Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
The role of accounting information in the sentiment-price relation
(USC Thesis Other)
The role of accounting information in the sentiment-price relation
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
THE ROLE OF ACCOUNTING INFORMATION IN THE
SENTIMENT-PRICE RELATION
by
Kun-chih Chen
A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(BUSINESS ADMINISTRATION)
May 2009
Copyright 2009 Kun-chih Chen
ii
Acknowledgements
I am indebted to my dissertation co-chairs: Drs. Sarah Bonner and Shiing-wu Wang and
to my other dissertation committee members: Drs. John Matsusaka, Nerissa Brown, and
Cheng Hsiao for their encouragement and guidance. Workshop participants at the
University of Southern California, the Pennsylvania State University, Hong Kong
University of Science and Technology, Chinese University of Hong Kong, Singapore
Management University, and INSEAD provided many useful comments. I am grateful to
the Leventhal School of Accounting and Marshall School of Business at the University of
Southern California for financial support. All remaining errors are mine.
iii
Table of Contents
Acknowledgements ii
List of Tables iv
Abstract v
Chapter 1: Introduction 1
Chapter 2: Literature review and hypotheses development 5
2.1 Sentiment-price relations in the behavioral finance literature 5
2.2 Effects of sentiment in the psychological literature 7
2.3 Hypotheses based on the effect of sentiment on judgment and
risk-preferences 8
Chapter 3: Measures and hypotheses testing 12
3.1 Measures for investor sentiment 12
3.2 Measures of uncertainties inherent in accounting information 14
3.3 Sample selection 18
3.4 Hypotheses testing methods 20
3.5 Results based on Fama-MacBeth regression 22
3.6 Results based on pooled regression 25
3.7 Supplemental tests in short-windows 28
3.8 Sensitivity tests based on firm characteristics 33
Chapter 4: Conclusions 36
References 39
Appendix: Hypotheses development based on log transformation of
the discount cash flow model 42
iv
List of Tables
Table 1: Factor analyses: measures of investor sentiment based surveys
and market Activities 13
Table 2: Earnings fluctuations 16
Table 3: Descriptive statistics of the public trading companies used in the study 18
Table 4: Long window Fama-MacBeth regression: the effect of sentiment on
price reactions to decomposed earnings 23
Table 5: Long window pooled regression: the interaction of sentiment and
decomposed earnings 27
Table 6: Short window test: sentiment-Price relation around earnings
announcements 31
Table 7: Sensitivity test: sentiment price relation of firms with different
degrees of return volatility and BM ratios 34
v
Abstract
This study reconciles inconsistent evidence on the sentiment-price relation in
prior studies by explicitly considering the effects of sentiment on both investor judgments
and risk preferences. Using the uncertainty in accounting information, I am able to
disentangle these two effects of sentiment and investigate the causes of the variations in
the sentiment-price relation. The results show that, under low uncertainty, the effect of
sentiment on risk preferences dominates in the sentiment-price relation, such that a
negative effect of sentiment on price is observed. In contrast, under high uncertainty, the
effect is less negative and, in fact, becomes positive. This suggests that in cases of high
information uncertainty, the effect of sentiment on judgments dominates in the sentiment-
price relation.
1
Chapter 1: Introduction
Most prior studies show that stock prices, on average, increase with increases in
investor sentiment. This positive effect of sentiment on price, however, is not universally
supported by empirical evidence. For example, Baker and Wurgler (2007) find a negative
relation between sentiment and price for stable stocks.
1
Because the underlying process
of the effect of sentiment on stock prices is not fully understood, researchers have only
been able to speculate on the causes for observed variations in the sentiment-price
relation. This study reconciles prior mixed findings by incorporating investors’
judgments and risk preferences in the sentiment-price relations. The results show that the
direction of the sentiment-price relation can be predicted when investors’ judgments and
risk preferences are considered.
In prior studies of the sentiment-price relation, sentiment has been defined as the
judgments of market participants, and high sentiment as optimistic judgments (e.g.,
Brown and Cliff 2004, Baker and Wurgler 2007). However, this definition can only
explain a positive sentiment-price relation; the existence of a negative sentiment-price
relation in some studies suggests that judgments and sentiment are not necessary the
same construct. In this study, I adopt a more comprehensive approach and, based on the
psychology literature, define sentiment as an affective phenomenon. Sentiment refers to
feelings of people, either good or bad. These feelings influence their judgments.
Some studies have documented a strong effect of sentiment on individual risk
preferences. However, the effect of sentiment on risk preferences has not often been
considered in studies of the sentiment-price relation, even though psychology studies
1
Baker and Wurgler (2007) rank stocks based on their past twelve-month return volatility and find that
stocks with low volatility have a negative sentiment-price relation.
2
have found that people are likely to make more optimistic judgments and are likely to be
more risk-averse during periods of high sentiment. Conversely, people are likely to make
more pessimistic judgments and to be less risk-averse during low sentiment periods.
These two effects of sentiment lead to different predictions regarding the sentiment-price
relation. Optimism (pessimism) brought about by high (low) sentiment will increase
(decrease) investors’ judgment of expected future cash flows, indicating a positive
sentiment-price relation. Meanwhile, high (low) sentiment will lead to an increase
(decrease) in risk-aversion, which in turn will lead to a higher (lower) discount rate. This
effect implies a negative sentiment-price relation.
2
It is likely that the variation in the
sentiment-price relation is attributed to the interplay of these two contradictory effects.
To correctly predict the sentiment-price relation, we need to disentangle the
effects of sentiment on judgments and risk preferences by identifying conditions in which
one effect predominates over the other. Prior studies suggest that the level of uncertainty
changes the magnitude of the effect of sentiment on judgments. When uncertainty is high,
the effect of sentiment on judgments is strongest (e.g., Hirshliefer 2001, Zhang 2006b),
because substantive processing is required for these judgments (Forgas 1995). High
uncertainty allows for greater variation in judgments to occur as a result of variations in
sentiment. In contrast, when uncertainty is low, the effect of sentiment on judgments is
likely to be small because little processing effort is required and little variation is allowed.
Hence, I hypothesize that the effect of sentiment on risk preferences will stand out in low
uncertainty cases and the sentiment-price relation will be negative. In contrast, in high
2
These two contradictory effects of sentiment are consistent with the anecdotal evidence from the vector
autoregressive studies (e.g., Vuolteenaho 2002) which suggest that positive cash flow shocks often times
come with positive discount rate shocks.
3
uncertainty cases, the effect of sentiment on judgments will be stronger and the
sentiment-price relation will be more positive.
To test the hypotheses, I focus on the level of information uncertainty inherent in
specific earnings components
3
and conduct two sets of empirical analyses to examine
how the sentiment-price relation differs between earnings components with high versus
low information uncertainty. The results from both empirical analyses are consistent with
my predictions. Specifically, for earnings components with low uncertainty, the effect of
sentiment on risk preferences prevails: when sentiment increases (decreases), stock price
decreases (increases). In other words, there is a negative sentiment-price relation. For
components with high uncertainty, the judgment effect is so strong as to dominate the
risk preference effect: when sentiment increases (decreases), price responses increase
(decrease). The results of low uncertainty components reveal the existence of usually
unobservable discount rate shocks. The results of high uncertainty components further
confirm that cash flow shocks are usually much larger than discount rate shocks
(Vuolteenaho 2002).
The revelation of the effect of sentiment on risk preferences helps explain some of
the findings in prior literature on behavioral finance. For example, the negative sentiment
beta for firms with low uncertainty (Baker and Wurgler 2007) and the strong reaction to
bad news during low sentiment periods (Mian and Sankaraguruswamy 2008) can both be
attributed to a depressed discount rate brought about by the declining risk-aversion of
investors during low sentiment periods. The existence of a risk preference effect further
3
Conceptually, uncertainty may arise from less than complete knowledge of one’s own relationship with
the environment, the relationship between actions and outcomes, or states of the future (Downey and
Slocum 1975). This paper examines uncertainties inherent in information which is “the ambiguity with
respect to the implications of new information for a firm’s value” (Zhang 2006b, p.567); therefore, it is
more consistent with the later two sources of uncertainty. Specifically, I examine the uncertainty of future
benefits generated by different earnings components.
4
suggests that, in addition to high uncertainty cases, low uncertainty cases can also
account for the effects of sentiment. Also, instead of relying on variations in the general
information environment (e.g., Zhang 2006a, Baker and Wurgler 2007), this paper shows
the influence of behavioral factors directly through investors’ reactions to specific
components of accounting information.
This paper also contributes to the accounting literature. I show that the effect of
sentiment explains part of the temporal variations of earnings response coefficients. The
effect is evident when I decompose earnings components based on the uncertainty of the
future benefits they generate. For earnings items with high (low) uncertainty, the effect of
sentiment on price is positive (negative). This paper responds to calls for research that
investigates behavioral factors to explain temporal variations in earnings response
coefficients (e.g., Kothari and Shanken 2003) and shows that, in order to conduct this
type of research, an understanding of the interactions between behavioral factors and the
characteristics of accounting information is required.
The remainder of the paper is structured as follows: Section 2 provides a brief
literature reviews and describes the development of the hypotheses; Section 3 presents
data used in the study and methods of analyses; Section 4 reports the results of the study,
discusses the findings, their implications, and suggests directions for future research.
5
Chapter 2: Literature Review and Hypotheses Development
In this section, I first review the behavioral finance literature on the sentiment-
price relation and then the psychology literature on the effects of sentiment, especially its
impact on judgments and risk preferences. I then develop hypotheses based on this
review of the literature.
2.1 Sentiment-price relations in the behavioral finance literature
Prior research shows that the sentiment-price relation is mostly positive (e.g.,
Saunders 1993, Hirshliefer and Shumway 2003). However, several studies find that both
the magnitude and direction of the effect of sentiment on stock prices vary greatly among
stocks. For example, the positive sentiment-price relation is stronger for stocks that are
hard to evaluate and difficult to arbitrage (e.g. Glushkov 2006, Baker and Wurgler 2006).
More important, the sentiment-price relation is negative for stable stocks (e.g., Baker and
Wurgler 2007). To explain this negative “sentiment beta” for stable stocks requires a
clear definition of sentiment as well as a theory about the effects of sentiment.
Baker and Wurgler (2007) define sentiment as pessimism or optimism about stock
prices in general. However, they argue that it is the changes in demand for both stable
and volatile stocks that create the negative sentiment-price relation they find.
Specifically, when sentiment is high (low), investors prefer volatile (stable) stocks to
stable (volatile) stocks; therefore, they buy volatile stocks and sell stable ones. An
intuitive follow-up question would then be: why do volatile stocks become more
attractive and stable stocks become less so when sentiment changes in this way? One
possibility is that, as investors become more optimistic, they think volatile stocks will
perform better. Another possibility is that investors become less risk-averse and find the
6
depressed price of volatile stocks attractive. Baker and Wurgler (2007) does not consider
the possibility for sentiment to affect risk preferences. In other words, they treat the first
explanation as the driving force; as such, they treat the issue of whether the sentiment
beta will become negative for stable stocks as an empirical question. They view it as an
empirical question because they presume that, since sentiment only changes judgments, a
negative sentiment beta will appear only when investors substitute stable stocks for
volatile stocks. No theory suggests that this substitution effect is sure to happen. In fact,
prior studies document that the demand curves for most stocks is downward sloping (e.g.,
Shleifer 1986, Wurgler and Zhuravskaya 2002), suggesting lack of close substitutes for
most of the stocks. Therefore, a negative sentiment-price relation cannot be attributed
fully to investors’ optimistic judgments and the substitution of one stock for another.
