The Market Impact of Corporate News Stories Werner ...assets.csom.umn.edu/assets/32652.pdf ·...
Transcript of The Market Impact of Corporate News Stories Werner ...assets.csom.umn.edu/assets/32652.pdf ·...
The Market Impact of Corporate News Stories
Werner Antweiler and Murray Z. Frank
December 6, 2004.
ABSTRACT
Fama (1998) points out that ‘splashy results’ get more attention and so there is an incentive
to find them. This suggests that published articles may reflect bias. Does splashy result bias
affect the usual understanding of the results from ordinary short-run event studies? We provide
some evidence that it does. The traditional idea is that once news is made public, it is fully
reflected in prices within at most a day or two. We test this idea using more than 180,000 Wall
Street Journal news stories from 1973 to 2001. Using computational linguistics methods we
classify the stories according to topic, and for each topic we run an event study. Our results
differ from popular impressions in several ways. 1. On average there is a reversal so that
pre-event and post-event abnormal returns have the opposite sign. 2. Statistically significant
return momentum is observed for many days after publication. 3. As a result, the inference
to be drawn from an event study is often very sensitive to the assumed event window. 4. The
average news story has a bigger and more prolonged impact during a recession than during an
expansion. (JEL classification: G12, G14)
Keywords: stock returns, news, event study, publication bias.
We would like to thank the following people and organizations. The SSHRC provided financial support. Helpful feed-
back and suggestions were provided to us by Henry Cao, Ming Huang, Jiang Wang, and the seminar audiences at the
Cheung Kong Graduate School of Business (Beijing), the Hong Kong University of Science and Technology, and the
University of British Columbia. Shih-Wei Yen provided excellent research assistance. Murray Frank also thanks the
B.I. Ghert Family Foundation for financial support, and the Cheung Kong Graduate School of Business for hospitality
while writing this paper. We alone are responsible for any errors.c©2004 by Werner Antweiler and Murray Z. Frank.
All rights reserved. Both authors are with the Sauder School of Business, University of British Columbia, Vancouver
BC, Canada V6T 1Z2. Werner Antweiler: Phone 604-822-8484, Fax 604-822-8477, [email protected].
Murray Frank: Phone 604-822-8480, Fax 604-822-8477, [email protected]. Corresponding author is
Frank.
I. Introduction
The fact that stock prices react reasonably and promptly to news ranks with the absence of
arbitrage as a key idea in modern finance. The literature contains many papers in which stock
prices are shown to react sensibly to news stories. This has had a fundamental influence on how
we think about financial markets. As a result event studies are a common tool in financial research.
There is however, reason to be concerned about our understanding of stock price reactions to
news stories. Fama (1998) argues that published studies may reflect bias. “Splashy results get
more attention and this creates an incentive to find them.” His idea is that many event studies are
initiated. Studies which obtain splashy results attract more attention and are more likely to be
published.
To see why this may be a serious problem, suppose that all events were truly insignificant. By
chance some will likely be found to be statistically significant, and these are the studies that will
tend to get published. Furthermore, suppose that a study starts to find results that the researcher
finds hard to understand or defend. Rather than continuing to waste time, the project may be
discontinued in favor of more promising topics. Not only might the anomalous results fail to get
through the editorial process, the results may never even be submitted for publication in the first
place. As a result, the published literature can create a seriously misleading impression of the
underlying evidence.1 In order to get around the bias, Fama (1998) suggests consideration of a
random sample of events. Of course, rather than limiting attention to a random sample, it is better
to examine the underlying population if possible.
In this paper we analyze the relationship between news stories and stock returns using con-
ventional event studies. Our data consists of 230,793 corporate news stories published in the Wall
Street Journal over the period 1973 to 2001. Initially the data is analyzed assuming that the simple
existence of a news story in the Wall Street Journal about a given firm on a given date constitutes
the event. Subsequently we follow Antweiler and Frank (2004) in using methods from computa-
tional linguistics to classify the news stories according to topic. An ordinary short-run event study
1This hypothesis has a significant history in statistics where it is known as ‘the file drawer problem,’ or ‘publica-tion bias.’ The file drawer problem has been shown to affect the published literatures in many academic disciplinesincluding medicine, psychology, economics, etc. Only successful studies are published. In order to be successful thestudy must not only have been carried out properly, but it must also be interesting. Insignificant results are generallynot interesting. Good references include Begg and Berlin (1988), DeLong and Lang (1992), and Sterling, Rosenbaum,and Weinkam (1995).
1
is conducted on each topic for which we were able to identify at least 50 news stories. Interaction
effects across topics are also studied. In addition to studying cumulative abnormal returns, we also
examine cumulative abnormal trading volume.
As discussed in many places such as Ross (2005) and Campbell, Lo and MacKinlay (1997), it
is widely believed that the typical event study pattern is fairly simple. The market is supposed to
jump right away when the news becomes public. It is well-known that information sometimes leaks
in advance. Thus before the event some drift of the same sign as the jump may occur. After the
event has become public thesubsequentcumulative abnormal returns should be zero since public
information is fully impounded right away.2 We call the conventional pattern the ‘once-and-for-all
jump’ hypothesis.
The once-and-for-all jump hypothesis has three stages: what happens before the event, what
happens at the moment of the event, and what happens after the event. We find that the once-
and-for-all jump hypothesis is only partly correct. The problem is with what happens after the
event.
Consider the three stages in sequence. Consistent with the leakage idea, there is often an
abnormal return observed the day before the news is reported. It is widely believed that when
information is revealed the market reacts strongly right away. For good news the abnormal return
is positive and for bad news it is negative. Many event studies have documented this kind of market
response. We find that this part of the once-and-for-all jump hypothesis is empirically correct.
It is also widely thought (e.g. Ross, 2005) that the market absorbs new information fully and
promptly, so that within at most a day or two, the news is fully impounded. However, this idea is
not commonly tested in ordinary event studies.3 We find that this part of the once-and-for-all jump
hypothesis does not appear to be empirically correct. In fact, the process does not typically end
within a day or two of the news being made public. The typical pattern is that after a day or two the
2When studying corporate news it is assumed that the events are small enough that they will have no effect on therisk-free rate, the equity risk premium or the risk profile of the firm. Boyd, Hu and Jagannathan (2004) discuss thedecomposition of stock returns when news arrives. Post-event momentum is generally regarded as inconsistent withinformation being fully impounded. Of course, as a matter of logic, post event momentum is not precluded by marketefficiency. If an event changes the riskiness of the firm, then average returns will be affected. This is usually regardedas a very minor effect over short periods of time.
3Long-run event studies have studied whether information can be used to predict returns many months or evenyears into the future. The interpretation of this evidence is controversial. Fama (1998) argues that the evidence reflectssplashy result bias. Loughran and Ritter (2000) are more sympathetic to the existence of long-run predictability as areal phenomenon.
2
returns start to drift in theoppositedirection from the initial jump. That drift commonly continues
for an extended period. During normal economic conditions the bulk of the adjustment is largely
completed after two or three weeks. During recessions the adjustment is more prolonged, and we
were not able to clearly identify the point at which it was typically completed.
Our findings about the typical abnormal return pattern have a number of potentially important
implications both for asset pricing, and for the event study literature. With respect to asset pricing
our evidence suggests that momentum is not only observed on the month-by-month time scales
studied by, for example, Jegadeesh and Titman (1993). It is also observable in daily data.4
Another way to describe our results is to say that the initial reaction to news is commonly
an overreaction—at least to judge by what typically follows over the next several days and
weeks. This kind of overreaction is quite different from the long-term overreaction documented
by DeBondt and Thaler (1985).5 Because it is a short-run phenomenon, the source may well stem
from some aspects of market microstructure.6 For example, it takes time for a trader with a large
portfolio to accumulate or unwind a large position in a stock. Griffin, Harris and Topaloglu (2003)
analyze some aspects of the dynamics of institutional trading.
The typical abnormal return pattern also carries an important implication for how we conduct
event studies. It is fairly common (e.g., Asquith, Mikhail, Au, 2004) for event studies to standard-
ize on a particular event window. Generally the window involves buying the stock a day or two
before the event and holding the stock until a day or two after the event has been made public.
However, because the typical pattern is a reversal, the conclusions of event studies are likely to
be sensitive to the event window. Of course not all events will be so sensitive. Nevertheless, re-
assurance on this front would be very helpful before placing too much confidence in the reported
abnormal returns from any particular event study.
4The fact that momentum is observable in daily frequency date may open the door to alternative approaches toinvestigating the sources of momentum. This stems from the fact that greater or lesser degrees of momentum may beassociated with different impulses or economic conditions. For example, we find greater momentum during recessions.
5DeBondt and Thaler (1985) found that NYSE stocks between 1926 and 1982 overreacted in the following sense.Stocks of firms that were doing particularly poorly during the previous three years were undervalued while stocks offirms that were doing particularly well were overvalued. This is on a completely different time scale from what we areobserving.
6This is why we describe the hypothesis being tested as the once-and-for-all jump hypothesis, rather than labellingit a test of the efficient market hypothesis.
3
Beyond the basic abnormal returns pattern, we also study the extent to which news reports are
independent and unpredictable. As might be expected, there is clear evidence that news coverage
is highly skewed towards large firms. There is also evidence that both financial factors and prior
news stories help to predict news reports. The most predictable kinds of news are news reports
related to operational issues such as capacity increases, signing of business contracts, and new
product announcements. Financial issues, corporate governance announcements, and issues related
to restructuring are much harder to predict.
While this predictability is statistically quite significant, the ability to predict particular kinds of
news in particular months is actually quite limited. There is also evidence of a lack of independence
across news topics, although the magnitudes of the conditionality are not very large.
The fact that most news stories are not very predictable is generally good news for event studies.
It implies that most of the stories that form the basis of analysis in event studies were probably
largely unexpected.
The rest of the paper is organized in the following manner. Prior research is discussed in
section II. Section III explains the computer linguistics methods used to classify the news stories
according to topic. In section IV the data are described. The main results are presented in V.
The effect of news on trading volume is considered in section VI. Differences across the business
cycle are considered in section VII. Our conclusions are presented in VIII. Some details of the
event study methodology are collected in Appendix A. Many tables that illustrate robustness and
examine predictability of the news stories are collected in a Technical Appendix.7
II. Prior Research
To the best of our knowledge the literature contains no previous attempt to examine the idea
that publication bias affects our understanding of how corporate news is reflected in short-run
event studies. Of course, the literature contains a very large number of individual event studies,
see Campbell, Lo and MacKinlay (1997). But by construction such studies focus on particular
types of events rather than studying all events at once as we do.
7The Technical Appendix is not intended for publication but will be made available on the authors’ web site.
4
We use computational linguistics methods to classify all corporate news from the Wall Street
Journal according to topic. Antweiler and Frank (2004) and Das and Chen (2004) both employ
computational linguistics methods to examine the role of internet stock message boards. Neither
of these papers use the algorithms to distinguish news according to topic. Instead they try to dis-
tinguish bullish from bearish messages. Both papers examine a limited number of firms, whereas
this current paper studies a broad population of publicly-traded American firms. Thus the focus,
scope, and results of these papers is quite different from the current paper.
Bhattacharya et al (2004) examine the market impact of news stories related to IPOs from
1996 through 2000. They conclude that the market tends to discount a fair bit of media hype. They
apparently used a large number of research assistants to classify the news rather than relying on
computer linguistics methods. They are interested in news related to IPOs rather than the broader
universe of topics that we consider.
Two previous studies have examined news headlines without classifying them according to
topic. Mitchell and Mulherin (1994) have studied the number of news headlines reported by Dow
Jones from 1983 to 1990. They have found that the number of news stories and aggregate mea-
sures of trading activity such as trading volume and the absolute value of firm-specific returns are
significantly related. Chan (2003) has studied the long-run effects of news reports in monthly data.
He has found negative momentum after bad news particularly in the smaller illiquid stocks. Nei-
ther of these studies has considered the role of the news topics, and neither has examined the daily
patterns that we focus on.
There have also been a number of papers that used news reports to examine the extent to which
news can account for the variation in stock returns. Particularly well-known papers here include
Schwert (1981), French and Roll (1986), Roll (1988), and Cutler, Poterba, and Summers (1989).
These studies have shown that news stories cannot account for a large proportion of the observed
stock returns. These studies have not asked about the typical stock returns and trading volume
patterns associated with various types of news.
Boyd, Hu, and Jagganathan (2004) find that the impact of news about the unemployment rate
differs depending on the business cycle. This result helped stimulate our interest in whether this
distinction might also prove to be important for the effects of corporate news. We find that there
5
are important and significant differences in how the market reacts to news during recessions and
boom periods.
III. Classifying the News
In order to analyze the news articles from theWall Street Journalwe need to classify a very
large number of text items that represent company-specific events into a set of event topics. Fol-
lowing Antweiler and Frank (2004) we use machine learning techniques for automated text classi-
fication developed by computational linguists.8 The Naıve Bayes classifier is among the most suc-
cessful known algorithms for learning to classify text documents, parametric or non-parametric.
