University of Southern California Dissertations and Theses

Dividend policy, earnings announcements, and analysts' earnings forecasts

(USC Thesis Other)

Dividend policy, earnings announcements, and analysts' earnings forecasts

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content DIVIDEND POLICY, EARNINGS ANNOUNCEMENTS, AND ANALYSTS'
EARNINGS FORECASTS
by
Perikles Fotiou Konstantinides
A Dissertation Presented to the
FACULTY OF THE GRADUATE SCHOOL
UNIVERSITY OF SOUTHERN CALIFORNIA
In Partial Fulfillment of the
Requirements for the Degree
DOCTOR OF PHILOSOPHY
(Finance and Business Economics)
January 1999
Copyright 1999 Perikles Fotiou Konstantinides
Dedication
This work is dedicated to my mother, Georgia, who made it all possible,
and to the memory of my father, Fotios, who would have been pleased.
11
Acknowledgements
This work would not have become possible without the tolerance, the persistence,
and the friendship of the Chairman of my Dissertation Committee, Professor Harry
DeAngelo. For all the years of all this, I cannot thank him enough.
I would also like to extend my warmest appreciation to the members of my
dissertation committee, Professors Mark Weinstein, Kevin Murphy, and George Pa
pavassilopoulos.
For their support and friendship, or both, Michael Konstantinides, David Shimko,
Helen Pitts, Marilyn Johnson, George Michailidis, Nicholas Kyriazis, Demetrios De mekas, Vera Wilhelm, Linda Stauffer, and Alex Morales.
The contribution of Deborah Gruenfeld in all this has been subtle and indirect ,
but essential.
lll
Contents
Dedication
Acknowledgements
List Of Tables
Abstract
1 Introduction
1.1 Background . . . . . . . . . . . . . . .
1.2 The Information Content of Dividends
1.2.1 The theory . . . . . . . . . . .
1.2.2 The empirical evidence on the ICD Hypothesis .
1.2.3 Dividend changes and earnings expectations .
1.2.4 Dividend changes and abnormal stock returns
1.3 Earnings forecasts and the ICD Hypothesis
1.3.1 Studies of the forecast revision .
1.3.2
1.3.3
Comments ........ .
The Ofer and Siegel study . . .
2 The sample of dividend changes and earnings forecasts
2.1 Screening criteria .
2.2 Sample description .....
3 The secular revision activity
3.1 The intra-year decline of forecast errors
3.2 The intra-year forecast revision .. . .
3.3 The sample statistics of forecast errors
3.3.1 The raw forecast errors ...
3.3.2 The absolute forecast errors
3.4 Conclusions . . .. . . ...... .
11
111
Vl
vu
1
4
9
9
12
12
14
17
18
22
26
33
33
37
44
44
46
49
50
53
55
IV
4 Forecast revisions after the dividend announcement 57
4.1 The measures of the forecast revision after the dividend announcement 57
4.1.1 The dollar forecast revision
4.1.2 The percentage forecast revision ..
4.2 The revision after dividend announcements
4.2.1 Discussion . . . . . . . . . . . . . .
4.3 Earnings forecasts and announcements . .
4 .3.1 The timing of the dividend change announcement
4.3.2 Discussion . ........... .. .... . .. .
4.4 The regression analysis of the excess forecast revision . .
4 .4 .1 Analysis of the one-month excess forecast revision
4.4.2 Analysis of the two-month excess forecast revision
5 Conclusions
61
63
66
73
76
78
82
83
89
92
94
V
List Of Tables
1.1 A summary of the empirical evidence on the ICD hypothesis . . . . . 13
1.2 The empirical evidence on the impact of announcements of dividend
changes on analysts ' earnings forecasts 19
2.1 Sample selection process . . . . . . . . 36
2.2 Sample descriptive statistics . . . . . . 39
2.3 Year, month, and industry sample breakdown 41
2.4 Year, month, and industry sample breakdown ( continued) . 42
3.1 The intra-year earnings forecast revision activity . . . . . . 47
3.2 The intra-year raw forecast error behavior of the full sample 51
3.3 The intra-year absolute forecast error behavior of the full sample . 52
4.1 Mean consensus forecast revision after the dividend announcement :
Dividend increases 67
4.2 Median consensus forecast revision after the dividend announcement :
Dividend increases 68
4.3 Mean consensus forecast revision after the dividend announcement :
Dividend cuts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
4.4 Median consensus forecast revision after the dividend announcement :
Dividend cuts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
4.5 Tabulation of the mean consensus earnings forecast revision around
the dividend announcement month . . . . . . . . . . . . . . . . . . . 79
4.6 Tabulation of the median consensus earnings forecast revision around
the dividend announcement month . . . . . . . . . . . . . . . . . . . 80
4. 7 OLS analysis of the excess mean consensus forecast revision around
the dividend month: one-month revision measure . . . . . . . . . . . 85
4.8 OLS analysis of the excess mean consensus forecast revision around
the dividend month: two-month revision measure . . . . . . . . . . . 86
4.9 OLS analysis of the excess median consensus forecast revision around
the dividend month: one-month revision measure . . . . . . . . . . . 87
4.10 OLS analysis of the excess mean consensus forecast revision around
the dividend month: two-month revision measure . . . . . . . . . . . 88
Vl
Abstract
This thesis examines the impact of announcements of dividend changes on analysts'
earnings forecasts . In particular, I inquire whether security analysts revise their
forecasts of annual earnings per share, after the announcement of shifts in divi
dend policy, significantly differently than in prior months, after controlling for other
relevant information sources such as contemporaneous quarterly earnings announce
ments. This is a direct test of the Information Content of Dividends Hypothesis
of Modigliani and Miller, who proposed that shifts in dividend policy should signal
changes in managerial expectations of future earnings.
I find that changes in corporate dividend policy, on average, do not generate
revisions in earnings forecasts which are significantly different from the documented
secular forecast revision activity in the months prior to the dividend announcement
month. The only instance in which forecast revisions are statistically significantly
different from the secular forecast revision activity is after dividend cuts which were
announced within a calendar month of a quarterly earnings announcement. The
findings of this thesis do not support the predictions of the Information Content of
Dividends Hypothesis .
\ill
Chapter 1
Introduction
This dissertation exammes the behavior of analysts' earnings forecasts after an
nouncements of corporate dividend changes. It tests directly Modigliani and Miller's
[58] Information Content of Dividends Hypothesis that changes in dividend payouts
signal managers ' expectations of a shift, in the same direction, in future permanent
earnings. Such a signal should cause professional market participants, such as finan
cial analysts , to revise their forecasts of firms ' earnings in the same direction as the
dividend change.
Denis, Denis and Sarin [26], Lang and Litzenberger [4 7], Ofer and Siegel [63] , and
Yoon and Starks [73] have investigated the impact of announcements of dividend
changes on analysts' earnings forecasts. With one exception (Lang and Litzenberger
[47]), these authors find (a) that analysts revise their forecasts of year-end earn ings after announcements of dividend changes, and (b) the revision is in the same
direction as the dividend change. The earlier papers have attributed the analysts'
revision to the dividend announcement but have ignored the effect of other informa
tion sources such as quarterly earnings announcements contemporaneous with the
dividend announcement.
I extend the current literature by analyzing analysts' responses to dividend an
nouncements after controlling for contemporaneous earnings announcements. I use
the I/B/E/S Inc. Historical tape which contains analysts ' earnings forecasts on a
monthly basis. These reflect the consensus forecast of year-end earnings released by
many of the largest investment houses in the United States. The forecasts, and their
1
changes , are likely affected by quarterly earnings announcements which not only ma
terially reduce uncertainty about the level of annual earnings but also may change
expectations about current and future earnings levels. Therefore, any examination
of the impact of announcements of dividend changes on monthly earnings forecasts
should control for the possibility that the average documented revision is due to the
quarterly earnings announcement rather than the dividend signal.
I examine:
• The average raw rev1s1on m analysts ' forecasts of annual earnmgs per share
after the announcement of a dividend change. I examine the raw revision
using both the mean and the median consensus analysts' earnings forecast. In
addition, I measure the revision over both the one month after, and the two
months around the dividend announcement month.
• The excess forecast revision after the dividend announcement month defined as
the raw forecast revision after the dividend announcement minus the average
forecast revision in up to three preceding months. Following Denis, Denis,
and Sarin [26], I adjust the measure of the forecast revision for the intra-year
secular earnings forecast revision activity.
• The statistical significance of the excess forecast rev1s10n m the subsample
of dividend announcements which occurred within a calendar month of the
nearest earnings announcement compared to the control subsample of dividend
announcements which occurred more than one month before or after such an
earnings announcement . This is a test of the marginal informational content
of the dividend change announcement after controlling for other information
sources because it separates the potential informational impact of earnings
announcements in a sample of a common treatment, in this case a sample of
dividend changes.
• The explanatory power of the dividend signal in linear regression models of
the excess forecast revision, and especially in models which include various
2
combinations of independent variables for the contemporaneous earnings m
formation.
My sample consists of 2,866 dividend changes of more than 10% in absolute value,
between 1979-90, announced by 909 New York Stock Exchange or American Stock
Exchange firms.
The bulk of my findings indicates that, after controlling for the contemporane
ous earnings announcement information, announcements of dividend changes , on
average, do not cause analysts to change their year-end earnings forecasts. The only
exception to this general trend are announcements of dividend cuts which were made
within a 40-trading day window around the nearest quarterly earnings announce ment. Furthermore, the magnitude of the dividend change is not a significant ex planatory variable of the excess forecast revision, both in models in which it is the
only independent variable, and in models which include variables for contemporane
ous earnings information.
More specifically, the main empirical findings of this study are:
• The average excess forecast revision after announcements of dividend changes
(i.e. the forecast revision after controlling for the secular forecast revision in
any month of the year), with only one exception, is statistically significantly
different than zero only after dividend cuts which were announced within a
40-trading day window around the nearest earnings announcement.
• The excess forecast revision is not correlated with the magnitude of the div
idend change. In linear regression models of the excess forecast revision, the
coefficient of the dividend change variable is not statistically significant , even
in a model where it is used as the only independent variable.
• The two significant explanatory variables in all regression models of the excess
forecast revision after dividend change announcements in my sample are (a)
the year-to-year change in quarterly Earnings per Share Before Extraordinary
Items (Quarterly Industrial COMPUSTAT Item 8) in the quarter of the divi
dend change, and (b) a dummy variable which takes the value 1 if the dividend
3
was cut and its announcement was within a 40-trading day window around the
nearest earnings announcement, and O otherwise.
• Analysts revise their year-end earnings forecasts throughout the fiscal year,
and using measures of the post-dividend announcement forecast revision which
do not control for this secular revision activity results in drawing erroneous
conclusions about the informational content of the dividend change, whether
in the presence of earnings information or not. In particular, all measures
of the forecast revision which do not control for the secular revision activity
document significant forecast revisions after both dividend increases and cuts,
and dividend announcements independently of their timing with respect to the
nearest earnings announcement.
1.1 Background
The assessment of the potential impact of announcements of dividend changes on an
alysts ' earnings forecasts is an extension of the research on the Information Content
of Dividends (ICD) hypothesis , or "cash flow signaling" hypothesis of Modigliani
and Miller [58]. The M & M ICD hypothesis posits that managers change dividend
policy only when their expectations of non-transitory future earnings change. Div
idend changes, then, should be a credible and costly managerial signal of shifted
expectations of future profitability.
The three main empirical implications of the ICD hypothesis have been the sub
ject of a large volume of research. First, dividend changes are predicted statistically
significant explanatory variables in linear statistical models of current and future
earnings . Second, the firm's stock price should adjust on the day of the announce ment to reflect the present value of the marginal future cash flows signalled by the
dividend change. Third, forecasts of current and future earnings should be revised
in the direction of the dividend change. Tables 1 and 2 summarize the main results
of some of the most important of these studies.
4
The empirical evidence on the link between the potential signaling function of
dividend changes and current and future earnings realizations is inconclusive (see
Gonedes [38] , Riding [67] , Watts [71], versus DeAngelo, DeAngelo and Skinner [25],
Penman [64], Pettit [65]) . As a result , Miller [53] has questioned whether dividends
can be viewed as predictors of future earnings, since a strong link between the two
has not been established.
The stock market reaction to announcements of dividend changes is significant
and well documented ( among others in Aharony and Swary [1] , Charest [18], Eades,
Hess and Kim [30], Pettit [65]). Consistent with expectations updating, dividend
increases on average generate abnormal returns of no more than 1 % around the
day of the dividend announcement , and dividend cuts cause stock prices to fall by
an average of around 5%. Nevertheless, in this case also, the connection between
the abnormal stock return around the dividend announcement day and the ICD
hypothesis of Modigliani and Miller [58] is not well established.
An alternative potential explanation for the abnormal stock price reaction around
the dividend announcement day was proposed by Jensen [43]. In what is called the
"overinvestment hypothesis" , Jensen postulates that firms may be subject to agency
costs, if managers control large amounts of free cash flows , because of the potential
divergence between the interests of shareholders and management, and the potential
misuse ( through "overinvestment") of these cash flows.
According to Jensen, dividend increases (cuts) signal reduced (increased) free
cash flow at managers' discretion and, as a result, a reduction (rise) in potential
agency costs. The firm 's stock price adjusts inversely to the direction of the shift
in potential agency costs. The overinvestment hypothesis has been tested and its
predictions confirmed in Lang and Litzenberger [47] and Bajaj, Vijh and Westerfield
[7]. As will be discussed in the next section, Denis, Denis and Sarin [26] question
the conclusions of [4 7] and provide different explanations for the documented stock
price behavior.
The third testing methodology of the ICD is to examine directly the potential
changes in market expectations of firms' future profitability after dividend changes.
5
This approach involves examining the behavior of financial analysts' forecasts of
annual earnings after the announcement of a dividend change.
Examining the revision of analysts ' earnings forecasts after a dividend change an
nouncement is a direct assessment of the impact of the dividend change on market
expectations of firms' future earnings for two reasons. First, analysts' earnings fore
casts are expectational data directly associated with earnings, whereas stock price
reactions can be attributed to both the ICD hypothesis' signaling predictions, and to
earnings-related information content consistent with the overinvestment hypothesis.
Second, market participants are likely to include these forecasts in determining their
investment strategies, as the best available predictions of firms ' earnings.
With the exception of Lang and Litzenberger [47] who find no statistically sig
nificant forecast revisions after the announcement of a dividend change, the other
papers in this line of research document significant analysts ' earnings forecast re
visions after announcements of dividend changes. In particular, Denis , and Sarin
[26] and Yoon and Starks [73] document that their measures of the standardized
median consensus forecast revision around the dividend announcement month are
significantly different than zero, after both dividend increases and dividend cuts.
These findings are at least prima facie evidence of a significant effect of the dividend
change on analysts ' earnings expectations.
The methodology used in the studies by Lang and Litzenberger, Yoon and Starks,
and Denis, Denis and Sarin does not control for other information sources, such as
earnings announcements, near the announcement date of the dividend change. Such
other sources convey material information on firms' year end earnings and are likely
to affect analysts' forecasts of these earnings. Therefore, earnings announcements
and other similar information signals contemporaneous with the dividend announce
ment should be considered jointly with the dividend signal when assessing the effect
of the dividend signal on the forecast revision.
In addition, the measures of the forecast revision used in [4 7] and [73] do not
control for the expected, normal forecast revision in any month of the year, which
is an empirical regularity documented in, among others, Brown, Hagerman, Griffin
6
and Zmijewski [14], and Elton, Gruber and Gultekin [34]. Furthermore, all the
above papers use a two-month period around the dividend announcement month in
which they assess the forecast revision. As discussed in Section 1.3.2 below, this
examination period could overstate the actual forecast revision.
The most-cited study in this area of research, and the only one focusing exclu sively on the impact of announcements of dividend changes on analysts' earnings
forecasts, is Ofer and Siegel [63]. Using a sample of 781 dividend change announce
ments, from 1976 to 1984 Ofer and Siegel assess the relation between a measure of
the dividend announcement surprise and the level of the forecast error ( defined as
actual minus forecasted earnings) in each of the six months surrounding the dividend
announcement month.
Specifically, they examine six cross-sectional regression models. The dependent
variable in each one is the forecast error in each month, from one month before to
four months after the dividend announcement month. The explanatory variable in
all six regressions is the same: a variable which accounts for the surprise in the
dividend announcement . They find that the coefficient of the dividend surprise
variable declines in successive regressions, i.e. successive months. Ofer and Siegel
say that this decline of the coefficient estimate is evidence that:
• The severity of the forecast error before the month of the dividend announce
ment is systematically related to the information content of the dividend an
nouncement, and
• The systematic relationship between the forecast error and the dividend sur
prise weakens after the dividend announcement month, i.e. analysts incorpo
rate the earnings-related information contained in the dividend announcement
in their earnings forecasts .
In this thesis , I will argue that:
7
1. The examination of the information content of dividend changes with respect
to analysts' earnings forecasts should include the potential effect of other con
temporaneous sources of earnings-related information, such as announcements
of quarterly earnings.
2. The appropriate measure of the forecast rev1s10n to be used in such an ex amination of the informational content of dividend changes with respect to
analysts ' earnings forecasts should account for the documented intra-year sec
ular forecast revision activity.
3. The appropriate period over which to assess the forecast rev1s10n after the
dividend announcement is the one month after the announcement month and
not the two months around the announcement month, which is the period used
in [47] , [26], and [73] . In this study I examine both one-month and two-month
rev1s1on measures.
4. The appropriate measure of analysts' earmngs forecasts when assessmg the
informational impact of dividend announcements on them is the mean con
sensus earnings forecast and not the median consensus earnings forecast. In
my study, I use both the mean and the median consensus analysts' earnings
forecast.
