The Best of Times, The Worst of Times: Testing Which Behavioral Biases Affect Analyst Forecasts

International Review of Finance, 2018
DOI: 10.1111/irfi.12168
The Best of Times, the Worst of

Times: Testing which Behavioral
Biases Affect Analyst Forecasts*
YUK YING CHANG AND WEI-HUEI HSU
School of Economics and Finance, Massey University, New Zealand
ABSTRACT
Mood-induced optimism, cognitive inaccuracy, and distraction can affect

analyst forecasts. This study compares and contrasts these influences. The
novelty of our approach is that we first show that these behavioral biases
have different implications for analysts’ forecast errors conditioned on the
errors being positive and negative. We then use proxies for positive and neg-
ative moods to empirically test the support for each of these biases. Consis-
tent with cognitive precision, we find that analysts make less (more) accurate
forecasts when they are in positive (negative) moods. We further show that
these results are driven neither by sentiment associated with contemporane-
ous economic or market conditions nor by under- or overreaction to more
bad news released on days immediately before weekends or holidays.
JEL Codes: G24; G14; D03; G02
Accepted: 29 November 2017
I. INTRODUCTION
A fairly large body of literature suggests that the moods or psychological states
of people affect their judgment. People in a more positive mood are more opti-
mistic and assign higher probabilities to positive events and lower probabilities
to negative events (Wright and Bower 1992). For example, a happier investor is
more optimistic about the expected performance of stock markets (Kaplanski
et al. 2015). A college student is also more optimistic about his own perfor-
mance after watching his college sports team win compared to when he
watches them lose (Hirt et al. 1992).
However, positive bias is also consistent with distraction. Sport events may
draw attention away (Gantz and Wenner 1991; Schmidt 2013; Ehrmann and
Jansen 2017), thereby increasing the likelihood that people will make more or
* We thank Sudipto Dasgupta and Ling Cen for their helpful suggestions. We also thank Kee-Hong
Bae, Michael Brennan, Bei Cui, Jie Gan, Gilles Hilary, Peter Joos, Hongping Tan, Xiaoyun Yu, and
the participants in the 2016 FMA Asia/Pacific Conference, the 2016 Asian Finance Association Con-
ference and the Massey University seminar. We acknowledge Andrea Bennett for proofreading our
drafts and Massey Business School for providing financial support.
© 2018 International Review of Finance Ltd. 2018

International Review of Finance
larger errors. For example, Drake et al. (2016) find that market reactions to earn-
ings news released during the NCAA basketball tournament are muted. In addi-
tion, positive bias may be explained by cognitive inaccuracy. Happy investors
may be easily swayed by emotion, which may decrease both their logical consis-
tency and their attention to detail (Schwarz and Bless 1991). Hence, such inves-
tors may be less analytical and their predictions about market performance may
be more biased. Meloy (2000) reports that mood-elevated consumers display
severely distorted evaluations of new product information. Therefore, it is
unclear whether positive bias is completely driven by optimism.
In this paper, we differentiate among explanations based on optimism, dis-
traction, and cognitive accuracy using analyst forecasts. Analysts produce fore-
casts higher and lower than realized earnings, which allow us to distinguish
between optimism and cognitive accuracy explanations, as discussed further
below. Moreover, analyst forecasts are widely followed and watched. Analyst
accuracy is an important issue, and related studies have long existed in the
mainstream literature (e.g., Basi et al. 1976).1 However, although optimism and
distraction are fairly well-known subjects and have been discussed in the litera-
ture, cognitive precision is less researched, except in the psychology field. This
study provides evidence consistent with cognitive accuracy.
Our arguments are as follows. When an analyst is in a more positive mood,
he is more optimistic, and his forecast errors are more positive on average. Con-
versely, when an analyst is in a more negative mood, he is more pessimistic,
and his forecast errors are more negative on average. Importantly, these mood-
induced biases are not ex post conditional on whether the errors are positive or
negative. For example, assume an unbiased analyst whose mean positive errors
average + 10¢ and whose mean negative errors average − 10¢ (the basis of com-
parison). In this case, when the analyst is more optimistic, the corresponding
positive and negative errors will average, for example, +12¢ and −9¢, respec-
tively. In contrast, when an analyst is more pessimistic, the positive and nega-
tive errors will average, for example, +9¢ and −12¢, respectively. We call this
“the optimism explanation.”2
When a happy mood negatively affects the accuracy with which an analyst
processes information, he will make larger errors on both sides, for example,
making average positive and negative errors of +12¢ and −12¢, respectively. In
contrast, when an analyst is in a negative mood, he will process information
more systematically and substantively (Martin and Clore 2013). Hence, he is
more accurate and his average positive and negative errors may be smaller, for
1 On September 1, 2017, there were 77,000 and 59,500 Google Scholar hits for “accuracy of
analyst forecasts” and “analyst forecast errors”, respectively.
2 This example compares a mood-affected analyst with an unbiased analyst. However, a similar
argument applies if we compare an analyst’s forecast errors when he is exposed to mood-
affecting events with his average positive and negative forecast errors. The latter comparison
is more consistent with our empirical methodology, which considers “within variation” and
subsumes time-invariant analyst effects with analyst fixed effects.
2 © 2018 International Review of Finance Ltd. 2018

The Best of Times, the Worst of Times
example, +9¢ and −9¢, respectively. We refer to this as “the cognition

explanation.”
However, when an event disturbs an analyst emotionally and distracts him
from his work, regardless of whether it is a happy or sad event, the analyst consis-
tently commits larger errors. In this case, regardless of whether the analyst is in a
positive or negative mood, the average positive and negative errors are larger, for
example, +12¢ and −12¢, respectively. We term this “the distraction explanation.”
Studies find that people are happier when they anticipate holidays (Gilbert
and Abdullah 2002; Van Boven and Ashworth 2007; Nawijn et al. 2010) and on
holidays (Bollen et al. 2011; Sharpe 2014). Achor and Gielan (2016) suggest that
taking more vacations leads to greater happiness. Using various daily mood
measures, including the Gallup Mood Survey and Facebook Gross National
Happiness Index, Autore and Jiang (2017) provide evidence of uplifted mood
swings when holidays approach. Good mood has also been proposed as a possi-
ble explanation for significant preholiday effects in finance (Thaler 1987;
Fabozzi et al. 1994; Frieder and Subrahmanyam 2004; Bialkowski et al. 2012;
Bergsma and Jiang 2015; Hirshleifer et al. 2016; Autore and Jiang 2017). Hence,
we hypothesize that analysts are in good moods when holidays are upcoming
and that the analysts become happier as holidays get closer.3
Conversely, studies suggest that disaster information causes a negative mood
or affects mood adversely (Johnson and Tversky 1983; Göritz and Moser 2006;
Västfjäll et al. 2008; Papousek et al. 2014). Kaplanski and Levy (2010a) find sig-
nificantly negative stock market returns following aviation disasters. Antoniou
et al. (2017) document more negative sentiment effects due to terrorist attacks
and mass shootings among managers of corporate policies, as well as an
increase in cash holdings and a decrease in R&D expenditure. We consider it
likely that disaster and holiday events will be associated with different moods
to distinguish between the distraction and cognitive accuracy explanations.
Studies also suggest that people tend to be in more negative moods in
autumn as a result of seasonal affective disorder (SAD) (Kamstra et al. 2003).
Evidence consistent with the SAD effect includes studies of seasonal anomalies
(Lakonishok and Smidt 1988), the Halloween indicator (Bouman and Jacobsen
2002), financial analysts and equity market returns (Lo and Wu 2010), mutual
fund flows (Kamstra et al. 2017), and the mood beta of stocks (Hirshleifer
et al. 2016). Therefore, we hypothesize that analysts will be in more negative
moods in October.4
3 If people have a variety of moods when there is a forthcoming holiday, there will be a bias
against finding significant preholiday effects of positive moods.
4 For the United States, fall is defined as September 21 to December 20. October is the first full
month of fall. In addition, Hirshleifer et al. (2016) find that the value-weighted CRSP return
is lowest in October, at −0.37%, in their sample period (1963–2015). We thus choose October
to test the SAD effect. We do not consider the effects of “blue Monday,” because several stud-
ies (e.g., Stone et al. 2012) suggest that blue Monday does not exist. Moreover, Wang et al.
(1997) find that negative average stock returns on Mondays are neither prevalent nor robust.
© 2018 International Review of Finance Ltd. 2018 3

Based on a sample of forecasts from US firms over the 1983 to 2014 period,
we find that forecasts released during calendar weeks that end on Thanksgiving
Day, Christmas Day, or New Year’s Day5,6 are more biased for forecasts both
higher and lower than actual earnings. We find similar results when we look at
the two business days immediately before and on 13 holidays throughout the
year.7 The more positive bias is consistent with all three explanations, whereas
the more negative bias supports only the cognition and distraction explana-
tions. In addition, we find that forecast accuracy improves when major disasters
involving significant fatalities occur in the United States. Hence, because posi-
tive forecast errors are less positive in this case, the optimism and cognition
explanations win out over the distraction explanation. Conversely, because
negative forecast errors are less negative, the cognition explanation is the only
winner. Similar to the disaster results, we also find higher forecast accuracy in
October than in other months.
In summary, only the cognition explanation is consistent with all of the
results.8 The above differential forecast errors are statistically significant at the
1% level when we control for at least 18 explanatory variables and for year,
industry, firm, and analyst fixed effects and when robust standard errors are
based on two-way clustering at the firm and analyst levels. The additional inac-
curacy and precision are also economically significant. The average preholiday
inaccuracy effect on positive errors is up to 21% of the unconditional mean of
positive errors, whereas the corresponding average effect on negative errors is
up to 28% of the corresponding mean. The magnitude of the disaster precision
effect on positive errors is 12% of the unconditional mean of positive errors on
average, whereas the corresponding average effect on negative errors is 10%.
The magnitude of the October precision effect on positive errors is 7% of the
unconditional mean of positive errors, on average, whereas the corresponding
average effect on negative errors is 15%.
Furthermore, we find that the disaster incremental precision effect is stronger
among analysts located in the area where the disaster occurs (local analysts)
than among nonlocal analysts. Likewise, accuracy improvement is generally
more pronounced among Chinese analysts than among non-Chinese analysts
when major disasters occur in China.
5 According to Gallup, holidays (particularly Thanksgiving Day and Christmas Day) are usually
the happiest days of the year (Sharpe 2014).
6 Van Boven and Ashworth (2007) provide evidence of greater happiness approximately 2 weeks
prior to Thanksgiving Day.
7 The design follows those presented in Hirshleifer et al. (2016) and Autore and Jiang (2017).
The 13 holidays are New Year’s Day, Valentine’s Day, Presidents’ Day, St. Patrick’s Day, Eas-
ter, Mother’s Day, Memorial Day, Father’s Day, Independence Day (Fourth of July), Labor
Day, Halloween, Thanksgiving, and Christmas.
8 Optimism, distraction, and cognitive accuracy are not mutually exclusive. Although our
results are more supportive of cognitive precision, they do not necessarily suggest the nonex-
istence of optimism and distraction.

In addition, preholiday incremental inaccuracy is positively related to the

closeness of a forecast release day to an upcoming holiday. Furthermore, we
obtain qualitatively identical results for several additional analyses, including
those that use personal income tax shocks as an alternative mood proxy.
We also show that our results are driven neither by the sentiment effect of the
contemporaneous economic conditions nor by under- or over-reaction to bad
news released just before weekends or holidays. Finally, we find that analysts are
as responsive to news during preholiday periods—regardless of whether it is the
day immediately before a weekend/holiday—as they are during other periods.
The remainder of this paper is organized as follows. Section II provides a brief
overview of related literature. Section III explains our hypotheses. Section IV
describes the data and variables. Section V explicates the methodology.
Section VI reports summary statistics and the results of a univariate analysis.
Section VII presents and discusses the results of our regression analyses.
Section VIII concludes.
II. RELATED LITERATURE
Studies have examined how stock returns are associated with mood proxy vari-
ables that reflect investor optimism or pessimism, such as biorhythms (e.g., SAD,
daylight savings time, and lunar phases) (e.g., Kamstra et al. 2000; Kamstra et al.
2003; Yuan et al. 2006; Kamstra et al. 2012) and weather (e.g., rain, sunshine,
temperature, wind, and storms) (e.g., Saunders 1993; Hirshleifer and Shumway
2003; Chang et al. 2008; Goetzmann et al. 2015; deHaan et al. 2016). These stud-
ies posit that certain variables, including sports outcomes (Edmans et al. 2007;
Kaplanski and Levy 2010b) and disastrous events (Kaplanski and Levy 2010a),
affect investors’ moods, which in turn affect their decisions. Consequently, asset
prices and returns vary with mood. Shu (2010) provides a theoretical model that
helps to explain the results of empirical studies. Studies also consider other
financial variables. For example, Cortés et al. (2016) suggest that more pessimis-
tic loan officers issue fewer credit approvals on cloudier days. Regarding analyst
forecasts, deHaan et al. (2016) find that analysts experiencing unpleasant
weather engage in fewer forecasting activities and make more pessimistic fore-
casts than those experiencing pleasant weather, which the authors conclude is
evidence of the impact of weather-induced moods. Dolvin et al. (2009) and Lo
and Wu (2010) document that SAD is associated with pessimism, which results
in lower analyst optimism. Bourveau and Law (2016) report more pessimistic
forecasts among analysts in New Orleans than among analysts outside of Louisi-
ana following Hurricane Katrina in 2005. In this study, we not only examine
pessimism/optimism, but also consider cognitive precision.
There is also a growing body of literature on how distraction or limited atten-
tion affects decision-making that is relevant to financial markets. For instance,
Hirshleifer and Teoh (2003) study investors’ limited attention and firms’ informa-
tion disclosure and financial reporting. More recently, Lu et al. (2016) conclude

that both marriages and divorces distract hedge fund managers, who consequently
deliver lower fund alphas. In the analyst forecast literature, in addition to pessi-
mism/optimism, studies also consider distraction. For example, Dong and Heo
(2014) show that a higher flu intensity is associated with a lower degree of dis-
agreement among analysts, overprediction (interpreted as greater optimism) of tar-
get prices for high-performing stocks, and underprediction (interpreted as greater
pessimism) of target prices for low-performing stocks. They suggest that their
results are driven by limited attention and effort due to analysts being distracted
by the flu symptoms of family members, colleagues, and themselves.
III. HYPOTHESES
In this section, we explain our main testable hypotheses. We first consider the
effect of good moods (more optimism) on analyst forecast errors. More optimism
is defined as a more positive bias in analysts’ forecasts. Specifically, if true earn-
ings are normalized to 0, an optimistic analyst is positively biased, and forecast
x has an expected value of μ > 0. Greater optimism implies a higher value of μ.
However, forecast errors can be either positive or negative ex post, that is,
x can be either positive or negative. We argue that it is useful to examine fore-
cast errors conditioned on x being positive or negative. Assume that x is drawn
from a distribution function F(x) with mean μ and standard deviation σ. For
most standard distributions, we would expect that as an analyst becomes more
optimistic, both E[X|x > 0] and E[X|x < 0] would increase in μ. Thus, if being in
a good PRE_3HOLIDAY mood makes an analyst more optimistic, the expected
forecast error conditioned on the error being positive or negative should
become more positive.
For exposition, we consider a uniform distribution F(x) over the interval [−δ1
+ κ, δ2 + κ], where κ ≥ 0 and δ2 > δ1 > κ > 0; this is trivial to verify. For this distri-
2
bution, the mean is μ = δ2 −δ21 + 2 > 0. The variance is ðδ1 12
+ δ2 Þ
. An increase in the
mean without any change in the variance is represented by an increase in κ.
Moreover, E[X|x > 0] = (δ2 + κ)/2 and E[X|x < 0] = (−δ1 + κ)/2. Thus, both E[X|
x > 0] and E[X|x < 0] would increase in κ and hence in μ.9 These observations
lead to our first hypothesis:
Hypothesis 1 (optimism hypothesis). If analysts become more optimistic

(more pessimistic) in good (bad) times, the values of the expected forecast errors
conditional on the errors being positive or negative both increase (decrease).
An alternative hypothesis is that major events—either good or bad—strongly

disturb analysts emotionally and distract them10; accordingly, the precision of
9 See Appendix A for the case of a normal distribution.

