Firm-Specific Investor Sentiment

David Aboody UCLA Anderson Graduate School of Management

daboody@anderson.ucla.edu

Omri Even-Tov UCLA Anderson Graduate School of Management

omri.even.tov.2014@anderson.ucla.edu

Reuven Lehavy Ross School of Business University of Michigan

e-mail: rlehavy@umich.edu

Brett Trueman UCLA Anderson Graduate School of Management

brett.trueman@anderson.ucla.edu

Current draft: July 2013 We thank Brad Barber, Avanidhar Subrahmanyam, and participants at the 2013 London Business School Transatlantic Doctoral Conference for their helpful comments. All remaining errors are our own.

Abstract

A significant body of research is devoted to examining the effect of investor sentiment on

the price response to firm-level disclosures. Absent a firm-level measure of sentiment, these

studies have used proxies of market-wide sentiment. In this paper we introduce a measure of

firm-specific investor sentiment and use it to gain insights into sentiment’s effect on the price

response to earnings surprises. We develop a simple model which predicts that the greater the

level of firm-specific sentiment, the weaker the relation between announcement return and

earnings surprise and the more negative the market response to earnings that just meet

expectations. Using our proxy for firm-specific investor sentiment, we find strong empirical

support for both of the model’s key predictions. In addition, we show that our firm-specific

measure has greater explanatory power for announcement returns than does the market-wide

sentiment measure used in prior literature.

1

Firm-Specific Investor Sentiment

Introduction

The effect of market sentiment on the cross-sectional and time-series properties of stock

returns is a topic of substantial research interest.1 Among the proxies that have been used to

measure market sentiment are NYSE share turnover, the closed-end mutual fund discount, the

degree of underpricing in initial public offerings, and the aggregate level of corporate

investment. There is also a significant body of work studying how investor sentiment affects

decision-making at the firm level and the price response to firm-level disclosures. While this

type of cross-sectional research naturally calls for the use of a measure of sentiment that is

calculated at the firm level, there are no measures of this sort in the literature. As a consequence,

these studies utilize proxies of market-wide sentiment which, although varying over time, are

invariant in the cross-section. In this paper we introduce a measure of firm-specific investor

sentiment and use it to gain new insights into sentiment’s impact at the individual firm level.

Baker and Wurgler (2006) discuss the potential weakness of using a market-wide

measure of sentiment in firm-level studies. The main focus of their paper is on the construction

of an index of market sentiment (hereafter the BW index), which they use to study the time-

series properties of stock returns. However, in a final analysis, they test whether their market-

wide sentiment index is useful for predicting the returns around individual firms’ earnings

announcements. While they find some evidence that it does, they caution against interpreting too

much into those results, acknowledging that measures of market-wide sentiment have limited

power to detect sentiment at the firm level. They write that “…our [time-series return] results

1 Papers include Arif and Lee (2013), Baker and Wurgler (2006), Brown and Cliff (2004, 2005), Lee et al. (1991), Lemmon and Portniaguina (2006), Ljungqvist et al. (2006), Neal and Wheatley (1998), Qui and Welch (2004), Ritter (1991), Stambaugh et al. (2012), and Yu and Yuan (2011).

2

are driven by the correlated correction of mispricing, but a firm’s announcement event return

picks up the expectational corrections that occur only to it alone, within its own announcement

window.” Similar limitations are expressed by Brown et al. (2012), who use the BW index to

examine the effect of sentiment on a firm’s decision whether to voluntarily release pro forma

earnings statements.2

Our measure of firm-specific sentiment is based upon the notion that investor sentiment

reflects either optimism or pessimism about a firm’s value that is not justified by the available

information.3 It has as its underlying premise that investors who are influenced by sentiment

would be most likely to establish their speculative positions, and thereby influence stock prices,

shortly before an earnings announcement, since that is when they expect their optimism or

pessimism to be confirmed. Establishing positions much earlier would needlessly expose them

to risks unrelated to the announcement. Those investors who are optimistic would tend to

purchase shares prior to the earnings announcement, causing prices to go up, while those who are

pessimistic would sell shares, depressing prices. Reflecting all this, our measure of firm-specific

investor sentiment is the market-adjusted return during the five trading days before the firm’s

earnings announcement.

We begin our analysis by developing a simple model of firm-level investor sentiment,

which generates theoretical predictions regarding the impact of sentiment on the stock reaction to

earnings surprises. Our model shows that the more positive (or less negative) the sentiment, the

less sensitive will returns be to unexpected earnings and the more negatively will the market 2 Other papers that examine firm-level issues using market-wide sentiment measures include Livnat and Petrovits (2009) and Mian and Sankaraguruswamy (2012), who employ the BW index to study the effect of sentiment on the sensitivity of stock prices to firm-specific earnings news, and Bergman and Roychowdhury (2008), who use the Michigan Consumer Confidence Index to investigate the effect of investor sentiment on managerial forecasting behavior. Mikhail et al. (2009) employ the Michigan Index of Consumer Expectations and Hribar and McInnis (2012) use the BW Index to study the impact of sentiment on analysts’ forecast errors. 3 This is similar to the definition of Baker and Wurgler (2007), that sentiment is “a belief about future cash flows and investment risks that is not justified by the facts at hand.”

3

react to an announcement that earnings exactly met expectations. Intuitively, when investors

have positive sentiment about a firm, the firm’s pre-earnings announcement stock price will

reflect an unjustifiably high earnings expectation. With expectations high, just meeting the prior

earnings forecast will be looked upon negatively. Moreover, with the stock price elevated, the

marginal return to a one-cent change in the earnings surprise will be muted. Conversely, when

investors have negative sentiment, the pre-announcement stock price will reflect an earnings

expectation that is unjustifiably low, and so meeting the prior forecast will be viewed positively.

Also, with the stock price depressed, the marginal return to a penny change in the earnings

surprise will be magnified.

Our analyses are conducted on a sample of more than 400,000 quarterly earnings

announcements made over the years 1973 through 2010. We first rank each quarter’s

announcements according to the five-day pre-announcement return (cumulated over days -6

through -2, where day 0 is the date of the earnings announcement). We then partition the

observations into deciles, with the highest (lowest) decile containing the stocks with the most

positive (negative) pre-announcement returns. Regressing the earnings announcement return

(cumulated over days -1 through +1) on unexpected earnings, we find that both the estimated

slope and intercept decline with the sentiment decile, as our model predicts. These results

support our conjecture that the five-day pre-announcement return serves as a measure of firm-

specific investor sentiment. We obtain similar results when we estimate the return-earnings

surprise regression each month over our sample period using a Fama-MacBeth approach.

Results also remain substantially unchanged when we add size, book-to-market, and momentum

as explanatory variables to control for differences in firm-specific characteristics across

sentiment deciles, along with a dummy variable for firm-quarters with losses.

4

Baker and Wurgler (2006), Hribar and McInnis (2012), Mian and Sankaraguruswamy

(2012), and Seybart and Yang (2012) all conjecture that sentiment will have a greater impact on

the returns of firms that are harder to value. The evidence they present is supportive of this

conjecture. If our measure captures firm-specific investor sentiment, then we would expect to

find a similar result. Specifically, the impact of the pre-announcement return on the slope and

intercept of the return-earnings surprise relation should be greater for harder-to-value firms.

To test this prediction, we employ four different proxies, widely used in the literature, to

measure the degree of difficulty in assessing firm value: firm size (smaller firms are likely harder

to value), age (younger firms are expected to be more difficult to value), earnings volatility, and

return volatility (higher volatility is expected to be associated with harder-to-value firms). For

all four proxies, the intercept on the return-earnings surprise regression is more sensitive to

changes in investor sentiment for harder-to-value firms. For three of the four proxies, the

regression slope is also more sensitive for the harder-to-value firms. These results provide

additional support for our conjecture that the five-day pre-announcement return captures firm-

specific investor sentiment.