Finally, my thesis does not address any aspect of market efficiency. In general,
the literature on the impact of announcements of dividend changes on analysts'
earnings forecasts is not directly related to the Efficient Markets Hypothesis (EMH).
The main prediction of the semi-strong form of the EMH with respect to dividend
announcements is that all information about a firm's value contained in the change
of dividend policy is fully incorporated in the company's stock price. As a result, no
abnormal returns can be earned by trading based on the dividend information. As
discussed above, the dividend information would be construed by investors either as
a signal of insiders ' expectations of future earnings, or as a signal of a shift in agency
costs.
The issues explored in my dissertation are unrelated to this question. In par
ticular, the question I am investigating is whether dividend change announcements
8
contain information signals related to future earnings at all, not how this informa
tion affects or does not affect security prices. In other words, the research program
on the effect of dividend changes on analysts' earnigs forecasts is concerned with the
existence of an information signal, not with examining the existence of profitable
trading strategies based on dividend signals. My thesis is similarly not examining
any aspects of market efficiency.
The thesis is organized as follows. In the remaining of this chapter I discuss the
ICD hypothesis, the methodology and results of the research on forecast revisions
around the dividend announcement month, the Ofer and Siegel study, and the ques
tions arising from their methodology and results. Chapter 2 presents my sample,
and the sample descriptive statistics. Chapter 3 discusses the literature on forecast
errors, and the behavior of forecast errors in my sample, and in a control sample
in fiscal years without a dividend change. Chapter 4 presents the methodology and
results of the examination of forecast revisions and concludes.
1.2 The Information Content of Dividends
1.2.1 The theory
In his 1956 survey of corporate managers on their firms' dividend policies, Lintner
[49] suggested that managers change their firms' dividend payouts based on current
net earnings, which is a measure of a firm's performance understood and followed
by the market.
1
More specifically, management decisions about cash distribution
levels to shareholders are made considering past, present and future earnings, and
1
From Lintner [49, p.100]: "[ ... ], it became clear that any reason which would lead manage
ment to decide to change an existing rate [of dividend payouts] -and any reason which would
be an important consideration in determining the amount of the change- had to seem prudent
and convincing to officers and directors themselves and had to be of a character which provided
strong motivations to management. Consequently, such reasons had to involve considerations that
stockholders and the financial community generally would know about and which management
would expect these outside groups to understand and find reasonably persuasive, if not compelling.
Current net earnings meet these conditions better than any other factor".
9
a target payout ratio which differs across firms. The actual decision to change the
dividend payout reflects current earnings levels. Lintner notes managers' "general
reluctance to make reductions in dividend rates, especially in 'regulars'," implying
that such a decision would have a grave impact on investors' assessment of a firm's
value and expectations of future prospects.
Modigliani and Miller [58], citing Lintner's findings as the empirical foundation
of their hypothesis, propose that managers in pursuit of a policy of "stabilizing
dividends" distinguish earnings into their permanent and transitory components,
and pay dividends based on permanent future income prospects.
2
Every shift in
dividend policy reveals to the market managers' expectations of current and fu
ture earnings and, therefore, is a potentially strong "insider" signal of the firm's
profitability.
The M & M dividend signaling proposition has been extended in a number of
theoretical models of dividend behavior which apply the tools introduced in the
seminal work of Spence [69] on labor market equilibrium under asymmetric infor
mation, and Ross [68], who presented a model of capital structure choice consistent
with costly insider signaling. These models ([9], [45], [56], [70]) explore the potential
cash flow signaling function of dividend policy in a world of asymmetric information
between the managers and the outside owners of a firm. Managers, in these mod
els, possess information about their firms' earnings prospects which can be credibly
communicated to outside shareholders only through a costly signaling mechanism.
Bhattacharya [9] develops a signaling equilibrium model, in which dividends are
taxed at a higher rate than capital gains. The dividend distribution maximizes
current shareholders' wealth because it conveys to the market manager's information
about future cash flows and, thus, the firm's "true value" consistent with those
managerial expectations. John and Williams [45] propose a signaling equilibrium
2
1n their reply to David Durand's criticism of their capital structure irrelevancy proposition,
Modigliani and Miller [58] write that "[ ... ] whenever corporations follow a policy of stabilizing
dividends-and the excellent studies of Lintner [ ... ] leave no doubt that the majority of publicly
held corporations usually do-dividends will contain considerable information about [the expected
value of the (uncertain) earnings of the assets currently held], possibly even more than [about
current earnings]."
10
in which managers pay costly, taxable dividends when the firm 's and shareholders'
cash needs exceed corporate, internally generated funds. The dividend payouts are
financed with new equity issues.
Miller and Rock [56] develop a signaling model in which managers possess inside
information about the firm 's prospects which cannot be credibly conveyed to out side shareholders costlessly. Managers , then, pay higher dividends than they would
have to pay in the absence of the informational asymmetries, in order to convince
shareholders about the firm's earnings prospects. Firms forego positive net present
value investment opportunities because of financing constraints caused by the higher
dividend levels. Dividend changes convey information about current earnings, and
possible stock price changes are not necesarily related to signals about future prof
itability. In particular, a dividend cut does not signal shifts in earnings expectations.
It is just an interruption of the signaling process of higher than otherwise dividend
levels.
In Wa1-ther [70] managers possess information which cannot be credibly conveyed
to shareholders. In equilibrium, dividends reveal information only when they are
reduced or omitted because such a decision enables outsiders to distinguish between
poorly and adequately managed firms. Announcements of unchanged or increasing
dividends are not information signals , because firms are perceived by investors to
choose a dividend policy such that they will almost always be able to maintain . Both
excellent and mediocre firms belong to the same information subset with respect to
their dividend payouts , i.e. both of these types of firms can meet their preset payout
targets. Only the worst firms cannot meet their dividend payout target , and their
dividend decision is the only one which is truly informative.
Dividends in the above models serve as costly signals of managers ' expectations
of firms' earnings prospects because of uncertainty and informational asymmetries
between managers and outside shareholders. These market failures cause corporate
dividend policy to depart from the irrelevancy Miller and Modigliani [55] proved
under perfect capital markets. Consequently, dividend policy, in the presence of in
formational asymmetries and other market imperfections, has a direct impact on firm
11
value through its effect on investors' information sets, and expectations of corporate
earnmgs.
1.2.2 The empirical evidence on the ICD Hypothesis
The information content of dividend policy changes has been examined in a large
number of empirical studies. In general, investigations of the ICD hypothesis have
focused on (a) the usefulness of the dividend-change signal in determining current
and predicting future earnings, (b) the firm's stock price reaction around the day
of the announcement of a dividend change, and ( c) the impact of the dividend
announcement on analysts' earnings forecasts.
The next two sections present the main results from the first two lines of research.
3
The evidence on the effect of announcements of dividend changes on analysts'
earnings forecasts is discussed in section 1.2. Table 1 presents a summary of the
research on the predictive power of dividend announcements with respect to current
and future earnings, and on the stock price reaction around dividend announcement
days.
1.2.3 Dividend changes and earnings expectations
The first empirical test of the signaling function of dividend changes with respect to
firms' profitability was by Watts [71] who finds no evidence that current dividends
help predict future earnings after controlling for current earnings. Healy and Palepu
[42] document that the dividend decisions of firms initiating or omitting dividends
support the Lintner findings, because they are consistent with managers considering
past, current and future earnings as determinants of the payout ratio. For example,
3
More complete surveys of the theory on dividend policy, its signaling implications, and the
empirical results can be found in Ang [3] and Allen and Michaely [2]. The papers cited here are
indicative of the general methodology and empirical findings and their selection is by no means
exhaustive.
12
Table 1.1: A summary of the empirical evidence on the ICD hypothesis
This table presents a summary of the most important empirical work on the fuformation Content of Dividends. The
hypotheses that have been tested in the listed papers can be roughly divided into two categories: (a) Dividends
are paid based on past and current earnings and a target payout ratio (Lintner [49]), or based on projected non
transitory, future earnings (Modigliani and Miller [58]), thus, providing information about the firm's current and
future performance, and (b) Dividends serve as signals of the true value of the firm in a world of asymmetric
information (Bhattacharya [9], John and Williams [45], etc.), The empirical studies on these hypotheses, listed
below, can be separated into papers which have examined the usefulness of the dividend announcement information
in predicting current and future earnings, and papers which have assessed the stock price reaction around the
dividend announcement day, as an indication of the informativeness of the dividend change.
Study
A. Dividends as signals of future earnings
Watts
Healy and Palepu
Brickley
DeAngelo et al.
DeAngelo et. al.
B. Price reactions around dividend dates
Pettit
Charest
Eades, Hess, and Kim
Aharony and Swary
Lang and Litzenberger
Healy and Palepu
Asquith and Mullins
Michaely, Thaler, and Womack
Findings
Using monthly return data, dividends do not predict future earnings
better than do current earnings
Payout ratio depends on past, current, and future earnings
Changes of regular dividends are more informative than changes
of specially designated dividends. Dividend changes predict
earnings changes.
Dividend cuts by firms experiencing significant financial distress
improve predictive ability with respect to future earnings
Dividend decisions of firms whose earnings drop after at least
9 years of continuing growth are not informative
of future earnings direction.
Announcements of dividend increases result in positive excess returns
(0.53%) while dividend cuts result in negative excess returns (-3.70%)
Confirmed the findings of Pettit [65].
Confirmed the findings of Charest [18], and Pettit [65].
Confirmed the findings of Charest [18], and Pettit [65].
Price reaction around a dividend change announcement is consistent
with overinvestment hypothesis of Jensen [43].
Abnormal returns of 3.9% around the announcement day of a dividend
initiation, and -9.5% around a dividend omission announcement.
Confirmed the findings of Healy and Palepu [42].
Confirmed the findings of Healy and Palepu [42] and documented
excess long term drifts in the months after the announcement.
13
they show that the earnings of the 131 firms in their sample that initiate regular
dividend payments consequently rise, whereas dividend omissions, in 172 cases, are
followed by negative earnings growth.
Brickley [11] finds that earnings tend to increase after regular dividend increases,
and that regular dividend changes are more helpful than specially designated divi
dends in predicting future earnings. By implication, managers seem to treat their
regular dividend decision consistent with (a) the findings of Lintner [49], in that
regular dividends are the more important and informative corporate payout policy
instrument , and (b) the cash flow signaling hypothesis of Modigliani and Miller [58],
in terms of the link between dividend changes and future earnings.
Additional support for this connection between dividend changes and future earn
ings is provided in DeAngelo, DeAngelo and Skinner [24] who find that dividend cuts
in firms experiencing losses , even when augmented with special accounting items,
significantly improve the ability to predict future earnings.
More recently, DeAngelo, DeAngelo, and Skinner [25] present evidence against
the cash flow signaling hypothesis. Of their 145 sample firms which experienced an
earnings decline after at least 9 years of uninterrupted earnings growth, 99 firms
increased their dividend, although their earnings decline was typically not reversed
in subsequent years. In this sample, the dividend signal is not informative about
firms' earnings prospects, after controlling for current earnings.
1.2.4 Dividend changes and abnormal stock returns
The second line of investigation of the ICD hypothesis is the assessment of the stock
price reaction around the announcement day of dividend changes. If the dividend
announcement triggers an abnormal stock price reaction (i.e. negative or positive
risk-adjusted stock return), then the dividend signal conveys information about the
firm's expected performance, which leads investors to revalue the firm 's market value
of equity. Nevertheless, the stock price reaction may be related to aspects of the
14
firm's current and expected performance which are indirectly related to earnmgs
expectations, such as agency costs, wealth transfers from bondholders to share
holders, etc. Then, abnormal returns around dividend announcement dates do not
necessarily support the ICD hypothesis, since the dividend announcement carries
information content which is related to earnings prospects indirectly and not in the
sense proposed by Modigliani and Miller.
In one of the first tests of the cash flow signaling hypothesis, Pettit [65] exam
ines 135 announcements of dividend changes by NYSE-listed firms, from 1967 to
1969. He documents that firms which announce dividend increases on average ex
perience positive abnormal returns of 0.53% on the day of the announcement, while
firms which cut their dividends experience negative mean excess returns of 3.69%,
consistent with the ICD hypothesis.
Charest [1 8] finds that the stock price of firms announcing a dividend increase
jumps by 1 % on the day of the announcement, and the stock price of firms cutting
their dividend drops by an average of 3.18%, also on the day of the announcement .
Eades, Hess and Kim [30] examine all 15,100 dividend changes contained in the
CRSP files, from 1962 to 1980. They report that the raw return on the day of
the announcement of a dividend increase is 0.63%, and on the day of a dividend
cut -1.14%. Aharony and Swary [1], in a sample of 384 dividend increases and
4 7 dividend cuts, document announcement day returns of 1.35% and -4.35% for
dividend increases and cuts, respectively.
Healy and Palepu [42], in their investigation of initiations and omissions of regu
lar dividends, examine the stock price reaction around the day of the announcement .
They document a 3.9% excess return around the dividend initiation announcement
day, and a -9.5% excess return around the day of the announcement of a divi
dend omission. These asymmetric abnormal stock price reactions to initiations and
omissions are also documented in Asquith and Mullins [4]. Michaely, Thaler, and
Womack [52] find excess returns of around 3% around initiations and -7% around
om1ss10ns.
15
In the overinvestment-hypothesis literature, the separating criterion that is used
to distinguish between "overinvesting" and "well-managed" firms is Tobin's Q-ratio
of the market value of the firm's assets (i.e. the firm's market capitalization) over
their replacement value. These papers (e.g. [26], [47] etc.) have compared the
stock price reaction around the dividend announcement day between the two types
of firms, and tested the hypothesis that dividend changes by what are prescribed
as "overinvesting" firms cause significantly larger excess returns. The implication is
that dividend changes in firms with agency problems (i.e. whose Q < 1) indicate
shifts in free cash flow available to managers, and cause relatively larger reactions
than the dividend changes of firms with lower agency costs (i.e. firms whose Q > 1).
Testing the overinvestment hypothesis, Lang and Litzenberger [47] analyze 429
dividend changes of more than 10% or less than -10%, between 1979 and 1984. They
find that the stock price reaction around the dividend announcement is consistent
with Jensen's [43] agency costs explanation of the signaling function of announce
ments of dividend changes. The stock price reaction after the dividend announce
ment in much larger in magnitude for firms with relatively more severe agency prob
lems ( Q < 1) than for firms whose managers do not "overinvest".
Bajaj, Vijh and Westerfield [7] examine the relationship between stock mar
ket reaction to dividend announcements and the ownership structure of the firms
announcing dividend changes. They find that agency cost issues and the level of
insiders ' shareholdings affect the magnitude of the market reaction on the day of the
dividend change announcement: the larger the insiders' shareholdings the less pro
nounced the stock market reaction, consistent with the overinvestment hypothesis.
Denis, Denis and Sarin [26] revisited the issue arguing that the excess returns
around the announcement day of dividend changes may be related to the level of
the dividend change and to the tax clientele effect identified by Bajaj and Vijh
[6], rather than Tobin's Q. The authors examine 6,777 dividend changes of more
than 10% in the period 1962-1988, and document abnormal returns around the
dividend announcement day of 1. 76% for dividend increases, and -5. 71 % for dividend
cuts. They attribute these stock price reactions to both cash flow signaling and the
16
dividend clientele effect, i.e. the preference some groups of investors have for taxable
dividends.
1.3 Earnings forecasts and the ICD Hypothesis
The third method of testing the cash flow signaling hypothesis of corporate divi
dend policy changes is to examine analysts' earnings forecasts. These forecasts are
a proxy for explicit market expectations of firms' earnings, because analysts are
professional market participants who regularly release forecasts of firms' expected
current and future earnings, and a large part of the investment community relies on
these forecasts to form earnings expectations and trading strategies.
4
Analysts' earnings forecasts are more accurate than a wide variety of statisti
cal models of earnings prediction that have been compared with them ( see, among
others, Brown and Roze:ff [16], and Givoly and Lakonishok [37]). Nevertheless,
they typically fail on average to predict year-end earnings within a 10-20% margin
around the actual earnings realization (see, among others, Barefield and Comiskey
[8], Dreman and Berry [28], and Richards [66]). In addition, earnings forecast errors
( defined as actual annual earnings per share minus the analysts' earnings forecast in
each month of the fiscal year) are statistically and economically significant through
out the year, but decline from the first to the last month of the year, as has been
documented in the literature (Brown et al. [14], and Elton et al. [34]) and confirmed
in my study.
4
Both the widespread use and investment value of analysts' earnings forecasts are well docu
mented in the finance and accounting literatures. O'Brien and Bhushan [62] find that institutional
investors tend to buy the common stock of firms followed by analysts. Brown and Rozeff [16] show
that analysts' forecasts of earnings are superior to simple, time-series earnings forecasts. Givoly
and Lakonishok [36, 37] document that buy-and-hold strategies are inferior to active portfolio
reshuffling based on analysts' recommendations. The superiority of investment strategies based on
earnings forecasts has been confirmed in, among others, Dimson and Marsh [27], Elton, Gruber
and Grossman [33], Brown, Richardson and Trzcinka [15], and Fried and Givoly [35]. Recently,
Womack [72] documents significant initial excess stock price reactions to analysts' "buy" (3%) and
"sell" (-4.7%) recommendations. He also finds excess stock price drifts or 2.4% in the one month
after the change of a recommendation to "buy", and -9.1% in the 6 months after the change of a
recommendation to "sell" .
17
In other words , analysts ' earnings forecasts ( a) are not very accurate in predict
ing actual earnings , but (b) they become more accurate toward the end of the year
compared to the beginning of the year. An implication of this intra-year improve
ment in accuracy ( or, alternatively, the intra-year decline in the forecast error) is
that analysts ' earnings forecasts typically change throughout the year, i.e. there is
considerable forecast revision activity.