This algorithm has been employed very successfully in numerous applications. It may be best
known today for use in spam filters that screen e-mail messages. While some of the competing
algorithms such as SVM also perform quite well, Naıve Bayes combines an appealing simplicity
with robustness in performance, and reduces the need to fine-tune the learning process. One of
the most important selection criteria for choosing a particular algorithm is the replicability of the
algorithm and absence of arbitrary fine-tuning parameters. Naıve Bayes fulfills these requirements.
The basic assumption of Naıve Bayes classifiers is that all attributes of the news stories are
independent of each other given the context of the event category. This is what is called the
“Naıve Bayes assumption.” Even though this assumption will typically be violated in real-world
applications, the Naıve Bayes classifier remains remarkably robust. In fact, by artificially trimming
the vocabulary of words that are used in the text classification, the naıve Bayes assumption is
violated a priori.
The Naıve Bayes classifier is also quite immune to linguistic problems such as negation of
words. For example, “earnings are good this quarter” and “earnings are not good this quarter” carry
very different meanings. As the word “good” may be associated strongly with the “earnings up”
category, its negation will tend to misclassify the particular news stories. However, if news stories
8Computer linguists use both parametric classifiers such as Naıve Bayes (NB) and non-parametric classifiers suchas Support Vector Machines (SVM), neural networks (NNet), and k-nearest neighbour (kNN). See Hastie, Tibshirami,and Friedman (2001) for a discussion of the underlying statistical concepts.
6
are long enough, a large-number effect tends to corrects the problem with occasional negations.9
Using a Naıve Bayes classifier proceeds in the following steps:
1. We acquire the data set of Wall Street Journal newspaper stories through Lexis-Nexis. Each
news story provides a title and the first paragraph of the news story. In the original data, each
news story is already classified by firm before we start our analysis.
2. We extract a random sample of 2,000 news items, about 1/125-th of our total data and call
this the “training data set.”
3. The training data set is classified manually, which serves two purposes: it identifies the event
space (the list ofC event categories); and it assigns the news stories from the training data
set to specific event categories.
4. The vocabulary ofM words is computed by applying the information gain (IG) criterion
described below. Limiting the vocabulary has been shown to greatly improve performance
of the Naıve Bayes classifier.
5. The Naıve Bayes classifier is applied to the entire data set to calculate posterior odds for all
news stories and event categories. For each of theN news stories there areC posterior odds.
The posterior odds are interpreted as binary state variables using a probability thresholdθ.
For Naıve Bayes text classification we have employed the Rainbow package developed by
McCallum (1996). This software can be downloaded freely for academic purposes from the web
at http://www.cs.cmu.edu/˜mccallum/bow/.
Our first task is to define these event categories based on reading a random sample of the news
articles. We defined 67 basic news topics and permitted 7 possible qualifiers. However many of
them are lightly populated. When we restrict attention to topics with at least 50 stories we only use
about 43 of these categories as shown in the various Tables. The requirement of 50 stories interacts
with the various requirements so that there is slight variation in the number of topics included
depending on the Table.
9This approach, an example of a “bag of words” approach to text classification, makes no direct use of the gram-matical structure. As an empirical matter it has been found that a surprisingly small amount is gained at substantial costby attempting to exploit grammatical structure in the algorithms. For a helpful discussion of the various approaches toanalyzing text see Manning and Schutze (1999).
7
Each newspaper article contains an information set composed ofm ≤ M wordsωj. Let
j ≡ {ω1, ..., ωj} denote the partial information sets forj ≤ m. A particularly convenient form of
Bayes Theorem makes use of logs of odd ratioso ≡ p/(1− p). We are interested in predicting the
odds of a news article belonging to news event categoryEi. The log posterior odds for information
setj+1 is the sum of log prior odds for information setj and the log odds of the new information
ωj.
ln o(Ei|j+1) = ln o(Ei|j) + ln o(ωj|Ei) (1)
Summation of log odds ratios is not only appealing from the point of view of algebraic simplicity,
it is also a preferred computational method that avoids overflow and underflow conditions when
dealing with sequences of very small probabilities.
Iterating over the elements of the information set for a given news article, the final posterior
odds that the news item belongs into event categoryEi is given by
ln o(Ei|m) = ln o(Ei|0) +m∑
j=1
ln o(ωj|Ei) (2)
whereo(Ei|0) = oi is the prior odds for the empty information set0 = ∅. Prior odds are derived
from the frequency of events in the training set for the algorithm:oj = ni/(N − ni), whereN is
the total number of news articles in the training set andni is the number of news articles classified
as event typeEi.
An obvious problem in the calculation of the odds ratioo(ωj|Ei) is the case where wordωj
only occurs in classEi. This would make wordωj a “perfect” predictor of classEi with an
infinite odds ratio because the denominator is zero. Likewise, if wordωj never occurs in class
Ei, its absence would also identify classEi with perfect precision, and in this case the odds ratio
would be zero. In either case,ln o(ωj|Ei) is undefined. To overcome this problem, computer
linguistics have introduced Laplace smoothing, which introduces a small “fudge factor” to avoid
zero or infinite odds ratios. Concretely, Laplace smoothing redefines frequentist probabilities such
that pt = nt/∑T
t=1 nt is transformed into~pt = (nt + κ)/(κT +∑T
t=1 nt) so that it still holds
8
that∑T
t=1 ~pt = 1. Hereκ is a very small positive number. In terms of the odds ratio, Laplace
smoothing can be expressed as redefining
~o ≡ ~p
1− ~p=
p + κ/T
1− p + κ(T − 1)/T(3)
It is easily seen that nowln o ∈ [κ0, κ1], whereκ0 ≡ ln(κ) − ln(T + κ(T − 1)) < 0 andκ1 ≡
ln(T + κ)− ln(κ)− ln(T − 1) > 0.
A further challenge is to infer a binary state variable from each odds ratio for a news event.
The posterior odds ratio is transformed into a ]0,1[-bounded probability throughp(Ei|m) =
o(Ei|m)/(1 + o(Ei|m)). It is necessary to define an arbitrary thresholdθ that mapsp to the
binary states 0 or 1. We chooseθ = 0.5. However, in our robustness checks (reported in the
Technical Appendix of our paper) we have also applied thresholds ofθ = 0.2 andθ = 0.8. This
particular parameter choice does not adversely affect our results.
The definition of the vocabulary (i.e., the wordsω that define the dictionary for text classifica-
tion) is an important practical matter in text classification. While all words can be used potentially
to classify text, truncation of the vocabulary is a proven method to improve the performance of
the Naıve Bayes classifier. The Naıve Bayes assumption states that words are independent of each
other. However, in practical applications words appear in context and thus tend to be highly corre-
lated with each other. The basic idea behind pruning the dictionary is the selection of an effective,
or representative set of words. The most valuable words for identifying eventEi are those that are
distributed most differently in the sets of positive and negative examples of eventEi. This idea
leads to the definition of the information gain (IG) for wordωj relative to a specific event category
Ei:
IG(ωj, Ei) =∑
ω∈{ωj ,�ωj}
∑E∈{Ei, �Ei}
p(ω,E) ln
[p(ω,E)
p(ω)p(E)
](4)
Herep(ω) andp(E) are simple frequentist probabilities of words and events in the training data
set, andp(ω,E) is the joint probability for wordω and eventE. The summations are over the
four permutations of eventEi and non-event�Ei, and presence and absence of wordωj. As we are
9
dealing with multiple event categories, it is necessary to compute a weighted sum of information
gains in order to calculate the information gain for a particular wordωj:
IG(ωj) =∑
i
p(Ei) IG(ωj, Ei) (5)
After computing the information gain for each word, only the words with the largest information
gain are included in the dictionary.10 In our particular application we use a dictionary of 1,000
words.
IV. Data Description
Figure 1 illustrates that the number of corporate news stories published by the Wall Street
Journal was higher in the 1990s than it was earlier. There was a drop in the number of stories
during 1988. The proportion of news stories that we are able to classify is fairly stable over time.
We are able to classify 186,073 news stories.
Figure 2 illustrates the number of stories according to topic for all topics for which we were
observed at least 100 stories. Corporate restructuring is discussed in 41,000 stories. Operational
issues are discussed in about 35, 000 stories. More than 30,000 stories are related to earnings
reports. Around 7,000 stories are related to firm financial issues. Around 6,000 stories relate to
legal issues. Of course, many other subject areas are also discussed. These subject areas are broken
down further in Figure 2 as well.
When we turn to specific topics for which we run event studies, the most common events
are news stories that report on a contract, earnings reported up, earnings reported down, merger
announcements, announcements that a unit of the firm has been sold, and stories that report on the
appointment or election of an executive.
Figure 2 also illustrates the fraction of each type of story that is attributable to each size decile.
We measure the size decile by the firm’s market capitalization at the time the story is reported. Not
too surprisingly larger firms get much more coverage. For all classified news stories, the top decile
of firms are the subject of 54.3% of the stories. The extremely high coverage of the top decile of
10The information gain criterion is not the only method for truncating the dictionary. Another popular approachinvolvesχ2 statistics.
10
firms is true for almost all classes of news. The only noteworthy exception is for stories that report
a listing change. In this case smaller firms also receive a fair bit of coverage.
Our empirical analysis uses average standardized cumulative abnormal returns (ASCARs), also
referred to as “abnormal returns” for short. The underlying methodology is described in Appendix
A. An extensive discussion of this methodology can be found in Campbell, Lo, and MacKinlay
(1997, chapter 4). “Abnormal” is relative to the specific market model. “Cumulative” is relative to
the event window, with cumulation commencing on a day defined relative to the event day. “Stan-
dardized” refers to the normalization of cumulative returns relative to the variance of returns for
each stock as determined through the market model regression. “Average” refers to the averaging
of standardized cumulative abnormal returns across all observations (stocks by event day) in our
sample. Similar to ASCAR we also define average standardized cumulative abnormal trading vol-
ume, or ASCATV. Trading volume undergoes the same normalization procedure as returns, with
logarithms of stock trading volume and market trading volume replacing the usual role of stock
returns and market returns.
V. Results
The day that news is actually reported in the Wall Street Journal is well defined. We call this
day zero, or the event day. It is not clear when the news first became known to at least some people
who trade. Accordingly, event studies often include a day or two before the actual event. Event
studies also differ in the event window considered. It is not unusual for a study to standardize on
a particular base day and a particular event window. This seems to reflect a belief that in general
these choices are not critical.
Table 1 presents overall results for all news treated together. Results are presented for event
windows that start from three days before the event date to three days after. Event windows of
5, 10, 20 and 40 trading days are presented. Most event studies use daily data, while others use
monthly data. The reported ranges of event windows extend beyond a month.
The Wall Street Journal is published in the morning. The prices used are standard CRSP closing
prices. So an event window that starts on the event date is based on information that should already
be reflected in the prices. Suppose the news did not leak. Then buying one day before the news
11
should be at prices that do not yet reflect the impact of the news. On the other hand the news
might also be reported by wire service or over the internet on one day and then published the
next morning in the Wall Street Journal. In that case the relevant date is shifted one day earlier.
Antweiler and Frank (2004) find evidence that suggests a one day earlier shift might be relevant.11
Table 1 shows that the choice of length and starting day of the event window is important.
Consider the 5-day event window. All event windows that start before the publication of the news
have statistically significant positive abnormal returns. On average news is good news. But the
picture changes sharply if we define a 20-day event window. Now all three starting dates have
negative and significant results. On average news is bad news. Thus there is no general answer to
the question of whether news is on average, good or bad. It depends on the precise specification of
the event window.
To understand this difference look at the 5-day event window returns associated with buying
after the news is already public. According to the once-and-for-all jump hypothesis these should all
have zero abnormal returns. They do not. In fact buying a day or two after news is reported in the
Wall Street Journal has a negative abnormal return. Under the once-and-for-all jump hypothesis we
might expect the abnormal returns to be either unchanged, or perhaps be gradually overwhelmed
by the underlying noise in the market as we lengthen the event window. In fact, the reverse is
observed as the event window is increased. The abnormal return generally becomes stronger as the
event window is extended to include more post-event days.
The evidence in Table 1 rejects the once-and-for-all jump hypothesis. Significant abnormal re-
turns are seen when buying on public information. The idea that—once made public—information
is immediately fully impounded in prices does not describe the evidence. This might reflect large
investors gradually adjusting their portfolios. It is known to take time to accumulate or unwind a
large position. The results in Table 1 are consistent with this fact.
The number of observations underlying Table 1 is fairly large, but one might still be concerned
that the distributions of returns might be non-normal. To mitigate this concern, Table 2 reports
median standardized cumulative abnormal returns in place of means. If the results reported in
Table 1 were being driven by the tails of the distribution, then the median SCARs should be more
11Mitchell and Mulherin (1994) compare news reports in the Wall Street Journal to news reports on the Dow JonesBroadtape over the period 1983-1990. They report that it is very rare for a news story to lag more than one day.
12
robust. Numerically the results differ a bit. The shift from positive to negative returns becomes
somewhat stronger. But the basic patterns from Table 1 are repeated. So the failure of the once-
and-for-all jump hypothesis seems to be robust in this sense.12
Tables 1 and 2 show that the typical response to news is first positive and then negative. This
seems like a process of overreaction. Much of the rest of the paper is devoted to decomposing this
result and checking it for robustness from various points of view.
