The Bright Side of Managerial Over-optimism
Gilles Hilary
[email protected]
INSEAD
Charles Hsu
[email protected]
HKUST
Benjamin Segal
[email protected]
INSEAD
Rencheng Wang
[email protected]
University of Queensland
Abstract: Human inference and estimation is subject to systematic biases. In particular, there is
a long literature showing that overconfidence can lead to sub-optimal decisions. We depart from
this research by showing empirically that a) over-optimism is a related but different bias, b)
importantly, it can improve firm’s welfare, and more specifically, firms’ profitability and market
value, and c) it can emerge dynamically in a rational economic framework.
JEL classification: G39
Corresponding author: Gilles Hilary, INSEAD, 1 Ayer Rajah Avenue, 138676 Singapore, Tel: (+65) 6799 5100,
[email protected]. We thank Denis Gromb, Kai Wai Hui, Tiphaine Jerome, Kirill Novoselov, Guochang
Zhang, and workshop participants at Hong Kong University of Science and Technology, INSEAD, National
University of Singapore, and Xiamen University for their helpful comments. Hsu acknowledges financial support
from the Hong Kong Research Grants Council (FSGRF13BM02).
Electronic copy available at: http://ssrn.com/abstract=2286226
The Bright Side of Managerial Over-optimism
Abstract: Human inference and estimation is subject to systematic biases. In particular, there is
a long literature showing that overconfidence due to cognitive biases can lead to sub-optimal
decisions. We depart from this research by showing empirically that a) over-optimism is a
related but different bias, b) importantly, it can improve firm’s welfare, and more specifically,
firms’ profitability and market value, and c) it can emerge dynamically in a rational economic
framework.
Key words: over-optimism, firm performance.
Electronic copy available at: http://ssrn.com/abstract=2286226
1. Introduction
One of the best-established stylized facts in the decision-making literature is that
individuals are over-optimistic about future outcomes (e.g., Weinstein, 1980). Over-optimism is
related to overconfidence but is distinct. To use the terminology of Hackbarth (2008), optimistic
managers overestimate the growth rate of earnings while overconfident managers underestimate
the riskiness of earnings. 1 In other words, over-optimism creates an upward bias in the mean of
the distribution while overconfidence creates an upward bias in its precision. Importantly, prior
analytical work also suggests the possibility that over-optimism can increase firm performance
(e.g., Van den Steen, 2004), but little empirical work exists on this topic. We investigate this
issue empirically.
We first note that individuals are over-optimistic in general, and particularly so regarding
the effect of their own actions (we call this phenomenon “static over-optimism”). In addition,
managers may suffer from a biased attribution of causality after a series of good performance
that leads them to under-estimate the role of random noise and over-attribute successes to their
own actions (we call this phenomenon “dynamic overconfidence”). The combination of these
two phenomena leads to an increase in over-optimism after a series of successes (we call this
phenomenon “dynamic over-optimism”): optimistically-biased managerial actions receive an
increasingly disproportionate weight in the overall estimation of the project success. Thus, our
first hypothesis posits that the degree of managerial over-optimism should increase after a series
of successes. Perhaps paradoxically, this miscalibration in managerial prediction may encourage
managers to exert a greater effort. Our second hypothesis conjectures that firm performance
1
Heaton (2002) defines managers as “optimistic” when they systematically overestimate the probability of good
firm performance and underestimate that of bad performance.
2
Electronic copy available at: http://ssrn.com/abstract=2286226
should increase following a series of successes, even if the increase remains below the level
expected by an over-optimistic manager.
We note that our hypotheses can be motivated in an economic framework in which
individuals are Bayesian and fully rational. For example, Van den Steen (2004) proposes a
model in which rational agents are endowed with different priors. In the most basic setting, a
manager can choose from a menu of actions. For each action, the manager's prior over the
likelihood of success equals the true probability plus a random error. The manager optimally
chooses the action for which her prior is highest, which is also the action whose probability of
success the manager is most likely to over-estimate. This generates the “static” over-optimism
we discuss above.
This framework also allows for a rational (i.e., Bayesian) but biased
attribution of performance that yields a form of dynamic overconfidence.
Combined with
“static” over-optimism, this phenomenon generates the “dynamic over-optimism” that is the core
of our first hypothesis. Our second hypothesis relies on the idea that over-optimism can elicit
higher effort from management. This emerges naturally if effort and probability of success are
complements (e.g., the payoff is increasing in effort when the action is successful).
To set the stage, we first explore the empirical relation between firm performance and
two measures of over-optimism: one based on options held by the CEO (Campbell et al., 2011)
and one based on the degree of optimism in firms’ press releases (based on textual analysis).
Both measures are shown to be positively related to performance in cross-sectional pooled
regressions. Although our preliminary cross-sectional pooled results are consistent with our
hypothesized model, this approach is not well suited for understanding the dynamics emerging
from our framework. To verify that our interpretation of the preliminary results is correct, we
empirically investigate whether past performance leads to managerial over-optimism and
3
whether this optimism encourages managers to exert greater effort. Since our predictions are
largely about dynamic over-optimism rather than static biases, our main tests focus on timeseries properties for a given manager rather than on cross-sectional ones. 2
Our main empirical findings are consistent with our analytical framework and support the
presence of dynamic over-optimism. First, we show that managers who experienced more
frequent successes in the recent past subsequently issue more optimistic forecasts and exhibit
more optimistic behaviors proxied by late option exercises and by the tone in earnings releases.
Importantly, these results hold in specifications with manager fixed effects. By showing that
over-optimism has an endogenous and dynamic component, our results suggest that, at least to
some extent, managers are made (not born) over-optimistic. Our results are robust to a number
of alternative definitions of past successes. Second, and importantly, we show that the overoptimism we document can increase firm performance and that managers appear to exert greater
effort to meet their own over-optimistic forecasts. Specifically, firm accounting performance
and quarterly market return increase as the degree of managerial optimism increases. In contrast,
measures of accruals or real earnings management are not affected by this optimism. These
results suggest that the increase in firm performance is genuine, and that managers, being overoptimistic regarding the likelihood of meeting the expectations they set, may not feel the need to
manage earnings to reach their forecasts. Third, using comparative statics, we show that our
empirical findings are consistent with our economic framework, which predicts that overoptimism increases with the diffusion of priors and decreases with managerial experience. Van
den Steen (2004) notes these comparative statics empirically distinguish this framework from
2
The fact that static optimism and dynamic overconfidence are intertwined provides the foundation for our analysis.
Section IV of Van den Steen (2004) elaborates on the theoretical links between these two biases in our framework.
The links are empirically important in our setting because, as we note in Section 2.2, "static" over-optimism is
difficult to test in a cross-sectional setting. The dynamic form of over-optimism allows us to use fixed effect
specifications.
4
others. Their added importance is minimizing the risk that our conclusions are driven by an
omitted variable as this unspecified variable not only would have to covary with our variables of
interest but also would have to conditionally covary in a way that is consistent with our
framework.
We contribute to the literature in several ways. First, we show that over-optimism can
increase firm performance.
Although prior analytical work suggests this possibility (e.g.,
Benabou, and Tirole, 2002; Compte and Postlewaite, 2004; Van den Steen, 2004; Hackbarth,
2008; Gervais, Heaton, and Odean, 2011), little empirical work exists on this topic. Second, we
present evidence consistent with the comparative statics and ancillary predictions of the
economic over-optimism framework.
In fact, this framework provides a well-integrated
motivation for our empirical tests and our study is able to shed light on most of the testable
implications that have been identified by the prior literature. Lastly, we empirically distinguish
between overconfidence and over-optimism. The importance of this distinction has been noted
by the prior literature. For example, Moore and Healy (2008) explain (p. 503) that the “first
problem with overconfidence research” is that the most popular research paradigm confounds the
overestimation of one’s actual ability, performance, level of control, or chance of success on the
one hand and excessive certainty regarding the accuracy of one’s belief on the other hand. There
is now an established literature on overconfidence in managerial behavior (see Glaser and Weber
2010 for a review) but much less is known about over-optimism in this context. 3
3
A small number of recent studies investigate the possibility that overconfidence is a dynamic phenomenon (e.g.,
Hilary and Menzly, 2006; Hilary and Hsu, 2011). Our study differs from those studies in three important ways.
First, we consider how the optimism bias affects the manager’s effort. We believe ours is one of the first empirical
studies showing that biased (optimistic) expectations lead to an improvement in firm performance. Second, we
show that past successes affect the level of the bias, while those studies focus on the precision of the forecast. In
other words, we distinguish between over-optimism and overconfidence. Third, those studies rely on research in
psychology to motivate their empirical analysis. We empirically explore a rational framework that does not rely on
5
The remainder of the paper proceeds as follows. In Section 2, we discuss our hypothesis
development and our empirical design.
In Section 3, we present our samples, descriptive
statistics and preliminary results. In Section 4, we provide our main empirical results and relate
them to Van den Steen (2004). In Section 5, we discuss different robustness checks. Section 6
concludes.
