Full text of DOI:10.1111/1475-679X.00040

Journal of Accounting Research
Vol. 40 No. 1 March 2002
Printed in U.S.A.
The Association between
Trading Recommendations and
Broker-Analysts’ Earnings Forecasts
M I C H A E L E A M E S ,∗
S T E V E N M . G L O V E R ,†
A N D J A N E K E N N E D Y‡
Received 9 February 1999; accepted 17 September 2001
ABSTRACT
This study examines analyst forecast errors within the context of stock re-
commendations. We predict positive forecast error (i.e., optimism) for buy
recommendations and negative forecast error (i.e., pessimism) for sell recom-
mendations. We offer two explanations for this prediction: (1) the unconscious
tendency to process information in a manner that supports one’s goal, which
we refer to as the “objectivity illusion” hypothesis, and (2) the economic in-
centive to boost trade, which we refer to as the “trade boosting” hypothesis.
The pattern of analyst forecast bias we predict (i.e., optimism for buys and pes-
simism for sells) is opposite in direction to that predicted by the management
relations hypothesis—a commonly cited hypothesis for analyst forecast bias.
We find broker-analyst earnings forecast errors are significantly op-
timistic for buy recommendations and significantly pessimistic for sell
∗
Santa Clara University; † Brigham Young University and PricewaterhouseCoopers;
‡
University of Washington. We are grateful for the comments of Sudipta Basu, Robert Bowen,
Dave Burgstahler, James Jiambalvo, Laureen Maines, Dawn Matsumoto, James Meyers, Eric
Noreen, Kay Stice and participants at the 1999 Universities of British Columbia, Oregon, and
Washington (UBCOW) Conference, the 1999 European Accounting Conference, the 1999
Western Regional Meeting of the AAA, the 1999 AAA Annual Meeting, the 2000 Globalization
Conference, and accounting workshops at Brigham Young University, Santa Clara University,
and the University of Utah. We thank the PricewaterhouseCoopers Foundation, the Mary and
Ellis Fund at Brigham Young University and the Accounting Development Funds at the Uni-
versity of Washington and Santa Clara University for financial support. We thank Ralph Arvizu
for his excellent research assistance.
8
5
Copyright ꢀC , University of Chicago on behalf of the Institute of Professional Accounting, 2002

86
M. EAMES, S. M. GLOVER, AND J. KENNEDY
recommendations, consistent with the objectivity illusion and trade boosting
hypotheses. Our study indicates that the pattern of results reported in prior re-
search (i.e., increasingly optimistic earnings forecasts as the stock recommen-
dation becomes less favorable) is likely driven by a correlated omitted variable,
actual earnings. Results of an analysis to distinguish between trade boosting
and objectivity illusion appear more consistent with the objectivity illusion.
1
. Introduction
This paper investigates the influence of two factors on the accuracy of
analyst earnings forecasts: the incentive to boost trading, and motivated
reasoning—the tendency to process information in a manner that supports
one’s goal. We predict these factors lead to forecasts that are greater than
actual earnings (i.e., positive forecast errors or optimistic forecasts) in as-
sociation with buy recommendations and forecasts that are less than actual
earnings (i.e., negative forecast errors or pessimistic forecasts) in association
with sell recommendations.
The pattern of forecast error we predict is opposite to that predicted by
the “management relations” hypothesis—a commonly cited hypothesis used
to explain analyst forecast bias (Francis and Philbrick [1993]). The man-
agement relations hypothesis suggests that analysts intentionally issue opti-
mistic earnings forecasts in order to curry favor with management, and this
forecast optimism is particularly pronounced in association with sell recom-
mendations because analysts wish to counter negative effects on the analyst-
management relation generated by sell recommendations. The empirical
evidence supporting the management relations hypothesis (i.e., increasingly
optimistic forecasts as the stock recommendation becomes less favorable)
appears to be driven by a correlated but previously omitted variable—actual
earnings. It is important to include earnings in the analysis of forecast errors
and recommendations because both forecast errors and recommendations
are significantly correlated with earnings. When we control for the relation
between earnings and forecast errors we obtain results strikingly different
from those reported in prior research.
Based on an extensive broker-analyst data set, we find evidence that ana-
lyst earnings forecast errors are significantly optimistic for buy recommen-
dations and significantly pessimistic for sell recommendations, consistent
with both the incentive to boost trading and a form of motivated reasoning
we label the objectivity illusion. Motivated reasoning, or the tendency to pro-
cess information in a manner that makes the desired outcome more likely,
is a well-known concept in the behavioral decision literature (e.g., Kunda
[
1990]).
2
. Background and Hypotheses
Research generally finds that analyst earnings forecasts are: (1) more ac-
curate than time-series predictions of earnings (e.g., Affleck-Graves, Davis,

TRADING RECOMMENDATIONS
87
and Mendenhall [1991], Brown, Griffin, Hagerman, and Zmijewski [1987]),
(
[
2) imperfect proxies for market expectations of earnings (Brown et al.
1987], O’Brien [1988]), (3) less accurate in predicting future earnings
than a simple model incorporating analyst forecasts and past time-series
properties of earnings (Ali et al. [1992]), and (4) optimistic on average (e.g.,
O’Brien [1988], Lys and Sohn [1990], Mendenhall [1991], Brown [1993],
Dugar and Nathan [1995], Das, Levine, and Sivaramakrishnan [1998]).
Attempts to explain this forecast optimism generally focus on analysts’
responses to institutional incentives.
2.1 MANAGEMENT RELATIONS AND TRADE BOOSTING
Francis and Philbrick [1993] hypothesize that analysts intentionally issue
optimistic earnings forecasts to curry favor with management and thereby
gain greater access to management’s private information (the “management
relations hypothesis”). Francis and Philbrick contend that such intentional
optimism is particularly pronounced in association with sell recommen-
dations because analysts wish to counter negative effects on the broker-
management relation generated by the sell recommendation. Consistent
with their hypothesis, Francis and Philbrick find that Value Line analysts’
earnings forecasts are optimistic on average and that the extent of optimism
in earnings forecasts increases as the Value Line timeliness ranking becomes
less favorable. Several subsequent studies rely on the management relations
hypothesis (e.g., Das et al. [1998], Lim [1998], Dugar and Nathan [1995])
to generate predictions of intentional analyst forecast bias.
Kim and Lustgarten [1998] suggest that broker-analysts have incentives to
bias their forecasts to both maintain favorable relations with management
and to stimulate stock trading. Their “trade boosting” hypothesis asserts that
broker-analysts’ trade boosting incentives dominate management relations
incentives, and result in more optimistically biased forecasts for buy than for
hold stocks and more pessimistically biased forecasts for sell than for hold
stocks. However, they find that broker-analysts’ earnings forecasts are more
optimistically biased for sell and hold stocks than for buy stocks and conclude
that, for broker-analysts, management relations incentives dominate trade
boosting incentives.
While the management relations hypothesis appears to be supported in
the extant literature, this hypothesis conflicts with important institutional
factors and recent academic research. Once actual earnings are announced,
forecast errors for all publicly issued forecasts are easily computed. If repu-
tation is important in the broker-analyst business, the negative reputation
effects associated with inaccurate forecasts should reduce the incentive to
1
issue intentionally biased forecasts. The management relations hypothesis
also fails to consider the fact that managers have incentives to avoid negative
1
Analyst pay is often based, in part, on analyst position on the Institutional Investor All-American
Research Team. Forecast accuracy is one of the four criteria used to determine membership on
the team.