A. Results by Topic
The first kind of decomposition is by topic. Obviously some kinds of news are good news,
while other types of news are bad news. The Wall Street Journal presumable chooses what to
report in an effort to make money. There is no apparent reason that this should lead to an equal
number of good news stories and bad news stories. It depends on what the editors think people
want to read.
As discussed in section III, we use the Naıve Bayes Algorithm to classify the news according
to topic. Table 3 presents basic decomposition results by topic. Only topics with at least 50 stories
are reported. The event window is started two days prior to the news being reported.
Many of the individual results match conventional expectations. For example, the five-day
response to ‘earnings reported down’ is negative while the response to ‘earnings reported up’ is
positive. Debt issues have no market effect while equity issues have a negative effect. Dividend
increases and stock splits have a positive five day effect on the market. As far as we know, the
literature does not contain an earlier comprehensive treatment such as that provided in Table 3.
Following Brown and Warner (1985), it is widely believed that short-run event study results
are pretty robust to minor variations of methodology. For instance, Campbell, Lo and MacKinlay
(1997) suggest that improved normalizing models might affect the variance slightly, but the gains
from an improved normalizing model are said to be small.
There are important implications of the belief that short-run event study results are quite robust.
Event studies are sometimes published without providing evidence of robustness to alternative
12This should help to mitigate any concern that the results are driven by thin trading. Further evidence that thintrading is not the key force is found in the technical appendix Table A-20. There are almost exactly as many statisticallysignificant abnormal returns in the top trading volume quartile as there are in the bottom trading quartile. The averagemagnitude is slightly greater in the bottom trading quartile.
13
event windows, to alternative normalizing models, etc. For example, consider the very interesting
recent study of equity analysts reports by Asquith, Mikhail and Au (2004). Quite properly their
paper spends a fair bit of time discussing the nature of the equity analysts reports. However, when
conducting the event study itself they standardize on a five day event window around the event,
and do not discuss the effects of alternative market models or alternative event windows. This is
fairly typical and it apparently grows out of the belief that short-run event studies are robust on
these dimensions.
The results in Table 3 reemphasize the findings from Tables 1 and 2 that ASCARs depend on
the length of the event window. For about half of the event types, the 5-day ASCARs differ either
in sign or in statistical significance from the 40-day ASCARs. Thus a lack of robustness across
event windows is a very common phenomenon. An extreme example is ‘merger announce’ which
is positive and significant with a 5-day event window, but negative and significant with a 40-day
window. Most instances are not so extreme in their differences.
Many event studies attempt to control for other kinds of news that might confound the results.
A common procedure is to discard events for which other kinds of news is reported around the
same period. Table 4 examines the effect of conditioning on other news reports concerning the
firm during the last two calendar weeks. We report results for another news report of the same
type (own), for some other kind of news about the same firm (other), and for no other news stories
(none). The number of stories under each heading is as follows: all classified has 181,684 stories,
own prior event has 34,283 news stories, other prior event has 82,907 stories, and no prior event
has 98,777 stories. Thus restricting attention to only those events for which there are no prior
stories costs almost half of the sample. It is usually assumed that ‘none’ is relatively clean when
compared to the other cases. If that is true, then the ‘none’ column ought to have the greatest
number of statistically significant results. This is not a strongly supported result. In Table 4 we
find that 21 classes of news are statistically significant under ‘none,’ while 18 are significant under
‘other,’ and 23 are significant under ‘own.’
In Table 1 we saw that there was typically a change of sign once the return became public. To
evaluate this phenomenon, Table 5 considers individual news types for event windows that start one
14
day after the news is published.13 Surely if anything is in the public information set, it is the news
that was published in yesterday’s Wall Street Journal. Thus under the classical interpretation of the
efficient market hypothesis all the abnormal returns in Table 5 should be statistically insignificant
apart from sampling variation. This idea could hardly be more at odds with the evidence. Overall
the average returns in Table 5 are negative, and they grow in magnitude as time accumulates. Many
of the individual types of news also become increasingly extreme as time is accumulated.
A common issue in carrying out an event study is the choice of proxy for the market. In most of
our analysis we have used the CRSP equal-weighted market index. The most common alternative
is the CRSP value-weighted index, which gives greater weight to larger firms. Both indicies have
advantages and disadvantages. Table 6 replicates Table 3 with the value-weighted index used in
place of the equal-weighted index. The key results from Table 3 do hold up in Table 6, but the
numerical magnitudes are affected. Once again a positive abnormal return is found for a five day
event window, but that reverses as the window is lengthened. The longer term negative return is
somewhat larger in Table 6. Most of the individual results are fairly similar in the two Tables.
Thus the basic results that we study do not seem to be a simple artifact of the choice of market
index. This is quite consistent with the common idea that the choice of market index is not all that
important in an event study.
It is well known that firm size is related to average stock returns. Accordingly, it is not surpris-
ing that firm size also matters in event studies. Comparing ASCARs by firm size tends to produce
results that are on average more negative for large firms and more positive for small firms.14 This
is just a reflection of the small firm effect as in Banz (1981). The standardization in ASCARs is
tantamount to averaging using inverse variances as weights. To the extent that inverse variances are
related positively to firm size, ASCARs tend to weight larger firms more strongly than small firms.
As our data set of WSJ stories contains a preponderance of large firms (see Figure 2), ASCARs
will exhibit a tendency to be slightly negative.
13There are subtle differences in the number of news observations from Table to Table as we change the eventwindow or make other changes to the specification. These differences originate from the matching of news eventswith returns data. For an observation to be included in our analysis there must be a sufficient number of observationsbefore and after the event in order to perform a market model regression and to calculate cumulative abnormal returns.There must also be nonzero variance in stock returns during the pre-event reference period.
14This result is reported in our Technical Appendix Table TA-9.
15
B. Not All News is Equally Important
For most of the analysis news of all directions and magnitudes are included. Thus we have
a mix of some effects that are positive, some that are unimportant, and some that are negative.
In Table 7 we try to disentangle these differences. For any given news story we can measure the
market return on the day the story is published. For simplicity in Table 7 we ignore the topic of
the news story. We sort the stories into deciles according to the return on the publication day. For
each decile we track out the typical returns starting one day later. For comparison we also took a
random sample of stocks (5%) that had no news.15
Under the once-and-for-all jump hypothesis, all ASCARs should be zero because we are condi-
tioning on public information. If the market underreacts to news then the sign of the returns should
be the same as the sign of the first day return for the given decile. It the market typically overreacts
then the sign of the returns should be the reverse of the news day return for the given decile.
Table 7 shows that the extreme positive return decile is followed by negative returns. The
extreme negative decile is followed by positive return.Big news is followed by reversals. This is
true both for large returns with news, and for large returns without news. When there is news, the
process seems to remain large for a longer period of time. In this sense there is more momentum
with news than without.
C. Predicting News
The predictability of news stories is potentially important. If news stories are predictable, then
under the idealized view of an event study the publication story will not affect returns. For the
event study method to be generally powerful, it is important that most news reports should really
be news to the market. Graham, Koski, and Loewenstein (2004) make a related point in a study
that compares anticipated and unanticipated dividend announcements. We put considerable effort
into attempts to predict the news stories using financial factors and other news stories. News is
statistically predictable. But the financial magnitudes of these effects are small.16
15Table TA-17 in the Technical Appendix also provides two more 5% random samples of no news returns resultswith cumulation commencing one day and three days before the event day, respectively. The no news returns resultsare pretty robust.
16Because the financial magnitudes are rather small, these tables have been relegated to the Technical Appendix.In Table TA-3 we have used a logit model with monthly data for predicting news, relying for our regressors on prior
16
VI. Trading Volume
Table 8 presents the abnormal trading volumes defined by analogy to the abnormal returns.17
We regress the log of trading volume of a particular stock on the log of total trading volume of the
market during a 120-day pre-event window.18 Conditioning on the market trading volume captures
day-of-week effects along with other common factors.
Abnormal trading volume usually tapers off over time. This can be seen by comparing the 5-day
post-event periods with longer horizons. Some event types are associated with very large additional
trading activity, such as reports of earning forecasts down, stock splits, merger announcements or
failed mergers, and ongoing lawsuits. News events are rarely associated with lower trading activity
in the short term.
An important issue is how the abnormal trading volume relates to abnormal returns.19 Accord-
ing to the once-and-for-all-jump hypothesis an event can be associated with either a positive or a
negative return, but once the news is public on average there should not be further accumulation
(momentum) nor should there be a reversal. Abnormal trading volume is not similarly restricted. It
is plausible that for many types of events there will be a surge of trading as people re-balance their
portfolios, followed by a return to normal trading volume. However some kinds of events such as
a stock split might be associated with a permanent increase in trading volume when compared to
what was perviously the norm for that firm.
In order to see how the cumulative abnormal trading volume is related to the cumulative ab-
normal returns we plot them against each other on a day by day basis starting two days before the
publication date and extending until 37 days after the publication date. Some examples of these
ASCAR/ASCATV plots are collected in Figure 3.
news (autoregression), financial observables (returns, volatility, trading volume), and firm size. Prior news increasesthe likelihood of further news. As observed in Figure 2, larger firms are also more likely targets of WSJ stories. Largertrading volume also precedes news.
17Table 8 with ASCATVs includes about 2.5% fewer event observations than the corresponding Table 3 with AS-CARs. This is due to the non-zero variance requirement for the pre-event reference period. While CRSP providesreturns based on bid/ask midpoint changes in the absence of transaction-based changes, trading volumes will remainzero when no trades occur.
18Trading volume for individual stocks is expressed as the number of shares, adjusted for stock splits. Total markettrading volume is the unadjusted sum of the traded number of shares of all individual stocks.
19Thin trading is often a concern is studies of stock returns. In order to examine whether thin trading is crucial forour results, we sorted the trading volumes into quartiles. If the results are driven by thin trading then the lowest tradingquartile should be most significant. This does not prove to be the case, and so the Table is relegated to the TechnicalAppendix as table TA-20.
17
For all news together Figure 3 illustrates that on average a news story in the Wall Street Journal
induces a positive abnormal return and a positive abnormal trading volume. However these positive
abnormal shocks are then reversed over then next several weeks. This pattern is by far the dominant
pattern in the data.
Some events have a sharply different response. The lower left panel of Figure 3 plots the
response to ‘earnings forecast down.’ In this case there is a sharp negative abnormal return initially
coupled with an increase in abnormal trading volume. But once again it appears that the initial
response is typically reversed over the next few weeks.
The typical pattern as in ‘all events’ and in ‘merger announce’ is also observed for: elect CEO,
elect director, elect executive, elect VP, earnings forecast up, earnings reported up, dividend up,
equity repurchase, listing change, miscellaneous, opinion, supply agreement, lawsuit by ongoing,
capacity up, product new, product new possible, joint venture restructuring, joint venture start,
merger announce, merger fail, merger fail possible, unit bought announce, unit sold announce, unit
sold possible. In each case there is an initial increase in both volume and in returns. In each case
a day or two after the news there is a negative drift in both the returns and in the trading volume.
The magnitudes differ from event type to event type.
The ‘bad news’ pattern associated with earnings forecast down, is also observed for: earnings
reported down, dividend down, rating down, and layoff. In these cases there is an initial negative
jump in returns and a positive jump in trading volume. In each case a day or two after the event the
trading volume starts to drop while the returns reverse direction and head in the positive direction.
The remaining event topics depart from these two common patterns. These departures take a
variety of forms. One of the more interesting examples is that of stock splits as depicted in Figure
3. As usual there is a jump followed by a returns reversal. But in this case after a few weeks the
trading volume starts to rise. This is consistent with the common idea that firms split their stocks
in order to generate more trading activity.
Since Banz (1981) we know that average stock returns differ between large and small firms. A
reasonable conjecture is that perhaps this difference can account for the ‘all events’ results depicted
in Figure 3. In order to examine this issue Figure 4 depicts scatter plots similar to Figure 3, but
for the top 500 firms by market capitalization (‘large’) separately from all other firms (‘small’).
Figure 4 shows that the initial sharp jump followed by lengthy drift in the opposite direction is true
18
both for large firms and for small firms. Small firms seem to exhibit greater leakage in advance
of publication, and after about three weeks the bulk of the reversal is over. Large firms exhibit
very little apparent leakage. However the reversal is much is much lengthier. It is not clear exactly
when it ends. Thus while the numerical magnitudes differ, the jump and then reversal pattern is not
limited to small firms. Indeed the momentum effect is, if anything, stronger for large firms than
for small firms.
VII. Business Cycle Effects
The typical event study does not distinguish news that take place during an expansion from
news that take place during a recession. Boyd, Hu and Jagganathan (2004) found this to be an
important distinction for the impact of news about unemployment. Accordingly Table 9 examines
whether this distinction is important for corporate news.20 It does turn out to be important.
On average news during an expansion is good new, whereas news during a recession is bad
news. In most cases the market response is stronger during a recession than it is during an expan-
sion. There are even instances in which the signs differ for a particular type of news according to
the phase of the cycle (listing change, new product). Some types of news are almost never observed
during recessions (e.g., dividend increases and stock splits).