2. Hypothesis development and empirical design
2.1. Hypothesis development
Economics has traditionally assumed that people begin the decision-making process with
subjective beliefs over the different possible states of the world and use a Bayesian approach to
update these beliefs as they receive additional information. Under such a framework, individuals
hold unbiased beliefs that typically converge toward the true parameter as more signals are
received. However, this view has been challenged as numerous systematic biases have been
identified. In particular, individuals have been shown to be over-optimistic and overconfident.
These two biases are similar but distinct. Over-optimism involves an excessive belief that future
events will be positive, while overconfidence involves placing too much weight on the accuracy
of private information and an excessive belief in one’s own skills. For example, Weinstein
(1980) notes in his seminal study on over-optimism that people believe negative events are less
likely to happen to them than to others whereas positive events are more likely to happen to them
than to others.
More recently, the literature has started to integrate the notion of static over-optimism in
an economic framework. For example, Van den Steen (2004) proposes a model in which
cognitive biases to motivate our hypotheses. To our knowledge, we provide the first archival study that is consistent
with a rational explanation of over-optimism.
6
rational (Bayesian) individuals are endowed with a menu of actions that will lead to either a
successful outcome or an unsuccessful outcome. The manager who selects actions has a prior
with respect to the likelihood of whether a given action will be successful. If, for exogenous
reasons, managers sometimes overestimate and sometimes underestimate the probability of an
action’s success (relative to other individuals such as analysts or relative to the true underlying
probability) and select the action with the highest perceived probability of success, then they are
more likely to select actions that overstate the probability of success the most. They will
therefore be over-optimistic about the success of the actions they take. This mechanism is
reminiscent of the winner’s curse (i.e., the winner of auctions in which bidders have incomplete
information is the bidder who has the most optimistic estimate of the asset’s value). Under this
framework, the random variation in priors together with systematic choice leads to systematic
bias. 4 Moore and Healy (2008) provide an alternative model in which Bayesian individual can
rationally overestimate their chance of success for difficult tasks while Benoit and Dubra (2011)
show how a majority of Bayesian individuals may rationally believe they have greater skills than
the majority of the population.
More than static over-optimism, our focus is on dynamic over-optimism.
A recent
empirical literature investigates the possibility that overconfidence (and managerial
overconfidence in particular) is a dynamic phenomenon (e.g., Hilary and Hsu, 2011). Research
on dynamic over-optimism is more limited although the dynamic aspect can be incorporated in
an economic framework. For example, Van den Steen’s (2004) model leads to biased attribution
of causality. In particular, the manager will rationally underestimate the role of exogenous
4
In our setting, these actions are related to the management of the firm (e.g., increasing the level of inventory,
starting an advertising campaign, and so forth) and we predict that managers subject to this over-optimism bias will
expect higher earnings than justified. These actions are context specific and their precise nature is not the subject of
our investigation.
7
factors in the case of success and overestimate their role in the case of failure. In the case of
success, this biased attribution will lead the manager to gradually put more weight on the
expected effect of her own actions (for which the prior distribution is optimistically biased) and
less weight on the effect of random noise (for which the prior distribution may not be similarly
biased). This change in weights will lead to a gradual increase in overall over-optimism even
though the perceived probability of success for each individual action does not change over time.
Rabin and Schrag (1999) and Gervais and Odean (2001) provide alternative theoretical economic
rationalizations for the existence of dynamic overconfidence. Based on this discussion, our first
hypothesis is that:
H1. Individuals become dynamically over-optimistic following a series of successes.
We then turn our attention to the effect of over-optimism on the firm’s welfare.
Economic theory suggests that managers should work harder if effort complements the
likelihood of success (e.g., Van den Steen, 2004). For example, suppose the strictly positive
payoff is increasing in effort in the case of success and is zero in the case of failure. Further
assume that the cost of effort is increasing and strictly convex.
The amount of effort is
increasing in the perceived probability of success: the greater the expected probability, the
greater the effort. To the extent that this effort increases firm performance, we expect an
increase in firm performance after a series of successes, even if the increase is not sufficient to
reach the level predicted in the forecast. Based on this discussion, our second hypothesis is that:
H2. The firm’s performance should gradually improve as managers become more optimistic.
8
2.2. Empirical design
To test these predictions, we focus on managerial forecasts of firm quarterly earnings. In
doing so we obtain a measure of managerial expectations about future firm performance that we
can benchmark against both subsequent realizations and common expectations at the time of
issuance (as proxied by analyst forecasts).
We decompose the public forecast F into two components, namely, the best estimate of
the realization privately observed by the manager and a strategic negative bias that allows the
manager to meet or beat her own forecast more easily.
F(X)i,t = E(X)i,t + bi,
(1)
where X is the realized earnings for firm-manager pair i in period t, F is the manager’s public
forecast of X, E(X) is the manager’s private expectation of X, and b is the manager’s strategic
bias. Following Hilary, Hsu, and Wang (2014), we assume that the strategic bias is constant (and
typically negative) for each firm-forecaster pair. We further assume that the manager’s private
expectation is a combination of personal and public expectations regarding the results of the
action, a, chosen by the manager. The combination of the two takes the form
E(X)i,t = αi,t Em(a)i + (1-αi,t) Ep(a)i,
(2)
where Em(a) represents the manager’s personal expectation (formed independently of the public
signal), Ep(a) represents the public expectation, and α represents the weight that the manager
places on her personal expectation. We next assume that the public expectation of the action
taken by the manager is well calibrated (i.e., Ep(a) is equal to the average realization of a, â). As
discussed before, we expect the manager to be over-optimistic with respect to the effect of her
actions (i.e., the manager’s personal expectation of a, Em(a), is greater than â). For convenience,
9
we also initially assume that this degree of static over-optimism, βi (i.e., Em(a)i – âi), remains
constant over time for a given firm-manager (we revisit this issue in Section 4.1). The overall
bias in the management forecast released by the manager, Bi,t, is equal to αi,t βi + bi. It is thus
impossible to know the average sign of the overall bias without knowing the precise value of the
three parameters. This makes cross-sectional predictions difficult. However, as previously
discussed, we expect that positive outcomes will make the manager more overconfident with
respect to the effect of her actions. In other words, α is increasing with past successes and the
overall bias is relatively more positive (or less negative) after a series of successes. This
expectation yields time-series predictions that we can test. 5
More specifically, to test our first hypothesis that managers become dynamically overoptimistic after a series of successes, we estimate the following model:
Optimi,t = αi + β1 MBSTRi,t + γk Xki,t + εi,t,
where Optim is a vector of optimism measures.
(3)
We use two direct measures of forecast
optimism. MFE (signed management forecast error) is defined as the management forecast
minus the earnings realization divided by the stock price two days before the issuance of the
forecast.
MFD (signed management forecast deviation from consensus) is defined as the
management forecast minus the pre-existing consensus analyst forecast divided by the stock
price two days before the issuance of the management forecast. We define the consensus
forecast as the median analyst forecast over the 90 days prior to the issuance of the management
forecast. In addition, we consider two additional proxies that are directly correlated with the
degree of optimism. Optim_Option is a measure based on the option portfolio of the manager.
5
The difference between two forecasts made by the same manager in two periods is equal to
F(X)2 - F(X)1 = [α2 Em(a) + (1-α2) Ep(a) + b] - [α1 Em(a) + (1-α1) Ep(a) + b] = (α2 - α1) (Em(a) - Ep(a)) = (α2 - α1) β. If
β is positive, the manager becomes more over-optimistic as she becomes more overconfident (i.e., as α increases).
10
We define a CEO as over optimistic if she postpones the exercise of vested options that are at
least 67% in the money. If a CEO is classified as over optimistic in a given year, the CEO
retains this type going forward unless she exhibits the opposite behavior according to the
measure. That is, we allow for the possibility that an over optimistic CEO becomes unbiased if
she exhibits the appropriate option exercise behavior (i.e., exercises vested options that are less
than 67% in the money) at least twice in the sample period after she has been defined as over
optimistic (Campbell et al. 2011). Finally, we consider Optim_PR, a measure of optimism based
on the textual analysis of firms’ press releases (the appendix details the construction of this
variable).
MBSTR, the streak of recent successes, is our key independent variable. It counts the
number of consecutive successes that CEO i enjoyed in the four quarters before the current
forecast was announced. For example, if the streak of recent successes is 1, 0, 0, and 1 (counting
backward from quarters t-1 to t-4), then MBSTR will be equal to one. If the streak is 1, 1, 0, and
1 (counting backward from quarters t-1 to t-4), then MBSTR will be equal to 2. 6 Initially, we
define a success as an earnings realization that is above the manager's expectation (i.e., the midpoint of a range forecast or the forecast for a point estimate), but the sensitivity analysis reported
in Section 5.1 shows that our results are not affected when we use multiple alternative proxies to
define success. Our key prediction is that β1 should be positive in both Model (3).
Xk represents a vector of k control variables suggested by prior literature. Size is the log
of market value at the beginning of the quarter (Baginski and Hassell, 1997). We control for
growth opportunities by including Book-to-Market, which is the book value of equity divided by
6
We consider the last four quarters as prior literature suggests that this is the optimal period length for measuring
short dynamics in biases (Hilary and Menzly, 2006; Hilary and Hsu, 2011).
