Вы находитесь на странице: 1из 26

International

Journal
016Y-207O/Y3/$06.00

9 ( 1993) 295-320
of Forecasting
@ 1093 Elsevicr Science Publishers

20s
B.V. All rights

reserved

Earnings forecasting research:


for capital markets research*
Lawrence

its implications

D. Brown

State University

of New

York

at Buffalo,

Buffalo,

NY

14260-4000,

USA

Abstract
Since the early 198Os, earnings forecasting
research has become much more closely aligned with
capital markets research.
Capital markets research requires a proxy for the (unobservable)
market
earnings
expectation
and earnings forecasting
research has provided such proxy measures.
Questions
considered
in this paper include: (1) if annual earnings follow a random walk or an IMA (1 ,I) model,
does this mean that earnings
changes cannot be predicted?
(2) Do stock prices act as if quarterly
earnings follow a seasonal random walk with drift process? (3) Is the predictive mode1 which is best on
the forecast accuracy dimension
also best on the market association
dimension?
(4) How do analysts
formulate
their earnings expectations?
(5) What is the role of earnings forecasting in earnings response
coefficient
and post-earnings
announcement
drift studies? (6) What is the likely role of earnings
forecasting
research in future capital market studies?
Keywords: Earnings
forecasting;
drift; Future research

Forecast

accuracy;

1. Introduction
When Paul Griffin and I issued a call for
papers for the Special Issue on Accounting
and
Finance
of the Journal of Forecasting
[Brown
and Griffin (1983)], most of the papers we received were concerned
about whether
annual
earnings follow a random walk (or random walk
with drift) process. Nevertheless,
this issue was
pretty much resolved
by the late 1970s. Little
(1962) provided evidence that British firms annual earnings follow a random walk process and
several
studies
of US firms [Ball and Watts
*Earlier versions of this paper were presented
at Columbia
University.
SUNY at Buffalo and the University of Waterloo.
I am grateful for the valuable comments
of Paul AndrC. Eli
Bartov, Sasson Bar-Yosef,
M. Daniel Ben&h,
Dan Coggin,
Robert Fildes, Bob Hagerman.
Jerry C. Y. Han, Tom Lechncr, John OHanlon,
Wonsun
Paek, Gordon
Richardson,
Kevin Sachs, Jake Thomas and Mark Zmijewski.

Market

association;

Post-earnings

announcement

(1972), Albrecht
et al. (1977), Watts and Leftwith (1977)] showed that Box-Jenkins
(1976)
autoregressive
integrated
moving
average
(ARIMA)
models confined
to annual
earnings
data did not generate more accurate forecasts. in
a holdout period, than those based on a random
walk model. While the papers published
in the
special issue did not focus on whether
annual
earnings follow a random walk (or random walk
with drift) process, most did focus on earnings
forecasting
and lacked implications
for capital
markets research.
A major change in the earnings forecasting
literature
took place in the
1980s; it became much more closely linked with
capital markets research.
Prior to Foster (1977), the earnings forecasting
literature
focused
on predictive
ability
and
evaluated
earnings expectation
models solely on
their ability to forecast future earnings
(hereafter, accuracy). Foster introduced
an alternative

way to evaluate
earnings
expectation
models,
namely contemporaneous
market association
of
abnormal
returns with earnings
surprise,
conditional upon the expectation
model. (Hereafter,
we shall refer to Fosters method as the association approach.)
He argued that. given the maintained
hypothesis
of an efficient
market,
the
strength of association
between abnormal returns
and earnings
surprise
indicates
how accurately
the model captures
the markets
expectation.
Since the capital markets
literature
requires
a
proxy for the markets earnings expectation.
the
association
approach made the earnings forecasting literature
important
for capital markets research. especially
when abnormal
returns were
measured
over a narrow window,
such as the
-day
period,
(-I .O), where
Day 0 is the
announcement
day [Foster ( I977)].
In an informationally
efficient market. ignoring transactions
costs and costs of generating
and
processing
information,
the association
and predictive ability approaches
should provide similar
results. The argument
that predictive ability and
association
arc two sides of the same coin is
based on the assumption
that the capital market
uses the best data available.
where best is defined as most accurate. This notion is similar to
the rational
expectations
principle
which maintains that the subjective
probability
distribution
of outcomes
is distributed
as the objective probability distribution
of outcomes
[Muth (1961 )].
In the early l%Os, researchers
began to use
the dual criteria of accuracy and association
to
evaluate
earnings forecasting
models. Fried and
Givoly (19X2) and Brown et al. (1%7a,b),
respectively.
used these same dual criteria to compare analysts
annual
and quarterly
earnings
forecasts
with those of univariate
time-series
models.
These
studies
showed
that analysts
earnings forecasts are better than those of timcseries models when both criteria are used and
suggested that analysts earnings forecasts should
be used in lieu of time-series
model forecasts,
when a proxy for the markets earnings cxpectations is required.
When analyst
data became
widely available
in machine-readable
form. academics began to use analysts earnings forecasts
in lieu of time-series
model forecasts to proxy for
the markets unobservable
earnings expectation.
The remainder
of this article proceeds as follows. Section 2 discusses a branch of the earnings

forecasting
literature
which continued
to evolve
without direct links to capital markets research.
It examines
the time-series
properties
of annual
and quarterly
earnings;
what happens to predictive
accuracy
when
additional
information,
beyond
earnings.
is allowed for in the information
set; the analysts
earnings
expectation
process; how analysts earnings forecasts can be
improved;
biases in analysts earnings forecasts;
determinants
of analyst following.
Section 3 examines
the relation
between
accuracy and association;
past quarterly
earnings
numbers
and
post-earnings
announcement
drift: analyst
bchavior
and post-earnings
announcement
drift;
forecast
convergence
an d
post-earnings
announcement
drift; the relation
of analysts
forecasts to abnormal
returns.
The earnings response coefficient literature,
discussed in Section
4, examines
proxies for earnings surprise; determinants
coefficients
of
earnings
response
(ERCs); the functional
form of the ERC cstimation equation;
stock-price
and earnings cxpectations.
Sections 2. 3 and 4 summarize
and critique the
and offer some guidance
for future
literature,
research.
Section 5 focuses on some directions
for future research.
Included therein is the role
of analysts earnings forecasts as inputs for stockprice recommendations:
the role of analyst expcctational
data in post-earnings
announcement
drift studies;
the role of cxpectational
data if
forecast accuracy and association
are (or are not)
two sides of the same coin.
Section 6 provides implications
for other branches of forecasting
research. Special attention
is
paid to two notions: ( 1) the forecasting
literature
has recently
developed
a number
of themes
which have been addressed
by the literature
surveyed
in this paper; (2) the intellectual
focus
of forecasting
research can be sharpened,
and its
impact enhanced.
if it broadens
its perspective
beyond
examining
forecasts
as ends in themselves.

2. A continuation

of the earnings

forecasting

literature
-7. 1. Tirnr-series

One

branch

properties

oJ earnings

of earnings

forecasting

research

L. D. Brown

I Earnings

focused on the time-series


properties of earnings
numbers
and lacked direct links to capital markets research.
Nevertheless,
the models
developed
by this literature
have been used as
measures of earnings expectations
in capital markets studies. Foster (1977), Griffin (1977), Watts
(1975) and Brown and Rozeff (197Ya) advocated
different Box-Jenkins
(1976) ARIMA models of
the quarterly
earnings
time-series
process. The
Box-Jenkins
model uses (pdq) X (PDQ)
notation,
the Foster
model
is the (100) x (010)
model,
plus a constant.
It is a first-order
autoregressive
model, determined
after seasonally
differencing
the time-series
of quarterly earnings
numbers.
The
Griffin-Watts
model
is the
(011) x (011) ARIMA
model. Both its ordinary
and seasonal
components
are moving average
processes
after
taking
first differences.
The
Brown-Rozeff
model,
expressed
as (100) X
(011) in ARIMA notation,
employs the ordinary
component
of the Foster model (the first term in
parentheses),
and the seasonal component
of the
Griffin model (the second term in parentheses).
Box and Jenkins
(1976) represent
the seasonal
time-series
process via the multiplicative
model:
f&(B)@@.+,

= B(B)O(B)cr,

+ e,,

where w, is a stationary
time-series;
B is the
backshift operator
such that Bh~lI = w,_~, 4(B)
and B(B) are polynomials
in B, namely 4(B) =
1 - &B- . . . -&B
and
0(B) = 1 - 0,B. -oqyBq, @(B)
and O(B),
the seasonal
parts of the model, have the form @(B) = 1 @,5Bs- . . -@,,,sBP
and
O(B) = 1 - O,sBs. . . - 0 ys .B, a, is an independent
and identitally distributed
white noise residual term with
mean zero and variance a,, l, and 0,, is a constant.
The model is of order (p,d,q)
x (P,D,Q).
Autoregressive
parameters
are denoted by p and P;
moving average parameters
by q and Q; d and D
are the degrees of ordinary
and seasonal
differencing
required to achieve stationarity.
Griffin (1977) and Watts (1975) proposed the
model:
(l-B)(l-B)X,=(l-O,B)(l-O,B)a,

per share.
the model:

207

research

(1 - $,B)(l

- B)X,

= (1 - &B4)q

(2)

The Foster (1977) model is a special case of the


Brown-Rozeff
(1979a) model, with 0, equal to
zero.
In contrast with the Griffin-Watts
and BrownRozeff models, which require the use of ARIMA
software to estimate model parameters,
the Foster models parameters
are estimable
using ordinary least squares (OLS). It can be estimated
as: X, - X,_, = L+ 4,(X,_, - X,_,) + Y, where
X, is quarterly
earnings
in period t, c is an
intercept
(constant term), 4, is a slope and cy, is
an error term. The ease with which these estimates can be made resulted in its being the most
often used of the three ARIMA
models.
These three ARIMA
models formed the core
of the quarterly
earnings
time-series
model literature.
Many studies
have examined
which
model
generates
the most accurate
quarterly
earnings forecast. Benston and Watts (1978) provided evidence
in favor of the Foster model,
Lorek (1979) argued on behalf of the GriffinWatts model, Collins and Hopwood
(1980) and
Bathke and Lorck (1984) provided evidence in
favor of the Brown-Rozeff
model. For over a
decade, no other quarterly ARIMA
model to be
introduced
was found to be more accurate, on
average. than these three models.
Lee and Chen (1990) showed that these three
ARIMA
models can be improved by incorporating temporary,
short-run
and long-run structural
changes. They argued that structural changes are
commonplace
and that statistical
models incorporating
them provide more accurate
earnings
forecasts
than can be generated
via the three
ARIMA
models.
However,
they did not find
that their procedure
was able to generate
more
accurate forecasts than those by Value Line analysts and they provided
no evidence
along the
association
dimension.
Thus, one cannot infer
that the Lee and Chen procedure generates better earnings
forecasts than can be provided by
analysts.
Although
it was commonly
believed in the late
1970s that annual earnings follow a random walk
Collins

(1)
where X, denotes quarterly
earnings
Brown and Rozeff (1979a) suggested

forrmsiing

et al. (1984) found a strong pattern


of higher
forecast errors in the fourth interim period for time-series
models and financial analysts forecasts,
suggesting
that a
different
forecasting
model may bc needed for the fourth
quarter than is used for the first three quarters.

(or random
walk with drift) model, some researchers
argued that annual earnings
with extreme year to year changes were better described
by a mean-reverting
model [Brooks and Buckmaster
(1976).
Salamon
and Smith (1977).
Beaver and Morse (1978)]. The mean-reverting
model, known as the ARIMA
(0.1 ,l) or integrated moving average (IMA) (1,l) model, has
been used in numerous
subsequent
studies by
accounting
researchers
[Beaver
et al. (lY80.
1987), Collins and Kothari (1989)].
Kendall and Zarowin
(1990), and Ramakrishnan and Thomas (1992) have shown that annual
earnings
are well-described
by a first-order
autoregressive
process in earnings
levels. In contrast with Brooks and Buckmaster
(1976). who
reject the random walk hypothesis
only for extreme earnings
changes,
these studies suggest
that the rejection
of the random walk model is
not confined
to cases of extreme
earnings
changes.
Ramakrishnan
and Thomas
(1992)
show that this behavior
has become more pronounced over time; the mean autocorrelations
at
the first lag in the excess earnings
time-series
(i.e. X, - X,_, - (R - 1)(X,_, - d,_ ,). where X
and d are net income and dividends,
respectively, and R. the required
return on equity, = 1.1)
are - 0.03
and - 0.16
for
the
two
20-year
periods,
19Sl-1970
and 1969-1988, respectively.
Nevertheless,
neither study provided
predictive
evidence
in a holdout
sample,3 and there are
Consider

the IMA (1 ,l) model:

X,( I - B) = tu,(l -

H,B) .

or

where X, is annual earnings


in year 1. a, is the shock
(random
error) in period ts earnings and N, is the moving
average coefficient.

v(a,,a,)

= 0, for all

f # s.