Figure 5 illustrates how abnormal trading volumes and abnormal returns are typically digested
by the market in an expansion as compared to a recession. The results differ sharply. In an expan-
sion the jump and reversal pattern seen in Figure 3 is repeated. In an expansion after about three
weeks the bulk of the returns reversal seems to have finished. The abnormal trading volume is con-
tinuing to decline. In a contraction the jump and reversal pattern from Figure 3 is also repeated.
However, in this case the return reversal continues beyond the three weeks. Indeed, we do not find
evidence of this negative return momentum finishing even by 37 days.
This evidence suggests that there is much stronger (negative) momentum during a recession.
During an expansion the return momentum seems to be shorter lived. This distinction may be
of some interest to those who are trying to provide an economic rationale for the observed stock
market momentum documented by Jegadeesh and Titman (1993) and others.
20We use the well-known NBER definitions of the phase of the business cycle; see http://www.nber.org/cycles.html.
19
VIII. Conclusions
Fama (1998) has pointed out that published event studies may reflect the incentive structure
faced by researchers in addition to the underlying true data generating process. We have studied
whether this publication bias affects what we think we know about ordinary event studies. To do
this we have gathered three decades worth of corporate news stories from the Wall Street Journal.
Each news story has been classified according to topic using the Naıve Bayes algorithm. For each
topic that has at least 50 stories we have run an event study. Publication bias appears to be real and
it appears to have affected the profession’s impression of how news is impounded in prices.
The usual view (e.g., Ross 2005) is that on an event date there is a once-and-for-all jump in the
stock price to reflect the new information. Since news may leak in advance there may be a run-up
beforehand. After the news is reported there should not be any statistically significant cumulative
abnormal return. The usual view is only partly correct.
The typical response to a news story is a strong and prompt reaction followed by a gradual and
lengthy reversal. The reversal often exceeds the magnitude of the initial jump. In other words,
overreaction to news is common. The initial returns jump is typically accompanied by a temporary
jump in trading volume. The reversal is typically accompanied by gradually declining trading
volume. During business cycle expansions the process is largely complete after two or three weeks.
During recessions the reversal is much more prolonged and we are not able to identify a clear point
at which the process ends.
Our evidence can be interpreted in at least two sharply different ways. Perhaps the lengthy
post-news momentum reflects something in human psychology, as stressed in the literature on
behavioral finance.21 From this point of view our evidence could be interpreted as a refutation
of traditional market efficiency. Alternatively, our results can be interpreted as simply a rejection
of the ‘once-and-for-all jump hypothesis,’ rather than as a rejection of market efficiency. This is
because the evidence might also be a reflection of market microstructure considerations. It is well-
known that it often takes time to accumulate or unwind a large portfolio position. Perhaps what
we see in the data is a reflection of the actions of institutional traders. We do not have the kind of21In fact, behavioral finance scholars such as Daniel, Hirshleifer, and Subrahmanyam (1998) have focused on ex-
plaining short-run underreaction to public news events. We have found that the typical pattern is actually a short-runoverreaction. This difference reflects the fact that, to the best of our knowledge, the short-run overreaction to newsfollowed by longer term reversal has not previously been documented.
20
data needed to empirically distinguish between these two interpretations. This issue seems worthy
of future research.
Our empirical evidence may open the door to alternative approaches to study stock market
momentum. We find that the drift after news events differs depending on the stage of the business
cycle, and momentum is stronger for large firms than for small firms. Furthermore, some types of
news are followed by more momentum than other types of news. These differences pose questions
for further study.
Furthermore, our empirical evidence also provides a serious cautionary note for event studies.
Different event studies define the event window differently. Sometimes a few extra days before
an event are included, sometimes they are not. Sometimes a few extra days after the event are in-
cluded, sometimes they are not. Most event studies use daily data, but some use monthly. Different
empirical models and differently-weighted market indices are used to normalize for overall trends
in the market. It is usually thought that these alternatives do not make too much difference in
general. This general impression is false. The definition of the event window and the specification
of the market model can make a huge difference.
In modern finance it is widely believed that short-run event studies have shown that information
is fully impounded in returns very rapidly. Thus the returns over the next few days following an
event should not be predictable. In fact, predictable structure is typical. It appears to take many
days or even weeks, before the NYSE, AMEX and NASDAQ fully digest the news reported in the
Wall Street Journal.
21
References
Antweiler, W. and M. Z. Frank, 2004. “Is All That Talk Just Noise? The Information Content of
Internet Stock Message Boards,” Journal of Finance, 59, 1259-1294.
Asquith, P., M.B. Mikhail and A.S. Au, 2004. “Information Content of Equity Analyst Reports,”
Journal of Financial Economics, forthcoming.
Banz, 1981. “The Relation between Return and Market Value of Common Stocks,” Journal of
Financial Economics, 9, 3-18.
Begg, C.B. and J.A. Berlin, 1988. “Publication Bias: A Problem in Interpreting Medical Data,”
Journal of the Royal Statistical Society. Series A, 151, 419-463.
Bhattacharya, U., N. Galpin, R. Ray, and X. Yu, 2004. “The Role of the Media in the Internet IPO
Bubble,” Working paper, Indiana University.
Boyd, J.H., J. Hu, and R. Jagannathan, 2004. “The Stock Market Reaction to Unemployment News:
Why Bad News is Usually Good for Stocks,” Journal of Finance, forthcoming.
Brown, S. and J. Warner, 1985. “Using Daily Stock Returns: The Case of Event Studies,” Journal
of Financial Economics, 14, 3-31.
Campbell, J.Y., A.W. Lo, and A.C. MacKinlay, 1997. The Econometrics of Financial Markets,
Princeton University Press, Princeton N. J.
Chan, W.S., 2003. “Stock Price Reaction to News and No-News: Drift and Reversal After Head-
lines,” Journal of Financial Economics, 70, 2, 223-260.
Cutler, D.M., J.M. Poterba, and L.H. Summers, 1989. “What Moves Stock Prices?” Journal of
Portfolio Management, 15, 4-12.
Daniel, K., D. Hirshleifer and A. Subrahmanyam, 1998. “Investor Psychology and Security Market
Under- and Over-reactions,” Journal of Finance, 53, 1839-1886.
Das, S.R. and M.Y. Chen, 2004. “Yahoo! for Amazon: Sentiment Extraction from Small Talk on
the Web,” Working paper, Santa Clara University.
DeBondt, W.F.M. and Thaler, R., 1985. “Does the Stock Market Overreact?,” Journal of Finance
40, 793-805.
DeLong, J.B. and K. Lang, 1992. “Are All Economic Hypotheses False?” Journal of Political
Economy, 100, 6, 1257-72.
22
Fama, E.F., 1998. “Market Efficiency, Long-term Returns, and Behavioral Finance,” Journal of
Financial Economics, 49, 3, 283-306.
French, K., and R. Roll, 1986. “Stock Return Variance: The Arrival of Information and the Reac-
tion of Traders,” Journal of Financial Economics 17, 99-117.
Graham, J.R., J.L Koski, and U. Loewenstein, 2004. “Information Flow and Liquidity around
Anticipated and Unanticipated Dividend Announcements,” Journal of Business, forthcoming.
Griffin, J.M., J.H. Harris, and S. Topaloglu, 2003. “The Dynamics of Insitutional and Individual
Trading,” Journal of Finance, 56,6, 2285-2320.
Hastie, T., R. Tibishirani, and J. Friedman, 2001. The Elements of Statistical Learning: Data
Mining, Inference, and Prediction. Springer, New York.
Jegadeesh, N. and S. Titman, 1993. “Returns to Buying Winners and Selling Losers: Implications
for Stock Market Efficiency,” Journal of Finance, 48, 65-91.
Loughran, T. and J.R. Ritter, 2000, “Uniformly Least Powerful Tests of Market Efficiency,” Journal
of Financial Economics, 55, 3, 361-389.
McCallum, Andrew, 1996. Bow: A toolkit for statistical language modeling, text retrieval, classi-
fication and clustering, Working paper, School of Computer Science, Carnegie-Mellon Uni-
versity.
Manning, Christopher D., and Hinrich Schutze, 1999. Foundations of Statistical Natural Language
Processing, (MIT Press, Cambridge, MA).
Mitchell, M.L. and J.H. Mulherin, 1994. “The Impact of Public Information on the Stock Market,”
Journal of Finance, 49, 923-950.
Roll, R., 1988. “R2,” Journal of Finance, 43, 2, 541-566.
Ross, S.A., 2005. Neoclassical Finance, Princeton University Press, Princeton, N.J.
Schwert, G.W., 1981.“The Adjustement of Stock Prices to Information about Inflation,” Journal of
Finance, 36, 15-29.
Sterling T.D., W.L. Rosenbaum, and J.J. Weinkam, 1995. “Publication Decisions Revisited: The
Effect of the Outcome of Statistical Tests on the Decision to Publish and Vice Versa,” The
American Statistician, 49, 108-112.
23
A. Cumulative Abnormal Returns
The basic method of carrying out an event study is spelled out in chapter 4 of Campbell, Lo
and MacKinaly (1997). There are a number of variations of the basic event study method that have
been employed in different empirical papers. For clarity this appendix spells out the steps in our
calculations in some detail.
Consider stocki’s scaled log return in period (day)t
rit = 100 ln(1 + Rit) (6)
whereRit is the simple return. Unlike the simple return, the log returnrit exhibits symmetry and is
additively separable within subperiods. Assume that this security follows a simple market model
process
rit = αL,it + βL,it rmt + εit (7)
whereαL,it andβL,it are allowed to vary over time and are estimated through rolling regressions
overL days preceding dayt. Subscriptm refers to the value-weighted market index. The error
term εit is distributed with zero mean and varianceσ2ε . It is convenient to define the following
expressions:
µL,it =1
L
t−L∑s=t−1
ris (8)
µL,mt =1
L
t−L∑s=t−1
rms (9)
σ2L,it =1
L
[t−L∑
s=t−1
r2is
]− µ2L,it (10)
σ2L,mt =1
L
[t−L∑
s=t−1
r2ms
]− µ2L,mt (11)
σL,mit =1
L
[t−L∑
s=t−1
ri,srm,s
]− µL,it µL,mt (12)
The expressions for the means and covariances can be calculated very efficiently in a single pass
through a vector ofri,t by updating the sums in the above equations when moving fromt = 1 to
24
t = T . When moving thek-day window from{t−L−1, .., t−2} to {t−L, .., t−1}, this amounts
to adding the observation fort − 1 and dropping the observation fort − L − 1. This procedure
keeps the computational complexity proportional toT rather thanTL. Using this computation
technique, the OLS estimates of the market model can be obtained as
βL,it = σL,mit/σ2
L,mt (13)
αL,it = µL,it − βL,it µL,mt (14)
σ2L,εit= σ2L,it − βL,it σL,mit (15)
Define the abnormal return during the event windows ∈ {t + 1, ..., t + M}t as
ais = ris − (αL,it + βL,it rms) (16)
Conditioned on the market returns over the event window, the abnormal return has an error
s2is = s2L,εit
{1 +
1
L
[1 +
(rms − µL,mt)2
σ2L,mt
]}(17)
which consists of a part due to the future disturbance and a part that is due to the sampling error
of αL,it andβL,it. As can be seen, this second component is diminishing asL becomes large. For
largeL it holds thats2is ≈ s2L,εit. Now the standardized abnormal return (SAR) can be defined as
~ais =ais
sis
(18)
The cumulative abnormal return (CAR) over1 + M periods, including the event periodt andM
future periods, is defined by
cit =t+M∑s=t
ais (19)
and thus the standardized cumulative abnormal return (SCAR) is defined as
~cit =cit√∑t+Ms=t s2is
≈ cit
sL,εit
√M + 1
(20)
25
The distribution of SCAR is Studentt with L− 2 degrees of freedom, which implies a variance of
(L− 2)/(L− 4). For largeL, the distribution will be approximately standard normal.
Finally, to assess the statistical significance of the eventδk,it ∈ {0, 1} of typek for companyi
on dayt, one needs to aggregate standardized cumulative abnormal returns over all
Nk =∑
i
∑t
δk,it (21)
events of typek. The average standardized cumulative abnormal return (ASCAR) is given by
�~ck =1
Nk
∑i
∑t
~citδk,it (22)
The test statistic
Jk = �~ck
√N
L− 4
L− 2∼ N(0, 1) (23)
is distributed approximately standard normal. Ignoring theL fraction in the above expression
will bias the test statistic downward, and thus does not inflate statistical significance. In a rolling
regression where trading days may be varying over any fixed-length calendar-day window,L will
somewhat vary for each company.
Several important design decisions need to be made when calculatingJk:
• the length of the pre-event periodL;
• the use of log returnsrit instead of simple returnRit;
• the use of a value-weightedrmt market index versus the use of an equal-weighted market
index.
• the choice of the underlying market model: CAPM, Fama-French, GARCH, etc.
There is also a significant statistical caveat. In an event study where events occur frequently for
a given type, the assumption of independence may be violated. The assumption that there is no
correlation across the abnormal returns of different securities may be violated by the fact that event
windows may sometimes overlap.
26
TABLES AND FIGURES
Figure 1: News Stories in the Wall Street Journal by Year
New
s S
torie
s (1
,000
)
0
2
4
6
8
10
12
14
16
73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 00 01
IdentifiedUnidentified
Note: “Identified” is the number of news articles that were classified into at least one event category. “Unidenti-fied” is the number of news articles that could not be classified into event categories.