11
the market value of equity at the beginning of the quarter (Bamber and Cheon, 1998). Leverage
is the ratio of total liabilities to total assets at the beginning of the quarter (Fombrun and Shanley,
1990; Malmendier, Tate, and Yan, 2011). StdEarn is the standard deviation of the quarterly
return on assets over at least six of the preceding eight quarters (Armor and Taylor, 2002).
RetVol is the standard deviation of the stock return six months before the management forecast
date (Armor and Taylor, 2002). Cover is the log of the number of analysts covering the firm in a
given quarter (Hilary and Hsu, 2011). Inst Ownership represents the percentage of shares owned
by institutional investors (Ajinkya, Bhojraj, and Sengupta, 2005). Hor is the forecast horizon,
measured as the log of the number of days between the management forecast date and the end of
the fiscal period (Johnson, Kasznik, and Nelson, 2001). Loss is an indicator variable that takes a
value of one if earnings are negative, and zero otherwise (Rogers and Stocken, 2005). All
continuous variables are winsorized at the 1% level.
To test our second hypothesis that firm performance increases after a series of successes
because managers become more optimistic, we estimate the following models:
ROAi,t = αi + β1 MBSTRi,t + γk Xki,t + εi,t,
(4a)
ROAi,t = αi + β1 Optimi,t + γk Xki,t + εi,t.
(4b)
where ROA is defined as operating income divided by total assets at the beginning of the quarter,
where operating income is earnings excluding special or non-recurring items such as gain or loss
from asset sales, asset write-downs or write-offs. MBSTR, Optim, and Xk are defined previously
after Model (3). We predict that β1 should be positive in both Models (4a) and (4b).
12
2.3. Estimation procedure
We employ a panel (fixed-effect) technique to estimate the equations (i.e., we use CEOspecific intercepts αi). The manager fixed-effect regression is particularly suitable for testing our
hypotheses, which focus on the short-term dynamics of managers’ forecasts, as it eliminates
cross-sectional variation in the means of the variables while leaving the time-series dynamics of
these forecasts intact. This approach allows us to treat over-optimism as a time-varying variable
whereas most of the prior empirical literature has treated optimism as a fixed or at least slow
moving characteristic. The use of manager fixed effects also provides a natural control for
omitted variables. For example, constant differences in managerial skill levels, prediction bias,
or constant firm characteristics such as industry classification are all controlled for.
Our
approach provides a direct estimate of the change in optimism and reduces the need for these
multiple robustness checks. In our main specifications, we employ CEO fixed effects. This
approach implicitly assumes that CEOs play a significant role in the issuance of forecasts and
that forecasts are important to them, an assumption that is consistent with prior literature. For
example, Lee, Matsunaga, and Park (2012) report the probability of CEO turnover to be
significantly higher when the magnitude of management forecast error is greater. Untabulated Ftests also indicate that CEO fixed effects are jointly significant in our ordinary least squares
specifications (the p-value is less than 0.01 in all cases).
We use an Ordinary Least Squares (OLS) approach when the dependent variable is
continuous and a logit specification when the dependent variable is binary. The standard errors
are calculated according to the procedure outlined in Cameron, Gelbach, and Miller (2011) and
are groupwise heteroskedasticity-consistent (i.e., adjusted simultaneously for heteroskedasticity
and the clustering of observations by CEO and year).
13
3. Samples, descriptive statistics and preliminary results
3.1. Samples
Accounting data come from Compustat’s quarterly data files, and stock price and return
data come from the daily files of the Center for Research in Security Prices (CRSP). We obtain
CEO information from the ExecuComp database. We drop firms for which we are unable to
obtain information such as the name of the CEO or the dates of her tenure. We obtain our first
(preliminary) sample of 80,122 observations. 7 We collect Optim_Option and Optim_PR based
on this sample. We use this broad sample in our preliminary analysis.
We also use a second, more focused, sample (i.e., management forecast sample) to test
our main hypothesis more directly. We obtain management forecast data from the FirstCall
database. To increase data consistency, we focus on quarterly predictions and exclude forecasts
made after the end of the fiscal period (i.e., pre-announcements). We include only the last
forecast made by a given manager before the end of the fiscal period because earlier predictions
may be drawn from a different distribution of forecasts. Our sample only includes point and
range forecasts, that is, we exclude qualitative forecasts that do not provide a numerical value of
earnings per share. We also require that managers issue at least five forecasts (four in the
previous eight quarters plus one in the current quarter) and that the same CEO manage the firm
at the time that these forecasts were made. Chuk, Matsumoto, and Miller (2013) document the
presence of several problems with the First Call CIG database, but these are mitigated by both
the time series and cross-sectional characteristics of our sample. First, Chuk, Matsumoto, and
Miller (2013) indicate that the problems in the First Call database are largely concentrated in the
7
Our first sample covers the 1998-2010 period, consistent with the period for our management forecast sample
described next.
14
pre-1998 period. To address this issue, we start our sample period in 1998. We stop our
sampling period in 2010 when FirstCall stops coverage.
Second, we require at least five
management forecasts for each CEO – it is unlikely that CIG omits a given CEO who issues a
series of forecasts. Our sample period thus covers the years from 1998 to 2010. These sampling
criteria yield 10,630 management forecasts. We match the forecast data with the corresponding
records of FirstCall reported earnings.
3.2. Preliminary descriptive statistics
We report descriptive statistics based on the Execucomp sample in Table 1. The mean
and median values of (quarterly) ROA are 0.034 and 0.032, respectively. The mean and median
values for the different variables are reasonably close to each other, suggesting that skewness is
not a major issue in our setting.
Table 2 presents a correlation matrix.
Optim_PR and
Optim_Option, our two measures of optimism independent of forecasts, are positively correlated
with each other. More importantly for our purpose, these two measures are positively and
significantly correlated with ROA. This result is consistent with our hypothesis that optimism
can enhance firm’s welfare.
The univariate correlations among the control variables are
relatively small, suggesting that multi-collinearity is not an issue in our setting.
3.3. Preliminary pooled cross-sectional regressions
Before testing our main hypotheses, we perform a preliminary analysis to establish an
empirical link between optimism and performance. We use the Compustat Execucomp sample to
estimate pooled cross-sectional specifications (with industry- and fiscal-quarter- but no CEO-
15
fixed effects) in which we regress ROA on either Optim_Option or Optim_PR (controlling for
Xk). 8
The preliminary empirical results are presented in Table 3. They indicate that both
Optim_PR and Optim_Option are positively associated with ROA. The relation is statistically
robust with z-statistics of 14.35, and 3.29, respectively.
The economic effect is such that
increasing Optim_Option from 0 to 1, or one standard deviation increase of Optim_PR improves
the median value of ROA by 26.31 percent and 2.63 percent, respectively. 9 This finding
provides preliminary but indirect support for our hypothesis that optimism is positively
associated with firm welfare. Turning attention to the control variables, we find that the signs
are consistent with those in the prior literature.
4. Main empirical results
We then turn to the more direct investigation of our hypotheses. To this end, we focus on
the second sample.
4.1. Main univariate correlations
We report univariate correlations based on the second sample in Table 4. The correlation
between our four measures of optimism is generally significantly positive but reasonably low.
Our first hypothesis predicts a positive correlation between recent past successes and our
measures of optimism.
We find that the correlation between MBSTR and either MFD,
8
Industries are defined as in Fama and French (1997).
9
0.842 (coefficient on Optim_Option in Table 3) ×1/0.032 (median ROA in Table 1) = 26.31% for Optim_Option
and 0.029 (coefficient on Optim_PR in Table 3) ×2.903 (standard deviation of Optim_PR in Table 1) /0.032 (median
ROA in Table 1) = 2.63% for Optim_PR.
16
Optim_Option or Optim_PR is indeed positive. The correlation between MBSTR and MFE is
negative, suggesting that managers who meet or beat with a longer streak in the past are more
likely to meet or beat in the current quarter than those who meet or beat with a shorter streak.
This result can be explained by the fact that some managers structurally display a greater ability
to meet or beat forecasts than others. Our framework does not preclude (nor does it predict)
cross-sectional differences in managers’ abilities.
Rather, it predicts that managers who
experienced recent successes in meeting their own public forecasts will issue more optimistic
forecasts than their characteristics would otherwise predict (that is, in the absence of the bias). In
other words, our framework describes time-series rather than cross-sectional behavior. This
finding highlights the importance of using fixed effects in our analysis (as we do in Model (3)).
Our second hypothesis predicts a positive relation between our measures of optimism and
of performance. This is generally the case for the univariate correlation as the correlation
between ROA and either MFD, MBSTR, Optim_Option or Optim_PR are all positive. The one
exception is the univariate correlation between MFE and ROA). However, there is a mechanical
relation between ROA (earnings scaled by assets) and earnings realizations, and thus between
ROA and MFE (the difference between the managerial forecast and the earnings realization).