The source of these revisions is direct ( e.g. quarterly earnings announcements)
and indirect ( e.g. investments in new business lines , restructurings, etc.) information
about firms ' performance and projected earnings that reach the market throughout
the year. This secular forecast revision activity is an important component of the
tests of the forecast revision after the announcement of a dividend change that I will
use in this study.
The potential experimental link between analysts ' earnings forecasts and the ICD
hypothesis is direct: if dividends contain information about firms' current and future
earnings , then analysts' forecasts of earnings should change after announcements of
dividend changes. This is a direct test of the cash flow signaling hypothesis because
the response variable of interest is the revision of the forecast of the earnings them
selves, and not a proxy like the stock price reaction to the dividend announcement.
1.3.1 Studies of the forecast revision
Table 2 summarizes the empirical findings on the impact of dividend announcements
on analysts' earnings forecasts. Lang and Litzenberger [47] examine 429 dividend
changes of more than ±10%, between 1979-1984, and the forecast revision, for each
of these dividend events , from the month before to the month after the dividend
announcement month. The analysts' earnings forecasts come from the 1/B/E/S Inc.
database.
90
18
Table 1.2: The empirical evidence on the impact of announcements of dividend
changes on analysts' earnings forecasts
A tabular presentation of the empirical evidence on the impact of announcements of dividend changes on analysts
earnings forecasts. This is an extension of the empirical testing of the cash flow signalling hypothesis of dividend
policy of Modigliani and Miller [61], in that the dividend change is directly linked to firms' earnings expectations,
i.e. analysts' earnings forecasts.
Authors
Ofer and Siegel
(JF, September 1987)
Lang and Litzenberger
(JFE, September 1989)
Denis, Denis, and Sarin
(JFQA, December 1994)
Yoon, and Starks
(RFS, Winter 1995)
Sample size Statistic used Authors' conclusions
781 The coefficient estimate of a series of six successive regressions of the
forecast error, in each of the six months around the month of the
dividend announcement, on a variable accounting for the dividend surprise Support for the ICD
declines in each successive regression. This is interpreted as evidence that
analysts revise their earnings forecasts after the announcement of the
dividend change.
429 The percentage change in the median earnings forecast in the two months
around the dividend announcement month, divided by the percentage No support for the ICD
dividend change is not statistically significant.
2,068 The percentage change in the median earnings forecast in the two months
around the dividend announcement month divided by the stock price two Support for the ICD
days before the announcement minus the typical two-month forecast revision
in the five months before the dividend announcement month is significant.
2,710 The percentage change in the median earnings forecast in the two months
around the dividend announcement month divided by the percentage Support for the ICD
dividend change is significant.
The measure of forecast rev1s10n Lang and Litzenberger use is the percentage
change in the median consensus earnings' forecast in the two months around the
month of the dividend announcement, standardized by the percentage dividend
change, i.e. the median consensus forecast revision elasticity with respect to the
dividend change. The authors find that the forecast revision elasticity is not sta
tistically significant, both for forecast revisions following dividend increases and for
revisions following dividend cuts.
Yoon and Starks [73] examine the earnings forecast revision after 2,710 announce
ments of dividend changes of more than ±10% (2,505 dividend increases, and 205
dividend cuts), in the period 1969-1988. They use the same statistic of forecast revi
sion used by Lang and Litzenberger. They document that analysts on average revise
their forecasts very significantly after dividend cuts, and somewhat after dividend
increases. In particular, for dividend increases the forecast revision elasticity is 8%
in the case of firms with Q < l (i.e. "overinvestors" ), and 1 % in the case of firms
with Q > l (i .e. firms without agency costs). In the subsample of the 205 dividend
cuts, for firms with Q < l the average revision elasticity is 56%, and for firms with
Q > l the elasticity is 27%.
The null hypothesis in Lang and Litzenberger [ 4 7], and Yoon and Starks [73] is
that the expected forecast revision in any month of the year is zero, and that any
observed forecast revision is unexpected and due to the dividend announcement. As I
discussed in the introduction, the two papers control neither for information sources ,
such as earnings announcements, which may be contemporeneous with the dividend
announcement , nor for the "normal", or "expected" secular forecast revision in the
month of the dividend announcement .
Denis, Denis and Sarin [26] , who account for the intra-year secular forecast revi
sion activity, find that analysts' earnings forecasts after announcements of dividend
changes significantly exceed the expected forecast revision in the same period. In
addition, the direction of the forecast revision is consistent with the sign of the div
idend change. The authors use a sample of 2,068 dividend changes of more than
20
±10% (1,865 increases and 203 cuts), between 1962 and 1988, and the correspond
ing monthly median consensus analysts' earnings forecasts from the 1/B/E/S Inc.
database.
Denis, Denis and Sarin argue that a test of the forecast revision after a dividend
announcement cannot ignore the documented secular revision activity of analysts'
earnings forecasts. To account for that secular forecast revision activity, they use
the methodology proposed by Brous [12] in a paper examining analysts' forecast
revisions following seasoned equity offerings.
In particular, they first calculate the change in the median consensus earnings'
forecast in the two-month period, from the month before to the month after the
dividend announcement month, standardized by the firm's stock price two days
before the day of the dividend announcement. Then, they adjust this revision for
the expected secular revision around the dividend announcement month.
The first step in assessing the expected forecast revision around the dividend an
nouncement month is to calculate the average two-month forecast revision through
out the year, for each firm in their sample, standardized by the firm's stock price in
the middle of each two-month period. Then, they subtract the average, intra-year
two-month forecast revision from the actual forecast revision in each two-month
period before the dividend announcement month (standardized by the firm's stock
price in the middle of each two-month period), in up to 5 preceding months. Finally,
they add this difference to the average full year revision and they use this figure as
their measure of the secular revision activity in the months preceding the dividend
announcement month.
Finally, they average these "excess" revisions in the period before the dividend
announcement, and they add this average "excess" revision to the average intra-year
two-month revision. For firm i in month m the expected forecast revision measure
(E[F ~,ml) is:
1 n-1
k + n ~ Ei,m-c
21
where k is the average two-month standardized revision in all months of the fiscal
year, t: is the average two-month standardized forecast revision in excess of k in the
5 months before the dividend announcement, and n is equal to five.
5
This is the
specification proposed by Brous [12].
The authors document that the "excess" two-month forecast rev1s10n (i.e. the
two-month forecast revision minus the expected forecast revision) from one month
before to one month after the dividend announcement month, significantly changes
by 1.4% of the preannouncement stock price in the case of dividend increases, and
-12% of the preannouncement stock price for dividend cuts.
1.3.2 Comments
The three papers discussed above leave room for two potential improvements in the
methodology that they use. The first is the issue of the marginal information impact
of the dividend change on analysts' earnings forecasts in the presence of contempo
raneous earnings information. The second is the issue of the period over which the
forecast revision is measured, and the statistic of analysts' earnings forecast used.
Perhaps the most important information that affects analysts' earnings forecasts
throughout the year is quarterly earnings announcements. Each one of these an
nouncements reduces uncertainty about annual earnings per share by around one
quarter of the actual figure, since annual earnings are the sum of the four quarterly
earnings realizations. In addition, earnings announcements are used by market par
ticipants to form expectations about current-year, and future earnings. Therefore,
each earnings announcement contains potentially significant, new or corroborating,
5
For example, if the dividend announcement occurred in month 8 of the fiscal year (in August,
if the fiscal year-end is in December), the benchmark forecast revision would be the average of the
forecast revisions in all two-month periods of the year, plus the mean of the differences between
this average and the actual forecast revisions between months 5 and 3, months 6 and 4, months
7 and 5, and months 8 and 6. The forecast revision around the dividend announcement month
(month 8) is calculated as the revision between months 9 and 7.
22
information about year-end earnings, and forecast accuracy should improve, on av
erage, after each quarterly earnings announcement .
The studies by Lang and Litzenberger, Yoon and Starks, Denis, Denis , and Sarin,
and Ofer and Siegel have not dealt with this potentially important marginal informa
tion source, although the joint informational effect of contemporaneous earnings and
dividend announcements has been pointed out and studied in the ICD Hypothesis
literature (see Griffin [39], and Leftwich and Zmijweski [48]) .
The second issue with regard to the evidence presented in [4 7], [73], and [26]
concerns the use of the two months around the dividend announcement month as
the relevant period of assessment of the forecast revision, and the use of the median
consensus analysts' earnings forecast as the relevant statistic.
By construction, the potential effect of a dividend announcement on analysts'
earnings forecasts cannot be easily isolated in the forecast revision response variable.
Earnings forecasts are reported once a month, and their changes are likely to reflect
all the relevant information released over the prior one-month period. In the event
study methodology, researchers typically use stock price changes over small windows
of two to three days around announcement days, and it is relatively easier to isolate
the information effect of the dividend policy announcement.
The observed analysts' earnings forecast revisions after dividend policy changes
are likely to be affected by the dividend change, but also by a number of other,
contemporaneous information sources, related to firm-specific economic and financial
factors as well as industry and economy-wide considerations. The release of the
I/B/E/S Inc. consensus earnings forecasts once a month makes the study of the
informational effect of any one of these sources difficult .
The ideal earnings forecasts ' data set would be comprised of individual analysts'
earnings forecasts on the day they are released. Then the researcher would perhaps
be able to isolate better, and assess individually, the potential effect of different
information sources. In addition, it would become relatively easier to determine the
23
impact of the information released between specific-date releases of forecasts. Such
a database, though, is not available.
The three studies on forecast revisions presented above measure the forecast re
vision over two months, from one month before to one month after the dividend
announcement month. This, according to Lang and Litzenberger [47], is because
I/B/E/S Inc. does not specify when the forecasts are issued, and therefore it can
not be determined that the consensus earnings forecast released in the dividend
announcement month does not include the impact of the dividend information.
6
If the revision were to be calculated from the month of the dividend announcement
to the next month, according to [4 7], it is possible that the dividend announcement
month earnings forecast will already have incorporated the dividend signal, if it were
announced early enough during that month, thus biasing downward the measure of
the forecast revision.
It is possible that assessing the forecast revision over two months exarcebates the
problem of separating the information signals and isolating the effects of each poten
tial information source on the forecast revision, especially if and when important and
informative events have occurred around the beginning of the dividend announce
ment month. Unfortunately, since the release date of the consensus forecasts is not
known, it is not possible to test whether this is true.
I do the next best thing. In particular, I examine measures of the forecast re
vision after the dividend announcement which are assessed over both ( a) the one
month from the dividend announcement month to the month following the dividend
announcement month, and (b) the two-month period from one month before the div
idend announcement month to the month after the dividend announcement month.
I find that the period over which the revision is measured affects the significance of
the documented forecast revision in only a few cases.
6
That would be the case, for example, if the dividend announcement occurred on May 5, and
the consensus earnings forecast includes individual analysts' forecasts released between May 6 and
May 25.
24
Finally, the use of the median consensus earnings forecast as the measure of the
earnings forecast in studies of the effect of dividend changes on analysts' earnings
forecasts may tend to underestimate the potential forecast revision. The three papers
presented above say that the median consensus forecast is the appropriate measure
because it mutes the effect of outlier forecasts. The argument can be made that,
exactly because of that, the median forecast may not be the appropriate statistic.
For the sake of completeness, I use both the mean and the median consensus earnings
forecast in my measures of the forecast revision.
Analysts revise their forecasts at a relatively leisurely pace. It has been docu
mented (see Brous [12]) that only about an average of 20% of forecasts are revised
in any given month, for each company on the 1/B/E/S Inc. database.
7
The
forecasts that are revised are likely to reflect current information. Forecasts that
remain unchanged can result from analysts either being late in updating and report
ing their forecasts, or from the information signal not being considered material and
warranting a forecast revision.
In either case, using the median earnings forecast may, on average, downplay the
potential information effect of the dividend change announcement, since the median
statistic is robust to "outliers", which in this case are likely to be the 20forecasts,
and it may dilute their effect on the consensus forecast. The study of the informa
tional effect of dividend changes on analysts' forecasts of earnings requires the use
of a response variable for the forecast revision which is sensitive to the release of
new information. The mean consensus earnings forecast may account for the magni
tude of forecast revisions after a dividend change better that the median consensus
earnings forecast because it is more likely to change when only a relatively small
percentage of analysts report a revised forecast.
As in the case of the one-month versus two-month forecast revision measurement
period, there is no apparent method to examine whether the mean consensus forecast
7
It is likely that the percentage of analysts releasing a revised forecast after an earnings or
a dividend announcement will be higher than the average 20% of all the analysts that follow a
particular company, which was documented in Brous [12]. Nevertheless, I do not have access to
this information.
25
revision is a superior statistic than its median counterpart. The reason is that my
database does not provide information on the individual forecasts which comprise
the reported consensus forecast.
One of the contributions of my study is addressing these issues in a comprehen
sive manner. In particular, I will examine the effect of announcements of dividend
changes on analysts ' earnings forecasts conditional on the presence of relevant con
temporaneous information sources such as quarterly earnings announcements. In
addition, I will measure the forecast revision both over one and two months , and
using both the mean and the median consensus analysts ' earnings forecast.
1.3.3 The Ofer and Siegel study
In a widely cited study on the impact of dividend announcements on analysts' earn ings forecasts , Ofer and Siegel [63] examine 781 announcements of dividend changes,
between 1976 and 1984, which survived a number of screening criteria. The firms
changed their regular, taxable, dollar dividend ( CRSP 1232 distribution code) by
more than 10% or less than -10% after at least three quarters of unchanged dividends.
In addition, there were no earnings announcements, cash or non-cash distributions,
or stories about the firm in the Wall Street Journal within a 12-day window, from
six days before to six days after the dividend announcement.
The Ofer and Siegel study is not directly comparable with the issues and method
ology of my dissertation for two main reasons:
• As I discuss next, they examme the behavior of forecast errors around the
dividend announcement month, as a proxy for analysts' forecast revisions,
whereas I use the actual forecast revision to assess the potential information
content of the dividend change.
• The authors have not addressed the issue of the impact of other contempora
neous information on the forecast revision.
26
Nevertheless, the Ofer and Siegel study has been widely cited, because it was
the first paper whose main topic was to address the issue of the impact of dividend
changes on analysts' earnings forecasts. Therefore, although the comparisons that
can be drawn between their study and mine, are only incidental ( e.g. both stud
ies examine earnings forecasts after announcements of dividend changes, but using
vastly different methodologies which make the respective results non-comparable, as
I discuss below), I present the Ofer and Siegel study here for the sake of complete
ness of presentation. The Lang and Litzenberger, Yoon and Starks, and Denis et.
al. studies examined forecast revisions after dividend changes in a related, but not
direct, context.
Ofer and Siegel assess the statistical relationship between the unexpected compo
nent of the dividend change and the level of analysts' earnings forecast errors around
the dividend announcement month, using instrumental variables . This is their test
of the potential informational impact of dividend changes on analysts' expectations
of annual earnings per share. They first separate the unexpected component of the
dividend announcement by regressing the dollar dividend change on the dollar stock
price change in a 3-day window around the day of the dividend announcement:
where -6.Di is the dollar dividend change for firm-event i (i = 1, . .. , 781), -6.Pi,(t- l ,t+l)
is the stock price change from the day before to the day after the dividend announce
ment for firm-event i, and Ei is the residual. The reported coefficient estimate /31
is 0.0133 , and the unadjusted R
2
is 0.097. The unexpected component of a given
dividend change (the "dividend surprise instrument" ), is the term /31 -6.Pi,(t-l ,t+l),
for each of the 781 observations.
In the second stage of their analysis , Ofer and Siegel apply six separate univariate
regression models . The dependent variable in each successive regression model is the
forecast error in a given month surrounding the dividend announcement month, and
it changes in each successive regression. In the first regression the dependent variable
is the forecast error in the month before the dividend announcement month. In the
27
second regression the dependent variable is the forecast error in the month of the
dividend announcement. And so on, up to the sixth regression whose dependent
variable is the forecast error in the fourth month after the dividend announcement
month.
The explanatory variable in all six regressions is the dividend surprise instrument
from the first-stage regression, $ 1~Pi,(t-l,t+l)· In other words, for each month, from
one month before the dividend announcement month to four months after it, they
separately fit the cross-sectional regression model:
where F Ei,m is the consensus analysts' earnings forecast error for firm-event i ( i =
1, ... , 781) in month m, defined as actual annual earnings per share minus the me
dian analysts' consensus earnings forecast in the corresponding month. This vari
able is indexed from the month before the dividend announcement month, to four
months after the dividend announcement month. $ 1~Pi,(t-l,t+l) is the dividend sur
prise instrument from the first stage regression for firm-event i. Finally, ei,m is a
heteroskedastic error term, corrected according to White's method, and indexed as
above.
Ofer and Siegel use this methodology as a test of the impact of dividend changes
on analysts' earnings forecasts. Their hypothesis is that if the dividend change
contains previously unknown information about current annual earnings then the
level of the forecast error in the months before the dividend announcement will be
systematically related to that information. The magnitude of the forecast error in
the prior months should be proportional to the information contained in the dividend
announcement.
After the dividend announcement, according to Ofer and Siegel (p. 897, par.1),
"the consensus forecast will include an increasing proportion of [individual analysts']
forecasts that fully incorporate the information conveyed by the dividend announce
ment". They say that the empirical implication of this hypothesis is that the coeffi
cient estimate of the dividend surprise variable should monotonically decline in each
28
successive month after the dividend announcement, i.e. the systematic relationship
between the severity of the forecast error and the informativeness of the dividend
surprise should weaken.
