magnify the estimated stan-
dard errors, particularly when return over a short leading period is included. As
T is increased, virtually all the effect of earnings anticipation is incorporated in
the price-earnings relation. Therefore, for t = 4, we expect cross-sectional
variation in the degree to which prices lead earnings will have relatively little
effect on the standard error.
The estimated average coefficient is closer, both in absolute terms and
statistically, to the predicted value of 6.83 as the leading-period return is
incorporated in the analysis. For all leading periods from one to three years,
however, the estimated earnings response coefficient is at least two standard
errors smaller than the predicted value of 6.83. For example, when three
leading-year returns are included, the earnings response coefficient estimate is
5.45, with a standard error of 0.1198. The null hypothesis that the estimated
average coefficient is equal to its predicted value of 6.83 is rejected at a p-value
less than 0.01.
“We report below results of regressions that include the market return as an additional indepen-
dent variable which controls the positive cross-sectional correlation among residuals. As a result, the
estimated standard error of the cross-sectional average earnings response coefficient will be less
biased.
S.P. Kothari and R.G. Sloan, 155
and draw inferences from the average earnings response coefficient estimates for
the following reason.
Expected-return variation through time for individual stocks is small com-
pared to the expected-return variation across stocks, particularly when
measurement intervals are long.” The effect of expected-return variation
through time is that the estimated earnings response coefficient, on average, is
approximately equal to (1 + l/ravg), where ravg is the average expected return
over the sample period. The reason for focussing on earnings response coeffic-
ients is that price-earnings regressions’ explanatory power is affected by the
expected-return variation and the problem is acute in a cross-sectional analysis.
The proportion of cross-sectional variation in returns (the dependent variable in
price-earnings regressions) attributable to risk differences is an increasing func-
tion of the measurement interval. Because the independent variable, X, _ ~,f/P, _~,
in price-earnings regressions is a risk proxy [e.g., Ball (1978)], the cross-
sectional association between returns and the X,_,.,/P,_, variable is due to both
unexpected earnings and risk differences. Variation due to unexpected earnings
yield, on average, will tend to be a smaller fraction of the cross-sectional
variation in X,-,.,/P, _rwhen earnings are aggregated, i.e., as the measurement
interval is increased. Variation due to expected earnings yield, on the other
hand, will constitute an increasing fraction of the cross-sectional variation in
X, r,, /Pr _r.Therefore the cross-sectional price-earnings association will also be
due increasingly to cross-sectional differences in expected earnings yields. As
noted above, variation in expected return (and therefore expected earnings yield)
does not bias earnings response coefficient estimates.
We estimate the following time series model for each firm i:
where r = one, two, or three quarters or one, two, three, or four years and
Xir_r,, = firm i’s earnings from t - T to t. When 5 = two quarters, earnings for
two quarters are summed, and when 5 = two years, annual earnings for two
years are summed.
In estimating (3), data availability requirements are more restrictive than
using only annual earnings. As a result, sample sizes decline substantially with
longer price-earnings measurement windows. To the extent expected rates of
return differ across the samples, the estimated earnings response coefficients are
not strictly comparable. Because the smaller samples at the longer measurement
intervals are likely to consist of larger-sized firms with relatively low expected
rates of return, ceteris parihus, their earnings response coefficients are expected
‘?3nce expected returns are positively autocorrelated and mean reverting [e.g., Fama and French
(1988) and Poterba and Summers (1988)], there is relatively less variation in expected returns
through time over longer measurement intervals.
S.P. Kothari and R.G. Sloan, Information in prices about future earnings 157
Table 3
Selected fractiles from the distribution of estimated parameters of firm-specific regressions of price
relatives on contemporaneous earnings deflated by price. Earnings measurement interval ranging
from one quarter to four years. One-, two-, and three-quarter priceearnings data are from 1979-88
and annual and longer-interval data are from 1950-88. Sample sizes are from 1,002 to 2,721 firms.”