4.2. Dynamic over-optimism
We next test our first hypothesis. The empirical results for the estimation of Model (3)
are reported in Table 5. In the first column, we use the realization of the forecasted earnings as a
benchmark. Results indicate that MBSTR is significantly associated with the degree of optimism
relative to the subsequent realization (MFE) with a z-statistic equal to 2.70. In other words, the
increase in optimism does not merely and unbiasedly reflect good news about the economic
17
situation of the firm.
In the second column, we use the consensus analyst forecast as a
benchmark (MFD). Results indicate that MBSTR is significantly associated with the degree of
optimism relative to the consensus forecast with a z-statistic equal to 6.67. In the third column,
we use the option-based measure of optimism (Optim_Option). Results indicate that MBSTR is
significantly associated with this measure with a z-statistic equal to 3.00. Finally, in the fourth
column, we use the measure of optimism based on press releases (Optim_PR). Results indicate
that MBSTR is significantly associated with this measure with a z-statistic equal to 2.63. The
economic effect of MBSTR on the continuous dependent variable (i.e., MFE, MFD, and
Optim_PR) is 72 percent on average the size of the effect of Book-to-Market. 10 The marginal
effect of increasing MBSTR from 0 to 4 on the binary dependent variable, Optim_Option is seven
percent.
These results are consistent with our first hypothesis that individuals become
dynamically over-optimistic following a series of successes. Turning attention to the control
variables, we do not observe a robust pattern across the four columns.
We find that our conclusions are not affected if we control for the amount of special
items reported in earnings, total accruals, sales growth, return momentum (i.e., the marketadjusted return over the six months prior to forecast issuance), forecast experience, and year (and
fourth quarter) indicator variables (e.g., Doyle, Jennings, and Soliman, 2013; Hilary, Hsu and
Wang, 2014). Finally, our results continue to hold if we use firm fixed effects or Chief Financial
10
It is customary to evaluate the economic effect by comparing it to the median of the dependent variable. However,
in our case, this would amount to divide by a number very close to zero. Instead, we compare the effect of MBSTR
to the most statistically robust variable in our specification. To do so, we multiply the coefficient on MBSTR by the
one standard deviation of MBSTR, and divide the product by the coefficient on Book-to-Market and one standard
deviation of Book-to-Market for columns “MFE”, “MFD”, and “Optim_PR”.
18
Officer (CFO) fixed effects instead of CEO fixed effects. 11 We note that it could be argued that
there is a mechanical relationship between past forecast error and contemporaneous forecast
error if the error process is mean-reverting.
Using Optim_Option and Optim_PR largely
mitigates this concern.
4.3. Firm’s welfare
We then test our second hypothesis and report the results from Model (4) in Table 6.
They confirm that the firm’s performance gradually improves as managers become more
optimistic. MBSTR is positively correlated with the measure of performance with a z-statistic of
3.90. The economic effect of MBSTR on ROA is 11 percent the size of the effect of Book-toMarket. 12 The positive coefficients associated with MBSTR are consistent with the presence of
complementary effort.
When we consider Optim_Option and Optim_PR, the results are
consistent with those presented in Table 3 and both variables have a positive effect on ROA (the
respective z-statistics are 5.16 and 1.95). Our conclusions also hold if we use MFD as a
treatment variable (the untabulated z-statistic is 8.98). 13 In essence, these results, combined with
those in the prior section, suggest that a dynamically over-optimistic manager delivers earnings
that are below her expectations but above what an unbiased manager would deliver in a similar
environment.
11
The results are also robust to using SqrMFE and SqrMFD (the square root of MFE and MFD while retaining their
sign) to mitigate any skewness in the MFE and MFD, and to using the mean values of MFE and MFD by industry
based on the Fama and French (1997) classification to control for industry shocks.
12
We multiply the coefficient on MBSTR (0.049) by the one standard deviation of MBSTR (1.586), and divide the
product by the coefficient on Book-to-Market (|-4.550|) and one standard deviation of Book-to-Market (0.155) =
11%.
13
We do not consider MFE, the difference between the managerial forecast and the earnings realization, as there is a
mechanical relation between ROA (earnings scaled by assets) and earnings realizations.
19
These results hold if we use alternative dependent variables. Specifically, the results
continue to go through if we define ROA as the ratio of net income (instead of operating income)
to total assets (z-statistics of 4.32, 5.05, and 1.89, respectively), use the change in ROA instead of
the levels (z-statistics of 4.23, 1.95, and 2.59, respectively), if we use a binary variable equal to
one if earnings are greater than prior-quarter earnings and zero otherwise (z-statistics of 2.10,
4.55, and 1.80, respectively), or if we use a binary variable equal to one if the firm reports a
profit and zero if it reports a loss (the z-statistics equal 2.35, 2.36, and 1.92, respectively for net
profit and 3.15, 1.90, and 1.84, respectively for operating profit). 14 The results also hold when
we control for the mean of ROA by industry in the regressions and if we use firm fixed effects or
CFO fixed effects instead of CEO fixed effects. 15
Turning our attention to the control variables, we find that firms with high book-tomarket, high leverage, negative earnings, small size, and low coverage observe lower growth in
profitability. Our conclusions are also not affected if we control for the amount of special items
reported in earnings, total accruals, sales growth, return momentum, forecast experience, number
of outside directors, litigation risk, internal control weaknesses, fourth quarter and year indicator
variables.
4.4. Comparative statics
We next investigate some comparative statics that are suggested by our analytical
framework.
Van den Steen (2004) note (p. 1142) these comparative statics empirically
distinguish this framework from others.
14
We use a logit specification when the dependent variable is binary.
15
We control for the mean of ROA by industry to avoid confounding effects of macroeconomic shocks.
20
We first consider prior dispersion.
As Van den Steen (2004) notes, over-optimism
increases in the mean-preserving spread of the distribution of priors. Intuitively, noise will
increase with more open disagreement with respect to the success of different actions. We proxy
for this disagreement by calculating Disp, the dispersion in analyst forecasts (e.g., Diether,
Malloy, and Scherbina, 2002), as the standard deviation of analyst forecasts 90 days prior to the
issuance of a management forecast divided by price. We then split the sample based on the
median value of Disp and re-estimate Model (3) using our two continuous variables (MFE and
MFD) as the dependent variables in each sub-sample.
Van den Steen (2004) also notes that the bias should weaken as managers gain more
experience about the probability of success associated with different possible actions. The effect
of experience results from a reduction in the heterogeneity of the priors rather than a decrease in
the underlying mechanism that generates over-optimism. To investigate this possibility, we
calculate Exp as the number of quarters for which a CEO has issued management forecasts prior
to the current quarter. We then partition the sample using the median value of Exp.
Results for the two partitions are reported in Table 7. As predicted, over-optimism is
stronger when there is greater dispersion in analyst priors and when managers have less
experience. Chi-square tests indicate that the difference between coefficients is statistically
significant across the two subsamples in three out of four partitions. Taken together, these
results support an explanation based on the economic framework.
21
4.5. Over-optimism and overconfidence
Our results so far indicate that CEOs become dynamically over-optimistic. As explained
above several times, over-optimism and overconfidence are two distinct notions. However, a
key part of the underlying mechanism is a biased attribution that leads the agent to become
dynamically overconfident in her skill and gradually puts more weight on both her own
information and her own actions. Although our focus is on over-optimism (i.e., the belief that
future events are more likely to be positive than is justified), we also investigate the importance
of dynamic overconfidence in our setting. Specifically, we examine whether past successes
affect three measures of overconfidence (defined as placing too much weight on the accuracy of
private information and an excessive belief in personal skills): the likelihood of issuing a forecast
(ISSUE) in the current quarter, the range of the forecast (RANGE), and the likelihood of being
more accurate than the consensus forecast (ACCD). To do this, we estimate the following
model:
Confi,t = αi + β1 MBSTRi,t + γk Xki,t + εi,t,
(5)
where Conf represents the measure of overconfidence (ISSUE, RANGE, and ACCD). ISSUE
equals one if the manager issues a management forecast in the current quarter (and we have
sufficient data to estimate MBSTR), and zero if the firm has issued forecasts in the past (i.e., we
have sufficient data to estimate MBSTR) but not in the current quarter. RANGE is the absolute
value of the difference between the upper and lower bounds of the forecast divided by the stock
price (point estimates have a range of zero). ACCD equals one if the management forecast is
more accurate than the consensus analyst forecast and zero otherwise. MBSTR is our previously
defined measure of prior successes and Xk is the vector of control variables.
22
The untabulated results are consistent with dynamic overconfidence increasing with overoptimism. In particular, MBSTR is positively associated with the likelihood of issuing a forecast
(z-statistic of 6.98), negatively associated with the range of the forecast (z-statistic of -2.58), and
negatively associated with the likelihood of being more accurate than the consensus analyst
forecast (z-statistic of -2.41). In other words, after a series of successes, managers are more
likely to issue a forecast that is more precise but less accurate. As noted by prior literature (e.g.,
Hilary and Hsu, 2011), this pattern can be explained by dynamic overconfidence paralleling the
dynamic over-optimism that we identify in Section 4.2.