The coefficient estimate of the unique dividend surprise variable in the regression
of the forecast error in the month before the dividend announcement is significant
and positive (12.469), and it declines in each successive regression after the dividend
announcement month (from 9.06 in the dividend announcement month forecast error
regression, to 4.306 in the regression of the forecast error in the fourth month after
the dividend announcement month) .
Like the other papers discussed in the previous section, the Ofer and Siegel study
does not account for contemporaneous information sources like earnings announce
ments around the dividend announcement . The authors screen their sample for an
nouncements of other information items, such as quarterly earnings announcements,
and Wall Street Journal stories, from 6 days before to 6 days after the dividend an
nouncement. This is aimed at separating the stock market price reaction around the
dividend announcement day from other information, given that the price reaction is
used as the first-stage instrument for the dividend surprise.
This separation of the dividend announcement from other information events in
a 12-day window around the announcement day does not address the issue of the
marginal information impact of the dividend announcement on analysts' earnings
forecasts in the presence of other information sources, such as earnings announce
ments. Ofer and Siegel do not account for the influence of such other sources on
analysts' earnings forecasts, only on the price reaction. Quarterly earnings announce ments which could have occurred outside the 12-day window, but still in the month
of the dividend change ( especially in the beginning of the month) are likely to influ
ence analysts' revisions of earnings forecasts, and the level of the forecast error in
subsequent months.
A possible method of controlling for such earnings announcements in the Ofer
and Siegel tests would be to include an earnings surprise variable in the second
stage estimation of the 6 regression models, along the lines of the dividend surprise
29
variable. Another approach could involve the inclusion of a dummy variable that
takes the value 1 if the dividend announcement occured within a 40-trading day
window around the nearest earnings announcement. My study uses such control
variables in linear regression models of the forecast revision.
The second issue with respect to the Ofer and Siegel study is related to the use
of the forecast error, and the instrumental variables methodology that they apply.
Ofer and Siegel say that their methodology tests the updating of market expectations
after announcements of dividend changes through the revision of analysts' earnings
forecasts, and that they provide evidence on analysts' forecast revisions. In the
abstract (p.889) they say:
" ... ; the a'Uthor-s prnvide evidence that analysts r·evise their· ear-nings for·e casts following the anno'Uncement of an 'Unexpected dividend change by an
amo'Unt positively r-elated to the size of the 'Unexpected dividend change.
They also prnvide evidence that these revisions ar-e positively related to
the change in eq'Uity value s'ur-r·o'unding the anno'Uncement. Pur·ther·; they
find that these r·evisions ar·e consistent with mtionality ... "
In the discussion of the second stage estimation of their model, Ofer and Siegel write
(Section B.l, p.901):
"These r-es'Ults indicate that; following the anno'Uncement of an 'Unex
pected change in dividends; analysts revise th eir- e:r;pectations in a manner
consistent with mtionality. "
In addition, they say (p. 891) that their tests provide direct evidence that "mar·
ket par·ticipants actually do r·evise their· expectations following [ announcements of
dividend policy changes]" .
As discussed in the previous section, the authors interpret the monotonic decline
of the unique dividend surprise instrument coefficient in each successive regression, as
evidence of analysts revising their forecasts after a dividend change announcement .
The apparent rationale behind their methodology is the belief that documenting a
decline in the dividend surprise variable coefficient after the dividend announcement
30
is equivalent to documenting that analysts revise their earnings forecasts after such
an event. But their tests do not assess forecast revisions.
The decline of the dividend surprise instrument coefficient estimate in each suc
cessive regression in the second stage of the Ofer and Siegel model may be attributed
to the intra-year secular decline in forecast errors documented in the literature. The
empirical evidence in the literature on forecast errors (see Brown et al. [14], and
Elton et al. [34]) , and the intra-year behavior of the forecast errors in the sample
I use in this study provide support to this explanation of the results reported by
Ofer and Siegel. Forecast errors on average decline during the fiscal year, in general
samples without the explicit presence of dividend changes.
To investigate the validity of the previous arguments, first I examine the intra
year behavior of forecast errors in my sample, and then I apply the methodology of
Ofer and Siegel on my data to confirm that my sample is comparable to theirs, and
therefore their findings can be replicated in my sample, which overlaps with theirs in
6 of its 9 years . I document that the forecast errors of the 909 sample firms decline
from the beginning to the end of the fiscal year. This regularity, furthermore, holds
not only in the 2,866 event-years of the sample, but also in all the years, between
1979 and 1990, in which there were no dividend changes of more than ±10%. These
results are presented and discussed in Chapter 3.
To replicate the Ofer and Siegel methodology on my sample, I use the 1,531
dividend announcements which were made at least four months before the fiscal
year end, to conform with their design. Then, I fit their model on my data. The
first stage regression for the extraction of the dividend surprise instrument yields
similar coefficient estimates as theirs: the intercept is 0.1925 (in Ofer and Siegel ( O
S) it is 0.0330), and the coefficient of the dollar stock price change is 0.0181 (O-S:
0.0133).
The coefficient estimates of the six second-stage regressions of the forecast errors
in each of the six months around the dividend announcement on the unique dividend
surprise instrument, also yield similar results. The coefficient in the month before
the dividend announcement is 9.61 (O-S: 12.47), in the month of the announcement
31
9.42 (O-S: 9.06), and in the four months after the announcement the coefficient is
8.68 (O-S:7.48) , 7.97 (O-S: 6.67), 7.53 (O-S: 5.27), and 6.82 (O-S: 4.30) , respectively.
The t-statistics are all significant, and range from around 5 to around 4.
To summarize, assessing the informational content of dividend changes through
the behavior of the level of the forecast error in the months around the dividend
announcement month, in the manner applied in Ofer and Siegel [63] may not be
entirely succesful for the following reasons:
• Information sources other than the dividend announcement are likely to have
an effect on analysts' earnings forecasts, and any examination of the effect
of dividend announcements on analysts' earnings forecasts should control for
such other sources.
• A direct test of the informational impact of the dividend change should be the
examination of forecast revisions themselves.
• Forecast errors decline during the year, in samples drawn independently of div
idend changes, and therefore linking the forecast error decline to the dividend
change is, at best, ambiguous.
32
Chapter 2
The sample of dividend changes and earnings
forecasts
2 .1 Screening criteria
My sample of 2,866 dividend changes, of more than 10% or less than -10%, announced
between 1979 and 1990, announced by firms listed either on the New York Stock
Exchange or the American Stock Exchange, was compiled from four main sources.
• The dividend dates and amounts, and the price data come from the 1991
version ( cutoff date: 12/31/1990) of the daily NYSE/ AMEX tape of the Center
for Research in Security Prices (CRSP), of the Graduate School of Business at
the University of Chicago.
• The annual earnings per share and the quarterly net income figures come from
the Primary, Secondary, and Tertiary and the Research Industrial tapes of the
COMPUSTAT service of Standard and Poor's.
• The analysts' earnings consensus forecast data were extracted from the His
torical Tape of 1/B/E/S International Inc.
• Finally, the dividend data were checked against the Wall Street Journal Index,
for each of the years in the period 1979-1990.
33
Dividend changes, from the CRSP database, were included in the sample if they
satisfied the following criteria:
1. The dividend announced is regular quarterly, payable in US dollars ( CRSP
distribution code 1232).
2. The announcement date is later than January 1, 1979.
3. The firm has not omitted a dividend in any of the previous three quarters.
4. The changed dividend is strictly positive, i.e. the change is not from a positive
dividend to an omitted one.
5. The firm has paid only quarterly dividends after January 1, 1979. This ex cludes semi-annual and annual dividends, but not special and extraordinary
distributions.
6. The firm has distributed at least 3 regular quarterly dividends (CRSP distri
bution code 1232) before the change.
7. The previous three dividends were unchanged.
8. The change in the split-adjusted dividend is greater than 10% or less than
-10%.
9. If there were two or more dividend changes from the same firm in the same
fiscal year, none of the announcements is included in the sample.
10. The dividend announcement occurred in the third month of the fiscal year or
later.
11. The firm 1s not included in one of the following SIC classification codes:
4900-5000 (Utilities), 6720 (Investment Offices), 6711 (Holding Offices), 6794
(Patent Owners and Lessors), 6798 (REITs), 6799 (Investors), 6211 (Security
Brokers and Dealers) .
As shown in Table 3, this procedure yielded 4,675 data points, of which 4,224
were dividend increases, and 451 dividend cuts . The hurdle of a dividend change of
34
at least ±10% was imposed (a) in an attempt to derive a sample of truly informative,
i.e. substantial, dividend changes, and (b) because it has become common practice
in the dividend literature (see, for example, [4 7], [63]) to use such a lower limit on
the percentage dividend change.
The I/B/E/S Inc. historical earnings forecast database was then searched for the
full year of earnings forecasts for each of these 4,675 firm-events. The I/B/E/S Inc.
database version that I used contains consensus earnings forecasts, for every month
from 1974 to 1992, released by the analysts of most major Wall Street investment
houses. The database contains two kinds of consensus earnings forecasts: the mean
of all the analysts' forecasts for each company in the particular month (from now
on the "mean consensus earnings forecast"), and the median of the individual ana
lysts' earnings forecasts for the particular month ( the "median consensus earnings
forecast").
This screening caused the elimination of 1,407 data points for one of the following
two reasons:
1. The I/B/E/S Inc. tape does not include forecasts for the firm-event year
because it does not follow the firm concerned.
2. The firm is included in the tape but there is no forecast listed for one or more
of the three months around the dividend announcement .
The remaining 3,268 observations were matched with the COMPUSTAT Quar
terly Primary, Secondary, and Tertiary industrial file , and the COMPUSTAT Quar
terly Research industrial file, for the quarterly Earnings per Share figure in the
quarter of the dividend announcement and the fiscal quarter one year before the
dividend announcement quarter. This caused the elimination of 77 more datapoints
because COMPUSTAT does not contain information for these firms.
Finally, all 3,191 firm-events which survived the COMPUSTAT screening were
checked against the Wall Street Journal Index. This screening ensured that:
35
Table 2.1: Sample selection process
This table presents the selection process of the sample of 2,866 announcements of a dividend change of more than
10% or less than -10%, between 1979-90. A total of 909 firms are represented in the sample with one or more
announcements of a dividend change. A firm-event is defined as the occurrence of an announcement of a dividend
change (increase or cut) of more than 10% or less -10% for each firm. Thus, a particular firm may be responsible
for one firm-event in the final sample, if it has only one dividend change included, or for many, if the selection
criteria picked numerous dividend changes of this firm. A dividend change event is included in the sample if it was
announced after 1979 by a firm which is not a utility and it is not included in the SIC categories 6211, 6711, 6720,
6794, 6798, 6799, and which has distributed only regular quarterly dividends after 1979.
Number of firms in CRSP Master tape 5968
Number of announcements of regular quarterly dividends 111,879
Number of announcements of regular quarterly dividends
excluding utilities and funds, trusts, etc. 99,013
All Increases Decreases
Dividend changes of more than 10% (Not utilities
or SIC 6211, 6711, 6720, 6794, 6798, 6799, with the
last distribution after 1979 and only regular
quarterly dividends after 1979) 4,675 4,224 451
Minus firm-events excluded due to the firm not being
listed in 1/B/E/S Inc. History Tape, or the three
months around the announcement month not being listed 1,407 1,230 177
Minus firm-events excluded because they are not
carried by the COMPUSTAT Annual Industrial Tape 77 51 26
Minus firm-events excluded because the Wall Street
Journal Index announcement date is more than four
days away from the CRSP announcement date 286 249 37
Minus firm-events excluded because the Wall Street
Journal carried stories of an imminent dividend change
months before it occured 39 20 19
Total firm-events excluded 1,809 1,550 259
Final Sample 2,866 2,674 192
36
1. The dividend announcement date listed by CRSP agrees with the publication
day of the announcement in the Wall Street Journal, i.e. the Wall Street Jour
nal publication date is the business day after the CRSP dividend announcement
day.
2. The dividend change was not pre-announced, days or months in advance, by
either management, the Board of Directors, or analysts.
In addition, all dividend announcements were thus classified with respect to
their timing compared to the nearest earnings announcement as recorded by the
Wall Street Journal: dividend announcements occurred either within 40 trading
days around the nearest earnings announcements, or outside of such a 40-trading
day window.
This screening against the Wall Street Journal caused the elimination of 325
observations. Of those, 286 observations were dropped because of a difference of 4
or more trading days in the exact announcement date, between CRSP and the Wall
Street Journal Index. The remaining 39 observations were dropped because officers
of the company and/ or the Board of Directors were quoted in reports months in
advance to predict a dividend change in the coming quarter( s). The final sample
consists of 2,866 observations, from 909 firms that announced a dividend change of
more than 10% or less than -10% between 1979 and 1990.
2.2 Sample description
Tables 4 and 5 contain the descriptive statistics and subgroup breakdowns of the
sample. Panel A of Table 4 presents the statistics of the full sample of 2,866 div
idend changes. The mean dollar dividend change (Column 1) is 20 cents (median
10%), and the mean percentage change 17% (median 17%).
1
The average stock
1
In the Ofer and Siegel sample, which also contains dividend changes of more than 10% or less
than -10%, the mean dollar dividend change is 3.6 cents. This difference is most likely due to
the different sizes of the two samples, and also to the fact that Ofer and Siegel's sample timing is
37
price change from two days before the dividend announcement, to the day after the
announcement, is 28 cents.
2
The mean excess stock return in the 3-day window
around the dividend announcement day is 1% (median: 1%). The mean excess
return from 6 to 2 days before the dividend announcement day is 0.3% (median:
0%).
Columns 6 to 8 of Table 4 present the descriptive statistics of the sample firms'
earnings performance in the fiscal quarter of the dividend announcement. Specif
ically, column 6 presents the Net Income Before Extraordinary Items (Quarterly
Industrial COMPUSTAT Item 8) measured in millions of dollars, in (a) the quar
ter of the dividend change, if the announcement occurred within a 40-trading day
window around the nearest earnings announcement, i.e. in a period from 20 trading
days before to 20 trading days after the nearest earnings announcement, or (b) the
previous fiscal quarter, if the dividend change was announced outside such a window.
Column 7 presents the Net Income figure in the fourth fiscal quarter before the
quarter reported in column 6. The last column presents the growth rate of quar
terly Net Income Before Extraordinary Items between the quarter of the dividend
announcement and the same quarter of the previous fiscal year, and is calculated
as the raw change in Net Income between the two quarters, divided by the income
figure of the prior year's quarter.
3
The mean growth rate between the dividend announcement fiscal quarter and
the quarterly income one year before was 15% and the median rate is 18%. This
growth is statistically significantly different from zero with t statistic equal to 2.23.
The dispersion of the growth rates is large. It ranges from 5,587% for the fastest
growing firm, to -9,514% for the worst performer.
centered around the recession of 1979, whereas my sample spans the economic growth years of the
1980's. The dividend increases in my sample are likely to be of higher magnitude than those in
Ofer and Siegel [63].
2
This is the stock market reaction variable that Ofer and Siegel use in their study. Its mean in
their sample is 20.3 cents.
3
When the current or the year-ago quarter income figure is negative, the growth rate is computed
such that it reflects the loss.
38
Table 2.2: Sample descriptive statistics
This table presents the descriptive statistics of the sample of 2,866 announcements of dividend changes of more than
±10%. The first variable is the dollar dividend change between the event month and the previous quarter. The
second variable is the percentage change of the dividend compared to the dividend in the previous quarter. The
third variable is the change in the firm's stock price from two days before the announcement day to one day after the
announcement day. The next two variables are the market-adjusted return in the periods from day t-6 to day t-2,
and from one day before the dividend announcement to one day after the announcement day, respectively. The last
three variables are: Net fucome Before Extraordinary Items (Quarterly fudustrial COMPUSTAT Item 8) in (a) the
quarter of the dividend change,if the dividend announcement occured within a 20-trading day window around the
dividend announcement,or (b) the previous quarter if the dividend announcement occured outside such a window
(Column 6), Net fucome Before Extraordinary Items four quarters before (Column 7), and the percentage change
in Net Income between q and q-4 (Column 8). Panel A contains the statistics for the full sample. Panel B contains
the statistics for the 192 dividend cuts, and Panel C contains the statistics for the 2,674 dividend increases. Finally,
Panel D presents the t and Z values for two-sample tests for differences between the means of the dividend cuts
versus dividend increases statistics.
Lagged Percentage
Dollar Percentage Price Return Return Quarterly Quarterly Quarterly
dividend dividend change between between Net Net Income
change change t-2,t+l t-6,t-2 t-1,t+l fucome fucome change
(1) (2) (3) (4) (5) (6) (7) (8)
A. Full Sample (N=2866)
Mean 0.20 17 0.28 0.30 1.00 28.04 31.71 15.12
Median 0.10 17 0.13 0.00 1.00 8.11 9.27 18.34
a 0.63 23 1.70 4.00 4.00 71.48 82.71 354.3
t-stat 16.66 40.07 8.84 3.12 7.30 20.69 20.22 2.23
Maximum 16.56 150 12.50 31.0 19 1,371.25 1,645.50 5,587.1
Minimum -4.80 -91 -24.13 -3.00 -37.00 -185. 75 -367.40 -9,514.4
B. Dividend cuts (N=192)
Mean -0.48 -47 -0.70 -2.00 -5.00 12.33 -9.56 -203.17
Median -0.26 -50 -0.50 -2.00 -4.00 2.36 -0.45 -103.33
a 0.61 17 1.56 6.00 7.00 62.23 43.79 503.20
t-stat -11.06 -39.09 -6.27 -3.83 -9.85 2.69 -2.96 -5.45
C. Dividend increases (N=2674)
Mean 0.24 22 0.35 0.00 1.00 29.15 34.63 30.14
Median 0.11 17 0.25 0.00 1.00 8.45 10.39 19.26
a 0.60 15 1.69 4.00 4.00 71.97 84.04 336.39
t-stat 21.03 75.41 10.78 4.68 13.40 10.64 21.00 4.61
D. Two-sample parametric t and Wilcoxon Z tests for differences between dividend cuts and increases
2-S t -16.24 -61.35 -8.43 -4.69 -11. 71 -3.42 -12.07 -6.06
Wile Z -23.24 -23.27 -10.70 -5.12 -13.07 -8.37 -17.25 -15.84
39
The sample is split into dividend cuts and dividend increases, in Panels B and
C. The mean dollar dividend cut is of a magnitude of 48 cents, or -47% (medians
of -26 cents, or -50%).