10 In addition to being emotionally disturbing, major events can draw the nonemotional atten-
tion of analysts.

their estimates will be lower around major events. In terms of the uniform dis-
tribution, this is captured by an increase in the spread of the range, keeping the
mean unchanged. Let the range of the distribution be given by [−δ1 + κ − ω, δ2 +
κ + ω], where ω > 0. The mean and variance are now given by μ = δ2 − δ21 + 2 and
ðδ1 + δ2 + 2ωÞ2
12 Moreover, E[X|x > 0] = (δ2 + κ + ω)/2 and E[X|x < 0] = (−δ1 + κ − ω)/2.
.
Thus, when precision falls (variance increases), E[X|x > 0] increases and E[X|
x < 0] decreases.11 These considerations lead to our second hypothesis:
Hypothesis 2 (distraction hypothesis). If analysts become more distracted

when exposed to good (bad) times, the value of the expected forecast error con-
ditional on the error being positive increases whereas the value of the expected
error conditional on the error being negative decreases.
Finally, we have the cognition hypothesis, which maintains that individ-

uals are more precise when exposed to negative experiences but may ignore
salient information when exposed to positive experiences. Thus, unlike the
first two hypotheses, in bad times, precision improves, so that E[X|x > 0]
decreases while E[X|x < 0] increases. According to the distraction hypothesis,
in good times, information processing is poorer, precision is lower, and abso-
lute errors are larger on both sides of zero.
Hypothesis 3 (cognition hypothesis). If analysts process information less

precisely in good times, the value of the expected forecast error conditional
on the error being positive increases, whereas the value of the expected
error conditional on the error being negative decreases. However, if analysts
process information more precisely in bad times, the value of the expected
forecast error conditional on the error being positive decreases, whereas the
value of the expected error conditional on the error being negative
increases.
To summarize,
Good times Bad times

Hypothesis E[X|x > 0] E[X|x < 0] E[X|x > 0] E[X|x < 0]
Optimism " " # #
Distraction " # " #
Cognition " # # "
11 See Appendix A for the case of a normal distribution.

IV. DATA AND VARIABLES
We obtain realized earnings and analyst earnings forecasts from I/B/E/S Detail
History files. The sample cover the 1982 to 2014 period. We only consider annual
earnings per share (EPS) forecasts released after the end of the last fiscal year and
before or at the end of the fiscal year for which the annual EPS is forecasted.12 We
use I/B/E/S actual earnings instead of Compustat earnings because I/B/E/S has a pol-
icy of reporting actual earnings numbers that are consistent with forecasts, that is, it
excludes the same items from actual EPS numbers that the majority of analysts
exclude from their forecasts (Christensen 2007). The sources of accounting and
financial data are Compustat and CRSP, respectively. The holiday dates are obtained
from http://www.timeanddate.com. Our definition of contraction periods is based
on National Bureau of Economic Research (NBER) business cycle reference dates.
We obtain monthly consumer sentiment data from University of Michigan surveys
and use the monthly consumer confidence index from Conference Board surveys.
The monthly Baker-Wurgler investor sentiment index (which ended in December
2010) was downloaded from Wurgler’s website (Baker and Wurgler 2006). The num-
ber of daily news items is obtained from Bloomberg. The dates of US disasters were
retrieved from the website below or from webpages linked to that website:
https://en.wikipedia.org/wiki/List_of_disasters_in_the_United_States_by_
death_toll.13
Our analyst location data are based on annual volumes of Nelson’s Directory
of Investment Research.14 We obtain location information for disasters from
12 Certain previous studies do not use forecasts released within 30 days before the end of the
fiscal year (e.g., Clement 1999; Clement and Tse 2005). Two main reasons are given. First,
these forecasts tend to be made by analysts who follow other analysts rather than the sub-
ject companies. Second, these forecasts may be for the next fiscal year. However, there are at
least three reasons why these factors do not concern us. First, as we show in Table 4, analysts
who create forecasts in the preholiday period generally make their first forecast (median of
day 65) no later than those who do not make forecasts in the preholiday period (median of
day 67). Second, there is no reason to expect that forecasts made by followers are more accu-
rate. O’Brien (1988) finds that more recent forecasts are more accurate and thus are unlikely
to be made by followers. Third, Guttman (2010) shows that analysts with higher learning
abilities tend to forecast later. Hence, recent forecasts tend to be made by capable forecasters
rather than by followers. Furthermore, there are studies that use the most recent forecasts
(Karamanou and Vafeas 2005; Hong and Kacperczyk 2010; Dhaliwal et al. 2012; Fong et al.
2014). Finally, we have a data field to ensure that the forecasts are for the current fiscal year.
13 Other researchers (Cameron et al. 2009; Kyselý and Kim 2009) use information regarding
disasters from Wikipedia. The presence of any incorrect classifications or omissions of disas-
ters will only increase noise and thus create a bias against finding significant effects.
14 We acknowledge Kee-Hong Bae and Hongping Tan for providing us with location data for
the period 1995–2010. The procedure used to identify analysts’ locations is the same as that
used in Bae et al. (2008a) and Bae et al. (2008b). Using Nelson’s Directories, we manually
check cases with the same analyst, the same research firm and multiple locations. We
exclude observations for which there is insufficient information to clearly identify the loca-
tion of the analyst. The exclusion of these observations does not materially affect our
results.

the websites of the Midwestern Regional Climate Center, Wikipedia, the Heat is
Online, and the National Oceanic and Atmospheric Administration (NOAA)
National Centers for Environmental Information, as well as two journal articles,
Palecki et al. (2001) and Wolfe et al. (2001). The dates of disasters in China over
the period 2005–2014 are from Sigma, an annual international insurance report
published by the Swiss Re Group.15 We use a list of Chinese surnames
obtained from
http://genealogy.familyeducation.com/browse/origin/chinese to identify
Chinese analysts.
We have five dummy variables of interest. First, PRE_3HOLIDAY takes a value
of 1 if the forecast release day falls in a week ending on Thanksgiving Day,
Christmas Day, or New Year’s Day and takes a value of 0 otherwise. Second,
PRE_13HOLIDAY takes a value of 1 if the forecast release day falls on one of two
business days immediately before or on New Year’s Day, Valentine’s Day, Presi-
dents’ Day, St. Patrick’s Day, Easter, Mother’s Day, Memorial Day, Father’s Day,
Independence Day (Fourth of July), Labor Day, Halloween, Thanksgiving or
Christmas and takes a value of 0 otherwise. Third, US_DISASTER takes a value
of 1 if the forecast release day is on the actual day or on any of the 5 days
immediately following a disaster in the United States that involves at least
100 fatalities, causes no significant economic damage, and does not involve ter-
rorism and takes a value of 0 otherwise.16,17,18 The list of these disasters is pro-
vided in Appendix B. Fourth, CN_DISASTER takes a value of 1 if the forecast
release day is on the actual day or on any of the 5 days immediately following a
disaster in China with at least 150 fatalities and takes a value of 0 otherwise.19
15 The reports cover natural catastrophes and man-made disasters, including floods, storms,
earthquakes, droughts, bush fires, heat waves, cold, frost, hail, tsunami, major fires, explo-
sions, aviation disasters, maritime disasters, rail disasters, mining accidents, and building/
bridge collapses.
16 According to Furin (2016), large casualty events are rare in the United States. Historically,
only 10 US disasters have resulted in more than 1000 casualties. The vast majority of US
events have resulted in fewer than 40 fatalities. The Centre for Research on Epidemiology of
Disaster use “ten or more people were killed” as one criterion for determining whether an
event is disaster (Kahn 2005). Sigma classifies a natural or man-made event as a catastrophe
if there are 20 deaths. Ho et al. (2013) suggest that 100 fatalities probably reflect a significant
level of severity perceived by the public and the media, which may also have a psychological
basis. Hence, a minimum of 100 fatalities is likely to be a significant cut-off.
17 We exclude disasters associated with significant economic loss or that are triggered by terror-
ism to avoid contamination by the impacts of possible contemporaneous changes in percep-
tions of uncertainty and/or firm prospects.
18 The 6-day window takes into consideration that news may be released to the public on day
t + 1 and it may take approximately 3–4 days to release a forecast. We thank Peter Joos and
Gilles Hilary for discussing the process of forecast publications.
19 We set a higher fatality hurdle for Chinese disasters for two reasons. First, the number of
fatalities caused by disasters in the United States is typically smaller than that caused by
disasters in China. Second, the proportion of Chinese analysts is relatively small, which
means that we need more significant disasters to identify incremental Chinese effects. None-
theless, as a robustness check, we also consider disasters in China with at least 100 fatalities.

Appendix C provides a list of these disasters. Fifth, OCTOBER takes a value of

1 if the forecast release day is in October and takes a value of 0 otherwise.
The other variables are defined in Appendix D. In the subsequent analysis,
we only consider those observations with no missing values for any variable
used in the baseline regressions. The sample attrition is reported in Appendix E.
In all regressions, all variables except count, time and dummy variables are win-
sorized at 1% and 99% to minimize the influence of outliers and errors in the
data, such as recording errors. We include year, three-digit SIC industry, firm,
and analyst fixed effects in all of our regressions.20 Our standard errors are
based on two-way clustering at the firm and analyst levels.
V. METHODOLOGY
A. The optimism, cognition, and distraction hypotheses

In this section, we describe the methods used to empirically test the hypothe-
ses. Our basic regression models are as follows:
POSERRORk, j, t = a1 + b1 × Zt + c1 × Controls + ek, j, t ð1aÞ

NEGERRORk, j, t = a2 + b2 × Zt + c2 × Controls + ek, j, t ð1bÞ
where subscript k indexes analysts, j indexes firms, and t indexes time. POSER-
ROR is the positive error in percentage terms, which is calculated as the annual
EPS forecast minus actual EPS, scaled by the market share price as of the end of
the last fiscal year, where the forecast is larger than actual EPS. NEGERROR is
the negative error as a percentage, which is calculated in the same manner as
POSERROR when the forecast is smaller than actual EPS. The last terms in both
equations are the error terms. The variable of interest Z is either PRE_3HOLIDAY
or PRE_13HOLIDAY, both of which are defined in Section IV. The optimism
hypothesis predicts that both b1 and b2 are positive. In contrast, the cognition
and distraction hypotheses predict that b1 is positive but b2 is negative.
Next, we take the opposite perspective and consider occasions when people
are likely in a negative mood. Specifically, we modify the models as follows:
POSERRORk, j, t = a1 + b1 × Zt + c1 × Controls + ϵk, j, t ð2aÞ

NEGERRORk, j, t = a2 + b2 × Zt + c2 × Controls + ϵk, j, t ð2bÞ
where the variable of interest Z is US_DISASTER or OCTOBER, which are defined

in Section IV. The optimism hypothesis predicts that both b1 and b2 are nega-
tive; the cognition hypothesis predicts that b1 is negative but b2 is positive;
and the distraction hypothesis for Z = US_DISASTER predicts that b1 is positive
but b2 is negative.
20 The firm and analyst fixed effects are nested within clustering of standard errors.

When the test variable is OCTOBER, we consider two specifications. First, we

only consider observations with fiscal years ending in March or April, such that
October is roughly in the middle of the fiscal year, which minimizes the hori-
zon effect, that is, earlier forecasts with less information about the entire fiscal
year tend to be less accurate whereas later forecasts with more information
about the entire fiscal year tend to be more accurate. Second, as a robustness
check, we contrast October forecasts with November and December forecasts
for all firms. If the horizon effect dominates, the November and December fore-
casts will be more accurate than the October forecasts because there will be
more cumulative information in November and December than in October.
However, if the cognition effect for negative mood dominates, October fore-
casts will be more accurate compared with November and December forecasts.
B. Difference-in-difference method
We further postulate that local people (i.e., those who live in states where disas-
ters occur) are more affected by disasters because they receive more information
about the events and because the events are more relevant to them (Kaplanski
and Levy 2010a). To test this hypothesis, we create a new dummy variable,
LOC_US_DISASTER, which takes a value of 1 if the analyst is located in a state
where and when a disaster occurs and takes a value of 0 otherwise. We then
add this dummy variable as an explanatory variable to equations (2a) and (2b).
This variable captures the differential mood effect of local analysts that is incre-
mental to that of nonlocal analysts. The three hypotheses predict that the coef-
ficients of LOC_US_DISASTER have the same predicted signs as those of
US_DISASTER.
Similarly, we posit that major disasters in China influence Chinese analysts
more than non-Chinese analysts. Empirically, the test variable is CN_DISASTER,
which takes a value of 1 if the forecast is released on the actual day or on any
of the 5 days immediately following a disaster in China with at least 150 fatali-
ties and takes a value of 0 otherwise. Like local versus nonlocal analysts, the
expected CN_DISASTER coefficients for the three hypotheses have a larger mag-
nitude for Chinese analysts than for non-Chinese analysts and have the same
signs as those for US_DISASTER.
C. Sentiment and mood effects

Around Thanksgiving Day, Christmas Day, and New Year’s Day, economic sta-
tistics are generally remarkably better (Barsky and Miron 1989). Meanwhile,
macroeconomic variables are typically strongly contemporaneously correlated
with consumer sentiment (Lemmon and Portniaguina 2006). Therefore, it is
plausible that people have generally higher sentiment during the preholiday
periods because they are affected by the prosperous economic conditions dur-
ing these times. It is also possible that people have coincidently differential sen-
timents associated with economic conditions when there are major disasters,

although this effect is less likely because disasters are typically random. To
investigate whether PRE_3HOLIDAY and US_DISASTER are driven by sentiment
associated with economic conditions, we incorporate an additional sentiment
measure into regressions (1a), (1b), (2a), and (2b) and examine whether PRE_3-
HOLIDAY and US_DISASTER lose explanatory power. In particular, we sepa-
rately consider the Michigan consumer sentiment index (CSENT),21 the
Conference Board consumer confidence index (CCON), and the Baker-Wurgler
investor sentiment index (SENT) as sentiment measures.
D. Under-reaction and mood effects

If managers take advantage of people’s limited attention prior to the holidays
by releasing more bad news during these times (DellaVigna and Pollet 2009),
they are most likely to do so on a Friday, just before the weekend, or on the
day immediately before a holiday, which is when people pay the least atten-
tion, rather than on any earlier days. If analysts are less attentive or more dis-
tracted prior to the holidays, such that they underreact or experience delayed
reactions to bad company news disclosed more during these times, and if the
bad news disclosed during these times is material to the corporate value being
forecasted, this may explain the larger errors—particularly positive forecast
errors—during weeks ending on Thanksgiving Day, Christmas Day, or New
Year’s Day.
To examine whether our results are driven by under-reaction, we add the fol-
lowing dummy variables as additional control variables. 1D_PRE_NON-B (the
business day immediately before a nonbusiness day) takes a value of 1 for the
day immediately before a weekend or a holiday, based on the trading days of
the S&P 500, and takes a value of 0 otherwise. It is important to note that it
may take approximately 3–4 days to release a forecast, although it will probably
take only a couple of hours to release new forecasts when important and mate-
rial news is released. For example, a forecast produced on a Friday may be pub-
lished on the following week, for example, on Wednesday. Hence, we also
include dummies for the third and fourth days after a weekend or a holiday,
that is, 3D_POST_NON-B (the third business day after a nonbusiness day) and
4D_POST_NON-B (the fourth business day after nonbusiness day). If under-
reaction drives the PRE_3HOLIDAY and US_DISASTER effects, then the coeffi-
cients of the corresponding test variables will be insignificant whereas at least
one of the new control variables will have a positive coefficient.
If holidays divert people’s attention away from work, analysts will pay inade-
quate attention and respond less to news released during preholiday periods.
Hence, we test whether the number of forecasts during preholiday periods is sig-
nificantly smaller in relation to the number of news items, as predicted by the
limited attention explanation. We thus study the incremental sensitivity of the
21 Walther and Willis (2013) consider CSENT a measure of investor expectations.

number of daily forecasts (FNUMt) associated with PRE_3HOLIDAY with respect

to the detrended number of daily news items (NEWSNUMt), μ2 in (3) below.22
FNUM t = μ0 + μ1 × NEWSNUM t + μ2 × NEWSNUM t × PRE_3HOLIDAY t

+ μ3 × PRE_3HOLIDAY t + et ð3Þ
Because it is expected that limited attention mainly occurs on the day just
before a weekend or holiday, we conduct the following robustness check. We
partition PRE_3HOLIDAY into several dummy variables in regression (3). PRE
takes a value of 1 if PRE_3HOLIDAY = 1 and the forecast is released on a Friday
or the day before a holiday and takes a value of 0 otherwise.23 FREE takes a
value of 1 if PRE_3HOLIDAY = 1 and the forecast is released on a weekend or
holiday. PRE_3HOLIDAY2 takes a value of 1 if PRE_3HOLIDAY = 1, PRE 6¼
1, and FREE 6¼ 1 and takes a value of 0 otherwise. PRE_3HOLIDAY2 is strictly
defined to include days that are likely to be influenced by a major holiday but
unlikely to be affected by under-reaction to bad news, which is disclosed more
on Fridays and on the days immediately before a holiday. Because the depen-
dent variable (the number of daily forecasts) is a count variable, we employ a
Poisson model. We also estimate a negative binomial model as a robustness
check. The limited attention explanation predicts that the coefficients of
NEWSNUMt × PRE_3HOLIDAYt and NEWSNUMt × PRE_3HOLIDAY2t are
negative.
In addition to the numbers (i.e., the levels) themselves, we also consider the
logarithmic transformation of the numbers LNFNUMt and LNNEWSNUMt,
where the slope coefficient of LNNEWSNUMt in the regression reflects the per-
centage change in the number of forecasts associated with a 1% increase in the
number of news items. The predictions for the limited attention and under-
reaction explanation are exactly the same as those for the level regressions dis-
cussed in the previous paragraph. Because it may take several days to release a
forecast, we also consider lagged numbers of daily news items.
VI. DESCRIPTIVE STATISTICS AND UNIVARIATE ANALYSIS
A. Descriptive statistics
Preholiday periods (PRE_3HOLIDAY) cover days in a calendar week that ends
on Thanksgiving Day, Christmas Day, or New Year’s Day from 1983 to 2014.24
There were 56,736 analyst forecasts of annual EPS released during these periods.
22 We detrend the time series of the number of news items because there is clearly a time trend
for the number of news items.
23 In addition, we create a dummy variable (FRI) exclusively for all forecasts released on Fridays
to examine whether there is generally under-reaction to more bad news disclosed on a
Friday.
24 Forecasts from earlier periods cannot be used because we employ lagged data to construct
certain control variables.