The above evidence suggests that sentiment-related changes in risk preferences
may play an important role in the negative price-sentiment relation observed in stable
stocks. Yet prior studies either ignore the effect of sentiment on risk preferences (e.g.
Brown and Cliff 2004, Baker and Wurgler 2007) or have the misperception that
sentiment changes risk-preferences in the same way as it changes judgments. I believe the
misperception comes mostly from the confusion between “judgments of risk” and “risk
preference.”
4
Indeed, when investors are more optimistic, they judge cash flow to be high
and risk low; however, these are still judgment effects instead of a risk preference effect.
5
4
In a traditional discount cash flow model, the judgment about risk was pushed down to the denominator in
the equation. The discount rate is thus a combination of judgments on risk and risk preferences.
Alternatively, we can keep judgment on risk in the numerator with judgment on cash flow and leave only
the risk preference in the denominator. This approach is similar to Rubinstein (1976). In this way, all
judgment effects are kept in the numerator and risk-preferences in the denominator. In this framework, the
confusion of “judgment on risk” and “risk preference” will be mitigated.
5
In a theoretical model, Jouini and Napp (2007) show that pessimism (optimism) does change the objective
judgment of the market price of risk just as it changes the judgment on expected cash flows. However, they
also point out changes in judgment are not due to pessimistic consumers being less risk-averse (p.1161) and
7
2.2 Effects of sentiment in the psychological literature
Findings from the psychology studies provide some guidance for the investigation
of the variation in sentiment effects. Before reviewing those studies, however, it is
helpful to define the term “sentiment” used in those studies. Most psychology studies use
the more general term affect instead of sentiment. Affect refers to an evaluative reaction
to a stimulus that has either positive or negative valence (Fiske and Taylor 1991).
Sentiment is one of four types of affective phenomena (Ben-ze’ev 2001). Compared with
the other three, sentiment has specific intentionality, is longer, more stable, and more
dispositional. Although differing on various dimensions, all four phenomena fall into the
basic positive-negative valence structure of affect.
6
In this study, I use only the basic
positive-negative valence structure of affect and thus consider the following theories
about “positive-negative affective states” or “good-bad mood” appropriate for “high-low
sentiment states”.
Psychology theories suggest that affective phenomena such as sentiment influence
both people’s judgments and risk preferences.
7
When sentiment is high, people are more
optimistic (e.g., Isen et al, 1978, Johnson and Tversky 1983, Loewenstein et al. 2001),
because they are primed by the sentiment state and are likely to retrieve affect-congruent
(positive) memories (e.g. Isen et al. 1978, Wright and Bower 1992). At the same time,
they want to maintain the positive sentiment state and therefore are more cautious about
suggest that the relation between risk preferences and pessimistic/optimistic judgments be examined by
behavioral or psychological empirical studies (p.1163).
6
Isen (2004) points out that “affect is used here as the most general term for the emotion domain,[…] other
terms, such as emotions or mood, are avoided here only because they have connotations that may confuse
the issue; ‘affect’ is simply the most general term, encompassing all of the others” (p.264).
7
Affect changes several aspect of behavior. For example, people are more cooperative under positive affect
(Forgas 1994). More and more evidence suggests that behavioral factors influence people’s risk
preferences. For example, prospect theory research shows that people are more risk-averse in the gain
domain and less risk averse in the loss domain.
8
taking risks and demand a higher risk premium for the same amount of risk (e.g., Isen
and Patrick 1983, Nygren and Isen 1985, Isen et al. 1988, Nygren et al. 1996). That is,
optimistic individuals tend to be more risk-averse. On the contrary, when sentiment is
low, people are more pessimistic in their judgments. Meanwhile, they are more willing to
take chances in order to lift their sentiment level. In other words, they become more risk-
seeking (e.g., Mano 1992, Mittal and Ross 1998).
8
2.3 Hypotheses based on the effect of sentiment on judgment and risk-preferences
These two effects of sentiment suggested in psychology theories bear different
implications for stock prices. First, investors’ relative optimism or pessimism would
affect their judgments about future cash flows. Second, their risk preferences would
affect the discount rates they demand. The combined effects can be illustrated by the
following simple discounted cash flow (DCF) valuation model:
∞
Value
t
= Σ (Expected cash flow
n
/ (1 + subjective discount rate)
n-t
)
n= t+1
As sentiment increases, optimism increases and, thus, judgments of future cash
flows increase. However, as sentiment increases, people also become more risk-averse,
thus raising the discount rate they demand. As a consequence, stock prices can either go
up or down, depending on which of these two effects dominates. Prior research finds that,
on average, the price effects of changes in expected cash flows are much larger than
changes in discount rate (Vuolteenaho 2002). Indeed, the sentiment-price relation is
positive in most prior studies. A positive relation suggests that the effect of sentiment on
8
One general concern of applying behavioral factors to the stock market is whether all investors display the
same behavioral patterns. Increasing evidence shows that, in addition to individual investors, behavioral
factors also affect sophisticated investors such as mutual fund managers (e.g., Frazzini 2006) and analysts
(e.g., Zhang 2006b). In the sentiment literature, Brown and Cliff (2004) separate sentiment measures for
institutional investors from measures for individual investors and find no evidence supporting the
conventional wisdom that sentiment primarily affects individual investors and small stocks.
9
judgments predominates over its effect on risk preferences. Therefore, a plausible way to
disentangle these two effects is to identify situations where the effect of sentiment on
judgments is reduced greatly or muted completely, making the effect on risk preference
stands out.
Prior studies show that the effects of behavioral factors are more evident for
stocks with high uncertainty (e.g., Zhang 2006a, Baker and Wurgler 2006, 2007). Here,
uncertainty specifically means “information uncertainty”
9
which is “the ambiguity with
respect to the implications of new information for a firm’s value” (Zhang 2006b, p.567).
More detailed examinations suggest that it is the effect of sentiment on judgments that
become more evident. Affect Infusion Model (AIM) asserts that the effects of affect (or
sentiment) tend to be exacerbated in judgments based on ambiguous stimuli that demand
substantial cognitive processing. This occurs because substantive and prolonged
processing leaves greater room for affect-priming effects to occur (Forgas 1995).
10
This
affect-priming effect is the exact mechanism that leads to optimistic or pessimistic
judgments (Wright and Bower 1992). Consistent with the AIM, later studies also points
out that “mood states tend to affect relatively abstract judgments more than specific ones
about which people have concrete information” (Hirshliefer 2001, p.1551). Even
analysts’ judgments are subject to greater behavioral biases when information uncertainty
is high (Zhang 2006b). This is an important finding because judgments (in this case,
9
Zhang (2006b) shows that risk and information uncertainty are two separate constructs in his empirical
test. Forecasts/judgments of EPS or cash flows are not subject to risk. However, analysts’ forecasts still
exhibit more behavioral biases when there is greater information uncertainty. The other factor contributing
to the information uncertainty is “quality of the information,” which is assumed to be constant in this study.
10
In contrast, there is no evidence that information uncertainty will change the effect of sentiment on risk
preference.
10
analysts’ forecasts), unlike stock prices, are not subject to risk or market frictions but
only uncertainty.
In contrast to prior studies that use firm characteristics as surrogates for
uncertainties in the information environment and examine only high uncertainty firms, I
use earnings components that generate either high or low uncertainty in this study. Firm
characteristics such as the book-to-market ratio or firm size capture certain aspects of the
information environment, but it is hard to determine which aspects actually are captured.
Unlike firm characteristics, earnings components themselves represent information, and
this provides an opportunity to measure the effects of information uncertainty directly.
Furthermore, ERCs are comprised of expected cash flow and discount rate, which
represent the effect of sentiment on judgments and risk preferences, respectively.
Therefore, ERCs are ideal vehicles to separate the two effects of sentiment.
Combining the above empirical findings with psychology theories, I make the
following predictions: the effect of sentiment on risk preferences predominates over its
effect on judgments in investors’ reactions to earnings components that generate future
benefits with low uncertainties. Therefore, the rising (falling) risk aversion under high
(low) sentiment period will increase (decrease) the required discount rate and thus lower
(raise) the price reaction to these components. Accordingly, the relation between
sentiment and price reaction to these components will be negative. In contrast, the effect
of sentiment on judgments is stronger in investors’ reactions to earnings components that
generate future benefits with high uncertainties than it is in their reactions to components
with low uncertainties. Therefore, we should observe a stronger counter-effect of
judgments on risk preferences in these items, making the sentiment-price relation more
11
positive. Those predictions generate the following testable hypotheses:
11
HYPOTHESIS 1: The relation between sentiment and investors’ reactions to
earnings components with low uncertainty is negative.
HYPOTHESIS 2: The effect of sentiment on investors’ reactions to earnings
components with high uncertainty will be more positive than its effect on earnings
components with low uncertainty.
11
Alternatively, we can test the relation between sentiment and the degree of risk aversion. However, there
is no commonly accepted measure of risk aversion. Also, it is hard to distinguish changes in “investors’ risk
appetite” from “changes in riskiness” of assets in an empirical setting (Kumar and Persaud 2003). Still,
Kliger and Levy (2003) find two measures of risk aversion from prior studies are negatively correlated with
investor sentiment measured by weather condition.
12
Chapter 3: Measures and Hypotheses Testing
3.1 Measures for investor sentiment
I considered possible measures of investor sentiment from a list of measures in
prior studies (Baker and Wurgler 2006, 2007; Brown and Cliff 2004, 2005; Glushkov
2006, Qiu and Welch 2005) and collected four survey-based and 11 market trading-based
measures.
12
The surveys include investor surveys from the American Association of
Individual Investors (AAII), the Michigan Consumer Sentiment Index (MCSI) from the
University of Michigan, the Index of Consumer Confidence (ICC) from the Conference
Board, and the Investor Intelligence Survey (II) from Chartcraft. The market trading
measures include NYSE turnover, dividend premium, odd-lot ratio, margin borrowing,
and short interests from the stock market; the closed-end-fund discount (CEFD), mutual
fund cash positions, net purchases of mutual fund from the mutual fund market; and
number of IPOs, share of equity in total aggregate issuing, and average first day IPO
return from the IPO market. Two of the four surveys each provide two different measures
of sentiment (current and expected); therefore, there are a total of 17 possible measures
for investor sentiment.
In order to include all 17 measures collected, the principal component analysis
runs first on a shorter time span from 1988 to 2005. This first principal component has an
Eigenvalue of 5.06 and can explain up to 30% of the variance of all the 17 variables.
Further rotation of the factor loadings suggests that five measures load on the first
principal component, including both the future and current indexes from MCSI, the future
and current indexes from ICC, and CEFD. Subsequently, I run the principal component
12
These 11 market-based measures cover trading activities from the stock market, the mutual fund market,
and the IPO market. Bond market and derivatives market data are not included either because data is not
publicly available or missing for a long time frame.
13
analysis from 1978 to 2005 with 10 measures that are available for this time period.
Seven measures, including the five measures from the first analysis, NYSE turnover
(TURN), and share of equity in total aggregate issuing (SERATIO) have the highest
factor loadings (>0.3) on the first principal component. Therefore, I use the first principle
component of those seven measures from 1978 to 2005 as the measure of investor
sentiment.