88
M. EAMES, S. M. GLOVER, AND J. KENNEDY
earnings surprises (i.e., managers seek to meet or beat analyst forecasts)
and thus may be displeased with optimistic earnings forecasts. Burgstahler
and Eames [2000] present evidence that managers alter both earnings and
forecasts to avoid negative earnings surprises. Matsumoto [2001] argues
that managers prefer relatively pessimistic forecasts for reasons including
the market’s reaction to negative earnings surprises, and presents results
consistent with this position. Numerous articles in the popular financial
press (e.g., Ip [1997a, b], McGee [1997]) reiterate that optimistic earnings
forecasts are not an effective way to curry management favor.
2
.2 MOTIVATED REASONING, FORECAST ERROR,
AND RECOMMENDATIONS
Individuals often make judgments in an environment in which they are
motivated to reach a particular conclusion, such as to support a project
or oppose a merger with another company. Normatively, an individual
should consider all relevant information in an unbiased manner and make
an accurate judgment without being influenced by motives to reach a
particular conclusion. In reality, it is difficult to set aside existing preferences
when processing information. Prior research suggests that motivation for a
particular conclusion (i.e., directional motivation) may bias the judgment
process. For example, individuals tend to evaluate information that is
consistent with the desired conclusion less critically than contradictory in-
formation (Festinger [1957], Ditto and Lopez [1992], Snyder and Swann
[
1978]). Directionally motivated individuals adopt information-processing
strategies most likely to yield the desired conclusion. This phenomenon has
been labeled motivated reasoning (Kunda [1990]).
Research on motivated reasoning suggests that, because individuals want
to construct rational justifications for conclusions they wish to draw, they
search for relevant information and construct beliefs that logically support
their desired conclusion (Kunda [1990], and Boiney, Kennedy, and Nye
[
1997]). If they succeed in finding such information, they draw the desired
conclusion while maintaining an illusion of objectivity. The objectivity of
this process is considered illusory because individuals do not realize that the
process is biased by their goals. They do not realize that they are accessing
only a subset of their relevant knowledge and that they would probably
access different beliefs and rules in the presence of different directional
goals. They are unaware that they constructed their decision process to
make the desired conclusion more likely.
Security analysis provides the basis for selecting stocks to buy and sell. The
ultimate judgment within a sell-side financial analyst’s formal report evalu-
ating a firm’s securities is a “buy, sell, or hold” recommendation. Givoly and
Lakonishok [1984] contend that “next to stock recommendations, earn-
ings forecasts are perhaps the most notable output of financial analysts,”
but forecasting earnings appears subordinate to the task of issuing stock
recommendations and writing reports supporting these recommendations.
Although a stock recommendation may follow a carefully prepared earnings

TRADING RECOMMENDATIONS
89
forecast, there is evidence that earnings forecasts play a relatively minor
role as input to analysts’ recommendation decisions (Balog [1991], Biggs
[
1984]).
We posit that, in generating earnings forecasts, analysts tend to process
information in a manner that biases forecasts in the direction that supports
their stock recommendation. Therefore, we predict that analyst earnings
forecasts will be optimistic for buy recommendations and pessimistic for
sell recommendations. We do not expect hold recommendations to result
in significant bias. We refer to the behavior leading to this pattern of bias as
the objectivity illusion, consistent with the idea that while analysts may perceive
their earnings forecasting behavior as objective, this objectivity is illusory.
Our hypotheses focus on analyst forecast errors conditioned on stock
recommendations. The null hypothesis is:
H1: Analyst earnings forecast errors are independent of stock recom-
mendations.
Our discussion of prior research on intentional biases (i.e., management re-
lations and trade boosting) and the objectivity illusion suggest the following
2
two competing alternative hypotheses:
H1a: Analysts’ earnings forecast optimism increases as the stock recom-
mendation becomes less favorable (consistent with the manage-
ment relations hypothesis).
H1b: Analyst earnings forecasts are optimistic for favorable stock recom-
mendations and pessimistic for unfavorable stock recommenda-
tions (consistent with the objectivity illusion and trade boosting
hypotheses).
3
. Data
We obtain individual analyst recommendations and actual and forecasted
earnings-per-share (EPS) values for the years 1988 to 1996 from the Zacks
Investment Research database.Recommendations in the database are ex-
pressed on a scale of 1.0 to 6.0 where: 1.0 = strong buy, 1.1 to 2.0 = moderate
buy, 2.1 to 3.0 = hold, 3.1 to 4.0 = moderate sell, 4.1 to 5.0 = strong sell, and
6.0 denotes that a recommendation is not available for the particular analyst
and time period. Zacks adds records to the database when an analyst initi-
ates coverage of a firm and issues a first recommendation, when an analyst
changes a previous recommendation, and when an analyst issues a recom-
mendation and the previously recorded recommendation in the database
is at least a year old.
2
None of the alternative hypotheses specifically mentions hold recommendations. However,
we anticipate that earnings forecast errors associated with such intermediate recommendations
will lie between observations for more extreme recommendations.