If 0, = 0. the stochastic
process is a random walk. Alternatively.
if H, = 1. the stochastic
process is strictly meanreverting.
Penman (lYY2a) shows that the extent of mean
reversion
can hc inferred by using price to earnings ratios
and price to book value ratios.
Each study estimated the autoregressive
parameters
in two
2%year periods,
hut neither examined
whether
the firmspecific parameter
estimate in the first period yielded more
accurate
forecasts than the random walk predictions
in the
second period.

numerous
ways to generate
a more accurate
annual earnings
forecast than the random walk
IMA (1 ,l) or (autoregressive)
AR (1) models,
by expanding
the information
set beyond
the
time-series
of past annual
earnings
numbers.
This literature
is discussed below.
2.2. Expunding
the information
set upon which
unnuul earnings expectations
are conditioned
Hopwood
et al. ( 1982) showed that the three
ARIMA
quarterly
earnings
time-series
models
can be used to improve the predictive
accuracy
of the annual earnings number.
More specitically, prior to observing the first interim (quarterly)
earnings
report of the fiscal year, one can use
these ARIMA
models to forecast the individual
quarterly
earnings
numbers
which, when summcd, equal an annual earnings forecast which is
1521%
more accurate
than is obtainable
by
extrapolating
from the firms past annual earnings numbers
alone. Hopwood
et al. define the
relative predictive
accuracy of the annual versus
quarterly
models as:

$, (F,,

- Actual, ),/$,

(Fo, - Actual,)

where F,+ is the forecast from the annual model


for year j, and Fo, is the forecast from the sum of
quarterly
forecasts for year j ( j = 1 - 5), based
on one of the three ARIMA
models.
Beaver
et al. (1980,
1987), Collins
et al.
(1987) and Freeman
(1987) show that the random walk model can be beaten by expanding
the information
set to include the firms stock
price.
Shroff (1992) shows that a composite
model forecast, utilizing random walk and pricebased model forecasts.
is more accurate than a
random walk model forecast when earnings variability is large relative to stock-price
variability.
These studies suggest that the firms stock price
is based partly on expectations
of future earnings. rather than simply on known, current and
past.
Consistent
with this finding,
earnings.
Brown et al. (1985) show that the firms stock
price is based on analysts expectations
of future
earnings
and that analysts
long-term
earnings
expectations
are more valuation-relevant
than
their short-term
earnings expectations.
Freeman et al. (1982) showed that the random

walk with drift model can be beaten by expanding the information


set beyond the firms past
annual
earnings
to include its past stock-price
and book rate of return. Based partly on these
findings, Ou and Penman (1989a,b),
Ou (1990),
Lev and Thiagarajan
(1991), Lev and Sougiannis
(1992) and Penman
(1992b) showed that nonearnings,
financial statement
data can be used to
predict future earnings.
Some of these studies
also showed that abnormal
returns can be obtained by using publicly available financial statement information,
giving rise to a financial statement information
anomaly.
Ball (1992) suggests
that the financial statement
information
anomaly
is due to accounting
ratios proxying for stocks
expected returns.
As evidence for this view, he
cites Stober (1992), who showed that the Ou and
Penman (1989a,b) algorithm can be used to generate abnormal
returns that continue
at an almost constant
rate for 6 years beyond the earnings prediction
date. Ball argues that if the abnormal returns persist for 6 years, they are likely
to be related
to risk factors,
not to financial
statement
data. Greig (1992) and Holthausen
and Larcker
(1992) also challenge
the Ou and
Penman results.
In summary.
time-series
models confined
to
past earnings generally can be beaten by expanding the conditioning
information
set to include
past quarterly
earnings,
stock prices, book rates
of return,
or other financial
statement
data.
Thus, while the random
walk with drift, IMA
(1,l) and AR(l)
models are reasonable
characterizations
of the annual
earnings
time-series
process, they can readily be improved
on once
additional
information
beyond past annual earnings is allowed
for in the conditioning
information set. More work is needed to determine
whether
a financial
statement
information
anomaly exists and if it does, whether it is more
pronounced
for certain firms and time periods
than for others.
2.3.

The analysts earnings expectation process

Research
has provided
insights into the process by which analysts formulate
their earnings
expectations.
Some studies have examined
why
analysts earnings
forecasts
are relatively
more
accurate than those made by time-series
models.
Brown and Zmijewski
(1987) show that the anal-

ysts earnings forecasting


advantage
is enhanced
at a time of strikes. Their finding is consistent
with the hypothesis
that analysts are relatively
better than time-series
models at distinguishing
among
permanent,
transitory
and price-irrele[Ramakrishnan
and
vant
earnings
shocks
Thomas (1991)]. H owever, they stack the decks
in favor of analysts by using as benchmarks
those
models
which
do not incorporate
structural
changes
into the quarterly
earnings
generating
process. While the Brown and Zmijewski
results
are intuitively
appealing,
their study should be
replicated
with a stronger
time-series
model
benchmark
and a more sophisticated
methodology before one can reliably conclude that analysts
are better than time-series
models at distinguishing between earnings shocks with different valuation implications.
Fried and Givoly (1982) and Brown et al.
(1987a) investigate
whether an analysts forecasting advantage
relative to that of a time-series
model is attributable
to two information
factors:
(1) analysts
have a contemporaneous
information
advantage
they utilize information
available
on the earnings
announcement
date
other than the sequence
of known quarterly
earnings
numbers;
(2) analysts
have a timing
advantage
they utilize information
acquired
after the time-series
model forecast
initiation
date (i.e. the earnings announcement
date), such
as management
forecasts,
macroeconomic
data
and industry variables.
Fried and Givoly (1982)
use annual earnings data from the, now defunct,
Standard
and Poors Earnings Forecaster, and
find that only the first (contemporaneous)
information
factor
is significant.
Brown
et al.
(1987a) use Value Line quarterly
forecasts,
and
find both the timing and contemporaneous
information
factors to be significant.
Nevertheless,
neither study provides insights into the relative
importance
of alternative
information
sources for
explaining
the analysts
forecasting
advantage.
Brown et al. (1987~) use both Institutional
Brokers Estimation
System (I/B/E/S)
consensus
(mean) forecasts and Value Line forecasts,
and
argue that the analysts
forecasting
advantage
relative to time-series
models is attributable
to
three characteristics
of the information
set conditioning
analysts expectations:
(1) the amount
of information
available;
(2) the precision of this
information;
(3) the correlation
among the in-

formation
variables.
Using firm size, the homogeneity of analysts earnings expectations
and the
number
of lines of business
(LOB),
as proxies
for the
amount,
precision,
and correlation
among the information
variables,
respectively,
Brown
et al. (1987~) show that the analysts
relative advantage
in forecasting
annual earnings
is positively
related to firm size and the homogeneity of analysts expectations.
Finding no relation
between
the analysts
relative
earnings
forecasting
advantage
and the number of LOB,
they suggest that the number
of LOB is too
crude a proxy for the correlation
among the
information
variables to allow for definitive findings. Their reasoning is as follows: Suppose that
(analysts
possess) tz information
signals (which)
relate to m lines of business. If m is small, the n
information
will
be
highly
corsignals
related.
. if wz is large and each of the n information
signals is randomly
assigned to one of
the m lines of business, the II information
signals
will tend to represent
more diverse types of
information.
This argument
is not very persuasive, as there is no reason to believe that each of
the y1 information
signals is randomly assigned to
one of the m lines of business.
Employing
Value Line quarterly
forecasts,
Kross et al. (1990) also use the number of LOB
to proxy for the correlation
variable and find it
to lack explanatory
power. One explanation
for
the poor LOB results is that the variable
is
significantly
correlated
with firm size and the
homogeneity
of analysts expectations
[Brown et
al. (1987c), p. 601. Another
explanation
for the
poor LOB results is that the correlation
among
the information
variables
is not an important
determinant
of analyst
forecast
superiority.
Nevertheless,
Branson
and Pagach (1993) show
that the Value Line analysts predictive
advantage vis-a-vis the Brown-Rozeff
(1Y79a) model.
is significantly
positively related to LOB for twoand three- (but not one) quarters-ahead
forecast
horizons.
Additional
research is needed to determine whether the Brown et al. (1987~) theory
is poorly specified, the number of LOB is a poor
proxy for the correlation
variable,
or both.
Kross et al. (1990) show that the analysts
advantage
in forecasting
annual
earnings
increases with the firms earnings
variability
and
the amount of coverage by the Wull Street Journal. This provides
additional
evidence that anal-

ysts possess an informational


advantage
vis-a-vis
time-series
Kim
models.
Similarly,
and
Schroeder
(1990) suggest that analysts utilize a
wide variety of non-earnings
information
sources
when forecasting
annual
earnings.
They show
that analysts anticipate
discretionary
accruals in
earnings
reports
of firms with earnings-based
bonus plans.
Collins et al. (1987) and Freeman
(1987) extend the work of Beaver et al. (1980, 1987) and
Beaver ct al. (1987) by examining
the relation
between
firm size and the predictive
advantage
of price-based
forecasts versus predictions
of a
naive earnings expectation
model. Arguing that
stock price is a more efficient
aggregator
of
information
for larger firms, Collins et al. and
Freeman
show that price-based
forecasts
for
larger firms have a predictive advantage
vis-a-vis
smaller
firms. Larger firms have richer information environments
than smaller firms, owing
to the information
acquisition
and dissemination
activities
of analysts and institutional
investors
[Atiase
( 198.5),
Bambcr
(1987),
Bhushan
(1989)]. Thus, it is relatively
more difficult for
larger firms managers
to withhold their private
information
[Brown and Kim (1993)], and larger
firms stock prices incorporate
information
relatively more rapidly.
Collins et al. (1987) and Freeman (1987) interpret their findings
as being consistent
with a
firm-size
differential
information
environment
hypothesis.
However, the random walk model is
a weak benchmark
for forecasting
annual earnings (e.g. it can be improved on simply by using
the firms past quarterly earnings numbers in lieu
of its past annual earnings numbers [Hopwood et
al. (1982)]). so these results should not be considered as definitive evidence of a firm-size based
explanation.
Before the evidence can be considered definitive,
it is necessary
to demonstrate
that the firm-size results are robust to stronger
time-series
benchmarks,
such as ones incorporating past quarterly
earnings numbers.
If the relative predictive advantage of using quarterly earnings in lieu of annual
earnings
is greater
for
smaller firms, the firm-size results may not be
robust to this stronger
time-series
benchmark.
Brown et al.s (1985) finding that stock price is
a leading indicator of analysts earnings forecast
revisions
suggests that the analysts information
advantage
may be attributable
to the incorpora-

L.

D. Brown

I Earnings forecmting

tion of lagged stock prices into their earnings


forecasts.
If so, the analysts comparative
advantage over univariate
time-series models may arise
from their utilization
of public rather than private information.
However, a price-based
explanation for the analysts comparative
advantage is
unlikely.
Stickel (1990) shows that the change in
the consensus
(mean) estimate and the analysts
deviation
from the consensus estimate are much
more important
determinants
of analysts earnings forecast revisions than is the firms lagged
stock price.
Thus,
the analysts
comparative
earnings forecasting
advantage
appears to result
at least partly
from the utilization
of firmspecific, private information,
not public information
reflected
in stock price. The evidence
that analysts possess private information
facilitating their earnings
predictions
is consistent
with
the finding that analysts possess private information enabling them to generate superior forecasts of firms future stock prices [Copeland and
Mayers (1982), Coggin and Hunter (1983). Dimson and Marsh
(1984),
Elton et al. (lY%),
Brown et al. (1991)]. However,
little is known
regarding
how much of the analysts predictive
advantage
is due to the utilization
of private
rather than public information.
In summary,
the analysts earnings forecasting
advantage
relative to time-series
models is partly
attributable
to their private information
acquisition activities,
which may enable them to better
distinguish
between
permanent,
transitory
and
price-irrelevant
earnings shocks. Their comparative advantage
is greater when the forecast is for
a larger firm, with more information
acquisition
activities,
greater earnings
variability
and more
homogenous
analyst
expectations.
Additional
work is needed to ascertain whether the correlation among
information
variables
is a viable
factor explaining
the analysts comparative
advantage.
Moreover,
the validity of the firm-size
explanation
and the role of public versus private
information
acquisition
in explaining
the analysts predictive
advantage
await future study.
2.4. Improving

analysts earnings forecasts

Several studies have examined


whether analysts earnings forecasts can be pooled with those
of time-series
models to generate relatively more
accurate earnings forecasts than can be made by

reseurch

301

either analysts or time-series


models individually. Newbold
et al. (1987) show that pooling
Value Line earnings forecasts with forecasts generated via the Brown and Rozeff (1979a) quarterly earnings time-series
model results in more
accurate
earnings
forecasts
than produced
by
either the Value Line analyst or the BrownRozeff model individually.
Conroy and Harris
(1987), Guerard
(1989), Lee and Chen (1990)
and Lobo ( 1991) use different
univariate
timeseries model forecasts, and show that predictive
ability can be enhanced
by pooling time-series
model forecasts with analysts forecasts. Elgers
and Murray (1992) show that I/B/E/S
consensus
forecasts,
when pooled with stock-price
based
forecasts
using the procedure
of Beaver et al.
(1980, 1987), generate
more accurate forecasts
than are available using either model individually. Stober (1992) shows that the summary financial statement
measure (Pr) derived by Ou and
Penman (1989a,b) works best when Pr indicates
buy and I/B /E/S consensus
analyst revisions
suggest sell.
Several studies interpret their own evidence as
suggesting
that analysts do not use all the information
in financial
statements
and stock
prices when formulating
their annual
earnings
expectations.
However,
one cannot
conclude
that this is the case since these studies use consensus, rather than individual,
analyst earnings
expectations.
As discussed below, some of the
forecasts included in the consensus
forecast are
not conditional
upon the financial statement
information
that these studies maintain
is ignored
by the analysts.
The contention
that analysts
disregard information
in financial statements
(or
stock prices) is warranted
only if the consensus
analyst forecast
is based on individual
analyst
forecasts made after the date that the financial
statement
(or stock-price)
data are available
[Brown and Han (1992)].
Lys and Sohn (1YYO) and Abarbanell
(1991)
use detail analyst forecasts rather than consensus
Conroy and Harris (1987) used I/B/E/S
consensus
forccasts, Guerard
( 1989) uses consensus forecasts of the Standard and Poors Earnings Forecaster. Lee and Chcn (1990)
use Value Line forecasts.
and Lobo (1991) uses Value Line
and I/B/E/S
consensus forecasts.
See Philbrick and Ricks
(1991) for a discussion of the advantages
and disadvantages
of these three sources of earnings
forecasts,
as well as
Zacks Investment
Research forecasts.

analyst forecasts and show that analysts earnings


forecasts do ignore information
reflected in stock
prices prior to the forecast-release
date. They
conclude that analysts earnings forecasts do not
incorporate
all the information
in prior stockprice changes,
suggesting
that analysts
are inefficient in collectingandior
interpreting
publicly
available
information.
A finding that analysts ignore publicly available information
is unsatisfactory
to capital markets researchers
who advocate semi-strong
form
market efficiency [Fama (1970)], but it is satisfactory to behavioral
researchers
who maintain
that people consistently
overweight
some cues
and underweight
others [Tvcrsky and Kahneman
(1984). Gilovich
et al. (l%S),
Mear and Firth
(1987),
Hunter
and Coggin
(1988).
DeBondt
(1993)]. Abarbanell
(1991) suggests that analysts
have incentives
to provide
new forecasts
only
when they obtain information
which is independent of prior stock-price
changes. Hc maintains
that this strategy makes it easier for investors to
infer the analysts private information:
. . private information
is more easily inferred by invcstors if it is not combined
with other signals
whose information
content is open to individual
interpretation
(p. 164). This view is controvcrsial as other studies have taken an opposite view.
For example, Kane et al. (1984) suggest that it is
easier to infer the valuation
implications
of dividend (earnings)
information
which is combined
with and corroborated
by earnings
(dividend)
information.
While the descriptive
validity
of
Abarbanells
scenario
versus the behavioralist
interpretation
awaits additional
research,
analysts do appear to ignore publicly available data
already incorporated
in past earnings and stock
prices.
Some studies have examined
ways to improve
analysts earnings forecasts without pooling them
with other information.
OBrien
(1988) shows
that the most recent I/B /E/S analysts earnings
forecast
is more accurate
than the consensus
(mean) forecast.
Although
this appears to contradict
the forecast
aggregation
principle
that
aggregation
improves predictive
accuracy by re

Lys and Sohn


search.
which
same

(1YYO) USC detailed

Inc. data. Abarbancll


represent
date.