Recognizing that a theory cannot determine the precise length of the pre-announcement

period that should be used for cumulating returns, we test the robustness of our results by

extending the window over which sentiment is measured to the two months prior to the earnings

announcement. In re-estimating our main regression using this longer window, we end the return

accumulation period six days before earnings are released. We do so in order to ensure that none

of our results are driven by the short-term return reversal pattern documented by Lehmann

(1990). In his paper, Lehmann (1990) finds that a portfolio long (short) in stocks with a positive

(negative) prior five-day return earns positive hedge returns over the next five days. While that

5

pattern cannot explain the inverse relation we document between investor sentiment and the

slope of the return-earnings surprise regression, it could partially contribute to the documented

negative relation between sentiment and the regression intercept. Using the longer window, we

find that the estimated regression slope and intercept remain decreasing functions of investor

sentiment. This is evidence that our results are robust to the length of the return accumulation

window and that they are not attributable to short-term return reversals.

Berkman et al. (2009) show that a portfolio long (short) in stocks with high (low)

dispersion of opinion generally earns a significantly positive hedge return prior to earnings

announcements and a significantly negative return afterwards. As with short-term return

reversals, this result cannot explain the negative relation we find between investor sentiment and

the slope of the return-earnings surprise regression. However, it could potentially be a

contributing factor to the negative relation between investor sentiment and the regression

intercept. To test whether differences of opinion around earnings announcements drives that

result, we re-estimate our return-earnings surprise regression, adding a proxy for the magnitude

of these differences. Including this control variable, we again find that the estimated slope and

intercept are decreasing functions of the level of investor sentiment, consistent with our model’s

prediction.

Using our theoretical model and data, we also re-examine whether the BW measure has

value in explaining the returns around earnings announcements. The evidence here is mixed.

Taking this analysis one step further, we find that our measure of firm-specific investor

sentiment has significant incremental explanatory power for earnings announcement returns over

the BW index. These results reinforce the importance of using a firm-specific measure to study

the effect of sentiment at the individual firm level.

6

The plan of this paper is as follows. In Section I we develop a model of firm-specific

investor sentiment. Our empirical methodology and hypothesis development appear in Section

II. This is followed in Section III by a description of our sample. Our full-sample empirical

results are presented in Section IV. An analysis of whether sentiment has the greatest effect in

hard-to-value firms appears in Section V. We present the results of robustness tests in Section

VI. Section VII provides a summary and conclusions.

I. Model development

In this section we develop a model which will serve as the basis of our empirical

analyses. We consider a one-period, three-date economy and two types of risky firms – one that

is traded solely by investors who are influenced by sentiment (as described below) and the other

that is traded exclusively by investors unaffected by sentiment. In all other respects the two

types of firms are indistinguishable. All investors in the economy are risk neutral and the one-

period discount rate is assumed equal to zero. The fraction of firms in the economy that are

traded by investors influenced by sentiment is denoted by p. Each firm’s earnings for the period,

X, are normally distributed with a mean of ( ) 0E X > and a standard deviation of 2Xσ , and are

assumed equal to the period’s cash flows.

At the beginning of the period, date 0, the price of each firm, 0P , is equal to the risk

neutral investors’ prior expectation of earnings, ( )E X . At the end of the period, date 2, each

firm’s earnings are formally announced and all firms are liquidated. Firm price at that time, 2P ,

is therefore equal to X. With probability 1 q− the value of the realized earnings leaks out at date

1, just before the formal announcement. In that case, the date 1 price of the firm, 1P , would also

equal X. The return at date 1, 1R , would be given by:

7

10

( )X E XRP

−= (1)

and the date 2 return, 2R , would equal zero. The returns at dates 1 and 2 would be unrelated

when there is information leakage. Note that the right hand side of (1) is just the firm’s

unexpected earnings (normalized by price). This quantity will sometimes be denoted by UE in

our analysis below.

With probability q there is no leakage of information at date 1. In that case, investors

who are influenced by sentiment receive in its place a completely noisy signal of earnings,

denoted by B, which they incorrectly believe provides information about the realized value of X.

(The investors who are not influenced by sentiment recognize that this signal is worthless.)

These investors conjecture that BB X ε= + , where Bε is normally distributed with a mean of zero

and a finite standard deviation of 2Bσ . Given their belief, the date 1 price of a firm traded by

these investors equals ( ) (1 )wE X w B+ − , where 2

2 2B

X B

w σσ σ

=+

. The return at date 1 for such a

firm is equal to:

1( ) (1 ) 1

( )wE X w BR

E X+ −

= − (2)

The realized value of the signal, B, determines whether the investors are optimistic or

pessimistic. If B is greater than the initial earnings expectation, ( )E X , then investors are

optimistic and the return at date 1 will be positive. If B is less than ( )E X , then they are

pessimistic and the return will be negative. Of course, B is not observable to the researcher.

However, the date 1 return, 1R , is observable and fully captures the information in B. It is our

measure of firm-specific sentiment. Values of 1R greater than zero imply investor optimism.

8

The more positive is 1R , the greater the level of optimism and the greater will be the positive

deviation of price from underlying firm value. Values of 1R less than zero imply investor

pessimism. More negative values reflect greater levels of pessimism and a greater negative

deviation of price from underlying firm value.

The date 2 return for a firm traded by investors influenced by sentiment is:

2[ ( ) (1 ) ]

[ ( ) (1 ) ]X wE X w BR

wE X w B− + −

=+ −

(3)

From equations (2) and (3), we can express 2R as a function of 1R and UE as follows:

21 1

1 11 1

UERR R

= + −+ +

(4)

The final possible scenario in this economy is one in which there is no leakage of

information and the investors trading in the firm are unaffected by sentiment. In this case, the

price of the firm at date 1 is simply equal to ( )E X and the date 1 return is zero. The date 2

return is equal to unexpected earnings, UE. As in the case of information leakage, the returns at

dates 1 and 2 would be unrelated.

Calculating the expected return at date 2, 2( )E R , over all possible scenarios in this

economy yields:

21 1

1( ) 1 (1 ) (1 )*0,1 1

UEE R q p p UE qR R

= ∗ + − + − + − + +

(5)

which can be rewritten as:

21( ) (1 ) ,

1 1UEE R qp q p UE qp qp

Sent Sent= − + − + +

+ + (6)

9

where we use Sent in place of 1R in (6) to emphasize that this is our measure of firm-specific

sentiment.

II. Empirical methodology and hypothesis development

The empirical analog to equation (6) is:

1 2 31 1 ,

1 ( ) 1 ( )jt jt jt jtc jt c jt

R UE UED Sent D Sent

α β β β ε= + + + ++ +

(7)

where:

jtR = the cumulative market-adjusted return for firm j (computed using the CRSP value-

weighted market index) in the three days surrounding the announcement of quarter t earnings

(trading days -1, 0, and 1, where day 0 denotes the date of the earnings announcement);4

4

4

jt jtjt

jt

X XUE

P−

−

−= , where jtX is earnings before extraordinary items for firm j in quarter t and

4jtP − is the share price of firm j at the end of quarter t-4;

jtSent = the cumulative market-adjusted return for firm j over the period from six trading days

before to two trading days before the announcement of quarter t earnings (sometimes referred to

as the pre-announcement return);

( )c jtD Sent = the decile rank of jtSent , where the rank is determined relative to the values of

itSent for all firms, i, that announce in the same calendar quarter c (higher deciles reflect greater

optimism/less pessimism); and

jtε = the regression residual for firm j and quarter t.