90
M. EAMES, S. M. GLOVER, AND J. KENNEDY
The Zacks earnings forecast and earnings surprise databases include ana-
lyst forecasts of EPS and actual EPS, in conformance with Zacks’ proprietary
definition of operating earnings-per-share before extraordinary and non-
recurring items. For each fiscal period, the earnings forecast database in-
cludes numerous forecasts released at various dates by a number of analysts.
Each annual EPS forecast record in the Zacks earnings forecast database
represents a forecast associated with the initiation of coverage, a change
in an analyst’s forecast for a fiscal period, or the reiteration of a previous
forecast that is at least 120 days old.
We scale total earnings and total forecast values by beginning of the year
3
market value of equity. Scaling by beginning of the year market value of equ-
ity provides for consistency with prior research (e.g., Francis and Philbrick
[
1993] and Kim and Lustgarten [1998]) and controls for firm size. We obtain
annual forecasted and actual earnings by multiplying Zacks annual EPS
forecasts and Zacks actual EPS values by the number of common shares
used to calculate EPS (Compustat annual data item # 54 × Compustat item
#
27). We obtain beginning of the year market value of equity by multiplying
common shares outstanding (Compustat item # 25) and price per share
Compustat item # 199), both obtained for the end of the prior fiscal year.
(
We ensure comparability of forecasted and actual earnings by obtaining
both of these numbers from Zacks.
We use only December fiscal year-end firms for the years 1988 to 1996,
and eliminate financial (SIC codes 4400 to 5000) and utility (SIC codes
6000 to 6500) firms. We match each individual analyst annual earning fore-
cast recorded in January, February, and March with the contemporaneously
reported recommendation by the same analyst if this is available. If there
is no such contemporaneous recommendation then we match the fore-
cast to the most recent preceding recommendation by the same analyst.4
For example, suppose an analyst issues a buy recommendation in August,
a sell recommendation in the following December, and a month later in
January issues an earnings forecast without an accompanying recommen-
dation. Then the January forecast is matched to the December recommen-
dation. However, had the analyst issued a recommendation on the same
date as the January forecast then the contemporaneous recommendation,
rather than the December recommendation, would have been matched to
the forecast. We calculate scaled forecast errors, defined as the difference
between forecasted and realized earnings divided by beginning of the year
market value of equity, for 62,908 observations. We further limit the data
to the 35,482 observations of annual earnings forecasts for the fiscal year
3
We obtain essentially equivalent results when we scale by beginning of the year book value
of equity and end of the year market value of equity.
4
We limit the sample to December year-end firms and January through March forecasts
for consistency with prior research and to obtain more powerful tests, since annual earnings
forecast accuracy is known to improve over time within the fiscal year (Burgstahler and Eames
[2000]).

TRADING RECOMMENDATIONS
91
ending from nine to eleven months after the date of forecast release (i.e.,
we eliminate forecasts issued after fiscal year-end and for years beyond the
current fiscal year). For all analyses requiring analyst recommendations,
we eliminate the 870 observations with unavailable recommendations (i.e.,
5
Zacks recommendation = 6) and employ a final sample size of 34,612.
Sample distributions for observations by month, year and recommenda-
tion are presented in table 1 (panels A, B, and C). Panels A and B illus-
trate that sample observations are relatively evenly distributed between the
months of the first quarter and approximately three quarters of the total
observations come in the last five years of the nine years included in the
sample. Panel C indicates that strong and moderate buy recommendations
are much more frequent than strong and moderate sells.
Table 1, panel D presents distribution statistics for calendar days between
earnings forecasts and the most recently preceding or contemporaneous
recommendation, firm market value of equity, scaled earnings, and scaled
forecasterror. Only15.7%oftheearningsforecastsrepresentajointissuance
of a recommendation and forecast, all other earnings forecasts are issued in
the context of a previously reported recommendation. Approximately 75%
of our earnings forecasts follow the currently outstanding recommendation
report date by more than 30 days. These results confirm that earnings fore-
casts are typically issued in the presence of an existing recommendation
that has been outstanding for at least one month. The second and third
columns of panel D indicate that our sample is comprised of predominately
large firms with an average return on equity of 6%. The last column of panel
D reports that the mean and median scaled forecast errors are .75 and .21
percent of market value of equity respectively, consistent with the findings
of analyst forecast optimism reported in numerous prior studies.
4
. Results
4.1 EARNINGS FORECAST ERROR AND THE LEVEL OF EARNINGS
Accumulating evidence shows that forecast errors are not the same across
levels of earnings scaled by market value of equity. For example, research
has documented different forecast error patterns for positive and negative
earnings (Brown [1999], Butler and Saraoglu [1999], Hwang, Jan, and Basu
[
1996]). Brown [1999] suggests that forecast errors should be investigated
separately for positive and negative earnings firms because the distributions
of earnings forecast errors differ substantially. Hwang et al. [1996] iden-
tify an inverse relationship between analyst earnings forecast errors and
reported earnings. Similarly, Eames, Glover, and Stice [2001] report that
earnings forecasts exhibit greater optimism as earnings decline below the
mean earnings for all firms and greater pessimism as earnings increase above
5
We also conducted analyses on samples that deleted forecast errors greater than 20% of
market value, and on samples that excluded the largest 5 and 10 percent of scaled forecast errors
(ranked using absolute values). The results were virtually identical to the results presented here.

92
M. EAMES, S. M. GLOVER, AND J. KENNEDY
T A B L E 1
Sample Characteristics: First Quarter Forecasts of Annual Earnings, Realized Annual Earnings,
Recommendations, and Earnings Forecast Errors, and Market Value of Equity Based on Zacks
Investment Research Values for December Fiscal Year End Firms
Panel A: Sample Distribution by Month
Month
Observationsa
% of Total
January
February
March
13,391
11,369
10,722
38%
32%
30%
Total
35,482
100%
Panel B: Sample Distribution by Year
Year
Observations
% of Total
1
1
1
1
1
1
1
1
1
988
989
990
991
992
993
994
995
996
2,009
1,627
2,713
3,159
4,010
4,433
5,147
5,701
6,683
5.7%
4.6%
7.6%
8.9%
11.3%
12.5%
14.5%
16.1%
18.8%
Total
35,482
100.0%
Panel C: Sample Distribution by Recommendation
Recommendation
Observations
% of Total
Strong Buy (1)
Moderate buy (2)
Hold (3)
Moderate Sell (4)
Strong Sell (5)
Not Available (6)
11,047
7,886
13,916
960
803
870
31.1%
22.2%
39.2%
2.7%
2.3%
2.5%
Total
35,482
100.0%
Panel D: Days Between Recommendation and Earnings Forecast, Market Value of Equity,
Actual Earnings, and Earnings Forecast Error Distribution Statistics
Earnings
Number of Days
Market Value Earnings as a Forecast Errors
Recommendation
Precedes Forecast
of Equity
($ millions)
Percent of
Market Value
as a Percent of
Market Value
b
Mean
Median
109
92
29
178
0
360
91
6,434
1,797
418
5,738
6
5.93
6.11
3.80
.75
.21
Quartile 1
Quartile 3
Minimum
Maximum
Standard Deviation
a
−.73
8.56
2.02
−83.97
70.38
7.41
−68.12
93.29
5.46
119,989
12,214
Sample observations comprise forecasts matched with the contemporaneous or the most recent preced-
ing recommendation by the same analyst, provided the recommendation does not precede the forecast by
more than one year. Forecast errors are defined as forecasted earnings less actual earnings.
b
Zacks Investment Research adds new recommendation records to the database when an analyst issues a
first recommendation or changes a recommendation or when the analyst reiterates a prior recommendation
that is at least a year old.