27
Figure 2: Newspaper Reports by Firm Size Decilese Description Ne Market Capitalizaton Deciles (%, top decile at left)ANY Any Classified Event . . . 186,081 54.3 14.5
Corporate GovernanceELC Elect CEO . . . . . . . . . . . . . . 3,794 41.0 13.7ELD Elect Director . . . . . . . . . . 5,630 30.4 15.6 10.8ELX Elect Executive . . . . . . . . . 14,321 41.1 13.8ELV Elect VP . . . . . . . . . . . . . . . . 3,129 61.8 14.2
Earnings ReportsEFD Earnings Forecast Down 3,395 46.9 15.3 10.1EFU Earnings Forecast Up. . . 960 53.6 14.1ERD Earnings Report Down . 10,341 40.7 18.0 10.4ERU Earnings Report Up . . . . 18,418 56.0 15.7
Financial IssuesDBI Debt Issue. . . . . . . . . . . . . . 528 69.1DBR Debt Repurchase . . . . . . . 101 54.5 16.8 11.9DVU Dividend Up . . . . . . . . . . . 422 45.5 16.6 11.8EQI Equity Issue . . . . . . . . . . . . 1,070 38.0 16.7 12.1EQR Equity Repurchase . . . . . 2,488 28.9 17.1 11.7 10.0LSC Listing Change . . . . . . . . . 1,163 13.8 10.4 13.2 14.0PRC Price Up . . . . . . . . . . . . . . . 1,395 52.6 24.7 14.2RTD Rating Down . . . . . . . . . . . 1,272 62.9 17.7SSP Stock Split . . . . . . . . . . . . . 115 33.0 14.8 18.3 13.0
General IssuesCOR Correction . . . . . . . . . . . . . 235 57.4 15.3MSC Miscellaneous . . . . . . . . . . 25,292 63.8 11.8OPI Opinion . . . . . . . . . . . . . . . . 3,561 71.9 12.2
Legal IssuesWAS Lawsuit Against Start . . 318 64.5 12.3WBE Lawsuit By End . . . . . . . . 4,741 59.8 14.6WBO Lawsuit By Ongoing . . . 249 93.6WBS Lawsuit By Start . . . . . . . 1,193 52.3 14.4LGE Legal Gov. End. . . . . . . . . 761 63.9 14.6LGS Legal Gov. Start . . . . . . . . 541 71.3 10.4
Operational IssuesCPU Capacity Up . . . . . . . . . . . 5,422 67.5 15.4CON Contract . . . . . . . . . . . . . . . 21,656 69.6 13.0LCA Labor Contract Accept . 541 84.3LAY Layoff . . . . . . . . . . . . . . . . . . 2,078 62.2 14.9PDN Product New . . . . . . . . . . . 5,607 79.2PDP Product New Possible . . 674 69.4SUA Supply Agreement . . . . . 1,399 73.8 11.1
Restructuring IssuesJVR J.V. Restructuring . . . . . . 127 49.6 13.4 13.4JVS Joint Venture Start . . . . . . 5,364 67.3 13.5MGA Merge Announce . . . . . . . 17,350 34.5 18.1 12.8MGZ Merge Complete . . . . . . . 945 35.0 18.3 13.9MGF Merge Fail . . . . . . . . . . . . . 652 57.8 11.0MGG Merge Fail Possible . . . . . 317 42.9 18.3MGP Merge Possible . . . . . . . . . 7,121 54.4 16.8UBA Unit Bought Announce . 1,022 48.0 15.1USA Unit Sold Announce. . . . 9,124 45.8 16.8USP Unit Sold Possible . . . . . . 488 58.0 13.7
Note: Ne is the total number of observations for event e. The deciles are based on the market capitalization of the company in which the reportappeared in the WSJ. The numbers are percentages of Ne; they are not shown if smaller than 10%.
28
Figure 3: ASCAR/ASCATV Scatter Plots for Selected Event Types
All Events (N=181615)
ASCARs
ASC
AT
Vs
−5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8 9 100
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
−2
−1
01
2345678910
15
2025
3035
Merge Announce (N=15970)
ASCARs
ASC
AT
Vs
−10 −8 −6 −4 −2 0 2 4 6 8 10 12 14 160
2
4
6
8
10
12
14
16
18
20
22
24
26
−2
−1
0
12
3
45678910
1520
2530
35
Earnings Forecast Down (N=3292)
ASCARs
ASC
AT
Vs
−50 −45 −40 −35 −30 −25 −20 −15 −10 −5 00
5
10
15
20
25
30
35
40
45
50
−2
−1
01
2
34
56 789 10
1520
2530
35
Stock Split (N=107)
ASCARs
ASC
AT
Vs
−20 −15 −10 −5 0 5 10 15 20 25 30 35 40 450
5
10
15
20
25
30
35
40
45
50
55
60
65
−2
−10
1234567
891015
20
2530
35
Note: The scatter plots show average standardized cumulative abnormal returns (ASCARs) and average standardized cumulativeabnormal trading volumes (ASCATVs), both scaled by 100 for easier readability. The event windows commence at two tradings daysbefore the event days (marked by –2 in the chart). Thus “0” indicates three days of cumulation as observed at the end of the event day.Cumulation of abnormal returns and volumes are continued for 37 trading days after the event day, for a total of 40 trading days. Thenumber of observations (events) for each event type is given in the title of each panel.
29
Figure 4: ASCAR/ASCATV Scatter Plots for Firm Size Groups
Large (Top 500) Firms (N=109981)
ASCARs
ASC
AT
Vs
−5 −4 −3 −2 −1 0 1 2 3 4 50
1
2
3
4
5
6
7
8
9
10
−2
−1
0123
45
67
89
10
15
20
25
30
35
Small Firms (N=71622)
ASCARs
ASC
AT
Vs
−5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8 9 100
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
−2
−1
0
12345678910
15
20
25
30
35
Note: Firm size has been defined by the market capitalization ranking in the month preceding the event. ‘Large’ refers to the top500 firms by market capitalization, and ‘small’ refers to all other firms. The scatter plots show average standardized cumulativeabnormal returns (ASCARs) and average standardized cumulative abnormal trading volumes (ASCATVs), both scaled by 100 foreasier readability. The event windows commence at two tradings days before the event days (marked by –2 in the chart). Thus“0” indicates three days of cumulation as observed at the end of the event day. Cumulation of abnormal returns and volumes arecontinued for 37 trading days after the event day, for a total of 40 trading days. The number of observations (events) for each eventtype is given in the title of each panel.
Figure 5: ASCAR/ASCATV Scatter Plots for Business Cycle Periods
Business Cycle Expansions (N=162155)
ASCARs
ASC
AT
Vs
−5 −4 −3 −2 −1 0 1 2 3 4 5 6 7 8 9 100
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
−2
−1
01
2345678910
15
20
25
3035
Business Cycle Contractions (N=19460)
ASCARs
ASC
AT
Vs
−20−19−18−17−16−15−14−13−12−11−10−9 −8 −7 −6 −5 −4 −3 −2 −1 00123456789
1011121314151617181920
−2
−1
01
2345678910
152025
3035
Note: The NBER definition of business cycles are used. See previous figure for additional information on the definition of ASCARsand ASCATVs.
30
Table 1: ASCARs Conditioned on the Inclusion of Days Surroundingthe Event Day, Covering All Classifiable Events
Base Day 5d 10d 20d 40d
Third trading day before event 2.0c 0.7b −0.9c −2.4c
Second trading day before event 1.7c 0.5 −1.1c −2.5c
First trading day before event 0.9c −0.1 −1.4c −2.7c
Event day −0.9c −1.5c −2.3c −3.4c
First trading day after event −1.6c −2.3c −2.7c −3.7c
Second trading day after event −1.7c −2.4c −2.8c −3.8c
Third trading day after event −1.7c −2.4c −2.8c −3.7c
Note: Average standardized cumulative abnormal returns (ASCARs) are scaled by100 for easier readability. The columns identify the length of the event window of #dtrading days, starting on and including the day stipulated in the column “Base Day.”Statistical significance at the 95%, 99%, and 99.9% levels of confidence is indicated bythe superscripts a, b, and c, respectively. The market model regressions are based on a120 calendar-day pre-event window and the equal-weighted market index.
Table 2: Median SCARs, Conditioned on the Inclusion of Days Sur-rounding the Event Day, Covering All Classifiable Events
Base Day 5d 10d 20d 40d
Third trading day before event 0.8c −0.0 −0.8c −1.6c
Second trading day before event 0.6a −0.4 −1.1c −1.9c
First trading day before event −0.4 −1.0c −1.4c −2.1c
Event day −2.7c −2.7c −2.7c −3.0c
First trading day after event −3.6c −3.7c −3.3c −3.3c
Second trading day after event −3.8c −3.4c −3.2c −3.4c
Third trading day after event −3.7c −3.4c −3.2c −3.5c
Note: ASCARs have been replaced by median SCARs, or MSCARs. The columns iden-tify the length of the post-event period of +#d trading days, starting on and includingthe day stipulated in the column “Base Day.” Statistical significance at the 95%, 99%,and 99.9% levels of confidence is indicated by the superscripts a, b, and c, respectively.The market model regressions are based on a 120 calendar-day pre-event window andthe equal-weighted market index.
31
Tabl
e3:
ASC
AR
sby
Even
tTyp
eW
ith
Cum
ulat
ion
Com
men
cing
Two
Trad
ing
Day
sBe
fore
Even
tDay
Even
tN
5d10
d20
d40
dEv
ent
N5d
10d
20d
40d
All
New
sEv
ents
....
...
185,
690
1.2c
0.0
−1.
4c−
2.7c
Lega
lIss
ues
Any
Cla
ssifi
edEv
ent.
..18
1,68
41.
2c0.
1−
1.3c
−2.
6cLa
wsu
itA
gain
stSt
art.
311
−8.
7−
10.4
0.9
9.8
Cor
pora
teG
over
nanc
eLa
wsu
itBy
End
....
...
4,64
5−
1.0
0.8
0.6
0.8
Com
pens
atio
n..
....
...
995.
115
.212
.1−
0.4
Law
suit
ByO
ngoi
ng..
245
−1.
0−
8.0−
13.7
a−
14.3
a
Elec
tCEO
....
....
....
..3,
701
3.8a
2.7
1.5
−0.
6La
wsu
itBy
Star
t...
...
1,16
2−
6.9a
−6.
2a−
2.0
−0.
3El
ectD
irec
tor
....
....
..5,
482
2.4
0.8
−1.
1−
3.8b
Lega
lGov
.End
....
...
741
−5.
7−
5.0
−7.
1−
6.7
Elec
tExe
cuti
ve..
....
...
14,0
100.
80.
2−
0.7
−2.
1aLe
galG
ov.O
ngoi
ng..
.80
−7.
43.
92.
74.
7El
ectV
P..
....
....
....
..3,
087
−3.
4−
4.7b
−9.
6c−
10.7
cLe
galG
ov.S
tart
....
...
522
−6.
7−
2.8
−0.
0−
2.5
Earn
ings
Rep
orts
Ope
rati
onal
Issu
esEa
rnin
gsFo
reca
stD
own
3,33
3−
41.8
c−
33.0
c−
22.5
c−
10.3
cC
apac
ity
Up
....
....
..5,
338
−0.
4−
1.2
−1.
2−
2.5
Earn
ings
Fore
cast
Up
...
945
5.3
3.2
−5.
9−
5.0
Con
trac
t...
....
....
...
21,3
960.
80.
3−
1.8a
−2.
6c
Earn
ings
Rep
ortD
own
.10
,101
−14
.6c−
9.7c
−4.
7c1.
5La
bor
Con
trac
tAcc
ept
539
1.3
−0.
43.
76.
4Ea
rnin
gsR
epor
tUp
....
18,0
884.
8c4.
4c2.
8c0.
9La
yoff
....
....
....
....
2,03
2−
4.0
−1.
00.
54.
1Fi
nanc
ialI
ssue
sPr
oduc
tNew
....
....
.5,
539
4.6c
2.8a
1.6
0.2
Deb
tIss
ue..
....
....
....
524
0.7
−1.
0−
3.9
0.1
Prod
uctN
ewPo
ssib
le.
669
−0.
4−
0.8
−5.
7−
12.7
b
Deb
tRep
urch
ase
....
...
98−
7.4
−8.
6−
1.5
1.4
Supp
lyA
gree
men
t...
.1,
373
6.3a
2.5
0.9
−4.
2D
ivid
end
Dow
n..
....
..80
−14
.5−
7.2
−3.
6−
0.1
Res
truc
turi
ngIs
sues
Div
iden
dU
p..
....
....
.41
824
.6c
16.7
c8.
52.
2J.V
.Res
truc
turi
ng..
...
126
−4.
0−
13.6
−21
.2a−
24.1
b
Equi
tyIs
sue
....
....
....
1,00
5−
7.1a
−8.
9b−
8.6b−
10.7
cJo
intV
entu
reSt
art
....
5,26
65.
1c4.
4b2.
50.
7Eq
uity
Rep
urch
ase
....