4 Our results are robust to the use of raw returns, rather than market-adjusted returns.

10

We use a three-day earnings announcement window in order to mitigate the effect of

possible database errors with respect to the exact date of the earnings announcement as well as to

account for uncertainty over whether earnings were announced after hours. The earnings

announcement date used for our analysis is that reported in Compustat, unless the announcement

date also appears in the I/B/E/S database and the I/B/E/S date is the earlier of the two. In that

case, the I/B/E/S date is used instead. If I/B/E/S also specifies the exact time of the

announcement and it is after trading hours, then the earnings announcement date is advanced by

one trading day. We use a firm’s four-quarters ago earnings, rather than the consensus analyst

forecast, as our expectation of current-quarter earnings in order to be able to retain in our sample

those firms without analyst coverage. Recognizing that the cross-sectional distribution of our

sentiment measure changes over time, we use the sentiment decile in our regression, rather than

the value of the measure, itself. This allows us to maintain comparability across our sample

period and is consistent with Baker and Wurgler (2006), Hribar and McInnis (2012), and Seybart

and Yang (2012).5 We use a relatively short period of five trading days prior to the earnings

announcement window to measure sentiment because individuals who believe they have private

information about upcoming earnings are more likely to trade on their information close in time

to the earnings announcement. Establishing positions earlier would subject the traders to a

greater level of risk unrelated to the forthcoming announcement.6

Our formal hypothesis follows directly from expression (7):

5 Using the actual value of sentiment, rather than the sentiment decile, does not qualitatively affect our main results. 6 Using similar reasoning, Berkman et al. (2009) hypothesize that firms subject to high dispersion of opinion should experience a share price run-up in the few days before their earnings announcements. Their empirical findings support this conjecture. As we report later, our main results are qualitatively unchanged when we control for dispersion of opinion.

11

Hypothesis: In a regression of earnings announcement return on earnings surprise, the intercept

and slope will be decreasing functions of firm-specific investor sentiment. That is, the

coefficients 2β and 3β in expression (7) will be positive.7

III. Data selection and descriptive statistics

Our initial sample is the set of quarterly earnings announcements contained in the merged

CRSP-Compustat quarterly file for the period from January 1973 through December 2010. We

end our sample period in December 2010 because some of our tests involve the BW index, for

which data is provided only through the end of that year.8 From this sample we remove those

announcements for which either the announcement date is not available or is more than 90 days

after quarter-end. We also delete any announcement of quarter t earnings by firm j if one or

more of the following variables are missing from the CRSP-Compustat dataset: (1) the earnings

of firm j for quarter t-4, (2) the stock price of firm j at the end of quarter t-4, (3) firm j’s market

capitalization at the end of quarter t, (4) monthly returns for firm j over the period from 12

months before the end of quarter t through two months before the end of that quarter (this data is

used to calculate return volatility, which is defined as the standard deviation of monthly stock

returns over those months, and is also used to calculate stock price momentum), (5) firm j’s book

value of equity at the end of quarter t, and (6) any day’s return during the period from six trading

days before to one trading day after firm j’s announcement of quarter t’s earnings.

We also remove announcements for which (a) the announcement date of quarter t

earnings recorded in Compustat and the date recorded in I/B/E/S differ by more than five trading

days, (b) firm j’s price per share at the end of quarter t is less than $1, or (c) firm j’s market value 7 Since our measure of sentiment appears in the denominator of the independent variables in (7), an inverse relation between sentiment and the slope and intercept of the return-earnings surprise relation requires that the two regression coefficients be positive. 8 Our main results are robust to extending our sample period to the end of 2012.

12

at the end of quarter t is less than $1 million. Finally, for each calendar quarter we truncate those

observations that are at the top and bottom one percent of the unexpected earnings distribution,

in order to lessen the impact of outliers. After removing all announcements that meet any of

these criteria, we are left with a final sample of 425,170 observations.

A few of our regressions include earnings volatility as an independent variable. We

define earnings volatility as the standard deviation of the ratio of quarterly operating income

before depreciation to average total assets, calculated over the prior 20 quarters. For these tests

we drop any earnings announcement for which this operating income-to-asset ratio is not

available for at least eight of the prior 20 quarters. This reduces the sample size in these tests to

332,033 observations.

Descriptive statistics for our variables are presented in Table I, panel A. The average

cumulative market-adjusted return over the three-day earnings announcement window (days -1,

0, and 1) is 0.22 percent, while the average value of our firm sentiment measure, the cumulative

market-adjusted return over the pre-announcement period (days -6 through -2), is 0.28 percent.

These positive average returns likely reflect the earnings announcement premium, which has

been documented by Ball and Kothari (1991), Cohen et al. (2007), Frazzini and Lamont (2007),

and Barber et al. (2012), among others. The mean (median) market value of the firms in our

earnings announcement sample is $2.11 billion ($214 million), while the mean (median) book-

to-market ratio is 0.99 (0.63).9 These numbers compare to a mean (median) end-of-quarter

market value for all New York Stock Exchange firms over our sample period of $3.13 billion

($490 million) and a mean (median) book-to-market ratio of 2.06 (0.68).10 The average age of

the firms in our sample (where age is defined as the number of years a firm’s shares have been

9 For firms with negative book value, we set the book-to-market ratio equal to zero. 10 As we demonstrate by adding controls for firm characteristics to our regression, our results are not driven by the specific size and book-to-market attributes of our sample.

13

publicly trading) is just under 17 years. Unexpected earnings has a mean of 0.34 percent of

stock price, reflecting growth in earnings, on average, over our sample period. Return volatility

has a mean of 12.4 percent, while earnings volatility averages 3.8 percent.

The pairwise Pearson correlation coefficients for the variables of interest are presented in

Table I, panel B. As expected, age and firm size are strongly positively correlated (correlation =

0.225). Earnings volatility and return volatility are also significantly positively correlated, as

expected; however, the magnitude of the correlation, 0.005, is low. The correlation between the

market-adjusted return during the earnings announcement window and the corresponding return

during the pre-announcement period is a significantly negative 0.087. The negative correlation

is consistent with the notion that the announcement of a firm’s earnings serves to partially

correct the misvaluations of investors who are influenced by sentiment.

IV. Full-sample results

We begin our empirical analysis by estimating equation (7). Results are presented in

column 1 of Table II. Consistent with our hypothesis, and the conjecture that pre-earnings

announcement returns are a measure of firm-specific investor sentiment, the coefficients on both

11 ( )c jtD Sent+

and 11 ( )jt

c jt

UED Sent+

in (7) are positive and significant. Alternatively stated,

the sensitivity of announcement returns to earnings surprises is decreasing in investor sentiment

as is the expected stock price reaction to an announcement that earnings just met expectations.11

To ensure that our results are not driven by differences in firm characteristics across

sentiment deciles, we add ln(size), ln(book-to-market ratio), momentum, and a dummy variable 11 Cohen et al (2007) show that unexpected delays in earnings announcements are accompanied by negative returns in the pre-announcement period. To determine whether this phenomenon has an impact on our results, we repeat our analysis, excluding any firm-quarter observation for which the earnings announcement date falls more than seven days after the announcement date in the corresponding quarter of the previous year. Untabulated results are similar to those presented in Table II.

14

equal to one for firm-quarters with losses as explanatory variables to (7).12 As reported in

column 2 of Table II, the coefficients on our independent variables, 11 ( )c jtD Sent+

and

11 ( )jt

c jt

UED Sent+

, remain significantly greater than zero and are of similar magnitudes to the

values reported in column 1 of the table. From this we conclude that our findings are not due to

differences in firm characteristics across the sentiment deciles.

In order to reduce the potential impact of influential observations of jtUE on our results,

we re-estimate (7) using the rank of jtUE , in place of jtUE :

1 2 31 1( ) ( ) .

1 ( ) 1 ( )jt jt jt jtc jt c jt

R Rank UE Rank UED Sent D Sent

α β β β ε= + + + ++ +

(8)

We do not replace the variable 11 ( )c jtD Sent+

by the rank of that variable since we are already

using the decile rank of jtSent , rather than jtSent , itself, in the regression. The results of

estimating this rank regression appear in column 3 of Table II. The coefficients on

11 ( )c jtD Sent+

and 1( )1 ( )jt

c jt

Rank UED Sent+

are significantly positive, confirming that

influential observations are not driving our results.