TRADING RECOMMENDATIONS
93
the mean earnings of all firms. They conclude this relation is due, in part,
to unanticipated earnings shocks. Unexpected positive (negative) earnings
shocks result in higher (lower) unanticipated earnings and will generally
be associated with more negative (positive) forecast errors. More formally,
suppose that reported scaled earnings at time t for firm i, Eit, can be decom-
posed into a predictable component, EPit, and an unpredictable component
or earnings shock, eUit ∼ N(0, σ). Then Eit = EPit + eUit. We can model the
analysts’ scaled forecasts, Fit, of Eit as a function of the predictable com-
ponent: Fit = EPit + fit, where a non-zero mean value for the error term,
fit, is consistent with forecast bias (either intentional or unintentional). We
define forecast error as: FEit = Fit − Eit = (EPit + fit) − (EPit + eUit), or FEit =
fit − eUit. Thus, assuming no forecast bias, unpredictable earnings shocks will
result in an inverse relation between forecast error and reported earnings.6
To plot the relationship between scaled forecast error and scaled earnings
we divide our sample of 35,482 observations (including the 870 observations
with unavailable recommendations) into 20 equally-sized portfolios based
on the magnitude of scaled earnings. Figure 1 illustrates portfolio mean and
median scaled forecast errors plotted by portfolio median scaled earnings.
The inverse relationship depicted in figure 1 is consistent with results in
Eames, Glover, and Stice [2001] that, on average, firms experiencing large
positive (negative) earnings shocks are more likely to have positive (nega-
7
tive) earnings and negative (positive) forecast errors. Regressing scaled
analyst forecast errors on scaled actual earnings results in an R-squared of
48 percent and a negative slope significant at the .001 level (table 2, panel B).
If earnings are also associated with analyst recommendations, earn-
ings could be an important correlated omitted variable in prior studies
examining the relationship between forecast errors and recommendations.
Figure 2 plots mean and median scaled earnings by recommendation and
6
Because earnings is included on both the right and left-hand side of the equation, i.e.,
FE = F − E = B0 + B1E + e , there may be concern that this linear model will force an
it
it
it
it
it
algebraic inverse association between FEit and Eit. However, an algebraic association between
forecast error and earnings does not necessitate an observed inverse association. Fit is a function
of Eit,thus we do not necessarily expect a negative value for B1, and only obtain a negative value
under restrictive assumptions regarding the behavior of Fit with respect to Eit. To see this, we
∂
∂
F
differentiate both sides of the preceding equation with respect to Eit, obtain
− 1 = B1, and
E
∂
F
∂F
∂E
note that B1 is negative only if
either of these conditions a priori. For our sample
< 1, and B1 = −1 only if
= 0. There is no basis for asserting
∂
E
∂
∂
F
E
= .48 and is significantly different from
zero at the 1% level. Thus, the inverse association between forecast error and earnings is
not algebraic. Rather, a systematic inverse relationship between forecast errors and earnings
requires systematic analyst behavior. For example, the inability of analysts to forecast some
component or portion of earnings. Because we focus on scaled earnings, we assume that the
scaled earnings shock has constant variance across firms.
7
The optimistic forecast errors associated with lower earnings are consistent with the man-
agement relations hypothesis given that sell recommendations are commonly associated with
low earnings. Although the management relations hypothesis could be contributing to the ob-
served relation between forecast error and earnings, it cannot explain the pessimism observed
at higher earnings observations because the management relations hypothesis predicts only
optimism.

94
M. EAMES, S. M. GLOVER, AND J. KENNEDY
FIG. 1.—Mean and median scaled earnings forecast error by mean scaled earnings for ob-
servations of first quarter forecasts of annual earnings with a contemporaneous or recently
preceding recommendation by the same analyst. Forecasts of annual earnings and actual an-
nual earnings values are from Zacks Investment Research. Forecasts are reported in first fiscal
quarter for December year-end firms. With 35,482 observations 20 portfolios are formed based
on earnings rank. The highest earnings portfolio has 1,776 observations and all others have
1
,774 observations. We scale earnings and forecast values by beginning of the year market value
of equity and measure forecast error as forecast earnings less actual earnings. Similar results
are obtained if beginning of the year book value and ending market value of equity are used
as the scaler.
illustrates the relationship between scaled earnings and recommendations.
The depicted relationship is a monotonic increase in mean scaled earnings
8
from sell to buy recommendations.
Panel A of table 2 reports the correlation matrix for scaled earnings, scaled
forecasts, scaled forecast errors, and recommendations. The observed nega-
tive significant correlation between scaled forecast error and scaled earn-
ings and the negative significant correlation between scaled earnings and
recommendation (where 1 = buy and 5 = sell) are consistent with figures
1
and 2. Collectively, the figures and significant correlations suggest that
earnings is a potential explanatory variable in the relation between forecast
errors and recommendations.
A further reason to consider earnings as an explanatory variable in ana-
lyzing the relationship between forecast errors and recommendations stems
from the potential for analyst and managerial behaviors to differ for profit
and loss firms. Analyst behavior may differ as a result of such factors as the
selective non-reporting of results for anticipated loss firms (McNichols and
O’Brien [1997]), while managerial differences may be associated with such
8
All paired comparisons of mean scaled earnings by recommendation category, with the
exception of the comparison of strong and moderate buy recommendations, are significant at
the .01 level. While the sample used to plot figure 2 is essentially the same as that used to plot
figure 1, there are no negative observations of scaled earnings in figure 2 because the mean
and median scaled earnings are positive for all categories of recommendations.

TRADING RECOMMENDATIONS
95
FIG. 2.—Mean and median scaled earnings by stock recommendation for observations of
first quarter forecasts of annual earnings with a contemporaneous or recently preceding recom-
mendation by the same analyst. Actual annual earnings values and recommendations are from
Zacks Investment Research. We limit observations to forecasts reported in first fiscal quarter
for December year-end firms and scale earnings values by beginning of the year market value
of equity. We match each forecast with the most recent preceding or contemporaneous recom-
mendation by the same analyst, provided the recommendation does not precede the forecast
by more than one year. Recommendations are available for 34,612 observations. Forecast er-
rors are defined as forecasted earnings less actual earnings. Mean and median earnings levels
are plotted by recommendation with the following samples: Strong Buy = 11,047, Moderate
Buy = 7,886, Hold = 13,916, Moderate Sell = 960, Strong Sell = 803.
factors as the propensity of loss firms to substantially lower earnings with a
big bath (Brown [1999]).
4.2 EARNINGS FORECAST ERROR AND RECOMMENDATIONS
The positive correlation between scaled forecast errors and recommen-
dations reported in panel A of table 2 is consistent with the results reported
in Francis and Philbrick [1993] and Kim and Lustgarten [1998]. However,
the analysis reported in panel B of table 2 highlights the importance of
including earnings as an explanatory variable. Table 2, panel B, reports
simple and multiple regression analyses of recommendations and scaled
earnings on scaled forecast errors. We find that the sign of the coefficient
on recommendation reverses from a significant positive value (t = 6.32, p <
.
001) in the simple regression to a significant negative value (t = −4.70,
p < .001) when scaled earnings is added as an independent variable in the
multiple regression. The coefficient estimates for scaled earnings in the
9
9
Before pooling the data across years we obtained regression estimates by year. Our results
were substantially similar to those presented above. For example, in equation 3 the coefficient
on earnings was negative and significant in all years, and the coefficient on recommendation
was negative and significant in six of the nine years, and not significant in the remaining
3
years.