Table 1: Factor analyses
Independent variable
(t-statistic)
Model1 Model2 Model3 Model4
MCSICI 0.895 0.885 0.938 0.769
MCSIEI 0.880 0.842 0.937 0.912
CBCI 0.859 0.851 0.748 0.597
CBEI 0.889 0.881 0.715 0.943
TURN 0.207 0.107 0.644 0.268
SERATIO -0.164 -0.151 -0.479 0.001
CEFD 0.676 0.707 -0.308 -0.133
NIPO 0.098 0.110 0.199
RIPO 0.343 0.219 0.174
PDND -0.400 -0.332 -0.064
MUTCA -0.261
ODDLOT 0.211
II 0.102
AAII 0.368
MUTFL -0.295
MARGIN 0.056
SHORTIN 0.010
% of variance explained 30% 43% 36% 51%
KMO 0.750 0.715 0.653 0.664
Eigen value 5.07 4.33 3.65 3.57
Time period 1988-2005 1988-2005 1978-2005 1978-2005
The component matrix is rotated by Varimax method with Kaiser Normalization. The measures included
are current index of Michigan consumer sentiment index (MCSICI), expectation index of Michigan
consumer sentiment index (MCSIEI), current index of conference board consumer confidence index
(CBCI), expectation index of conference board consumer confidence index (CBEI), NYSE turnover
(TURN), dividend premium (PDND), odd-lot ratio (ODDLOT), margin borrowing (MARGIN), and short
interests from NYSE (SHORTIN); the closed-end-fund discount (CEFD), mutual fund cash positions
(MUTCA), net purchases of mutual fund (MUTFL), number of IPOs (NIPO) share of equity in total
aggregate issuing (SERATIO), average first day IPO return (RIPO), investor surveys from American
Association of Individual Investors (AAII), Investor Intelligence Survey from Chartcraft (II).
14
Note that this sentiment measure from the principle component analyses relies
heavily on surveys of consumers. Prior research (e.g., Lemmon and Portniaguina 2006,
Qiu and Welch 2006) shows that consumer sentiment surveys capture a construct
different from the sentiment measures in Baker and Wurgler (2006) and the closed-end-
fund-discount (CEFD) in Lee et al. (1991). The differences are quite apparent: surveys
ask for people’s judgments only
13
, while market activities represent the effects of
sentiment on both judgments and risk preferences. Surveys better capture investor
sentiment as defined in this study because theories suggest that the relation between
sentiment and judgment is monotonic. In contrast, measures based on market activities
will capture either the effect of sentiment on judgments or its effect on risk preferences,
depending on which effect predominates over the other. Second, Qiu and Welch (2006)
find that, although these surveys do not ask for subjects’ views on securities prices
directly, changes in the MCSI are highly correlated with changes in the UBS/Gallup
survey, which does ask for investors’ views on securities prices. Furthermore,
consumption-based valuation models (Cochrane 2005) suggest a positive relation
between consumption and investment. Therefore, a sentiment measure relied heavily on
surveys of consumers is considered adequate for this study.
3.2 Measures for uncertainties inherent in accounting information
I propose capital expenditures, advertising expenses, and R&D expenses to be
earnings components with high information uncertainty. Consistent with this intuition,
13
Baker and Wurgler (2007) acknowledge that sentiment might show up first in investor judgments, which
could be surveyed. However, they point out “economists always treat surveys with some degree of
suspicion, because of the potential gap between how people respond to a survey and how they actually
behave” (p.135). Sharing the same concern but from another perspective, this is exactly why survey
measures are preferred in this study. Sentiment has a monotone effect on people’s judgments. In contrast,
for market activities, the effect of sentiment on judgments is confounded by its effect on preferences. The
gap between surveys and behavior is the effect of sentiment on preferences.
15
prior studies choose these three as earnings components that “bring future benefits with
high uncertainty” (Kothari et al. 2002). As for earnings components that generate future
benefits with low uncertainty, I choose extraordinary items, non-operating income, and
income from discontinued operations, because theoretically, they only have one-time
effects on cash flows. Several measures of uncertainty of accounting information are used
in prior studies, e.g., analysts’ forecast dispersion (e.g., Zhang 2006b), earnings
fluctuations (e.g., Kothari et al. 2002), and the standard deviation of residuals from a time
series forecast model (e.g., Lipe 1990). Whereas most measures do not disaggregate
earnings, Kothari et al. (2002) alone look at the uncertainty of future benefit from
disaggregated earnings components. Therefore, to validate my selections, I follow
Kothari et al. (2002) by regressing the standard deviation of five years of earnings per
share on the six earnings components.
[Model 1]
STD (EPS
t+1-t+5
) = a
1
+b
1
RD
t
+ b
2
ADV
t
+ b
3
CAPEX
t
+ b
4
EXTRA
t
+ b
5
EXTRA
t
*D +
b
6
DISCON
t
+ b
7
DISCON
t
*D + b
8
SPECIAL
t
+ b
9
SPECIAL
t
*D + b
10
SIZE
+ b
11
LEVERAGE +ε
RD is R&D expense (COMPUSTAT data item 46); ADV is advertising expense
(item 45); CAPEX is capital expenditures (item 128); EXTRA is extraordinary items
(item 192); SPE is special items (item 17); and DISCON is income from discontinued
operations (item 66). All of the above items are divided by shares outstanding (item 54)
and then deflated by price (item 199). LEV is leverage, calculated as long term debt
(item 9) plus short term debt (item 34) deflated by total assets (item 6). Size is the
logarithm of market capitalization as of the end of the fiscal year, calculated as the log
value of price (item 199) times total shares outstanding (item 54). Accounting numbers
16
with dummy variables (*D) are smaller than or equal to zero and numbers without
dummies are larger than zero. Dummy variables are added for all three transitory items
because they are directional (they can be either positive or negative); however, the
dependent variable (standard error) is always positive.
The coefficients of earnings components that bring future benefits with high
uncertainties (b
1
, b
2
, and b
3
) should be positive and significant. The coefficients of
components that bring low uncertainties should be negative for variables without a
dummy (b
4
, b
6
, and b
8
) and positive for variables with a dummy (b
5
, b
7
, and b
9
). Also,
earnings of larger firms should have less fluctuation than those of smaller firms (b
10
should be negative), and high-leverage firms should have more fluctuation in earnings
per share than firms with low debt (b
11
should be positive).
Table 2: Earnings fluctuations
STD (EPS
t+1-t+5
) = a
1
+b
1
R&D
t
+ b
2
ADV
t
+ b
3
CAPEX
t
+ b
4
EXTRA
t
+ b
5
EXTRA
t
_D +
b
6
DISCON
t
+ b
7
DISCON
t
_D + b
8
SPECIAL
t
+ b
9
SPECIAL
t
_D+ b
10
SIZE + b
11
LEVERAGE
+ε
Independent variable
(t-statistic)
Sign
prediction
Model 1
Deflated by Price
Model 2
Deflated by BVE
Intercept
0.0460** (8.92) 0.0805** (11.83)
R&D
+
0.1003** (6.19) 0.0915** (11.26)
ADV
+
0.0343** (3.28) 0.0167* (1.51)
CAPEX
+
0.0053 (0.76) -0.0144** (-2.72)
EOI
-
0.1734** (3.82) 0.1173** (4.73)
EOI_d
+
-0.0028** (-3.41) -0.0021 (-1.15)
SPE
-
-0.148** (-9.06) -0.1797** (-10.12)
SPE_d
+
0.0016** (2.55) 0.0022** (3.80)
DISCON
-
-0.1924** (-3.58) -0.1686** (-3.21)
DISCON_d
+
0.0009** (2.18) 0.0024* (1.47)
lmv
-
-0.0198** (-20.54) -0.0117** (-9.91)
lev
+
0.1020** (10.46) 0.0953** (7.83)
Adj. R squared 0.154 0.105
Sample size (firm-year) 49719 46067
**significant at one-tail 5% level *significant at one-tail 10% level
17
Table3: Continued
Note: I obtain data from the Compustat Annual Industrial and Research files 1978~2007. Dependent
variable SD(Et+1,t+5) is the standard deviation of earnings per share before extraordinary items and
discontinued operations (data item 58); the standard deviation is calculated using five annual earnings
observations for years t + 1 through t + 5. Each of the following variables except MV and Leverage is
deflated by book value of equity per share (item216/item54) or price (item199) at the end of the fiscal year
t − 1. Observations with negative book values are excluded. R&D (item46), CAPEX (item128), EOI
(item192), DISCON (item66), SPE (item17), and ADV (item45) are deflated by shares outstanding
(item54); MV is the natural log of the Market Value (item199*item54) at the end of fiscal year t; LEV is
the book value of debt (item9+item34) divided by total assets (item6) at the end of fiscal year t. Only
observations with all variables required are included in the model. Follow Kothari et al. (2002), I winsorize
all variables one percent each tail. Also, observations with deflated EPS values of less than −1 or greater
than 1 are winsorized at −1 and +1. In cases where EPS data are absent in any of the years from t + 1
through t + 5, SD(Et+1,t+5) is set equal to the mean SD(Et+1,t+5) of the firms in the same Altman Z-Score
decile portfolio. Z-score = 1.2*Working Capital (COMPUSTAT item179)/Total Assets (item6) +
1.4*Retained earnings (item36)/Total Assets (item6) + 3.3*Earnings before interest and Tax
(item170+item5-item62)/Total Assets (item6) + 0.6*Market value of Equity (item199*item25)/Book Value
of Total Liabilities (item181) + 0.999*Sales (item12)/Total Assets (item6). Reported t-tests are adjusted for
both time-series (year) and cross-sectional (firm) correlations as in Petersen (2008).
Table 2 presents the results of the regression. As predicted, R&D and
advertising expenses increase the standard deviations of five-year earnings per share.
Consistent with Kothari et al. (2002), R&D expenses bring the most fluctuation in future
earnings. Although the coefficient for capital expenditures is positive when deflated by
price, it turns into negative when deflated by book value of equity. By contrast,
discontinued operations and special items decrease the fluctuation of future earnings per
share. However, extraordinary items bring more fluctuations to future earnings. One
possibility is that firms with extraordinary items are extreme performers. Based on the
results of the above analyses, I select R&D and advertising expenses as earnings
components that generate high uncertainty and special items and discontinued operations
as components that generate little uncertainty for future earnings
18
3.3 Sample selection
Earnings data is obtained from COMPUSTAT and price data is obtained from
CRSP. Both surveys used in this study are not available on a monthly basis prior to 1978;
therefore, only data from 1978 to 2005 is included. Table 3, Panel A, presents the sample
selection criteria. The main tests include only firm-years for which net income
( COMPUSTAT data item 172), book value (item 60), special items (item 17),
discontinued operations (item 66), common shares (item 25), R&D expenses (item 46),
and advertising expenses (item 45) are available. Firms in the finance industry and utility
industry are excluded because these two industries are highly regulated and lack R&D
and special items data. This selection procedure results in 43,598 firm-year observations.
The average size of the firms in this sample is larger than the average firm size in
COMPUSTAT during the same period. Because behavioral biases are harder to find for
larger firms (e.g., Baker and Wurgler 2006), this sampling bias should work against
finding the predicted results.