We also employ a Fama-MacBeth approach and estimate regression (8) for each month

of our sample period.13 The averages of the monthly coefficients on 11 ( )c jtD Sent+

and

1( )1 ( )jt

c jt

Rank UED Sent+

are reported in column 4 of Table II. Both are reliably positive.

12 For firms with negative book value, we set ln(book-to-market ratio) equal to zero. 13 We exclude from our regressions the 12 months at the beginning of our sample period that have fewer than 100 observations.

15

Figure I, panel A (panel B) plots the estimated value of 2β ( 3β ) for each month of our sample

period. As is clear from the figure, the months with positive coefficients are not concentrated

within any narrow time frame. Rather, they are spread out over our entire sample period. The

coefficient, 2β , is positive in 91.1 percent of the months, while 3β is positive for 53.7 percent of

the months.14

We alternatively calculate unexpected earnings using as our expectation the consensus

analyst forecast, computed as the average of all valid individual forecasts in I/B/E/S that are

outstanding seven days before the earnings announcement.15 Unexpected earnings in this case is

equal to the I/B/E/S reported earnings minus the consensus analyst forecast. (We normalize each

earnings surprise by the stock price at the end of the prior quarter.) As before, we truncate the

highest and lowest one percent of the unexpected earnings observations. The requirement that

I/B/E/S consensus forecasts be available, along with the truncation of the top and bottom one

percent of the unexpected earnings observations, reduces our sample to slightly more than

220,000 announcements. The results of re-estimating regression (8) using this alternative

calculation of unexpected earnings are presented in panel A of Table III. Both the coefficient on

11 ( )c jtD Sent+

and the coefficient on 1( )1 ( )jt

c jt

Rank UED Sent+

remain significantly greater than

zero.

14 When estimating panel regressions it is common practice to adjust the standard errors for correlation over time and across firms (usually referred as double clustering). This procedure is used because market-wide shocks may induce correlation between firms at a given point in time, and persistent firm-specific shocks may induce correlation across time for a given firm. In addition, researchers use the White correction to account for potential cross-sectional heteroskedasticity (see Thompson, 2011). Our research design, though, eliminates the need to adjust our standard errors for auto-correlation and heteroskedasticity. Our firm-specific sentiment measure is the pre-earnings announcement return; as such, the errors should not be correlated over time. Moreover, we use the rank of the sentiment measure; hence, market-wide shocks, while affecting all firms, should not affect their decile ranking. This allays concerns over the correlation between firms at a given point in time and heteroskedasticity. Nevertheless, we replicated our analysis using both double clustering (over time and across firms) and employing the White correction and obtained qualitatively similar results. 15 A valid forecast is one that is issued no more than 90 days prior to the earnings announcement.

16

The I/B/E/S sample also gives us a convenient alternative means of testing the hypothesis

that the greater the level of firm-specific sentiment, the less positive (or more negative) the stock

price reaction to an announcement of earnings that just meets expectations (UE=0). Using the

subsample of observations for which realized earnings equal the consensus forecast, we regress

the market-adjusted announcement return on the announcement’s sentiment decile. As we report

in Table III, panel B, the coefficient on ( )c jtD Sent in this regression is negative and significant,

consistent with our conjecture.

While not the main focus of their paper, Baker and Wurgler (2006) examine whether

their market-wide sentiment index is useful for predicting the returns around individual firms’

earnings announcements. Although they find some evidence that it does, they acknowledge that

their methodology has limited power in this setting since their index is not explicitly designed to

detect sentiment at the firm level. We re-examine the effectiveness of their measure at the firm

level using our theoretical model and our data, by substituting the BW index for our measure of

sentiment in expression (7):

1 2 31 1 ,

1 1jt jt jt jtt t

R UE UEBWSent BWSent

α β β β ε′ ′ ′′= + + + ++ +

(9)

where tBWSent is the value of the BW index during the last month of quarter t. As reported in

Table IV, column 1, the regression results are mixed. Consistent with the hypothesis, the

coefficient, 2β ′ , on 11 tBWSent+

is significantly positive. However, contrary to the hypothesis,

the coefficient, 3β ′ , on 11jt

t

UEBWSent+

is insignificantly different from zero.

We take this analysis one step further and test whether our measure of firm-specific

investor sentiment has incremental explanatory power for earnings announcement returns over

17

the BW index. We do so by adding the two independent variables, 11 tBWSent+

and

11jt

t

UEBWSent+

, to regression (7):

1 2 3

2 3

1 11 ( ) 1 ( )

1 1 .1 1

jt jt jtc jt c jt

jt jtt t

R UE UED Sent D Sent

UEBWSent BWSent

α β β β

β β ε

= + + ++ +

′ ′+ + ++ +

(10)

Coefficient estimates from this regression are reported in Table IV, column 2. The coefficients

on 11 ( )c jtD Sent+

and 11 ( )jt

c jt

UED Sent+

remain positive and significant, providing evidence

that our firm-specific sentiment measure has incremental explanatory power for announcement

window returns over the BW index. As before, the coefficient 2β ′ is significantly positive,

while the coefficient 3β ′ is insignificantly different from zero. These results reinforce the

importance of using a firm-specific measure to study the effect of sentiment at the individual

firm level.

V. The impact of sentiment on hard-to-value firms

Baker and Wurgler (2006), Hribar and McInnis (2012), Mian and Sankaraguruswamy

(2012), and Seybart and Yang (2012) all conjecture that sentiment will have a greater effect on

returns for subsets of firms that are harder to objectively value. The empirical evidence they

present is consistent with this conjecture. If our measure captures firm-specific investor

sentiment, then we should find a similar result. Specifically, the impact of the pre-announcement

return on the slope and intercept of the return-earnings surprise relation should be greater for

harder-to-value firms.

18

To test this we estimate the following regression:

0 1 2 3

1 2 3

1 11 ( ) 1 ( )

1 1( ) ( ) ( ) ,1 ( ) 1 ( )

jt jt jtc jt c jt

c jt c jt c jt jt jtc jt c jt

R UE UED Sent D Sent

D HTV D HTV D HTV UED Sent D Sent

β β β β

γ γ γ ε

= + + ++ +

+ + + ∗ ++ +

(11)

where:

jtHTV = the value of the proxy (see below) used to measure the difficulty of valuing firm j in

quarter t and

( )c jtD HTV = the decile rank of jtHTV , where the rank is determined relative to the values of

itHTV for all firms, i, that announce in the same calendar quarter c.16 If the market-adjusted

return over days -6 through -2 is, indeed, a measure of firm-level investor sentiment, then the

coefficients 2γ and 3γ in expression (11) should be positive. We proxy for the extent to which a

firm is difficult to value by four different variables: firm size (smaller firms are likely harder to

value), age (it is expected that younger firms will be more difficult to value), earnings volatility,

and return volatility (higher volatility is likely associated with harder-to-value firms).17 We

expect that firm-specific investor sentiment will have the greatest impact on the return-earnings

surprise relation for the smallest and youngest firms, and for those with the greatest earnings

volatility and return volatility.

Table V reports the results of estimating (11) over our entire sample period for each hard-

to-value proxy. The coefficient 2γ is significantly positive for all the proxies – age (column 1),

size (column 2), earnings volatility (column 3), and return volatility (column 4). The coefficient

3γ is positive and significant for three of the four proxies (all but return volatility, where it is 16 The deciles are constructed such that firms in higher deciles are more difficult to value. 17 These proxies are also used by Baker and Wurgler (2006), Hribar and McInnis (2012), Mian and Sankaraguruswamy (2012), and Seybert and Yang (2012).