96
M. EAMES, S. M. GLOVER, AND J. KENNEDY
T A B L E 2
Correlations, Regression Analysis and Distribution Statistics for Recommendations and Market Value
of Equity Scaled Measures of Forecasts, Earnings, and Forecast Errors Based on Zacks Investment
a
Research Values for December Fiscal Year End Firms
Panel A: Correlation Matrix (Pearson (Spearman) Correlation Coefficients Below
(
Above) Diagonal, n = 34,612)
Scaled Forecast
Error
Recommendation Scaled Earnings
Forecast
∗
∗∗
∗∗∗
∗∗∗
∗∗∗
∗∗∗
∗∗∗
Scaled Forecast Error
Recommendation
Scaled Earnings
Forecast
.025
−.584
−.066
.047
∗
∗
∗
∗∗
∗∗
∗∗
.034
−.696
.056
−.072
∗∗∗
∗∗∗
−.075
−.069
.673
∗∗∗
.663
.
Panel B: Regression Analyses (n = 34,612)
Forecast Error = β0 + β1Recommendation +e
Forecast Error = β0 + β2 Earnings + e
(1)
(2)
(3)
Forecast Error = β0 + β1 Recommendation +β2 Earnings +e
β0
β1
t-value
β2
t-value
R2
Equation 1
Equation 2
Equation 3
.0034
.0378
.0400
.002
6.32
.001
.484
.485
−.511
−.512
−180.33
−180.23
−.001
−4.70
Panel C: Market Value of Equity Scaled Forecast Error Statistics
((Forecast Earnings-Actual Earnings)/(Beginning of Fiscal Period Market Value of Equity))
Unadjusted
Data Scaled
Forecast Error
Mean Earnings
Adjusted Scaled
Forecast Error
Median Earnings
Adjusted Scaled
Forecast Errorb
Central Region
Scaled Forecast
Errorc
b
Mean
Median
.75
.21
.00
−.28
.36
.00
.24
.06
Quartile 1
Quartile 3
Minimum
Maximum
Standard
Deviation
−.73
−1.45
−1.05
−.51
2.02
1.21
1.46
.96
−68.12
93.29
5.46
−74.71
82.58
4.19
−72.92
84.37
4.21
−18.75
19.37
2.08
a
Sample observations comprise forecasts matched with the contemporaneous or the most recent pre-
ceding recommendation by the same analyst, provided the recommendation does not precede the forecast
by more than one year. Forecast errors are defined as forecasted earnings less actual earnings. Recommen-
dation data are coded 1 = strong buy, 2 = moderate buy, 3 = hold, 4 = moderate sell, 5 = strong sell, and
6
= not available.
b
Mean (median) earnings adjusted forecast errors are obtained by identifying 20 equal sized portfolios of
observations based on ranks of market value of equity scaled earnings, measuring forecast error as the forecast
less actual earnings, again scaled by beginning-of-the-year market value of equity, and then subtracting the
relevant portfolio mean (median) forecast error from the observed forecast error.
c
The central region includes all observations with earnings scaled by beginning-of-the-year market value
of equity ranging from 5% to 10%.
∗
∗∗
Significant at the .001 level.
simple and multiple regressions are nearly identical and not significantly
different.
Equation 3 in panel B of table 2 assumes linearity. However, inspection
of figure 1 suggests the relation between scaled earnings level and fore-
cast error is non-linear. To control for the relation between scaled earnings
and forecast errors without assuming linearity, and to limit the impact of

TRADING RECOMMENDATIONS
97
outliers, we obtain earnings-portfolio adjusted scaled forecast errors using
1
0
the same 20 equal-sized portfolios formed to plot figure 1. We obtain mean
median) adjusted scaled forecast errors by subtracting the relevant port-
(
folio mean (median) scaled forecast error from observed scaled forecast
errors. The resulting adjusted scaled forecast errors are similar to residuals
from a regression of scaled forecast error on actual scaled earnings, however
the portfolio adjustment does not assume linearity and is not significantly
1
1
impacted by outliers.
A third method of “controlling” for the relationship between earnings
and forecast errors is to restrict our analysis to the 15,701 observations in
a “central region” of scaled earnings where visual inspection of figure 1
suggests the relation between scaled earnings and scaled forecast error is
relatively weak. We define this central region to be all observations with
1
2
earnings of 5 to 10 percent of beginning-of-year market value of equity.
We select this 5 to 10 percent region for analysis because: (1) there is a
high density of forecasts in this area (15,701 of the 34,612 observations),
(
2) dispersion in analysts forecast errors is relatively low for the region, (3)
this region includes the mean and median scaled earnings (6% of market
value of equity) and excludes extreme values of earnings, and (4) earnings
1
3
level is only moderately related to forecast error in this region. Regressing
scaled forecast error on scaled earning in the central region results in an
R-squared of only 6%, although the relation is still negative and significant
(
p < .001).
Table 2, panel C presents distribution statistics for the mean and median
earnings-portfolio-adjusted and central region forecast errors. For ease of
comparison, this panel also repeats the values for unadjusted scaled forecast
1
0
Another advantage of the portfolio adjustment technique over regression is that it lends
itself to the analysis of absolute forecast optimism and pessimism across recommendation
categories. Analysis via regression coefficients focuses only on the trend in forecast error across
recommendations.
1
1
Our results are robust to alternative methods of controlling for earnings. All reported re-
sults examining the portfolio adjusted forecast errors are similar to results obtained examining
residuals from a regression of forecast error on actual earnings. Because the plot illustrated
in figure 1 suggests the relationship between forecast error and earnings is nonlinear, we also
used a three-part piecewise regression with breaks in the earnings value at 5 and 10 percent of
market value. The results of analyzing residuals from the piecewise regressions are also simi-
lar to analysis of portfolio earnings level adjusted forecast errors. Because we believe that our
earnings-portfolio adjusted forecast errors are superior to regression residuals, we limit our
discussion to analysis of the earnings-portfolio adjusted forecast errors.
1
2
We also analyzed various central regions with ranges of actual earnings of from as small as
to 8 percent to as large as 1 to 13 percent. Results for all these ranges displayed qualitatively
7
similar relationships between forecast error and recommendation to the relationship reported
for the 5 to 10 percent range.
1
3
As an alternative measure of how extreme large (small) observations in figure 1 are, data
points can be converted to a ratio approximating PE ratios (recall that we scale by beginning
market value). In other words, firms reporting realized earnings that are 20% of beginning
market value have a PE ratio of 5.