It~~~e.rrmet~/

Zacks

Investment

date.

week

lb-

uses Value Line data.

one or two analyst forecasts

approximately
Sur,ry

(1091)

ducing idiosyncratic
error. studies finding evidence consistent
with the aggregation
principle
[Ashton
and Ashton
(19X5), Libby and Blashheld (1978).
Winkler
and Makridakis
(1983)]
examined
equally recent forecasts.
Analysts
earnings
forecasts
are not usually
equally
recent.
They
make
their
forecasts
throughout
the year, rather than at any set time,
such as on earnings
announcement
dates. As
some forecasts included in the consensus may be
more than 6 months old [OBrien (19X8)]. the
analyst ignores two quarterly
earnings
reports.
Since analysts
earnings
forecasts
can bc improved simply by substituting
in known quarterly
earnings
numbers
for their previous
expected
values, leaving intact forecasts for the remainder
of the year [Brown
and Rozeff
(1979c)].
it
should
be possible
to improve
analysts
consensus forecasts by deleting old estimates.
Brown
(1991) shows that earnings
forecast
accuracy
can be improved
by discarding
old
earnings
forecasts.
Using detail Zacks Investment Research
data, Brown finds that each of
three *timely composites
(i.e. the most recent
forecast,
an average of the three most recent
forecasts and the 30-day average) is more accurate than the mean forecast. Moreover,
the comparative advantage
of each timely composite depends on firm size; the most recent forecast is
approximately
as accurate as the 30-day average
for small firms, but the 30-day average is significantly
more accurate
than the most recent
forecast
for large firms. These results can be
interpreted
as follows. The aggregation
principle
pertains
to the 30-day average
of large firms
because these firms are followed by many analysts. In contrast, small firms are followed by few
analysts
and even fewer analysts
make their
earnings forecasts within 30 days of each other.
Browns
results
suggest
that the aggregation
principle pertains to earnings forecasts which arc
equally timely, but not to those made at different
points of time.
Several studies have examined
if there exists a
superior
analyst. Using a small sample of Value
Line
earnings
forecasts,
Brown
and Rozeff
(1980) conclude
that superior forecast ability is

before

made on the

the

Vuluc

Lirw

Brown does not

directly

following

the

firm.

positively

correlated

examine

However.

the number

of analysts

analyst hollowing

with firm size [Shores (IYYO)].

is highly

not evident.
Using a much larger sample of
I/B /E/S data and a more powerful methodology, OBrien (1990) reaches the same conclusion.
However,
Stickel (1992) uses a large sample of
Zacks Investment
Research
data and concludes
that members
of the Institutional Investor All
American
Research Team are superior earnings
forecasters.
Unlike
OBrien
(1990), he uses a
matched
pair design which fully controls
for
forecast
recency.
Using I/B/E/S
data and a
more powerful control for forecast recency than
OBrien (1990) and Sinha et al. (1993) conclude
that superior analysts do exist.
Using detail Zacks Investment
Research data,
Stickel (1990) shows that analysts annual earnings forecasts can be improved by updating them
with publicly
available
data.
Indeed,
Stickel
(1993) demonstrates
that these updated forecasts
are more accurate than those of the three timely
composites
examined
by Brown (1991). Thus,
the aggregation
principle
pertains
to earnings
which are made at different points in time, provided that they are updated with publicly available data.
In summary,
there are numerous
ways to improve analysts
earnings
forecasts,
such as by
pooling them with time-series
model forecasts,
stock prices, or financial statement
data. Consensus estimates
can be improved
by discarding
old forecasts and by including updated forecasts
in the consensus.
Analysts
do not appear
to
process publicly available information
efficiently
and there appears to be some difference in their
abilities
to forecast earnings.
Why analysts do
not process information
efficiently
and what is
the nature of the information
they underweight/
overweight
are interesting
topics
for future
research.
2.5.

Biases in analysts

earnings

forecasts

Several studies have examined


biases in analysts earnings forecasts. Fried and Givoly (1982)
and OBrien (1988) show that analysts earnings
forecasts generally
are overly optimistic.
Butler
and Lang (1991) find the analysts
degree of

optimism
or pessimism to persist over time and
that the degree of persistence
is independent
of
the level of the firms earnings predictability
(as
defined by Value Lines earnings
predictability
ranking).
Biddle and Ricks (1988) show that
analysts
were overly optimistic
regarding
the
earnings
of firms that adopted
last-in-first-out
(LIFO) in 1974. All of these studies suggest that
analysts
may intentionally
generate
optimistic
forecasts.
However,
Affleck-Graves
et al.
(1990) show that disinterested
judges exhibit
significant
optimistic
biases. This suggests that
analysts biases may be attributable
to judgmental heuristics
employed
in earnings
per share
predictions,
and thus are, partly, unintentional.
Sell-side analysts are more optimistic when the
firms they work for either underwrite
the securities of the companies
whose earnings
they
estimate
[Lin and McNichols
(1991)] or act as
the companies
investment
bankers
[Dugar and
Nathan (1992)]. Sell-side analysts are employed
by brokerage
Firms and their forecasts are reported
externally.
Buy-side
analysts
are employed by banks. insurance
companies
and pension funds, and their forecasts are used internally
by the firms portfolio managers.
Sell-side analysts are relatively
more likely to possess an
optimistic bias because they are paid, in part, by
the amount of commissions
they generate [Dorfman (1991)]. It is much easier to get clients to
buy stocks than to sell them (especially to convince clients to short-sell
stocks they do not
own). Moreover,
if the firms that employ them
underwrite
securities or act as the firms investment bankers,
sell-side analysts have additional
incentives
to provide optimistic
forecasts so as
not to upset their firms clients [Siconolh (1992)].
Dugar and Nathan (1992) show that the stock
market
is aware of and adjusts for this bias.
However,
they do not determine
whether
the
adjustment
is complete,
so it is unclear whether
capital markets fully extract the bias.
Francis and Philbrick (1992) show that Value
Line analysts are overly optimistic, in spite of the
fact that Value Line Inc. neither
underwrites
securities
nor provides investment
banking
ser-

Alternative

interpretations
of the evidence
are that the
testing methodologies
lack sufficient
power or improper
assumptions
are made regarding
market
efficiency
(e.g.
zero information
acquistion.
dissemination
and processing
costs).

Fried

and

Forecaster

l/B/E/S
and Poors

Givoly utilize Standard


and Poors Eurnin~.s
data; OBrien and Butler and Lang use detail
data; Biddle and Ricks use data from Standard
Earnings

Forecaster.

vices. They argue that analysts arc pressured by


managers
to bc overly optimistic
in order to
remain
privy to managers
asymmetric
information,
but an alternative
explanation
for their
result is the Aflleck-Graves
et al. (1990) finding
that
the optimistic
bias is unintentional.
A
stronger
test of analyst optimism
would be to
examine
the biases, if any, of buy-side analysts,
as it is not evident
that these analysts
have
incentives
to be overly optimistic.
Using l/B/E/S
consensus
data. DeBondt
and
Thaler (1990) conclude
that analysts over-react
to extreme
earnings
changes
and returns.
Although
this is not mentioned
by the authors,
analysts
are more likely to over-react
if they
believe
that annual
earnings
follow a random
walk with drift process (e.g. analysts ignore the
evidence
that annual
earnings
revert to their
mean values, especially when year to year carnings changes are extreme).
Klein (1990) argues
that analysts over-reaction
to extreme earnings
changes
and returns
requires
them to undcrpredict
annual
earnings
after large stock price
declines
and over-predict
earnings
after large
stock-price
increases.
Her findings. however, are
the opposite (i.e. analysts earnings estimates are
overly optimistic
after stock-price
declines).
She
interprets
her evidence as suggesting that managers whose firms face adverse conditions
pressure analysts to make overly optimistic forecasts.
However.
in the light of my conversations
with
Value Line personnel,
I would suggest that analysts tend to be overly optimistic
when a firm
expcrienccs
adverse times without
being pressured by managers.
Analysts do not often predict
that firms will fail; they are more likely to expect
a reversion
to the mean
(i.e.
better
times
ahead). Behaviorists
would characterize
such behavior as underweighting
(overweighting)
recent
(past) cues.
In summary,
the evidence suggests that analysts do possess a positive bias; their degree of
optimism
or pessimism
persists through
time;
and their optimistic
bias is greater when firms
employing
them are underwriters
or investment
bankers
of the companies
whose earnings
they
estimate.
Further research is needed to ascertain
how much of the bias is unintentional;
the causes
of intentional
analyst bias; whether capital markets fully extract these, intentional
and unintentional, biases.

2.6.

Modeling

analyst

following

Bhushan (1989) concludes that the number of


analysts following a firm is positively
related to
the number of institutions
holding its shares, the
percentage
of its shares held by institutions,
the
variance of its daily returns, the R-squared
from
the market model regression of its returns on the
market return, and the market value of its equity. His arguments
regarding an expected positive
relation between analyst following and both the
variance
of the firms daily stock-price
returns
and the R-squared
from the market model regression
(i.e. the regression
of the firms daily
stock-price
returns
on the return
of a valueweighted
market portfolio)
are as follows. Private information
is useful and it is more useful
for firms with higher return variability.
The cxpetted trading profit is an increasing function of
the probability
of obtaining
a large deviation
between
the expected
return conditional
on all
known information
and the expected return conditional
on public information.
The expected
trading
profit based on private information
is
higher for firms with higher return variability and
the aggregate
demand for analyst services is an
increasing
function
of the firms return
variability.
Bhushan
argues that a firms marginal
information
acquisition
cost is lower if it has a
higher R-squared
from a market model regression. and that an increase in the R-squared
betwcen firm and market returns
is likely to increase
total expenditures
on analyst
services.
However,
his use of cross-sectional
data to test a
temporal
model makes it difficult to interpret his
results.
Brennan
and Hughes
(1991) show that the
number of analysts following a firm is positively
related to firm size and the variance of return.
and negatively
rclatcd to share price. They contend that managers
with favorable
mm-specific
private information
split their shares in order to
reveal their favorable
private information
to investors: lower priced shares cause increased following by investment
analysts and increased following by analysts
increases
the timeliness
of
firm-specific
information
disclosures.
Brennan
and Hughess hypothesis has not yet been tested.
One way to test it would be to examine whether
firms, after splitting their shares, provide invcstors with more timely information,
such as mak-

L. Il. Brown I Emzings

ing their quarterly earnings announcements


relatively closer to the fiscal quarter-end
to which
the earnings
pertain.
OBrien
and Bhushan
(1990) investigate
the
joint
analyst-institution
decision
process,
and
show that the relation between analyst following
and tirm size is eliminated
once analysts decisions to follow firms and institutional
investors
decisions
to hold these firms in their portfolios
are jointly determined.
They find that changes in
analyst
following
are positively
related to net
entry of firms into the industry and regulation,
and negatively
associated
with the amount
of
pre-existing
analyst following and changes in volatility. Volatility is defined as the residual standard error from a market-model
regression.
The
authors maintain
that analysts prefer regulated
industries
because regulatory
oversight and disclosures provide operating
information
that supplements
financial disclosures,
reducing analysts
information
acquisition
costs. While not mentioned by the authors.
increased
disclosures
by
regulated
firms are likely to reduce
the incremental
value of analysts information
gathering activities.
making it less attractive
for them
to follow these firms. OBrien
and Bhushans
results suggest that analyst following should be
modellcd
jointly
with institutions
decisions
to
avoid drawing
erroneous
inferences
regarding
the determinants
of analyst following.
The lack of robustness
of the Bhushan (1989),
Brennan
and Hughes (1991), and OBrien and
Bhushan (1990) findings to the manner in which
the decision
process is specified suggests that
analyst following
may be related to firm size if
factors additional
to the analyst-institution
decision process are considered.
The lack of robustness to model specification
implies that the theoretical
relation
between
analyst following
and
firm size needs further elaboration.

3. Some evidence of the impact of earnings


forecasting research on capital markets
research
3.1. Accuracy
versus association
surprise with abnormal returns

of forecast

There is mixed evidence


as to whether
the
model that is best in the accuracy dimension
is

forecusring

research

305

best in the association


dimension
(i.e. has the
strongest
association
with abnormal
returns).
Fried and Givoly (1982) focus on annual earnings, and show that forecasts generated
by the
Standard
and Poors Earnings Forecaster outperform
two univariate
time-series
models
in
both dimensions.
Bathke and Lorek (1984) examine the quarterly
earnings
forecasts
of five
univariate
time-series
models, and conclude that
the Brown and Rozeff (1979a) ARIMA model is
best in both dimensions.
Brown et al. (1987a,b)
show that quarterly
earnings forecasts by Value
Line analysts are better than those of five univariate time-series
models in both dimensions.
Brown (1991) and Brown and Kim (1991) use
Zacks Investment
Research
data to show that
three timely composites
are better than the consensus (mean) quarterly earnings forecast in both
dimensions.
Thus, several studies have obtained
similar results using both the accuracy and association
metrics.
Other studies find the opposite,
namely that
more accurate earnings forecasts do not generate
earnings
surprises which are more highly correlated with contemporaneous
abnormal
returns.
In a seminal study, Foster (1977) examined
six
univariate
time-series
models of quarterly
earnings and found the (100) X (010) model plus a
constant
to be best in the predictive
ability dimension,
but the seasonal
random
walk with
drift model (i.e. the (000) x (010) plus constant
model)
to be best in the market
association
dimension.
Hughes
and
Ricks
(1987)
and
OBrien
(19&S), respectively,
use Standard
and
Poors Earnings
Forecaster and IiBlElS
consensus data. and find analysts annual earnings
forecasts
to be better than time-series
model
forecasts in the predictive ability dimension.
but
not in the association
dimension.
Philbrick
and Ricks
(1991)
examine
four
sources of analyst data: Value Line, I/B/E/S,
Standard
and Poors Earnings Forecaster,
and
Zacks Investment
Research.
They find that the
smallest absolute forecast errors are obtained by
pairing
Value Line forecasts
with Value Line
actuals and that the strongest associations.
using
a 3-day window, are obtained
using Value Line
actuals with Value Line or I/B/E/S
forecasts.
Their results suggest that the source of actual
earnings per share (EPS) data is more important
than the source of forecasted
EPS data. This

finding may explain the disparity between studies


using the same sources of forecasts and actuals
[e.g. Brown et al. (1987a,b)]
with those using
different
sources of forecasts
and actuals [e.g.
Hughes and Ricks (19X7)]. The former, but not
the latter, studies have found predictive
ability
and association
to be two sides of the same coin.
Stickel (1990) shows that the accuracy of analysts annual earnings forecasts can be improved
by updating
them with publicly available
data.
Indeed, Stickel (1993) shows that these updated
forecasts are more accurate than the three timely
composites
examined
by Brown (1991). Stickcl
( 1991) shows that earnings-forecast
errors conditional
upon these updated
forecasts
are less
highly associated with abnormal
returns than arc
forecast errors conditioned
on the non-updated
forecasts.
Hopwood
and McKcown
(1990) find
Value Line forecasts.
for quarterly
and annual
earnings.
to be more accurate
than those of
time-series
models and the association
evidence
to be in favor of the Value Line analyst for
annual.
but not quarterly,
earnings.
Wicdman
(1993) obtains the opposite
result, that is, she
finds a higher association
for time-series
model
forecasts than for I/B/E/S
forecasts for annual,
but not quarterly,
earnings.
In summary,
the findings of Fried and Givoly
(1982). Bathke and Lorek (1984). and Brown et
al. (1987a,b)
suggest that accuracy and association arc two sides of the same coin. However.
the evidence of Foster (1977). Hughes and Ricks
(1987).
OBrien
(198X).
Hopwood
and McKeown
(1990).
Stickel (1991),
and Wiedman
(1993) suggests otherwise.
The issue of whether
the model that is best in the accuracy dimension
is best in the association
dimension
is unresolved. If the best model in the association
dimension
is not best in the accuracy dimension.
capital market efficiency in a conventional
sense
appears to be violated, provided that analysts arc
the market, that is they arc the marginal price
setters.
However,
such evidence
is consistent
with behavioral
research showing that individuals
improperly
weight
information.
Alternatively,
the evidence
is consistent
with a less stringent
definition
of market
efficiency
than the one

Wiedman
made more
&dines
casts.