The numbers of annual EPS forecasts in November and December are 168,947
and 108,316, respectively. Hence, forecasts during preholiday mood periods
account for approximately 20.5% of forecasts in November and December.25
PRE_13HOLIDAY has 123,310 observations, accounting for nearly 12% of the
sample. Regarding proxies for negative mood occasions, 17,169 analyst fore-
casts were released on the actual day of or on any of the 5 days immediately fol-
lowing a disaster in the United States with at least 100 fatalities (US_DISASTER),
which amounts to 30.3% of forecasts released in preholiday mood periods.
Local analysts account for 43% of US_DISASTER observations. The correspond-
ing number of forecasts around Chinese disasters with at least 150 fatalities
(CN_DISASTER) is 44,802. Chinese analysts account for 3.6% of these observa-
tions. OCTOBER observations represent approximately 11.3% of the sample.
Table 1 reports the summary statistics, means, standard deviations (SD), 25th
percentiles (25%), medians, and 75th percentiles (75%) of the variables used in
this study. The variables are defined in Appendix D. The mean (median) of pos-
itive and negative forecast errors is 2.45% and −1.11% (0.81% and −0.44%),
respectively.26 Table 2 shows the correlation matrix of the explanatory vari-
ables. Most correlation coefficients are small in magnitude. The two highest cor-
relation coefficients are 0.72527 (between LNSIZE and COVERAGE) and 0.723
(between NCO and NSIC3). The next three strongest correlation coefficients are
0.643 (between SP500 and COVERAGE), 0.633 (between SP500 and LNSIZE),
and −0.496 (between PROFIT and LOSS). All (untabulated) variance inflation
factors for the specifications used are smaller than 10. These values suggest that
multicollinearity is not an issue.
B. Univariate analysis
In this section, we conduct univariate analysis to test the optimism, cognition,
and distraction hypotheses. The major problem using raw forecast data for
these tests is that the horizon, that is, the number of days between the date
when the forecast is released and date of the fiscal year end for which earnings
are being forecast, is systematically different between PRE_3HOLIDAY and
NON-PRE_3HOLIDAY. Over 70% of the observations have fiscal year ends in
December, which is closer to the time when forecasts are released during preho-
liday periods than to the time when forecasts are released during the other
25 The number of forecasts during preholiday periods is approximately 3% of the total number
of forecasts in the sample.
26 The untabulated mean (median) of forecast signed error, which is calculated by subtracting
the realized value from the forecast value, and absolute error, which is the absolute differ-
ence between the forecast value and the realized value, is 0.72% and 1.84% (0.019% and
0.554%), respectively. These statistics are consistent with certain summary statistics in the
literature (e.g., Bushee et al. 2010, with a median absolute error of 0.4%, and Walther and
Willis 2013, with a mean error of 0.76%).
27 This is highly comparable to the estimated correlation coefficient between Log(Assets) and
Number of Analysts, 0.72, in Chang et al. (2006).

Table 1 Summary statistics

Mean SD 25% Median 75%
POSERROR 2.452 4.587 0.257 0.806 2.349
NEGERROR −1.107 1.821 −1.159 −0.437 −0.162
COVERAGE 20 12 11 18 27
LNSIZE 7.738 1.735 6.554 7.710 8.917
LNBM −0.820 0.688 −1.230 −0.760 −0.350
RET 0.014 0.034 −0.005 0.013 0.030
SIGMA 0.105 0.057 0.066 0.091 0.129
VOLROE 0.072 0.359 0.002 0.006 0.022
PROFIT 0.131 0.218 0.067 0.142 0.214
SP500 0.451 0.498 0.000 0.000 1.000
MNMD −0.001 0.021 −0.004 0.000 0.005
LOSS 0.073 0.260 0.000 0.000 0.000
FIRM_MERROR 0.096 0.608 −0.060 0.000 0.110
ANALYS_MERROR 0.044 0.224 −0.030 0.001 0.060
HORIZON (days) 184 97 90 176 260
RETTODATE 0.030 0.265 −0.080 0.033 0.167
NCO 29 100 13 17 24
NSIC3 8 16 3 5 9
EXPERIENCE_WITH_FIRM (years) 3.5 3.5 1.1 2.4 4.9
GENERAL_EXPERIENCE (years) 7.2 5.5 2.9 5.8 10.1
Notes: This table reports summary statistics. POSERROR (NEGERROR), in percentage terms, is fore-
cast EPS for a given firm j in month m in fiscal year t by analyst k, minus actual EPS [where the
forecast is larger (smaller) than actual EPS] over the market share price at the end of fiscal year t
−1. COVERAGE is the number of analysts covering firm j in fiscal year t−1. LNSIZE is the natural
logarithm of firm j’s market capitalization at the end of fiscal year t−1. LNBM is the natural loga-
rithm of firm j’s book value divided by its market capitalization at the end of fiscal year t−1. RET
is the average monthly stock returns for firm j for the past 12 months in relation to the data of
the last actual annual earnings. SIGMA is the standard deviation of the raw monthly stock
returns for firm j for the past 12 months in relation to month m. VOLROE is the variance of the
residuals from an AR(1) model for firm j’s annual ROE using the past 10-year series. ROE is calcu-
lated as the ratio of earnings to the beginning book value of equity. PROFIT is the operating
income of firm j for fiscal year t−1 over the book value of assets of the firm as of the end of fiscal
year t−2. SP500 equals one if firm j is in the S&P 500 index on the day that the forecast is released
and equals 0 otherwise. MNMD is the difference between the mean and median of EPS, scaled by
the market share price for the end of fiscal year t−1 for firm j for the period between fiscal year t
−4 and fiscal year t + 4, excluding fiscal year t. LOSS equals one if the forecast made by analyst
k is negative and equals 0 otherwise. FIRM_ MERROR is the median error of all forecasts for firm
j for fiscal year t−1. ANALYS_MERROR is the median error of all forecasts for analyst k for fiscal
year t−1. HORIZON is the number of days between the release of analyst k’s earnings forecast for
firm j and the data of the earnings being forecast. RETTODATE is the cumulative stock returns
(using monthly data) for firm j between the last annual earnings and the release of the earnings
forecast by analyst k. NCO is the number of firms for which analyst k has generated annual EPS
forecasts in fiscal year t. NSIC3 is the number of three-digit SIC industries for which analyst k has
generated annual EPS forecasts in fiscal year t. EXPERIENCE_WITH_FIRM is the number of years
since the release of the first annual EPS forecast for firm j generated by analyst k. GENERAL_EX-
PERIENCE is the number of years since the release of the first annual EPS forecast generated by
analyst k.

16
Table 2 Correlation matrix of explanatory variables
PRE_3HOLIDAY (2) (3) (4) (5) (6) (7) (8) (9) (10)
US_DISASTER (2) −0.001
COVERAGE (3) 0.000 −0.005
LNSIZE (4) −0.011 −0.007 0.725
LNBM (5) 0.004 −0.010 −0.168 −0.357
RET (6) 0.003 −0.003 −0.055 0.066 −0.344
SIGMA (7) −0.004 0.017 −0.178 −0.310 −0.019 −0.018
VOLROE (8) −0.005 −0.004 −0.067 −0.072 −0.162 0.027 0.157
PROFIT (9) 0.006 0.004 0.147 0.272 −0.140 0.169 −0.335 −0.134
LOSS (10) 0.003 −0.004 −0.111 −0.222 0.045 −0.147 0.384 0.171 −0.496
MNMD (11) 0.000 0.005 0.070 0.134 −0.119 0.074 −0.111 −0.023 0.147 −0.135
SP500 (12) 0.008 0.003 0.643 0.633 −0.135 −0.046 −0.203 −0.071 0.152 −0.124
FIRM_MERROR (13) 0.001 −0.002 −0.033 −0.148 0.159 −0.267 0.168 0.007 −0.299 0.231
ANALYS_MERROR (14) 0.002 0.001 −0.019 −0.111 0.175 −0.228 0.122 −0.009 −0.149 0.141
HORIZON (15) −0.284 0.001 0.015 0.025 −0.018 −0.009 0.012 0.004 0.000 −0.042
RETTODATE (16) −0.013 0.020 0.017 −0.028 0.094 −0.041 −0.085 −0.022 0.022 −0.148
NCO (17) 0.007 0.009 −0.001 −0.030 0.027 0.009 −0.022 −0.010 0.005 −0.014
NSIC3 (18) 0.013 0.009 −0.037 −0.067 0.026 0.010 −0.026 −0.022 0.024 −0.032
LNEXPERIENCE_ 0.002 −0.002 0.119 0.149 0.006 −0.052 −0.066 −0.031 0.024 −0.014
WITH_FIRM (19)
LNGENERAL_ 0.008 0.002 −0.021 0.147 −0.045 −0.024 −0.017 0.006 0.016 −0.004
EXPERIENCE (20)

Table 2 (continued)
(11) (12) (13) (14) (15) (16) (17) (18) (19)
US_DISASTER (2)
COVERAGE (3)
LNSIZE (4)
LNBM (5)
RET (6)
SIGMA (7)
VOLROE (8)
PROFIT (9)
LOSS (10)

MNMD (11)
SP500 (12) 0.051
FIRM_MERROR (13) −0.228 −0.030
ANALYS_MERROR (14) −0.089 0.001 0.378
HORIZON (15) 0.003 −0.004 −0.005 −0.011
RETTODATE (16) 0.043 0.024 −0.063 −0.042 −0.013
NCO (17) −0.006 0.003 0.005 −0.014 −0.010 0.009
NSIC3 (18) 0.002 0.006 0.013 0.012 −0.017 0.013 0.723
LNEXPERIENCE_WITH_FIRM (19) 0.020 0.126 −0.004 −0.024 −0.046 0.002 −0.003 −0.004
LNGENERAL_EXPERIENCE (20) 0.023 0.041 −0.042 −0.044 −0.06 0.013 0.080 0.078 0.424
Note: This table reports the correlation coefficients between the explanatory variables. Refer to Table 1 for variable definitions.
17
periods. Hence, forecasts for PRE_3HOLIDAY typically have a shorter horizon

than those for NON-PRE_3HOLIDAY. Studies suggest that forecasts made on a
longer horizon are generally more optimistic because, for example, analysts
trade positive forecast bias (i.e., less accurate forecasts) for access to manage-
ment when information about companies’ earnings prospects is less readily
available (Lim 2001; Beyer 2008). The shorter the horizon, the more informa-
tion has been revealed over time. Therefore, forecasts issued closer to the fiscal
year end will be more accurate. As suggested by Byard et al. (2011), Jacob
et al. (1999), and Clement and Tse (2003), to adjust the systematic difference in
horizon between PRE_3HOLIDAY and NON-PRE_3HOLIDAY observations, we
calculate horizon-adjusted positive (negative) errors as the residuals from the
regression of raw/unadjusted positive (negative) errors on LNHORIZON.
Panel A of Table 3 reports the test results based on these horizon-adjusted
forecast performance variables. For the overall sample, consistent with all three
hypotheses, the positive errors for PRE_3HOLIDAY are, on average, more posi-
tive than those for NON-PRE_3HOLIDAY. The difference in the horizon-
adjusted bias is 0.465%, which is significant at the 1% level.28 Consistent with
the cognition and distraction hypotheses, the negative errors for PRE_3HOLI-
DAY are generally more negative than those for NON-PRE_3HOLIDAY; the dif-
ference is −0.225%, which is significant at the 1% level.
We next consider negative mood. The results are shown in Panel B. In line
with the predictions of the optimism and cognition hypotheses, the positive
errors for US_DISASTER are generally more negative than those for NON-US_DI-
SASTER; the difference is −0.488%, which is significant at the 1% level. Support-
ing the cognition hypothesis alone, the negative errors for US_DISASTER are
typically more positive than those for NON-US_DISASTER; the difference is
0.210%, which is significant at the l% level.
Furthermore, local people are plausibly in more negative moods compared
with nonlocal people when there is a disaster. Therefore, we expect that the
disaster mood effect on forecast precision is more pronounced among local ana-
lysts than among nonlocal analysts. The results are reported in Panel B1. Con-
sistent with the above expectation, local analysts’ forecasts are more negative
for positive errors and more positive for negative errors compared with the fore-
casts of nonlocal analysts; the differences are −0.395% and 0.164%, which are
both significant at the 1% level.
Likewise, Chinese analysts are likely to be in more negative moods compared
with non-Chinese analysts when a major disaster occurs in China. Therefore,
we expect that the Chinese disaster-induced mood effect on forecast precision
is stronger among Chinese analysts than among non-Chinese analysts. Panel
B2 displays results consistent with this expectation.29 Chinese analysts’
28 Because the horizon-adjusted variables are regression residuals, their average value is zero.
29 Psychology studies (e.g., Diener et al. 1995) suggest that the Chinese generally have more
negative emotions compared with White Americans. Therefore, we allow for an average dif-
ferential negative error for Chinese analysts by including a dummy with a value of 1 for Chi-
nese analysts and a value of 0 for non-Chinese analysts.