19

negative). These findings provide support for the hypothesis that the intercept and slope of the

return-earnings surprise relation are more sensitive to firm-specific investor sentiment for firms

that are harder to value. As such, the results are also supportive of our conjecture that the five-

day pre-announcement return captures firm-specific investor sentiment.

VI. Robustness tests

Recognizing that theory cannot determine the precise length of the pre-announcement

period that should be used to measure sentiment, we test the robustness of our results by

extending the return measurement window to the two months prior to the earnings

announcement. In re-estimating our main regression we exclude from this return accumulation

period the returns over days -6 through -2. We do so in order to ensure that none of our results

are driven by the short-term return reversal pattern documented by Lehmann (1990). In his

paper, Lehmann (1990) finds that stocks whose prices increase (decrease) over the previous five

days tend to decline (rise) over the next five. While this pattern cannot account for the inverse

relation we find between investor sentiment and the sensitivity of announcement returns to

earnings surprises, it could potentially be a contributing factor to the negative relation between

investor sentiment and the intercept of the return-earnings surprise regression.

We estimate the following regression using the longer return accumulation window:

1 2 3 41 1 ,

1 ( ) 1 ( )jt jt jt jt jtc jt c jt

R UE UE SentD Sent D Sent

α β β β β ε= + + + + +′ ′+ +

(12)

where jtSent′ is the cumulative market-adjusted return for firm j over the period from 60 trading

days before to seven trading days before the announcement of quarter t earnings and ( )c jtD Sent′

is the decile rank of jtSent′ , where the rank is determined relative to the values of itSent′ for all

firms, i, that announce in the same calendar quarter c. As a control variable, we also include in

20

the regression the cumulative market-adjusted return over days -6 through -2. Results of

estimating (12) are reported in Table VI, panel A. Although smaller than their counterparts in

Table II, column 1, both 2β and 3β are reliably greater than zero. This is evidence that our

results are robust to the length of the pre-announcement return accumulation window and that

they are not attributable to short-term return reversals.

That the cumulative market-adjusted return over days -60 through -7 reflects firm-

specific investor sentiment invites the question of whether our five-day sentiment measure

provides incremental explanatory power over this longer-window measure. We expect that it

will have additional explanatory power, given that investors who are influenced by sentiment are

arguably more likely to establish their speculative positions shortly before an earnings

announcement. To test this conjecture, we add our measure as a main effect to regression (12)

and also interact it with the longer-term sentiment measure:

1 2 3 4

5 6

1 1 11 ( ) 1 ( ) 1 ( )

1 1 1 1x x1 ( ) 1 ( ) 1 ( ) 1 ( )

.

jt jt jtc jt c jt c jt

jt jtc jt c jt c jt c jt

R UE UED Sent D Sent D Sent

UED Sent D Sent D Sent D Sent

α β β β β

β β ε

= + + + +′ ′+ + +

+ + +′ ′+ + + +

(13)

Results of estimating (13) are presented in Table VI, panel B. As expected, the coefficient on the

five-day measure of firm-specific investor sentiment, 4β , is significantly greater than zero, as

are the coefficients on the interaction terms between the short-window and longer-window

measures, 5β and 6β . Moreover, the coefficient on 11 ( )c jtD Sent′+

, 2β , is no longer positive,

consistent with the market-adjusted return over days -6 through -2 being the more powerful of

the two sentiment measures.

21

We next test whether differences of opinion could be driving any of our results. In the

presence of differences of opinion and short-sales constraints, Miller (1977) conjectures that

stock prices will be positively biased; investors with favorable information will bid prices up,

while those with unfavorable information will be constrained from fully realigning prices to

fundamental value. Berkman et al. (2009) posit that this bias will be stronger just before

earnings announcements (when investors trade on their private information), but will dissipate

once earnings are announced and differences of opinion are reduced. Consistent with their

conjecture, they find that a hedge portfolio long in stocks with high dispersion of opinion and

short in stocks with low dispersion generally earns a significantly positive return during the two

weeks before the earnings announcement and a significantly negative return during the

subsequent two weeks. As with short-term return reversals, this phenomenon cannot account for

the negative relation we find between investor sentiment and the sensitivity of returns to earnings

surprises. However, it, too, could partially contribute to the documented negative relation

between sentiment and the intercept of the return-earnings surprise regression. To test whether

differences of opinion is driving the intercept result, we re-estimate (7), adding an explanatory

variable to capture the degree to which there are differences of opinion prior to an earnings

announcement:

1 2 3 41 1 ( ) ,

1 ( ) 1 ( )jt jt jt c jt jtc jt c jt

R UE UE D DiffD Sent D Sent

α β β β β ε= + + + + ++ +

(14)

where jtDiff is our proxy (see below) for differences of opinion prior to firm j’s announcement

of quarter t earnings and ( )c jtD Diff is the decile rank of jtDiff , where the rank is determined

relative to the values of itDiff , for all firms, i, that announce earnings in calendar quarter c

(higher deciles reflect greater differences of opinion). We employ two separate proxies,

22

earnings volatility and return volatility. Regression results are reported in Table VI, panel C.

Consistent with Berkman et al. (2009), the coefficient on the earnings volatility variable (column

1) and on the return volatility variable (column 2) are significantly negative. More importantly

for our purposes, the coefficients, 2β and 3β , remain significantly greater than zero and are of a

magnitude similar to those estimated for regression (7), when no difference of opinion proxy is

included. We infer from this that our results are not driven by differences of opinion.

VII. Summary and conclusions

This paper tests the conjecture that firm-specific investor sentiment can be measured by a

firm’s pre-earnings announcement return. To guide our empirical analysis, we develop a simple

theoretical model to show how firm-level investor sentiment is expected to impact the relation

between earnings announcement returns and unexpected earnings. The model leads to the

predictions that the return-earnings surprise relation will be weaker the greater the level of

investor sentiment and that the announcement return in cases where earnings just meet

expectations will be more negative the greater the level of investor sentiment. These predictions

are a consequence of the role that earnings play in mitigating investors’ misvaluation of the firms

in which they invest. Our empirical analysis provides support for each of these predictions. We

also conjecture that the impact of investor sentiment on the return-earnings surprise relation will

be greater for harder-to-value firms than for those that are easier to value. Our empirical results

are supportive of this prediction as well.

Market-wide measures of sentiment have been used in the past to examine cross-sectional

and time-series properties of stock returns. Absent a firm-specific measure of sentiment,

researchers have also employed these market-wide measures to study the effect of sentiment on

23

decisions and prices at the individual firm level. By developing a firm-level measure of

sentiment, we provide a tool that can be used in future research on this topic.

24

References

Arif, Salman, and Charles Lee, 2013, Does aggregate investment reflect investor sentiment?, Indiana University working paper. Baker, Malcolm, and Jeffrey Wurgler, 2006, Investor sentiment and the cross-section of stock returns, Journal of Finance, 61(4), 1645–1680. Baker, Malcolm, and Jeffrey Wurgler, 2007, Investor sentiment in the stock market, Journal of Economic Perspectives, 21(2), 129–151. Ball, Ray, and S.P. Kothari, 1991, Security returns around earnings announcements, The Accounting Review, 66(4), 718–738. Barber, Brad, Emmanuel De George, Reuven Lehavy, and Brett Trueman, 2013, The earnings announcement premium around the globe, Journal of Financial Economics, 108(1), 118-138. Bergman, Nittai, and Sugata Roychowdhury, 2008, Investor sentiment and corporate disclosure, Journal of Accounting Research, 46(5), 1057–1083. Berkman, Henk, Valentin Dimitrov, Prem Jain, Paul Koch, and Sheri Tice, 2009, Sell on the news: Differences of opinion, short-sales constraints, and returns around earnings announcements, Journal of Financial Economics, 92(3), 376-399. Brown, Gregory, and Michael Cliff, 2004, Investor sentiment and the near-term stock market, Journal of Empirical Finance, 11(1), 1-27. Brown, Gregory, and Michael Cliff, 2005, Investor sentiment and asset valuation, Journal of Business, 78(2), 405-440. Brown, Nerissa, Theodore Christensen, W. Brooke Elliott, and Richard Mergenthaler, 2012, Investor sentiment and pro forma earnings disclosures, Journal of Accounting Research, 50(1), 1-40. Cohen, Daniel, Aiyesha Dey, Thomas Lys, and Shyam Sunder, 2007, Earnings announcement premia and the limits to arbitrage. Journal of Accounting and Economics, 43(2-3), 153-180. Frazzini, Andrea, and Owen Lamont, 2007, The earnings announcement premium and trading volume, National Bureau of Economic Research working paper. Hribar, Paul, and John McInnis, 2012, Investor sentiment and analysts’ earnings forecast errors, Management Science, 58(2), 293–307. Lee, Charles, Andrei Shleifer, and Richard H. Thaler, 1991, Investor sentiment and the closed-end fund puzzle, Journal of Finance 46(1), 75–109.