USC\ I/R/E/S
than 2 months

a consensus

detail
prior

data,

discards

forecasts

to the quarter-end

as the median of the individual

and
forc-

which is generally
employed.
Additional
restarch
is needed to determine
which of these
scenarios
is the most descriptively
valid.
3.2. Past quarterly earnings numbers
eurnings unnouncement
drift

and post-

The concept
of post-earnings
announcement
drift has a long history in accounting
and finance
literature
[Ball and Brown (1968), Joy et al.
(1977), Watts (197X). Foster et al. (1984)]. Ball
and Brown showed that stock-price
movements
after annual earnings announcements
are in the
same direction
as consecutive
annual
earnings
changes.
Thus. if annual earnings
in year t exceed, or fall short of, those of year t - 1. stock
prices move up. or down, after year t earnings
arc announced.
Stock-price
movements
after
year t earnings
are announced
are known as
post-earnings
announcement
drift. Rendleman
et
al. ( 1987) showed that stock prices fail to capture
fully the implications
of current quarterly
earnings for future earnings.
Observing
significantly
positive adjacent yuartcrly autocorrelations.
they
argued that investors fail to recognize the implications of quarter
t earnings
for quarter
t+ 1
earnings.
Bernard
and
Thomas
(1990)
extend
the
Rendlcman
et al. analysis by incorporating
implications of quarter t earnings for quarter t + 1 to
t + 4 earnings.
They show that the 3-day stockprice reaction around the time of quarterly earnings reports
is predictable,
conditional
upon
knowledge
of the past four quarterly
earnings
numbers.
More specifically. they argue that stock
prices erroneously
act as if quarterly
earnings
follow a seasonal random walk with drift model,
in fact, quarterly
earnings
follow the
when.
(011) x (011) Box-Jenkins
ARIMA process [i.e.
the model
suggested
by Brown
and Rozeff
( 1979a)] Thus. stock prices mistakenly
behave
as if the current quarterly
earnings shock has a
positive impact on earnings four quarters hence.
I ~lheconventional
ing.

generating

explanations

definition

absumes zetw Costa of transact-

and processing

for model A king

information.
better

the accuracy hut not association


curacy and/or

association

both dimensions
included

than tither

in the study.

dimensions

are not correctly

els A and B arc not indepcndcnt:

Alternative

than model B on
includc:
defined:

ac-

mod-

model C is better

on

model A or B. hut it is not

and no impact on earnings one, two and three


quarters
hence. If stock-price
movements
conformed
with the implications
of the BrownRozeff model,
the stock-price
reaction
to the
quarter t earnings number would incorporate
the
positive impact of the current shock on the next
three quarterly
earnings
numbers,
each with a
lesser influence,
and possibly a negative impact
on earnings
four quarters
hence. The current
shock will exert a negative impact on earnings
four quarters
hence if the moving-average
seasonal component
more than offsets the autoregressive
ordinary
component.
[Recall that the
Brown-Rozeff
model in Box and Jenkins (1976)
notation
is (100) X (01 l)].
Re-examining
the Bernard and Thomas (1990)
result on a holdout sample, Bartov (1992) provides predictive
evidence
which validates
their
findings. He shows that once the implications
of
past quarterly earnings for future quarterly earnings are controlled,
the positive relation between
unexpected
earnings in quarter t and stock-price
drift in quarter
t + 1 disappears.
He concludes
that post-earnings
announcement
drift is fully
explained
by the failure of the market to correctly characterize
the
time-series
process
of
earnings.
The failure of stock prices to fully capture the
time-series
properties
of the quarterly
earnings
generating
process gives rise to the cockroach
theory of earnings
surprises - goodlbad-news
quarterly
earnings
reports precede other good/
bad-news
quarterly
earnings
reports. Similar to
cockroaches,
you never find just one of them!
Bernard
and Thomas
(1990) maintain
that the
expected number of cockroaches
is five - the
current one and the four subsequent
ones, where
the second and third subsequent
earnings shocks
are smaller
in size than their immediate
predecessors,
and the fourth subsequent
earnings
shock is opposite in sign to the current one.
In summary,
there is considerable
evidence
that stock prices fail to capture the implications
of current quarterly earnings for future quarterly
earnings.
The failure of stock prices to fully
capture the time-series
properties of the quarter In a somewhat

similar view. Freeman


and Tse (1989)
found at least a SO% decline in post-earnings
announccment drift following
quarter
ts earnings
announcement
when observations
were matched on quarter t + Is carnings innovations.

ly earnings generating
process appears to violate
semi-strong
market
efficiency
[Fama
(1970)).
Whether
the anomaly
occurs, and, if so, why,
are interesting
topics for future research.
3.3. Analyst behavior and post-eurnings
announcement drift
Mcndenhall
(1991) shows that: (I) analysts
underestimate
the serial correlation
in quarterly
earnings,
(2) investors
use analysts
earnings
forecast revisions to reassess the persistence
of
their previous
forecast errors and (3) investors
systematically
underestimate
the persistence
of
their previous
forecast
errors.
Assuming
that
stock prices capture security analysts earnings
expectations,
Mendenhalls
results are consistent
with those of Bernard and Thomas.
Similar
to Mendenhalls
results
regarding
quarterly earnings, but in contrast with an earlier
study of annual earnings by Givoly (1985), Ali et
al. (1992) find that analysts generally
underestimate the permanence
of the current annual earnings surprise
with regard
to future
annual
earnings. More specifically,
Ali et al. show that
analysts generally are overly optimistic regarding
the subsequent
years annual earnings and that
their annual earnings forecast errors display significantly
positive
serial
correlation.
Givoly
found analysts forecasts to be unbiased and their
prediction
errors
to be serially
uncorrelated.
However,
his sample is much smaller than that
of Ali et al. (i.e. Givolys results were based on
as few as 378 firm-years; Ali et al.s findings were
based on 3184-5305
firm-years).
The latter rcsults appear to be relatively more credible. given
the studys larger sample and more powerful
methodology.
Using Value Line data, Abarbanell
and Ber Stock prices do not fully capture
analysts expectations,
but the degree of proximity
should be close enough to
make the Mendenhall
(IYYI) and Bernard
and Thomas
(1989. 1990) results reasonably
consistent with each other.
As evidence that stock prices do not fully capture analysts
expectations,
see Ahdel-khalik
and Ajinkya
(1982) and
Elton et al. (1981).
who show that analysts
forecast
revisions led stock price changes. Their respective sources
of analyst forecasts are Merrill Lynchs Option Alerr and
I/B/E/S
consensus
data.
Mendenhall
u5es
Value Line data; Givolv uses Standard
and Poors Eurnings Forecaster detail data; Ali et al. use
I/B/E/S
consensus
(median) data.

nard (1992) show that analysts earnings


forecasts under-react
to recent earnings
announcements in a manner
similar to the -?-day stockprice under-reaction
observed
by Bernard
and
Thomas
(1990).
They conclude
that security
analysts behavior
is, at best, only a partial explanation
for stock-price
under-reaction
to earnings and. hence,
post-earnings
announcement
drift. Their conclusion
is consistent
with the
cvidencc
of Brown and Rozeff (197Yb), who
showed that analysts do not act as if the quarterly earnings generating
process is a seasonal random walk with drift. In order for analysts behavior to explain
post-earnings
announccmcnt
drift, analysts would have to act as if the carnings generating
process was a seasonal random
walk with drift. As mentioned
earlier, Bernard
and Thomas
(1YYO) maintain
that stock prices
erroneously
act as if quarterly
earnings follow a
seasonal random walk model with drift (seasonal
submartingale)
model. Both Brown and Rozcff
( lY7Yb) and Abarbanell
and Bernard
( lYY2)
show that, similar to the Foster (1977) model,
analysts
act as if the quarterly
earnings
timeseries process is autoregressive.
Thus, it is not
surprising
that security
analysts behavior
only
partly explains the Bernard and Thomas ( 1990)
results. Additional
work is needed to further our
understanding
of the relation
between
analyst
behavior
and post-earnings
announcement
drift.

Some studies have examined


the relation between earnings
announcements
and the convergence of analysts expectations
of future carnings.
Bayesian
theory
suggests
that earnings
announcements
generally
increase
the
precision of individuals
estimates of future earnings
numbers.
The dispersion
among
individuals
earnings
expectations
is often used to proxy for
the precision
of individuals
estimates
of future
earnings
[Ajinkya
and Gift (lYX5), Brown ct al.
( lYX7c)]. Morse et al. ( 199 1) examine this issue
with l/B/E/S
consensus
(mean) forecasts
and
conclude
that earnings announcements
decrease
the homogeneity
of analysts cxpcctations
of future earnings.
Their results are inconsistent
with
their intuition,
Bayesian
theory and the timcseries properties
of earnings.
The joint evidence

that the quarterly


earnings generating
process is
autoregressive
[ Lipe and Kormendi
( lYY3)], and
that analysts
act as if the quarterly
earnings
generating
process is autoregressive
[Brown and
Rozcff ( lY79b)], implies that earnings announcements should increase the precision of analysts
earnings expectations.
The intuition
is that each
individual
revises his/her
forecast in the direction of his/her error (actual minus forecast).
As
shown by Morse et al. ( lY91), unless the errors
are very large. the homogeneity
of analysts forccasts of future earnings
will increase after observing current earnings.
Brown and Han (IY92) argue that the Morse
ct al. results arc due to their use of consensus
(mean) analyst forecasts. Since analysts earnings
forecasts arc not all made at the same point in
time [OBrien
(19xX)], some of the individual
forecasts
included
in the consensus
obtained
after the earnings
announcement
were dated
prior to the earnings
announcement.
In other
words, not all forecasts included in the Morse et
al. study
were conditioned
on the earnings
announcement
that the analysts supposedly
had
as inputs into their decision models. Brown and
Han ( lYY2) USC detail I/B/E/S
forecasts dated
after the earnings announcement,
and show that
earnings
announcements
generally
increase
the
homogeneity
of analysts expectations
of future
earnings.
Consistent
with Bayesian
theory and
Morse et al.s intuition.
Brown and Han find that
analysts expectations
of future earnings become
more dispersed
only when earnings
surprise is
very large (i.e. for the top decilc of earnings
surprise).
The joint evidence
that timely composites
of
analyst
earnings
estimates
are more accurate
than
the consensus
(mean)
analyst
estimate
[OBrien (IYXX), Brown (1991)]. that errors conditional
upon timely composites
of earnings are
relatively
more highly associated with abnormal
returns [Brown and Kim (19Yl)] and that analysts expectations
of future
earnings
become
more homogenous
upon observing current earnings when using individual.
but not consensus,
analyst forecasts [Brown and Han (lY92)]. suggests that researchers
should use individual analysts forecasts in lieu of consensus analysts forecasts when requiring
a proxy for the markets
earnings
expectation
or the precision
of individuals expectations.
Given that l/B/E/S
and

L, D. Rrown

Zacks
access
search

I Earnings forecurting

Investment
Research now have increased
to these data, it is likely that future rewill rely more on individual
analyst data.

3.5. Analyst forecasts and abnormal

returns

Elton et al. (1981) provide evidence that investors cannot earn abnormal
returns by selecting stocks based on analysts consensus
growth
forecasts.
In contrast,
Givoly and Lakonishok
(1979), Abdel-khalik
and Ajinkya (1982), Hawkins et al. (1984), Freeman and Tse (1989), and
Beneish (1991) show that consensus analyst forecast revisions can bc used to predict subsequent
A potential
reason for the
abnormal
returns.
discrepancy
is that Elton et al. did not examine
forecast
revisions.
It would be instructive
to
investigate
whether
this is the reason for the
discrepancy
in results between
Elton et al. and
the other studies,
and if larger profits can be
made by using timely analysts forecasts,
those
pooled with time-series
model forecasts and/or
those
updated
using the Stickel (1990) procedure.
It also would
be interesting
to determine
whether
larger profits can be made by using
timely, or updated,
analyst estimates
in lieu of
the consensus estimate. Stickels (1991) evidence
may be instructive
in this regard. Showing that
the analysts
current
outstanding
forecast
is a
better measure of the markets earnings expectation, in the association
dimension,
than is an
updated forecast,
he uses this result to create a
profitable
trading strategy that predicts changes
in outstanding
forecasts.
Additional
research is
needed to reveal how and why analyst expectational data can be used to produce
abnormal
returns.