Table 3 Univariate tests of analysts’ positive and negative errors
Horizon-adjusted POSERROR Horizon-adjusted NEGERROR
(1) (2) (1)–(2) (3) (4) (3)–(4)
Panel A: PRE_3HOLIDAY versus NON-PRE_3HOLIDAY

Full sample PRE_3HOLIDAY NON-PRE_3HOLIDAY DIFF PRE_3HOLIDAY NON-PRE_3HOLIDAY DIFF
0.453 −0.012 0.465*** −0.219 0.006 −0.225***
Panel B: US_DISASTER versus NON-US_DISASTER

Full sample US_DISASTER NON-US_DISASTER DIFF US_DISASTER NON-US_DISASTER DIFF
−0.484 0.004 −0.488*** 0.208 −0.002 0.210***
Panel B1: Local analyst versus Nonlocal analyst

1995–2010 US_DISASTER Local analysts Nonlocal analysts DIFF Local analysts Nonlocal analysts DIFF

−0.742 −0.347 −0.395*** 0.302 0.138 0.164***
Panel B2: Chinese analyst versus Non-Chinese analyst

2005–2014 CN_DISASTER Chinese analysts Non-Chinese analysts DIFF Chinese analysts Non-Chinese analysts DIFF
−0.514 −0.321 −0.193**** 0.335 0.112 0.223***
Notes: This table reports the results of t-tests where POSERROR (NEGERROR) differs between PRE_3HOLIDAY and NON-PRE_3HOLIDAY observations,
between US_DISASTER and NON-US_DISASTER observations, between local analyst US_DISASTER and nonlocal analyst US_DISASTER observations, and
between Chinese analyst CN_DISASTER and non-Chinese analyst CN_DISASTER observations. Horizon-adjusted POSERROR (NEGERROR) is the residual
from the regression of unadjusted/raw POSERROR (NEGERROR) on LNHORIZON, where unadjusted/raw POSERROR (NEGERROR), in percentage form, is
the forecast EPS minus the actual EPS [where the forecast is larger (smaller) than the actual EPS] over the market share price at the end of the last fiscal
year. LNHORIZON is the natural logarithm of 1 plus the number of days between the release of analyst k’s earnings forecast for firm j and the data date of
the earnings being forecast. PRE_3HOLIDAY takes a value of 1 if the forecast release day is in a week ending on Thanksgiving Day, Christmas Day, or New
Year’s Day and takes a value of 0 otherwise. US_DISASTER takes a value of 1 if the forecast release day is on the day of or on any of the 5 days immediately
following the day when a disaster occurs in the United States that involves at least 100 fatalities, causes no significant monetary damage, and does not
involve terrorism and takes a value of 0 otherwise. CN_DISASTER takes a value of 1 if the forecast release day is on the day of or on any of the 5 days
immediately following the day when a disaster occurs in China that involves at least 150 fatalities and takes a value of 0 otherwise. Local analysts are ana-
lysts located in a state where a disaster occurs. Chinese analysts are analysts with Chinese surnames. *** and **** indicate the 1% and one-sided 10% levels
of significance, respectively.
19
forecasts are more negative for positive errors and more positive for negative
errors compared with the forecasts of non-Chinese analysts; the differences are
−0.193% and 0.223%. The latter is significant at the 1% level.
One reason why forecasts for PRE_3HOLIDAY have larger errors than those
for NON-PRE_3HOLIDAY may relate to differences in analyst characteristics. For
example, it is possible that senior analysts, who have higher professional stan-
dards, avoid making forecasts during the holiday periods when they are more
distracted and less attentive. In contrast, junior analysts may be more active
during these quiet times, when it is easier to attract attention and thus poten-
tially to increase their profiles. Due to career concerns, forecasts made by these
junior analysts will tend to be more optimistic. In addition, there will be fewer
senior analysts generating useful information during these holiday periods,
which means that forecasts made during these times will have larger errors.
Consequently, we will observe less accuracy for PRE_3HOLIDAY than for NON-
PRE_3HOLIDAY. Therefore, we investigate whether analyst characteristics differ
between PRE_3HOLIDAY observations and NON-PRE_3HOLIDAY observations.
Panel A of Table 4 shows that analysts who release forecasts during preholi-
day periods generally cover more firms and more industries, are slightly more
likely to be a star analyst, and have similar general and firm-specific forecasting
experience.30 It is plausible that PRE_3HOLIDAY analysts are busier, given that
their coverage is broader (4.4 more firms and 1.2 more three-digit SIC indus-
tries). However, there is no evidence that the PRE_3HOLIDAY analysts are either
forecast followers or more junior than those for OTHER (i.e., NON-PRE_3HOLI-
DAY and NON-US_DISASTER). In addition, analysts releasing forecasts following
major US disasters generally cover more firms and more industries, are slightly
more likely to be a star analyst, and have similar general and firm-specific fore-
casting experience. Nonetheless, analyst characteristics are taken into consider-
ation in the subsequent regression analyses.
Regarding firm characteristics, we find that PRE_3HOLIDAY, US_DISASTER,
and OTHER firms generally have similar characteristics, with the exception of
the size and distribution of certain industries. As shown in Panel B of Table 4,
PRE_3HOLIDAY and US_DISASTER firms are generally slightly smaller than
OTHER firms. Nevertheless, we control for firm size in all of our regressions. We
also conduct a robustness check for the results of the baseline regression based
on firms with a share price above $5. With respect to industry distribution, as
shown in Panel C of Table 4, PRE_3HOLIDAY firms are more often in the retail
industry compared with OTHER firms (14.57% versus 10.01%) and less often in
the finance (12.09% versus 14.73%) and services industries (9.45% versus
10.94%). Holiday season sales are probably important for the retail industry,
and there are more material updates during the preholiday periods for this
industry. Meanwhile, there is no noticeable difference in industry distribution
between US_DISASTER firms and OTHER firms, except that the former group
30 We thank Ling Cen for providing us with the data for regarding star analysts, which cover
the 2002–2012 period.

includes slightly more manufacturing firms (44.42% versus 41.97%). We

include three-digit SIC fixed effects in all of our regressions.
VII. EMPIRICAL RESULTS
A. Baseline results
Table 5 reports our baseline results for PRE_3HOLIDAY and US_DISASTER. In
columns 1 and 2, consistent with the optimism, cognition, and distraction
hypotheses, the estimated coefficients of PRE_3HOLIDAY are positive and sig-
nificant at the 1% level for POSERROR regressions. In columns 3 and 4, in line
with the cognition and distraction hypotheses, the estimated coefficients of
PRE_3HOLIDAY are negative and significant at the 1% level for NEGERROR
regressions. In column 2, in line with the optimism and cognition hypotheses,
the estimated US_DISASTER coefficient is negative and significant at the 1%
level for the POSERROR regression. In column 4, as predicted by the cognition
hypothesis, the estimated US_DISASTER coefficient is positive and significant at
the 1% level for the NEGERROR regression. Overall, the results support the cog-
nition hypothesis, that is, analysts in a more positive mood are less analytical
and make less accurate forecasts, whereas analysts in a more negative mood
process information more systematically and substantively and therefore make
more precise forecasts.
These coefficients are also economically significant. The preholiday incre-
mental positive and negative errors are, on average, 0.51% and −0.31%, repre-
senting 21% and 28% of the unconditional mean positive and negative
errors.31 Hong and Kacperczyk (2010) report an estimated increase of 0.13% in
mean forecast optimism for a decrease in coverage by one analyst. In contrast,
the reduction in positive and negative errors associated with disasters is typi-
cally −0.31% and 0.11%, equivalent to 12% and 10% of the magnitude of the
corresponding unconditional means. Dolvin et al. (2009) find an increase of
0.0220% to 0.0302% in forecast accuracy during SAD periods. Dhaliwal
et al. (2012) report an improvement in forecast accuracy of 0.435% for firms
that make corporate social responsibility disclosures compared with those that
do not make such disclosures. Regarding the explanatory power or goodness of
fit of our baseline models, the adjusted R2s, 0.505 and 0.472, are comparable to
those in Walther and Willis (2013), that is, 0.356 and 0.500, and are generally
higher than those in other studies (e.g., 0.23–0.28 in Liang and Riedl 2014;
0.071–0.081 in So 2013; and 0.036–0.124 in Gu and Wu 2003).
Table 6 reports results for PRE_13HOLIDAY and OCTOBER. Consistent with
PRE_3HOLIDAY in Table 5, PRE_13HOLIDAY in columns 1 and 4 in Table 6
31 Kasser and Sheldon (2002) suggest that older people engage more and thus have happier
holidays compared with younger people. Hence, we expect optimism associated with PRE_3-
HOLIDAY to be more pronounced among more senior analysts. This expectation is con-
firmed when seniority is measured by GENERAL_EXPERIENCE (the number of days for
which the analyst has been a forecaster) in logarithmic form.

show that forecasts both higher and lower than actual earnings are less accurate
prior to or on the 13 holidays. Consistent with US_DISASTER in Table 5, OCTO-
BER in columns 2, 3, 5, and 6 in Table 6 show that analyst forecasts are more
accurate in October than in the other 11 months, including November and
December. Hence, these results further support the cognition hypothesis.
B. Local versus nonlocal analysts

We expect major disasters to have a greater impact on local people (i.e., those
who are located in the area where a disaster occurs) than on people in other
places because of greater media coverage, more information, more relevance,
and greater attention. Therefore, we predict that when major US disasters occur,
local analysts (i.e., those in states where disasters occur) are normally more
affected. Hence, the LOC_US_DISASTER coefficients, which capture the incre-
mental disaster effects on local analysts, should have the same signs as the
US_DISASTER coefficients. The results are shown in Table 7. Consistent with
the cognition hypothesis, the estimated coefficients of LOC_US_DISASTER are
significantly negative and positive for the POSERROR and NEGERROR regres-
sions, respectively. Similarly, US_DISASTER remains significantly negative and
positive for the POSERROR and NEGERROR regressions, respectively. Hence,
disaster-induced accuracy is more pronounced among local analysts.
We perform several robustness checks and obtain similar results.32 In particu-
lar, we add alternative market-/economic-based sentiment controls, the Michi-
gan consumer sentiment index, and the Baker-Wurgler investor sentiment
index. Moreover, we include and exclude the 1994 disaster, for which we have
less analyst location data. In addition, when constructing LOC_US_DISASTER,
we include and exclude the states where the 1999 heat wave occurred but is
mentioned in only one source. Finally, we consider only the set of firms that
have forecasts from both local and nonlocal analysts when disasters occur.
C. Chinese versus non-Chinese analysts

Similar to the above discussion, we expect major disasters in China to affect
Chinese analysts more than non-Chinese analysts. Hence, although the CN_DI-
SASTER coefficients should have the same signs as the US_DISASTER
coefficients, their magnitudes should be larger for Chinese analysts than for
non-Chinese analysts. Table 8 reports the results. In line with the cognition
hypothesis, the estimated coefficients of CN_DISASTER are significantly nega-
tive and positive for the POSERROR and NEGERROR regressions, respectively,
for both Chinese and non-Chinese analysts. Consistent with expectations, the
magnitude of the coefficients is larger for Chinese analysts. The CN_DISASTER
difference in negative errors between Chinese and non-Chinese analysts is
32 These results are not tabulated but are available from the authors upon request.

significant at 5%. As a robustness check, we consider disasters in China with at

least 100 fatalities and obtain similar results.
D. Sentiment and mood effects

Table 9 shows the results after controlling for economic-/market-related senti-
ment. The PRE_3HOLIDAY coefficients have the same signs and are as statisti-
cally and economically significant as they were before, with comparable
magnitude. They range from 0.597% to 0.649% (0.512% in Tables 5 and 6) for
POSERROR and from −0.310% to −0.324% (−0.306% in Tables 5 and 6) for
NEGERROR. The US_DISASTER coefficients also remain similar to those in
Tables 5 and 6. Hence, holiday inaccuracy and disaster precision effects are not
driven by economic-associated sentiment. In line with the notion that analysts
are more optimistic when they have higher sentiment due to better economic/
market conditions, the estimated coefficients of all three-sentiment measures
are generally positive.
Because economic conditions are typically poorer during contractions, ana-
lysts’ moods may be adversely affected. Hence, we further allow for differential
mood effects during contractions. Consistent with expectations, contractions’
incremental mood effects (C × PRE_3HOLIDAY) are qualitatively similar to
those of US_DISASTER. Meanwhile, standalone contractions (C) are associated
with larger errors, likely as a result of higher uncertainty during these times.
E. Under-reaction and mood effects

The under-reaction explanation suggests that the positive forecast bias during
preholiday periods is driven by analysts’ under-reaction to more bad company
news released just before weekends or holidays. To examine whether our PRE_3-
HOLIDAY results are driven by under-reaction, we control for effects on Friday
or the day immediately before a holiday using the dummy variable 1D_PRE_-
NON-B, which takes a value of 1 if the forecast is released on the day just before
a nonbusiness day and takes a value of 0 otherwise. Because it may take
3–4 days to release forecasts, we also control for the effects of 3–4 days after a
nonbusiness day by including two dummy variables, 3D_POST_NON-B and
4D_POST_NON-B, for the third and fourth days, respectively, following a non-
business day. We study whether PRE_3HOLIDAY effects disappear with the pres-
ence of these additional control variables. Table 10 reports the results. We find
that the estimated coefficients of PRE_3HOLIDAY have the same signs and high
level of significance as they did before. These estimated coefficients are nearly
the same as before, within the range of 0.599% to 0.652% (0.597% to 0.649%
in Table 9) for the POSERROR regressions and within the range of −0.312% to
−0.327% (−0.310% to −0.324% in Table 9) for the NEGERROR regressions. These
results suggest that our PRE_3HOLIDAY effects are not driven by the under-
reaction of analysts to bad news on the days just before weekends or holidays.

Table 4 Analyst characteristics, firm characteristics, and industry distribution
24
PRE_3HOLIDAY PRE_3HOLIDAY US_DISASTER US_DISASTER OTHER OTHER
Mean Median Mean Median Mean Median
Panel A: Analyst characteristics
FIRST FORECAST DAY 102.0 65 92 68 97 67
GENERAL_EXPERIENCE 7.2 5.9 7.2 5.8 7.2 5.8
EXPERIENCE_WITH_FIRM 3.6 2.6 3.7 2.3 3.5 2.4
NCO 33.5 18 40.1 18 29.1 17
NSIC3 9.2 6 9.7 6 8.0 5
STARS (proportion) 10.5% 10.7% 9.5%
Panel B: Firm characteristics

COVERAGE 19.80 19 19.05 18 19.76 18
SIZE 8548 2070 8409 2065 9644 2222
BM 0.545 0.474 0.507 0.444 0.545 0.467
RET 0.014 0.013 0.012 0.012 0.014 0.013
SIGMA 0.104 0.091 0.117 0.099 0.105 0.091
VOLROE 0.063 0.005 0.058 0.006 0.075 0.006
PROFIT 0.139 0.146 0.140 0.150 0.128 0.141
MNMD −0.001 0.000 0.000 0.000 −0.001 0.000
FIRM_MERROR 0.099 0.000 0.088 0.003 0.098 0.000
PRE_3HOLIDAY US_DISASTER OTHER
Proportion (%) Proportion (%) Proportion (%)

Panel C: Industry distribution
Agriculture, forestry, and fishing (SIC = 0xxx) 0.13 0.14 0.13
Mining (SIC = 10xx–14xx) 7.49 8.23 7.92
Construction (SIC = 15xx–17xx) 0.87 1.03 0.99
Manufacturing (SIC = 20xx–39xx) 41.94 44.42 41.97
Transportation, communications, and utilities (SIC = 4xxx) 10.84 11.65 10.98
Wholesale Trade (SIC = 50xx or 51xx) 2.33 2.02 2.12

Table 4 (continued)
PRE_3HOLIDAY US_DISASTER OTHER
Proportion (%) Proportion (%) Proportion (%)
Panel C: Industry distribution
Retail Trade (SIC = 52xx–59xx) 14.57 9.33 10.02
Finance, insurance, and real estate (SIC = 6xxx) 12.09 13.61 14.74

Services (SIC = 70xx–89xx) 9.45 9.28 10.96
Nonclassifiable firms (SIC = 99xx) 0.28 0.28 0.17
Notes: This table reports summary statistics of analyst characteristics, firm characteristics, and industry distribution of PRE_3HOLIDAY, US_DISA-
STER, and OTHER (i.e., NON-PRE_3HOLIDAY and NON-US_DISASTER) observations. FIRST FORECAST DAY is the number of days between the
release date of an analyst’s first forecast for a particular firm for a particular fiscal year and the beginning of the fiscal year. STARS is the propor-
tion of forecasts that are made by star analysts. STARS data are limited to the 2002–2012 period. SIZE (in millions of dollars) is firm j’s market cap-
italization at the end of fiscal year t−1. BM is firm j’s book value divided by its market capitalization at the end of fiscal year t−1. Refer to Table 1
for the definitions of the other variables.
25
In addition, we consider a prediction based on the limited attention explana-

tion by testing whether analysts react less to news during preholiday periods,
thereby generating less accurate forecasts. We thus study the incremental sensi-
tivities of the number of forecasts with respect to the number of news items
associated with PRE_3HOLIDAY and separately with PRE_3HOLIDAY2 (the
“strict” preholiday periods). Columns 4–6 of Panel A in Table 11 report the
results. We also consider the natural logarithmic transformation of the num-
bers. The corresponding results are shown in columns 1–3 of Panel A in
Table 11. The limited attention explanation predicts that the incremental sensi-
tivities are negative (i.e., for a given increase in the number of news items, there
will be a smaller increase in the number of forecasts). However, we find that the
differential sensitivities associated with both PRE_3HOLIDAY and PRE_3HOLI-
DAY2 are insignificant, with magnitudes close to zero. Hence, the number of
forecasts in relation to the number of news items is not significantly smaller for
PRE_3HOLIDAY and PRE_3HOLIDAY2, suggesting a lack of evidence for limited
attention during the holiday periods and “strict” preholiday periods.33,34,35,36
F. Robustness checks and additional analyses

We perform several robustness checks and additional analyses. First, we con-
sider three alternative nearby non-PRE_3HOLIDAY periods37 and compare posi-
tive and negative forecast errors during each of these periods with those of
PRE_3HOLIDAY periods. Second, we examine whether the PRE_3HOLIDAY
33 We have additional evidence that contradicts the limited attention explanation. First, we
have qualitatively the same incremental findings for PRE as for PRE_3HOLIDAY2. Second,
there is significantly more news sensitivity on Fridays. The magnitude is also large, at 2.47 –
2.67 times the sensitivity of other days. Hence, there are greater news reactions on Friday.
34 The number of contemporaneous and lagged news items is highly correlated. Hence, we esti-
mate their impacts in separate regressions. We consider the number of one-, two-, three-,
four-, and five-day lags of news items. We find that the number of three- and four-day lags
of news items is most positively related to the number of forecasts (with the largest and
most significant coefficients and highest R2s), which is consistent with the notion that it
typically takes 3–4 days to release a forecast. The results for four-day lags are presented in
Panel B of Table 11. The results for one-, two-, three-, and five-day lags are not tabulated but
are available from the authors upon request.
35 We also find that the estimated coefficients of the interaction between FRI and NEWNUM
and the interaction between FRI and LNNEWNUM are significantly positive. These results
are contrary to the under-reaction explanation, which predicts negative coefficients.
36 The results of the negative binomial models are essentially the same as those of the Poisson
models and thus are not tabulated. These results are available from the authors upon
request.
37 The alternative non-PRE_3HOLIDAY periods are as follows. The NOVEMBER non-PRE_3HOLI-
DAY period covers forecasts released in November but not during PRE_3HOLIDAY periods.
The JANUARY non-PRE_3HOLIDAY period covers forecasts released in January but not during
PRE_3HOLIDAY periods. The final BNA (before and after) alternative non-PRE_3HOLIDAY
period covers forecasts released between the week immediately before the week ending on
Thanksgiving Day and the week immediately after the week ending on New Year’s Day,
inclusively, but not during PRE_3HOLIDAY periods.