Table3: Descriptive statistics
Panel A
Sample selection Criteria
firm size (MM)* MB ratio* Firm-year Observations 1978-2005 No. of
obs
Mean Median Mean Median
Total observations from COMPUSTAT 609162 927.91 69.80 3.12 1.65
Exclude financial industry and utility industry 458351 935.82 56.24 3.44 1.80
Exclude observations with missing price info 144920 1052.62 76.51 3.07 1.79
Exclude observations with missing accounting items 43899 1451.20 62.68 4.15 1.70
Exclude observations with missing control variables 43598 1026.03 64.86 2.93 1.70
*winsorized at 1% each tail
19
Table3: Continued
Panel B
Descriptive Statistics
Variable Mean S.D. Min 10th 25
th
Median 75
th
90th Max
Non-
zero %
PRC
bv
ni
ni_adj
rd
adv
special
discon
lmv
bm
16.44
9.34
0.75
1.96
0.57
0.54
-0.14
-0.02
-2.55
0.71
18.41
11.36
2.16
2.92
1.04
1.27
0.65
0.63
2.26
2.38
0.04
-10.66
-10.26
-3.70
0.00
0.00
-4.99
-22.94
-10.52
-242.72
1.50
0.49
-0.94
-0.27
0.00
0.00
-0.38
0.00
-5.28
0.14
4.10
2.22
-0.12
0.17
0.00
0.00
-0.01
0.00
-4.19
0.30
10.50
5.84
0.50
1.17
0.19
0.07
0.00
0.00
-2.77
0.57
22.63
12.26
1.50
2.77
0.67
0.47
0.00
0.00
-1.11
0.98
38.13
22.60
2.97
5.29
1.49
1.52
0.02
0.00
0.51
1.55
364.00
142.77
18.06
16.85
8.59
12.69
1.35
43.21
6.05
205.56
71%
70%
38%
9%
Note: The full sample consists of 43598 firm-year observations on the Compustat Annual Industrial
Research files from 1978–2005 with non-missing price information from CRSP and non-missing
accounting items including: BV (data60), NI (data172), R&D (data46), ADV (data45), DISCON (data66),
and SPE (data17). Ni_adj=data172+data46+data45-data17. All numbers are on per share basis. PRC is
price three month after fiscal year end. LMV is the natural log of the Market Value at the end of fiscal year
t; BM is the book value (data60) divided by market value (data199*data25) of equity at the end of fiscal
year t
Panel C
Pearson (above)/Spearman(below) correlation
Variable Sent price Eps Adj_eps bv adv rd Special discon
Sent 0.05 -0.06 -0.05 -0.07 0.01 -0.02 0.00 0.00
Price 0.02 0.54 0.68 0.64 0.27 0.45 -0.05 0.03
Eps -0.06 0.69 0.68 0.68 0.22 0.22 0.37 0.24
Adj_eps -0.05 0.76 0.81 0.77 0.63 0.59 -0.07 0.02
bv -0.07 0.75 0.60 0.72 0.30 0.47 -0.04 0.02
adv 0.06 0.23 0.17 0.39 0.22 0.00 -0.04 0.00
rd -0.02 0.33 0.12 0.33 0.26 0.14 -0.14 0.01
special 0.00 0.04 0.28 0.05 0.07 -0.04 -0.08 -0.01
discon 0.01 0.06 0.13 0.05 0.04 0.03 0.02 0.01
Correlations in bold are significant at the 5% level
Strong correlation (0.50 ~ 1.00) Median correlation (0.30 ~ 0.49)
Note: Sent is the sentiment score one month before the price date from this study. Price information is
taken from CRSP database. Price is the stock price three month after the fiscal year end date. Accounting
information is taken from COMPUSTAT annual database as defined in Panel B.
Table 3, Panel B, displays descriptive statistics for these accounting numbers.
These statistics indicate that over seventy percent of the firm-year observations have non-
zero R&D expenses and advertising expenses. Also, collectively, more than forty percent
20
of the firm-year observations have either special items or income from discontinued
operations. Panel C shows the correlation matrix. Among the accounting items, the
adjusted EPS has the highest correlation with stock price, suggesting that it captures
“core earnings” better than unadjusted EPS. Correlations among accounting numbers are
negligible except that both RD and ADV expenses have weak to medium positive
correlations with EPS. In addition, there is no significant correlation between sentiment
and individual accounting numbers.
3.4 Hypotheses testing methods
To observe how sentiment changes investors’ reactions to accounting information,
I take a two-step approach. First, I measure investors’ reactions to accounting information
in a cross-sectional regression. Second, I measure the effects of sentiment on investors’
reactions to accounting information in a time-series regression. In the following cross-
section equation, investors’ reaction for earnings component n of firm i, in time period t
is β
i,t,n
P
i,t
= α
i,t
+ β
n,i,t
× E
n,i,t,
+ ε
i,t
……………………………………………..…….(2)
To observe the effect of sentiment on investors’ reactions to accounting
information over time, I regress β
n,t
on the corresponding sentiment scores.
β
n,t
= a+ b
n,t
× sentiment
t
+ ε
n,t
…………………………………………….(3)
β
n,t
equals investors’ judgments of expected cash flow from earnings component n
divided by the objective discount rate. If the effect of sentiment on risk preferences
predominates over its effect on judgments, during high sentiment periods, the increase in
21
the discount rate will predominate over the increase in the cash flow forecast. Therefore,
sentiment will decrease price response to that earnings component and b
n,t
should be
negative. If the effect of sentiment on judgments predominates over its effect on risk-
preferences for component n during high sentiment periods, the increase in cash flow
forecasts will predominate over the increase in the discount rates. Therefore, sentiment
will increase the price response to that earnings component and b
n,t
should be more
positive compared with the case above. Appendix C provides a more detailed illustration
of the sign prediction for b
n,t
.
Running equation (2) and (3) separately is similar to the Fama-MacBeth two stage
regression approach. We can also combine both equations into a pooled regression.
Replacing β in equation (2) with equation (3), I obtain:
P = α + (a+ b × sentiment) × E + ε
Rearranging the terms, I obtain:
P = α + a× E + b × sentiment × E + ε…………………………………..(4)
Coefficient b, the effect of sentiment on people’s reactions to accounting
information, becomes the coefficient of the interaction term in the above pooled time-
series cross-section regression. In the following sections, I first run the two-stage model
(Equations 2 and 3), then examine the interaction terms in the pooled regression
(Equation 4).
22
3.5 Results based on Fama-MacBeth regression
This two-stage method has been adopted in prior research exploring temporal
variation in ERCs (e.g., Kothari and Shanken 2003). Following prior long-window
studies, the dependent variable in the first stage cross-section regression is price three
months after the fiscal year end date. This will give enough time for accounting
information to disseminate in the stock market. The independent variables are R&D
expense, advertising expenses, income from discontinued operations, special items, and
the adjusted EPS. Book value also is included to avoid a potential misspecification
problem (Penman and Yehuda 2004). Therefore, there is a total of six different
accounting numbers on the right hand side of the equation.
[Model 1-1]
P
t+3
= a
1
+ b
1
BV+ b
2
adj_EPS
t
+ b
3
RD
t
+ b
4
ADV
t
+ b
5
DISCON
t
+ b
6
SPE
t
+ controls + ε
BV is book value of the firm (COMPUSTAT data item 60) from the annual
statement; RD is R&D expense (item 46); ADV is advertising expense (item 45); SPE is
special item (item 17); and DIS is income from discontinued operations (item 66). The
above four earnings components are then divided by shares outstanding (item 25).
Adj_EPS is EPS (item 58) plus RD plus ADV minus SPE minus DIS. Control variables
include firm size, BM ratio, and future earnings growth ((EPS
t+3
-EPS
t
)/EPS
t
). Interaction
terms of all accounting numbers with firm size, BM ratio, future earnings growth, and the
absolute value of EPS are also included. Running this cross-sectional regression across
time, a time-series of ERCs for all six accounting items is obtained in the first stage.
23
In the second stage, those ERCs are regressed on the sentiment score from the
prior month and a linear time-series variable.
[Model 1-2]
Coefficient
t
= α
1
+ β
1
Sentiment
t-1
+ γ
1
time
t
+ ε
This analysis not only allows ERCs to vary across time, but also measures ERCs
at the market level. Therefore, a relatively large sample for each period is needed in order
to get a representative market-level coefficient. I include two sets of tests using months
with sample sizes larger than 200 and 300 respectively.
14
Table 4 shows the betas and
R
2
’s from these two time series regressions:
Table 4: Fama-MacBeth test
Step1: Cross-sectional regression
P
t+3
=a
1
+b
1
BV
t
+ b
2
Adj_EPS
t
+ b
3
RD
t
+ b
4
ADV
t
+ b
5
DISCON
t
+ b
6
SPE
t
+ control
variables+ ε
Step2: Time-series regression
Slope coefficient from step one (b
n,t
) = α
1
+ β
1
sentiment
t-1
+ β
2
time
t
+ ε
Dependent variable: coefficient
from step1 regression
BV EPS RD ADV DIS SPE
Panel A
β
1
in step 2
0.24**
(2.62)
-0.80*
(-1.95)
2.04†**
(3.06)
0.96†*
(1.55)
-16.37†
(-1.11)
-0.36†
(-0.63)
Adjusted R squared 0.20 0.09 0.18 0.05 0.03 0.04
Panel B
β
1
in step 2
0.20**
(2.31)
-0.86
(-1.57)
2.02†**
(2.23)
0.72†
(1.11)
-3.32†*
(-1.52)
-0.35†
(-0.71)
Adjusted R squared 0.24 0.09 0.17 0.05 0.17 0.02
**significant at 5% level *significant at 10% level †One tail tests
Note: dependent variable in the cross-sectional regression is stock price three month after the fiscal year
end date from CRSP database. BV (data item 60), R&D (data item 46), AdvEx (data item 45), SPE (data
item 17), and DISCON (data item 66) are from COMPUSTAT annual database and deflated by shares
outstanding (data item 25). Adj_EPS= EPS (data item 58)-SPE+ADV+RD All accounting numbers are
winsorized at 1% each tail. The model includes firm size (item199*item25), BM ratio
(item60/item199*item25), earnings growth in the next three years as control variables. All accounting
numbers are also interacted with firm size, BM ratio, earnings growth, and the absolute value of EPS.
14
If the threshold is set at 100, several months will have regressions with multi-collinearity problems.
Setting the threshold at 400 observations per month gives the same result as setting the threshold at 300.
24
Table4: Continued
In the time-series regressions, the variables are assumed to follow AR(1) process. All estimates are Yule-
Walker estimates. Dependent variables are the ERCs from the above cross-sectional equations. Investor
sentiment is from one month before the stock price date. Panel A includes only 46 months where the
number of observations each month is greater than 200. Panel B includes only 29 months where the
number of observations each month is greater than 300.
When sentiment goes up, investors are more risk-averse. However, their
judgments on the expected cash flows from special items and income from discontinued
operations remain intact. Because little cognitive processing effort is needed for those
judgments, the affect-priming effect is minimal; thus sentiment has little impact on
judgments here. Consequently, the effect of sentiment on risk preferences predominates
over its effect on judgments and thus the slope coefficients should be negative for special
items and income from discontinued operations. Indeed, I find β
1
to be negative for those
two items.