.2,
450
32.3
c28
.2c
21.3
c16
.2c
Mer
geA
nnou
nce
....
.16
,623
11.0
c5.
7c0.
3−
6.2c
List
ing
Cha
nge
....
....
.1,
126
1.5
−9.
6b−
12.7
c−
16.5
cM
erge
Com
plet
e..
....
912
3.5
2.8
3.7
1.2
Pric
eU
p..
....
....
....
.1,
373
3.5
−1.
5−
5.8a−
12.3
cM
erge
Fail
....
....
....
640
−3.
4−
8.5a−
15.3
c−
27.1
c
Rat
ing
Dow
n..
....
....
.1,
236
−7.
1a−
3.7
0.7
6.9a
Mer
geFa
ilPo
ssib
le..
.30
8−
2.9
−3.
4−
12.8
a−
15.6
b
Stoc
kSp
lit..
....
....
...
113
37.0
c23
.8a
4.0−
18.1
Mer
gePo
ssib
le..
....
..6,
810
0.5
−3.
9b−
5.9c−
11.1
c
Gen
eral
Issu
esU
nitB
ough
tAnn
ounc
e1,
004
2.2
−0.
60.
12.
0C
orre
ctio
n..
....
....
...
228
9.0
5.8
8.5
−0.
6U
nitS
old
Ann
ounc
e..
8,91
06.
9c4.
0c1.
1−
1.8
Mis
cella
neou
s...
....
...
24,5
81−
0.3
−1.
7b−
2.7c
−3.
0cU
nitS
old
Com
plet
e..
.58
1.5−
18.8
−18
.1−
4.7
Nam
eC
hang
e...
....
...
6916
.4−
6.5−
12.4
−15
.4U
nitS
old
Poss
ible
....
.47
617
.0c
15.8
c7.
6−
1.2
Opi
nion
....
....
....
....
3,45
83.
4a−
0.1
−4.
8b−
7.6c
Not
e:A
vera
gest
anda
rdiz
edcu
mul
ativ
eab
norm
alre
turn
s(A
SCA
Rs)
are
scal
edby
100
for
easi
erre
adab
ility
.Th
eco
lum
nsid
enti
fyth
ele
ngth
ofth
eev
ent
win
dow
mea
sure
din
trad
ing
days
,sta
rtin
gtw
otr
adin
gda
ysbe
fore
the
even
tday
.Ev
entt
ypes
wit
hfe
wer
than
50ob
serv
atio
nsar
esu
ppre
ssed
.St
atis
tica
lsig
nific
ance
atth
e95
%,9
9%,a
nd99
.9%
leve
lsof
confi
denc
eis
indi
cate
dby
the
supe
rscr
ipts
a,b
,and
c,r
espe
ctiv
ely.
The
mar
ketm
odel
regr
essi
ons
are
base
don
a12
0ca
lend
ar-d
aypr
e-ev
entw
indo
wan
dth
eeq
ual-
wei
ghte
dm
arke
tind
ex.
32
Tabl
e4:
10-D
ayA
SCA
Rs
byEv
entT
ype
and
Con
diti
oned
onD
iffer
entP
rior
Even
tsD
urin
gth
eLa
stTw
oC
alen
dar
Wee
ksEv
ent
Nal
low
not
her
none
Even
tN
all
own
othe
rno
neC
orpo
rate
Gov
erna
nce
Law
suit
Aga
inst
Star
t.31
1−
10.4
−10
.8−
9.9
Com
pens
atio
n..
....
...
9915
.216
.513
.1La
wsu
itBy
End
....
...
4,64
50.
84.
1b3.
2a−
1.3
Elec
tCEO
....
....
....
..3,
701
2.7
3.3a
3.6a
2.3
Law
suit
ByO
ngoi
ng..
245
−8.
0−
20.6
b−
12.6
7.3
Elec
tDir
ecto
r..
....
....
5,48
20.
814
.5c
2.2
0.3
Law
suit
BySt
art.
....
.1,
162
−6.
2a6.
0a−
4.9
−7.
6a
Elec
tExe
cuti
ve..
....
...
14,0
100.
24.
1c1.
7−
0.6
Lega
lGov
.End
....
...
741
−5.
0−
15.3
c−
4.8
−5.
3El
ectV
P..
....
....
....
..3,
087
−4.
7b−
10.7
c−
4.3a
−5.
1bLe
galG
ov.O
ngoi
ng..
.80
3.9
2.7
2.9
Earn
ings
Rep
orts
Lega
lGov
.Sta
rt..
....
.52
2−
2.8
−0.
51.
5−
9.8a
Earn
ings
Fore
cast
Dow
n3,
333−
33.0
c−
30.9
c−
25.3
c−
37.5
cO
pera
tion
alIs
sues
Earn
ings
Fore
cast
Up
...
945
3.2
−2.
66.
9aC
apac
ity
Up
....
....
..5,
338
−1.
2−
2.4
−1.
1−
1.4
Earn
ings
Rep
ortD
own
.10
,101
−9.
7c−
9.1c
−7.
6c−
10.6
cC
ontr
act.
....
....
....
.21
,396
0.3
−0.
4−
0.7
2.5c
Earn
ings
Rep
ortU
p..
..18
,088
4.4c
2.7c
3.3c
4.9c
Labo
rC
ontr
actA
ccep
t53
9−
0.4
3.3
−1.
22.
0Fi
nanc
ialI
ssue
sLa
yoff
....
....
....
....
2,03
2−
1.0
1.3
−0.
5−
1.6
Deb
tIss
ue..
....
....
....
524
−1.
04.
21.
9−
6.3
Prod
uctN
ew..
....
...
5,53
92.
8a2.
12.
63.
0a
Deb
tRep
urch
ase
....
...
98−
8.6
2.6−
14.7
Prod
uctN
ewPo
ssib
le.
669
−0.
8−
14.7
c4.
4−
4.7
Div
iden
dD
own
....
....
80−
7.2
−5.
3Su
pply
Agr
eem
ent.
...
1,37
32.
57.
3b3.
80.
7D
ivid
end
Up
....
....
...
418
16.7
c13
.5b
18.0
cR
estr
uctu
ring
Issu
esEq
uity
Issu
e..
....
....
..1,
005
−8.
9b−
24.3
c−
5.6−
10.7
cJ.V
.Res
truc
turi
ng..
...
126−
13.6
−30
.6c−
0.4
Equi
tyR
epur
chas
e..
...
2,45
028
.2c
21.5
c18
.4c
31.1
cJo
intV
entu
reSt
art
....
5,26
64.
4b5.
9c4.
2b4.
7c
List
ing
Cha
nge
....
....
.1,
126
−9.
6b−
11.8
c−
12.1
c−
8.7b
Mer
geA
nnou
nce
....
.16
,623
5.7c
−0.
60.
57.
7c
Pric
eU
p..
....
....
....
.1,
373
−1.
5−
11.7
c−
0.5
−2.
2M
erge
Com
plet
e..
....
912
2.8
−9.
9b7.
1a
Rat
ing
Dow
n..
....
....
.1,
236
−3.
7−
1.7
−0.
8−
7.4a
Mer
geFa
il..
....
....
..64
0−
8.5a−
20.0
c−
13.6
c1.
6St
ock
Split
....
....
....
.11
323
.8a
23.0
aM
erge
Fail
Poss
ible
...
308
−3.
4−
13.5
a6.
5G
ener
alIs
sues
Mer
gePo
ssib
le..
....
..6,
810
−3.
9b−
7.8c
−7.
7c2.
4a
Cor
rect
ion
....
....
....
.22
85.
85.
75.
9U
nitB
ough
tAnn
ounc
e1,
004
−0.
6−
2.9
−7.
7a2.
5M
isce
llane
ous.
....
....
.24
,581
−1.
7b−
4.1c
−3.
0c0.
1U
nitS
old
Ann
ounc
e..
8,91
04.
0c−
7.4c
−0.
76.
4c
Nam
eC
hang
e...
....
...
69−
6.5
−4.
5U
nitS
old
Com
plet
e..
.58
−18
.8O
pini
on..
....
....
....
..3,
458
−0.
1−
5.4b
−1.
82.
9U
nitS
old
Poss
ible
....
.47
615
.8c
16.9
c14
.7b
Lega
lIss
ues
Not
e:A
vera
gest
anda
rdiz
edcu
mul
ativ
eab
norm
alre
turn
s(A
SCA
Rs)
are
scal
edby
100
for
easi
erre
adab
ility
.Th
eco
lum
nsid
enti
fyth
ele
ngth
ofth
eev
entw
indo
wm
easu
red
intr
adin
gda
ys,s
tart
ing
two
trad
ing
days
befo
reth
eev
entd
ay.
The
leng
thof
the
even
twin
dow
is10
trad
ing
days
incl
udin
gan
dfo
llow
ing
the
even
tday
.The
colu
mns
refe
rto
the
cond
itio
ning
even
ts:
colu
mn
‘all’
repo
rts
resu
lts
for
the
unco
ndit
ione
dca
se;c
olum
n‘o
wn’
cond
itio
nsth
eA
SCA
Rs
ona
prio
rev
ento
fth
esa
me
type
wit
hin
the
last
14ca
lend
arda
ys;
colu
mn
‘any
’co
ndit
ions
the
ASC
AR
Son
apr
ior
even
tof
anot
her
type
wit
hin
the
last
14ca
lend
arda
ysif
ther
eis
no‘o
wn’
prio
rev
ent;
and
colu
mn
‘non
e’co
ndit
ions
the
ASC
AR
son
nopr
ior
even
tw
hats
oeve
rdu
ring
the
last
14da
ys.
Even
tty
pes
wit
hfe
wer
than
50ob
serv
atio
nsar
esu
ppre
ssed
.St
atis
tica
lsig
nific
ance
atth
e95
%,9
9%,a
nd99
.9%
leve
lsof
confi
denc
eis
indi
cate
dby
the
supe
rscr
ipts
a,b
,and
c,r
espe
ctiv
ely.
The
mar
ket
mod
elre
gres
sion
sar
eba
sed
ona
120
cale
ndar
-day
pre-
even
tw
indo
wan
dth
eeq
ual-
wei
ghte
dm
arke
tind
ex.
33
Tabl
e5:
ASC
AR
sby
Even
tTyp
eW
ith
Cum
ulat
ion
Com
men
cing
onth
eTr
adin
gD
ayFo
llow
ing
the
Even
tDay
Even
tN
5d10
d20
d40
dEv
ent
N5d
10d
20d
40d
All
New
sEv
ents
....
...
185,
774
−1.
7c−
2.5c
−2.
8c−
3.8c
Lega
lIss
ues
Any
Cla
ssifi
edEv
ent.
..18
1,76
7−
1.6c
−2.
4c−
2.8c
−3.
7cLa
wsu
itA
gain
stSt
art.
312
−3.
63.
37.
914
.6a
Cor
pora
teG
over
nanc
eLa
wsu
itBy
End
....
...
4,64
6−
0.4
−0.
10.
20.
8C
ompe
nsat
ion
....
....
.99
4.9
4.5
5.1
−5.
4La
wsu
itBy
Ong
oing
..24
6−
12.9
a−
12.1
−17
.3b−
14.5
a
Elec
tCEO
....
....
....
..3,
698
0.1
0.2
−1.
1−
2.4
Law
suit
BySt
art.
....
.1,
163
−4.
1−
0.3
3.8
1.9
Elec
tDir
ecto
r..
....
....
5,49
8−
1.6
−3.
0a−
4.2b
−6.
0cLe
galG
ov.E
nd..
....
.74
2−
1.7
−1.
5−
4.5
−4.
6El
ectE
xecu
tive
....
....
.14
,029
−0.
4−
0.8
−1.
3−
2.7b
Lega
lGov
.Ong
oing
...
790.
46.
016
.43.
1El
ectV
P..
....
....
....
..3,
091
−2.
3−
5.7b
−8.
7c−
10.4
cLe
galG
ov.S
tart
....
...
523
−0.
21.
90.
80.
1Ea
rnin
gsR
epor
tsO
pera
tion
alIs
sues
Earn
ings
Fore
cast
Dow
n3,
333
4.1a
4.7b
9.7c
13.8
cC
apac
ity
Up
....
....
..5,
344
−0.
7−
1.0
−1.
5−
3.0a
Earn
ings
Fore
cast
Up
...
951
−1.
3−
4.9
−8.
2a−
7.5a
Con
trac
t...
....
....
...
21,4
14−
0.4
−2.
5c−
2.4c
−3.
1c
Earn
ings
Rep
ortD
own
.10
,111
3.3b
3.9c
6.5c
9.6c
Labo
rC
ontr
actA
ccep
t54
0−
7.9
−3.
32.
33.
3Ea
rnin
gsR
epor
tUp
....
18,1
020.
30.
60.
3−
1.1
Layo
ff..
....
....
....
..2,
028
0.1
1.9
2.4
6.8b
Fina
ncia
lIss
ues
Prod
uctN
ew..
....
...
5,53
8−
1.4
−2.
7a−
1.3
−1.
3D
ebtI
ssue
....
....
....
..52
3−
0.5
−2.
8−
1.9
0.4
Prod
uctN
ewPo
ssib
le.
669
−4.