4. Earnings
4.1.

response

coefficient

literature

rc~earch

309

in accounting
earnings
to abnormal
stock returns [Easton and Zmijewski
(1989)]. The ERC
directly
links earnings
forecasting
research
to
capital markets
research because the definition
of earnings surprise is conditional
upon an earnings expectations
model. ERC studies often require a proxy for the earnings persistence
factor,
which relates changes in expected future earnings to current
earnings
surprise,
because,
in
theory, the valuation
implication
of an earnings
number
(the ERC) is positively
related to its
degree of permanence
(the earnings persistence
parameter).
Studies have not been uniform regarding the ARIMA
model chosen to estimate
the earnings persistence
parameter.
Some of the
assumed
ARIMA
models and their advocates
are: ARIMA
(O,l,l)
[Beaver
et al. (1980)],
ARIMA
(0,O.l)
[Miller
and
Rock
(1985)].
ARIMA
(2,1,(l) [Kormendi
and Lipe (1987)],
and
ARIMA
(1 ,O,O) [Ramakrishnan
and
Thomas
( 1992)].15 Surprisingly,
little work has
been done examining
the sensitivity
of ERC
estimates to alternative
ARIMA model specifcations. [See Kendall and Zarowin (1990) for some
limited
evidence
on this issue.] The external
validity of this research would be enhanced
(or
lessened)
considerably
by showing that the results are (or are not) robust
to alternative
ARIMA
model specifications.
Easton and Zmijewski
argue that the analysts
earnings
revision coefficient
(i.e. the coefficient
relating current earnings to analysts estimates of
future earnings)
is an explanatory
variable
of
ERC. They document
a positive association
betwecn the ERC and the analysts earnings revision coefficient,
and a negative association
be If earnings
can be modclled
as an AROMA process and
changes
in expectations
about future earnings
produce
dollar for dollar changes
in expectations
about future
dividends,
Flavin (1081) shows that the theoretical
ERC is
(I + persistence factor) where the persistence factor is:

Proxies for earnings surprise

The earnings
efficient relating

response
coefficient
is the cothe surprise (new information)

Elton

ct al. use I/B/E/S


consensus
data. Abdel-khalik
and Ajinkya USC Merrill Lynchs Oprions Alert data. Givoly and Lakonishok
use Standard
and Poors Eumings
Fbrecouter data,
Hawkins et al. use I/B/E/S
consensus
data. aand Ben&h
uses analysts information
reported
in
the WuN Street Journals Heard on the Street column.

where 4, is an autoregressive
coefficient
moving average
coefficient
of order i;
orders
of the autoregressive
process,
moving average processes,
respectively;
of capital.

of order C; 0, is a
p. d and q arc
differencing
and
R is 1 + the cost

tween ERC and systematic risk. The authors rely


on the earnings
capitalization
model of Miller
and Modigliani
(1961) to test their hypotheses
that ERCs are increasing
in the revision parameter and decreasing
in risk. The revision parameter and risk are positively
correlated
with the
numerator
and denominator
of the earnings
capitalization
model, respectively.
The authors
include
firm size to control
for measurement
error in their proxy for unexpected
earnings and
not to test a hypothesis.
They use the analysts
consensus
(mean) earnings estimate to proxy for
expected
earnings
and current earnings surprise
to estimate the ERC. Both of these metrics may
need reconsideration
in light of recent evidence.
Regarding
the use of consensus
forecasts, the
evidence
of Brown (1991) and Brown and Kim
(1991) suggests that expected earnings and earnings forecast
revisions
may be sensitive
to the
utilization
of timely composite forecasts in lieu of
consensus
estimates.
With regard to the use of
current
earnings
surprise to estimate the ERC,
Bernard
and Thomas
(1990), Wiggins (1990).
and Bartov
(1992) show that abnormal
stockprice
movements
at the
time
of earnings
announcements
are related to current earnings
and the four most recent quarterly earnings numbers. Cornell and Landsman
(19X9) and Abdelkhalik (1990) show that forecast revisions provide incremental
explanatory
power in a regression of abnormal
returns
on earnings
surprise.
These
papers
use I/B/E/S
consensus
(mean)
forecasts and Value Line forecasts as proxies for
the markets
(unobservable)
earnings
expectation. Kothari
and Sloan (1992) show that the
ERC from a regression
of returns on contemporaneous
earnings
changes
is biased towards
zero and that the bias is mitigated
by including
leading-period
returns
in the ERC estimation
equation.
Thus, ERC studies confined to current
earnings
[e.g. Kormendi
and Lipe (1987), Collins and Kothari (1989), Easton and Zmijewski
(1989)]
are likely to suffer from an omitted
variable problem, causing inefficient and possibly
biased ERC estimates.
In summary.
ERC studies using a short window [e.g. Easton and Zmijewski
(1989)] require
a precise measure
of earnings
response.
(The
issue of narrow versus wide windows is discussed
further
in Section 4.3.) Additional
research
is
needed
to determine
whether
the results
of

studies modeling
ERCs are robust to ERC estimates incorporating
timely earnings
forecasts.
lagged quarterly
earnings,
revisions
of future
quarterly
earnings
and alternative
specifications
of the earnings
time-series
process.
4.2.

Determinants of ERG

Collins
and Kothari
(1989) find that ERG
increase in economic
growth and/or time-series
persistence
[Kormendi
and Lipe (1987)],
and
decrease
in interest rates and risk. Lipe (1990)
shows that the ERC is positively related to both
the predictability
of the earnings series and the
time-series
persistence
of earnings. The forecasting literature
as it pertains
to the time-series
properties
of earnings
is relevant to the Collins
and Kothari
(1989) and Lipe (1990) studies.
These studies, respectively,
assume, but do not
test, that the time-series
process of annual earnings follow an ARIMA
(O,l,l)
[Beaver et al.
( 19SO)] and ARIMA
(2,l .O) [ Kormendi
and
Lipe (1987)] process. However,
Lipe and Kormendi (1993) show that annual earnings follow a
fourth-order
autoregressive
process and that a
dollar of year t earnings
surprise
leads to an
approximately
8O-cent revision
in year t + 1
earnings.
The robustness
of the Collins and
Kothari (1989) and Lipc (1990) results to alternative definitions
of earnings persistence.
such as
the one advocated
by Lipe and Kormendi
(1993), awaits further research.
Brown and Zmijewski
(1987) find that ERCs
are smaller during strikes than at other times.
This finding is consistent
with the joint hypothesis that price-irrelevant
and transitory
earnings
are higher during strikes, and that the capital
market assigns a smaller multiplier to these earnings components
[Ramakrishnan
and Thomas
(1991)]. Collins and DeAngelo
(1990) find ERCs
to be higher during proxy contests
and lower
after proxy contests than at other times. They
consider the former result as surprising,
because
earnings
management
during
proxy
contests
should lower ERCs and the latter result as con Lipe and Kormendi consider the general ARIMA ( p,d,q)
model. set tl = I, q = 0. and focus on the effects of changing /7. the order of the ARIMA
(p.1.0) t~~odel.They
cstimnted
the parameters
for the first through tenth orderh. but found most of the improvements
to occur hy the
fourth order.

L. D. Brown

i Earnings

sistent with firms taking a big (earnings)


bath
after experiencing
management
changes [Elliott
and Shaw ( 19SS)]. Choi and Jeter ( 1992) show
that a firms ERC declines
significantly
after
issuing qualified
audit reports for a sample of
subject-to
and consistency
qualifications.
Their
results are consistent
with the hypothesis
that
audit qualifications
reduce the markets responsiveness to earnings
announcements
by altering
perceptions
of earnings noise and/or persistence.
These studies suggest that a temporal estimate of
a firms ERC should adopt a methodology
incorporating systematic non-earnings
shocks, such as
that employed
b.y Lee and Chen (1990).
Using the variance in analysts earnings forecasts, prior to a firms annual earnings announcements, to proxy for its ex ante uncertainty,
Imhoff and Lobo (1992) examine the effect of ex
ante earnings
uncertainty
on ERCs.
Finding
firms with relatively
high ex ante uncertainty
to
exhibit little or no systematic stock-price change
when earnings
are announced,
they conclude
that dispersion
in pre-announcement
analysts
earnings
forecasts proxies for noise in the earnings signal. However, Kim and Verrecchia (1991)
maintain
that expected stock-price
change is related to the resolution
of ex ante uncertainty,
not
to its amount.
The resolution
of uncertainty
is
related to both the amount of ex ante uncertainty and the precision
of the earnings
announcement. A commonly
used proxy for resolution
of
uncertainty
is the change in dispersion
of analysts earnings
forecasts
[Morse et al. (1991),
Brown and Han (1992)]. The robustness
of the
Imhoff and Lobo results should be examined
using the change in dispersion
of analysts earnings forecasts
at the time
of an earnings
announcement,
in lieu of the level of analyst
dispersion
prior to the announcement.
In summary,
ERCs are positively
related to
economic growth and/or time-series
persistence,
and negatively
related to interest rates and risk.
They are smaller when earnings
noise is high,
earnings
persistence
is low and ex ante uncertainty is high. The impact of the resolution
of ex
ante uncertainty
on ERCs awaits further study.
4.3. Functional
equation
Abdel-khalik

form

of the ERC

(199(l),

estimation

Das and Lev (1991)

and

forecasting

research

311

Freeman
and Tse (1992) show that the relation
between abnormal
returns and unexpected
earnings is non-linear.
Abdel-khalik
(1990) shows
that a mode1 relating
abnormal
returns to unexpected earnings is better specified if it includes
the past and future revisions of earnings expectations in the estimation
equation.
Das and Lev
(1991) show that the non-parametric
procedure,
locally weighted
regression,
yields significantly
more explanatory
power
than ordinary
least
squares
(OLS) in an abnormal
return on unexpected earnings regression,
especially for small
earnings
surprises.
Freeman
and Tse (1992)
show that the marginal response of stock price to
unexpected
earnings
declines
as the absolute
value of unexpected
earnings increases.
Assuming that the absolute value of unexpected
earnings is negatively
related to earnings persistence
and that it is relatively easier to forecast permanent
earnings
than transitory
earnings,
they
maintain
that these two assumptions
imply that
transitory
earnings surprises are concentrated
in
the tails of the unexpected
earnings distribution.
Their evidence
is consistent
with the view that
the earnings-return
relation
is S-shaped
(i.e.
convex
for bad news and concave
for good
news).
Cheng et al. (1992) evaluate the specification
of the commonly
used method of estimating
the
ERC, namely
a linear regression
of abnormal
returns
on unexpected
earnings.
They include
three samples of earnings forecasts - the random walk mode1 forecast
and two sources of
analyst forecasts,
I/B/E/S
and Value Line and examine returns for both short (2-day) and
long (more than 1 year) event windows. Testing
for non-linearity,
heteroscedasticity,
residual
non-normality,
omitted variables and coefficient
variation
across firms, they generally
reject the
classical normal linear regression
model assumption (i.e. in almost all cases, they find the relation to be non-linear).
The Cheng et al. (1992)
and Freeman and Tse (1992) studies suggest that
ERC estimation
models based on the linearity
assumption
are misspecified
and may lead to
inconsistent
inferences.
Cheng et al. replicate
Cornell and Landsman
(1989), and find that the
specification
problems are severe enough to lead
to substantial
instability
in the linear regression
models inferences.
Their OLS analysis of Cornell and Landsmans
data revealed
problems

with heteroscedasticity,
non-linearity
and nonnormality.
The model with the ranked forecast
errors (the Cheng et al. approach)
produced
a
considerably
higher degree of coefficient stability, R, and larger t-values than the conventional,
unranked
forecast error approach
(used by Corncll and Landsman).
In contrast with Abdel-khalik
(19SO), Das and
Lev (1991) and Freeman
and Tse (1992), who
conduct
cross-sectional
and in-sample
analyses,
Beneish
and Harvey
(1993) conduct
temporal
and cross-sectional
analyses,
and provide fitting
and predictive
evidence.
Their temporal
and
cross-sectional
analyses
on seven models (one
linear. four non-linear
and two non-parametric)
suggest
that linear
models
arc inferior
when
using in-sample
data, but not when using out-ofsample data. Providing evidence that some of the
non-linear
models overfit the data. they suggest
that temporal
and cross-sectional
ERC changes
should be identified
with a methodology
that is
robust to influential
observations
(c.g. adaptive
non-parametric
regression).
Mendenhall
and Nichols (1988) find the ERC
response to bad-news earnings to depend on the
announcements
fiscal quarter. More specifically,
the ERC is greater for interim than for annual
(fourth
quarter)
reports.
Their results suggest
that ERC models should allow for differential
rcsponscs
to bad news for interim versus annual
earnings announcements.
However. their empirical tests arc incomplete
as they exclude goodnews earnings
[Palcpu (1%X)]. Additional
theoretical
development
and empirical
tests are
necessary
before definitive
conclusions
can be
drawn
Ohlson
(lC)Yla,b)
and Ohlson
and Shroff
(1992) argue that earnings levels, deflated by the
beginning-of-period
price, often arc more appropriate
than
are
properly
deflated
earnings
changes as proxies for unexpected
earnings when
modeling
rates of return.
Easton
and Harris
(1901) and Ali and Zarowin
(1992) use long
windows
(12-month
returns).
and provide cvidence consistent
with these models. Their tindings suggest
that long-window
ERC
studies
should include earnings
levels to appropriately
estimate
ERCs.
E&ton and Harris
earnings
Y.,lhfi,3/P,,

introduce

a multivariate

analysis

ot

R,, = r,!, + y,,[ A,,il,,


,I +
,) + E!, where: R,,. return on a share of firm j

levels and changes:

Lev ( 1989), Easton et al. ( 1992) and Warfield


and Wild (1992) show that earnings
levels and
stock prices are more highly correlated over long
(4-10 years) windows than short ones (one quarter to 1 year). Kothari
(1992) also uses long
windows and examines
alternative
specifications
of ERC estimation
procedures
when stock prices
lead earnings.
He shows that when stock prices
impound
information
that is not reflected in the
time-series
of past earnings,
earnings levels provide relatively
more explanatory
power than
earnings changes (both deflated by beginning-ofperiod stock price). and that including
earnings
levels in the ERC estimation
equation
may lessen the ERC bias. The need for a precise cxpcctations model is diminished
when a long-window
approach
is used. With a longer window,
the
generally accepted accounting
principle (GAAP)
conservatism
problem is less severe because what
is reported
in financial statements.
via the realization
and matching
principles,
is more congruent with what is impounded
in share price.
In summary.
while the evidence
regarding
whether
the returns-earnings
relation
is nonlinear is mixed, the relation
is better specified
with both (appropriately
deflated) earnings levels
and changes than with earnings
changes alone.
The need for a precise earnings
expectations
model, such as recent analyst forecasts.
is less
important
when a long-window
approach is used.
Additional
research
is needed to ascertain
the
impact on ERC estimates of both the fiscal quarter and the nature of the news (i.c. whether it is
good or bad).
4.4. Stock-price

expectutions

md

earnings

eqectations

Bandyopadhyay
et al. (1993) examine whether
analysts incorporate
their earnings forecasts into
their stock-price forecasts and if the usefulness of
analysts earnings
forecasts for their stock-price
forecasts
is related to the length of their joint
stock price to earnings
forecast horizon.
They
show that analysts act as if earnings forecasts are

over

the

months
earnings

I2 months
after

extending

9 months
annual

per share of tirm j for period

share of firm j at time I time I.

from

the tiscal year end: A,,.