In the case of R & D and advertising expenses, the effect of sentiment on
judgments should be stronger. Because investors’ judgments on the expected cash flow
from R&D and advertising expenses requires much more processing effort, the strong
affect-priming effects make these judgments more easily swayed by sentiment. This
stronger effect on judgment then counters the effect on risk preferences and increases
investors’ reaction to R&D and advertising expenses when sentiment is high. The results
from Table 4 show that the results are consistent with this prediction. In fact, the effect on
judgments is so strong that it predominates over the effect on risk-preferences for those
two items, so that β
1
is positive for both R&D and advertising expenses. F statistics
further shows that β
1
for R&D is more positive than those of special items and income
from discontinued operations (F=6.33 p=0.02 and F=3.81 p=0.06, respectively). β
1
for
25
advertising expenses is also marginally more positive (F=1.79 p=0.19 and F=3.24
p=0.08, respectively). It is worth noting that the sign on adjusted EPS is negative and
significant. It suggests that adjusted EPS brings little uncertainty to future earnings after
adjusting for R&D, advertising expenses, and the transitory items.
Prior studies find that the value relevance of accounting information seems to
change across time (e.g., Core et al. 2003) and suggest that other behavioral factors
possibly play a role (e.g., Kothari and Shanken 2003). This analysis shows that sentiment
is one of the behavioral factors that explain the time-series properties of ERCs. Although
in both panels, advertising expenses, special items, and income from discontinued items
have the expected signs, they are not all significant. This might be due to the small
sample size used in the time-series regression. The combined regression in the next sub-
section helps address this problem.
3.6 Results based on pooled regression
Instead of running the cross-sectional regression and the time-series regression
separately, the following model combines those two regressions into one. The
combination resembles the model used in the prior section, where the second stage time-
series regressions are applied back to the first stage cross-sectional regression and the
coefficients of interest become the coefficients on the interaction terms.
[Model 2]
Price = a
1
+ b
1
Sent + b
2
BV + b
3
BV*sent + b
4
Adj_EPS + b
5
Adj_EPS*sent + b
6
RD
+ b
7
RD*sent + b
8
ADV + b
9
ADV*sent + b
10
SPE + b
11
SPE*sent + b
12
DIS
+ b
13
DIS*sent + Control variables +ε
26
Sent is the sentiment score two month after the fiscal year end date (which is
one month before the price date); BV is book value of the firm (COMPUSTAT data item
60) from the annual statement; RD is R&D expense (item 46); ADV is advertising
expense (item 45); SPE is special item (item 17); DIS is income from discontinued
operations (item 66). The above four earnings components are then divided by shares
outstanding (item 25). Adj_EPS is EPS (item 58) plus RD plus ADV minus SPE minus
DIS.
I expect the effect of sentiment on risk-preferences to predominate over its effect
on judgments in the price response to special items and income from discontinued
operations; therefore, b
11
and b
13
should be negative. Meanwhile, the effect of sentiment
on judgments should be stronger and counter the effect on risk preference for R&D and
advertising expenses. Therefore, the price response to those two items should be more
positive than that of the one-time items.
The model also controls for firm specific variables suggested by the prior ERC
literature: future earnings growth ((EPS
t+3
-EPS
t
)/EPS
t
), absolute value of EPS, sign of
EPS, and firm size. In addition to those firm characteristics, a linear time trend variable is
also included. To test the effect of sentiment on price response to earnings items, I use the
sentiment score one month before the stock price date in the model.
15
To test the model in
a panel data setting, the standard errors for all estimated coefficients are clustered in both
time-series (year) and cross-sectional (firm) to correct for unobserved time-series and
firm correlation patterns (Petersen 2008).
15
Untabulated tests show that replacing lagged sentiment with concurrent sentiment generates similar
results.
27
Table5: Pooled regression test
Price = a
1
+b
1
Sent +b
2
BV +b
3
BV*sent +b
4
Adj_EPS +b
5
Adj_EPS*sent +b
6
RD +b
7
RD*sent
+b
8
ADV +b
9
ADV*sent +b
10
SPE +b
11
SPE*sent +b
12
DIS +b
13
DIS*sent +Control variables +ε
Independent variable
(Petersen t-statistic)
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6
Intercept 16.5364**
(33.65)
7.5093**
(11.69)
18.0538**
(18.21)
17.7160**
(16.69)
17.9434**
(17.37)
16.3986**
(13.92)
Sentiment 0.8687**
(1.98)
1.7582**
(2.41)
-0.2641
(-0.66)
-0.2580
(-0.65)
-0.2126
(-0.52)
-0.2501
(-0.63)
BV 0.8049**
(9.40)
0.4642**
(4.93)
0.3877**
(4.24)
0.2752**
(3.21)
0.2053**
(2.06)
BV*sent 0.1141**
(3.33)
0.1154**
(3.35)
0.1151**
(3.95)
0.1442**
(4.96)
NI 2.1434**
(10.82)
2.3945**
(7.59)
3.1753**
(9.28)
2.3022**
(9.66)
3.7136**
(8.31)
NI*sent -0.1940
(-1.58)
-0.1220
(-0.64)
-0.2694**
(-2.99)
-0.4671**
(-2.37)
R&D -0.0333
(-0.0572)
-1.7633**
(-2.59)
RD*sent 0.2400†
(0.90)
0.5173†*
(1.41)
ADV -1.2912**
(-2.34)
-3.2899**
(-4.94)
ADV*sent 0.3333†**
(2.36)
0.5566†**
(3.05)
SPE 0.4443
(0.86)
1.1064**
(2.67)
SPE*sent -0.2584†*
(-1.60)
-0.2132†*
(-1.62)
DISCON 0.4654*
(1.74)
0.4640*
(1.78)
DISCON*sent -
0.2174†**
(-2.49)
-
0.2495†**
(-2.48)
Adj. R squared 0.0022 0.4588 0.6834 0.6890 0.6921 0.7067
Sample size (firm-
year)
43598 43598 43598 43598 43598 43598
**significant at 5% level *significant at 10% level †One tail tests
Dependent variable is stock price three month after the fiscal year end date from CRSP database. Investor sentiment
is one month before the stock price date. BV (data item 60), R&D (data item 46), AdvEx (data item 45), SPE (data
item 17), and DISCON (data item 66) are from COMPUSTAT annual database and deflated by shares outstanding
(data item 25). Adj_EPS= EPS (data item 58)-SPE+ADV+RD All accounting numbers are winsorized at 1% each
tail. Model 3~6 include untabulated control variables: firm factors (firm size, book to market ratio, future earnings
growth), a linear time trend variable. In addition, model 3~6 include interaction terms of all accounting numbers
with six control variables: future earnings growth ((EPS
t+3
-EPS
t
)/EPS
t
), absolute value of EPS, sign of EPS, firm
size, and a linear time trend variable. Standard errors of the coefficient are adjusted for time-series (year) and cross-
sectional (firm) correlations as suggested by Petersen (2008).
28
Table 5 presents the results of this analysis. The signs of the coefficients on the
interaction terms from Table 5 are all consistent with findings from the last section: both
b
7
and b
9
are positive, and b
11
and b
13
are negative.
16
The p-value from a one-tailed Wald
test confirms that both b
7
and b
9
are more positive than b
11
and b
13
. The results show that
people are more risk-averse when the sentiment score goes up; however, just as the result
in prior section, their optimistic forecasts on the benefit of R&D and advertising
overshadow the rise in risk-aversion.
There are some other interesting findings from Table 5. First, when interaction
terms are not added, the coefficient on sentiment is positive. This is consistent with prior
studies finding a positive correlation between sentiment and stock prices. However, when
the interaction terms are introduced, the coefficient of sentiment becomes negative. This
might imply that the positive correlation between sentiment and stock price mainly comes
from people’s reactions to accounting information. A path analysis is needed to validate
this speculation. Second, the interaction between sentiment and adjusted EPS is negative
and significant. It suggests that, after taking RD and ADV out of EPS, the uncertainty of
future benefit from this adjusted EPS is small. Therefore, the effect of sentiment on risk
preference prevails on the adjusted EPS.
3.7 Supplemental test in short-windows
The short-window test uses quarterly numbers rather than annual earnings figures.
The model is a regression of three day [-1, +1] cumulative abnormal return on earnings
innovation numbers and their interactions with sentiment. For firm i, the three day
abnormal return is calculated as the cumulative, size-adjusted return over a three-day
16
An F-test indicates that the coefficient on the interaction terms of RD and ADV are both statistically
significantly greater than that of SPE and DISCON at a p-value less than 0.0001.
29
window starting one day before the filing date:
1 1
ab_ret
i
[-1, +1] = ∏ (1+RET
i, t
) - ∏ (1+DECRET
i, t
)
t= -1 t= -1
where RET
i,t
is the return of stock i on date t relative to the firm’s quarterly earnings
announcement date, and DECRET
i,t
is the date t average return of all firms in the
corresponding size group.
Following Mian and Sankaraguruswamy (2008), the model transforms the
sentiment measure into a dichotomous variable (0 for negative sentiment score and 1 for
positive sentiment score) and includes only firms with positive income to increase the
power of the test. Following their analysis, the model also separates positive earnings
innovations from negative ones. Specifically, the regression is as follows:
ab_return[+1, -1] =a
1
+b
1
Sent +b
2
BV +b
3
BV*sent +b
4
P∆Adj_EPS +b
5
P∆Adj_EPS*sent
+b
6
N∆Adj_EPS +b
7
N∆Adj_EPS*sent +b
8
∆RD +b
9
∆RD*sent +b
10
P∆SPE
+b
11
P∆SPE*sent +b
12
N∆SPE +b
13
N∆SPE*sent +b
14
P∆DIS
+b
15
P∆DIS*sent +b
16
N∆DIS +b
17
N∆DIS*sent +Control variables +ε
where Sent is a dichotomous variable based on the sentiment score one month before the
month of the quarterly earnings announcement (0 for negative sentiment score and 1 for
positive sentiment score); BV is book value of the firm (COMPUSTAT data item 60)
from the quarterly statement; ∆Adj_EPS is adjusted EPS calculated as EPS (item 7) plus
RD (item 4/item 15) minus SPE (item32/item15) minus DIS (item 33/item 15) from
current quarter minus the adjusted EPS from the same quarter of the previous year; ∆RD
is R&D expense (item 4) from the earnings announcement divided by shares outstanding
(item 15) minus the corresponding R&D expense from the same quarter of the previous
30
year; ∆SPE is special item (item 32) from the quarterly statement divided by shares
outstanding minus the corresponding special items from the same quarter of the previous
year; ∆DIS is income from discontinued operations (item 33) from the quarterly
statement divided by shares outstanding minus the corresponding R&D expense from the
same quarter of the previous year. In addition to a linear time trend variable, four
variables suggested by Mian and Sankaraguruswamy (2008) are added for control
purposes: earnings growth (EPS this quarter minus EPS from last year for the same
quarter), earnings fluctuation (standard deviation of 4 quarterly EPS from previous year),
firm size, and absolute value of EPS.
Again, variables of interest are the interaction terms between accounting
numbers and sentiment. For transitory items, because the effect of sentiment on risk
preferences predominates over its effect on judgments, signs of the coefficient should be
negative. In contrast, judgments based on R&D require much more processing effort, and
therefore b
9
should be more positive.
Although sentiment effects should also exist in the short windows, this test
may not be powerful enough. Ideally, the model should include consensus forecasts as
the benchmark to calculate earnings innovations. However, separate forecasts for
decomposed earnings information do not exist. Therefore, I use a random-walk model to
calculate earnings innovations as the difference between the earnings component for the
current quarter and the same quarter of the previous fiscal year. While this approach
might introduce noise to the model, I contend that this is the most practical method.