9−
8.9a−
11.9
b−
15.5
c
Deb
tRep
urch
ase
....
...
98−
9.9
−1.
70.
6−
0.9
Supp
lyA
gree
men
t...
.1,
373
−0.
7−
2.0
−1.
9−
7.9b
Div
iden
dD
own
....
....
815.
14.
75.
510
.9R
estr
uctu
ring
Issu
esD
ivid
end
Up
....
....
...
418
1.6
0.3
−1.
0−
5.8
J.V.R
estr
uctu
ring
....
.12
6−
17.1
−23
.3b−
21.2
a−
20.5
a
Equi
tyIs
sue
....
....
....
1,00
6−
4.7
−7.
5a−
7.6a
−9.
2bJo
intV
entu
reSt
art
....
5,27
00.
4−
0.1
−0.
9−
1.8
Equi
tyR
epur
chas
e..
...
2,45
37.
8c7.
8c5.
3b5.
4bM
erge
Ann
ounc
e..
...
16,6
20−
6.0c
−8.
7c−
10.3
c−
14.6
c
List
ing
Cha
nge
....
....
.1,
124−
19.5
c−
18.6
c−
18.9
c−
21.3
cM
erge
Com
plet
e..
....
918
1.6
0.1
0.0
−0.
2Pr
ice
Up
....
....
....
...
1,37
5−
4.8
−6.
3a−
7.9b−
13.4
cM
erge
Fail
....
....
....
639
−9.
4a−
13.3
c−
20.1
c−
31.6
c
Rat
ing
Dow
n..
....
....
.1,
237
3.7
6.3a
8.4b
11.5
cM
erge
Fail
Poss
ible
...
307
1.3
−7.
3−
11.5
a−
14.5
a
Stoc
kSp
lit..
....
....
...
113
4.2
−5.
6−
23.6
a−
24.5
aM
erge
Poss
ible
....
....
6,79
2−
7.7c
−8.
8c−
11.1
c−
15.2
c
Gen
eral
Issu
esU
nitB
ough
tAnn
ounc
e1,
002
−3.
6−
3.9
−1.
7−
0.2
Cor
rect
ion
....
....
....
.22
73.
06.
84.
9−
0.9
Uni
tSol
dA
nnou
nce
..8,
916
−2.
0−
2.9b
−3.
9c−
5.0c
Mis
cella
neou
s...
....
...
24,5
80−
3.1c
−3.
4c−
3.8c
−3.
8cU
nitS
old
Com
plet
e..
.57
−30
.2a−
23.1
−22
.8−
10.4
Nam
eC
hang
e...
....
...
68−
35.8
b−
35.2
b−
36.4
b−
26.0
aU
nitS
old
Poss
ible
....
.47
60.
4−
1.6
−4.
4−
9.0
Opi
nion
....
....
....
....
3,45
8−
5.4b
−6.
5c−
8.9c
−9.
8c
Not
e:A
vera
gest
anda
rdiz
edcu
mul
ativ
eab
norm
alre
turn
s(A
SCA
Rs)
are
scal
edby
100
for
easi
erre
adab
ility
.Th
eco
lum
nsid
enti
fyth
ele
ngth
ofth
eev
ent
win
dow
mea
sure
din
trad
ing
days
,sta
rtin
gon
the
first
trad
ing
day
afte
rth
eev
ent
day.
Even
tty
pes
wit
hfe
wer
than
50ob
serv
atio
nsar
esu
ppre
ssed
.St
atis
tica
lsi
gnifi
canc
eat
the
95%
,99%
,and
99.9
%le
vels
ofco
nfide
nce
isin
dica
ted
byth
esu
pers
crip
tsa
,b,a
ndc,r
espe
ctiv
ely.
The
mar
ket
mod
elre
gres
sion
sar
eba
sed
ona
120
cale
ndar
-day
pre-
even
tw
indo
wan
dth
eeq
ual-
wei
ghte
dm
arke
tind
ex.
34
Tabl
e6:
ASC
AR
sby
Even
tTy
peW
ith
Val
ue-W
eigh
ted
Mar
ket
Inde
xU
sed
Inst
ead
ofth
eEq
ual-
Wei
ghte
dM
arke
tIn
dex
and
Wit
hC
umul
atio
nC
omm
enci
ngTw
oTr
adin
gD
ays
Befo
reEv
entD
ayEv
ent
N5d
10d
20d
40d
Even
tN
5d10
d20
d40
d
All
New
sEv
ents
....
...
185,
635
0.8c
−0.
8c−
2.1c
−3.
5cLe
galI
ssue
sA
nyC
lass
ified
Even
t...
181,
629
0.8c
−0.
7b−
2.0c
−3.
4cLa
wsu
itA
gain
stSt
art.
311−
11.0
−14
.1a−
5.0
2.5
Cor
pora
teG
over
nanc
eLa
wsu
itBy
End
....
...
4,64
4−
2.0
0.0
−0.
3−
0.4
Com
pens
atio
n..
....
...
992.
511
.511
.2−
2.1
Law
suit
ByO
ngoi
ng..
245−
13.5
a−
15.2
a−
14.2
a−
9.8
Elec
tCEO
....
....
....
..3,
700
5.5c
2.8
1.5
−1.
0La
wsu
itBy
Star
t...
...
1,16
2−
9.2b
−5.
1−
1.3
0.9
Elec
tDir
ecto
r..
....
....
5,48
00.
9−
1.8
−2.
3−
5.4c
Lega
lGov
.End
....
...
740
−3.
0−
4.1
−5.
7−
5.6
Elec
tExe
cuti
ve..
....
...
14,0
071.
50.
5−
0.0
−1.
4Le
galG
ov.O
ngoi
ng..
.80
−9.
76.
0−
3.4
−1.
4El
ectV
P..
....
....
....
..3,
087
−1.
8−
2.9
−4.
0a−
3.1
Lega
lGov
.Sta
rt..
....
.52
2−
4.8
0.2
9.5a
5.0
Earn
ings
Rep
orts
Ope
rati
onal
Issu
esEa
rnin
gsFo
reca
stD
own
3,33
3−
46.1
c−
35.1
c−
20.3
c−
5.6b
Cap
acit
yU
p..
....
....
5,33
8−
0.2
−0.
4−
0.1
0.4
Earn
ings
Fore
cast
Up
...
945
6.6a
3.4
−5.
1−
3.8
Con
trac
t...
....
....
...
21,3
941.
3−
0.8
−2.
1b−
2.6c
Earn
ings
Rep
ortD
own
.10
,099
−14
.0c−
8.3c
−2.
9b2.
3aLa
bor
Con
trac
tAcc
ept
539
−3.
2−
5.4
−4.
4−
1.3
Earn
ings
Rep
ortU
p..
..18
,085
4.3c
3.0c
0.7
−3.
8cLa
yoff
....
....
....
....
2,03
0−
3.6
1.2
3.4
11.2
c
Fina
ncia
lIss
ues
Prod
uctN
ew..
....
...
5,53
84.
8c1.
70.
1−
1.5
Deb
tIss
ue..
....
....
....
524
−2.
1−
5.5
−6.
6−
5.9
Prod
uctN
ewPo
ssib
le.
669
−0.
2−
2.1
−5.
3−
8.5a
Deb
tRep
urch
ase
....
...
98−
13.3
−10
.1−
11.1
−7.
3Su
pply
Agr
eem
ent.
...
1,37
38.
7b3.
50.
2−
4.8
Div
iden
dD
own
....
....
80−
6.0
−4.
46.
617
.4R
estr
uctu
ring
Issu
esD
ivid
end
Up
....
....
...
418
22.9
c18
.3c
10.6
a3.
8J.V
.Res
truc
turi
ng..
...
126−
11.1
−15
.9−
26.3
b−
26.1
b
Equi
tyIs
sue
....
....
....
1,00
5−
9.9b−
11.6
c−
14.9
c−
19.3
cJo
intV
entu
reSt
art
....
5,26
65.
3c3.
8b1.
3−
0.5
Equi
tyR
epur
chas
e..
...
2,45
032
.7c
26.7
c22
.0c
18.4
cM
erge
Ann
ounc
e..
...
16,6
129.
2c3.
2c−
3.1c−
10.1
c
List
ing
Cha
nge
....
....
.1,
126
−3.
8−
13.7
c−
15.7
c−
18.0
cM
erge
Com
plet
e..
....
912
8.4a
4.3
3.6
−0.
0Pr
ice
Up
....
....
....
...
1,37
32.
5−
4.7
−5.
8a−
6.6a
Mer
geFa
il..
....
....
..64
0−
5.9
−9.
8a−
15.1
c−
29.3
c
Rat
ing
Dow
n..
....
....
.1,
236−
10.2
c−
1.7
3.2
11.7
cM
erge
Fail
Poss
ible
...
308
0.3
−4.
1−
15.7
b−
19.0
c
Stoc
kSp
lit..
....
....
...
113
36.3
c16
.1−
4.1−
25.4
bM
erge
Poss
ible
....
....
6,80
4−
1.9
−6.
7c−
9.8c−
15.8
c
Gen
eral
Issu
esU
nitB
ough
tAnn
ounc
e1,
004
4.2
0.4
0.6
0.1
Cor
rect
ion
....
....
....
.22
89.
10.
25.
70.
8U
nitS
old
Ann
ounc
e..
8,90
58.
4c4.
7c0.
9−
2.3a
Mis
cella
neou
s...
....
...
24,5
66−
1.7b
−2.
6c−
3.6c
−3.
7cU
nitS
old
Com
plet
e..
.58
19.3
3.5
−3.
117
.7N
ame
Cha
nge.
....
....
.69
−12
.7−
38.9
b−
31.7
b−
23.1
Uni
tSol
dPo
ssib
le..
...
476
18.2
c16
.1c
5.6
−3.
8O
pini
on..
....
....
....
..3,
458
1.9
−1.
2−
6.4c
−7.
8c
Not
e:A
vera
gest
anda
rdiz
edcu
mul
ativ
eab
norm
alre
turn
s(A
SCA
Rs)
are
scal
edby
100
for
easi
erre
adab
ility
.Th
eco
lum
nsid
enti
fyth
ele
ngth
ofth
eev
ent
win
dow
mea
sure
din
trad
ing
days
,sta
rtin
gtw
otr
adin
gda
ysbe
fore
the
even
tday
.Ev
entt
ypes
wit
hfe
wer
than
50ob
serv
atio
nsar
esu
ppre
ssed
.St
atis
tica
lsig
nific
ance
atth
e95
%,9
9%,a
nd99
.9%
leve
lsof
confi
denc
eis
indi
cate
dby
the
supe
rscr
ipts
a,b
,and
c,r
espe
ctiv
ely.
The
mar
ketm
odel
regr
essi
ons
are
base
don
a12
0ca
lend
ar-d
aypr
e-ev
entw
indo
wan
dth
eva
lue-
wei
ghte
dm
arke
tind
ex.
35
Table 7: ASCARs, Cumulation Commencing on the Trading Day Following the Event Day, Disag-gregated by Deciles of the Return on the Event Day
News Events No Events (5% Sample)Decile / Return Range +5d +10d +20d +40d +5d +10d +20d +40d shareAll Events −1.6c −2.3c −2.7c −3.7c −1.1c −1.0c −0.9c −0.9c 100
1. −∞ −3.4 6.3c 4.4c 3.0c 1.7a 10.0c 7.8c 5.5c 3.4c 11.32. −3.4 −1.8 3.6c 2.5b 2.0b 0.4 1.3c 1.0c 0.3 −0.9c 9.43. −1.8 −1.0 0.3 −0.1 −0.9 −3.2c −1.9c −1.8c −1.9c −2.7c 8.24. −1.0 −0.4 −1.6a −1.3 −2.2b −3.5c −4.8c −4.1c −4.5c −5.8c 6.95. −0.4 0.0 −2.1b −2.3b −2.7c −3.0c −1.8c −1.1c 0.1 1.5c 14.96. 0.0 +0.4 −0.7 −2.1b −3.1c −4.1c −1.1c −0.1 0.8c 2.2c 15.07. +0.4 +1.0 −2.8c −4.7c −5.0c −6.2c −1.3c −2.3c −3.4c −4.9c 7.08. +1.0 +1.9 −4.0c −4.2c −3.8c −4.9c −2.6c −2.5c −2.4c −2.4c 8.09. +1.9 +3.6 −6.5c −7.2c −6.7c −7.8c −4.1c −3.9c −3.6c −3.5c 8.7
10. +3.6 +∞ −8.3c −7.9c −7.4c −6.8c −7.2c −5.9c −4.3c −2.4c 10.6Observations 171,360 1,410,328
Note: Average standardized cumulative abnormal returns (ASCARs) are scaled by 100 for easier readability. The columnsidentify the length of the event window measured in trading days, starting on the first trading day after the event day. Decileswere formed by ranking the nominal returns on the event day. The no-news columns are based on randomly sampling 5% ofall companies and trading days, excluding days identified with WSJ news events. The column ‘share’ indicates the percentagedistribution of the returns of the no-news events when the decile boundaries of the news events are applied to them. Thiscolumn highlights differences between the distribution of news-day returns relative to the population of returns. Statisticalsignificance at the 95%, 99%, and 99.9% levels of confidence is indicated by the superscripts a, b, and c, respectively. Themarket model regressions are based on a 120 calendar-day pre-event window and the equal-weighted market index.