25

Lehmann, Bruce, 1990, Fads, martingales, and market efficiency, Quarterly Journal of Economics, 105(1), 1-28. Lemmon, Michael, and Evgenia Portniaguina, 2006, Consumer confidence and asset prices: some empirical evidence, Review of Financial Studies, 19(4), 1499-1529. Livnat, Joshua, and Christine Petrovits, 2009, Investor sentiment, post-earnings announcement drift, and accruals, New York University working paper. Ljungqvist, Alexander, Vikram Nanda, and Rajdeep Singh, 2006, Hot markets, investor sentiment, and IPO pricing, Journal of Business, 79(4), 1667-1702. Mian, G. Mujtaba, and Srinivasan Sankaraguruswamy, 2012, Investor sentiment and stock market response to earnings news, The Accounting Review, 87(4), 1357-1384. Mikhail, Michael, Beverly Walther, and Richard Willis, 2009, Does investor sentiment affect sell-side analysts’ forecast accuracy?, Arizona State University working paper. Miller, Edward, 1977, Risk, uncertainty, and divergence of opinion, Journal of Finance, 32(4), 1151–1168. Neal, Robert, and Simon Wheatley, 1998, Do measures of sentiment predict returns?, Journal of Financial and Quantitative Analysis, 33(4), 523-535. Qiu, Lily, and Ivo Welch, 2004, Investor sentiment measures, NBER Working Paper No. 10794. Ritter, Jay, 1991, The long-run performance of initial public offerings, Journal of Finance, 46(1), 3-27. Seybart, Nicholas, and Holly Yang, 2012, The Party’s Over: The role of earnings guidance in resolving sentiment-driven overvaluation, Management Science, 58(2), 308-319. Stambaugh, Robert, Jianfeng Yu, and Yu Yuan, 2012, The short of it: investor sentiment and anomalies, Journal of Financial Economics, 104(2), 288-302. Thompson, Samuel, 2011, Simple formulas for standard errors that cluster by both firm and time, Journal of Financial Economics, 99(1), 1-10. Yu, Jianfeng, and Yu Yuan, 2011, Investor sentiment and the mean-variance relation, Journal of Financial Economics, 100(2), 367-381.

Panel A: Monthly coefficients on

Panel B: Monthly coefficients on

Figure I Monthly coefficients from regressions of earnings announcement return on firm-specific sentiment

-0.1

-0.05

0

0.05

0.1

0.15

0.2

0.25

0.3

Dec

-73

Dec

-74

Dec

-75

Dec

-76

Dec

-77

Dec

-78

Dec

-79

Dec

-80

Dec

-81

Dec

-82

Dec

-83

Dec

-84

Dec

-85

Dec

-86

Dec

-87

Dec

-88

Dec

-89

Dec

-90

Dec

-91

Dec

-92

Dec

-93

Dec

-94

Dec

-95

Dec

-96

Dec

-97

Dec

-98

Dec

-99

Dec

-00

Dec

-01

Dec

-02

Dec

-03

Dec

-04

Dec

-05

Dec

-06

Dec

-07

Dec

-08

Dec

-09

Dec

-10

-6

-4

-2

0

2

4

6

Dec

-73

Dec

-74

Dec

-75

Dec

-76

Dec

-77

Dec

-78

Dec

-79

Dec

-80

Dec

-81

Dec

-82

Dec

-83

Dec

-84

Dec

-85

Dec

-86

Dec

-87

Dec

-88

Dec

-89

Dec

-90

Dec

-91

Dec

-92

Dec

-93

Dec

-94

Dec

-95

Dec

-96

Dec

-97

Dec

-98

Dec

-99

Dec

-00

Dec

-01

Dec

-02

Dec

-03

Dec

-04

Dec

-05

Dec

-06

Dec

-07

Dec

-08

Dec

-09

Dec

-10

11 ( )D Sent+

This figure reports the 2β (panel A) and 3β (panel B) coefficients for monthly regressions of

1 2 3

1 1,

1 ( ) 1 ( )jt jt jt jtc jt c jt

R UE UED Sent D Sent

α β β β ε= + + + ++ +

where jtR is the cumulative market-

adjusted return (computed using the CRSP value-weighted market index) over the three-day window

(days -1, 0, and 1) surrounding the announcement of quarter t earnings for firm j; jtUE is the difference

between firm j's earnings before extraordinary items for quarter t (Compustat item EPSPXQ) and for quarter t-4, normalized by the share price of firm j at the end of quarter t-4, and then multiplied by 100 (UE is

truncated at the top and bottom one percent of its distribution); jtSent is the cumulative market-adjusted

return for firm j over the period from six trading days before to two trading days before the announcement

of quarter t earnings; ( )c jtD Sent is the decile rank of jtSent , where the rank is determined relative to the

values of itSent for all firms, i, that announce in calendar quarter c. Our sample period spans the years

1973 through 2010.

11 ( )

UED Sent+

VariableNo. of

observationsMean Median Std. Dev.

25th percentile

75th percentile

Earnings announcement return 425,170 0.22% 0.000 0.082 -0.032 0.033

Sentiment 425,170 0.28% -0.001 0.070 -0.029 0.028

Market value 425,170 2,105 214 11,238 55 903

Book-to-market 425,170 0.987 0.632 7.057 0.371 1.012

Age 425,170 16.832 12.166 14.981 6.333 21.750

Unexpected earnings 425,170 0.335 0.186 3.917 -0.577 0.923

Return volatility 425,170 0.124 0.105 0.086 0.072 0.151

Earnings volatility 332,033 0.038 0.014 2.788 0.007 0.025

Earnings announcement

returnSentiment

Market value

Book-to-market

AgeUnexpected

earningsReturn

volatility

Sentiment -0.0868 1

Market value -0.0023 -0.0042 1

Book-to-market 0.0028 0.0030 -0.0143 1

Age -0.0004 -0.0153 0.2251 -0.0075 1

Unexpected earnings 0.1397 0.0494 -0.0049 -0.0036 -0.0116 1

Return volatility -0.0124 0.0451 -0.0792 -0.0042 -0.2170 0.0720 1

Earnings volatility 0.0006 -0.0013 -0.0017 -0.0009 -0.0072 0.0017 0.0048

Panel B: Pairwise Pearson correlation coefficients

Table IDescriptive statistics and correlation coefficients

Panel A: Descriptive statistics

This table provides sample descriptive statistics for our variables (panel A) and the average of the quarterly pairwise Pearson correlation coefficients (panel B). For a given quarter, t, earnings announcement return is equal to the cumulative market-adjusted return (using the CRSP value-weighted market index) over the three-day window (days -1, 0, and 1) surrounding the announcement of quarter t earnings. Sentiment is defined as the cumulative market-adjusted return over the pre-announcement period (days -6 through -2), prior to the announcement of quarter t earnings. Market value is the end-of-quarter t market value of equity. Book-to-market is the end-of-quarter t book value of equity divided by the end-of-quarter t market value of equity (we set this variable to zero for firms with negative book value). Age is defined as the number of years a firm’s shares have been publicly trading, as of the end of quarter t. Unexpected

earnings is computed as the difference between earnings before extraordinary items for quarter t (Compustat item EPSPXQ) and for quarter t-4, normalized by the share price at the end of quarter t-4, and then multiplied by 100. Unexpected earnings is truncated at the top and bottom one percent of its distribution. Return volatility is computed as the standard deviation of monthly stock returns over months t-12 through t-2. Earnings volatility is defined as the standard deviation of the ratio of quarterly operating income before

depreciation to average total assets, calculated over the 20 quarters prior to quarter t (with a minimum of 8 out of 20 quarters required for computation). Our sample period spans the years 1973 through 2010. In Panel B, numbers in bold are significant at the 10% level or better.