4.6. Link with Van den Steen (2004)
Van den Steen (2004) note (p.1143) that “the contribution of [his] paper is to formalize
this overoptimism mechanism in a general form, to demonstrate how it can cause several other
well-known judgment biases, and to derive comparative statics that determine when these biases
will be most pronounced and allow verification of the model.” Our empirical approach mirrors
his theoretical development.
His Section I (“Choice Driven Overoptimism”) states that "rational agents with differing
priors tend to be overoptimistic about their chance of success". This provides support for our
first hypothesis. Section II provides comparative statics that is the basis for our Section 4.4.
Section II.B (“Variance of Belief Distribution”) states that "the overoptimism also increases in
the mean-preserving spread of distribution of the priors".
comparative statics.
This is the basis for our first
Section II.C (“Previous Experience”) explains that "the bias would
disappear as the agents gain more experience". This is the basis for our second comparative
23
statics. 16 Section II.D (“Importance of Success in the Presence of Complementary Effort”)
shows that the bias is increasing in the benefit of success. This is the basis for our second
hypothesis. Section III discusses the assumption of differing priors, why debiasing does not
occur and why comparative statics is a good way to test the theory. This further supports our
analysis based on comparative statics. Section IV (“Relationship to Other Biases”) discusses
links with other biases such as asymmetric attribution of causality of success and failure, the
over-estimation of control and overconfidence. The importance of this section is twofold. This
provides a direct motivation for our Section 4.5. More importantly, it allows us to combine the
"static" overoptimism (discussed in his Section I) with a dynamic form of overconfidence. This
allows us to use panel specifications rather than a cross-sectional (or pooled) approach.
5. Robustness tests
5.1. Alternative definitions of successes
In our main specifications, we define a managerial success as a situation in which a
manager announced earnings that were above her own expectations. Our results hold if we
calculate MBSTR over three, five, or six quarters instead of four, and if we define MBSTR as the
number of successes over the last four forecasts instead of the number of successes in a row.
Our results are also robust to alternative definitions of successes.
Specifically, we
consider six additional specifications for MBSTR. In the first (STRMin or STRMax), we consider
the lower (or upper) bound of a range forecast (instead of the mid-point) to evaluate if a forecast
16
Section II.A (“Number of distinct Alternatives”) indicates that “Agents' overoptimism increase in the number of
distinct alternatives.” This is the only comparative statics we do not test for lack of a good empirical proxy for the
number of alternatives the manager faces.
24
is classified as “successful” (Ciconte, Kirk, and Tucker, 2014). In the second (STRGrowth), we
define a success as a realization that is above the prior-quarter realization (Miller, 2002). In the
third (STRDebias), we define a success as a realization that is above the forecast adjusted for the
systematic bias in the managerial forecast (i.e., forecast minus average forecast error for the CEO
over the entire sampling period) (Hilary, Hsu, and Wang, 2014). In the fourth, we define a
success as a realization that is above the manager's forecast and is positive. In the fifth, we
define a success as a realization that is greater than the prior-quarter realization and is positive.
Finally, in the sixth, we define a success as a realization that is above both the manager’s
forecast and the prior-quarter realization.
When we re-estimate Models (3) and (4a) using these six alternative measures of MBSTR,
our conclusions are not affected: the z-statistics range between 1.74 and 2.62 when MFE is the
dependent variable, between 3.26 and 8.78 when MFD is the dependent variable, between 2.12
and 5.16 when Optim_Option is the dependent variable, between 1.79 and 3.15 when Optim_PR
is the dependent variable; between 3.04 and 6.71 when ROA is the dependent variable.
5.2. Is the improvement in performance genuine?
As noted above, results from Table 6 are consistent with the idea that optimism and effort
are complements. However, an alternative explanation is that managers manipulate reporting to
deliver on the optimistic forecast once it has been announced. This could be done through either
accrual manipulation or real earnings management (e.g., Myers, Myers, and Skinner, 2007). To
investigate these possibilities, we first look for evidence of manipulation. We then examine
whether markets take the improvement in performance to be genuine.
25
To determine whether there is evidence of manipulation, we regress a measure of accrual
management (AccrMgt) and a measure of real earnings management (RealMgt) on MBSTR and
our usual control variables:
Mgti,t = αi + β1 MBSTRi,t + γk Xki,t + εi,t,
(6)
where Mgt represents earnings management as proxied by AccrMgt or RealMgt. We calculate
AccrMgt as the residual from a specification that regresses total accruals on assets, change in
sales minus change in accounts receivables, plant, property, and equipment, and lagged
performance (e.g., Kothari, Leone, and Wasley, 2005; Brown and Pinello, 2007). We calculate
RealMgt as a combination of the abnormal levels of cash flow from operations (OCF),
discretionary expenses, and production costs (e.g., Roychowdhury, 2006; Cohen and Zarowin,
2008).
We provide additional details on the construction of AccrMgt and RealMgt in the
Appendix.
Results for Model (6) are presented in the first two columns of Table 8. Inconsistent with
the presence of manipulation, the coefficients on AccrMgt and RealMgt are insignificantly
different from zero in both regressions, with z-statistics of 0.27 and 0.88, respectively. We also
use a performance-adjusted abnormal accrual measure for AccrMgt. For each quarter-industry
(two-digit SIC code) pair, we create four portfolios by sorting the data into quartiles of ROA
measured four quarters prior to the quarter of portfolio formation (Kothari, Leone, and Wasley,
2005). The abnormal accrual for a given firm is the unexplained accrual for that firm minus the
average (excluding the sample firm). Our conclusions are not affected. Our results continue to
hold if we adjust for seasonal effects in quarterly earnings (Louis and Robinson, 2005), or
control for lagged accruals in estimating discretional accruals (Gong, Louis, and Sun, 2008).
These alternative measures of AccrMgt could help mitigate the concern that MBSTR might be
26
potentially correlated with performance for other reasons. 17 Our results are also similar when we
use the three individual real earnings management proxies (R_OCF, R_PROD, and R_DISX) in
our test.
To examine whether markets take the improvement in performance to be genuine, we
regress AdjRet, the buy-and-hold market-adjusted return for the quarter, on MBSTR and our usual
control variables:
AdjReti,t = αi + β1 MBSTRi,t + γk Xki,t + εi,t.
(7)
If the improvement in performance is genuine, we expect MBSTR to be significantly positive.
The results, presented in Column 3 of Table 8, are consistent with this prediction. In
particular, after controlling for other factors (e.g., growth, investor sophistication) that make
MBSTR be potentially correlated with performance (e.g., Barth, Elliott, and Finn, 1999; Ke and
Petroni, 2004; Koonce and Lipe, 2010; Shanthikumar, 2012), the coefficient on MBSTR is
positive and significant with a z-statistic equal to 5.41. The economic effect of MBSTR on
AdjRet is 29 percent the size of the effect of Book-to-Market. 18 The results continue to hold if
we use raw returns or if we start the return accumulation period after the management forecast
announcement. They also hold if we define MBSTR as the number of successes over the last four
forecasts instead of the number of successes in a row. Last, our results remain similar if we
control for current performance (ROA) in the AdjRet specification (Kasznik and McNichols,
17
Untabulated results show that some of the alternative measures of AccrMgt are negatively associated with MBSTR
instead. It is consistent with the notion that firms with sustained superior earnings performance have better earnings
quality (Ghosh, Gu, and Jain, 2005).
18
We multiply the coefficient on MBSTR (0.761) by the one standard deviation of MBSTR (1.586), and divide the
product by the coefficient on Book-to-Market (26.731) and one standard deviation of Book-to-Market (0.155) =
29%.
27
2002). Overall, the lack of evidence in support of managerial manipulation and the presence of
positive returns suggest that the improvement in performance is genuine.
5.3. Increase in over-optimism or decrease in over-pessimism?
The results from Table 5 are consistent with the presence of dynamic over-optimism.
However, it could be argued that these results reflect a decrease in over-pessimism rather than an
increase in over-optimism. We investigate this possibility by turning our attention to analyst
reactions.
If there is a reduction in pre-existing (suboptimal) over-pessimism, managerial
forecasts should come closer to optimality and financial analysts should give more weight to
forecasts after a streak of successes. To test this conjecture, we regress REV on MBSTR and our
usual control variables, where REV is defined as the ratio of an individual analyst forecast
revision to the difference between the management forecast and the consensus forecast before
the issuance of a new management forecast, 19 and an individual analyst forecast revision is
defined as the difference between an analyst’s first forecast issued within 30 days after the
management forecast date and the latest forecast issued within 90 days before the management
forecast date.
In untabulated results, we find the coefficient associated with MBSTR is
significantly negative with a z-statistic of -2.93. This finding indicates that analysts give less
weight to these forecasts, which suggests that these forecasts are further from optimality than the
managerial forecasts unaffected by past successes. In other words, the positive effect of MBSTR
on our measures of optimism reflects an increase in over-optimism rather than a decrease in
over-pessimism.
19
We treat analysts who did not revise their forecasts within 30 days of a management forecast as missing
observations, although setting REV to zero for these analysts does not change our conclusions (untabulated).