98
M. EAMES, S. M. GLOVER, AND J. KENNEDY
errors that are reported in table 1. Both mean and median adjusted scaled
forecast errors and central region scaled forecast errors are significantly
less optimistic than unadjusted scaled forecast errors. Note that while we
replicate previous studies (e.g., Francis and Philbrick [1993]) by analyzing
unadjusted scaled forecasts errors, our primary focus is on adjusted scaled
forecast errors and observations in the central region because we argue these
two latter data sets appropriately consider the association between forecast
error and earnings.
Table 3, panel A presents mean and median unadjusted scaled forecast
errors (i.e., no control for the relation between actual earnings and fore-
cast errors) by analyst recommendation. The mean scaled forecast error
for each recommendation is significantly greater than zero, and the means
are increasingly positive as the recommendation moves from buy to sell.
The mean scaled forecast error is significantly more optimistic for strong
sell than for strong buy (t = 4.14, p = .000), for hold than for moderate
buy (t = 3.86, p = .000), and for strong sell than for moderate sell recom-
1
4
mendations (t = 1.99, p = .047). Thus, when we omit earnings from the
analysis, we are able to replicate prior findings of increasing analyst fore-
cast optimism as recommendations become less favorable (e.g., Francis and
Philbrick [1993], Kim and Lustgarten [1998]). Median scaled forecast er-
rors, reported in panel A, generally follow a similar pattern. However, the
medians do not follow a monotonic increase moving from strong buy to
strong sell recommendation. While the unadjusted data reject H1 and sup-
port H1a, the management relations hypothesis, these data do not appro-
priately consider the association between earnings and both forecast errors
and recommendations.
Panel B presents a similar analysis using mean and median adjusted
scaled forecast errors. The analysis of adjusted scaled forecast errors (i.e.,
controlling for the relationship between actual earnings and forecast error)
yields dramatically different results from those reported in panel A. The
mean adjusted scaled forecast error for the strong buy recommendation
is positive (t = 1.65, p = .098) while the mean adjusted scaled forecast
errors for both moderate and strong sell recommendations are negative
(
t = −1.91, p = .056; t = −2.49; p = .013, respectively). Mean-adjusted
scaled forecast errors shift consistently and monotonically from optimism
to pessimism as recommendations move from strong buy to strong sell.
The mean-adjusted scaled forecast error is significantly more optimistic for
strong buy than for strong sell (t = 2.75, p = .006), for moderate buy than
for moderate sell (t = 1.98, p = .048), and for hold than for moderate sell
recommendations (t = 1.77, p = .077). Analysis of median-adjusted scaled
forecast errors yields a similar pattern of forecast optimism in association
1
4
All p values relating to t statistics in the text represent the probability of a larger abso-
lute t value. The t values for comparisons are approximate t values under the assumption of
unequal variances. The p values represent the probability via the Cochran and Cox [1950]
approximation for the approximate t test.

TRADING RECOMMENDATIONS
99
T A B L E 3
Market Value of Equity Scaled Annual Earnings Forecast Errors by Analyst Recommendation
for the Sample of Forecast Errors with Contemporaneous or Recently Preceding
a
Recommendation by the Same Analyst
Mean Scaled
Median
Analyst
Recommendation
Sample
Size
Forecast Error in
% (t-value)
Scaled Forecast
Mean Comparisons
b
Error in %b
c
(t-value)
Panel A: Scaled Earnings Forecast Errors using Unadjusted Data
∗
∗
∗
∗
∗
∗∗
∗∗
∗∗
∗∗
∗∗
Strong Buy (SB)
Moderate Buy (MB)
Hold (H)
11,047
7,886
13,916
960
.587
12.52)
.191
.172
.243
.403
.340
SB v. MB (.17)
SB v. SS (4.14)
MB v. H (3.86)
∗
∗∗
∗∗
∗∗
∗∗
∗∗
∗∗∗
∗∗∗
(
(
(
.599
∗
10.79)
.886
MB v. MS (1.61)
H v. MS (.41)
∗
18.03)
.982
∗∗
Moderate Sell (MS)
Strong Sell (SS)
MS v. SS (1.99)
∗
(4.24)
1.668
803
∗
(6.49)
Panel B: Adjusted Scaled Forecast Errorsd
∗
∗∗
Strong Buy
Moderate Buy
Hold
11,047
7,886
13,916
960
.061
1.65)
.022
.040
.005
SB v. MB (.66)
SB v. SS (2.75)
∗
∗∗∗
(
MB v. H (.69)
MB v. MS (1.97)
H v. MS (1.77)
∗
∗
(
0.49)
∗
−.018
−.49)
−.325
−.015
(
Moderate Sell
Strong Sell
−.167∗∗
MS v. SS (.62)
∗
(
−1.91)
803
−.484
−.314∗∗∗
∗
∗
(
−2.49)
Panel C: Central Region Scaled Forecast Errorse
∗
∗
∗
∗∗
∗∗
∗∗
∗∗∗
SB v. MB (2.69)
SB v. SS (4.15)
MB v. H (1.09)
MB v. MS (1.06)
H v. MS (.68)
Strong Buy
Moderate Buy
Hold
5,195
3,650
6,261
336
.349
12.44)
.115
.043
.042
.015
∗
∗∗
∗∗
∗∗
∗∗∗
(
.231
∗
(
(
(
6.89)
.185
∗
6.93)
.105
∗∗
MS v. SS (2.10)
Moderate Sell
Strong Sell
a
0.92)
259
−.296
−.277∗∗∗
∗
∗
(
−1.94)
Forecasts, realized earnings, and recommendations are from the Zacks Investment Research database.
The sample is based on first quarter forecasts of annual earnings for December fiscal year end firms. We
match each forecast with the most recent preceding or contemporaneous recommendation by the same
analyst, provided the recommendation does not precede the forecast by more than one year (n = 34,612).
Forecast error is defined as forecasted earnings less actual earnings. Both forecasts and actual earnings are
scaled by beginning of the year market value of equity.
b
Mean and median tests from zero are based on a parametric t-test and a nonparametric sign test,
respectively.
c
SB = strong buy, MB = moderate buy, H = hold, MS = moderate sell, SS = strong sell.
d
Forecast errors are adjusted to control for the observed association between earnings and forecast
errors (see figure 1). Mean (median) earnings adjusted forecast errors are obtained by identifying 20 equal
sized portfolios of observations based on ranks of earnings scaled by market value of equity, measuring
forecast error as the forecast less actual earnings, again scaled by beginning-of-the-year market value of
equity, and then subtracting the relevant portfolio mean (median) forecast error from the observed forecast
error.
e
The central region includes all observations with earnings scaled by beginning-of-the-year market value
ranging from 5% to 10% (see section 4.1).
∗
∗∗ ∗∗∗
/ ) Designate two-tailed statistical significance at the 10% (5%/1%) level.
(