I : E,,is

random

I: I,,

prior

to 3

accounting

,,

price pel-

error of firm j at

important
for their stock-price forecasts and that
earnings
are relatively more important
for price
in the longer term. Their finding that long-term
ex ante ERCs arc relatively greater than shortterm ex ante ERCs is consistent with Lev (1989),
Kothari and Sloan (1992), and Ohlson and Penman (1992), who show that long-term
ex post
ERCs
are relatively
greater
than short-term
ones. (Ex ante ERC studies relate expected returns to expected earnings;
cx post ERC studies
relate actual returns to actual earnings.)
There are two factors at work regarding
the
relation between the ERC and the length of the
horizon.
On the one hand, GAAP conservatism
suggests a higher ERC when the joint earnings
and price horizon is lengthened
because cumulated reported
earnings
include relatively
more
of what is impounded
in price. On the other
hand, the ERC over the life of the firm (birth to
death) is 1.0. As a longer horizon is more akin
to the life of the firm, the ERC may be smaller
over a longer horizon (i.e. it may approach
1.0
asymptotically
from above).
Consistent
with
Levs (1989) argument
that GAAP conservatism
in measuring
earnings causes stock prices to lead
accounting
earnings,
the Bandyopadhyay
et al.
(1993) evidence
suggests that non-earnings
information
is relatively
more
important
for
shorter-term
price estimates.
However,
Bandyopadhyay
et al. (1993) are silent regarding the
nature of the non-earnings
information,
and they
do not say how analysts stock-price
forecasts,
earnings forecasts and non-earnings
forecasts relate to the capital markets expectation
of these
variables.
Govindarajan
(1980) uses content analysis to
examine
the importance
given by security analysts to earnings
versus cash flows when they
make their recommendations.
Security analysts
comments
on companies
in 976 reports were
scrutinized
using the Wall Street Transcript, and
rated from one to six, according to the authors
interpretation
of the commentary,
as follows: (1)
no mention
of earnings,
(2) earnings analysis is
less important
than cash flow analysis, (3) earnings analysis and cash flow analysis are equally
important,
(4) earnings
analysis
is more important
than cash flow analysis,
(5) earnings
analysis only, except for mention
of dividends
and (6) no mention of cash flows. Security analysts attached more weight to earnings than cash

flows in 86.6% of the total number


of reports
examined.
Consistent
with Bandyopadhyay
et al.
(1993), Govindarajan
(1980) finds earnings to be
an important
source of analysts recommendations. In contrast,
Bouwman
et al. (1987) use
protocol analysis and conclude that income statements do not play a dominant
role in analysts
selection of investment
candidates.
More specifically, the authors provided professional
financial
analysts
with a package
of financial
materials
that reflected
the magnitude
and types of information
they normally
review when screening
prospective
investments.
The
analysts
were
asked to verbalize whatever came to mind during their evaluations.
The authors resulting transcripts. called protocols,
served as the basis for
their analyses.
In summary,
capital markets research suggests
that analysts incorporate
their earnings forecasts
into their stock-price
forecasts
and that their
earnings forecasts are relatively more important
for their stock-price
forecasts in the longer term.
The evidence from behavioral
research regarding
the importance
of earnings
forecasts
is mixed.
Little is known from either branch of research
about the way in which analysts formulate
their
stock-price
forecasts
or about the role of accounting
and non-accounting
information
in this
process.

5. Some directions for future research in the


capital markets/earnings
forecasting areas
Sections 2 to 4 discuss numerous
directions for
future research so this section discusses only a
few such directions.
Several studies have examined
the relation
between
analysts earnings
forecasts and firm-specific
factors, such as stock
prices, changes in consensus
earnings estimates,
deviation from the consensus and forecast errors.
However,
with limited exceptions
[Stone (1977),
Fildes and Lam (1990)], the macroeconomic
and
industry
factors asserted by analysts to be important
cues to their decisions
have been ignored.
This appears
to be a fruitful area for
future
research.
Moreover.
there has been a
scarcity of research examining
linkages between
analysts stock-price
forecasts and other factors
of interest
(e.g. earnings
forecasts,
past stock
prices, and macroeconomic
and industry data).

As earnings forecasts are not ends in themselves,


but
are inputs
into stock
recommendations
[Schipper (1991)], future research should explore
these linkages.
A second direction
for future research
is to
link post-earnings
announcement
drift with convergence of analysts earnings expectations.
Bernard and Thomas (1990) suggest that drift is a
delayed
response
to information.
Verrecchia
(1980) argues that the rapidity
of stock-price
adjustment
to information
increases as the precision of the information,
as determined
by the
consensus judgment
of investors, increases. If so,
drift should be most evident when analysts forecasts converge
slowly after earnings
announcements. This issue should be addressed
using
individual,
rather than consensus,
analysts earnings forecasts,
allowing the researcher
to ascertain if the forecast is made conditional
upon a
particular
earnings announcement.
A third direction
for future
research
is to
examine
the role of analysts forecasts if capital
markets are efficient compared
with their role if
capital markets are inefficient.
If capital markets
arc informationally
efficient in the conventional
sense, and if the analyst is the market. forecast
accuracy and market association
are two sides of
the same coin, and the role of analysts is rather
limited. More specifically,
in this case, a composite of analysts (and perhaps other) forecasts can
proxy for the markets (unknown)
earnings
expectation
and examination
of the analysts expectation formulation
process is not especially informative. However.
if capital markets are not efficient in the conventional
sense, the role of analysts forecasts
is much broader,
provided
that
analysts
are reasonably
efficient
processors
of
financial information
and stock prices ultimately
reflect fundamentals.
In a return to the world of
Graham
et al. ( 1962), analyst behavior
merits
investigation
to determine
how fundamental

The relation
hetwcen
post-earnings
announcement
drift
and the rapidity of convergence
of analysts expectations
is
best examined for positive earnings surprises. Stock prices
should he lower when uncertainty
is heightened
[Vxian
(19X5)], unconditional
on the sign of the earnings surprise.
Thus. stock prices should drift upwards
after earnings
announcement
drift for both good news and bad news. An
increase
in stock prices owing to uncertainty
resolution
may provide an explanation
for positive earnings surprises
being followed by positive drift.

financial statement
analysis can be used to detect
mispricing.
[Lev and Thiagarajan
(19YI), Penman (1992~). and others have recently advocated
a return
to fundamentals.]
Such research
should examine how (and why) analysts use (and
ignore) certain financial statement
data.
The manner
in which analysts
use/neglect
earnings
and non-earnings
information
is crucial
for a better understanding
of how stock prices
are formulated.
If capital markets
are not informationally
efficient, it is important
to understand how analysts
process earnings
and. nonearnings
information
to estimate intrinsic value,
which is not necessarily
impounded
in stock
price. Behavioral
research,
which has played a
limited role to date in understanding
how analysts formulate
their expectations,
is likely to play
a more prominent
role in the future.

6. Implications
areas

for forecasting

research

in other

In the past 1.5 years, forecasting


literature
has
developed
a number of themes which have been
addressed
by the literature
surveyed
in this
paper. Some of these themes are: (i) the ex ante
comparison
of accuracy and the stability of models out-of-sample:
(ii) the choice of error measure and if the particular
error measure chosen
matters; (iii) the comparative
analysis of method
performance
using comparable
information
sets;
(iv) heuristics
and biases of judgment
and their
incompatibility
with rational expectations.
It has been shown that complex earnings forecasting models often fail in out-of-sample
comparisons
[Albrecht
et al. (1977).
Watts and
Leftwich
(1977).
Foster
(1977),
Beneish
and
Harvey
(1993)].
It is easy to over-identify
a
structural
model using in-sample
data, but simpler models have proven to be more robust in
holdout periods. It appears that Ockhams razor
is valid - parsimonious
models often are preferable to complex ones.
With accounting
data, the choice of error measure is important,
as the denominator
can be
close to zero or even negative [Lev and Sunder
(1979), Frecka and Hopwood ( 1983)]. Thus, the
results of parametric
tests may rely heavily on
outliers.
In order to reduce the outlicr problem,
many studies
deflate by stock price [Christie

L. D. Brown

I Earnings forecasting resecrrch

(1987)], trim the sample [Foster (1986)], or use


non-parametric
tests [Brown and Rozeff (1978)].
Earnings forecasting
research has occasionally,
but not often, provided comparative
analyses of
method
performance
with comparable
information sets. For example,
Brown et al. (1987~)
provide similar findings with each of two sources
of analysts
earnings
forecasts.
However,
the
ERC literature
has typically
provided
results
conditional
upon a particular
assumed
timeseries process. Given the unfortunate
aversion to
replication
in the social sciences (cf. the hard
sciences where replication
is the norm), it behooves social science researchers
to provide results with multiple data sources.
The earnings forecasting
literature
has recently, but belatedly,
begun to examine
heuristics
and biases of judgment,
and its incompatibility
with rational expectations
[Affleck-Graves
et al.
(1990), Ah et al. (1992), DeBondt
and Thaler
(1990)]. This research is still in its infancy. Joint
efforts by capital market
researchers
and behavioralists
to examine
these issues more thoroughly would considerably
enhance
our understanding
of the role of analysts
in the price
formation
process.
Earnings
forecasting
research appeared
to be
at a dead end in the late 1970s. Annual earnings
were known to follow a random walk with drift
process,
except when annual
earnings
changes
were extreme,
and three univariate
ARIMA
models of the quarterly earnings time-series process were introduced.
Little work has been done
since the late 1970s to develop more accurate
time-series
models. Two events transpired
in the
late 1970s which provided an impetus for earnings forecasting
research.
First, analysts earnings forecast
data became readily available
in
machine-readable
form. Second, capital markets
research
began to look to forecasting
research
for proxies of the (unobservable)
market earnings expectation.
Earnings
forecasting
research
has flourished
since its use of analyst expectational
data and its
linkage
to capital markets
research.
Increased
information
availability,
via analysts
earnings
estimates,
and increased
focus on information
usage, via the capital markets, has greatly aided
earnings
forecasting
research.
Other disciplines
utilizing
forecast data may be able to sharpen
their intellectual
foci and broaden their perspec-

tives by looking for ways to enhance


availability
and information
usage.

31s

information

References
Abarbanell.
J.S., 1991. Do analysts earnings forecasts incorporate
information
in prior stock price changes?,
Journul of Accounting und Economics, 14, 147-165.
Abarbanell,
J.S. and V.L. Bernard,
1992, Tests of analysts
overreactioniunderreaction
to earnings information
as an
explanation
for anomalous
stock price behavior,
Journul
of Finance, 47, 1181-1207.
Ahdel-khalik,
A.R.,
1990, Specification
problems
with information
content of earnings:
revisions and rationality
of
expectations
and self-selection
bias, Conternporq
Accounting Research, 7. 142-172.
Abdel-khalik.
A.R. and B.B. Ajinkya.
19X2, Returns
to
informational
advantages:
the case of analysts
forecast
revisions.
Accounting Review. 57. 661-680.
AfHeck-Graves.
J., L.R. Davis and R.R. Mendenhall,
1990.
Forecasts
of earnings per share: possible sources of analyst superiority
and bias, Contemporary Accounting Review. 6. 501-517.
Ajinkya.
B.B. and M.J. Gift, 1985, Dispersion
of financial
analysts earnings forecasts and the (option model) implied
standard
deviations
of stock returns,
Jourrrul of Finance.
40. 1353-1365.
Albrecht.
W.S., L. Lookabill
and J. McKeown,
1977, The
time series properties
of annual
earnings,
Journal of
Accounting Re.yeurch. IS, 226-244.
Ah, A. and P. Zarowin,
1992. The role of earnings levels in
annual earnings-returns
studies,
Journal of Accounting
Research, 30. 2X6-296.
Ah, A.. A. Klein and J. Rosenfeld,
lY92. Analysts
use of
information
about permanent
and transitory earnings components
in forecasting
annual EPS, Accounting Review,
67, 183-398.
Ashton, A.H. and R.H. Ashton,
1985, Aggregating
subjective forecasts:
some empirical
results,
Management Science 1 31 . 1499%1508.
Atiase, R.K.. 1985, Prcdisclosure
information.
firm capitalization,
and security
price
behavior
around
earnings
announcements.
Journal of Accounting Research. 23. 2136.
Ball. R., 1992, The earnings-price
anomaly,
Journul of
Accounting und Economics, 15, 319-345.
Bail. R. and P. Brown, 196X. An empirical evaluation
of
accounting
income numbers.
Journal of Accounting Research. 6, 159-178.
Bail. R. and R. Watts, 1972, Some time series properties
of
accounting
income, Journal of Finunce, 27, 663-681.
Bamber,
L.S.. 1987, Unexpected
earnings,
firm size, and
trading
volume
around
quarterly
earnings
announcements, Accounting Review, 62, 510-532.
Bandyopadhyay.
S.P., L.D. Brown and G.D. Richardson,
1993, The importance
of analysts earnings forecasts
to
their stock price forecasts,
Working paper, University
of
Waterloo.

Bartov.

E.,

IYY2.

cxplanatirm
Ikll.

Patterns

in unexpected

for post-announcement

earnings

drift..

as an

Accoutttin~

Rr-

67. 610-622.

Bathkc.

Jr. A.W.

between

Lorek.

models

of quarterly

lY84.

The

relationship

and the security

earnings.

markets

A~outt/it~g

ex-

Re~~ir~~, SY.

163-176.
Beaver.

W. and D.

Morse.

ratios?,

W.H.,

W.H..

R.A.

information

D.

price-

33. 65-76.

Morse.

prices.

1080.

Journul

The
of

AC

mrl

and S.G.

Ryan.

prices:

Economics.

1087,

a second

The
look.

Y. 13Y- 157.

M.D.

and C.R.

the earnings-return

Harvey.

1993. The

relation,

Working

specification

paper.

Duke

of

Un-

versitv.
G.J.

and R.L.

of earnings.
Bernard,
mium?.
Bernard.

Watts.

Working

V.L.

and

announcement
Jourrtul
V.L.

paper.

J.K.

drift:

197X. The
University

Thomas,

delayed

for

future

EcotKttnic.s,
Bhushan.
ing,

response

R..