PeIP,,+, = ‘ii07 + BiI~Xit-LIIPil-i + sitb.C
“For the one-, two-, and three-quarter interval price-earnings data, any firm that had earnings
data on the Compustat quarterly tape and return data on the Center for Research in Security Prices
monthly tape for five nonoverlapping measurement intervals during 1979-88 is included. For the
annual and longer measurement intervals, the Compustat Annual Industrial and the Compustat
Annual Research tapes are used from 1950-88.
bX.,,_,,,/Pi,_, = primary earnings per share before extraordinary items and discontinued opera-
tions, summed over the relevant measurement interval r and then divided by the price at the
beginning of the earnings measurement interval. Price relative P,/Pt_, = one plus the buy-and-hold
return, inclusive of dividends, over the interval t - Tto t. All earnings and price data are adjusted for
stock splits, stock dividends, and dividend payments,
‘The model is estimated by modeling the residuals from the ordinary least squares regressions as
a first-order autoregressive process, and then ordinary least squares parameters are reestimated
using transformed variables.
“Annualized slope coefficient = [1 + r(y^,, - l)], where )l,r is the estimated coefficient from the
regressions. The annualized slope coefficients are deflated by 1.039 to control for bias stemming from
using dividend-reinvested returns.
158 S.P. Kothari and R.G. Sloun, I~formaiion in prices about future earnings
“Somewhat surprisingly. the average coefficient for the three-quarter measurement interval is
2.97 which is greater than that for the annual measurement interval. This is due to a different sample
and time period for the quarterly earnings data compared to the annual and longer-interval earnings
data. The median earnings response coefficients, however, increase monotonically with the measure-
ment interval.
S.P. Korhari and R.G. Sloan, Information in prices about future earnings 159
a factor of (1 + dividend yield)‘- ’ and the bias arises because of the asymmetric
effect of dividend payments on returns versus earnings. This bias should be
reduced if returns exclusive of dividends are used. We therefore repeat the entire
analysis using price relatives exclusive of dividends. If all firms’ dividend payouts
are lOO%, then it is easy to show that using returns excluding dividends leads to
an earnings response coefficient of l/r [see e.g., Beaver, Lambert, and Morse
(1980)-J. For a less than 100% divided payout, they predicted earnings response
coefficient is (1 - dividend payout + l/r). The estimated average coefficient
using returns excIuding dividends for none to three leading years are 2.41,4.35,
4.57, and 4.74. These magnitudes are uniformly smaller than those reported
above for the corresponding return measurement intervals in table 2, which
range from 2.56 to 5.45. As expected, the estimates using returns exclusive of
dividends are not smaller by one than those using with-dividends returns
because firms’ dividend payouts are less than 100%. The pattern of the earnings
response coefficient estimates as a function of leading-period returns, however, is
similar to that noted earlier.
We also estimate regressions by including up to nine leading years’ returns. As
more than three leading years’ returns are included in the analysis, the average
coefficient estimated using returns exclusive of dividends shows little systematic
increase. It changes from 4.74 when three leading years’ returns are included to
4.70 when six leading years’ returns are included to 4.64 when nine leading years’
returns are included. The corresponding medians are 3.75, 3.78, and 3.64. The
average and median coefficients are thus largely insensitive to including addi-
tional leading-period returns.
When more than three leading years’ returns inclusive of dividends are added,
the coefficient increases from 5.45 to 6.13 to 6.83 as three, six, and nine leading
years’ returns are included. This means the coefficient, on average, increases by
3.8% per year beyond three leading years. While it is not obvious whether this
small increase is a spurious one (because the sample size declines with additional
leading-period returns), it is too small to explain the increase in the average
coefficient from 2.56 when no leading-period return is included to 5.45 when
three leading years’ returns are included.12
We also assess whether the increase in the average coefficient is due to the
changing sample composition as leading period returns are included in the
analysis. We estimate earnings response coefficients for the subsample for which
price-earnings data are available over all the return mesurement intervals. We
obtain a sample of 1,669 firms. As none to three leading years’ returns are
included, the average coefficients for this sample are 2.52, 4.60, 4.88, and 5.07.