31
Table 6: Short-window test
ab_return[+1, -1] =a
1
+b
1
Sent +b
2
BV +b
3
BV*sent +b
4
P∆Adj_EPS +b
5
P∆Adj_EPS*sent
+b
6
N∆Adj_EPS +b
7
N∆Adj_EPS*sent +b
8
∆RD +b
9
∆RD*sent +b
10
P∆SPE
+b
11
P∆SPE*sent +b
12
N∆SPE +b
13
N∆SPE*sent +b
14
P∆DIS +b
15
P∆DIS*sent +b
16
N∆DIS
+b
17
N∆DIS*sent +Control variables +ε
Independent
variable
(t-statistic)
Model1 Model2 Model3 Model4 Model5
Intercept 0.0115**
(33.84)
0.011**
(22.27)
0.0111**
(30.24)
0.0111**
(30.30)
-0.0127**
(-5.31)
Sent 0.0011*
(1.76)
0.0002
(0.31)
0.0006
(0.55)
P∆Adj_EPS 0.0004*
(1.60)
0.0004*
(1.6)
0.0002
(0.68)
0.0002
(0.69)
0.0091**
(7.66)
P∆Adj_EPS*se
nt
0.0021**
(3.09)
0.0007**
(2.57)
0.0017
(1.14)
N∆Adj_EPS 0.001
(0.51)
0.0007
(0.33)
0.0015
(0.64)
0.0014
(0.63)
0.0009
(0.22)
N∆Adj_EPS*s
ent
-0.0041
(-0.81)
-0.004
(-0.79)
-0.0098
(-1.34)
∆R&D -0.342*
(-1.86)
∆RD*sent 0.2723†**
(3.54)
P∆DISCON 0.3054
(0.43)
P∆DISCON*se
nt
-0.0688†
(-0.19)
N∆DISCON 0.8180
(1.15)
N∆DISCON*s
ent
-0.1976†
(-0.59)
P∆SPE -0.2059*
(-1.68)
P∆SPE*sent 0.0656†*
(1.37)
N∆SPE -0.0447
(-0.32)
N∆SPE*sent -0.0551†
(-0.55)
Adj. R squared 0.0001 0.0001 0.0001 0.0001 0.0291
Sample size
(firm-quarter)
99393 99312 99312 99312 61048
**significant at 5% level *significant at 10% level †One tail tests
32
Table6: Continued
Note: dependent variable is three day cumulative abnormal return around quarterly earnings announcement
date. Sent is a dichotomous variable based on the sentiment score one month before the month of the
quarterly earnings announcement (0 for negative sentiment score and 1 for positive sentiment score); BV is
the book value of the firm (COMPUSTAT data item 60) from the quarterly statement; ∆Adj_EPS is the
adjusted EPS calculated as EPS (item 7) plus RD (item 4/item 15) minus SPE (item32/item15) minus DIS
(item 33/item 15) from current quarter minus the adjusted EPS from the same quarter of the previous year;
∆RD is the R&D expense (item 4) from the earnings announcement divided by shares outstanding (item 15)
minus the corresponding R&D expense from the same quarter of the previous year; ∆SPE is the special
item (item 32) from the quarterly statement divided by shares outstanding minus the corresponding special
items from the same quarter of the previous year; ∆DIS is the income from discontinued operations (item
33) from the quarterly statement divided by shares outstanding minus the corresponding R&D expense
from the same quarter of the previous year. All earnings numbers are on per share basis and deflated by
stock price two days before the announcement date.
Prior studies on sentiment effects also obtain little success in the short window
test. Short window (event) studies mainly test the effect of “information innovations”.
However, Baker and Wurgler (2006) argue that sentiment affects “information as a
whole” rather than “information innovations”. Nevertheless, they claim the short window
test still serves as a lower bound for the effects of sentiment. Brown and Cliff (2004) find
little evidence of a relationship between investor sentiment and short-term stock market
performance. Brown and Cliff (2005) argue that “arbitrage forces are likely to eliminate
short-run mispricing but may break down at longer horizons”
17
(p.407) and they find
sentiment effects to be prevalent in the long window.
Table 6 presents the results of the short-window tests. Despite the low power of
these tests, the interaction between sentiment and R&D expenses is still positive and
significant. Three of the four interactions on transitory items are negative. However, all
of them are not significant. Consistent with prior research, the R
2
’s are low in these
regressions. The signs on positive earnings surprise and negative earnings surprise are the
17
Two examples of the limits to arbitrage in the long run are noise trader risk (DeLong et al. 1990) and the
interaction of agency costs and capital constraints (Shleifer and Vishny 1997).
33
same as in Mian and Sankaraguruswamy (2008). There is a jump in R
2
’s when the effect
of sentiment on different accounting numbers are added to the model. In general, the
results are consistent with the hypotheses.
3.8 Sensitivity test
Prior studies find that the effects of sentiment are stronger for large firms (Brown
and Cliff 2005) and firms with volatile stock returns (Baker and Wurgler 2007).
Specifically, the stock prices of large/volatile firms are higher during high sentiment
periods. I follow their analyses and rank the stocks based on firm size or the standard
deviation of monthly returns for the past 12 month. By running the regression separately
for each rank, I can observe how the sentiment effect based on accounting characteristics
changes with firm characteristics.
As Table 7, Panel A shows that both the effects that are conditional on firm
characteristics and those conditional on accounting characteristics exist. The coefficient
on sentiment alone is positive for large/volatile stocks and negative for small/stable
stocks. Consistent with prior studies, this result indicates that sentiment effects are
conditional on firm characteristics. At the same time, the relationship between accounting
items and sentiment generally stays the same across the rankings. In Panel A, this
relationship is more evident for firms in the middle rank (rank 2, 3, and 4). It shows that
accounting characteristics complement firm characteristics in showing the effects of
sentiment. In panel B, the pattern is not as strong in panel A. This might be due to stock-
return variability potentially removing the treatment effect of earnings components on
information uncertainty, as suggested by Kothari et al. (2002) and Beaver et al. (1970).
34
Table 7: Sensitivity test
Panel A (By Firm Size)
Price = a
1
+ b
1
Sent + b
2
BV + b
3
BV*sent + b
4
Adj_EPS + b
5
Adj_EPS*sent + b
6
RD+
b
7
RD*sent + b
8
ADV + b
9
ADV*sent + b
10
SPE + b
11
SPE*sent + b
12
SPE*D + b
13
SPE*D*sent +
b
14
DIS + b
15
DIS*sent + b
16
DIS*D + b
17
DIS*D*sent + Control variables +ε
Ranking
I.V.
1
Small
2 3 4 5
Large
Intercept 6.45**
(26.35)
14.38**
(19.26)
18.81**
(21.15)
21.530**
(33.71)
26.815**
(34.69)
Sentiment -0.240**
(-6.32)
-0.558**
(-7.61)
-0.956**
(-8.39)
-0.958**
(-6.24)
1.088**
(3.76)
BV 0.504**
(12.71)
0.047
(0.91)
-0.154**
(-2.71)
-0.225**
(-6.71)
0.939**
(24.59)
BV*sent 0.069**
(9.61)
0.104**
(14.09)
0.102**
(9.01)
0.133**
(10.00)
0.007
(0.42)
Adj_EPS 1.799**
(17.08)
1.673**
(9.09)
3.469**
(15.68)
4.824**
(23.29)
1.452**
(8.11)
Adj_EPS*sent 0.027
(0.87)
-0.227**
(-6.11)
-0.080
(-1.26)
-0.479**
(-6.42)
-0.128
(-1.45)
R&D 2.216**
(7.38)
-0.204
(-0.19)
-0.182
(-0.20)
-0.944**
(-2.73)
0.126
(0.40)
RD*sent -0.006†
(-0.09)
0.444†**
(4.53)
0.277†**
(2.29)
0.277†**
(2.19)
0.201†*
(1.46)
ADV -1.369**
(-12.38)
-0.718**
(-3.90)
-3.244**
(-11.17)
-4.011**
(-14.54)
-0.960**
(-2.96)
ADV*sent 0.039†
(0.83)
0.169†**
(3.24)
0.225†**
(2.33)
0.603†**
(5.58)
-0.142†
(-0.95)
SPE -0.223
(-0.90)
-1.296**
(-5.85)
-0.963**
(-2.75)
1.407**
(4.90)
-0.077
(-0.22)
SPE*sent -0.153†**
(-3.09)
-0.110†**
(-2.02)
-0.090†
(-1.03)
-0.255†**
(-2.48)
-0.187†*
(-1.63)
DISCON -0.395
(-0.47)
7.841**
(5.86)
-0.385**
(-0.45)
1.241*
(1.73)
0.178
(0.21)
DISCON*sent 0.083†
(0.64)
-0.013†
(-0.14)
-0.297†**
(-1.97)
-0.281†*
(-1.41)
0.227†
(0.64)
Adj. R squared 0.6244 0.6212 0.5670 0.5486 0.6198
**significant at 5% level *significant at 10% level †One tail tests
R&Dt, CapExt, AdvExt, Specialt, discont, and extrat are from the annual financial statement and deflated
by shares outstanding. Those accounting items are winsorized at 1% each tail. All models include
untabulated control variables: macroeconomic factors (change in durable goods consumption, non-durable
goods consumption, service, CPI, industrial production, unemployment rate), firm factors (log market
value, book to market ratio), year, and month dummy variables. In addition, we also include interaction
terms of all earnings items with four control variables: future earnings growth ((EPS
t+3
-EPS
t
)/EPS
t
),
absolute value of EPS, firm size, and a linear time trend variable.
future earnings growth ((EPS
t+3
-EPS
t
)/EPS
t
), absolute value of EPS, sign of EPS, firm size, book to market
ratio, and a linear time trend variable
35
Panel B (By Return Volatility)
Price = a
1
+ b
1
Sent + b
2
BV + b
3
BV*sent + b
4
Adj_EPS + b
5
Adj_EPS*sent + b
6
RD+
b
7
RD*sent + b
8
ADV + b
9
ADV*sent + b
10
SPE + b
11
SPE*sent + b
12
SPE*D + b
13
SPE*D*sent +
b
14
DIS + b
15
DIS*sent + b
16
DIS*D + b
17
DIS*D*sent + Control variables +ε
Ranking
I.V.