36
Tabl
e8:
Ave
rage
Stan
dard
ized
Cum
ulat
ive
Abn
orm
alTr
adin
gVo
lum
eby
Even
tTyp
eW
ith
Cum
ulat
ion
Com
men
cing
Two
Trad
ing
Day
sBe
fore
the
Even
tDay
Even
tN
5d10
d20
d40
dEv
ent
N5d
10d
20d
40d
All
New
sEv
ents
....
...
180,
919
9.5c
8.0c
6.0c
3.1c
Lega
lIss
ues
Any
Cla
ssifi
edEv
ent.
..17
6,93
79.
3c7.
9c5.
8c2.
9cLa
wsu
itA
gain
stSt
art.
303
2.3
−0.
4−
5.5
−8.
8C
orpo
rate
Gov
erna
nce
Law
suit
ByEn
d..
....
.4,
538
4.1b
3.0a
1.1
−1.
9C
ompe
nsat
ion
....
....
.98
−3.
1−
4.7
−0.
9−
0.5
Law
suit
ByO
ngoi
ng..
245
24.6
c23
.7c
27.4
c14
.2a
Elec
tCEO
....
....
....
..3,
580
8.7c
6.9c
4.5b
−0.
3La
wsu
itBy
Star
t...
...
1,12
32.
31.
2−
2.8
−6.
5a
Elec
tDir
ecto
r..
....
....
5,17
72.
21.
40.
1−
1.1
Lega
lGov
.End
....
...
727
4.9
5.2
2.5
−0.
5El
ectE
xecu
tive
....
....
.13
,495
3.7c
2.5b
1.0
−1.
5Le
galG
ov.O
ngoi
ng..
.77
0.5
4.3
10.8
18.8
Elec
tVP
....
....
....
....
3,01
11.
81.
10.
4−
1.2
Lega
lGov
.Sta
rt..
....
.50
02.
60.
10.
3−
2.9
Earn
ings
Rep
orts
Ope
rati
onal
Issu
esEa
rnin
gsFo
reca
stD
own
3,28
529
.5c
26.4
c21
.4c
15.4
cC
apac
ity
Up
....
....
..5,
172
4.8c
4.3b
3.7b
3.2a
Earn
ings
Fore
cast
Up
...
897
9.7b
6.1
3.0
−0.
1C
ontr
act.
....
....
....
.21
,087
1.7a
0.7
−1.
3−
4.5c
Earn
ings
Rep
ortD
own
.9,
911
14.6
c10
.3c
5.4c
0.2
Labo
rC
ontr
actA
ccep
t52
67.
29.
3a11
.4b
11.6
b
Earn
ings
Rep
ortU
p..
..17
,779
13.0
c9.
3c5.
8c3.
2cLa
yoff
....
....
....
....
1,99
56.
0b5.
1a2.
6−
1.9
Fina
ncia
lIss
ues
Prod
uctN
ew..
....
...
5,51
16.
1c5.
0c2.
9a1.
3D
ebtI
ssue
....
....
....
..51
25.
56.
78.
112
.4b
Prod
uctN
ewPo
ssib
le.
667
13.8
c11
.6b
8.5a
6.6
Deb
tRep
urch
ase
....
...
981.
21.
0−
0.4
−8.
6Su
pply
Agr
eem
ent.
...
1,35
47.
6b6.
8a5.
6a5.
2D
ivid
end
Dow
n..
....
..73
22.4
15.3
9.1
1.8
Res
truc
turi
ngIs
sues
Div
iden
dU
p..
....
....
.39
49.
25.
91.
3−
0.5
J.V.R
estr
uctu
ring
....
.11
98.
09.
67.
13.
7Eq
uity
Issu
e..
....
....
..98
018
.7c
18.6
c20
.4c
22.1
cJo
intV
entu
reSt
art
....
5,17
74.
0b3.
8b3.
7b4.
2b
Equi
tyR
epur
chas
e..
...
2,29
210
.1c
8.8c
6.2b
1.9
Mer
geA
nnou
nce
....
.15
,875
19.2
c18
.0c
16.7
c14
.1c
List
ing
Cha
nge
....
....
.1,
041
3.3
3.1
1.1
−1.
8M
erge
Com
plet
e..
....
879
6.1
6.7a
8.5a
7.0a
Pric
eU
p..
....
....
....
.1,
348
4.7
4.1
5.7a
8.4b
Mer
geFa
il..
....
....
..62
218
.9c
20.0
c17
.3c
10.1
a
Rat
ing
Dow
n..
....
....
.1,
222
7.3a
6.8a
5.4
2.3
Mer
geFa
ilPo
ssib
le..
.29
25.
62.
30.
4−
3.9
Stoc
kSp
lit..
....
....
...
107
18.1
18.9
19.3
a41
.3c
Mer
gePo
ssib
le..
....
..6,
620
21.1
c21
.3c
18.5
c11
.0c
Gen
eral
Issu
esU
nitB
ough
tAnn
ounc
e95
66.
5a6.
15.
72.
9C
orre
ctio
n..
....
....
...
221
−0.
3−
0.7
0.3
−5.
4U
nitS
old
Ann
ounc
e..
8,60
16.
6c5.
6c4.
1c0.
9M
isce
llane
ous.
....
....
.24
,097
8.7c
7.8c
6.2c
3.2c
Uni
tSol
dC
ompl
ete
...
57−
3.9
−0.
01.
4−
3.1
Nam
eC
hang
e...
....
...
61−
6.3−
11.1
−17
.2−
26.7
aU
nitS
old
Poss
ible
....
.46
515
.3b
12.1
b9.
03.
3O
pini
on..
....
....
....
..3,
432
19.8
c17
.7c
15.5
c12
.4c
Not
e:A
bnor
mal
trad
ing
volu
me
isca
lcul
ated
asth
ere
sidu
alfr
omth
ere
gres
sion
oflo
gtr
adin
gvo
lum
eof
com
pany
ion
log
mar
kett
radi
ngvo
lum
edu
ring
trad
ing
days
ofth
ela
st12
0ca
lend
arda
ys.A
vera
geSC
ATV
sar
esc
aled
by10
0fo
rea
sier
read
abili
ty.T
heco
lum
nsid
enti
fyth
ele
ngth
ofth
eev
entw
indo
wof
#d
trad
ing
days
,whi
chin
clud
esth
eev
entd
ayan
dth
etw
otr
adin
gda
ysbe
fore
the
even
tday
.Eve
ntty
pes
wit
hfe
wer
than
50ob
serv
atio
nsar
esu
ppre
ssed
.St
atis
tica
lsig
nific
ance
atth
e95
%,9
9%,a
nd99
.9%
leve
lsof
confi
denc
eis
indi
cate
dby
the
supe
rscr
ipts
a,b
,and
c,r
espe
ctiv
ely.
37
Tabl
e9:
10-D
ayA
SCA
Rs
byEv
entT
ype
and
Con
diti
oned
onBu
sine
ssC
ycle
Even
tN
All
Con
.Ex
p.Ev
ent
NA
llC
on.
Exp.
All
New
sEv
ents
....
...
185,
690
0.0
−4.
9c0.
6bLe
galI
ssue
sA
nyC
lass
ified
Even
t...
181,
684
0.1
−4.
7c0.
7bLa
wsu
itA
gain
stSt
art.
311−
10.4
−9.
4C
orpo
rate
Gov
erna
nce
Law
suit
ByEn
d..
....
.4,
645
0.8
−5.
1c1.
5C
ompe
nsat
ion
....
....
.99
15.2
16.6
Law
suit
ByO
ngoi
ng..
245
−8.
0−
8.8
Elec
tCEO
....
....
....
..3,
701
2.7
1.4
2.8
Law
suit
BySt
art.
....
.1,
162
−6.
2a−
18.9
c−
4.9
Elec
tDir
ecto
r..
....
....
5,48
20.
8−
2.3
1.1
Lega
lGov
.End
....
...
741
−5.
0−
9.0a
−4.
6El
ectE
xecu
tive
....
....
.14
,010
0.2−
10.3
c0.
9Le
galG
ov.O
ngoi
ng..
.80
3.9
4.7
Elec
tVP
....
....
....
....
3,08
7−
4.7b−
11.1
c−
3.8a
Lega
lGov
.Sta
rt..
....
.52
2−
2.8−
10.2
a−
1.7
Earn
ings
Rep
orts
Ope
rati
onal
Issu
esEa
rnin
gsFo
reca
stD
own
3,33
3−
33.0
c−
29.8
c−
33.4
cC
apac
ity
Up
....
....
..5,
338
−1.
2−
3.0a
−0.
9Ea
rnin
gsFo
reca
stU
p..
.94
53.
2−
0.5
4.7
Con
trac
t...
....
....
...
21,3
960.
3−
3.2c
1.2
Earn
ings
Rep
ortD
own
.10
,101
−9.
7c−
17.0
c−
8.8c
Labo
rC
ontr
actA
ccep
t53
9−
0.4
19.6
c−
3.0
Earn
ings
Rep
ortU
p..
..18
,088
4.4c
1.4
4.7c
Layo
ff..
....
....
....
..2,
032
−1.
0−
16.8
c2.
5Fi
nanc
ialI
ssue
sPr
oduc
tNew
....
....
.5,
539
2.8a
−2.
9a3.
3a
Deb
tIss
ue..
....
....
....
524
−1.
0−
5.8
−0.
3Pr
oduc
tNew
Poss
ible
.66
9−
0.8−
29.1
c2.
0D
ebtR
epur
chas
e..
....
.98
−8.
6−
7.7
Supp
lyA
gree
men
t...
.1,
373
2.5
−3.
03.
0D
ivid
end
Dow
n..
....
..80
−7.
2−
5.7
Res
truc
turi
ngIs
sues
Div
iden
dU
p..
....
....
.41
816
.7c
18.6
cJ.V
.Res
truc
turi
ng..
...
126−
13.6
−15
.4Eq
uity
Issu
e..
....
....
..1,
005
−8.
9b5.
7−
10.1
bJo
intV
entu
reSt
art
....
5,26
64.
4b0.
04.
9c
Equi
tyR
epur
chas
e..
...
2,45
028
.2c
37.4
c27
.0c
Mer
geA
nnou
nce
....
.16
,623
5.7c
0.9
6.0c
List
ing
Cha
nge
....
....
.1,
126
−9.
6b7.
2a−
11.7
cM
erge
Com
plet
e..
....
912
2.8−
21.3
c4.
4Pr
ice
Up
....
....
....
...
1,37
3−
1.5
0.2
−2.
2M
erge
Fail
....
....
....
640
−8.
5a−
20.7
c−
7.8
Rat
ing
Dow
n..
....
....
.1,
236
−3.
7−
19.7
c−
2.0
Mer
geFa
ilPo
ssib
le..
.30
8−
3.4
10.9
−5.
1St
ock
Split
....
....
....
.11
323
.8a
25.1
bM
erge
Poss
ible
....
....
6,81
0−
3.9b
−0.
1−
4.2c
Gen
eral
Issu
esU
nitB
ough
tAnn
ounc
e1,
004
−0.
6−
7.3a
0.1
Cor
rect
ion
....
....
....
.22
85.
85.
8U
nitS
old
Ann
ounc
e..
8,91
04.
0c−
7.4c
5.1c
Mis
cella
neou
s...
....
...
24,5
81−
1.7b
−8.
7c−
1.0
Uni
tSol
dC
ompl
ete
...
58−
18.8
−14
.4N
ame
Cha
nge
....
....
..69
−6.
5−
0.8
Uni
tSol
dPo
ssib
le..
...
476
15.8
c17
.3c
Opi
nion
....
....
....
....
3,45
8−
0.1
0.5
−0.
1N
ote:
Ave
rage
stan
dard
ized
cum
ulat
ive
abno
rmal
retu
rns
(ASC
AR
s)ar
esc
aled
by10
0fo
rea
sier
read
abili
ty.T
heco
lum
nsid
enti
fyth
ele
ngth
ofth
eev
entw
in-
dow
mea
sure
din
trad
ing
days
,sta
rtin
gtw
otr
adin
gda
ysbe
fore
the
even
tday
.Th
eco
lum
nsre
fer
toth
eco
ndit
ioni
ngev
ents
:‘A
ll’re
fers
toth
eun
cond
itio
ned
case
;‘C
on.’
refe
rsto
the
cont
ract
ion
ofth
ebu
sine
sscy
cle,
and
‘Exp
.’re
fers
toth
eex
pans
ion
ofth
ebu
sine
sscy
cle.
Even
ttyp
esw
ith
few
erth
an50
obse
rvat
ions
are
supp
ress
ed.
Stat
isti
cals
igni
fican
ceat
the
95%
,99%
,and
99.9
%le
vels
ofco
nfide
nce
isin
dica
ted
byth
esu
pers
crip
tsa
,b,a
ndc,r
espe
ctiv
ely.
The
mar
ket
mod
elre
gres
sion
sar
eba
sed
ona
120
cale
ndar
-day
pre-
even
twin
dow
and
the
equa
l-w
eigh
ted
mar
keti
ndex
.
38