These coefficient values are very similar to those reported in table 2 where the
‘*A portion of the increase in the earnings response coefficient estimates could arise if the same
dividend yield is greater than the dividend yield for the CRSP equal-weighted portfolio that we use
in our adjustment.
160 S.P. Kothari and R.G. Sloan, Information in prices about future earnings
We next report the results of estimating a model similar to (2) except that the
return on the market is also included as an explanatory variable.14 The motiva-
tion for including the market return variable is threefold: (i) The average
cross-correlation among the residuals will be closer to zero when the market
return is included as an independent variable. The standard error of the sample
mean of the earnings response coefficients will then be less biased due to
cross-sectional correlation, which will enable us to perform more reliable statist-
ical significance tests of the average earnings response coefficient. (ii) If some of
the price movement is not explained by the earnings variable, but is related to
the market return, then potentially statistically more precise estimates of the
earnings response coefficients are possible by including the market return
variable [see Sloan (1993) for evidence]. The standard error of the average
earnings response coefficient is therefore likely to decrease with the inclusion of
the market return variable as an additional independent variable. (iii) Since
many researchers use market- and risk-adjusted returns in estimating the
earnings response coefficient [e.g., Kormendi and Lipe (1987)], our analysis
facilitates comparison with their results.
Table 4 reports results of including R,,_,,, as an additional explanatory
variable, where R,, _ I. t is the CRSP equal-weighted return including dividends
from t - 5 to t, obtained by summing monthly returns over the period from
t - t to t. The results in table 4 are similar to those reported in table 2. The
average earnings response coefficient increases from 1.96 when no leading-
period return is included to 4.49 when the three-year leading return is included.
The average and median earnings response coefficients are smaller than those in
table 2 at all lengths of leading-period returns. The standard errors of the
average earnings response coefficients are very similar to the corresponding
standard errors reported in table 2.
The average coefficient on the market return, yZr, is the estimated relative risk
(beta) of the average stock in the sample. At first glance, it is puzzling to find the
average security in our sample is considerably less risky than the market,
especially when leading-period returns are included. For example, the average
jZr is 0.92 for the annual contemporaneous regression and it is only 0.66 when
three-year leading returns are included. It is likely that glr and fZr are affected by
14An alternative approach would be to regress market and risk adjusted returns on correspond-
ingly adjusted earnings. Our approach is econometrically preferable to the alternative two-stage
approach which is likely to yield biased estimates [see Beaver (1987)]. Moreover, in the alternative
two-stage approach the test period would be shorter than that examined here. The reason is that the
two-stage approach entails estimating the market model parameters separately for the security-
return-generating process and the earnings-generating process. This estimation will use up data over
the initial few years from 1950 which then cannot be included in the test period.
162 S.P. Kothari and R.G. Sloan, Information in prices about,future earnings
Table 4
Selected fractiles from the distribution of estimated parameters of firm-specific regressions of price
relatives on earnings deflated by price and market return including dividends. Earnings measure-
ment interval is one year and price relatives are for the contemporaneous one year and for periods r
that include leading years. Annual earnings data from 1950-88. Sample sizes from 2,319 to 2,721
“Sample selection: Any firm that has earnings data on the Compustat Annual Industrial or
Research tape for five consecutive years during 1950-88 and return data for the earnings measure-
ment years and the three leading years is included. The return observations contain overlapping
periods when one or more years of leading-period returns are included.
bEarnings yield X,/P,_, = primary annual earnings per share before extraordinary items and
discontinued operations for year t. divided by the price at the beginning of the return measurement
interval. All earnings per share numbers and prices are adjusted for stock splits and stock dividends.
Price relative P,/P,_, = one plus the buy-and-hold return, inclusive of dividends, over the years
t - T to t. R,,_,,, = CRSP equal-weighted return including dividends over years t - T to t.
‘The model is estimated by modeling the residuals from the ordinary least squares regressions as
a first-order autoregressive process and then ordinary least squares parameters are reestimated
using transformed variables. The estimated slope coefficients are deflated by (1.039Y’ to control for
bias stemming from using dividend-reinvested returns.