Table 5 Preholidays, disasters, and positive and negative analyst errors

(1) (2) (3) (4)
POSERROR POSERROR NEGERROR NEGERROR
PRE_3HOLIDAY 0.479*** 0.512*** −0.289*** −0.306***
(0.039) (0.039) (0.021) (0.021)
US_DISASTER −0.305*** 0.107***
(0.050) (0.024)
COVERAGE 0.015*** 0.014*** −0.007*** −0.007***
(0.005) (0.005) (0.002) (0.002)
LNSIZE −1.047*** −1.046*** 0.671*** 0.679***
(0.075) (0.072) (0.029) (0.029)
LNBM 0.936*** 0.929*** −0.149*** −0.146***
(0.094) (0.091) (0.031) (0.031)
RET −7.630*** −8.330*** 0.193 0.212
(0.929) (0.907) (0.344) (0.338)
SIGMA 7.636*** 7.214*** −3.288*** −3.072***
(0.699) (0.702) (0.281) (0.278)
VOLROE −0.174**** −0.174**** −0.027 −0.029
(0.113) (0.110) (0.045) (0.044)
PROFIT −0.441** −0.509*** 0.435*** 0.485***
(0.192) (0.177) (0.077) (0.076)
SP500 0.468*** 0.432*** −0.209*** −0.205***
(0.128) (0.125) (0.047) (0.046)
MNMD 4.985* 5.823** 4.094*** 4.039***
(2.757) (2.744) (1.022) (1.030)
LOSS 0.475*** 0.434*** −1.045*** −1.040***
(0.132) (0.131) (0.076) (0.076)
FIRM_MERROR 0.450*** 0.597*** −0.142*** −0.146***
(0.087) (0.090) (0.030) (0.032)
ANALYS_MERROR 0.532*** 0.571*** 0.028 −0.012
(0.130) (0.117) (0.045) (0.043)
LNHORIZON 0.911*** 0.921*** −0.399*** −0.404***
(0.024) (0.024) (0.012) (0.012)
RETTODATE −0.723*** −0.731*** −0.128*** −0.126***
(0.086) (0.085) (0.031) (0.030)
NCO −0.001* −0.001* 0.000**** 0.000****
(0.000) (0.000) (0.000) (0.000)
NSIC3 0.008* 0.008* −0.002 −0.001
(0.005) (0.005) (0.002) (0.001)
LNEXP_WITH_FIRM 0.018*** 0.018*** 0.002 0.002
(0.005) (0.005) (0.002) (0.002)
LNGENERAL_EXP −0.098*** −0.094*** 0.013 0.009
(0.030) (0.030) (0.013) (0.014)
Observations 1,041,230 1,041,230 959,518 959,518
Adjusted R2 0.503 0.505 0.472 0.472
Year fixed effects Y Y Y Y
Three-digit SIC fixed effects Y Y Y Y

Table 5 (continued)
(1) (2) (3) (4)
Firm fixed effects Y Y Y Y
Analyst fixed effects Y Y Y Y
SE clustering Firm Firm Firm Firm
Analyst Analyst Analyst Analyst
Notes: This table reports the results of baseline regressions regarding the incremental effects of
three major holidays and disasters on analysts’ positive and negative errors. The dependent vari-
able is either POSERROR or NEGERROR. POSERROR (NEGERROR), in percentage terms, is the fore-
cast EPS for a given firm j in month m in fiscal year t by analyst k minus the actual EPS [where
the forecast is larger (smaller) than the actual EPS] over the market share price at the end of fiscal
year t−1. PRE_3HOLIDAY takes a value of 1 for forecast release days that fall within a week ending
on Thanksgiving Day, Christmas Day, or New Year’s Day and takes a value of 0 otherwise. US_DI-
SASTER takes a value of 1 if the forecast release day is on the day of or on any of the 5 days
immediately following the day when a disaster occurs in the United States that involves at least
100 fatalities, causes no significant monetary damage, and does not involve terrorism and takes a
value of 0 otherwise. Refer to Table 1 for the definitions of the other variables. NCO and NSIC3
are in hundreds. Estimated coefficients and the robust standard errors are reported in parenthe-
ses. ***, **, *, and **** indicate the 1%, 5%, 10%, and one-sided 10% levels of significance,
respectively.
effects are stronger when the forecast is released on a day closer to a holiday. Third,
we rerun the baseline regressions only for firms with a share price above $5. Fourth,
because forecasts were often delivered to I/B/E/S in batches, not daily, before 1994
and thus the forecast publication dates for pre-1994 forecasts may be inaccurate
(Hilary and Hsu 2013), we rerun the baseline regressions only for forecasts released
in and after 1994. Fifth, because Regulation Fair Disclosure was promulgated in
August 2000 and the Global Research Analyst Settlement was finalized on April
28, 2003, we test whether the PRE_3HOLIDAY effects on inaccuracy remain for fore-
casts released after August 2000 or April 2003. Sixth, we consider personal income
tax shocks as another mood proxy.38 Finally, we study other holidays (considering
three holidays, Independence Day, Martin Luther King Day, and Memorial Day,
together and then considering only the latter two holidays together). Our results are
robust with respect to these checks.
Although it is generally thought that people are in more negative moods
in January (the January blues) because the holidays are behind them (post-
Christmas blues), the nights are long and the weather is cold, some studies
suggest that people are more optimistic in January (Hirshleifer et al. 2016).
Accordingly, we also examine forecast accuracy in January relative to other
months.39 To neutralize the horizon effect, we only consider earnings forecasts
for fiscal years ending in June or July, such that January is approximately in the
middle of the fiscal year. We find that positive forecast errors are less positive
and negative forecast errors are more positive in January than in other months.
38 Our data for personal income tax shocks come from Mertens and Ravn (2013).
39 The results are not tabulated but are available from the corresponding author upon request.

Table 6 Preholidays, October, and positive and negative analyst errors
(1) (2) (3) (4) (5) (6)
POSERROR POSERROR POSERROR NEGERROR NEGERROR NEGERROR
PRE_13HOLIDAY 0.149*** −0.053***
(0.017) (0.008)
OCTOBER −0.172** −0.199*** 0.161*** 0.125***
(0.068) (0.033) (0.032) (0.016)
NOVDEC 0.071** −0.056***
(0.036) (0.019)
COVERAGE 0.015*** 0.003 0.015*** −0.007*** −0.003 −0.007***
(0.005) (0.015) (0.005) (0.002) (0.008) (0.002)
LNSIZE −1.046*** −0.488* −1.047*** 0.671*** 0.574*** 0.671***
(0.075) (0.257) (0.075) (0.029) (0.091) (0.029)
LNBM 0.937*** 1.247*** 0.937*** −0.149*** −0.308*** −0.148***
(0.094) (0.238) (0.094) (0.031) (0.091) (0.031)

RET −7.639*** −8.507** −7.627*** 0.191 −0.578 0.197
(0.929) (3.504) (0.928) (0.344) (1.040) (0.344)
SIGMA 7.641*** 0.994 7.618*** −3.294*** −3.001*** −3.285***
(0.699) (2.745) (0.699) (0.281) (0.992) (0.281)
VOLROE −0.172**** −1.496** −0.174**** −0.028 −0.009 −0.027
(0.113) (0.643) (0.113) (0.045) (0.525) (0.045)
PROFIT −0.440** 0.140 −0.439** 0.436*** 0.565**** 0.434***
(0.192) (0.491) (0.192) (0.077) (0.420) (0.077)
SP500 0.469*** −0.046 0.468*** −0.210*** −0.248*** −0.210***
(0.129) (0.422) (0.129) (0.047) (0.083) (0.047)
MNMD 4.988* 9.337 4.995* 4.093*** −4.263 4.093***

(2.758) (7.871) (2.756) (1.022) (3.605) (1.021)
LOSS 0.471*** −0.653* 0.471*** −1.044*** −0.995*** −1.043***
(0.132) (0.367) (0.132) (0.076) (0.198) (0.076)
FIRM_MERROR 0.450*** 1.323*** 0.450*** −0.141*** −0.044 −0.142***
(0.087) (0.400) (0.087) (0.030) (0.143) (0.030)
ANALYS_MERROR 0.532*** 0.314 0.532*** 0.028 0.013 0.028
(0.130) (0.465) (0.130) (0.045) (0.112) (0.045)
29
Table 6 (continued)
30
(1) (2) (3) (4) (5) (6)
LNHORIZON 0.882*** 0.836*** 0.876*** −0.381*** −0.381*** −0.379***
(0.023) (0.083) (0.025) (0.011) (0.035) (0.012)
RETTODATE −0.728*** 0.073 −0.726*** −0.122*** −0.191** −0.125***
(0.086) (0.249) (0.086) (0.031) (0.092) (0.030)
NCO −0.001* 0.000 −0.001* 0.000**** 0.000 0.000****
(0.000) (0.001) (0.000) (0.000) (0.000) (0.000)
NSIC3 0.008* −0.000 0.008* −0.002 −0.004 −0.002
(0.005) (0.012) (0.004) (0.002) (0.007) (0.002)
LNEXP_WITH_FIRM 0.017*** 0.009 0.018*** 0.002 0.006 0.002
(0.005) (0.024) (0.005) (0.002) (0.008) (0.002)
LNGENERAL_EXP −0.102*** 0.047 −0.099*** 0.016 0.010 0.014
(0.030) (0.088) (0.030) (0.013) (0.037) (0.013)
Observations 1,041,230 39,665 1,041,230 959,518 37,711 959,518
Adjusted R2 0.503 0.586 0.503 0.472 0.535 0.472
Year fixed effects Y Y Y Y Y Y
Three-digit SIC fixed effects Y Y Y Y Y Y
Firm fixed effects Y Y Y Y Y Y
Analyst fixed effects Y Y Y Y Y Y
SE clustering Firm Firm Firm Firm Firm Firm
Analyst Analyst Analyst Analyst Analyst Analyst
Notes: This table reports the results of baseline regressions regarding the incremental effects of 13 holidays and October on analysts’ positive and
negative errors. The dependent variable is either POSERROR or NEGERROR. POSERROR (NEGERROR), in percentage terms, is the forecast EPS for a
given firm j in month m in fiscal year t by analyst k, minus the actual EPS [where the forecast is larger (smaller) than the actual EPS] over the mar-
ket share price at the end of fiscal year t−1. PRE_13HOLIDAY takes a value of 1 if the forecast release day is on one of two business days immedi-
ately before or on New Year’s Day, Valentine’s Day, Presidents’ Day, St. Patrick’s Day, Easter, Mother’s Day, Memorial Day, Father’s Day,
Independence Day (Fourth of July), Labor Day, Halloween, Thanksgiving Day or Christmas Day and takes a value of 0 otherwise. OCTOBER takes
a value of 1 if the forecast release day is in October and takes a value of 0 otherwise. NOVDEC takes a value of 1 if the forecast release day is in
November or December and takes a value of 0 otherwise. Refer to Table 1 for the definitions of the other variables. NCO and NSIC3 are in hun-
dreds. Columns 2 and 5 include only earnings forecasts for which the fiscal year ends in March or April. Estimated coefficients and the robust
standard errors are reported in parentheses. ***, **, *, and **** indicate the 1%, 5%, 10%, and one-sided 10% levels of significance, respectively.

Table 7 Disasters and positive and negative errors of local versus nonlocal
analysts
(1) (2)
POSERROR NEGERROR
US_DISASTER −0.170*** 0.079**
(0.065) (0.031)
LOC_US_DISASTER −0.275*** 0.084**
(0.094) (0.042)
Observations 508,985 541,038
Adjusted R2 0.518 0.495
Notes: This table reports the regression results of incremental disaster effects of local analysts over
nonlocal analysts on positive and negative errors. The dependent variable is either POSERROR or
NEGERROR. POSERROR (NEGERROR), in percentage terms, is the forecast EPS for a given firm j in
month m in fiscal year t by analyst k minus the actual EPS [where the forecast is larger (smaller)
than the actual EPS] over the market share price at the end of fiscal year t−1. US_DISASTER takes
a value of 1 if the forecast release day is on the day of or on any of the 5 days immediately fol-
lowing the day on which a disaster occurs in the United States that involves at least 100 fatalities,
causes no significant monetary damage, and does not involve terrorism and takes a value of
0 otherwise. LOC_US_DISASTER takes a value of 1 if the analyst is located in a state where and
when a disaster occurs and takes a value of 0 otherwise. Disasters are listed in Appendix B. For
the sake of brevity, the results for the other control variables are not reported. The regressions
include year, three-digit SIC, firm, and analyst fixed effects. Estimated coefficients and robust
standard errors based on two-way clustering at the firm and analyst levels are reported in paren-
theses. *** and ** indicate the 1% and 5% levels of significance, respectively.
These results are consistent with “the January blues” and the cognition hypothesis
that analysts are generally in more negative moods in January than in other months
and thus produce more accurate forecasts in January. However, there may be a con-
cern that the accuracy of January results is driven by the increased availability of
information at the beginning of the year. To address this concern, we exclude fore-
casts released in the first half of January and repeat the analysis. We continue to find
that forecasts in the second half of January are more accurate than in other months,
which is consistent with the cognition hypothesis.
VIII. CONCLUSIONS
Using US data over the 1982 to 2014 period, we find that analysts generally
produce forecasts with larger errors when holidays are approaching and when
there is a personal income tax cut shock, which are proxies for favorable
moods. Moreover, when major disasters occur and during October, which are
proxies for negative moods, we find more analyst accuracy, particularly among
local analysts and analysts who originate from the country where a disaster
occurs. Further analyses suggest that the additional forecast inaccuracy is
explained neither by higher sentiment due to different economic or market
conditions nor by analyst under-reaction to more bad news released just before
weekends or holidays. We also find that relative to the number of news stories,
there are as many forecasts during the preholiday periods as there are during

Table 8 Chinese disasters and positive and negative errors of Chinese versus non-
Chinese analysts
Non-Chinese Chinese Non-Chinese Chinese
analysts analysts analysts analysts
(1) (2) (3) (4)
CN_DISASTER −0.118*** −0.249** 0.068*** 0.212***
(0.038) (0.101) (0.014) (0.062)
Observations 302,600 9875 345,258 12,929
Adjusted R2 0.559 0.525 0.529 0.535
This table reports the regression results of incremental Chinese disaster effects of Chinese and
non-Chinese analysts on positive and negative errors. The dependent variable is either POSERROR
or NEGERROR. POSERROR (NEGERROR), in percentage terms, is the forecast EPS for a given firm
j in month m in fiscal year t by analyst k minus the actual EPS [where the forecast is larger (smaller)
than the actual EPS] over the market share price at the end of fiscal year t−1. CN_DISASTER takes a
value of 1 if the forecast release day is on the day of or on any of the 5 days immediately following
the day on which a disaster with at least 150 fatalities occurs in China and takes a value of 0 other-
wise. Chinese analysts are analysts with Chinese surnames. The disasters are listed in Appendix C.
For the sake of brevity, the results for the other control variables are not reported. The regressions
include year, three-digit SIC, firm, and analyst fixed effects. Estimated coefficients and robust stan-
dard errors based on two-way clustering at the firm and analyst levels are reported in parentheses.
*** and ** indicate the 1% and 5% levels of significance, respectively.
other periods. Overall, our results are consistent with the notion of a mood-
induced cognitive influence whereby analysts generally make less- (more-) accu-
rate forecasts when they are in a more positive (negative) mood.
Yuk Ying Chang

School of Economics and Finance
Massey University, Manawatu
Private Bag 11 222, Palmerston North 4442
New Zealand
y.chang@massey.ac.nz
REFERENCES
Achor, S., and M. Gielan (2016), ‘The Data-Driven Case for Vacation’, Harvard Busi-
ness Review Digital Articles, 13 July 2016, 2-6. Available at https://hbr.org/2016/
07/the-data-driven-case-for-vacation, retrieved on September 3, 2017.
Antoniou, C., A. Kumar, and A. Maligkris (2017), ‘Terrorist Attacks, Sentiment, and
Corporate Policies’, SSRN Working Paper. Available at https://ssrn.com/abstract=
2566740 or https://dx.doi.org/10.2139/ssrn.2566740, retrieved on December 18,
2017.
Autore, D. M., and D. Jiang (2017), ‘The Preholiday Corporate Announcement Effect’,
SSRN Working Paper. Available at https://ssrn.com/abstract=2520489 or https://
doi.org/10.2139/ssrn.2520489, retrieved on December 18, 2017.