28
6. Conclusion
Human inference and estimation is subject to systematic biases. In particular, there is a
long literature showing that overconfidence due to cognitive biases can lead to sub-optimal
decisions. We depart from this research by showing empirically that a) optimism is a related but
different bias, b) it can improve firm’s welfare, and c) it can emerge dynamically in a rational
framework.
Specifically, we show that managers of firms that have experienced recent successes are
more likely to subsequently issue forecasts that are more optimistic than they would have been
absent this bias. They are more likely to overinvest in their company options. Further, press
releases become increasingly optimistic as the firm experiences short-term successes.
Importantly, we also find that managers appear to exert greater effort to meet their overoptimistic forecasts: contemporaneous performance increases as managers become optimistic.
This is true for both ROA and market returns. In contrast, contemporary measures of accruals or
real earnings management are not affected by this optimism, suggesting that the increase in
performance is genuine and not the byproduct of managerial manipulation. Comparative statics
suggest that these results can be explained by an economic framework. In particular, consistent
with Van den Steen (2004), optimism increases with the diffusion of managerial priors and
decreases with managerial experience.
29
Appendix
Estimation of Optim_PR, AccrMgt, and RealMgt
Optim_PR
Optim_PR is the optimism score for the text of earnings related press releases in 8-K
filings as generated by DICTION software (www.dictionsoftware.com).
DICTION uses
dictionaries (word lists) to search a text for qualities such as certainty, realism, and optimism and
scores these qualities based on occurrence. The dictionary approach to textual analysis is well
established, dating back several decades (see, for example, the Harvard-IV-4 and Lasswell
dictionaries used in the General Inquirer: www.wjh.harvard.edu/~inquirer). DICTION is a more
recent application of this approach and has been applied in various scholarly fields to examine a
wide variety of public communications, including newspaper articles, political speeches,
corporate filings, and company press releases (for a list of scholarly research employing the
Diction language analysis program, see http://www.dictionsoftware.com/research-summaries).
Full texts of press releases are obtained from 8-K filings via SEC’s EDGAR. We average
optimism scores where more than one filing occurred within the same quarter for the same
company.
AccrMgt
We follow Kothari, Leone, and Wasley (2005) and Brown and Pinello (2007) to calculate
AccrMgt as the residual from the following regression estimated with quarterly data for all
COMPUSTAT firms in each industry-year (industry and firm subscripts omitted):
30
TAt / Assetst-1 = β0 + β1(1 / Assetst-1) + β2(ΔSALEt−ΔRECt) / Assetst-1) +β3(PPEt / Assetst-1) +
β4(NIt-1 / Assetst-1) + εt ,
(a)
where TAt is total accruals for quarter t, defined as earnings before extraordinary items and
discontinued operations minus operating cash flows.
Assetst-1 represents total assets at the
beginning of the quarter. ΔSALEit is the change in revenues from quarter t-1 to quarter t. ΔRECt
is the change in accounts receivable from quarter t-1 to quarter t, PPEt is net property, plant, and
equipment at the end of quarter t.
NIt-1 is the earnings before extraordinary items and
discontinued operations of quarter t-1. We require at least 15 observations to estimate the
regression.
RealMgt
We follow Roychowdhury (2006) and Cohen and Zarowin (2008) to calculate RealMgt.
Specifically, we consider the abnormal levels of cash flow from operations (OCF), production
costs, and discretionary expenses. We estimate the normal level of OCF using the following
cross-sectional regression for each industry-year:
OCFt / Assetst-1 = β1(1 / Assetst-1) + β2(SALEt) / Assetst-1 + β3(ΔSALEt / Assetst-1) + εt.
(b)
We estimate the normal level of production costs as:
PRODt / Assetst-1 = β1(1 / Assetst-1) + β2(SALEt) / Assetst-1 + β3(ΔSALEt / Assetst-1)
+ β4(ΔSALEt-1) / Assetst-1] + εt.
(c)
31
The normal level of discretionary expenses can be expressed as a linear function of sales:
DiscExpt / Assetst-1 = β1(1 / Assetst-1) + β2(SALEt-1) / Assetst-1 + εt.
(d)
In the above equations OCF is cash flow from operations in quarter t; Prod represents production
costs in quarter t, defined as the sum of the cost of goods sold and the change in inventories; and
DiscExp represents discretionary expenditures in quarter t, defined as the sum of R&D expenses
and SG&A. Abnormal OCF (R_OCF), abnormal production costs (R_PROD), and abnormal
discretionary expenses (R_DISX) are computed as the difference between the actual values and
the normal levels predicted from equations (b), (c), and (d). We require at least 15 observations
to estimate each regression.
We compute RealMgt by summing the three individual real earnings management
variables. Specifically, we multiply R_OCF and R_DISX by negative one so the higher the
amount of R_OCF and R_DISX, the more likely it is that the firm is engaging in sales
manipulation through price discounts and discretionary expense reductions. We do not multiply
R_PROD by negative one since higher production costs, as noted earlier, are indicative of
overproduction to reduce cost of goods sold.
32
References
Ajinkya, B., Bhojraj, S. and Sengupta, P., 2005. The association between outside directors,
institutional investors and the properties of management earnings forecasts. Journal of
Accounting Research 43, 343–376.
Armor, D.A., Taylor, S.E., 2002. When predictions fail: The dilemma of unrealistic optimism. in
Heuristics and Biases, The psychology of Intuitive Judgment, Editor, Thomas Gilovich, Dale
Griffin and Daniel Kahneman, Cambridge University Press, 334–347.
Baginski, S.P., Hassell, J.M., 1997. Determinants of management forecast precision. The
Accounting Review 72, 303–312.
Bamber, L.S., Cheon, Y.S., 1998. Discretionary management earnings forecast disclosures:
antecedents and outcomes associated with forecast venue and specificity choices, Journal of
Accounting Research 36, 167–190.
Barth, M.E., Elliott, J.A., Finn, M.W., 1999. Market rewards associated with patterns of
increasing earnings. Journal of Accounting Research 37, 387-413.
Benabou, R., Tirole, J., 2002. Self-confidence and personal motivation. Quarterly Journal of
Economics 117, 871–915.
Benoit, J.P., Dubra, J., 2011. Apparent overconfidence. Econometrica 79, 1591–1625.
Brown, L.D., Pinello, A.S., 2007. To what extent does the financial reporting process curb
earnings surprise games? Journal of Accounting Research 45, 947–981.
Campbell, T.C., Gallmeyer, M., Johnson, S.A., Rutherford, J., Stanley, B.W., 2011. CEO
optimism and forced turnover. Journal of Financial Economics 101, 695–712.
Cameron, A.C., Gelbach, J.B., Miller, D.L., 2011, Robust inference with multi-way clustering.
Journal of Business and Economic Statistics 29, 238–249.
Chuk, E., Matsumoto, D., Miller, G.S., 2013. Assessing methods of identifying management
forecasts: CIG vs. researcher collected. Journal of Accounting and Economics 55, 23–42.
Ciconte III, W., Kirk, M., Tucker, J.W., 2014. Does midpoint of range earnings forecasts
represent managers’ expectations? Review of Accounting Studies 19, 628–660.
Cohen, D.A., Zarowin, P., 2008. Accrual-based and real earnings management activities around
seasoned equity offerings. Journal of Accounting Economics 50, 2-19.
Compte, O., Postlewaite, A., 2004. Confidence-enhanced performance. American Economic
Review 94, 1536–1557.
Diether, K., Malloy, C.J., Scherbina, A., 2002. Differences of opinion and the cross section of
stock returns. Journal of Finance 57, 2113–2141.
33
Doyle, J.T., Jennings, J.N., Soliman, M.T., 2013. Do managers define non-GAAP earnings to
meet or beat analyst forecasts? Journal of Accounting and Economics 56, 40–56.
Fama, E., French, K., 1997. Industry costs of equity. Journal of Financial Economics 43, 153–
193.
Fombrun, C., Shanley, M., 1990. What’s in a name? Reputation building and corporate strategy.
The Academy of Management Journal 33, 233–258.
Gervais, S., Odean, T., 2001. Learning to be overconfident. Review of Financial Studies 14, 1–
27.
Gervais, S., Heaton, J.B., Odean, T., 2011. Overconfidence, compensation contracts, and capital
budgeting. Journal of Finance 65, 1735–1777.
Ghosh, A., Gu, Z., Jain, P.C., 2005. Sustained earnings and revenue growth, earnings quality,
and earnings response coefficients. Review of Accounting Studies 10, 33–57.
Glaser, M., Weber, M., 2010. Overconfidence. Behavioral Finance, Kolb Series in Finance,
Wiley, H. Kent Baker and John R. Nofsinger, editors, 241–258.
Gong, G. Louis, H., Sun, A.X., 2008. Earnings management and firm performance following
open-market repurchases. Journal of Finance 63, 947–986.
Hackbarth, D., 2008. Managerial traits and capital structure decisions. Journal of Financial and
Quantitative Analysis 43, 843–881.
Heaton, J.B., 2002. Managerial Optimism and Corporate Finance, Financial Management 31,
33–45.
Hilary, G., Hsu, C., 2011. Endogenous overconfidence in managerial forecasts. Journal of
Accounting and Economics 51, 300–313.