S.P. Korhari and R.G. Sloan, Information in prices about future earnings 163
a positive correlation between the market return and the earnings variable,
particularly when leading-period return is included. To verify that the sample
firms’ average beta risk is not unusually small, we reestimate firm-specific
regressions without the earnings variable included as an independent variable.
That is, we estimate the market model for each firm using annual returns. The
average market model beta of the sample firms then is 1.01.
This section assesses whether our inferences are sensitive to alternative es-
timation procedures and statistical tests. We estimate ordinary least squares
coefficients without making an adjustment for autocorrelated errors. Since the
ordinary least squares estimates are unbiased, the average coefficients should be
similar to those reported earlier. The average coefficients when one to three
leading years’ returns are included are 4.67, 5.07, and 5.67.
Earlier in this section, we tested whether the average estimated coefficient
equals its predicted value using a t-test. An alternative test is based on the
sample distribution of the firm-specific estimated t-statistics. For each firm
a t-statistic is calculated to test the null hypothesis that the firm’s earnings
response coefficient equals the predicted value of 6.83. By aggregating these
t-statistics, we then calculate a Z-statistic to test the null hypothesis that the
average estimated coefficient equals its predicted value of 6.83 [see, e.g., Heady,
Kang, and Palepu (1987) for a description of this test]. This test weights each
coefficient by the precision with which it is estimated, as compared to an equal
weight in the t-test of the sample mean coefficient. The estimated Z-statistic also
rejects the hypothesis that the estimated average coefficient equals its predicted
value when none to three leading-year returns are included.
Appendix
and
E(Y,) = (1 + l/r),
where E($,) = (1 + l/r) in the absence of any bias. The estimated coefficient
from (A.2) is
4
, 1 = coWIPt-,, P,lP,-,)lvar(X,lP,-,)
= covc(x,~p,-1)(p,-llp,-,)~(p~lp,-l)(p,-llp,-,)1
(A.4)
varC(X,/P,- 1 HP,- I P-,)1
S.P. Kothari and R.G. Sloan. Irzformarion in prices about future earnings 167
- E(X,l~,-,)E(P,l~,-,)CE(P,,/P,_,)12.
UX,IP,- 1 I= r,
W/P,- I) = (1 + r),
The ratio of (A.5) to (A.6) yields the earnings response coefficient estimate whose
expectation is
In the more realistic case of firms paying dividends, earnings through time
represent earnings on the initial equity and earnings on the increase in equity
through earnings retention (i.e., less than 100% dividend payout). The return
variable, on the other hand, assumes dividends are reinvested and thus is
unaffected by dividend payments (except for a small effect of the assumption
with respect to the rate of return on reinvested dividends). Under these circum-
stances, the estimated coefficient is
(A.81
where dividend yield is the realized dividend yield over one period, which is
assumed constant through time. Eq. (A.12) reveals that the earnings response
coefficient estimate will be biased upwards and the bias is approximately equal
to the compounded dividend yield over the leading-return period. We adjust the
estimated earnings response coefficients for the bias stemming from using
returns inclusive of dividends.
S.P. Korhari and R.G. Sloan, Information in prices about &we earnings 169
References
Albrecht, William, L. Lookabill, and James McKeown, 1977, The time series properties of annual
earnings, Journal of Accounting Research, 226244.
Ali, Ashiq and Paul Zarowin, 1992, Permanent versus transitory components of annual earnings and
estimation error in earnings response coefficients, Journal of Accounting and Economics, this
issue.
Anthony, Joseph H. and Krishnamoorthy Ramesh, 1992, Association between accounting perform-
ance measures and stock price: A test of the life cycle hypothesis, Journal of Accounting and
Economics, this issue.
Ball, Raymond, 1978, Anomalies in reltionships between securities’ yields and yield surrogates,
Journal of Financial Economics 6, 103-126.
Ball, Raymond and Philip Brown, 1968, An empirical evaluation of accounting income numbers,
Journal of Accounting Research, 159-178.
Ball, Raymond and Ross L. Watts, 1972, Some time series properties of accounting income, Journal
of Finance 27, 6633682.