1
Stable
2 3 4 5
Volatile
Intercept 17.24**
(30.04)
17.39**
(34.94)
15.75**
(35.14)
16.253**
(37.47)
15.43**
(27.92)
Sentiment -1.26**
(-5.67)
-1.451**
(-7.70)
-1.624**
(-10.41)
-0.615**
(-4.45)
0.329**
(2.04)
BV -0.105**
(-3.96)
0.426**
(13.69)
0.229**
(6.96)
0.292**
(7.62)
0.413**
(6.97)
BV*sent 0.132**
(9.71)
0.074**
(5.11)
0.138*
(8.57)
0.071**
(3.82)
0.183**
(6.24)
Adj_EPS 3.983**
(24.8)
2.665**
(16.75)
4.044**
(23.72)
3.499**
(22.15)
2.130**
(10.26)
Adj_EPS*sent -0.431**
(-4.99)
-0.018
(-0.21)
0.187*
(2.40)
0.300**
(3.51)
-0.453**
(-5.66)
R&D -0.963**
(-3.18)
-0.430
(-1.50)
-0.814**
(-2.44)
-0.944**
(-2.73)
2.281**
(3.48)
RD*sent 0.171†
(1.23)
0.417†**
(3.33)
0.367†**
(2.83)
-0.096†
(-0.66)
0.908†**
(4.35)
ADV -3.780**
(-16.63)
-1.643**
(-6.22)
-3.592**
(-14.22)
-2.846**
(-10.45)
-2.071**
(-3.37)
ADV*sent 0.701†**
(6.09)
0.341†**
(2.75)
0.002†
(0.99)
-0.279†**
(-1.91)
0.132†
(0.67)
SPE -0.285
(-0.84)
1.504**
(4.51)
-0.212
(-0.65)
-0.173
(-0.54)
0.924**
(2.35)
SPE*sent -0.433†**
(-3.52)
0.266†**
(2.59)
-0.247†**
(-2.66)
0.219†**
(2.05)
-0.189†*
(-1.51)
DISCON 1.830**
(2.55)
-1.622**
(-3.08)
3.136**
(3.11)
3.613**
(5.60)
3.398**
(3..22)
DISCON*sent 0.030†
(0.12)
-0.182†
(-0.65)
-0.171†
(-0.84)
-0.876†**
(-3.31)
0.358†
(1.05)
Adj. R squared 0.754 0.715 0.710 0.637 0.457
**significant at 5% level *significant at 10% level †One tail tests
R&Dt, CapExt, AdvExt, Specialt, discont, and extrat are from the annual financial statement and deflated
by shares outstanding. Those accounting items are winsorized at 1% each tail. All models include
untabulated control variables: macroeconomic factors (change in durable goods consumption, non-durable
goods consumption, service, CPI, industrial production, unemployment rate), firm factors (log market
value, book to market ratio), year, and month dummy variables. In addition, I also include interaction terms
of all earnings items with four control variables: future earnings growth ((EPS
t+3
-EPS
t
)/EPS
t
), absolute
value of EPS, firm size, and a linear time trend variable.
36
Chapter 4: Conclusion
Psychology theories suggest that, during high (low) sentiment periods, investors
are optimistic (pessimistic) but more risk-averse (risk-seeking). I use the inherent
information uncertainty of different earnings components to disentangle these two effects
of sentiment. Findings in this paper are consistent with the predictions that the effect of
sentiment on judgments is more prominent for “high uncertainty” cases and the effect of
sentiment on risk preferences is more prominent for “low uncertainty” cases. In the price-
earnings equation, price responses to earnings items bringing “high uncertainty” increase
when sentiment increases; conversely, price responses to earnings items bringing “low
uncertainty” decrease when sentiment increases. The effects of sentiment that I document
also are consistent with findings from studies that use the vector autoregressive technique
(VAR) to separate cash flow shocks from discount rate shocks (e.g., Vuolteenaho 2002).
Vuolteenaho (2002) shows that cash flow shocks usually predominate over discount rate
shocks and the correlation between these two shocks is positive. I find that, during high
(low) sentiment periods, a positive (negative) cash flow shock is usually discounted by a
higher (lower) required rate of return. Furthermore, the effect of sentiment on discount
rate stands out only when the effect on judgments is reduced by low information
uncertainty. Therefore, sentiment could be one of the exogenous shocks that generate the
result in the VAR studies.
The results of this study can be used to reconcile the mixed findings of prior
studies. For example, Baker and Wurgler (2007) find that the price of stable stocks is
lower in high sentiment periods. According to the current study, it is likely that the effect
of sentiment on judgments is swamped by the effect of sentiment on risk preferences for
37
stable stocks, and thus produces the negative sentiment beta. In another study, Mian and
Sankaraguruswamy (2008) find that, during low sentiment periods, investors have larger
responses to bad news compared with their response during high sentiment periods.
However, they rule out the possibility that investors’ risk preferences could have
explained their results because “during these (low sentiment) times, investors’ increased
risk aversion pushes up the discount rate”. Yet, this study and psychology studies both
provide evidence consistent with investors tending to be less risk averse when they are
pessimistic. Therefore, it is possible that investors are more risk-seeking during low
sentiment times and thus lower the discount rate.
This paper also enhances our understanding of sentiment effects. First, studies
based on characteristics of firms show only hard-to-value and difficult-to-arbitrage firms
are subject to the effects of sentiment. By introducing different earnings components to
the analyses, I am able to examine investors’ reactions to information directly and show
that the effects of sentiment exist across firms with different characteristics. Second, I
provide an alternative explanation as to why stock prices of many firms are subject to
little or no sentiment effect in some of the prior studies. “Limits to arbitrage” theory
argues that “irrational” responses are arbitraged away for these firms. However, I show
that it also is possible that the conflicting effects of sentiment on judgments and risk
preferences cancel each other out for these firms. Third, “limits to arbitrage” theory only
explains why volatile stocks are subject to sentiment effects but not why stable stocks
show sentiment effects in the opposite direction. The combination of shifts in judgments
and risk preferences provides an answer to firms on both ends: judgments overshadow
risk preferences under high uncertainty situations and vice-versa.
38
Future studies can expand this line of research by investigating the effect of
sentiment on different market participants such as analysts and managers. For example,
does market sentiment change managers’ decisions on disclosures? Do analysts take
sentiment into consideration when they make recommendations? Of note, both Fessler et
al. (2004) and Lerner and Ketler (2001) argue that valence-based (either positive or
negative) emotion is not precise enough to investigate this issue. For example, if bad
earnings news is attributed to management misconduct, investors may feel angry. If bad
earnings news is attributed to the economic downturn, investors may feel apprehensive.
Future research can expand this line of research by looking into specific emotions.
39
References:
Baker, M. and J. Wurgler. 2006. Investor sentiment and the cross-section of stock returns. Journal of
Finance. 64:1645-1680
Baker, M. and J. Wurgler. 2007. Investor Sentiment in the Stock Market. Journal of Economic Perspectives
21(2):129-151.
Beaver,W., P. Kettler and M. Scholes. 1970. The Association Between Market Determined and Accounting
Determined Risk Measures.” The Accounting Review 45:654–682
Ben-Ze’ev, A. 2001.The Subtlety of Emotions. MIT press, MA.
Brown, G.W. and M. T. Cliff. 2004. Investor sentiment and the near-term stock market. Journal of
Empirical Finance. 11:1-27
Brown, G.W. and M. T. Cliff. 2005. Investor sentiment and asset valuation. Journal of Business. 78:405-
440
Campbell, J. Y. and Ammer, J. 1993. What moves the stock and bond market? A variance decomposition
for long-term asset returns. Journal of Finance. 48:3-37
Cochrane, J.H. 2005. Asset Pricing. Princeton University Press, NJ.
Core, J. E., W. R. Guay, and A.V. Buskirk. 2003. Market valuations in the New Economy: an investigation
of what has changed. Journal of Accounting and Economics. 34:43-67
Curtin, R. 2000. Psychology and macroeconomics: fifty years of the survey of consumers. University of
Michigan
Downey, H. K. and J. W. Slocum. 1975. Uncertainty: Measures, Research, and Sources of Variation. The
Academy of Management Journal. 18:562-78.
Fama, E. 1970. Efficient capital markets: a review of theory and empirical work. Journal of Finance
52:383-417
Fessler, D.W.T., E.G. Pillsworth, and T.J. Flamson. 2004. Angry man and disgusted women: an
evolutionary approach to the influence of emotions on risk taking. Organizational Behavior and Human
Decision Process. 95:107-123
Fiske, S. and S. Taylor. 1991. Social Cognition. McGraw-Hill, NY.
Forgas, J.P. 1995. Mood and judgment: the affect infusion model (AIM). Psychological Bulletin.
117(1):39-66
Franzzini, A. 2006. The disposition effect and underreaction to news. Journal of Finance. 64:2017-2046
Glushkov, D. 2006. Sentiment Beta. University of Texas-Austin working paper.
Hirshleifer, D. 2001. Investor psychology and asset pricing. Journal of Finance. 56:(4)1533-1597
Hirshleifer, D. and T. Shumway. 2003. Good Day Sunshine: Stock Returns and the Weather. Journal of
Finance. 58(3):1009
Isen, A.M. 1984. Toward understanding the role of affect in cognition. In: Handbook of Social Cognition
Ed. by R.S. Wyer and T.K. Srull. Erlbaum, Hillsdale, NJ. pp.179–236.
Isen, A.M. 2004. Some perspectives on positive feelings and emotions: positive affect facilitates thinking
and problem solving. In Feeling and Emotions. Ed. by A.S.R. Mansted, N.H. Frijda, and A. Fischer.
Cambridge University Press. pp.263-281
Isen, A.M. and R. Patrick. 1983. The effect of positive feelings on risk taking: When the chips are down,
Organizational Behavior and Human Decision Processes 31:194–202
Isen, A. M, T.E. Nygren, and G.F. Ashby. 1988. Influence of Positive Affect on the Subjective Utility of
Gains and Losses: It Is Just Not Worth the Risk. Journal of Personality and Social Psychology. 55(5):710
40
Isen, A.M., T.E. Shalker, M. Clark, and L. Karp. 1978. Affect, accessibility of material in memory, and
behavior: a cognitive loop? Journal of Personality and Social Psychology. 36(1):1-12
Johnson, E. J. and Tversky, A. 1983. Affect, generalization, and the perception of risk. Journal of
Personality and Social Psychology. 45:20–31
Jouini, E. and Napp, C. 2007. Consensus Consumer and Intertemporal Asset Pricing with Heterogeneous
Beliefs. Review of Economic Studies 74(4):1149-1174
Kliger, D. and O. Levy. 2003. Mood-induced variation in risk preferences. Journal of Economic Behavior
and Organization. 52(4):573-584
Kothari, S.P., T. E. Laguerre, and A. J. Leone. 2002. Capitalization versus Expensing: evidence on the
uncertainty of future earnings from capital expenditures versus R&D outlays. Review of Accounting
Studies. 7:355-382
Kothari, S. P. and J. Shanken. 2003. Time-series coefficient variation in value-relevance regressions: a
discussion of Core, Guay, and Van Buskirk and new evidence. Journal of Accounting and Economics
34(1): 69-87
Kumar, S.M. and A. Persaud. 2003. Pure contagion and investors’ shifting risk appetite: analytical issues
and empirical evidence. International Finance. 5:401-426
Lee, Charles M.C., Andrei Shleifer, and Richard H. Thaler. 1991. Investor sentiment and the closed-end
fund puzzle. The Journal of Finance. 46:75-109
Lemmon, M. and E. Portniaguina. 2006. Consumer confidence and asset prices: some empirical evidence.
Review of Financial Studies. 19(4):1499-1596
Lerner, J.S. and D. Keltner. 2001. Fear, anger, and risk. Journal of Personality and Social Psychology.
81(1):146-159
Lipe, R. 1990. The relation between stock returns and accounting earnings given alternative information.
The Accounting Review. 65(1):49-71
Loewenstein, G.F., E.U. Weber, C.K. Hsee, and N. Welch. 2001. Risk as feelings. Psychological Bulletin.
127(2):267-286
Mano, H. 1992. Judgments under distress: Assessing the role of unpleasantness and arousal in judgment
formation. Organizational Behavior and Human Decision Processes. 52:216-246
Mian, M.G. and S. Sankaraguruswamy. 2008. Investor Sentiment and Stock Market Response to Corporate
News. National Singapore University working paper.