Hilary, G., Hsu, C., Wang, R., 2014. Management forecast consistency. Journal of Accounting
Research 52, 163–191.
Hilary, G., Menzly, L., 2006. Does past success lead analysts to become overconfident?
Management Science 52, 489–500.
Johnson, M., Kasznik, R., Nelson, K., 2001. The impact of securities litigation reform on the
disclosure of forward-looking information by high technology firms. Journal of Accounting
Research 3, 297–327.
Kasznik, R., McNichols, M.F., 2002. Does meeting earnings expectations matter? Evidence from
analyst forecast revisions and share prices. Journal of Accounting Research 40, 727–759.
34
Ke, B., Petroni, K., 2004. How informed are actively trading institutional investors? Evidence
from their trading behavior before a break in a string of consecutive earnings increases. Journal
of Accounting Research 42, 895–927.
Kothari, S.P., Leone, A.J., Wasley, C.E., 2005. Performance matched discretionary accrual
measures. Journal of Accounting Economics 39, 163–197.
Koonce, L., Lipe, M.G., 2010. Earnings trend and performance relative to benchmarks: How
consistency influence their joint use. Journal of Accounting Research 48, 859–884.
Lee, S., Matsunaga, S. R., Park, C.W., 2012. Management forecast accuracy and CEO turnover.
Accounting Review 87, 2095–2122.
Louis, H., Robinson, D., 2005. Do managers credibly use accruals to signal private information?
Evidence from the pricing of discretionary accruals around stock splits. Journal of Accounting
and Economics 39, 361–380.
Malmendier, U., Tate, G.A., Yan, J., 2011. Overconfidence and early-life experiences: The effect
of managerial traits on corporate financial policies. Journal of Finance 66, 1687–1733.
Miller, G.S., 2002. Earnings performance and discretionary disclosure. Journal of Accounting
Research 40, 173–204.
Moore, D.A., Healy, P.J., 2008. The trouble with overconfidence, Psychological Review 115,
502–517.
Myers, J., Myers, L., Skinner, D.J., 2007. Earnings momentum and earnings management.
Journal of Accounting, Auditing and Finance 22. 249-284.
Rabin, M., Schrag, J.L., 1999. First impressions matter: A model of confirmatory bias. Quarterly
Journal of Economics 114, 37–82.
Rogers, J.L., Stocken, P.C., 2005. Credibility of management forecasts. The Accounting Review
80, 1233–1260.
Roychowdhury, S., 2006. Earnings management through real activities manipulation. Journal of
Accounting and Economics 42, 335–370.
Shanthikumar, D.M., 2012. Consecutive earnings surprises: Small and large trader reactions. The
Accounting Review 87, 1709-1736.
Van den Steen, E., 2004. Rational overoptimism (and other biases). The American Economic
Review 94, 1141–1115.
Weinstein, N.D., 1980. Reducing unrealistic optimism about illness susceptibility. Health
Psychology 2, 11–20.
35
Table 1
Descriptive Statistics
ROA is defined as operating income divided by total assets at the beginning of the quarter, where operating income
is earnings excluding special or non-recurring items such as gain or loss from asset sales, asset write-downs or writeoffs. Optim_Option is option-based measure of over-optimism. Optim_PR is the optimism score for the text of
earnings related press releases in 8-K filings. Size is the log of the market value at the beginning of the quarter.
Book-to-Market is the book value divided by the market value of the firm's equity at the beginning of the quarter.
Leverage is the ratio of total liabilities to total assets at the beginning of the quarter. Inst Ownership is the
percentage of institutional ownership at the beginning of the quarter. StdEarn is the standard deviation of the
quarterly return on assets over at least six of the preceding eight quarters. RetVol is the standard deviation of the
stock return six months before the beginning of the quarter. Cover is the log of the number of analysts covering the
firm in a quarter. The means, medians, and standard deviations are computed for the Execucomp sample, which
begins in the first quarter of 1998 and finishes in the last quarter of 2010.
Variable
ROA
Optim_Option
Optim_PR
Size
Book-to-Market
Leverage
Inst Ownership
StdEarn
RetVol
Cover
N
Mean
Median
Std. Dev.
80,122
70,591
44,027
80,122
80,122
80,122
80,122
80,122
80,122
80,122
0.034
0.346
48.960
7.412
0.545
0.538
0.632
0.016
0.029
1.414
0.032
0.000
49.140
7.264
0.453
0.546
0.706
0.008
0.025
1.609
0.029
0.476
2.903
1.570
0.408
0.225
0.301
0.023
0.015
1.050
36
Table 2
Preliminary correlation table
ROA is defined as operating income divided by total assets at the beginning of the quarter, where operating income is earnings excluding special or non-recurring
items such as gain or loss from asset sales, asset write-downs or write-offs. Optim_Option is option-based measure of over-optimism. Optim_PR is the optimism
score for the text of earnings related press release in 8-K filings. Size is the log of the market value at the beginning of the quarter. Book-to-Market is the book
value divided by the market value of the firm's equity at the beginning of the quarter. Leverage is the ratio of total liabilities to total assets at the beginning of the
quarter. Inst Ownership is the percentage of institutional ownership at the beginning of the quarter. StdEarn is the standard deviation of the quarterly return on
assets over at least six of the preceding eight quarters. RetVol is the standard deviation of the stock return six months before the beginning of the quarter. Cover
is the log of the number of analysts covering the firm in a quarter. The Pearson correlations are computed for the Execucomp sample and are all significant at
the 5% level or less.
(1)
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
(9)
(10)
ROA
Optim_Option
Optim_PR
Size
Book-to-Market
Leverage
Inst Ownership
StdEarn
RetVol
Cover
0.245
0.076
0.201
-0.409
-0.225
0.153
-0.097
-0.224
0.099
(2)
0.047
0.061
-0.271
-0.087
0.034
-0.043
-0.078
-0.011
(3)
(4)
0.103
-0.076
0.040
0.022
-0.082
-0.105
0.050
-0.385
0.213
0.050
-0.215
-0.382
0.467
(5)
(6)
0.067
-0.072
0.033
0.353
-0.138
-0.126
-0.188
-0.109
-0.012
(7)
(8)
-0.088
-0.120
0.254
(9)
0.364
-0.065
-0.111
37
Table 3
Preliminary analysis of contemporaneous performance
This table reports regressions of contemporaneous firm performance (ROA) on measures of over-optimism based on
the Execucomp sample. Variables are defined in Table 1. Industries fixed effects and fiscal quarter fixed effects are
included. We define industries using the Fama and French (1997) 48 industry classification. For readability, all of
the coefficients are multiplied by 100. The z-statistics, which are reported in parentheses, are calculated using
double clustering by CEO and year to control for heteroskedasticity-consistent standard errors. ***, **, and *
indicate significance at the 0.01, 0.05, and 0.10 two-tailed levels, respectively.
Dependent Variable = ROA
Optim_Option
0.842***
(14.25)
Optim_PR
Size
Book-to-Market
Leverage
Inst Ownership
StdEarn
RetVol
Cover
Number of observations
Adj. R2
0.076*
(1.75)
-2.220***
(-9.70)
-2.040***
(-6.91)
0.584***
(4.63)
-9.551***
(-4.48)
-22.509***
(-3.75)
0.044
(0.84)
0.029***
(3.29)
0.090**
(2.14)
-2.530***
(-8.87)
-2.442***
(-8.62)
0.573***
(4.23)
-9.468***
(-3.85)
-17.411***
(-2.63)
0.002
(0.05)
70,591
33.48
44,027
32.12
38
Table 4
Correlation table based on management forecast sample
ROA is defined as operating income divided by total assets at the beginning of the quarter, where operating income
is earnings excluding special or non-recurring items such as gain or loss from asset sales, asset write-downs or writeoffs. MFE is the management forecast minus earnings realization divided by the stock price two days before the
issuance of the forecast. MFD is the management forecast minus the pre-existing analyst consensus forecast divided
by the stock price two days before the issuance of management forecast. MBSTR is the number of consecutive
successes that CEO i enjoyed over the last four quarters. Optim_Option is option-based measure of over-optimism.
Optim_PR is the optimism score for the text of earnings related press releases in 8-K filings. The Pearson
correlations are computed for the management forecast sample and are all significant at the 5% level or less.
MFE
MFD
MBSTR
Optim_Option
Optim_PR
ROA
MFE
MFD
MBSTR
Optim_ Option
-0.084
0.266
0.171
0.235
0.063
0.290
-0.121
-0.019
0.053
0.080
0.093
0.037
0.101
0.053
0.070
39
Table 5
Dynamic over-optimism
This table reports regressions of managerial optimism (MFE, MFD, Optim_Option, and Optim_PR) on the
number of past successes. Hor is the forecast horizon measured as the log of the number of days between the
management forecast date and the end of the fiscal period. Loss is an indicator variable that takes a value of one
if the earnings are negative, and zero otherwise. Other variables are defined in Table 1. CEO fixed effects are
included. For readability, the coefficients in columns “MFE” and “MFD” are multiplied by 100. The zstatistics, which are reported in parentheses, are calculated using double clustering by CEO and year to control
for heteroskedasticity-consistent standard errors. ***, **, and * indicate significance at the 0.01, 0.05, and 0.10
two-tailed levels, respectively.