Coefficient estimates Panel regressionPanel regression including control

variablesRank regression

Fama-MacBeth monthly rank

regressions(1) (2) (3) (4)

Intercept -0.007*** -0.00003 -0.0319*** -0.035***-32.0 -0.7 -67.7 -41.2

UE 0.277*** 0.245***48.0 42.0

Rank(UE) 0.011E-5*** 0.013E-5***60.3 35.6

0.044*** 0.049*** 0.040*** 0.043***43.5 48.5 21.2 12.7

0.096*** 0.093***4.3 4.2

0.044E-6*** 0.036***5.7 2.5

ln(Market value ) -0.0004***-7.1

ln(Book-to-market ) .0025***15.8

Momentum -0.003***-15.3

Dloss -0.018***-57.8

Adjusted R-squared 0.024 0.033 0.038 0.046No. of observations 425,170 425,170 425,170 436

Table IIEarnings announcement returns and firm-specific sentiment

11 ( )D Sent+

11 ( )

UED Sent+

This table reports the coefficients for regressions of 1 2 3

1 1,

1 ( ) 1 ( )jt jt jt jtc jt c jt

R UE UED Sent D Sent

α β β β ε= + + + ++ +

where jtR is the cumulative market-adjusted return (using the CRSP value-weighted market index) over the three-day

window (days -1, 0, and 1) surrounding the announcement of quarter t earnings for firm j; jtUE is the difference

between firm j's earnings before extraordinary items for quarter t (Compustat item EPSPXQ) and for quarter t-4, normalized by the share price of firm j at the end of quarter t-4, and then multiplied by 100 (UE is truncated at the top

and bottom one percent of its distribution); jtSent is the cumulative market-adjusted return for firm j over the period

from six trading days before to two trading days before the announcement of quarter t earnings; ( )c jtD Sent is the decile

rank of jtSent , where the rank is determined relative to the values of itSent for all firms, i, that announce in calendar

quarter c; ln( )jtMarket value is the log of the end-of-quarter t market value of equity of firm j; ln( - - )jtBook to market is

the log of the ratio of end-of-quarter t book value of equity of firm j to end-of-quarter t market value of equity of firm j

(we set the value of this variable to zero for firms with negative book value); jtMomentum is the cumulative monthly

stock return for firm j over months t-12 through t-2. DLoss is equal to one if the earnings of firm j in quarter t are

negative and zero, otherwise. For column 4, jtUE and 1

1 ( )jtc jt

UED Sent+

are replaced by the rank of each of these

variables. Below each coefficient value is the corresponding t-statistic. Our sample period spans the years 1973 through 2010. In column 2, significance is determined based on the time-series distribution of equally weighted monthly coefficient estimates. ***=Significant at the 1% level; **=significant at the 5% level; *=significant at the 10% level (all one-tailed).

1( )1 ( )

Rank UED Sent+

Panel A: I/B/E/S subsample resultsCoefficient estimates Panel regression

Intercept -0.041***-64.3

Rank(UE) 0.030E-5***60.0

0.034***13.3

0.014E-5***6.8

Adjusted R-squared 0.070No. of observations 220,495

Panel B: Results for I/B/E/S subsample with UE = 0

Coefficient estimates Panel regressionIntercept 0.011***

6.9

-0.003***-11.1

Adjusted R-squared 0.012No. of observations 9,700

Table IIIEarnings announcement returns and firm-specific sentiment for I/B/E/S subsample

11 ( )D Sent+

1( )1 ( )

Rank UED Sent+

Panel A reports the coefficients for regressions of

1 2 3

1,

1 ( )1( ) ( )

1 ( )jt jt jtc jt

jtc jt

R UED Sent

Rank Rank UED Sent

α β β β ε= + + + ++ +

for all observations with valid

individual forecasts in I/B/E/S that are outstanding seven days before the earnings announcement, where jtR is

the cumulative market-adjusted return (using the CRSP value-weighted market index) over the three-day window

(days -1, 0, and 1) surrounding the announcement of quarter t earnings for firm j; jtUE is the difference between

firm j's earnings as reported on I/B/E/S and the I/B/E/S consensus forecast (computed as the average of all valid individual forecasts in I/B/E/S that are outstanding seven days before the earnings announcement), normalized

by the share price of firm j at the end of quarter t-1, and then multiplied by 100 (in this case, UE is truncated at

the top and bottom one percent of its distribution); ( )jtRank UE is the rank of jtUE ; jtSent is the cumulative

market-adjusted return for firm j over the period from six trading days before to two trading days before the

announcement of quarter t earnings; ( )c jtD Sent is the decile rank of jtSent , where the rank is determined

relative to the values of itSent for all firms, i, that announce in calendar quarter c. Panel B reports the coefficient

for a regression of jtR on ( )c jtD Sent for the subsample of I/B/E/S observations with 0jtUE = . Below each

coefficient value is the corresponding t-statistic. Our sample period spans the years 1983 through 2010. ***=Significant at the 1% level; **=significant at the 5% level; *=significant at the 10% level (all one-tailed).

( )D Sent

With BW index onlyWith BW index and firm-specific

sentiment measure(1) (2)

Intercept 0.001*** -0.007***8.7 -32.2

UE 0.29*** 0.27***89.4 47.6

0.0443***43.5

0.096***4.3

0.00012*** 0.00013***3.5 3.9

-0.001 -0.00126-1.4 -1.3

Adjusted R-squared 0.0195 0.0240No. of observations 425,170 425,170

Table IVEarnings announcement returns, Baker-Wurgler (BW) index, and firm-specific sentiment

Columns (1) and (2) report the coefficients for regressions of 1 2 3

1 11 1jt jt jt jt

t t

R UE UEBWSent BWSent

α β β β ε′ ′ ′′= + + + ++ +

and 1 2 3 2 3

1 1 1 11 ( ) 1 ( ) 1 1

,jt jt jt jt jtc jt c jt t t

R UE UE UED Sent D Sent BWSent BWSent

α β β β β β ε′ ′= + + + + + ++ + + +

respectively, where Rjt is the cumulative market-adjusted return (using the CRSP value-weighted market index) over the three-day window (days -1, 0, and 1) surrounding the announcement of quarter t earnings for firm j; UEjt is the difference between firm j's earnings before extraordinary items for quarter t (Compustat item EPSPXQ) and for quarter t-4, normalized by the share price of firm j at the end of quarter t-4, and then multiplied by 100 (UE is truncated at the top and bottom one percent of its distribution); BWSentt is the value of the Baker-Wurgler (BW) index of market sentiment during the last month of quarter t. Sentjt is the cumulative market-adjusted return for firm j over the period from 6 trading days before to two trading days before the announcement of quarter t earnings; Dc(Sentjt) is the decile rank of Sentjt, where the rank is determined relative to the values of Sentit for all firms, i, that announce in calendar quarter c. Our sample period spans the years 1973 through 2010. Below each coefficient value is the corresponding t-statistic. ***=Significant at the 1% level; **=significant at the 5% level; *=significant at the 10% level (all one-tailed).