Mittal, V. and W. Ross. 1998. The Impact of Positive and Negative Affect and Issue Framing on Issue
Interpretation and Risk Taking. Organizational Behavior and Human Decision Processes. 76(3):298-364
Mossin. 1969. Security pricing and investment criteria in competitive markets. The American Economic
Review. 59(5):749-756
Nygren, T. E. and A. M. Isen. 1985. Examining Probability Estimation: Evidence for Dual Subjective
Probability Functions. Paper presented at the meeting of the Psychonomics Society, Boston.
Nygren, T. E., A.M. Isen, P.J. Taylor, and J. Dulin. 1996. The influence of positive affect on the decision
rule in risk situations: Focus on outcome (and especially avoidance of loss) rather than probability.
Organizational Behavior and Human Decision Processes. 66(1):59-72
Qiu, L. and I. Welch. 2006. Investor sentiment measures. NBER working paper no. 10794
Penman, S.H. and N. Yehuda. 2004. The Pricing of Earnings and Cash Flows and an Affirmation of
Accrual Accounting. Columbia University working paper.
Petersen, M.A. 2008. Estimating standard errors in finance panel data sets: comparing approaches.
Forthcoming Review of Financial Studies.
41
Rubinstein, M. 1976. The valuation of uncertain income streams and the pricing of options. The Bell
Journal of Economics 7(2):407-725
Saunders, E. M., Jr. 1993. Stock prices and Wall Street weather. American Economic Review 83:1337-1345
Shleifer, A. 1986. Do demand curves for stocks slope down? Journal of Finance 41:579–90.
Vuolteenaho, Tuomo. 2002. What drives firm-level stock returns? Journal of Finance. 57(1):233-264
Wright, W.F. and G.H. Bower. 1992. Mood effects on subjective probability assessment. Organizational
Behavior and Human Decision Process. 52(2):276-291
Wurgler, J. and Zhuravskaya E. 2002. Does arbitrage flatten demand curves for stocks? Journal of
Business. 75(4):583-608
Zhang, F. 2006a. Information Uncertainty and Stock Returns. Journal of Finance 61(1): 105-137
Zhang, F. 2006b. Information Uncertainty and Analyst Forecast Behavior. Contemporary Accounting
Research 23(2):565-590
42
Appendix: Hypotheses development based on log transformation of the discount
cash flow model
The effect of sentiment on stock prices can be illustrated by a simple discounted
cash flow (DCF) valuation model:
∞
Value
t
= Σ (Expected cash flow
n
/ (1 + subjective discount rate)
n-t
)…(1)
n= t+1
Sentiment changes both investors’ judgments on future cash flow (the numerator)
and their risk preferences as presented in the discount factor (the denominator). To
disentangle those two effects, I take a log transformation of the DCF model
18
:
P = α + β × E + γ × D + ε……………………………………………..…….(2)
Where P is the log of stock price, E is the log of expected value of future cash
flow, and D is the log of risk adjusted discount factor. By definition, β>0 and γ<0.
Some studies have adopted a similar log-linear transformation that separate those
two effects. Vuolteenaho (2002) investigates the magnitude of cash-flow shocks (E) and
expected-return shocks (D) and the correlation between the two. He finds that cash-flow
shock is usually more than two-times larger than expected-return shock and drives stock
price most of the time. He also finds that the correlation between those two streams of
shocks is 0.41. That means a positive cash-flow shock usually comes with a positive
expected return shock. In other words, when people think the cash flow is larger, they
also demand a higher rate of return. Although those two effects work against each other,
cash flow shock is usually stronger and prevails.
18
Both Mossin (1969) and Fama (1970) show how different types of multi-period valuation problems can
be collapsed into a properly specified single period problem with risk adjusted discount rate. Also,
Campbell and Ammer (1993) show similar log-linear approximation is feasible with an adjustment factor. I
did not include the adjustment factor here because this model is meant for illustration of ideas and I am
more interested in the signs of the coefficients rather than the magnitude.
43
His result is consistent with the findings from the psychology theory. Investors’
sentiment can be the exogenous force that changes both expected cash-flow and required
returns. The fact that cash flow shock dominates required return shock may also explain
why most of the time we see the effect of sentiment on judgment.
Next, based on psychology theories, I construct expected value of future cash
flow (E) and risk adjusted discount factor (D) as functions of sentiment.
Expected cash flow (E) is a function of both sentiment and accounting
information. Prior studies show that investors are optimistic when sentiment level is high.
Accounting literature also document that different accounting information has different
implications on future cash flow. In addition to those two main effects, I argue that one of
the characteristics (i.e., uncertainty) of the accounting information moderates the effect of
sentiment. A valuation task based on accounting numbers bring future benefits that are
highly uncertain demands more effort on judgment. In other words, uncertainty
establishes the room for judgments. In the opposite extreme cases, items generate no
uncertainties need no judgment. In other words, high uncertainty generated by accounting
information will enlarge the effect of sentiment. Low uncertainty generated by
accounting information will mute the effect of sentiment. Expressed in a simple function:
E = a
0
+ a
1
×S + a
2
× ACCT(V) + a
3
× S × ACCT(V)………………………(3)
Where S is the investor sentiment and ACCT(V) is the accounting information
with uncertainties (V) as one of its characteristics. Sentiment effect is large for
accounting information generating highly uncertain future benefit; therefore, a
3
>0 for
ADV and R&D expenses. When the uncertainties generate by the accounting information
44
is very little, sentiment will not change the influence of accounting information on
expected cash flow. Consequently, a
3
=0 for one-time items.
Following the CAPM framework, I define subjective risk adjusted discount factor
as one plus risk free rate plus the product of amount of risk and the price of risk:
D = 1 + R
f
+ Pr × V………………………………………………………….(4)
Where R
f
is the risk free rate. V is the amount of risk; in this case it is the
uncertainty of the future benefit generated by accounting information. Pr is the price of
risk or the risk premium. Based on the psychology literature, the price of risk is a
function of risk preference or degree of risk aversion which changes with sentiment. I
simply assume Pr= b
0
× S. When sentiment level is high, investors demand high premium
for the same amount of risk and therefore the price of risk will be higher. The theory
indicates that b
0
>0.
Apply equations (3) and (4) back to equation (2) in the empirical setting:
P=δ
0
+ δ
1
× (a
0
+ a
1
×S + a
2
×ACCT(V) + a
3
×S×ACCT(V)) + δ
2
× (1+ R
f
+ b
0
× S × V )
Where δ
1
represents β and δ
2
represents γ. Simply rearrange the terms:
P= δ
0
+δ
1
×a
0
+ δ
2
× (1+ R
f
)
+(a
1
× δ
1
) ×S
+ (δ
1
×a
2
) ×ACCT(V)
+ (a
3
×δ
1
+ δ
2
×b
0
×ø) ×S×ACCT(V)
45
The coefficient on the interaction of sentiment and accounting information
contains two terms: a
3
×δ
1
and δ
2
×b
0
×ø. The first term represents the effect of sentiment
on judgments and the second term represents the effect of sentiment on risk preferences.
For one time items, a
3
=0 (or very close to zero), and the effect of sentiment on judgment
is zero (or minimum). Therefore, the coefficient is just δ
2
×b
0
×ø. Because δ
2
(which
corresponds to γ) is negative and both b
0
and ø are positive
19
, the sign of the coefficient
should be negative. For ADV or R&D expenses, a
3
×δ
1
(a positive number) is added to the
coefficients. Therefore, the coefficients for ADV and R&D expense should be larger than
the coefficients for one time items.
19
ø is just a transformation of ACCT(V) into V. In general, the larger the number the larger the standard
deviation of that number will be. ACCT′(V) should be positive; therefore, ø is positive.
Abstract (if available)
Abstract
This study reconciles inconsistent evidence on the sentiment-price relation in prior studies by explicitly considering the effects of sentiment on both investor judgments and risk preferences. Using the uncertainty in accounting information, I am able to disentangle these two effects of sentiment and investigate the causes of the variations in the sentiment-price relation. The results show that, under low uncertainty, the effect of sentiment on risk preferences dominates in the sentiment-price relation, such that a negative effect of sentiment on price is observed. In contrast, under high uncertainty, the effect is less negative and, in fact, becomes positive. This suggests that in cases of high information uncertainty, the effect of sentiment on judgments dominates in the sentiment-price relation.
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
“What you see is all there is”: The effects of media co‐coverage on investors’ peer selection
PDF
Leveling the playing field: unbiased tests of the relative information content of book income and taxable income
PDF
Does fair value accounting exacerbate the pro-cyclicality of bank lending?
PDF
Does mandatory adoption of international financial reporting standards decrease the voting premium for dual-class shares: theory and evidence
PDF
The effects of accounting performance and professional relationships on promotion, dismissal, and transfer decisions in a conglomerate
PDF
The role of social media in influencing consumer opinions and preferences in the fintech industry
PDF
Essays on the empirics of risk and time preferences in Indonesia
PDF
Shortcomings of the genetic risk score in the analysis of disease-related quantitative traits
PDF
Essays in asset pricing
PDF
Choice biases in making decisions for oneself vs. others
PDF
CEO reputation: who benefits -- the firm and the CEO?
PDF
The interactive effects of incentive threshold and narcissism on managerial decision-making
PDF
Who manages the firm matters: the incremental effect of individual managers on accounting quality
PDF
The risk estimates of pneumoconiosis and its relevant complications: a systematic review and meta-analysis
PDF
Leadership, accountability, and decision-making: An examination of first-responder experience and self-efficacy in real-time decision-making to improve community outcomes
PDF
The acute relationship between affective states and physiological stress response, and the moderating role of moderate-to-vigorous physical activity
PDF
The role of depression symptoms on social information processing and tobacco use among adolescents
PDF
Essays in panel data analysis
PDF
Mindfulness and resilience: an investigation of the role of mindfulness in post-9/11 military veterans' mental health-related outcomes
PDF
The role of Rho-GEF signaling in synapse function and autism-related disorders
Asset Metadata
Creator
Chen, Kun-chih
(author)
Core Title
The role of accounting information in the sentiment-price relation
School
Leventhal School of Accounting
Degree
Doctor of Philosophy
Degree Program
Business Administration
Degree Conferral Date
2009-05
Publication Date
05/09/2009
Defense Date
04/10/2009
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
accounting information,investor sentiment,judgment and decision making,OAI-PMH Harvest,risk preference
Language
English
Contributor
Electronically uploaded by the author
(provenance)
Advisor
Bonner, Sarah (
committee chair
), Wang, Shiing-wu (
committee chair
), Brown, Nerissa (
committee member
), Hsiao, Cheng (
committee member
), Matsusaka, John (
committee member
)
Creator Email
kunchih@gmail.com,kunchihc@usc.edu
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-m2225
Unique identifier
UC1276740
Identifier
etd-Chen-2787 (filename),usctheses-m40 (legacy collection record id),usctheses-c127-236518 (legacy record id),usctheses-m2225 (legacy record id)
Legacy Identifier
etd-Chen-2787.pdf
Dmrecord
236518
Document Type
Dissertation
Rights
Chen, Kunchih
Type
texts
Source
University of Southern California
(contributing entity),
University of Southern California Dissertations and Theses
(collection)
Repository Name
Libraries, University of Southern California
Repository Location
Los Angeles, California
Repository Email
cisadmin@lib.usc.edu
Tags
accounting information
investor sentiment
judgment and decision making
risk preference