Table 9 Sentiment and positive and negative analyst errors
(1) (2) (3) (4) (5) (6)
US_DISASTER −0.286*** −0.288*** −0.248*** 0.115*** 0.116*** 0.121***
(0.051) (0.051) (0.050) (0.024) (0.024) (0.024)
PRE_3HOLIDAY 0.614*** 0.597*** 0.649*** −0.310*** −0.315*** −0.324***
(0.041) (0.041) (0.045) (0.022) (0.022) (0.023)
C × PRE_3HOLIDAY −0.867*** −0.769*** −0.846*** 0.038 0.065**** 0.103**
(0.120) (0.118) (0.124) (0.041) (0.041) (0.042)
C 0.300*** 0.332*** 0.158** −0.383*** −0.360*** −0.358***
(0.075) (0.076) (0.076) (0.038) (0.038) (0.036)
CSENT 1.277*** −0.168*
(0.198) (0.086)
CCON 1.028*** 0.100*

(0.146) (0.051)
SENT 0.442*** 0.119***
(0.053) (0.030)
Observations 1,041,230 1,041,230 885,869 959,518 959,518 790,846
Adjusted R2 0.503 0.503 0.510 0.473 0.473 0.479
Notes: This table reports the regression results of the incremental effects of major holidays and disasters on analysts’ positive and negative errors,
controlling for sentiment measure. The dependent variable is either POSERROR or NEGERROR. POSERROR (NEGERROR), in percentage form, is
the forecast EPS for a given firm j in month m in fiscal year t by analyst k minus the actual EPS [where the forecast is larger (smaller) than the
actual EPS] over the market share price at the end of fiscal year t−1. US_DISASTER takes a value of 1 if the forecast release day is on the day of or
on any of the 5 days immediately following the day that a disaster occurs in the United States that involves at least 100 fatalities, causes no sig-
nificant monetary damage, and does not involve terrorism and takes a value of 0 otherwise. Disasters are listed in Appendix B. PRE_3HOLIDAY
takes a value of 1 if the forecast release day is in a week ending on Thanksgiving Day, Christmas Day, or New Year’s Day and takes a value of
0 otherwise. C takes a value of 1 if the forecast is released in any month in an NBER contraction period, that is, December 1973–March 1975,
February 1980–July 1980, August 1981–November 1982, August 1990–March 1991, April 2001–November 2001, or January 2008–June 2009,
inclusive, and takes a value of 0 otherwise. CSENT is the Michigan consumer sentiment index in month m. CCON is the Conference Board con-
sumer confidence index in month m. SENT is the Baker-Wurgler investor sentiment index in month m. For the sake of brevity, the results for the
other control variables are not reported. The regressions include year, three-digit SIC, firm, and analyst fixed effects. Estimated coefficients and
the robust standard errors based on two-way clustering at the firm and analyst levels are reported in parentheses. ***, **, *, and **** indicate the
1%, 5%, 10%, and one-sided 10% levels of significance, respectively.
33
Table 10 Pre-nonbusiness days and positive and negative analyst errors
34
(1) (2) (3) (4) (5) (6)
US_DISASTER −0.287*** −0.289*** −0.248*** 0.115*** 0.116*** 0.121***
(0.051) (0.051) (0.050) (0.024) (0.024) (0.024)
PRE_3HOLIDAY 0.616*** 0.599*** 0.652*** −0.312*** −0.317*** −0.327***
(0.041) (0.041) (0.045) (0.022) (0.022) (0.023)
C × PRE_3HOLIDAY −0.867*** −0.769*** −0.846*** 0.038 0.065**** 0.103**
(0.120) (0.118) (0.124) (0.041) (0.041) (0.043)
C 0.300*** 0.332*** 0.158** −0.383*** −0.361*** −0.359***
(0.075) (0.076) (0.076) (0.038) (0.038) (0.036)
CSENT 1.274*** −0.172**
(0.198) (0.086)
CCON 1.027*** 0.098*
(0.146) (0.051)
SENT 0.442*** 0.119***
(0.053) (0.030)
1D_PRE_NON-B −0.014 −0.014 −0.015 0.017** 0.016** 0.012*
(0.013) (0.013) (0.016) (0.007) (0.007) (0.007)
3D_POST_NON-B 0.013 0.013 0.008 0.004 0.004 0.005
(0.010) (0.010) (0.011) (0.005) (0.005) (0.005)
4D_POST_NON-B −0.003 −0.002 0.001 0.017*** 0.017*** 0.010*
(0.010) (0.010) (0.011) (0.005) (0.005) (0.006)
Observations 1,041,230 1,041,230 885,869 959,518 959,518 790,846

Adjusted R2 0.503 0.503 0.510 0.473 0.473 0.479
Notes: This table reports the regression results of the incremental effects of major holidays and disasters on analysts’ positive and negative errors,
controlling for sentiment measure, and days just before nonbusiness days. The dependent variable is either POSERROR or NEGERROR. POSERROR
(NEGERROR), in percentage terms, is the forecast EPS for a given firm j in month m in fiscal year t by analyst k minus the actual EPS [where the
forecast is larger (smaller) than the actual EPS] over the market share price at the end of fiscal year t−1. US_DISASTER takes a value of 1 if the fore-
cast release day is on the day of or on any of the 5 days immediately following the day on which a disaster occurs in the United States that
involves at least 100 fatalities, causes no significant monetary damage, and does not involve terrorism and takes a value of 0 otherwise. The disas-
ters are listed in Appendix . PRE_3HOLIDAY takes a value of 1 for days in a week ending on Thanksgiving Day, Christmas Day, or New Year’s
Day and takes a value of 0 otherwise. 1D_PRE_NON-B takes a value of 1 for days just before nonbusiness days, based on the nontrading days of

the S&P 500, and takes a value of 0 otherwise. 3D_POST_NON-B takes a value of 1 for the third day after nonbusiness days and takes a value of
0 otherwise. 4D_POST_NON-B takes a value of 1 for the fourth day after nonbusiness days. C takes a value of 1 if the forecast is released in any

month in an NBER contraction period, that is, December 1973–March 1975, February 1980–July 1980, August 1981–November 1982, August
1990–March 1991, April 2001–November 2001, or January 2008–June 2009, inclusive, and takes a value of 0 otherwise. CSENT is the Michigan
consumer sentiment index in month m. CCON is the Conference Board consumer confidence index in month m. SENT is the Baker-Wurgler
investor sentiment index in month m. For the sake of brevity, the results for the other control variables are not reported. The regressions include
year, three-digit SIC, firm, and analyst fixed effects. Estimated coefficients and the robust standard errors based on two-way clustering at the firm
and analyst levels are reported in parentheses. ***, **, *, and **** indicate the 1%, 5%, 10%, and one-sided 10% levels of significance, respectively.
35
36
Table 11 Number of forecasts regressions on number of news items
Panel A: Number of contemporaneous news items
Ordinary least squares (OLS) regressions Poisson regressions
(1) (2) (3) (4) (5) (6)
LNFNUM LNFNUM LNFNUM FNUM FNUM FNUM
LNNEWSNUM 0.913*** 0.910*** 0.718***
(0.021) (0.022) (0.023)
NEWSNUM 0.103*** 0.102*** 0.084***
(0.003) (0.003) (0.003)
PRE_3HOLIDAY × LNNEWSNUM −0.024
(0.108)
PRE_3HOLIDAY × NEWSNUM −0.013
(0.015)
PRE_3HOLIDAY2 × LNNEWSNUM −0.051
(0.126)
PRE_3HOLIDAY2 × NEWSNUM −0.013
(0.017)
PRE × LNNEWSNUM −0.095
(0.176)
PRE × NEWSNUM −0.041
(0.028)
FREE × LNNEWSNUM 1.118

(1.460)
FREE × NEWSNUM 0.124
(0.352)
FRI × LNNEWSNUM 1.058***
(0.053)
FRI × NEWSNUM 0.140***
(0.008)
PRE_3HOLIDAY −0.447*** −0.462***
(0.153) (0.073)

Table 11 (continued)
Panel A: Number of contemporaneous news items
Ordinary least squares (OLS) regressions Poisson regressions
(1) (2) (3) (4) (5) (6)
PRE_3HOLIDAY2 −0.400** −0.464***
(0.182) (0.084)
PRE −0.395* −0.379***
(0.237) (0.128)
FREE −1.728 −1.595**
(1.281) (0.686)
FRI −1.552*** −0.643***
(0.075) (0.037)

CONSTANT 4.049*** 4.075*** 4.379*** 5.227*** 5.249*** 5.341***
(0.033) (0.034) (0.036) (0.015) (0.015) (0.016)
Observations 7862 7862 7862 7862 7862 7862
Pseudo R2 0.159 0.169 0.206
Adjusted R2 0.224 0.228 0.293
Panel B: Number of news items as of Day t−4
OLS regressions Poisson regressions
(1) (2) (3) (4) (5) (6)
LNNEWSNUML4 1.251*** 1.252*** 1.074***

(0.029) (0.030) (0.036)
NEWSNUML4 0.127*** 0.126*** 0.102***
(0.004) (0.004) (0.004)
PRE_3HOLIDAY × LNNEWSNUML4 −0.065
(0.137)
PRE_3HOLIDAY × NEWSNUML4 −0.017
(0.013)
37
38
(1) (2) (3) (4) (5) (6)
PRE_3HOLIDAY2 × LNNEWSNUML4 −0.097
(0.167)
PRE_3HOLIDAY2 × NEWSNUML4 −0.011
(0.014)
PRE × LNNEWSNUML4 −0.042
(0.220)
PRE × NEWSNUML4 −0.039
(0.024)
FREE × LNNEWSNUML4 0.797
(1.240)
FREE × NEWSNUML4 0.059
(0.251)
FRI × LNNEWSNUML4 0.569***
(0.061)
FRI × NEWSNUML4 0.092***
(0.009)
PRE_3HOLIDAY −0.349* −0.401***
(0.202) (0.071)
PRE_3HOLIDAY2 −0.114 −0.284***
(0.255) (0.080)
PRE −0.637** −0.540***
(0.304) (0.131)
FREE −1.738 −1.505**
(1.347) (0.689)
FRI −0.760*** −0.255***
(0.089) (0.041)

(1) (2) (3) (4) (5) (6)
CONSTANT 3.442*** 3.461*** 3.705*** 4.952*** 4.971*** 5.034***
(0.044) (0.045) (0.056) (0.017) (0.017) (0.019)
Observations 4350 4350 4350 4350 4350 4350
Pseudo R2 0.222 0.231 0.260
Adjusted R2 0.333 0.336 0.356
Notes: This table reports the regression results of incremental sensitivity of the detrended number of daily news items on the number of daily
forecasts during nonstrict and strict holiday PRE_3HOLIDAY periods. The dependent variable is either FNUM or LNFNUM. FNUM is the number
of daily forecasts. LNFNUM is the natural logarithm of 1 plus the number of daily forecasts. NEWSNUM is the detrended number of daily contem-

poraneous news items (in thousands). LNNEWSNUM is the natural logarithm of 1 plus the detrended number of daily contemporaneous news
items (in thousands). NEWSNUML4 is the detrended number of daily news items (in thousands) as of day t−4. LNNEWSNUML4 is the natural log-
arithm of 1 plus the detrended number of daily news items (in thousands) as of day t−4. PRE_3HOLIDAY takes a value of 1 for the days in a week
ending on Thanksgiving Day, Christmas Day, or New Year’s Day and takes a value of 0 otherwise. PRE takes a value of 1 for Fridays and the days
just before holidays with PRE_3HOLIDAY = 1 and takes a value of 0 otherwise. FREE takes a value of 1 for weekends and holidays with PRE_3HO-
LIDAY = 1 and takes a value of 0 otherwise. PRE_3HOLIDAY2 takes a value of 1 if PRE_3HOLIDAY = 1, PRE 6¼ 1, and FREE 6¼ 1 and takes a value
of 0 otherwise. FRI takes a value of 1 if the day is Friday and takes a value of 0 otherwise. Estimated coefficients and robust standard errors are
reported in parentheses. ***, **, and * indicate the 1%, 5%, and 10% levels of significance, respectively.
39
Bae, K.-H., R. M. Stulz, and H. Tan (2008a), ‘Do Local Analysts Know More? A Cross-
Country Study of the Performance of Local Analysts and Foreign Analysts’, Journal
of Financial Economics, 88, 581–606.
Bae, K.-H., H. Tan, and M. Welker (2008b), ‘International GAAP Differences: The
Impact on Foreign Analysts’, Accounting Review, 83, 593–628.
Baker, M., and J. Wurgler (2006), ‘Investor Sentiment and the Cross-Section of Stock
Return’, Journal of Finance, 61, 1645–80.
Barsky, R. B., and J. A. Miron (1989), ‘The Seasonal Cycle and the Business Cycle’,
Journal of Political Economy, 97, 503–34.
Basi, B. A., K. J. Carey, and R. D. Twark (1976), ‘A Comparison of the Accuracy of Cor-
porate and Security Analysts’ Forecasts of Earnings’, Accounting Review, 51, 244–54.
Bergsma, K., and D. Jiang (2015), ‘Cultural New Year Holidays and Stock Returns
around the World’, Financial Management, 45, 3–5.
Beyer, A. (2008), ‘Financial Analysts’ Forecast Revisions and Managers’ Reporting
Behaviour’, Journal of Accounting and Economics, 46, 334–48.
Bialkowski, J., A. Etebari, and T. P. Wisniewski (2012), ‘Fast Profits: Investor Sentiment
and Stock Returns during Ramadan’, Journal of Banking and Finance, 36, 835–45.
Bollen, J., A. Pepe, and H. Mao (2011), ‘Modeling Public Mood and Emotion: Twitter
Sentiment and Socio-Economic Phenomena’. Proceedings of the Fifth International AAAI
Conference on Weblogs and Social Media (ICWSM 2011), 17–21 July 2011, Barcelona,
Spain.
Bouman, S., and B. Jacobsen (2002), ‘The Halloween Indicator, ’Sell in May and Go
Away’; Another Puzzle’, American Economic Review, 92, 1618–35.
Bourveau, T., and K. Law (2016), ‘Katrina and Analysts’, SSRN Working Paper. Avail-
able at https://ssrn.com/abstract=2782316 or https://dx.doi.org/10.2139/ssrn.
2782316, retrieved on December 18, 2017.
Bushee, B. J., J. E. Core, W. Guay, and S. J. Hamm (2010), ‘The Role of the Business
Press as an Information Intermediary’, Journal of Accounting Research, 48, 1–19.
Byard, D., Y. Li, and Y. Yu (2011), ‘The Effect of Mandatory IFRS Adoption on Finan-
cial Analysts’ Information Environment’, Journal of Accounting Research, 49, 69–96.
Cameron, P. A. et al. (2009), ‘Black Saturday: The Immediate Impact of the February
2009 Bushfires in Victoria, Australia’, Medical Journal of Australia, 191, 11–6.
Chang, X., S. Dasgupta, and G. Hilary (2006), ‘Analyst Coverage and Financing Deci-
sions’, Journal of Finance, 61, 3009–48.
Chang, S.-C., S.-S. Chen, R. K. Chou, and Y.-H. Lin (2008), ‘Weather and Intraday Patterns
in Stock Returns and Trading Activity’, Journal of Banking & Finance, 32, 1754–66.
Christensen, T. E. (2007), ‘Discussion of “Letting the `Tail Wag the Dog’: The Debate over
GAAP Versus Street Earnings Revisited,”’, Contemporary Accounting Research, 24, 741–62.
Clement, M. B. (1999), ‘Analyst Forecast Accuracy: Do Ability, Resources, and Portfo-
lio Complexity Matter?’, Journal of Accounting and Economics, 27, 285–303.
Clement, M. B., and S. Y. Tse (2003), ‘Do Investors Respond to Analysts’ Forecast Revi-
sions as if Forecast Accuracy Is all that Matters?’, Accounting Review, 78, 227–49.
Clement, M. B., and S. Y. Tse (2005), ‘Financial Analyst Characteristics and Herding
Behaviour in Forecasting’, Journal of Finance, 60, 307–41.
Cortés, K., R. Duchin, and D. Sosyura (2016), ‘Clouded Judgment: The Role of Senti-
ment in Credit Origination’, Journal of Financial Economics, 121, 392–413.
DellaVigna, S., and J. M. Pollet (2009), ‘Investor Inattention and Fridays Earnings
Announcements’, Journal of Finance, 64, 709–49.

