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, werner.antweiler@sauder.ubc.ca.

Murray Frank: Phone 604-822-8480, Fax 604-822-8477, murray.frank@sauder.ubc.ca. 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