4
The mean price change in the three days around the
announcement of the dividend cut is -70 cents, or -5%. These returns are consistent
with the empirical regularities established in the literature on stock price reactions
around dividend announcements (see [18], [47], [50], etc.). The mean excess return
from day 6 to day 2 before the announcement day is -2% (median: -1.5%).
The mean earnings growth rate for dividend cutters is -203% and the median
growth rate is -103%. In other words, dividend cutters saw their earnings plunge
to losses, on average. The mean earnings reported in the quarter of the dividend
announcement is -$9.56 million, compared to an average of $12.33 in the same quarter
of the previous year. The growth rate is statistically significant with a t statistic
of -5.45. This earnings performance of dividend cutting firms is consistent with the
findings of DeAngelo, DeAngelo and Skinner [24] who report that an annual loss is
a necessary condition for a firm to decide to reduce its dividend.
Firms that increase their dividend (Panel C), on the other hand, do so by an
average of 24 cents, or 22%, and experience stock price increases of an average of
35 cents, or 1 %, in the three days around the dividend announcement. The sample
is consistent with prior empirical evidence on the asymmetric stock price reactions
around dividend announcements. Dividend increasing firms' announced earnings are
higher than the same quarter of the previous year by an average of 30% (median of
19%).
In Table 5, the sample is examined with respect to possible concentrations in
particular years, months, and industries. Panel A presents the breakdown of the
sample into year of occurrence of the dividend announcement. There are no sig
nificant differences across the twelve years, with the possible exception of 1979, in
which there are 410 announcements of a dividend change, and 1990 that witnessed
only 149 such events.
4
The relatively large size of the average dividend cut is due not only to the cutoff point of 10%
imposed in the sample construction, but also to the empirical regularity ( discussed in Chapter 1)
of firms deciding to cut regular dividends only under severe distress, which leads to large cuts.
40
Table 2.3: Year, month, and industry sample breakdown
This table presents breakdowns of the sample of 2,866 firm-years of a dividend change of more than 10% or less
than -10%, from 1979 to 1990. The first panel breaks down the sample into calendar year of occurrence of the
dividend announcements. The calendar year need not necessarily coincide with the fiscal year of the announcement
(for example, an April 1985 announcement of a firm with fiscal year ending in May, would have occured in the 1984
fiscal year, but in the 1985 calendar year). The second panel shows the month-of-the-year clustering of the dividend
announcements, by size, and dividend increases vs. cuts. The designation of a firm-event in one of the three size
portfolios (Large, Medium, Small) are per CRSP, which assigns firms to ten market capitalization portfolios as of
the last trading day of the previous calendar year. Firm-events in the three largest CRSP size portfolios (10, 9, and
8) are assigned in the Large size portfolio, firm-events in CRSP portfolios 7, 6, 5, and 4 are assigned to the Medium
size portfolio, and firm-events in the smallest three CRSP capitalization portfolios (1, 2, and 3) are assigned to
the Small size portfolio. The third panel, in the next page, provides the sample breakdown in terms of industries
represented in the final sample. The industry designation is per CRSP.
Panel A. Sample breakdown into calendar year of occurrence of the dividend announcement.
1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990
410 320 259 215 236 236 187 185 209 236 224 149
Panel B. Sample breakdown into dividend increases versus cuts, and firm size in the year of the dividend announce
ment, and into the month of the fiscal year the dividend change was announced.
3 4 5 6 7 8 9 10 11
A. Increases 191 270 303 194 334 537 264 503 71
B. Cuts 17 31 32 15 24 32 19 15 1
E. Large firms 108 134 181 118 192 318 168 335 43
F. Medium firms 91 149 139 86 148 232 102 168 29
. Small firms 9 19 17 5 18 18 13 14 2
41
Table 2.4: Year, month, and industry sample breakdown ( continued)
This table concludes the presentation of the breakdowns of the sample of 2,866 firm-years of a dividend change of
more than 10% or less than -10%, from 1979 to 1990. The third panel, contained in this page, provides the sample
breakdown in terms of industries represented in the final sample. The industry designation is per CRSP.
Panel C. Sample breakdown by 2-digit SIC industry codes
Industry Industry Number of Industry Industry Number of
Code Name observations Code Name observations
10 Metal-gold mining 18 44 Water transportation 1
12 Coal-Lignite 12 45 Airlines 15
13 Oil drilling 85 47 Transport services 6
14 Non-mineral mining 4 48 Telephone 24
15 Builders 26 50 Durable goods-wholesale 60
16 Heavy construction 15 51 roceries-whollsale 53
17 Special construction 1 52 Building materials-retail 10
20 Foods-beverages 138 53 Retail stores 75
21 Tobacco 14 54 roceries-retail 48
22 Textiles 54 56 Apparel stores 20
23 Apparel 70 57 Home furniture stores 22
24 Lumber-wood fabrication 26 58 Restaurants 30
25 Furniture 18 59 Misc retail 61
26 Paper products 87 60 Banking 79
27 Printing-publishing 109 61 Financial services 32
28 Chemicals-Drugs 258 63 Insurance 41
29 Oil-integrated 44 64 Insurance agents 9
30 Tires and rubber 49 65 Homebuilding 10
31 Shoes 19 67 Other financials 17
32 lass containers 69 70 Hotels, motels 15
33 Furnaces-Aluminum 104 72 Personal svcs 14
34 Hardware 134 73 Data processing 80
35 Computers 203 75 Auto repair 13
36 Electronics 195 76 Misc repair services 7
37 Transport vehicles 104 78 Motion pictures 16
38 Toys-Leisure 97 79 Amusement 2
39 Jewelry, etc. 43 80 Health svcs 42
40 Railroads 17 82 Educational services 9
41 Passenger transportation 1 89 Architecture services 20
42 Trucking 15
42
Panel B presents the sample breakdown into the month of the fiscal year m
which the dividend change was announced, and into ( a) dividend increases versus
dividend cu ts, and (b) firm size in the year of the dividend change, determined by
the CRSP-assigned size portfolio the firm was in in the year of the dividend change.
Months 3, 6 and 9 have less dividend activity than the other months of the
year. This is to be expected, given that the vast majority of the sample firms have
December 31 fiscal year ends, and they tend to announce quarterly earnings one or
two months after the end of the fiscal quarter (i.e. in months 4 or 5, for the first
quarter, in months 7 or 8, for the second quarter, in months 10 or 11 for the third
quarter, and in January or February of the next fiscal year for the last quarter) .
Finally, Panel C presents the industry concentrations in the sample of 2,866
firm-events , according to the 2-digit SIC code. The largest group belongs to the
Chemicals-Drugs industry (SIC 28), with 258 observations, and the lowest to Spe
cial Construction (SIC 17), Passenger Transportation (SIC 41), and Water Trans
portation (SIC 44), with 1 observation each. Heavily represented industries include
Computers (203 observations) , Electronics (195 observations), Foods-Beverages (138
observations), and Hardware (134 observations).
43
Chapter 3
The secular revision activity
In this chapter I examme the intra-year secular forecast revision activity in my
sample of 2,866 dividend changes between 1979 and 1990. I conduct this analysis in
order to establish the need for adjusting for the intra-year forecast revision activity
when assessing the forecast revision after announcements of dividend changes. In
particular, I examine the intra-year behavior of forecast revisions, and the intra-year
behavior of forecast errors. First, I review the literature on the intra-year behavior of
forecast errors . Since there is no published research on analysts' intra-year forecast
revision behavior, to my knowledge, I do not conduct a literature review on this
subject. Then, I present a detailed examination of the forecast revision behavior
and the forecast error behavior in my sample of dividend changes.
3.1 The intra-year decline of forecast errors
Researchers have examined the behavior of forecast errors from the first to the last
month of the year, i.e. the intra-year evolution of forecast accuracy. This accuracy
has been assessed both relative to other predictive models of earnings, and by itself,
i.e. by assessing the intra-year evolution of the level of the forecast error.
The magnitude of the intra-year forecast revision activity per· se has not been the
subject of published research. Nevertheless , the forecast error behavior is indicative
of the forecast revision behavior since the relative, intra-year change of the forecast
44
error is exactly equal to the forecast revision. Note that the difference between the
forecast error in the end of the fiscal year and the forecast error in the beginning of
the year is, by definition, equal to the intra-year forecast revision.
The first group of papers on the behavior of forecast errors have focused on the
relative predictive power of analysts' earnings forecasts when compared to a number
of statistical forecasting models of earnings. Crichfield, Dyckman, and Lakonishok
[20] , examine forecasts of earnings of 46 firms for the period 1967-76, published
in Standard & Poor's Ear-nings For·ecaster·. They compare the errors of analysts'
forecasts against four time series and additive earnings forecast models. They report
that the accuracy of the forecasts improves , compared to the models they use as
alternative strategies, as the fiscal year progresses. This indicates that forecast
errors in their sample change throughout the year, as they become more accurate.
Other papers with similar methodology and results include Cragg and Malkiel [22],
Brown and Roze-ff [16], and Elton, Gruber, and Gultekin [34].
The second line of research has investigated the intra-year behavior of analysts'
earnings forecasts independently of a benchmark, such as other models of earnings
forecasts. These papers examine whether analysts' earnings forecasts become more
accurate predictors of annual earnings as the year progresses. Elton, Gruber and
Gultekin [34] regress the absolute value of the raw forecast error (i.e. actual year-end
EPS minus monthly forecast: IEP S-Fml) against time and find that errors decrease
by an average of 3.5% a month, for the approximately 1,200 earnings forecasts in
their 3-year sample. Brown, Hagerman, Griffin, and Zmijewski [14] use the relative
forecast error ( defined as the raw forecast error divided by the absolute value of
earnings per share, (EPS - Fm) + IEPS I), for 233 firms over 24 quarters, and
document that errors decrease in successive fiscal quarters , due to analysts using
contemporaneous information to update their forecasts.
All the above papers document that forecast errors become more accurate through
out the year, indicating that analysts, on average, revise their earnings forecasts from
the first to the last month of the year. This empirical regularity suggests that the as
sessment of the forecast revision after the announcement of dividend changes should
45
account for the secular forecast revision activity. If the measure of the post-dividend
announcement forecast revision fails to account for this secular revision activity it
should be expected to yield estimates of the true effect of the dividend signal on
analysts' expectations which contain the noise of the secular revision activity.
The measures of the forecast error used in this study draw from the methodology
presented in the two previous papers. I use the raw forecast error, and the absolute
value of the raw forecast error divided by the absolute value of annual EPS, as
measures of the forecast error.
3.2 The intra-year forecast revision
Table 6 presents the intra-year forecast revision activity in my sample of 2,866 event
years of a dividend change. The analysts whose earnings forecasts are included in
the sample revise significantly their expectations of year-end earnings throughout
the year. I examine the raw intra-year forecast revision, and the percentage intra
year forecast revision. The raw forecast revision is defined as the earnings forecast
in the last month of the fiscal year minus the earnings forecast in the first month
of the year. The percentage forecast revision is equal to the raw forecast rev1s1on
divided by the earnings forecast in the first month of the year.
Panel A of Table 6 presents the statistics of the full sample of intra-year forecast
revisions. Earnings forecasts are lower, on average, at the end of the year by 12
cents ( or 4% ), compared to the beginning of the year. The revision is statistically
significantly different from zero: the t-statistic is -5.82 for the raw and -3.87 for
the percentage forecast revision, and the Wilcoxon Z values are -12. 39 and -7 .01 ,
respectively. The medians of the two measures are 1 cent, and 1 %, respectively.
This is evidence of intra-year forecast revision activity in my sample, supported by
both the parametric and the non-parametric tests.
Panel B presents the statistics of the subsample of 192 dividend cuts , and Panel
C the statistics of the subsample of 2,674 dividend increases. In Panel D, I report the
46
Table 3.1: The intra-year earnings forecast rev1s10n activity
This table presents the descriptive statistics of the two measures of the intra-year forecast revision activity of my
sample of 2,866 dividend change event-years. The first variable, the raw intra-year forecast revision, is equal to the
raw change in the monthly earnings forecast between the first and the last month of the fiscal year. The second
variable is the percentage forecast revision, and it is equal to the raw forecast revision divided by the earnings
forecast in the first month of the fiscal year. Panel A presents the statistics for the full sample, Panels B and C for
the subsamples of dividend cuts and increases, and Panel D presents the two-sample test statistics for differences in
the means of the subsamples of dividend cuts and increases.
Raw forecast
revision
(1)
A. Full sample (N=2,866)
Mean -0.12
Median 0.01
a 1.09
t -5.82
Wile Z -12.39
Max 6.81
Min -18.99
B. Dividend cuts (N=192)
Mean -2.19
Median -1.35
a 2.84
t -10.51
Wile Z -12.10
C. Dividend increases (N=2,674)
Mean 0.03
Median 0.02
a 0.64
t 2.27
Wile Z 4.22
Percentage forecast
revision
(2)
-0.04
0.01
0.48
-3.87
-7.01
3.05
-8.68
-1.07
-0.79
1.20
-12.16
-12.72
0.04
0.01
0.27
7.35
9.12
D. Two-sample tests for differences between cuts and increases
Two-sample t
Wilcoxon Z
-10.70
-18.11
-12.57
-18.50
47
two-sample t and Wilcoxon Z statistics of the tests for differences between the two
subsamples. The means of the two subsamples are statistically significantly different
at the 5% significance level, with two-sample t-statistics of -10.70 and -12.57, and
Wilcoxon Z values of -18.11 and -18.50, for the raw and the percentage revision
measures , respectively.
The earnings forecasts of dividend cutting firms in my sample are revised down
ward by a highly significant $2.19, or -107% , from the first to last month of the year
(t-statistics of -10.51 and -12.16, and Wilcoxon Z values of -12.10 and -12.72, for the
two revision measures, respectively) . The median revision is -$1.35 or 79%.
The earnings forecasts of dividend increasing firms is revised by 3 cents, or 4%
(medians of 2 cents or 1 % ). The test statistics for differences from zero are also
significant: the t-statistics are 2.27 and 7.35, and the Wilcoxon Z values 4.22 and
9.12, for the raw and the percentage forecast revision measures, respectively.
In other words , firms that cut their dividend during the fiscal year cause analysts
to revise their year-end forecasts more actively than do firms which increased their
dividend. This is consistent with the empirical evidence in the literature that divi
dend cuts are surprising, rare, and traumatic corporate financial decisions. There
fore, analysts should be expected, on average, to be taken by surprise by dividend
cuts, and to revise their year-end earnings forecasts by more, in percentage terms ,
than after dividend increases.
The asymmetric revision behavior between dividend increases and cuts notwith
standing, the findings of Table 6 establish significant intra-year forecast revision
activity in my sample of 2,866 dividend changes . This regularity justifies the need
for a measure of the secular revision activity to adjust the statistic that I will use
for the forecast revision after a dividend announcement in my tests of Chapter 4.
48
3.3 The sample statistics of forecast errors
In this section I analyze the intra-year behavior of forecast errors in my sample, to
link my research with previous research. As mentioned above, the intra-year forecast
revision has not been a subject of interest in the literature, which has focused on
examinations of the forecast error. I analyze the intra-year forecast error behavior
in my sample to compare my -findings and sample characteristics with the empirical
regularities in the literature. The conclusions of the previous section do not change
here. This section presents summary statistics of the two measures of forecast errors
and discusses various issues concerning their statistical and economic significance.
In the next section I present a more detailed evaluation of the intra-year forecast
accuracy.
Tables 7 and 8 contain the full sample descriptive statistics of the measures of
the forecast error in my sample. I define the raw earnings forecast error, in each
month of a fiscal year for which 1/B/E/S Inc. provides a consensus forecast, as the
actual annual Earnings per Share before Extraordinary Items minus the 1/B/E/S
Inc. consensus forecast error in the month. This statistic has been used in, among
others, O'Brien [60, 61] , Elton, Gruber and Gultekin [34] , and Fried and Givoly [35].
The EPS figure used is the COMPUSTAT Annual Industrial tape Item 58, Earnings
Per Share Before Extraordinary Items.
I use Earnings per Share before Extraordinary Items as the statistic forecast by
analysts because it is the earnings statistic most often used in the literature, and in
forecast publications like the S&P Earnings Forecaster-. In addition, I assume that
investors care about, and corporations base their dividend decisions on permanent
earnings, consistent with the M & M ICD hypothesis and the empirical -findings of
De Angelo, De Angelo, and Skinner [24]. All the tests presented in this chapter were
also conducted using Earnings per Share Including Extraordinary Items (COMPUS
TAT Annual Industrial tape Item 57) , for completeness . The -findings do not change
in any significant way.
49
The analysts' earnings forecasts are the mean consensus forecasts contained in
the I/B/E/S Inc. History tape. The tests were also conducted using the median
consensus earnings forecast without any material change in the findings reported
next .