Dhaliwal, D. S., S. Radhakrishnan, A. Tsang, and Y. G. Yang (2012), ‘Nonfinancial Dis-

closure and Analyst Forecast Accuracy: International Evidence on Corporate Social
Responsibility Disclosure’, Accounting Review, 87, 723–59.
Diener, E., M. Diener, and C. Diener (1995), ‘Factors Predicting the Subjective Well-
Being of Nations’, Journal of Personality and Social Psychology, 69, 851–64.
Dolvin, S. D., M. K. Pyles, and Q. Wu (2009), ‘Analysts Get SAD Too: The Effect of Sea-
sonal Affective Disorder on Stock Analysts’ Earnings Estimates’, Journal of Behavioral
Finance, 10, 214–25.
Dong, G. N., and Y. Heo (2014), ‘Flu Epidemic, Limited Attention and Analyst Fore-
cast Behavior’, Working Paper, University of Columbia.
Drake, M. S., K. H. Gee, and J. R. Thornock (2016), ‘March Market Madness: The
Impact of Value-Irrelevant Events on the Market Pricing of Earnings News’, Contem-
porary Accounting Research, 33, 172–203.
Edmans, A., D. García, and . Norli (2007), ‘Sports Sentiment and Stock Returns’, Jour-
nal of Finance, 64, 1967–98.
Ehrmann, M., and D. J. Jansen (2017), ‘The Pitch Rather than the Pit: Investor Inat-
tention, Trading Activity, and FIFA World Cup Matches’, Journal of Money, Credit
and Banking, 49, 807–21.
Fabozzi, F. J., C. K. Ma, and J. E. Briley (1994), ‘Holiday Trading in Futures Markets’,
Journal of Finance, 49, 307–24.
Fong, K., H. Hong, M. Kacperczyk, and J. D. Kubik (2014), ‘Do Security Analysts Disci-
pline Credit Rating Agencies?’, AFA 2013 San Diego Meetings Paper. Available at
https://ssrn.com/abstract=1970963 or https://dx.doi.org/10.2139/ssrn.1970963,
retrieved on December 18, 2017.
Frieder, L., and A. Subrahmanyam (2004), ‘Nonsecular Regularities in Returns and Vol-
ume’, Financial Analysts Journal, 60, 29–34.
Furin, M. A. (2016), ‘Disaster Planning, Medscape’. Available at http://emedicine.
medscape.com/article/765495-overview, retrieved on September 5, 2017.
Gantz, W., and L. A. Wenner (1991), ‘Men, Women, and Sports: Audience Experi-
ences and Effects’, Journal of Broadcasting & Electronic Media, 35, 233–43.
Gilbert, D., and J. Abdullah (2002), ‘A Study of the Impact of the Expectation of a Holiday
on an Individual’s Sense of Well-Being’, Journal of Vacation Marketing, 8, 352–61.
Goetzmann, W. N., H. Hong, M. Kacperczyk, and J. D. Kubik (2015), ‘Weather-Induced
Mood, Institutional Investors, and Stock Returns’, Review of Financial Studies, 28, 73–111.
Göritz, A. S., and K. Moser (2006), ‘Web-Based Mood Induction’, Cognition and Emo-
tion, 20, 887–96.
Gu, Z., and J. S. Wu (2003), ‘Earnings Skewness and Analyst Forecast Bias’, Journal of
Accounting and Economics, 35, 5–29.
Guttman, I. (2010), ‘The Timing of Analysts’ Earnings Forecasts’, Accounting Review,
85, 513–45.
deHaan, E., J. Madsen, and J. D. Piotroski (2016), ‘Do Weather-Induced Moods Affect
the Processing of Earnings News?’ Stanford University Graduate School of Business
Research Paper No. 15–47.
Hilary, G., and C. Hsu (2013), ‘Analyst Forecast Consistency’, Journal of Finance, 68, 271–97.
Hirshleifer, D., and T. Shumway (2003), ‘Good Day Sunshine: Stock Returns and the
Weather’, Journal of Finance, 58, 1009–62.
Hirshleifer, D., and S. H. Teoh (2003), ‘Limited Attention, Information Disclosure and
Financial Reporting’, Journal of Accounting and Economics, 36, 337–86.

Hirshleifer, D. A., D. Jiang, and Y. Meng (2016), ‘Mood Beta and Seasonalities in Stock
Returns’, SSRN Working Paper, Available at https://ssrn.com/abstract=2880257,
retrieved on December 18, 2017. (also see https://www.ai-cio.com/news/
researchers-find-link-between-investor-moods-and-stock-performance/)
Hirt, E. R., D. Zillmann, G. A. Erickson, and C. Kennedy (1992), ‘Costs and Benefits of
Allegiance: Changes in Fans’ Self-Ascribed Competencies after Team Victory versus
Defeat’, Journal of Personality and Social Psychology, 63, 724–38.
Ho, J. C., M. Qiu, and X. Tang (2013), ‘Do Airlines Always Suffer from Crashes?’, Eco-
nomics Letters, 118, 113–7.
Hong, H., and M. Kacperczyk (2010), ‘Competition and Bias’, Quarterly Journal of Eco-
nomics, 125, 1683–725.
Jacob, J., T. Z. Lys, and M. A. Neale (1999), ‘Expertise in Forecasting Performance of
Security Analysts’, Journal of Accounting and Economics, 28, 51–82.
Johnson, E. J., and A. Tversky (1983), ‘Affect, Generalization, and Perception of Risk’,
Journal of Personality and Social Psychology, 45, 20–31.
Kahn, M. E. (2005), ‘The Death Toll from Natural Disasters: The Role of Income, Geog-
raphy, and Institutions’, The Review of Economics and Statistics, 87, 271–84.
Kamstra, M. J., L. A. Kramer, and M. D. Levi (2000), ‘Losing Sleep at the Market: The
Daylight-Savings Anomaly’, The American Economic Review, 90, 1005–11.
Kamstra, M. J., L. A. Kramer, and M. D. Levi (2003), ‘Winter Blues: A SAD Stock Mar-
ket Cycle’, The American Economic Review, 93, 324–43.
Kamstra, M. J., L. A. Kramer, and M. D. Levi (2012), ‘A Careful Re-Examination of Sea-
sonality in International Stock Markets: Comment on Sentiment and Stock
Returns’, Journal of Banking & Finance, 36, 934–56.
Kamstra, M. J., L. A. Kramer, M. D. Levi, and R. Wermers (2017), ‘Seasonal Asset Allo-
cation: Evidence from Mutual Fund Flows’, Journal of Financial and Quantitative
Analysis, 52, 71–109.
Kaplanski, G., and H. Levy (2010a), ‘Sentiment and Stock Prices: The Case of Aviation
Disasters’, Journal of Financial Economics, 95, 174–201.
Kaplanski, G., and H. Levy (2010b), ‘Exploitable Predictable Irrationality: The FIFA
World Cup Effect on the U.S. Stock Market’, Journal of Financial and Quantitative
Analysis, 45, 535–53.
Kaplanski, G., H. Levy, C. Veld, and Y. Veld-Merkoulova (2015), ‘Do Happy People Make
Optimistic Investors?’, Journal of Financial and Quantitative Analysis, 50, 145–68.
Karamanou, I., and N. Vafeas (2005), ‘The Association between Corporate Boards,
Audit Committees, and Management Earnings Forecasts: An Empirical Analysis’,
Journal of Accounting Research, 43, 453–86.
Kasser, T., and K. M. Sheldon (2002), ‘What Makes for a Merry Christmas?’, Journal of
Happiness Studies, 3, 313–29.
Ke, B., and Y. Yu (2006), ‘The Effect of Issuing Biased Earnings Forecasts on Analysts’
Access to Management and Survival’, Journal of Accounting Research, 44, 965–99.
Kyselý, J., and J. Kim (2009), ‘Mortality during Heat Waves in South Korea, 1991 to
2005: How Exceptional Was the 1994 Heat Wave?’, Climate Research, 38, 105–16.
Lakonishok, J., and S. Smidt (1988), ‘Are Seasonal Anomalies Real? A ninety-Year Per-
spective’, Review of Financial Studies, 1, 403–25.
Lemmon, M., and E. Portniaguina (2006), ‘Consumer Confidence and Asset Prices:
Some Empirical Evidence’, Review of Financial Studies, 19, 1501–29.
Liang, L., and E. J. Riedl (2014), ‘The Effect of Fair Value versus Historical Cost Report-
ing Model on Analyst Forecast Accuracy’, Accounting Review, 89, 1151–77.

Lim, T. (2001), ‘Rationality and Analysts’ Forecast Bias’, Journal of Finance, 56, 369–85.
Lo, K., and S. S. Wu (2010), ‘The Impact of Seasonal Affective Disorder on Financial
Analysts and Equity Market Returns’, SSRN Working Paper, Available at https://
ssrn.com/abstract=1321808, retrieved on December 18, 2017.
Lu, Y., S. Ray, and M. Teo (2016), ‘Limited Attention, Marital Events and Hedge
Funds’, Journal of Financial Economics, 122, 607–24.
Martin, L. L., and G. L. Clore (2013), Theories of Mood and Cognition: A user’s Guide-
book, Vol. 2013. New York, NY: Psychology Press.
Meloy, M. G. (2000), ‘Mood-Driven Distortion of Product Information’, Journal of Con-
sumer Research, 27, 345–59.
Mertens, K., and M. O. Ravn (2013), ‘The Dynamic Effects of Personal and Corporate
Income Tax Changes in the United States’, American Economic Review, 103, 1212–47.
Nawijn, J., M. A. Marchand, R. Veenhoven, and A. J. Vingerhoets (2010), ‘Vacationers
Happier, but Most Not Happier after a Holiday’, Applied Research Quality Life, 5,
35–47. (also see http://www.practicallypositive.com/do-holidays-affect-your-
happiness, retrieved on December 18, 2017)
O’Brien, P. C. (1988), ‘Analysts’ Forecasts as Earnings Expectations’, Journal of Account-
ing and Economics, 10, 53–83.
Palecki, M. A., S. A. Changnon, and K. E. Kunkel (2001), ‘The Nature and Impact of
the July 1999 Heat Wave in the Midwestern United States: Learning from the Les-
sons of 1995’, Bulletin of the American Meteorological Society, 82, 1353–67.
Papousek, I. et al. (2014), ‘Prefrontal EEG Alpha Asymmetry Changes while Observing
Disaster Happening to Other People: Cardiac Correlates and Prediction of Emo-
tional Impact’, Biological Psychology, 103, 184–94.
Saunders, E. M. Jr. (1993), ‘Stock Prices and Wall Street Weather’, The American Eco-
nomic Review, 83, 1337–45.
Schmidt, D. (2013), ‘Investors’ Attention and Stock Covariation’, Working paper, HEC Paris.
Schwarz, N., and H. Bless (1991), ‘Happy and Mindless, but Sad and Smart? The
Impact of Affective States on Analytic Reasoning’, in J. P. Forgas (ed), Emotion and
Social Judgments. Oxford, England: Pergamon Press, pp. 55–71.
Sharpe, L. (2014, January 28), ‘Americans’ Moods Still Improve on Holidays and
Weekends’, Gallup. Available at http://news.gallup.com/poll/167060/americans-
moods-improve-holidays-weekends.aspx, retrieved on December 18, 2017.
Shu, H. C. (2010), ‘Investor Mood and Financial Markets’, Journal of Economic Behavior
and Organization, 76, 267–82.
So, E. C. (2013), ‘A New Approach to Predicting Analyst Forecast Errors: Do Investors
Overweight Analyst Forecasts?’, Journal of Financial Economics, 108, 615–40.
Stone, A. A., S. Schneider, and J. K. Harter (2012), ‘Day-of-Week Mood Patterns in the
United States: On the Existence of ‘Blue Monday’, ‘Thank God It’s Friday’and
Weekend Effects’, The Journal of Positive Psychology, 7, 306–14.
Thaler, R. (1987), ‘Anomalies: Seasonal Movements in Security Prices II: Weekend, Holiday,
Turn of the Month, and Intraday Effects’, Journal of Economic Perspectives, 1, 169–77.
Van Boven, L., and L. Ashworth (2007), ‘Looking Forward, Looking Back: Anticipation
Is more Evocative than Retrospection’, Journal of Experimental Psychology: General,
136, 289–300.
Västfjäll, D., E. Peters, and P. Slovic (2008), ‘Affect, Risk Perception and Future Opti-
mism after Tsunami Disaster’, Judgement and Decision Making, 3, 64–72.
Walther, B. R., and R. H. Willis (2013), ‘Do Investor Expectations Affect Sell-Side Ana-
lysts’ Forecast Bias and Forecast Accuracy?’, Review of Accounting Studies, 18, 207–23.

Wang, K., Y. Li, and J. Erickson (1997), ‘A New Look at the Monday Effect’, The Journal
of Finance, 52, 2171–86.
Wolfe, M. I. et al. (2001), ‘Heat-Related Mortality in Selected United States Cities,
Summer 1999’, American Journal of Forensic Medicine and Pathology, 22, 352–7.
Wright, W. F., and G. H. Bower (1992), ‘Mood Effects on Subjective Probability Assess-
ment’, Organizational Behavior and Human Decision Processes, 52, 276–91.
Yuan, K., Z. Lu, and Q. Zhu (2006), ‘Are Investors Moonstruck? Lunar Phases and
Stock Returns’, Journal of Empirical Finance, 13, 1–23.
APPENDIX A
NORMAL DISTRIBUTION, OPTIMISM HYPOTHESIS, AND

DISTRACTION HYPOTHESIS
Consider a normal distribution F(x) with mean μ and standard deviation σ. The
expressions for E[X|x > 0] and E[X|x < 0] are obtained from those corresponding
to a truncated normal distribution, as follows:

ϕ −σμ
E½Xjx > 0 = μ + σ ðA:1Þ
1 − Φ −σμ
and
−μ
ϕ σ
E½Xjx < 0 = μ − σ −μ ðA:2Þ
Φ σ
where ϕ(•) and Φ(•) are, respectively, the standard normal density and
distribution.
Figure A.1 Normal distribution and optimism hypothesis.

Notes: This figure plots E[X|x > 0] (in black) and E[X|x < 0] (in red) as functions of μ for σ
= 1. Both curves increase in μ over the range shown. In particular, as seen in Figure A.1,
even for a value of μ that is five times the standard deviation of the distribution, both
functions increase in μ. [Color figure can be viewed at wileyonlinelibrary.com]

Figure A.2 Normal distribution and distraction hypothesis.