Jourrwl

of

carn-

Accorrt~fit~g

of

G.C.

characteristics

und

Accounting

and W.E.

anti

Ricks,

and analyst follow-

Economic~.s.

1988, Analyst

Journal

adopters.

1 I, 255-274.
forecast errors

dates of LIFO

announcement

of Accounring

Rcwurch.

l(,Y-194.

Bouwman,

P. Frishkoff

analysts make

investment

and P. Frishkoff.
decisions!

screening decision.

lYX7. How

do

a process model of the

Accounting.

Orqmizurions

LsoC~icfy. 12. l-30.


G.E.P.

and

Francisco.

CA).

Branson,

B.C.

prediction

1076. 7yrrte Serks

2nd edition

Anrrl,vsi.v:

(Holden-Day.

San

of

D.P.

Pagach.

earnings

Working

IV%.

per

paper

Examination

share

(North

and

forecast

Carolina

origin

State

Uni-

versity).
Brennan,

M.J.

and P.J. Hughes,

supply of information,
Brooks.

L.D.

and

Journul

D.A.

IYYI.

of

Journd

of Fimmce.
L.D..

lY76.

Further

evi-

income.

31. 1359-1373.

L.D.

on Accounting
Brown.

L.D.

earnings
liefs.

all forecasts
of Forecusr-

Griffin

(Editors).

IY8.1. Special

Issue

and Forecasting.

and

J.

Han.

announcements

Accounting
L.D.

and M.S.

Review.

and K.J.

Kim.

Journul of Forecasting.
2.
lYY2. The impact of annual

on convcrgencc

of analysts

be-

of

The
33.

of

evidence

l-16.

lY7Ya.

of

superiority

expectations:

Finunw.

IJnivariate

time

earnings per share: a

Accounting

Timely

aggrcgatc

analyst

Rozeff,

of i2crounling

of interim

reports

terly earnings.
Brown,

L.D.

Rrsearch.

R~.wrtdt.

labor

Accounring

and

17.

.wrtrdt.

between

value

of future

quar-

1.
can forecast

Munagettwn/,
1087,

risk-adjusted

6. 31-34.

The

effect

superiority

of
and

stock returns and

Conternporary

Accounting

Rc-

3. 61-75.
L.D..

jcwski.

P.A.

IYX7a.

univariatc

P.A.

lYX7b.

the markets

ttrg crtd
Brown.

Griffin,

An

to

earn-

Y, 6147.

and M.E.

Zmi-

proxies

for

of Auvutl/-

Journul

Y. 159-193.
and S.J. Schwager,

interpretation

in forecasting

L.D..

ti.D.

Strong-term
counting

of

earnings.

1087~.

financial

Journal

Richardson

efficiency

and C.A.

on the Toronto

analyst

of Accounting
Trzcinka.

IYYI.

stock exchange:

of analyst price forecasts,

Conkmporury

an
AC,-

Rc.wurc~h, 7. 32%5th.

P.. G. Foster and E. Noreen.


Ywr

Eurning.s

(American

Forecrrsrs

Accounting

K.C.

Association,

and L.H.

Lang.

analysts:

evidence

pessimism

. .lournul

of Accounting

C.S.A..

Non-linearity
earnings

W.S.

19X5. Swuri/y

und

individual
Chcny.

quarterly

of alternative

of earnings.

Zmi-

relative

. 25. -1%67,

R~scwrch

Brown.

Hagerman

Richardson

information

Brown.

and M.E.

superiority

and Economics.

R.L.

evaluation

Economic.s.
G.D.

Hagerman

analyst

models in forecasting

expectation

L.D..

An

R.L.

ofAccounfing

L.D..

jewaki.

Griffin.
Security

timc-series

ings. Journd

Butler.

I.

54. X-59

analysts forecast

earnings,

revision,

predictive

1980, Analysts

Zmijcwski,

strikca on security

unexpected

forecasts

of Iortfolio

cxpecta-

forecast

The

Rrvirw,

Rozcff,

M.E.

Adaptive

17, 311-X

1970,.

for improving

Journul

L.D.

lY7Yh.

and analyst

Rozeff.

and M.S.

accurately!.

the

Cupid

Sarasota.

Anu/y.sl
Market

FL).

IYYI. The forecast accuracy of


of systematic

Hopwood

and specification

optimism

Rwwch.

and J.C.

McKcown.

problems

response rcgreshion models.

and

20, I SO- 156.


lYY2.

in unexpected

AccowfitzgKcview.

67. 57%5Y8.
S.K.

and D.C.

Jetcr.

audit opinions

on earnings

oj Accounting

und

A.A..

ing research.

lYY2, The

Economics.

of

of qualified
Jourflnl

IS. 22Y-247.

1987, On cross-sectional
Journd

effects

response coefficients.

Axounting

analysis in accountund

Econotnic.~.

Y(

231-258.
Coggin, T.D. and J.E. Hunter. 198.3. Problems in measuring
the quality of investment information: the perils of the information

67, X62-875.
IYYI.

between

68. 668-W).

accounting

models.

L..D. and M.S.

Christie.

and P.A.

cxpecta-

29, 382-1X5.

association

1078.

Rozeff,

Journal

timeseries

Brown.

Choi.

Forecast selection when


recent. Internationd
Journul

7, 34Y-356.

Brown.

Brown.

prices and the


10. 166%16YI.

of accounting

1901.

are not equally


ing,

Stock

Finance.

Buckmaxter.

dence of the time series properties


Brown.

L.D.

tions,

Mulri-

and

superiority,

Jenkins,

Control,

M.S.

model,

examination

and G.M.

Fotwasting

of

Journal

and

superiority

M.,

financial

Box.

that stock

of current

near the earnings

untl

Brown.

Brown.

IYXY, Firm

Rozeff.

series models of quarterly

Brown.

27. I-38.

IYYO. Evidence

and stock price behavior


26.

or risk prc-

13. 305-330.

Journd

Biddle.

earnings.

1993. The

as measures

on the association

Post-earnings-

Resmrch.

Thomas,

forecast

of Rochester.

lY8Y.

price

of Accout~tinl:

and J.K.

markets

prices do not fully reflect the implications


ings

L.D.

proposed

earnings

Research,

RwiCw.

M.S.

from earnings.

Brown.

Bcnston.

Kim.

Accounfing
and

forecasts

Brown.

for market

disclosures by small firms and positive abnor-

L.D.

Journd

IYYI.

Stock prices and the dissemination of


analysts recommend~ltiona,
Journal of Busitwss. 64. 3Y3416.

Bcncish.

and K.J.

mal returns.
Brown.

proxies

of Accounting

17Y-l8Y.

of security

.Iorrrtlcrl of Accourr/itqy

Journul,

2. 3-2X.

Lambert

content

M.D..

and

of security

Ecotwrnic.s.

detcrmincs

Anulysts

Lambert

content

trrd

Ben&h.

lY7X. What

Financiul

R.A.

information
cour1ring

Beaver.

L.D.

nonearnings

analyst

earnings
Beaver.

as better

. Journul

Brown.

and K.S.

time-series

pectation

forecasts
tions

33.

coefficient.

Finuncd

Analysts

Jourturl,

39.

2S-

L. D. Brown

I Earnings forecasting research

Collins.
D.W. and L. DeAngelo.
1990. Accounting
information
and corporate
governance:
market and analyst
reactions
to earnings of firms engaged in proxy contests.
Journal of Accounting and Economics. 13. 213-247.
Collins. D.W. and S.P. Kothari,
1989. An analysis of intertemporal
and cross-sectional
determinants
of earnings response
of
Accounting
and
coefficients,
Journal
Economics, 11, 143-181,
Collins. D.W., S.P. Kothari and J.D. Rayburn,
1987, Firm
size and the information
content of prices with respect to
earnings.
Journal of Accounting and Economics, 9, 1 I l138.
Collins,
W.A. and W.S. Hopwood,
1980, A multivariate
analysis of annual earnings forecasts generated
from quarterly forecasts
of financial
analysts and univariate
timeseries models. Journal of Accounting Research. 18. 390406.
Collins,
W.A., W.S. Hopwood
and J.C. McKeown,
1984,
The predictability
of interim earnings
over alternative
quarters,
Journal of Accounting Research. 22. 467-479.
Conroy,
R. and R. Harris,
1987. Consensus
forecasts
of
corporate
earnings:
analysts
forecasts
and time series
methods,
Management Science, 33, 725-738.
Copeland,
T. and D. Maycrs, 1982. The Value Line enigma
(lY6551978):
a case study of performance
evaluation
issues. Journal of Financial Economics, 10, 289-322.
Cornell,
B. and W.R. Landsman,
1989, Security
price response to quarterly
earnings announcements
and analysts
forecast revisions.
Accounting Review, 64, 680-692.
Das, S. and B. Lev. IYYI. Evidence on the nonlinearity
of
the returns-earnings
relation,
Working paper (University
of California,
Bcrkelcy).
DeBondt.
W.F., 1993, Betting on trends: intuitive forecasts
of financial risk and return, International Journul of Forecasting. 9, 3555371.
DeBondt.
W.F. and R.M. Thaler,
1990. Do security analysts overreact?.
American Economic Review, 80. 52-57.
Dimson, E. and P. Marsh, 1984, An analysis of brokers and
analysts unpublished
forecasts of UK stock returns, Journul of Finance. 39. 1257-1292.
Dorfman.
J.R., 19Yl. Analysts
devote more time to selling
as firms keep scorecard on performance,
Wull Street Journal. 29 October,
Cl-2.
Dugar, A. and S. Nathan,
1992. The effect of investment
banking relationships
on financial analysts earnings forecasts and investment
recommendations.
Working paper
(Michigan
State University).
Easton. P.D. and T.S. Harris. 1991, Earnings as an explanatory variable for returns, Journal of Accounting Research,
2Y, 19-36.
Easton.
P.D. and M.E. Zmijewski,
1989, Cross-sectional
variation
in the stock market response to accounting
earnings
announcements.
Journal
of Accounting
und
Economics, 11, 117-141.
Easton,
P.D.. T.S. Harris and J.A. Ohlson,
1992, Aggregate accounting
earnings
can explain
most of security
returns:
the case of long event windows,
Journal of
Accounting and Economics, 15. llY-142.
Elgers.
P. and D. Murray,
1992, The relative and complementary
performance
of analyst
and security-pricebased measures of expected earnings,
Journal of Accounting and Economics, 15, 303-316.

317

Elliott, J.A. and W.H. Shaw, 1988. Write-offs


as accounting
procedures
to manage perceptions,
Journal of Accounting
Resettrch. 26, Yl-119.
Elton, E., M. Gruber and M. Gultekin.
1981, Expectations
and share prices. Management Science, 27, Y75-987.
Elton,
E., M. Gruber
and S. Grossman,
1986. Discrete
expectational
data and portfolio performance,
Journul of
Finance, 3. 609-712.
capital markets:
a review of
Fama, E.F..
1970, Efficient
theory and empirical work, Journal of Finance. 25. 383417.
Fildes, R. and K. Lam, 1990, Efficient
use of information
by financial analysts in the formation
of short term earnings forecasts,
Working
paper
(Manchester
Business
School).
Flavin, M.. 1981, The adjustment
of consumption
to changing expectations
about future income, Journul of folifical
Economy. 8Y, 974-1009.
Foster,
G., 1977. Quarterly
accounting
data: time-series
properties
and predictive-ability
results,
Accounting Revjiew. 52. l-21.
Foster, G., 1986. Financial Statement Analysis, 2nd edition
(Prentice-Hall,
Englewood
Cliffs. NJ).
Foster.
G., C. Olsen and T. Shevlin.
lYX4, Earnings
releases, anomalies.
and the behavior
of security returns,
Accounting Review. 59, 574-603.
Francis. J. and D. Philbrick,
1992, The association
between
Value
Line earnings
forecasts
and Value Line recommendations,
Working paper (University
of Chicago).
Frecka,
T.J. and W.S. Hopwood,
1983. The effects of
outliers on the cross-sectional
distributional
properties
of
financial ratios, Accounting Review. 58, 115128.
Freeman,
R.N., 1987, The association
between accounting
earnings
and security returns for large and small firms,
Journttl of Accounting tmd Economics. 9, 1955228.
Freeman,
R.N. and S.Y. Tse, 198Y, The multiperiod
information
content
of accounting
earnings:
confirmations
and contradictions
of previous earnings reports,
Journal
of Accounting Reseurch. 27. 4Y-84.
Freeman.
R.N. and S.Y. Tse, lYY2, A nonlinear
model of
security price responses
to unexpected
earnings,
Journal
of Accounting Research. 30. 1X5-209.
Freeman,
R.N..
J.A.
Ohlson,
and S.H. Penman.
19X2,
Book rate-of-return
and prediction
of earnings changes:
an empirical
investigation,
Journal of Accounting Research, 20. 639-653.
Fried, D. and D. Givoly, 1982, Financial
analysts forecasts
of earnings:
a better surrogate
for market expectations,
Journul of Accounting and Economics, 4, 85-107.
Gilovich,
T., R. Vallone and A. Tversky,
1985, The hot
hand in basketball:
on the misperception
of random scquences,
Cognitive Psychology,
17, 295-314.
Givoly,
D.,
1985,
The
Formation
of
Earnings
Expectations.
Accounting Review, 3, 372-386.
Givoly. D. and J. Lakonishok,
1979, The information
content of financial analysts forecasts of earnings:
some evidence on semi-strong
inefficiency,
Journal of Accounting
und Economics. 3, 1655185.
Govindarajan,
V., 1980. The objectives
of financial statcmerits: an empirical
study of the use of cash flow and
earnings by security analysts,
Accounting, Orguniztttions
and Society. 5, 383-392.

Graham.

B..

Irincipks

D. Dodd and S. Cottlc.


rmd Techniques.

lY62. Srcuriry Arw!,sis:


4th edition (McGraw-Hill,
Ncu

York, NY).
Grcig.
A.C..
1002. Fundamental
analysis and subsequent
stock returns.
Journd
of Accounring
cmd Economic.v.
15.

4 I3-J-12.

11unce. 5. 330-368.

Hopwood.
W.S.. J.C. McKeown and P. Newbold,
19X2. The
additional
information
content
of quarterly
earnings
rcporth: intertemporal
disaggregation.
Journd
of Accow~ting Re.wcrrc~h. 20. 343-339.
between
Hughes.
J.S. and W.E. Ricks, lY87. Associations
forecast
errors
and excess
returns
near
to earnings
announcements,
Accounting
Ret+w.
62. 15X-175.
Hunter. J.E. and T.D. Coggin. 1YXX.Analyst judgment:
the
efticicnt market hypothesis versus a psychological
theory of
Orpnizuiiond
Behavior
und Human
human judgment.
lIeci.tion
Processes. 42, 2X4-302.
Imhoff. E.A. and G.J. Loho. lYY2, The effect of cx ante
carninga
uncertainty
on earnings
response
coefficients,
Rrvie~~,

67. 427-439.