Dependent Variable
MBSTR
Size
Book-to-Market
Leverage
Inst Ownership
StdEarn
RetVol
Cover
Hor
Loss
Number of
observations
Adj. R2
MFE
MFD
Optim_Option
Optim_PR
0.013***
(2.70)
0.153***
(4.12)
-0.279***
(-2.67)
-0.218**
(-2.19)
-0.047
(-0.71)
-0.478
(-0.94)
0.104
(0.20)
-0.003
(-0.70)
0.021**
(2.11)
0.461***
(6.60)
0.021***
(6.67)
-0.012
(-0.47)
-0.584***
(-8.20)
-0.127
(-1.55)
-0.169***
(-2.84)
1.010
(1.39)
-1.638***
(-3.82)
-0.004
(-0.80)
0.011
(1.35)
-0.611***
(-9.42)
0.142***
(3.00)
2.840***
(4.65)
-6.651**
(-2.54)
-7.932***
(-4.53)
-0.425
(-0.47)
11.910
(1.36)
7.170
(1.27)
0.001
(0.01)
0.102
(1.06)
-1.073**
(-2.57)
0.031***
(2.63)
0.223***
(3.72)
0.244
(0.65)
-0.384
(-0.89)
-0.030
(-0.10)
-0.040
(-0.04)
-0.912**
(-2.32)
-0.039
(-0.72)
-0.001
(-0.03)
-0.070
(-0.51)
10,630
10,032
9,225
6,295
33.87
37.78
41.37
35.22
40
Table 6
Contemporaneous performance based on management forecast sample
This table reports regressions of contemporaneous firm performance (ROA) on the number of past successes
(MBSTR), option-based measure of over-optimism (Optim_Option), and the measure of optimism based on press
releases (Optim_PR). Variables are defined in Table 1. CEO fixed effects are included. For readability, all of
the coefficients are multiplied by 100. The z-statistics, which are reported in parentheses, are calculated using
double clustering by CEO and year to control for heteroskedasticity-consistent standard errors. ***, **, and *
indicate significance at the 0.01, 0.05, and 0.10 two-tailed levels, respectively.
Dependent Variable = ROA
MBSTR
0.049***
(3.90)
Optim_Option
0.355***
(5.16)
Optim_PR
0.162*
(1.70)
-4.550***
(-9.47)
-1.599***
(-3.38)
-0.144
(-0.95)
-1.231
(-0.44)
-0.920
(-0.78)
0.059***
(3.01)
0.045**
(2.46)
-2.345***
(-18.51)
00.137
(1.49)
-4.380***
(-10.61)
-1.691***
(-3.75)
0.030
(0.19)
-1.234
(-0.41)
-1.039
(-0.80)
0.061***
(2.93)
0.050**
(2.50)
-2.357***
(-15.71)
0.014*
(1.95)
0.323***
(3.11)
-4.360***
(-9.72)
-1.255***
(-2.64)
-0.271**
(-2.37)
-1.468
(-0.39)
-1.734
(-1.36)
0.046**
(1.97)
0.074**
(2.35)
-2.392***
(-13.83)
Number of observations
10,626
9,221
6,293
Adj. R2
64.53
64.22
66.84
Size
Book-to-Market
Leverage
Inst Ownership
StdEarn
RetVol
Cover
Hor
Loss
41
Table 7
Comparative statics
This table reports regressions of managerial optimism (MFE and MFD) on the number of past successes using
partitions suggested by the economic framework. Exp is the number of quarters for which a CEO has issued
management forecasts prior to the current management forecast. Disp is the standard deviation of analyst
forecasts 90 days prior to the issuance of management forecast divided by price. Other variables are defined in
Table 1. CEO fixed effects are included. For readability, the coefficients are multiplied by 100. The zstatistics, which are reported in parentheses, are calculated using double clustering by CEO and by year to
control for heteroskedasticity-consistent standard errors. ***, **, and * indicate significance at the 0.01, 0.05,
and 0.10 two-tailed levels, respectively.
Panel A: MFE
MBSTR
Size
Book-to-Market
Leverage
Inst Ownership
StdEarn
RetVol
Cover
Hor
Loss
High Disp
Dependent Variable
MFE
Low Disp
High Exp
Low Exp
0.015**
(2.47)
0.227***
(4.27)
-0.116
(-1.14)
-0.139
(-0.90)
-0.107
(-1.16)
-0.399
(-0.55)
0.206
(0.31)
0.004
(0.25)
0.027
(1.44)
0.305***
(7.91)
0.000
(0.02)
0.037*
(1.80)
-0.288**
(-2.40)
-0.131
(-1.13)
0.080
(1.21)
-0.303
(-0.40)
0.413*
(1.72)
-0.006
(-0.80)
0.010*
(1.67)
0.036
(1.00)
0.024***
(5.00)
0.190***
(3.86)
-0.182
(-1.16)
-0.182
(-0.79)
-0.048
(-0.47)
-0.979
(-0.88)
0.804
(1.40)
0.003
(0.25)
0.021***
(2.59)
0.262***
(5.88)
χ2 test p-value
Number of
observations
Adj. R2
0.000
(0.18)
0.111***
(2.77)
-0.199**
(-2.40)
-0.156**
(-2.26)
-0.025
(-0.39)
-0.155
(-0.23)
0.062
(0.16)
-0.005
(-0.66)
0.023*
(1.75)
0.199***
(5.86)
0.001
0.031
4,588
4,589
5,882
4,744
31.28
41.26
37.60
33.30
42
Panel B. MFD
MBSTR
Size
Book-to-Market
Leverage
Inst Ownership
StdEarn
RetVol
Cover
Hor
Loss
High Disp
Dependent Variable
MFD
Low Disp
High Exp
Low Exp
0.029***
(6.43)
-0.043
(-0.97)
-0.336***
(-4.02)
-0.041
(-0.23)
-0.157
(-1.44)
1.743*
(1.87)
-2.150***
(-4.37)
-0.008
(-0.63)
-0.001
(-0.08)
-0.347***
(-7.09)
0.011***
(2.16)
-0.041*
(-1.94)
-0.334***
(-3.90)
-0.106
(-1.52)
-0.011
(-0.21)
-0.071
(-0.13)
-0.412
(-1.50)
-0.002
(-0.20)
0.021***
(5.11)
-0.120***
(-4.28)
0.030***
(6.04)
-0.073
(-1.34)
-0.471***
(-2.81)
-0.082
(-0.65)
-0.011
(-0.19)
0.461
(0.66)
-1.402**
(-2.27)
0.013*
(1.65)
0.006
(0.66)
-0.314***
(-5.52)
χ2 test p-value
Number of
observations
Adj. R2
0.022***
(4.30)
-0.0231
(-0.70)
-0.318***
(-5.16)
-0.061
(-0.90)
-0.160**
(-2.47)
1.060
(1.02)
-1.344***
(-2.69)
-0.017**
(-2.38)
0.018
(1.38)
-0.253***
(-6.80)
0.328
0.003
4,588
4,589
5,882
4,744
31.95
40.84
36.66
40.84
43
Table 8
Is the performance genuine?
This table reports regressions of earnings management measures (AccrMgt and RealMgt) and the marketadjusted return (AdjRet) on the number of past successes. AccrMgt is the residual from a specification that
regress total accruals on assets, change in sales minus change in accounts receivable and plant, property and
equipment. RealMgt combines three normalized metrics: the abnormal levels of cash flow from operations, of
discretionary expenses, and of production costs. AdjRet is the market adjusted return for the quarter. Other
variables are defined in Table 1 and Table 5. CEO fixed effects are included. For readability, all the
coefficients are multiplied by 100. The z-statistics, which are reported in parentheses, are calculated using
double clustering by CEO and year to control for heteroskedasticity-consistent standard errors.
Dependent Variable
MBSTR
Size
Book-to-Market
Leverage
Inst Ownership
StdEarn
RetVol
Cover
Hor
Loss
Number of observations
Adj. R2
AccrMgt
RealMgt
AdjRet
0.022
(0.27)
-0.034
(-0.04)
-1.472
(-0.30)
-2.200
(-0.85)
-3.103*
(-1.93)
24.302
(1.42)
-3.225
(-0.35)
0.037
(0.10)
-0.073
(-0.29)
0.160
(0.12)
0.063
(0.88)
1.736***
(3.28)
10.099***
(4.33)
7.681***
(3.55)
-0.691
(-0.75)
-11.303**
(-2.26)
-1.529
(-0.17)
-0.093
(-0.81)
-0.216
(-1.59)
1.140**
(2.03)
0.761***
(5.41)
-12.752***
(-9.26)
26.731***
(7.11)
10.902***
(3.53)
2.201
(0.79)
-13.619
(-0.61)
-48.037***
(-4.62)
-0.562
(-1.28)
0.915***
(2.61)
-10.265***
(-8.48)
10,271
4.33
9,700
71.75
10,597
10.10
44