Notes: For the normal distribution, this figure plots equations (A.1) and (A.2) for μ =
1. The black line shows that E[X|x > 0] increases when precision falls (variance
increases), and the red line shows that E[X|x < 0] decreases when precision falls
(variance increases). [Color figure can be viewed at wileyonlinelibrary.com]
APPENDIX B
US DISASTER LIST
This appendix lists US disasters that had at least 100 fatalities, did not cause sig-
nificant monetary damage, and were not triggered by terrorism during the sam-
ple period.
Date Type Description Fatalities Location [State(s)]

12–16 July Heat wave Chicago Heat 739 Chicago, Illinois [IL]
1995 Wave of 1995
19–30 July Heat wave 271 Midwest and Northeast

1999 [CA*, CT*, DE*, GA*, IA*,
IN, IL, KY, MD, MI, MN,
NJ, NY, OH, OK*, PA, RI*,
TN*, VA*, WA*, WI, WV]
12 November Accident– American Airlines 265 Queens, New York [NY]
2001 Aircraft Flight 587
17 July 1996 Accident– TWA Flight 800 230 Long Island, New York [NY]
Aircraft

Appendix B (continued)
Date Type Description Fatalities Location [State(s)]

6 August 1997 Accident– Korean Air Flight 228 Nimitz Hill, Guam
Aircraft 801
31 October Accident– EgyptAir Flight 217 Atlantic Ocean near
1999 Aircraft 990 Nantucket, Massachusetts
[MA]
16 August Accident– Northwest 156 Detroit, Michigan [MI]
1987 Aircraft Airlines Flight
255
2 August 1985 Accident– Delta Air Lines 137 Dallas, Texas [TX]
Aircraft Flight 191
8 September Accident– USAir Flight 427 132 Pittsburgh, Pennsylvania
1994 Aircraft [PA]
19 July 1989 Accident– United Airlines 111 Sioux City, Iowa [IA]
Aircraft Flight 232
11 May 1996 Accident– ValuJet Flight 110 Florida Everglades [FL]
Aircraft 592
20 February Fire The station 100 West Warwick, Rhode
2003 (building) nightclub fire Island [RI]
*Denotes single-source information.
APPENDIX C
CHINESE DISASTER LIST

This appendix lists Chinese disasters with at least 150 fatalities over the period
2005–2014.
Date Description Fatalities Location
12 May 2008 Earthquake (Mw 7.9), 69,227 Sichuan, Wenchuan,
severe aftershocks Beichuan, Deyang,
Mianyang, Yingxiu,
Mianzhu, Chengdu,
Aba
14 April 2010 Earthquake (Mw 6.9), 2698 Tibet, Qinghai, Yushu,
aftershocks Jiegu
29 May–31 August Floods and landslides 1724+ Fujian, Jiangxi, Hubei,
2010 caused by heavy Hunan, Yunnan,
monsoonal rain Guangdong, Guangxi,
Sichuan, Guizhou,
Anhui, Shaanxi,
Gansu

Appendix C (continued)
Date Description Fatalities Location

8–9 August 2010 Mudslide caused by 1481 Gansu, Zhouqu
heavy rain
11–17 July 2006 Typhoon Bilis/No. 4, 637 Fujian, Guangdong,
floods Guangxi, Hunan,
Jiangxi
3 August 2014 Earthquake (Mw 6.1), 617 Wenping (Yunnan)
aftershocks and
landslides
6–11 August 2006 Typhoon Saomai/No. 8 441 Fujian, Zhejiang,
with winds up to Cangnan
216 km/h, torrential
rain
3 June–17 July Floods caused by heavy 350 Anhui, Zhejian, Jiangxi,
2011 monsoonal rains, Hubei, Hunan,
landslides Sichuan, Chongqing,
Guizhou
8 Sep 2008 Mudslide caused 271 Shanxi, Linfen,
collapse of dam at Xiangfen
Tashan ore mine
14 February 2005 Gas explosion at coal 213 Liaoning, Fuxin
mine
5–10 July 2013 Severe floods 200 Sichuan
5 July 2009 Riots in Urumqi City 197 Xinjiang
20 April 2013 Earthquake Mw 7.0 196 Lushan (Sichuan
Province)
27 November Explosion at coal mine 170 Heilongjiang, Qitaihe
2005
16–30 June 2005 Floods caused by heavy 165 Guangdong, Guangxi,
rain Fujian, Jiangxi,
Hunan, Zhejiang
27 June–17 August Heavy rainfall, floods, 154 Anhui, Hubei, Shaanxi,
2007 and landslides Henan, Jiangsu,
Shandong
18–22 July 2007 Floods caused by heavy 150 Yunnan, Tengchong
rain; mudslide at
Xiaojiangping dam
APPENDIX D
VARIABLES
This appendix lists the definitions of all variables used in this paper.

D1. Dependent variables

Positive forecast error (in percentage terms):
F −A
POSERROR = × 100, F > A ðD1:aÞ
P
where POSERROR is positive forecast error, estimated as the difference between

F and A, scaled by P, when F is larger than A. F is an annual EPS forecast. A is
the actual value of the EPS being forecast. P is the market share price as of the
end of the last fiscal year.
Negative forecast error (in percentage terms):
F −A
NEGERROR = × 100, F < A ðD1:bÞ
P
where NEGERROR is negative forecast error.

FNUM: the number of daily forecasts.
LNFNUM: the natural logarithm of 1 plus the number of daily forecasts.
D2. Test variables

CN_DISASTER takes a value of 1 if the forecast release day falls on the day of
or on any of the 5 days immediately following a disaster in China with at least
150 fatalities and takes a value of 0 otherwise.
FREE takes a value of 1 if the forecast release day is a weekend or a holiday in a
week ending on Thanksgiving Day, Christmas Day, or New Year’s Day and
takes a value of 0 otherwise.
LOC_US_DISASTER takes a value of 1 if the analyst is located in a state in
which a US disaster occurs that involves at least 100 fatalities, causes no signifi-
cant economic damage, and does not involve terrorism and takes a value of 0
otherwise.
LNNEWSNUM is the natural logarithm of 1 plus the number of daily news
items (in thousands) in the United States on the same day as the number of
forecasts.
LNNEWSNUML4 is the natural logarithm of 1 plus the number of daily news
items (in thousands) in the United States on day t−4, where day t is the day for
which the number of forecasts is calculated.
OCTOBER takes a value of 1 if the forecast release day is in October and takes
a value of 0 otherwise.
PRE_13HOLIDAY takes a value of 1 if the forecast release day is on one of two
business days immediately before or on New Year’s Day, Valentine’s Day, Presi-
dents’ Day, St. Patrick’s Day, Easter, Mother’s Day, Memorial Day, Father’s Day,
Independence Day (Fourth of July), Labor Day, Halloween, Thanksgiving Day
or Christmas Day and takes a value of 0 otherwise.

PRE_3HOLIDAY takes a value of 1 if the forecast release day is in a week end-

ing on Thanksgiving Day, Christmas Day, or New Year’s Day and takes a value
of 0 otherwise.
PRE_3HOLIDAY2 takes a value of 1 if the forecast release day is characterized
by PRE_3HOLIDAY = 1, PRE 6¼ 1, and FREE 6¼ 1 and takes a value of 0
otherwise.
US_DISASTER takes a value of 1 if the forecast release day falls on the day of
or on any of the 5 days immediately following the day a disaster occurs in the
United States that involves at least 100 fatalities, causes no significant economic
damage, and does not involve terrorism and takes a value of 0 otherwise.
NEWSNUM is the number of daily news items (in thousands) in the United
States on the same day as the number of forecasts.
NEWSNUML4 is the number of daily news items (in thousands) in the United
States on day t−4, where day t is the day for which the number of forecasts is
calculated.
PRE takes a value of 1 if the forecast release day is on the day just before a
weekend or a holiday and in a week ending on Thanksgiving Day, Christmas
Day, or New Year’s Day and takes a value of 0 otherwise.
D3. Control variables

Following the literature (e.g., Clement 1999; Gu and Wu 2003; Ke and Yu 2006;
Hong and Kacperczyk 2010), we employ the following firm-specific, analyst-
specific and time-specific control variables for the baseline regressions.
Firm-specific control variables:
COVERAGEj,t−1
A measure of analyst coverage, defined as the number of analysts covering firm
j for fiscal year t−1 and constructed using I/B/E/S data.
FIRM_MERRORj,t−1
The median error of all forecasts for firm j in fiscal year t-1, constructed using
I/B/E/S data.40
LNBMj,t−1
The natural logarithm of firm j’s book value of equity divided by its market cap-
italization at the end of fiscal year t−1.
LNBM = ln(ceq/(csho × prcc_f )) (D.2, from Compustat)
LNSIZEj,t−1
The natural logarithm of firm j’s market capitalization at the end of fiscal year
t−1.
LNSIZE = ln(csho × prcc_f )(D.3, from Compustat)
MNMDj,t
A skewness measure, defined as the difference between the mean and median
of EPS, scaled by the market share price for the end of fiscal year t−1 for firm j
40 FIRM_MERROR is included as a control variable because it captures different degrees of fore-

casting difficulty associated with different firms.

for the period between fiscal year t−4 and fiscal year t + 4, excluding fiscal year
t, using I/B/E/S data.
PROFITj,t−1
The operating income of firm j for fiscal year t−1 over the book value of firm
assets as of the end of fiscal year t−2.
PROFIT = ib./lagged ceq(D.4, from Compustat)
RETj,t−1
The average monthly stock returns for firm j for the past 12 months in relation
to the date of the last actual annual earnings, constructed using I/B/E/S and
CRSP data.
SIGMAj,t
The variance of the raw monthly stock returns for firm j for the past 12 months
in relation to the month in which the forecast is released, constructed using
I/B/E/S and CRSP data.
SP500j,t
An indicator that equals one if firm j is in the S&P 500 index when the forecast
is released and equals 0 otherwise.
VOLROEj,t−1
The variance of the residuals from an AR(1) model for firm j’s annual ROE using
the past 10-fiscal-year series. ROE is calculated as the ratio of earnings to the
beginning book value of equity.
ROE = ib./lagged ceq(D.5, from Compustat)
Analyst-specific control variables:
ANALYS_MERRORk,t−1.
The median error of all forecasts made by analyst k for fiscal year t−1, con-
structed using I/B/E/S data.
LNGENERAL_EXPk,t.
The natural logarithm of 1 plus the number of days since the release day of the
first annual EPS forecast made by analyst k, constructed using I/B/E/S data.
LNEXP_WITH_FIRMk,j,t
The natural logarithm of 1 plus the number of days since the release day of the
first annual EPS forecast for firm j made by analyst k, constructed using I/B/
E/S data.
LOSSk,j,t
An indicator that equals one if the forecast made by analyst k is negative and
equals 0 otherwise.
NCOk,t
The number of firms for which analyst k generated annual EPS forecasts in fiscal
year t, constructed using I/B/E/S data.
NSIC3k,t
The number of three-digit SIC industries for which analyst k made annual EPS
forecasts in fiscal year t, constructed using I/B/E/S data.
Time-specific control variables:
C

A dummy that takes a value of 1 if the month is in an NBER contraction period,

that is, December 1973–March 1975, February 1980–July 1980, August 1981–
November 1982, August 1990–March 1991, April 2001–November 2001, or
January 2008–June 2009, inclusive, and takes a value of 0 otherwise.
CCONt (based on Conference Board surveys)
The consumer confidence index for the month in which the forecast is
released.
CSENTt (based on University of Michigan surveys)
The consumer sentiment index for the month in which the forecast is released.
D1_PRE_NON-B
A dummy variable that takes a value of 1 for the days just before nonbusiness
days, based on nontrading days for the S&P 500, and takes a value of 0
otherwise.
D3_POST_NON-B
A dummy variable that takes a value of 1 for the third day after nonbusiness
otherwise.
D4_POST_NON-B
A dummy variable that takes a value of 1 for the fourth day after nonbusiness
otherwise.
FRI
A dummy variable that takes a value of 1 if the forecast release day is a Friday
and takes a value of 0 otherwise.
LNHORIZONk,j,t
The natural logarithm of 1 plus the number of days between the release date of
analyst k’s earnings forecast for firm j and the data date of the earnings being
forecast, constructed using I/B/E/S data.
NOVDEC
A dummy variable takes a value of 1 if the forecast release day is in November
or December and takes a value of 0 otherwise.
RETTODATEk,j,t
The cumulative stock returns (using monthly data) for firm j between the data
date of the last annual earnings and the date on which the earnings forecast by
analyst k is released, constructed using I/B/E/S and CRSP data.
SENTt
The Baker-Wurgler investor sentiment index for the month in which the fore-
cast is released.

APPENDIX E
SAMPLE ATTRITION
Nonmissing Observations
observations remaining Attrition
Step 1. POSERROR and 3,236,914 3,236,914
NEGERROR
Step 2. COVERAGE 3,534,052 3,085,141 151,773
Step 3. LNSIZE 3,393,093 3,083,363 1778
Step 4. LNBM 3,324,938 3,023,699 59,664
Step 5. RET 3,089,228 2,720,480 303,219
Step 6. SIGMA 3,127,157 2,720,053 427
Step 7. VOLROE 2,785,911 2,298,295 421,758
Step 8. PROFIT 3,191,491 2,287,207 11,088
Step 9. MNMD 3,379,846 2,287,185 22
Step 10. FIRM_MERROR 3,514,291 2,285,613 1572
Step 11. ANALYS_MERROR 3,514,291 2,285,613 0
Step 12. HORIZON 3,805,300 2,064,761 220,852
Step 13. RETTODATE 3,805,300 2,064,761 0
Step 14. NCO 3,805,300 2,064,761 0
Step 15. NSIC3 3,805,300 2,064,761 0
Step 16. 3,805,300 2,064,761 0
EXPERIENCE_WITH_FIRM
Step 17. 3,805,300 2,064,761 0
GENERAL_EXPERIENCE

The Best of Times, The Worst of Times: Testing Which Behavioral Biases Affect Analyst Forecasts

Загружено:

Сведения о документе

Оригинальное название

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

The Best of Times, The Worst of Times: Testing Which Behavioral Biases Affect Analyst Forecasts

Загружено:

Авторское право:

Доступные форматы

International Review of Finance, 2018

The Best of Times, the Worst of

Mood-induced optimism, cognitive inaccuracy, and distraction can affect

JEL Codes: G24; G14; D03; G02

Accepted: 29 November 2017

© 2018 International Review of Finance Ltd. 2018

2 © 2018 International Review of Finance Ltd. 2018

example, +9¢ and −9¢, respectively. We refer to this as “the cognition

© 2018 International Review of Finance Ltd. 2018 3

4 © 2018 International Review of Finance Ltd. 2018

In addition, preholiday incremental inaccuracy is positively related to the

II. RELATED LITERATURE

© 2018 International Review of Finance Ltd. 2018 5

Hypothesis 1 (optimism hypothesis). If analysts become more optimistic

An alternative hypothesis is that major events—either good or bad—strongly

9 See Appendix A for the case of a normal distribution.

6 © 2018 International Review of Finance Ltd. 2018

Hypothesis 2 (distraction hypothesis). If analysts become more distracted

Finally, we have the cognition hypothesis, which maintains that individ-

Hypothesis 3 (cognition hypothesis). If analysts process information less

Good times Bad times

11 See Appendix A for the case of a normal distribution.

© 2018 International Review of Finance Ltd. 2018 7

IV. DATA AND VARIABLES

8 © 2018 International Review of Finance Ltd. 2018

© 2018 International Review of Finance Ltd. 2018 9

Appendix C provides a list of these disasters. Fifth, OCTOBER takes a value of

A. The optimism, cognition, and distraction hypotheses

POSERRORk, j, t = a1 + b1 × Zt + c1 × Controls + ek, j, t ð1aÞ

POSERRORk, j, t = a1 + b1 × Zt + c1 × Controls + ϵk, j, t ð2aÞ

where the variable of interest Z is US_DISASTER or OCTOBER, which are deﬁned

10 © 2018 International Review of Finance Ltd. 2018

When the test variable is OCTOBER, we consider two speciﬁcations. First, we

C. Sentiment and mood effects

© 2018 International Review of Finance Ltd. 2018 11

D. Under-reaction and mood effects

21 Walther and Willis (2013) consider CSENT a measure of investor expectations.

12 © 2018 International Review of Finance Ltd. 2018

number of daily forecasts (FNUMt) associated with PRE_3HOLIDAY with respect

FNUM t = μ0 + μ1 × NEWSNUM t + μ2 × NEWSNUM t × PRE_3HOLIDAY t

VI. DESCRIPTIVE STATISTICS AND UNIVARIATE ANALYSIS

© 2018 International Review of Finance Ltd. 2018 13

14 © 2018 International Review of Finance Ltd. 2018

Table 1 Summary statistics

© 2018 International Review of Finance Ltd. 2018 15

© 2018 International Review of Finance Ltd. 2018

© 2018 International Review of Finance Ltd. 2018

periods. Hence, forecasts for PRE_3HOLIDAY typically have a shorter horizon

18 © 2018 International Review of Finance Ltd. 2018

(1) (2) (1)–(2) (3) (4) (3)–(4)

Panel A: PRE_3HOLIDAY versus NON-PRE_3HOLIDAY

Panel B: US_DISASTER versus NON-US_DISASTER

Panel B1: Local analyst versus Nonlocal analyst

© 2018 International Review of Finance Ltd. 2018

Panel B2: Chinese analyst versus Non-Chinese analyst

20 © 2018 International Review of Finance Ltd. 2018

includes slightly more manufacturing ﬁrms (44.42% versus 41.97%). We

VII. EMPIRICAL RESULTS

© 2018 International Review of Finance Ltd. 2018 21

B. Local versus nonlocal analysts

C. Chinese versus non-Chinese analysts

22 © 2018 International Review of Finance Ltd. 2018

signiﬁcant at 5%. As a robustness check, we consider disasters in China with at

MNMD 4.988* 9.337 4.995* 4.093* −4.263 4.093*

LNNEWSNUML4 1.251* 1.252* 1.074***