Joy. 0..
R. Litzcnherger
and R. McEnally.
1977. The
adjustment
of stock prices to announcements
of unanticipated changes in quarterly
earnings.
Journal of Accounring Rrseurch,
1.5. 207-225.
Kant.
A.. Y.K. Lee and A. Marcus.
1083. Earnings
and
dividend announcements:
is there a corroboration
effect?.
Jourd
of Finance. 30, lOYl-1OYY.
Kendall.
C.S. and P. Zarowin.
1990. Time-series
models of
earnings
persistence
and earnings
reannual
earnings,
sponse
coefficients.
Working
paper
(New York University).
use of maniiKim. K. and D. Schroeder.
IYYO. Analysts
gerial bonus incentives in forecasting
earnings.
Journul of
A~ounfing
und Economics.
13, 3-23.
Kim. 0. and R.E. Vcrrecchia.
1991, Trading
volume and
price reactions
to public announcements.
Journd
of Acwunring
Reseurch, 20, 302-321
Klein, A.. IYYO. A direct test of the cognitive bias theory of
share
price
reversals.
Journul
of
Accounting
and
Economics.

Kormcndi.
earnings

specifications

Accounting

und

and alternative

EconomicA.

deflators.

Journal

of

15, 173-202.

Kothari.
S.P. and R.G. Sloan, 1992. Information
in prices
about future earnings:
implications
for earnings response
cocfiicients,
Jownal
of Accounfing
cd
Economics,
143171.

Gritfin.
P.A.. lY77. The timeseries
behavior
of quarterly
earnings:
preliminary
evidence,
Jownd
of Accmrrriing
Rrseirrc~h
. 15. 7 l-83.
Gucrard.
Jr. J.B..
IYXY. Combining
time-series
model
forccasts
and analysts
forccnsts
for superior
forecasts
of annual earnings.
Fincmcid
Ancr~yvts Journtrl.
35. hY71.
Hawkins.
E.. S.C. Chamberlain
and W.E. Daniel.
IYX4.
Earnings
expectations
and security
prices.
I:incl,~c,itrl
Anulvs/s Jourrrd.
74. 24-29,
30-38. 74.
Holthausen.
R.W. and D.F. Larcker.
IYYZ, The prediction
of stock rcturna
using financial
statement
information.
Jolrnicrl of Accounting
and Economics.
15. 373-41 I.
Hopwood,
W.S. and J.C. McKeown.
1YYO. Evidence
on
surrogates
for earnings cxpcctations
within a capital marJournd
of Accounting,
Auditing
cmtl Fiket context.

Accounting

change

13. ISS-166.

R. and R. Lipe. lY87, Earnings


innovations.
pcrsistcnce.
and stock returns.
Journul of Bu.riness, 60. 323-346.
regressions
in the preaKothari.
S.P., lYY2. Price-earnings
cnce of prices lading
earnings:
earnings
level versus

Kross. W.. B. Ro and D. Schroeder,


1990, Earnings
cxpectationc: the analysts
information
advantage,
Accounting
Rrtiew.
6.51 461-476.
Lee. C.W.J. and C. Chcn, IYYO. Structural
changes and the
forecasting
of quarterly
accounting
earnings in the utility
industry.
Journd
of Accounfing
und Economic.s.
13. Y3122.
Lcv. B., IYXY. On the usefulness of earnings and earnings
research:
lessons and directions
from two decades of empirical
research.
Journd
of Accounting
Resrurch.
27.
15%201.

Lev. B. and S. Sunder.


lY79. Methodological
issues in the
use of financial
ratios.
Journul
of Accounting
rmd
Ecollomic~s.

I,

1X7-2 10.

Lcv. B. and S. R. Thiagarajan,


lY91, Fundamental
information
analysis.
Working paper (University
of California. Berkeley).
Lev, B. and T. Sougiannis.
IYY2. The capitalization,
amortiLation.
and value-relevance
of R&D.
Working
paper
(University
of California,
Bcrkelcy).
.ibhy. R. and R.K. Blashficld.
1978. Performance
of a
composite
as a function of the number of judges. Orpnizut~onul

Behavior

und Humun

Performance.

21,

l21-12Y.

in. H.W. and M. McNichols.


1YYI, Underwriting
relationships and analysts
earnings
forecasts.
Working
paper
(Stanford
University).
.ipe. R.. lYY0, The relation
hetwcen
stock returns and
accounting
earnings
given alternative
information.
Acwrtnring

Reviebv,

65. 49-7

Lipc. R. and R. Kormendi,


1993. The implications
of the
higher-order
properties
of annual
earnings
for security
valuation,
in: S. A. Butler (Editor).
Earnings
Qudity
(University
of Oklahoma,
Center for Economic
and Management
Research),
142-168.
Little. I.M.D.. 1Yh2, Higgledy
piggledy growth. Blrlletin
of
the Oxford
342.

Institute

of Economics

and Stutisrics.

24. MY-

Loho. G.J.. 1YY1. Alternative


methods of combining sccurity analysts
and statistical
forecasts
of annual corporate
earnings,
Internutionnl
Journal
of Forecusting,
7. S7-63.
Lorck.
K.S..
lY79, Predicting
annual
net earnings
with
quarterly
earnings
time-series
models.
Journul
of Accounting
Resecrrch. 17. 190-204.
Lys. 7. and S. Sohn. IYYO, The association
between rcvisions of financial analysts earnings forecasts and sccurityprice changes.
Journd
o,f Accounting
und Economic.s,
13.
34

l-363.

Mear. R. and M. Firth. 1987. Cue usage and self-insight of


financial analysts,
Accounfing
Review. 62, 176-182.
Mendenhall.
R.R.. 1991. Evidence
on the possible underweighting of earnings-related
information,
Jourrlul of Accounting
Reserrrch ~2Y. 170- 179.
Mendenhall,
R.R. and W.D. Nichols, lY88. Bad news and
differential
market reactions to annouhcements
of carlierquarters
versus fourth-quarter
earnings,
Journal
of Accounting

Resewch,

26. 63-86.

L. D. Brown

I Earnings forecasting

Miller.
M. and M. Modigliani.
1961, Dividend
policy.
growth, and the valuation of shares, Journal of Business.
34. 41 l-433.
Miller.
M. and K. Rock.
198.5. Dividend
policy under
asymmetric
information.
Journal of Finance. 40. 10311052.

Morse.
D., J. Stcphan
and E. Stice,
1991, Earnings
announcements
and the convergence
(or divergence)
of
beliefs, Accounting Review, 66, 376-388.
Muth, J.F.. 1961. Rational
expectations
and the theory of
price movements.
Econometrica,
29, 315-335.
Newbold,
P., J.K. Zumwalt and S. Kannan,
lY87. Combining forecasts to improve earnings per share prediction:
an
examination
of electric utilities.
International
Journal of
Forecasting,

3. 220-238.

OBrien,
P.. 1088. Analysts
forecasts as earnings expcctations, Journal of Accounting and Economics, 10. 53-83.
OBrien.
P.. 19Y0, Forecast
accuracy of individual analysts
in nine industries.
Journal of Accounting Research. 28,
286-304.

OBrien.
P. and R. Bhushan.
IYYO. Analyst
following and
institutional
ownership.
Journal of Accounting Research 1
2x. 55-76.
Ohlson,

J.A.. 1991a. The theory of value and earnings and


an introduction
to the Ball-Brown
Analysis.
Confernporury Accounting Research, 8, I-19.
Ohlson, J.A., lY91h. Earnings,
book values, and dividends
in security
valuation.
Working
paper (Columbia
University).
Ohlson.
J.A. and S.H. Penman,
1092, Disaggregated
accounting data as explanatory
variables for returns.
Journcrl of Accounting. Auditing and Finance. 7. 553-573.
Ohlson. J.A. and P.K. Shroff, 1992, Changes
versus levels
in earnings
as explanatory
variables
for returns:
some
theoretical
considerations,
Journal
of Accounting
Research.

30. 210-226.

J. A..
accounting

1990. The information


numbers
as earnings

Ou.

Accounting

Ou. J.A.
analysis

Research.

content of nonearnings
predictors,
Journal of

28. 144-163.

S.H. Penman.
1989a, Financial
statement
the prediction
of stock returns,
Journal of
Accounting and Economics, 11, 295-329.
Ou, J.A. and S.H. Penman.
lY89b. Accounting
measurement. price-earnings
ratio, and the information
content of
security prices, Journal of Accounting Research. 27. 11 I1.52.
Palepu. K.. 1988, Discussion
of bad news and differential
market reactions to announcements
of earlier-quarters
versus fourth-quarter
earnings.
Journal of Accounting
Research,

and
and

26, X7-90.

Penman,
S.H..
lY92a. The articulation
of price-earnings
ratios and market
to book ratios and the evaluation
of
growth.
Working paper (University
of California,
Bcrkeley).
Penman,
S.H., 1992b, Financial
statement
information
and
the pricing of earnings changes,
Accounting Review. 67.
563-577.

Penman.

319

research

Ramakrishnan,
R. and
permanent,
transitory
of reported
earnings.
sity).
Ramakrishnan,
R. and J.
the past: market value,
valuation
and sufficient
Journal
of Accounting,

J. Thomas,
19Yl. Valuation
of
and price-irrelevant
components
Working paper (Columbia
UniverThomas, 1992, What matters from
book value, or earnings? earnings
statistics for prior information,
Auditing

and

Finance.

7,

423%

464.

Rendleman.
R.J.. C.P. Jones and H.A. Latani,
lY87, Further insight into the standardized
unexpected
earnings
anomaly:
size and serial correlation
effects.
Financial
Review,

22,

131-144.

Salamon,
G. and E. Smith. lY77. Additional
evidence on
the time series properties
of reported
earnings per share:
Comment.
Journal of Finance, 32, 17YS-1801.
on analysts forecasts.
Schipper.
K.. IYYl, Commentary
Accounting

Horizons.

5, 105-121.

Shores,
D.. 1900. The association
between
interim
information
and security
returns
surrounding
earnings
announcements.
Journal
of Accounting
Research.
28,
164-181.

Shroff.
P.K.. lYY2. The variability
of earnings
and nonearnings
information,
Working paper (Ohio State University).
Siconolfi,
M.. lYY2. At Morgan
Stanley,
analysts
were
urged to soften harsh views. Wall Street Journal. 14 July
1902. Al and A6.
Sinha. P., L.D. Brown. and S. Das, 1993. Forecast
accuracy of individual
analysts
in nine industries:
a second
look.
Working
paper (State University
of New York.
Buffalo).
Stickel. S.E., I9Y0, Predicting
individual
analyst earnings
forecasts,
Journal of Accounting Research, 28, 409-317.
Stickel,
S.E..
19Yl. Common
stock returns
surrounding
earnings forecast revisions: more puzzling evidence.
Accounting

Review.

66. 402-416.

Stickel.
S.E.,
1992, Reputation
and performance
among
security analysts,
Journul of Finance, 47, 181 I-1836.
Stickel.
S.E..
1993, Accuracy
improvement
from a consensus of updated forecasts,
International Journal of Forecusting, 9, 345-353.

Stober. T.L., 19Y2, Summary


financial statement
measures
and analysts forecasts of earnings.
Journal of Accounting
and Economics.

IS. 347-372.

Stone, F.. 1977. Information


requirements
of security analysts, in: R. Abdel-khalik
and T. Keller (Editors),
Information
Requirements
of Security Analysts (Duke University Press. Durham,
NC).
Tversky,
A. and D. Kahneman.
1984, Judgment
under
uncertainty:
heuristic
and biases,
Science, 185. 11241131.

Varian.
H.R.,
1985. Divergence
of opinion
in complete
markets:
a note. Journal of Finance, 40, 309-317.
Verrecchia.
R.E., 1980, The rapidity of price adjustments
to
information.
Journal of Accounting
and Economics.
2.
63%Y2.

S.H.,

Accounting.

lYY2c. Return
Auditing

to fundamentals.

and Finance,

Journal

of

7, 465-483.

Philbrick,
D.R. and W.E. Ricks, 1991, Using Value Line
and IBES analyst forecasts in accounting
research,
Journcrl of Accounfing Research, 29, 397-417.

Warfield, T.D. and J.J. Wild. 1992. Accounting


recognition
and the relevance
of earnings as an explanatory
variable
for returns.
Accounting Review, 67. X21-842.
Watts, R.L.,
1975, The time-series
behavior
of quarterly
earnings,
Working paper (University
of Newcastle).

Watts.

R.L..

terly

lY78.

EwwmicS.
Watts.

Systematic

earnings

and R.W.

annual

accounting

scurch,

15, 253-271,

role

returns after quar-

Jourtml

of

Firwrlcitrl

6. 127-1.50.

R.L.

Wicdman.

abnormal

announcements,

C..
of

Lcftwich,

lYY3. Proxies

loss

lY77.

Journcrl

earnings.

The

for earnings

functions.

time scriea of

c!f Accounrirtg

Working

Kc>-

expectations:
paper

the

(Cornell

University).
Wiggins.
ings

J.B..

Working
Winkler.

contribute

paper
R.L.

forecasts.
1.57.

1000, Do

process

(Cornell

misrepresentations
to

drift?,

University).

and S. Makridakis.

Jorrrtd

about the earn-

post-announcement

1983, The combination

of the Royul Siurisricul Society. 2.

of

150-

Biography: Lawrence D. BROWNs


principal research interests are in the use of financial information by capital market
participants.
primarily security analysts. Dr. Brown has over
SO publications in the areas of accounting. finance and forecasting. and he has presented his research at over 35 universities. His research in the topical area of this paper has
been published in 71~~ Acwrrn~ing Revkrr~, Conremportw~

Accowttittg
Resctrrch. f~~tetxutiot~crl Journd
Op Fotwmrin~.
Jotrrmd of Accouttting Reseurch, Jourmrl of Ru.sincs.r
Foreusting.
Jowxul
c!f Fintrtwe, Journal
of f~orecusting. and
Jonrmrf of Portfolio
Matqemenr.
He is an Associate Editor
of the ft~frrmr~ionul Journul
of Forecusring
and Review o/
Qucrnrirtrrivr Fituncr cd Accounting;
he servcs on the editorial board of Conternftorury Accoudng
Rrxurch and is a
memhcr of the Exccutivc Committee of the Annual
cnce on Financial Economics and Accounting.

Confer-