The second measure of forecast error is the absolute earnings forecast error which
is defined as the absolute value of the raw forecast error divided by the absolute value
of the annual Earnings per Share before Extraordinary Items. The absolute forecast
error is the measure most often used in the literature (see, for example, Richards
[66] , Barefield and Comiskey [8], Brown, Hagerman, Griffin, and Zmijewski [14],
Fried and Givoly [35], Elton, Gruber and Gultekin [34]). The absolute forecast error
measures the magnitude of the forecast error as a percentage of earnings , regardless
of the sign of the error, and is a measure of the severity of the forecast error.
3.3.1 The raw forecast errors
The descriptive statistics of the full-sample raw forecast errors are presented in Table
7. All analysts ' earnings forecasts used in computing the forecast error statistics are
the mean I/B/E/S Inc. consensus forecasts. All references to the "median forecast
error" in this section refer to the median of the sample of mean consensus forecasts
and their errors, not to the median consens'Us analysts' .forecast.
The first three columns in Table 7 contain the statistics for the raw forecast
error in the first, the dividend announcement, and the last months of the fiscal year.
Columns 4, 5 and 6 of Table 7 present the t statistics ( first line) and Wilcoxon
Z statistics ( second line) of the two-sample tests for differences between the mean
forecast errors in the first and the last month of the year (Column 4) , the first and
the dividend announcement month (Column 5), and the dividend announcement
month and the last month of the year (Column 6).
The mean raw forecast error rises from $0. 76 in the first month of the fiscal year,
to $0.78 in the month of the dividend announcement , to $0.87 on the last month of
50
Table 3.2: The intra-year raw forecast error behavior of the full sample
This table contains the descriptive statistics of the raw measure of the forecast error for the sample of 2,866 dividend
changes of more than ±10%, between 1979-1990. The first three columns present the raw forecast error defined as
the annual Earnings per Share Before Extraordinary Items (Compustat Item 58) minus the mean consensus 1/B/E/S
analysts' forecast for the corresponding month (l=first month of the fiscal year, D.A.M.=dividend announcement
month, 12=last month of the fiscal year). Column 4 contains the t and the Wilcoxon Z statistics for the two-sample
parametric and nonparametric tests, respectively, between the raw error in Month 1 and the raw error in Month
12. Column 5 contains the t and the Wilcoxon Z statistics for the two-sample parametric and nonparametric tests,
respectively, between the raw error in Month 1 and the raw error in the D.A.M. Column 6 contains the t and the
Wilcoxon Z statistics for the two-sample parametric and nonparametric tests, respectively, between the raw error
in the D.A.M. and the raw error in Month 12. The statistics reported are the mean, the median, the standard
deviation, and Student's t and Wilcoxon Z statistics for differences from zero.
Raw Raw Raw t&Z t & z t&Z
error error error stats stats stats
in in in for for for
Month 1 D.A.M. Month 12 1-12 1-D.A.M. D.A.M.-12
(1) (2) (3) (4) (5) (6)
Mean 0.76 0.78 0.87 -1.75 -0.28 -1.77
Median 0.67 0.56 0.49 -0.20 -1.00 -1.32
a 2.56 2.28 2.03
t 15.62 17.90 22.58
Wile Z 28.55 28.73 31.52
Max 46.65 46.29 46.24
Min -38.93 -32.12 -26.70
51
Table 3.3: The intra-year absolute forecast error behavior of the full sample
This table contains the descriptive statistics of the absolute measure of the forecast error for the sample of 2,866
dividend changes of more than ±10%, between 1979-1990, expressed as percent of the absolute value of annual
earnings per share. The first three columns present the absolute forecast error defined as the absolute value of the
raw forecast error (i.e. annual Earnings per Share Before Extraordinary Items (Compustat Item 58) minus the mean
consensus 1/B/E/S analysts' forecast for the corresponding month) divided by the absolute value of the annual EPS,
for the corresponding months (l=first month of the fiscal year, D.A.M.=dividend announcement month, 12=last
month of the fiscal year). Column 4 contains the t and the Wilcoxon Z statistics for the two-sample parametric
and nonparametric tests, respectively, between the absolute error in Month 1 and the absolute error in Month 12.
Column 5 contains the t and the Wilcoxon Z statistics for the two-sample parametric and nonparametric tests,
respectively, between the absolute error in Month 1 and the absolute error in the D.A.M. Column 6 contains the
t and the Wilcoxon Z statistics for the two-sample parametric and nonparametric tests, respectively, between the
absolute error in the D.A.M. and the absolute error in Month 12. The statistics reported are the mean, the median,
the standard deviation, and Student's t and Wilcoxon Z statistics for differences from zero.
Absolute Absolute Absolute t & z t & z t & z
error error error stats stats stats
in in in for for for
Month 1 D.A.M. Month 12 1-12 1-D.A.M. D.A.M.-12
(1) (2) (3) (4) (5) (6)
Mean 81.2 69.2 55.1 3.24 1.48 1.99
t 12.48 13.84 12.63
a 339.0 260.0 228.0
Max 7,675 5600.0 5833.0
Min 1.0 1.0 1.0
52
the fiscal year. The median raw forecast error drops from $0.67 in the first month of
the fiscal year, to $0.56 in the dividend announcement month, and to $0.49, in the
last month of the fiscal year. As indicated by the t and Z statistics, the differences
between the forecast errors in the first, the dividend announcement , and the last
months of the year are not statistically significant , at a minimum 5% level.
Nevertheless, while the maximum value remains about the same throughout the
year, the minimum value declines from -$38.93 to -$26. 7, or by about 50%. The initial
minimum forecast error is highly negative ( analysts are excessively overoptimistic
for the particular firm) , and it ends up much less negative, in the end of the year.
This is indicative of at least some revision activity in my sample.
Graph 1 presents the frequency distributions of the raw forecast error in the first,
the dividend announcement (D.A.M.), and the last months of the fiscal year, and
provides a visual confirmation of the patterns identified in the previous paragraph.
The two panels graph the same frequency distributions, rotated to show both the
positive and the negative tails. Note that negative values on the horizontal axis
are indicated with numbers in parentheses. The mean raw forecast error drops,
the amount of negative errors declines, and the concentration around zero increases
from the first to the last month of the fiscal year. The upper graph shows that
whereas the positive forecast errors do not drop dramatically during the fiscal year,
highly negative initial errors become much smaller as the fiscal year winds down.
Nevertheless, a certain degree of intra-year forecast revision activity is evident.
3.3.2 The absolute forecast errors
Table 8 presents the descriptive statistics of the intra-year behavior of the absolute
forecast error. When measured as a percentage, the forecast errors in my sample
exhibit even more pronounced revision activity throughout the year. The absolute
forecast error drops from 81 % of EPS, in the first month of the year, to 69% in the
dividend announcement month, to 55% of EPS in the last month of the year. The
median absolute forecast error declines from 4 7% of EPS in the first month of the
53
year, to 46% of EPS in the dividend announcement month, and to 40% of EPS in
the last month of the year.
The absolute forecast error descriptive statistics document that in the last month
of the fiscal year the forecast error is a full 55% of earnings per share. This may
seem too high, particularly when compared to findings like those in Brown et . al.
[14], where it is documented that this measure of forecast error drops from 33% to
28% , in their sample of forecasts of 212 firms over the six year period 1975-80. The
difference between their results and mine is that the forecast errors in their sample
are truncated to 100%, when above it.
O ' Brien [60], who does not truncate her data, reports that her measure of the
absolute forecast error drops by more than 61 % from just after the first quarterly
earnings announcement date to just before the annual earnings announcement date,
results more similar to mine. As a comparison, when truncated to 100%, my sample
absolute forecast errors drop from 44% (median of 45%), to 36% (median of 35%),
in line with the documented error magnitude of previous research.
The t statistic and Wilcoxon Z for two-sample tests for differences between the
mean absolute forecast error on the first and the last month of the year, are reported
in columns 4 to 6 of Table 8. Column 4 contains the Student 's t and Wilcoxon Z
values for the two-sample tests for differences between the forecast errors in the
first (Month 1) and the last month (Month 12) of the fiscal year (1-12). Column 5
presents the t and Z values for the difference between the first month and the month
of the dividend announcement (1-D.A.M.) , and Column 6 contains the statistics
between the month of the dividend announcement and the last month of the year
(D .A.M.-12).
The Student's t and Wilcoxon Z values for the two-sample tests for differences
between the absolute error in the first and the last months of the year are significant
at the 5% level, indicating that the average absolute forecast error declines from the
first to the last month of the fiscal year, and the intra-year secular revision activity
is statistically significant. In Column 4, the t and Z statistics for differences between
the absolute error in the first and the last months of the fiscal year are 3.24 and 9. 76,
54
respectively. The Z statistics for the two-sample non-parametric tests for differences
within the two subperiods are also significant (3 .43 for the difference between the
first and the dividend announcement month, and 6.48 for the difference between the
dividend announcement month and the last month of the fiscal year) , in support
of the hypothesis that the forecast error decline in the period from the dividend
announcement month to the last month of the year.
Graph 2 presents the frequency distribution of the absolute forecast error in the
first , the dividend announcement , and the last months of the fiscal year. The two
panels graph the same frequency distributions, rotated to show both the positive
( upper panel) and the negative tails (lower panel). The absolute forecast error
declines markedly from the first to the last month of the fiscal year. Not only does
the concentration near zero increase dramatically, but the number of extreme errors
(above 130% of the year-end EPS) also declines. This indicates that the general
trend in my sample is for forecast errors to decline during the year, which leads to
the need to adjust for this secular revision activity in my tests of the forecast revision
after the dividend announcement .
3.4 Conclusions
I argued in Chapter 1 that testing for the impact of announcements of dividend
changes on analysts ' earnings forecasts should not involve examining the level of
forecast errors around the month of the dividend announcement , because ( a) assess
ing the level of forecast errors from month to month is not equivalent to examining
the forecast revision, and (b) the intra-year decline of forecast errors has been docu
mented in the literature, for samples without the presence of dividend changes and
therefore linking it to the potential information content of the dividend change is
problematic.
The bulk of the evidence presented in this chapter is supportive of the docu
mented secular intra-year decline of forecast errors, and establishes the need for
adjusting the measure of the forecast revision after a dividend announcement for
55
this secular revision activity. First, I document that the consensus analysts' earn
ings forecast in my sample is significantly revised throughout the fiscal year, using
a raw and a percentage measure of the intra-year forecast revision .
Then, employing two of the most widely used measures of forecast errors, the
raw forecast error and the absolute forecast error, I document a significant intra-year
decline in the sample mean of the absolute forecast error, and the sample medians of
both measures of the forecast error, and at least some revision activity in the sample
of the raw forecast errors . The frequency distributions of the two forecast error
measures support the hypothesis that forecast errors decline and their frequency
distributions concentrate around zero from the beginning to the end of the year.
In Chapter 4, next , I examine the forecast revision after dividend announcements.
56
Chapter 4
Forecast revisions after the dividend
announcement
This chapter presents the analysis of revisions of earnings forecasts after announce
ments of dividend changes. The sample was described in Chapter 2. The discussion
here begins with the two measures of the post-dividend announcement forecast re
vision. Sections 4.2 and 4.3 present the descriptive statistics of the forecast revision
after the dividend announcement in the subsamples of dividend increases and cuts,
and dividend announcements contemporaneous with a quarterly earnings announce
ment , and dividend changes which were not announced near such an announcement .
Section 4.4 presents the statistical analysis of the forecast revision. Finally, section
4.5 discusses the empirical results and concludes. The appendix provides details of
numerous sensitivity tests which are discussed in section s 4.3-4.5.
4.1 The measures of the forecast revision after the
dividend announcement
I examine the revision in analysts' earnings forecasts after the announcement of a
dividend change using:
• The excess forecast revision, which I define as the forecast revision after the
announcement of the dividend change minus the average forecast revision in
57
the period preceding the dividend announcement month. This is a measure
of the forecast revision after controlling for the secular, or "expected", or
"normal" forecast revision in the dividend announcement month. To compare
my findings and my sample with Lang and Litzenberger [4 7], and Yoon and
Starks [73], I also report the unadjusted forecast revision, i.e. the forecast
revision before controlling for the secular revision.
• Both the mean and the median consensus earnings forecast from the I/B/E/S
Inc. history tape. As I discussed in Chapter 1, previous researchers used the
median consensus forecast as their measure of analysts earnings forecast. For
the reasons explained there, I employ both measures.
• The excess forecast revision and the unadjusted forecast revision over both the
one-month period after the dividend announcement month, and the two-month
period around the dividend announcement month.
• Two measures of the excess and the unadjusted forecast revision: (a) the excess
and unadjusted absolute forecast revision and (b) the excess and unadjusted
percentage forecast revision.
I use such a range of measures and assessment periods of the forecast rev1s1on
first in order to provide as complete an assessment as possible of the impact of
announcements of dividend changes on analysts ' earnings forecasts. Second, I will
validate my findings by making them comparable to the findings in the literature
on analysts' forecast revisions after dividend announcements. Third, this variety of
measures and assessment periods , combined with a number of other parallel tests,
that I will describe and justify below , will serve as sensitivity checks on my findings.
I impose the restriction that the dividend announcement occurred in the third
month of the fiscal year or later, in order to have at least one prior two-month pe
riod over which to measure the secular revision. I measure the secular, or expected,
forecast revision in the one month after the dividend announcement as the average
one-month forecast revision in up to three months before the dividend announcement
month. Likewise, I use the average forecast revision in up to three prior two-month
58
periods to control for the secular forecast revision in the two months around the div
idend announcement month, i.e. from the month before the dividend announcement
month, to the month after the dividend announcement month.
When the dividend announcement month is the third month of the year, the
two-month secular revision is measured as the forecast revision in the previous
three-month period, i.e. between the last month of the previous fiscal year, and
the second month of the current fiscal year. For example, if the dividend announce
ment occurred in March, the two-month secular revision is measured as the revision
between December and February.
The 1/B/E/S Inc. database includes consensus analysts' forecasts of annual
earnings of New York Stock Exchange-listed and American Stock Exchange-listed
companies for all fiscal years between 1974 and 1992. The forecasts for each fiscal
year's earnings ( e.g. fiscal year 1986) typically begin in the last three months of the
previous fiscal year (in this case, there would be a consensus fiscal year 1986 earnings
forecast in each of the last three months of fiscal year 1985), and end in the third
month of the next fiscal year (in this case, there would be a consensus fiscal year
1986 earnings forecast in each of the first three months of fiscal year 1987), since
firms announce final annual results well into the next fiscal year.
When the dividend announcement occurred in the fourth month of the year, the
secular revision is measured as the average of the prior two two-month periods, i.e.
the average of the revision between the last month of the previous year and the
second month of the current year, and the revision between the first month and the
third month of the current year.
The two-month periods over which I compute the secular revision activity are
overlapping for two reasons. First, because this is the methodology used by Denis,
Denis and Sarin [26]. Second, in order to maximize the number of observations
included in each benchmark revision measure, and, therefore, to obtain a measure
less sensitive to extreme observations.
59
The difference between my specification of the secular forecast revision and that
of Denis , Denis and Sarin [26], is that (a) I do not impose the restriction that the
dividend announcement occurred at least 5 months after the beginning of the year,
which they do impose in order to ensure that enough two-month periods exist for
the fourth order moving average process of their specification of the secular revision
activity to be calculated, and (b) the secular forecast revision measure that I use is
not calculated by Brous' [12] method, but rather as the simple average of the prior
months ' forecast revision.
I do not emphasize the Brous, and Denis , Denis, and Sarin specification of the
secular forecast revision as the main secular revision measure in my study because
their measure is based on the average intra-year, two-month forecast revision. It
seems natural to think that as the fiscal year progresses, more quarterly earnings
announcements are made, and analysts generally obtain more information about
companies ' profitability. As a result, the use of a secular revision measure such
as the one proposed in Brous [12] may overestimate, on average, the "expected",
or secular forecast revision at the announcement of a dividend change if revisions
in the second half of the year are on average higher than the revisions in the first
part of the year. In addition, if the dividend announcement contains significant
information content it may lead to post-announcement revision activity which is
higher than before the announcement, and an intra-year measure of the secular
revision may lead to under-estimation of the significance of the revision after the
dividend announcement .
I estimate the benchmark, or secular, forecast revision over at most three prior
one-month periods, and over at most three two-month periods before the dividend
announcement month. I only require that the dividend announcement month is at
least the third month of the fiscal year or later, so that the secular two-month revision
be calculated over at least one prior two-month period. Using this less restrictive
screening criterion, my sample has 798 observations more than the sample used by
Denis , Denis, and Sarin [26] .
60
Nevertheless, to validate my findings, I computed the Brous measure of the secu
lar forecast revision and subtracted it from my measures of the unadjusted forecast
revision. These results are reported in Tables A.1 through A.6 of the Appendix.
I find that the choice of the secular forecast revision measure does not affect the
findings. In particular, the signs and the significance levels of the measures of the
excess forecast revision remain essentially identical when using either the Brous or
my specification of the expected forecast revision. In addition, the magnitudes of
the excess forecast revision are very similar. This is true for both the mean and the
median consensus analysts' earnings forecast measure.
4.1.1 The dollar forecast revision
I examine both the unadjusted dollar forecast revision and the excess dollar forecast
revision after the dividend announcement . The unadjusted dollar forecast revision
(UDRV) is defined as the dollar forecast revision after the dividend announcement .
The dollar forecast revision is assessed over both the one month after the dividend
announcement month (i.e. as the consensus earnings forecast released in the month
following the month of the dividend announcement minus the consensus earnings
forecast released in the month of the dividend announcement), and the two months
around the dividend announcement month (i.e. as the consensus earnings forecast
released in the month following the month of the dividend announcement minus the
consensus earnings forecast released in the month before the month of the dividend
announcement). The exact date of the release, or of the compilation, of the consensus
earnings forecast every month is not cited in the database. The use of the timing of
the earnings forecast follows the practice in the literature (see Lang and Litzenberger,
Yoon and Starks, and Denis et . al.).