Banz, Rolf, 1981, The relationship between return and market value of common stock, Journal of
Financial Economics 9, 3- 18.
Beaver, William H., 1987, The properties of sequential regressions with multiple explanatory
variables, The Accounting Review, 1377144.
Beaver, William H., 1989, Financial accounting: An accounting revolution (Prentice Hall, Engle-
wood Cliffs, NJ).
Beaver, William H., Roger Clarke, and William Wright, 1979, The association between unsystematic
security returns and the magnitude of earnings forecast errors, Journal of Accounting Research
17, 316-340.
Beaver. William H., Richard Lambert, and Dale Morse, 1980, The information content of security
prices, Journal of Accounting and Economics 2, 3-28.
Beaver, William H., Richard Lambert, and Stephen Ryan, 1987, The information content of security
prices: A second look, Journal of Accounting and Economics 9, 139-157.
Bernard, Victor L., 1987, Cross-sectional dependence and problems in inference in market-based
accounting research, Journal of Accounting Research 25, l-48.
Bernard, Victor L. and Jacob Thomas, 1989, Post-earnings-announcement drift: Delayed price
response or risk premium, Supplement of Journal of Accounting Research, l-36.
Biddle, Gary and Gim Seow, 1990, The estimation and determinants of association between returns
and earnings: Evidence from cross-industry comparisons, Working paper (University of Wash-
ington, Seattle. WA).
Blackwell, David, M. Wayne Marr, and Michael F. Spivey, Plant-closing decisions and the market
value of the firm, Journal of Financial Economics 26, 277-288.
Brealey, Richard and Stewart Myers. 1988, Principles of corporate finance (McGraw-Hill, New
York, NY).
Brickley, James and Leonard VanDrunen, 1990, Internal corporate restructuring: An empirical
analysis, Journal of Accounting and Economics 12, 251-280.
Brooks, L. and D. Buckmaster, 1980, First difference signals and accounting income time series
properties, Journal of Business, Finance and Accounting, 437-454.
Brown, Larry, Paul Griffin, Robert Hagerman, and Mark Zmijewski, 1987, An evaluation of
alternative proxies for the market’s assessment of unexpected earnings, Journal of Accounting
and Economics 9, 153-193.
Brown, Philip, George Foster, and Eric Noreen, 1985, Security analyst multi-year earnings forecasts
and the capital market, Studies in accounting research no. 21 (American Accounting Association,
Sarasota, FL).
Chan. Su Han, John Martin, and John Kensinger, 1990, Corporate research and development
expenditures and share value, Journal of Financial Economics 26, 255-276.
Cochrane, D. and G. Orcutt. 1949, Application of least squares regressions to relationships
containing autocorrelated error terms, Journal of the American Statistical Association, 32-61.
Collins, Daniel and S.P. Kothari, 1989, An analysis of the intertemporal and cross-sectional
determinants of the earnings response coefficients, Journal of Accounting and Economics 11,
143-181.
Collins. Dante]. S.P. Kothari. and Judy Rayburn, 1987. Firm size and the information content of
prtces with respect to earnings. Journal of Accountrng and Economics 9, 1 I I-138.
Easton. Peter and Tremor Harris. 1991. Earnings as an explanatory vartable for returns, Journal of
Accountmg Research 29, 19-36.
Easton. Peter and Mark Zmijewski. 1989a. Cross-sectional variation in the stock market response to
the announcement of accounting earnings, Journal of Accounting and Economics 11. 117~141.
Eaaton. Peter and Mark Zmijewskt, I989b. On the estimation of earnings response coefficients,
Working paper (University of Chicago. Chicago. IL).
Easton. Peter, Trevor Harris. and James Ohlson. 1992. Aggregate accounting earnings can explain
most of security returns: The case of long return mtervals, Journal of Accounting and Economics,
thts Issue.
Fama. Eugene and Kenneth French, 198X. Permanent and temporary components of stock prices.
Journal of Political Economy 96, 246~ 273.
Foster, George. Chris Olsen. and Terry Shcvlin. 1984. Earnmgs releases, anomalies and the behavior
of security returns. The Accounting Review. 574-603.