11 ( )D Sent+

11 ( )

UED Sent+

11 BWSent+

11

UEBWSent+

Age Size Earnings volatility Return volatility(1) (2) (3) (4)

Intercept -0.01*** -0.013*** -0.002*** -0.002***-20.34 -25.09 -3.37 -3.6

UE 0.28*** 0.29*** 0.28*** 0.28***48.13 49.79 40.93 48.13

0.047*** 0.069*** 0.042*** 0.036***23.4 34.1 15.0 14.1

0.13*** 0.508*** -0.007 0.294***4.23 18.39 -0.15 6.17

-0.0004*** -0.0010*** -0.0010*** -0.0010***

-5.54 -11.82 -10.9 -12.6

0.0008** 0.0051*** 0.0006* 0.0015***2.31 14.06 1.46 4.17

0.007* 0.121*** 0.013*** -0.026***

1.68 25.14 2.45 -5.05

Adjusted R-squared 0.0241 0.0259 0.023 0.0248No. of observations 425,170 425,170 332,033 425,170

Coefficient estimates

Table VEarnings announcement returns, firm-specific sentiment, the BW index, and hard-to-value firms

This table reports the coefficients for regressions of 0 1 2 31 1

1 ( ) 1 ( )jt jt jtc jt c jt

R UE UED Sent D Sent

β β β β= + + ++ +

1 2 31 1( ) ( ) ( ) ,

1 ( ) 1 ( )c jt c jt c jt jt jtc jt c jt

D HTV D HTV D HTV UED Sent D Sent

γ γ γ ε+ + + ∗ ++ +

where jtR is the cumulative market-adjusted return

(using the CRSP value-weighted market index) over the three-day window (days -1, 0, and 1) surrounding the announcement of quarter t earnings for firm j; jtUE is the difference between firm j's earnings before extraordinary items for quarter t (Compustat item EPSPXQ) and for

quarter t-4, normalized by the share price of firm j at the end of quarter t-4, and then multiplied by 100 (UE is truncated at the top and bottom one percent of its distribution); jtSent is the cumulative market-adjusted return for firm j over the period from six trading days

before to two trading days before the announcement of quarter t earnings; ( )c jtD Sent is the decile rank of jtSent , where the rank is

determined relative to the values of itSent for all firms, i, that announce in calendar quarter c; HTVjt is the value of the proxy (Age, Size,

Earnings volatility, and Return volatility) used to measure how hard it is to value firm j in quarter t; ( )c jtD HTV is the decile rank of jtHTV ,

where the rank is determined relative to the values of jtHTV for all firms, i, that announce in the same calendar quarter c (higher deciles

mean that the firms are more difficult to value). Age is defined as the number of years a firm’s shares have been publicly trading, as of the end of quarter t; Size is the end-of-quarter t market value of equity; Earnings volatility is defined as the standard deviation of the ratio of quarterly operating income before depreciation to average total assets, calculated over the 20 quarters prior to quarter t (with a minimum of 8 out of 20 quarters required for computation); Return volatility is computed as the standard deviation of monthly stock returns over months t-12 through t-2. Our sample period spans the years 1973 through 2010. Below each coefficient value is the corresponding t-statistic. ***=Significant at the 1% level; **=significant at the 5% level; *=significant at the 10% level (for t-statistics, all significance levels are one-tailed).

11 ( )

UED Sent+

11 ( )D Sent+

( )D HTV

1( )1 ( )

D HTVD Sent+

1( )1 ( )

D HTV UED Sent

∗+

Intercept -0.0005**-2.1

UE 0.287***49.8

0.010***9.9

0.081***3.7

Sent -0.110***-62.4

Adjusted R-squared 0.0286No. of observations 425,170

Panel A: Extending sentiment measurement window to (-60, -7) prior to earnings announcements

Table VIRobustness tests

Panel A reports the coefficients for regressions of 1 2 3 41 1 ,

1 ( ) 1 ( )jt jt jt jt jtc jt c jt

R UE UE SentD Sent D Sent

α β β β β ε= + + + + +′ ′+ +

where Rjt is the cumulative market-adjusted return over the three-day window (days -1, 0, and 1) surrounding the announcement of quarter t earnings for firm j; Sent’jt is the cumulative market-adjusted return for firm j over the period from 60 trading days before to seven trading days before the announcement of quarter t earnings; Dc(Sent’jt) is the decile rank of Sent’jt, where the rank is determined relative to the values of Sent’it for all firms, i, that announce in calendar quarter c; Sentjt is the cumulative market-adjusted return for firm j over the period from six trading days before to two trading days before the announcement of quarter t earnings; and UEjt is the difference between firm j's earnings before extraordinary items for quarter t (Compustat item EPSPXQ) and for quarter t-4, normalized by the share price of firm j at the end of quarter t-4, and then multiplied by 100 (UE is truncated at the top and bottom one percent of its distribution). Panel B reports the coefficients for regressions of

1 2 3 4 51 1 1 1 1x

1 ( ) 1 ( ) 1 ( ) 1 ( ) 1 ( )jt jt jtc jt c jt c jt c jt c jt

R UE UED Sent D Sent D Sent D Sent D Sent

α β β β β β= + + + + +′ ′ ′+ + + + +

61 1x

1 ( ) 1 ( )jt jtc jt c jt

UED Sent D Sent

β ε+ +′+ +

, where Dc(Sentjt) is the decile rank of Sentjt, where the rank is determined relative to

the values of Sentit for all firms, i, that announce in calendar quarter c. Panel C reports the coefficients for regressions of

1 2 3 41 1 ( )

1 ( ) 1 ( )jt jt jt c jt jtc jt c jt

R UE UE D DiffD Sent D Sent

α β β β β ε= + + + + ++ +

, where Diffjt is a proxy for difference of opinion

prior to firm j’s announcement of quarter t earnings (the two proxies used are earnings volatility and return volatility); Dc(Diffjt) is the decile rank of Diffjt, where the rank is determined relative to the values of Diffit, for all firms, i, that announce earnings in calendar quarter c; Earnings volatilityjt is the standard deviation of the ratio of quarterly operating income before depreciation to average total assets for firm j, calculated over the 20 quarters prior to quarter t; Return volatilityjt is the standard deviation of monthly stock returns for firm j (calculated over months t-12 through t-2). Our sample period spans the years 1973 through 2010. Below each coefficient value is the corresponding t-statistic. ***=Significant at the 1% level; **=significant at the 5% level; *=significant at the 10% level (all one-tailed).

11 ( )D Sent′+

11 ( )

UED Sent′+

Intercept -0.006***-14.0

UE 0.279***48.2

-0.005***-2.9

0.059**2.1

0.0323***16.9

0.055***7.5

0.175**2.2

Adjusted R-squared 0.0242No. of observations 425,170

Earnings volatility Return volatility(1) (2)

Intercept -0.002*** -0.004***-7.7 -12.0

UE 0.27*** 0.283***40.9 49.0

0.046*** 0.046***39.2 45.4

0.092*** 0.087***3.5 3.9

-0.0009*** -0.0007***-18.7 -18.0

Adjusted R-squared 0.023 0.0247No. of observations 332,033 425,170

Panel C: Incorporating proxy for dispersion of opinion

Panel B: Incorporating both long and short sentiment measurement windows

Table VI - Continued

11 ( )D Sent+

11 ( )

UED Sent+

( )D diff

11 ( )D Sent′+

11 ( )

UED Sent′+

11 ( )D Sent+

1 1x1 ( ) 1 ( )D Sent D Sent′+ +

1 1x1 ( ) 1 ( )

UED Sent D Sent′+ +