The excess dollar forecast revision after the dividend announcement is computed
as the unadjusted dollar forecast revision minus the benchmark (i.e. secular, or
expected), forecast revision. I calculate the secular forecast revision as the average
forecast revision in up to a maximum of three one- and two-month periods before
the dividend announcement month. I use both the I/B/E/S Inc. mean and median
61
consensus analysts' earnings forecast for both the unadjusted and the excess forecast
rev1s10n.
The unadjusted dollar forecast revision (UDRV) after the dividend announcement
month m is equal to:
UDRVm,m+l
UDRVm-l,m+l
where the first equation is the one-month measure and the second equation is the
two-month measure. Fm is the I/B/E/S Inc. (mean or median) consensus analysts'
earnings forecast in the base month of the dividend announcement, Fm+I is the
(mean or median) consensus analysts' earnings forecast in the month after the divi
dend announcement month, and Fm-l is the (mean or median) consensus analysts'
earnings forecast in the month before the dividend announcement month.
Assuming, for example, that the dividend announcement was on June 15 of a
given year, the dividend announcement month (month m) is June, the previous
month (month m-1) is May, and the month after the dividend announcement month
is July. The one-month forecast revision, then, would be equal to the July earnings
forecast minus the June earnings forecast, and the two-month revision would be
equal to the July forecast minus the May forecast.
The excess dollar forecast revision (EDRV) is defined as the unadjusted dollar
forecast revision minus the "expected" forecast revision, over the one month after,
and the two months around the dividend announcement month, m:
EDRVm,m+l
62
EDRVm-1,m+l
where the first bracketed term in the right-hand side of each of the equations is the
unadjusted forecast revision, and the second bracketed term is the expected forecast
revision in the dividend announcement month (i.e. the secular revision activity in up
to three prior one- and two-month periods). Fi is the (mean or median) consensus
analysts' earnings forecast in month i, where i is measured such that there are at
least one and at most three one- and two-month periods in the calculation of the
"expected" forecast revision. The number of months over which the secular revision
is measured, x, depends on the month of the fiscal year in which the dividend change
was announced, and is at least equal to one and at most equal to three.
Using the June dividend announcement month example from above, the one
month secular revision activity would be estimated as the average of three prior
forecast revisions: May forecast minus April forecast, April forecast minus March
forecast, and March forecast minus February forecast. Likewise, the two-month
expected revision would be the average of: May forecast minus March forecast,
April forecast minus February forecast, and March forecast minus January forecast.
The unadjusted dollar forecast revision measures the dollar earnings forecast
change after the dividend announcement. The excess dollar forecast revision mea
sures the unexpected consensus earnings forecast revision after the announcement
of a dividend change. The null hypothesis is that the analysts' consensus earnings
forecast revision after the dividend announcement is not different from the expected
forecast revision in that period.
4.1.2 The percentage forecast revision
The second measure of the forecast revision I use is the unadjusted and the excess
percentage forecast revision (PRY and EPRV). Again, I assess the two measures
over the one month after the dividend announcement month, and the two months
63
around the dividend announcement month, usmg both the mean and the median
consensus analysts' earnings forecast.
The unadjusted percentage forecast rev1s10n after the dividend announcement
month, m, is defined as:
PRVm,m+l
PRVm-I,m+l
Fm+i-Fm-1
Fm-1
where Fm is the 1/B/E/S Inc. (mean or median) consensus analysts' earnings fore
cast in the base month of the dividend announcement, Fm+I is the (mean or median)
consensus analysts' earnings forecast in the month after the dividend announcement
month, and Fm-I is the (mean or median) consensus analysts' earnings forecast
in the month before the dividend announcement month. The interpretation of the
months and the estimation periods is as explained in the previous section.
1
The excess percentage forecast revision (EPRV) after the dividend announcement
month, m, is defined as:
EPRVm,m+l
Fm+l -Fm _ [l "°'m-1 F;-Fi-1]
Fm x L..i=m-x Fi-1
EP RVm-l,m+l
1
The algorithm used to compute the percentage forecast revision takes into account all potential
changes in signs in the level of the earnings forecast. For example, the algorithm recognizes that
an analyst who issues an earnings forecast of -$2 after forecasting earnings of $1 in the previous
month has revised her forecast by -300%. This adjustment in the sign of the forecast revision was
applied to all earnings forecast sign reversals.
64
where the interpretation of the symbols is as follows . The first term in the right-hand
side of the equation is the unadjusted percentage forecast revision, measured over
one month, from the dividend announcement month to the next month (in the first
equation), and the two months , from the month before the dividend announcement
month to the month after the month of the dividend announcement (in the second
equation) . The second, bracketed, term is the measure of the expected percentage
forecast revision in the dividend announcement month, i.e. the secular percentage
forecast revision in the months prior to the dividend announcement month. The
number of months over which the secular revision is measured, x, depends on the
month of the fiscal year in which the dividend change was announced, and it is at
least one and at most equal to three.
This set of measures is roughly similar to the one used by Lang and Litzenberger
[47] and Yoon and Starks [73]. The differences between the percentage forecast
revision and the measures used in [47], and [73] are that:
• In the tests that I report and discuss below, I don't divide the percentage fore
cast revision by the dividend percentage change, which is the revision measure
used by Lang and Litzenberger and Yoon and Starks. Nevertheless, as a sensi
tivity check, I conducted the univariate tests that I review in the next sections
using the Lang and Litzenberger unadjusted forecast revision measure, and I
find results similar to the ones reported by Yoon and Starks, but which also
confirm my conclusions. In particular, I find that the measure used by Lang
and Litzenberger and Yoon and Starks, when applied to my sample, yields sig
nificant forecast revisions after the announcement of both dividend increases
and dividend cuts . Nevertheless, when adjusting for the secular revision activ ity ( using a benchmark revision measure which is computed similarly to the
Lang and Litzenberger measure, i.e. standardized by the dividend percentage
change) , the post-dividend announcement revision is statistically significantly
different from zero only after dividend cuts. These results are, of course, avail
able upon request.
• Lang and Litzenberger and Yoon and Starks use only the first part of this
statistic (i .e. the unadjusted revision, which, again, they standardize by the
65
dividend percentage change) as the measure of the forecast revision around the
dividend announcement month, without controlling for the expected forecast
rev1s1on.
• Both pairs of authors use only the median consensus earnings forecast , whereas
I use both the mean and the median consensus earnings forecast .
• Finally, Lang and Litzenberger and Yoon and Starks use only the two-month
period around the dividend announcement month to assess the forecast revi
sion, whereas I use both the one-month after, and the two-month around (from
the month before to the month after) the dividend announcement month pe
riods .
The null hypothesis is that the percentage forecast rev1s1on after the dividend
announcement is equal to the benchmark percentage forecast revision.
4.2 The revision after dividend announcements
Tables 9 and 10 present the univariate statistics of the mean (Table 9) , and the
median (Table 10) consensus forecast revision around the dividend announcement
month for the subsample of 2,674 announcements of a dividend increase. Tables
11 and 12 present the univariate statistics of the mean (Table 11) and the median
(Table 12) consensus forecast revision for the subsample of the 192 announcements
of a dividend cut .
I separate the subsamples of dividend increases and cuts to test the dividend sig
nalling hypothesis. According to the Modigliani and Miller [58] Information Content
of Dividends Hypothesis, dividend increases should generate upward revisions in ex
pectations of firms' future earnings , which should translate into significant upward
earnings forecast revisions, i.e. forecast revisions with a positive sign. Likewise,
dividend cuts should signal lower current and future earnings, which should lead to
significant downward earnings forecast revisions .
66
Table 4.1: Mean consensus forecast rev1s10n after the dividend announcement: Div
idend increases
This table contains the test of the significance of the analysts' mean consensus earnings forecast revision after the
announcement month of the dividend change, for the sample of 2,674 announcements of a dividend increase of
more than 10%, between 1979-90. The first three columns present the statistics for the one-month forecast revision
measured from the month of the dividend announcement to the month after the dividend announcement month.
Columns 4, 5, and 6 contain the statistics for the two-month forecast revision measured from the month before
the month of the dividend announcement to the month after the dividend announcement month. Column 1 is the
unadjusted earnings forecast revision measure after the month of the dividend announcement. Column 2 is the
average monthly forecast revision in the period up to 3 months before the dividend announcement month. Column 3
is the excess forecast revision, calculated as the difference between the unadjusted forecast revision and the average
forecast revision in the prior months (Column 1 minus Column 2). Columns 4 to 6 present the two-month measures
of the unadjusted forecast revision, the average prior months' revision and the excess forecast revision. The two
month unadjusted forecast revision (Column 4) is measured from one month before the dividend announcement
month to one month after the dividend announcement month. The average prior months' forecast revision ( Column
5) is measured over two-month intervals up to 4 months before the dividend announcement month. The excess
forecast revision (Column 6) is the difference between the unadjusted and the average prior months' revision.
The first panel (A) presents the statistics of the dollar forecast revision defined as Fm+l - Fm, where F is the
consensus analysts' earnings forecast in the dividend announcement month m. The second panel (B) presents the
statistics of the percentage forecast revision, defined as the dollar forecast revision divided by the base month's
earnings forecast. The numbers in the parentheses under the measurement period headings indicate the number of
positive/zero/negative revisions in the sample.
One-month statistics Two-month statistics
(1,043/888/743) (1,538/424/712)
Unadjusted Average Excess Unadjusted Average Excess
forecast prior months' forecast forecast prior months' Forecast
revision revision revision revision revision revision
(1) (2) (3) (4) (5) (6)
A. Dollar forecast revision
Mean 0.011 0.016 -0.005 0.029 0.037 -0.009
Median 0.000 0.003 0.000 0.000 0.01 -0.001
a 0.18 0.08 0.18 0.26 0.17 0.24
t-stat 2.69 7.99 -1.12 4.17 8.48 -1.38
Wile Z 7.77 11.05 -2.09 9.08 11.29 -1.40
B. Percentage forecast revision
Mean 0.70 0.80 -0.09 1.70 1.80 -0.10
Median 0.00 0.20 0.00 0.00 0.60 0.00
a 7.00 4.21 8.45 10.22 8.20 9.10
t-stat 4.81 9.69 -0.07 6.89 10.5 -0.77
Wile Z 7.78 11.37 -3.39 9.08 11.81 -3.14
67
Table 4.2: Median consensus forecast rev1s10n after the dividend announcement:
Dividend increases
This table contains the test of the significance of the analysts' median consensus earnings forecast revision after
the announcement month of the dividend change, for the sample of 2,674 announcements of a dividend increase of
more than 10%, between 1979-90. The first three columns present the statistics for the one-month forecast revision
measured from the month of the dividend announcement to the month after the dividend announcement month.
Columns 4, 5, and 6 contain the statistics for the two-month forecast revision measured from the month before
the month of the dividend announcement to the month after the dividend announcement month. Column 1 is the
unadjusted earnings forecast revision measure after the month of the dividend announcement. Column 2 is the
average monthly forecast revision in the period up to 3 months before the dividend announcement month. Column 3
is the excess forecast revision, calculated as the difference between the unadjusted forecast revision and the average
forecast revision in the prior months (Column 1 minus Column 2). Columns 4 to 6 present the two-month measures
of the unadjusted forecast revision, the average prior months' revision and the excess forecast revision. The two
month unadjusted forecast revision (Column 4) is measured from one month before the dividend announcement
month to one month after the dividend announcement month. The average prior months' forecast revision ( Column
5) is measured over two-month intervals up to 4 months before the dividend announcement month. The excess
forecast revision (Column 6) is the difference between the unadjusted and the average prior months' revision.
The first panel (A) presents the statistics of the dollar forecast revision defined as Fm+l - Fm, where F is the
consensus analysts' earnings forecast in the dividend announcement month m. The second panel (B) presents the
statistics of the percentage forecast revision, defined as the dollar forecast revision divided by the base month's
earnings forecast. The numbers in the parentheses under the measurement period headings indicate the number of
positive/zero/negative revisions in the sample.
One-month statistics Two-month statistics
(554/1,301/819) (793/884/997)
Unadjusted Average Excess Unadjusted Average Excess
forecast prior months' forecast forecast prior months' Forecast
revision revision revision revision revision revision
(1) (2) (3) (4) (5) (6)
A. Dollar forecast revision
Mean 0.011 0.014 -0.002 0.028 0.032 -0.004
Median 0.000 0.003 0.000 0.000 0.005 0.000
a 0.19 0.09 0.19 0.26 0.18 0.25
t-stat 2.77 7.15 -0.60 4.64 7.66 -0.58
Wile Z 7.47 11.21 -2.55 9.48 11.73 -1.82
B. Percentage forecast revision
Mean 0.80 0.80 -0.00 1.80 1.10 0.70
Median 0.00 0.00 0.00 0.00 0.10 0.00
a 8.25 4.50 8.00 11.10 7.60 11.23
t-stat 4.43 9.60 -0.27 6.96 6.48 2.82
Wile Z 7.47 11.47 -4.22 9.48 10.71 2.12
68
Table 4.3: Mean consensus forecast rev1s10n after the dividend announcement: Div
idend cuts
This table contains the test of the significance of the analysts' mean consensus earnings forecast revision after the
announcement month of the dividend change, for the sample of 192 announcements of a dividend cut of less than
-10%, between 1979-90. The first three columns present the statistics for the one-month forecast revision measured
from the month of the dividend announcement to the month after the dividend announcement month. Columns 4,
5, and 6 contain the statistics for the two-month forecast revision measured from the month before the month of the
dividend announcement to the month after the dividend announcement month. Column 1 is the unadjusted earnings
forecast revision measure after the month of the dividend announcement. Column 2 is the average monthly forecast
revision in the period up to 3 months before the dividend announcement month. Column 3 is the excess forecast
revision, calculated as the difference between the unadjusted forecast revision and the average forecast revision in
the prior months (Column 1 minus Column 2). Columns 4 to 6 present the two-month measures of the unadjusted
forecast revision, the average prior months' revision and the excess forecast revision. The two-month unadjusted
forecast revision (Column 4) is measured from one month before the dividend announcement month to one month
after the dividend announcement month. The average prior months' forecast revision (Column 5) is measured over
two-month intervals up to 4 months before the dividend announcement month. The excess forecast revision (Column
6) is the difference between the unadjusted and the average prior months' revision. The first panel (A) presents the
statistics of the dollar forecast revision defined as Fm+l - Fm, where F is the consensus analysts' earnings forecast
in the dividend announcement month m. The second panel (B) presents the statistics of the percentage forecast
revision, defined as the dollar forecast revision divided by the base month's earnings forecast. The numbers in the
parentheses under the measurement period headings indicate the number of positive/zero/negative revisions in the
sample.
One-month statistics Two-month statistics
(18/34/140) (30/14/148)
Unadjusted Average Excess Unadjusted Average Excess
forecast prior months' forecast forecast prior months' Forecast
revision revision revision revision revision revision
(1) (2) (3) (4) (5) (6)
A. Dollar forecast revision
Mean -0.337 -0.229 -0.108 -0.702 -0.459 -0.242
Median -0.160 -0.145 -0.023 -0.340 -0.298 -0.089
a 0.49 0.29 0.38 1.03 0.60 0.72
t-stat -8.57 -9.96 -3.58 -7.84 -8.74 -3.89
Wile Z -12.55 -12.99 -2.37 -13.27 -12.94 -3.99
B. Percentage forecast revision
Mean -18.80 -10.45 -8.40 -40.23 -22.74 -17.55
Median -11.00 -7.90 -4.59 -27.61 -16.50 -13.15
a 119.00 59.50 135.00 108.25 53.70 130.00
t-stat -2.09 -2.35 -2.83 -4.62 -5.42 -4.69
Wile Z -8.82 -8.81 -3.96 -11.14 -11.47 -7.67
69
Table 4.4: Median consensus forecast rev1s10n after the dividend announcement:
Dividend cuts
This table contains the test of the significance of the analysts' median consensus earnings forecast revision after the
announcement month of the dividend change, for the sample of 192 announcements of a dividend cut of less than
-10%, between 1979-90. The first three columns present the statistics for the one-month forecast revision measured
from the month of the dividend announcement to the month after the dividend announcement month. Columns 4,
5, and 6 contain the statistics for the two-month forecast revision measured from the month before the month of the
dividend announcement to the month after the dividend announcement month. Column 1 is the unadjusted earnings
forecast revision measure after the month of the dividend announcement. Column 2 is the average monthly forecast
revision in the period up to 3 months before the dividend announcement month. Column 3 is the excess forecast
revision, calculated as the difference between the unadjusted forecast revision and the average forecast revision in
the prior months (Column 1 minus Column 2). Columns 4 to 6 present the two-month measures of the unadjusted
forecast revision, the average prior months' revision and the excess forecast revision. The two-month unadjusted
forecast revision (Column 4) is measured from one month before the dividend announcement month to one month
after the dividend announcement month. The average prior months' forecast revision (Column 5) is measured over
two-month intervals up to 4 months before the dividend announcement month. The excess forecast revision (Column
6) is the difference between the unadjusted and the average prior months' revision. The first panel (A) presents the
statistics of the dollar forecast revision defined as Fm+l - Fm, where F is the consensus analysts' earnings forecast
in the dividend announcement month m. The second panel (B) presents the statistics of the percentage forecast
revision, defined as the dollar forecast revision divided by the base month's earnings forecast. The numbers in the
parentheses under the measurement period headings indicate the number of positive/zero/negative revisions in the
sample.