100
M. EAMES, S. M. GLOVER, AND J. KENNEDY
with buy recommendations and pessimism for sell recommendations.
Consequently, when we analyze scaled forecast errors, adjusted to control
for the relationship between earnings level and forecast error, we again
reject hypothesis H1, but now the results support the objectivity illusion
and trade boosting hypotheses (H1b) and not the management relations
hypotheses (H1a). These results also provide evidence that the relation
between recommendation and forecast error is not driven solely by the
relation between scaled earnings and forecast error. If that were the case, we
would expect no relation between recommendations and adjusted forecast
errors.
Results for scaled forecast errors in the central region (i.e., where the rela-
tion between scaled earnings and scaled forecast error is weak) are presented
in panel C, and are consistent with results presented in panel B for mean and
median adjusted scaled forecast errors. Again we observe optimistic mean
scaled forecast error for strong buy recommendations and pessimistic mean
scaled forecast error for strong sell recommendations, with scaled forecast
errors for intermediate recommendations monotonically aligning between
the extremes. The mean scaled forecast error is significantly more optimistic
for strong buy than for strong sell (t = 4.15, p = .000), for strong buy than
for moderate buy (t = 2.69, p = .007), and for moderate sell than for strong
sell recommendations (t = 2.11, p = .036). Median scaled forecast errors
decrease in magnitude going from strong buy to strong sell recommenda-
tions, and exhibit optimism for buy recommendations and pessimism for
strong sell recommendations. Results in panel C provide a control for the
association between earnings and forecast errors by examining a subset of
the unadjusted data reported in panel A.15 These results support the trade
boosting and objectivity illusion hypotheses (H1b), but not the manage-
ment relations hypotheses (H1a). A comparison of panels A and C also
highlights the key role extreme earnings observations likely play in prior
research. Panels B and C suggest that evidence in previous studies support-
ing the management relations hypothesis (panel A) is driven by the relation
between analyst forecast error and scaled earnings levels outside the central
region.
1
5
Additional evidence that we have not “adjusted away” management relations is obtained
by investigating the relation between forecast errors and forecasts. If recommendations induce
biased forecast errors and if recommendations are associated with the level of forecast (i.e.,
buy (sell) recommendations are generally associated with higher (lower) forecasted earnings),
we expect to observe a significant relation between forecast error and forecast. For trade
boosting and the objectivity illusion (management relations) we expect a positive (negative)
association between forecast error and forecast. As expected, the relation between forecast
error and forecasts is positive and highly significant (R-squared = 67%). Just as earnings are
significantly correlated with recommendations (see figure 2), we find a similar correlation
(
p < .001, see also panel A, table 2) between forecasted earnings and recommendations. The
observed positive relationship is consistent with the pattern predicted by objectivity illusion and
trade boosting but not management relations and provides additional evidence that analysts’
earnings forecasts are systematically biased. We thank the reviewer for suggesting this analysis.

TRADING RECOMMENDATIONS
101
FIG. 3.—Mean scaled earnings forecast error by recommendation for observations of first
quarter forecasts of annual earnings with a contemporaneous or recently preceding recommen-
dation by the same analyst. We obtain individual analyst recommendations, actual earnings,
and forecasted earnings from the Zacks Investment Research database. All forecast error ob-
servations are scaled by beginning-of-the-year market value of equity. “Unadjusted” forecast
errors are defined as forecasted earnings minus actual earnings. The unadjusted and earnings-
adjusted plots are based on 34,612 observations. The “Earnings Adjusted” data are obtained
by identifying 20 equal-sized portfolios of observations based on ranks of market-value scaled
earnings and then subtracting the relevant portfolio mean forecast error from the observed
forecast error. The “Central Region” plot is based on 15,701 unadjusted-forecast-error observa-
tions where earnings range from 5 to 10 percent of market value of equity. We obtain individual
analyst recommendations, actual earnings, and forecasted earnings from the Zacks Investment
Research database. All forecast error observations are scaled by beginning-of-the-year market
value of equity. “Unadjusted” forecast errors are defined as forecasted earnings minus actual
earnings. The unadjusted and earnings-adjusted plots are based on 34,612 observations. The
“
Earnings Adjusted” data are obtained by identifying 20 equal-sized portfolios of observations
based on ranks of market-value scaled earnings and then subtracting the relevant portfolio
mean forecast error from the observed forecast error. The “Central Region” plot is based on
1
5,701 unadjusted-forecast-error observations where earnings range from 5 to 10 percent of
market value of equity.
The bar chart in figure 3 summarizes results presented in table 3 and il-
lustrates the relation between forecast error and stock recommendation for
the three data sets we analyze (i.e., unadjusted data, earnings adjusted data
and central earnings region data). Figure 3 illustrates that when we exam-
ine the data in a manner consistent with prior studies (i.e., using the entire
range of earnings with no control) we find results consistent with the man-
agement relations hypothesis. When we control for the relation between

102
M. EAMES, S. M. GLOVER, AND J. KENNEDY
earnings and forecast errors by examining the “adjusted” and “central re-
gion” data we find strikingly different results; the relation between forecast
errors and recommendations no longer supports the management relations
hypothesis.1
6
4.3 OBJECTIVITY ILLUSION AND TRADE BOOSTING
Both the objectivity illusion and trade boosting hypotheses predict op-
timistic forecast errors for buys and pessimistic forecast errors for sells.
While both factors could certainly exist simultaneously, we examine whether
our results are more supportive of the trade boosting or objectivity illusion
hypothesis.
Analysts’ incentives to intentionally bias earnings forecasts to boost trade
are likely greater for strong recommendations than for moderate recom-
mendations, since analysts can boost trading on stocks that do not have an
extreme recommendation with a shift to a more extreme recommendation,
as well as by intentionally biasing the earnings forecast. Analysts likely prefer
recommendation-related trade boosting over earnings-forecast-based trade
boosting because recommendation shifts are a stronger signal to investors
than an unspecified level of forecast error at the time of forecast issuance
and because forecast errors are more readily measurable than are recom-
mendation errors. Therefore, for the trade-boosting hypothesis we expect
no significant relation between forecast error and recommendations for in-
termediate recommendation levels, significant optimistic forecast bias for
strong buy recommendations, and significant pessimistic forecast bias for
strong sell recommendations.
The objectivity illusion predicts forecast errors associated with both mod-
erate and strong recommendations and the degree of bias should increase
as recommendations become stronger. Thus, greater absolute forecast error
for strong buy and strong sell recommendations and a significant relation
between forecast bias and recommendation for intermediate recommenda-
tions would lend support for the objectivity illusion.
Panel B of table 3 reports that mean and median adjusted scaled fore-
cast errors are significantly optimistic for strong buy recommendations,
significantly pessimistic for strong sell recommendations, and exhibit a con-
tinuum in forecast bias across all levels of recommendations. Consistent
with the objectivity illusion hypothesis and not with trade boosting, mean
adjusted forecast errors differ significantly between moderate buy and mod-
erate sell recommendations (t = 1.97, p = .048) and between moderate sell
and hold recommendations (t = 1.77, p = .077).
1
6
To preclude the possibility that a correlated error structure in our data might be driving
the significant results we obtain in panels B and C of table 3, we also examined our data on
a year-by-year basis. Although significance was generally reduced due to the smaller samples
employed, the year-by-year results were highly consistent with results obtained for the entire
sample.