Freeman. Robert, 1987. The association between accounting earnings and security returns for large
and small firms, Journal of Accounting and Economics9, 1951228.
Handa. Puneet. S.P. Kothari. and Charles Waslev. 19X9. The relation between the return interval
and betas: Implications for the size effect, Journal of Financial Economics 23. 799100.
Healy. Paul. 1985. The effect of bonus schemes on accounting decisions. Journal of Accounting and
Economics 7. 85-107.
Healy. Paul, Sok-Hyon Kang, and Krishna Palepu, 1987, The effect of accounting procedure
changes on CEO’s cash salary and bonus compensation, Journal of Accounting and Economics
9. 7734.
Ibbotson. Roger and Rex Sinqueheld. 1989, Stocks, bonds, bills and inflation: Historical returns
(1926-1987) (The Research Foundation of the Institute of Chartered Financial Analysts, Char-
lottesville. VA).
Kendall. C.S. and Paul Zarowin. 1990. Time series properties of earnings. earnings persistence and
earnings response coeffictents. Working paper (New York University, New York, NY).
Kormendi. Roger and Robert Lipe, 19X7. Earnings innovations, earnings persistence and stock
returns, Journal of Business 60, 3233345.
Kormendi. Roger and Robert Lipe. 1991. The tmplicattons of the higher-order properties of annual
earnings for security valuation. Working paper (University of Michigan. Ann Arbor, MI).
Landsman. Wayne and Joseph Magliolo, 1988. Cross-sectional capital market research and model
specification. The Accounting Review 64. 586-604.
Lang. Mark. 1991. Time-varying stock-prtce responses to earnings induced by uncertainty about the
time-series properties of earnings, Working paper (Graduate School of Business, Stanford
University. Stanford, CA).
Lev. Baruch. 1989. On the usefulness of earnings: Lessons and directions from two decades of
emptrical research, Supplement of Journal of Accounting Research, 1533192.
Lev. Baruch and S. Ramu Thiagarajan. 1991, Financial statement information. Working paper
(University of California. Berkeley. CA).
Lipe. Robert, 1990, The relatton between stock returns and accounting earnings given alternative
information. The Accounting Review 65. 49~-71.
Maddala. G.S.. 1977. Econometrics (McGraw-Hill. New York, NY).
McConnell. John and Chris Muscarella. 1985. Corporate capital expenditure decisions and the
market value of the firm. Journal of Financial Economics 14, 3999422.
Ohlson. James A., 1991, The theory and value of earnings and an introduction to the Ball-Brown
analysis. Contemporary Accounting Research, I-19.
Ou, Jane and Stephen H. Penman, 1989. Financial statement analysis and the prediction of stock
returns. Journal of Accounting and Economics 11, 295-329.
Penman. Stephen H.. 1990. Financial statement information and the pricing of earnings, Working
paper (University of California, Berkeley. CA).
Poterba. James and Larry Summers, 19X8, Mean reversion in stock prices: Evidence and implica-
ttons. Journal of Fmanctal Economics 22, 27-59.
Ramesh, Krishnamurthi, 1990. The existence of a unit root in earnings as a determinant of the
price-earnings relation. Working paper (Michigan State University, East Lansing, MI).
S.P. Korhari and R.G. Slam, It$rmarion in prices about future earnings 171
Rao, Gita R., 1991. The relation between stock returns and earnings: A study of newly-public firms,
Working paper (University of Illinois, Urbana-Champaign, IL).
Ross, Stephen A.. Rudolf W. Westerfield, and Jeffrey F. Jaffe, 1990, Corporate finance (Richard D.
Irwin, Homewood, IL).
Sloan, Richard, 1992. Accounting earnings and top executive compensation, Journal of Accounting
and Economics, forthcoming.
Watts, Ross L. and Richard Leftwich, 1977, The time series of annual accounting earnings, Journal
of Accounting and Research 15, 253-27 1.
Watts, Ross L. and Jerold L. Zimmerman, 1986, Positive accounting theory (Prentice-Hall, Engle-
wood Cliffs, NJ).