TRADING RECOMMENDATIONS
103
5
. Summary and Conclusion
We examine the relation between analysts’ earnings forecast errors and
stock recommendations. Prior research has generally found both broker
and non-broker-analyst earnings forecasts to be optimistic on average, and
increasingly optimistic as the stock recommendation becomes less favorable.
Prior research posits that analysts intentionally issue optimistic forecasts in
order to curry management favor (the management relations hypothesis).
However, this explanation contradicts conventional wisdom reported in the
financial press as well as recent academic research suggesting managers may
actually prefer slightly pessimistic forecasts.
We believe that prior results supporting the management relations hy-
pothesis are driven by omitting an important correlated variable, earnings.
When we control for the relation between forecast error and earnings, we
find that analyst forecast errors are optimistic for buy recommendations and
pessimistic for sell recommendations. These results are not consistent with
the management relations hypothesis but are consistent with both the trade
boosting and objectivity illusion hypotheses. The trade boosting hypothesis
contends that analysts intentionally bias their forecasts to boost trade. The
objectivity illusion hypothesis contends that analysts unintentionally bias
their forecasts to achieve consistency with their stock recommendations.
Results of an analysis to distinguish between trade boosting and objectivity
illusion appear more consistent with the objectivity illusion.
REFERENCES
AFFLECK-GRAVES, J.; L. DAVIS; AND R. R. MENDENHALL. “Forecasts of Earnings Per Share: Possible
Sources of Analyst Superiority and Bias.” Contemporary Accounting Research (Spring 1990):
5
01–17.
ALI, A.; A. KLEIN; AND J. ROSENFELD. “Analysts’ Use of Information about Permanent and Tran-
sitory Earnings Components in Forecasting Annual EPS.” The Accounting Review 67 (1992):
1
83–98.
BALOG, S. “What an Analyst Wants from You.” Financial Executive (July/August 1991): 47–52.
BIGGS, S. “Financial Analysts’ Information Search in the Assessment of Corporate Earnings
Power.” Accounting, Organizations and Society 9 (1984): 313–23.
BOINEY, L.; J. KENNEDY; AND P. NYE. “Instrumental Bias in Motivated Reasoning: More When
More is Needed.” Organizational Behavior and Human Decision Processes 72 (October 1997):
1
–24.
BROWN, L. “Earnings Forecasting Research: Its Implications for Capital Markets Research.”
International Journal of Forecasting 9 (1993): 295–320.
BROWN, L. “Managerial Behavior and the Bias in Analysts’ Earnings Forecasts.” Working paper,
Georgia State University, 1999.
BROWN, L.; P. GRIFFIN; R. HAGERMAN; AND M. ZMIJEWSKI. “An Evaluation of Alternative Proxies
for Market’s Assessment of Unexpected Earnings.” Journal of Accounting and Economics 11
(1987): 159–93.
BURGSTAHLER, D., AND M. EAMES. “Management of Earnings and Analyst Forecasts.” Working
paper, University of Washington, 2000.
BUTLER, K. C., AND H. SARAOGLU. “Improving Analysts’ Negative Earnings Forecasts.” Financial
Analysts’ Journal (May/June 1999): 48–56.
COCHRAN, W., AND G. COX. Experimental Designs. New York. John Wiley and Sons, Inc.

104
M. EAMES, S. M. GLOVER, AND J. KENNEDY
DAS, S.; C. B. LEVINE; AND K. SIVARAMAKRISHNAN. “Earnings Predictability and Bias in Analysts’
Earnings Forecasts.” The Accounting Review 73 (April 1998): 277–94.
DITTO, P. H., AND D. F. LOPEZ. “Motivated Skepticism: Use of Differential Decision Criteria for
Preferred and Nonpreferred Conclusions.” Journal of Personality and Social Psychology (1992):
5
68–84.
DUGAR, A., AND S. NATHAN. “The Effect of Investment Banking Relationships on Financial
Analysts’ Earnings Investment Recommendations.” Contemporary Accounting Research 12 (Fall
1
995): 131–60.
EAMES, M.; S. GLOVER; AND K. STICE, “The Relation between Earnings and Earnings Forecast
Error.” Working Paper, Brigham Young University, 2001.
FESTINGER, L. “A Theory of Cognitive Dissonance.” Stanford, CA: Standford University Press,
1
957.
FRANCIS, J., AND D. R. PHILBRICK. “Analysts’ Decisions as Products of a Multi-Task Environment.”
Journal of Accounting Research 31 (Autumn 1993): 216–30.
GIVOLY, D., AND J. LAKONISHOK. “Properties of Analysts’ Forecasts of Earnings: A Review and
Analysis of the Research.” Journal of Accounting Literature 3 (Spring 1984): 117–52.
HWANG, L.; C. JAN; AND S. BASU. “Loss Firms and Analysts’ Earnings Forecast Errors,” The Journal
of Financial Statement Analysis (Winter 1996): 18–29.
IP, G. “Traders Laugh Off the Official Estimate on Earnings, Act on Whispers.” Wall Street Journal,
January 16 (1997a).
IP, G. “Rise in Profit Guidance Dilutes Positive Surprises.” Wall Street Journal, June 23 (1997b).
KIM, C., AND S. LUSTGARTEN. “Broker-Analysts’ Trade-Boosting Incentive and Their Earnings
Forecast Bias.” Working paper, Queens College of the City University of New York, 1998.
KUNDA, Z. “The Case for Motivated Reasoning.” Psychological Bulletin 108 (1990): 480–98.
LIM, T. “Rationality and Analyst’s Forecast Bias.” Working Paper, Dartmouth College. 1998.
LYS, T., AND S. SOHN. “The Association Between Revisions of Analysts’ Earnings Forecasts and
Security-Price Changes.” Journal of Accounting and Economics 13 (December 1990): 341–63.
MATSUMOTO, D. A. “Management’s Incentives to Avoid Negative Earnings Surprises.” Working
paper, University of Washington, 2001.
MCGEE, S. “As Stock Market Surges Ahead, ‘Predictable’ Profits are Driving It.” Wall Street
Journal, C1, May 5 (1997).
MCNICHOLS, M., AND P. C. O’BRIEN. “Self-Selection and Analyst coverage.” Journal of Accounting
Research (Supplement 1997): 167–99.
MENDENHALL, R. R. “Evidence on the Possible Underweighting of Earnings-Related Informa-
tion.” Journal of Accounting Research 29 (Spring 1991): 170–79.
O’BRIEN, P. C. “Analysts’ Forecasts as Earnings Expectation.” Journal of Accounting and Economics
1
0 (1988): 53–83.
SNYDER, M., AND W. B. SWANN, JR. “Hypothesis-Testing Processes in Social Interaction.” Journal
of Personality and Social Psychology 36 (1978): 1202–12.