One-month statistics Two-month statistics
(26/41/125) (23/39/130)
Unadjusted Average Excess Unadjusted Average Excess
forecast prior months' forecast forecast prior months' Forecast
revision revision revision revision revision revision
(1) (2) (3) (4) (5) (6)
A. Dollar forecast revision
Mean -0.330 -0.216 -0.114 -0.595 -0.445 -0.150
Median -0.132 -0.149 -0.063 -0.350 -0.272 -0.069
a 0.48 0.32 0.45 0.82 0.67 0.74
t-stat -6.97 -9.20 -2.52 -9.04 -8.31 -2.53
Wile Z -11.39 -12.83 -1.69 -13.35 -13.08 -3.41
B. Percentage forecast revision
Mean -20.25 -12.43 -7.81 -46.90 -9.76 -37.20
Median -9.19 -8.42 -3.00 -25.63 -5.54 -19.15
a 161.00 38.20 165.20 89.15 16.00 90.00
t-stat -2.67 -4.33 -3.63 -6.59 -7.54 -5.21
Wile Z -8.22 -10.91 -2.97 -11.31 -12.28 -8.04
70
Analysts, in my sample, on average, do not revise their year-end earnings fore
casts, differently than in any other month, after announcements of dividend in
creases . The bulk of the evidence presented next suggests that analysts , on aver
age, revise their forecasts of year-end earnings significantly differently than in prior
months only after dividend cuts. These general findings hold for both the mean and
the median consensus I/B/E/S Inc. earnings forecast.
Tables 9, and 10 show that the unadjusted forecast revision after dividend in
creases is significantly positive, for both measures and for both measurement periods
( Columns 1 and 4, of Panel A for the dollar forecast revision, and of Panel B for the
percentage forecast revision) . The statistical significance of the unadjusted mean
consensus forecast revision, which seems to indicate significant forecast revision, is
misleading because the unadjusted revision does not account for normal secular re
vision activity. The unadjusted forecast revision is the measure used by Lang and
Litzenberger [47] , and Yoon and Starks [73] .
In contrast, with the exception of the two-month percentage median revision
(Table 10, Panel B, Column 6) , the unadjusted forecast revision is not significantly
higher than the secular forecast revision, whether the revision measure is assessed
over a one-month or a two-month period. For the excess forecast revision after a
dividend increase to support the predictions of the Information Content of Dividends
Hypothesis its sign must be positive: if a dividend increase signals managers' expec tations of permanently higher future earnings, analysts should revise their forecasts
upward by more than in prior months when this information signal was not yet
released. The findings of Table 9 are inconsistent with this hypothesis.
The average excess mean consensus forecast revision after dividend increases
(Table 9, Panels A and B, Columns 3 and 6) is either lower in magnitude than the
typical prior months' forecast revision, or not significantly different from zero, at the
5% significance level. The important statistics of interest here are the means of the
excess revision measures and the t and Wilcoxon Z statistics for differences of these
means from zero.
71
The t statistics of the excess revision measures (Columns 3 and 6) are either
not significant , or they are significantly negative. The significant negative Wilcoxon
Z values of the one-month measures are the opposite of what the theory predicts,
because they indicate that the forecast revision after the announcement of a dividend
increase is lower than the secular forecast revision, i.e. analysts revise their forecast
either upward but less than in previous months, or downward.
The evidence presented in Table 9 with respect to the mean consensus forecast
revision is largely repeated in the case of the median consensus revision, in Table
10. The only exception is the two-month measure of the percentage median forecast
revision (Panel B, Column 6). This may be due to the fact that this measure of
the forecast revision is likely to capture the effect of information signals above and
beyond the dividend announcement, since it is assessed over two months . Given
that this is the only apparent instance of disagreement witht the main tendencies of
Tables 9 and 10, I point it out but I do not consider it enough evidence to overturn
the main patterns discovered so far.
In contrast to dividend increases, dividend cuts, in my sample, generate signif
icant reductions in analysts' earnings forecasts , across the board, consistent with
the predictions of the ICD hypothesis. This is true for both the mean and the
median consensus earnings forecast measures, and for both the one-month and the
two-month measures. In addition, this pattern is supported by both the parametric,
and the non-parametric test statistics, at the 5% significance level.
Table 11 presents the univariate statistics of the mean consensus forecast revision
measures of dividend cutting firms. Without exceptions the forecast revision after
dividend cuts is statistically significant at the 5% level, as shown by the the Student
t and Wilcoxon Z statistics in Columns 3 and 6, of Panels A and B of the Table.
Analysts, in my sample, reduce their mean forecast revision after dividend cuts by
10.8 cents, or 8.4%, more than in prior months, when the revision is measured from
the dividend announcement to the next month, and by 24.2 cents, or 17.5%, more
when the revision is measured over the two months around the month of the dividend
announcement.
72
Table 12 presents the statistics for the median consensus forecast revision mea
sures of dividend cutting firms. Analysts revise their median consensus forecast
downward. Columns 3 and 6 of Panels A and B contain the Student t and Wilcoxon
Z statistics of the median excess forecast revision, which show that the forecast revi
sion after the announcement of a dividend cut is significantly lower than the secular
forecast revision, at the 5% significance level.
The average one-month median consensus forecast is revised by an average of 11
cents, or 7.8%, more than the median consensus forecast revision in any prior month,
and by 15 cents , or 37.2%, more when the revision is measured over the two months
around the dividend announcement month. The means are statistically significant,
with t-statistics ranging from -2.52 to -5.21. These results are consistent with the
predictions of the ICD hypothesis of Modigliani and Miller: analysts in my sample
lower their expectations of the future earnings of firms that cut their dividend.
4.2.1 Discussion
The bulk of the evidence presented in this section indicates that the analysts whose
forecasts are included in my sample revise significantly their earnings forecasts after
the announcement of a dividend change only after dividend cuts. Denis, Denis and
Sarin [26], the only published research that adjusts the forecast revision measure
for the secular revision activity, find that the two-month excess percentage forecast
revision divided by the stock price two days before the dividend announcement is
significant for both dividend increases and cuts. The measure used in the previous
section which comes closest to the one used by Denis, Denis and Sarin [26], i.e.
the two-month median consensus excess percentage forecast revision, is significantly
different from the secular forecast revision, for both dividend increases and dividend
cuts , consistent with their findings.
Nevertheless, given that all other measures of the forecast revision I use indicate
that only dividend cuts have significant information content , there is a disagreement
between the two sets of findings. In addition, when I use the Brous measure of the
73
secular forecast revision in my sample (Tables A.1 and A.3 in the Appendix) , the
excess percentage forecast revision after dividend increases is not significant, using
either the median or the mean consensus earnings forecast. In particular, of the eight
measures of the excess forecast revision using Brous ' method, six measures are not
significantly positive according to both parametric and non-parametric tests , while
two measures are positive according to only the non-parametric test for differences
from zero. I consider these findings as essentially confirming the general conclusions
from the tests presented in Tables 9 to 12.
There are two possible explanations for the disagreement between my findings
and the evidence presented by Denis et . al. First, the results may be affected by
the size (2,866 observations) of my sample which is by 798 observations larger than
the one used in Denis , Denis and Sarin [26] . Second, as discussed in Chapter 2, my
sample was screened against the Wall Street Journal Index for the correct dividend
announcement date, and for absence of any kind of anticipatory announcements or
reports of the dividend change This last screening caused the discarding of 325 data
points. Denis, Denis, and Sarin do not report having conducted such a screening.
Given the lag that has been documented in reporting revised forecasts on the
I/B/E/S Inc. tape (e.g. O ' Brien [60], and Brous [12]) , the announcement of a
dividend change, on average, will not lead to revisions by all the analysts following
a company. The possible updated expectations will be incorporated in the I/B/E/S
Inc. database only gradually. Anticipatory dividend change announcements two or
three months before the actual dividend announcement month will lead to revisions
which are likely to be reflected on earnings forecasts by the month of the dividend
announcement . It is possible that Denis, Denis , and Sarin document such an effect,
whereas my sample, cleansed of such anticipated dividend announcements does not.
The findings of the previous section, in addition, are in disagreement with the
conclusions of both Lang and Litzenberger [47], and Yoon and Starks [73]. Given
that these two papers report contradicting results on the significance of the same
measure of the forecast revision after a dividend change announcement (Lang and
74
Litzenberger do not find significant revisions, whereas Yoon and Starks find signifi
cant revisions after both dividend increases and cuts), my findings must be consid ered as additional evidence, rather than as a new development.
Lang and Litzenberger, who do not control for the secular revision activity, do not
find significant forecast revisions after the dividend announcement. I find that the
unadjusted forecast revision measure that I use is statistically significant after both
dividend increases and cuts, but that the forecast revision measure which accounts
for the secular revision activity is significant only after dividend cuts.
Yoon and Starks report significant unadjusted revisions after both increases and
cuts, which agrees with the findings I report here. Nevertheless, their conclusions
on the significance of the post-announcement revision may not hold after controlling
for the secular revision activity. In particular, although the unadjusted revision
measures in my sample are significant, and of a sign consistent with the predictions
of the Information Content of Dividends Hypothesis, after both dividend increases
and cuts, adjusting for the secular revision activity reveals that only after dividend
cuts do analysts revise their forecasts significantly differently than in prior months,
and in a direction consistent with the predictions of the ICD hypothesis.
My findings are consistent with the dividend signalling model of Wa1-ther [70],
who posits that only dividend cuts are informative signals of firms' earnings prospects
because they define a credible signalling separation between good and distressed
firms.
Finally, the empirical evidence presented here on the asymmetric reaction of ana
lysts to dividend cuts versus dividend increases ( a) agrees with the findings of De An gelo, DeAngelo and Skinner [24] who showed that dividend cuts contain marginal
information about future earnings prospects beyond that contained in current earn
ings, and (b) is consistent with the findings of DeAngelo, DeAngelo and Skinner's
later paper [25] which documents that the dividend decisions of firms experienc ing declining earnings after 9 years of earnings growth (many of which increased
dividends) are not informative of earnings prospects.
75
4.3 Earnings forecasts and announcements
One of the most important sources of information that affects market expectations,
and analysts' forecasts , of firms' year-end earnings should be the quarterly earnings
announcements. Quarterly earnings announcements materially reduce uncertainty
about the annual earnings realization, since annual earnings are the sum of the
quarterly earnings figures. In addition, current earnings are well documented to
be able to predict future earnings (see, among others, Watts [71]) . Each quarterly
announcement of earnings decreases, on average, the proportion of annual earnings
that is unknown. The information content of potential signals (such as dividend
announcements) of firms' profitability should be considered in conjunction with po
tential contemporaneous earnings announcements, because the informational content
of the earnings announcement is obvious and direct.
The marginal information content of dividend announcements in the presence of
contemporaneous earnings announcements has been assessed in several event studies.
The evidence is mixed. One of the first papers that examined the information content
of dividend policy changes in the presence of earnings signal, Gonedes [38] finds that
the dividend announcement signal does not improve the power of a linear model
which controls for current earnings to predict future earnings. DeAngelo, DeAngelo
and Skinner [25] document that the dividend decisions of firms that experience an
earnings decline after at least nine years of growing earnings do not contain marginal
information content beyond that contained in current earnings.
Penman [64] examined the predictive accuracy of dividend announcements and
managers ' earnings forecasts with respect to firms ' future earnings realizations . He
reports that the dividend change is informative only when it signals expectations of
high earnings levels corroborated by a similar earnings prediction. In those cases
where the dividend change is not commensurate with the earnings prediction, the
dividend announcement does not carry measurable information content .
In one of the first papers that document marginal dividend signal in the pres
ence of an earnings announcement , Kane, Lee, and Marcus [46] examine the stock
76
price reaction around contemporaneous earnings and dividend announcements and
find that both cause abnormal price movements. They interpret their results as
indicating that the market uses the information in one announcement to corrobo rate the signal contained in the other. Similarly, DeAngelo, DeAngelo and Skinner
[24] document that the dividend decisions of the 167 firms with at least one annual
loss between 1980-1985 in their sample contain information content beyond that of
current earnings.
Aharony and Swary [1] examine non-contemporaneous dividend and earnings an
nouncements and find that dividend announcements on average convey information
about future earnings above and beyond that contained in the quarterly earnings
realization. Griffin [39] examined the long run stock price reaction to dividend and
earnings announcements, and the release of analysts' revised forecasts . He docu
ments that announcements of dividend changes ( and especially cuts) cause abnormal
stock price drifts, independently of earnings announcements.
Leftwich and Zmijewski [48] measure abnormal stock returns around the day of
both contemporaneous (within 5 trading days of each other) and non-contemporaneous
( outside a 10-trading day window of each other) earnings and dividend announce
ments . They find that the dividend signal contains relevant information only when
it is of opposite sign to the earnings surprise in one particular case, namely when a
favorable earnings announcement coincides with a dividend reduction. In all other
cases , dividend announcements around earnings announcements do not cause abnor
mal stock price reactions.
The studies on the impact of announcements of dividend changes on analysts'
earnings forecasts have not examined the marginal information content of dividend
changes in the presence of earnings announcements . This is an important issue
because the information content attributed to dividend signals, in previous papers,
may be due to the information in the nearest earnings announcement. If that is the
case, then the dividend announcement may add no information that is relevant to
the analysts whose forecasts are included in my sample.
77
To examine this question, I divide my sample into dividend announcements which
occurred within 20 trading days (1 calendar month) before or after the nearest
quarterly earnings announcement (which I label "contemporaneous with the nearest
earnings announcement") , and dividend changes which were announced outside such
a 40-trading day window around the nearest quarterly earnings announcement ("non
contemporaneous with the nearest earnings announcement") .
The determination of the announcement day of the nearest earnings announce
ment is based on the Wall Street Journal Index, for all 2,866 firm-events. There
are 1,891 dividend announcements contemporaneous with the nearest earnings an
nouncement, and 975 dividend announcements non-contemporaneous with the near
est earnings announcement in my sample.
4.3.1 The tin1ing of the dividend change announcen1ent
Tables 13 and 14 present the one- and two-month mean (Table 13) and median (Ta
ble 14) consensus forecast revision statistics after dividend announcements, split into
dividend increases and dividend cuts , and dividend announcements contemporane ous with the nearest earnings announcement , and dividend announcements outside
a 40-trading day window around the nearest earnings announcement .
The statistics reported, for both the dollar and the percentage excess forecast
revision , are the mean, the median, the Student 's t statistic, and the Z value of the
Wilcoxon non-parametric test for differences from zero. The subsample sizes are as
indicated in the parentheses under each subsample heading.
The bulk of the evidence in Tables 13 and 14 indicates that analysts, in my
sample, revise significantly their annual earnings forecasts after the announcement
of a dividend change only after dividend cuts which were announced within a 40-
trading day window around the nearest earnings announcement . This is true for
both the mean and the median consensus earnings forecast revision .
78
Table 4.5: Tabulation of the mean consensus earnings forecast rev1s10n around the
dividend announcement month
This table presents the measures of the excess mean consensus earnings forecast revision after the dividend announce
ment month, for the 2,866 firm-events in the sample of dividend announcements, broken down into (a) dividend
increases versus dividend cuts, and (b) into dividend announcements within a 40-trading day window around the
nearest earnings announcement, and dividend announcements outside such a 40-trading days window around the
nearest earnings announcement. The statistics reported are the mean, the median, and the test statistics of the
one-month and the two-month excess forecast revision measures (i.e. the dollar and the percentage forecast revi
sion). Columns 1 through 4 report the dividend increases statistics, and Columns 5 through 8 report the dividend
cuts statistics. Columns 1 and 2, and Columns 5 and 6 report the statistics for the excess revision measures after
dividend announcements within a 40-trading day window around the nearest earnings announcement. Columns 3
and 4, and Columns 7 and 8 contain the statistics for the excess revision measures after dividend announcements
outside such a 40-trading day window around the nearest earnings announcement. The one-month measures are
reported in columns 1, 3, 5, and 7, and the two-month measures are reported in columns 2, 4, 6, and 8. The numbers
in the parentheses under the headings indicate subsample sizes.
Dividend fucreases Dividend Cuts
(2,674) (192)
Same month Not in same month Same month Not in same month
with earnings with earnings with earnings with earnings
announcement announcement announcement announcement
(1,763) (911) (128) (64)
1 month 2 month 1 month 2 month 1 month 2 month 1 month 2 month
measure measure measure measure measure measure measure measure
(1) (2) (3) (4) (5) (6) (7) (8)
Dollar Mean 0.005 -0.006 -0.010 -0.004 -0.170 -0.294 0.002 -0.120
Revision Median 0.000 0.000 -0.000 0.000 -0.042 -0.110 0.000 -0.001
t-stat 0.21 -0.82 -1.12 -0.49 -4.30 -4.08 0.05 -1.67
Wilcoxon Z 1.35 -2.11 -1.19 -1.29 -7.41 -7.05 2.08 -1.79
Percentage Mean 0.30 -0.20 -0.40 -0.20 -5.60 -9.60 2.00 -3.80
revision Median 0.00 0.00 -0.10 -0.10 -4.20 -8.12 0.00 0.00
t-stat 1.13 -0.64 -1.89 -0.47 -3.31 -3.48 0.34 -2.82
Wilcoxon Z 2.34 -1.15 -1.91 -2.28 -10.95 -11.17 2.00 -2.21
79
Table 4.6: Tabulation of the median consensus earnings forecast revision around the
dividend announcement month
This table presents the measures of the excess median consensus earnings forecast revision after the dividend an
nouncement month, for the 2,866 firm-events in the sample of dividend announcements, broken down into (a)
dividend increases versus dividend cuts, and (b) into dividend announcements within a 40-trading day window
around the nearest earnings announcement, and dividend announcements outside such a 40-trading days window
around the nearest earnings announcement. The statistics reported are the mean, the median, and the test statistics
of the one-month and the two-month excess forecast revision measures (i.e. the dollar and the percentage forecast
revision). Columns 1 through 4 report the dividend increases statistics, and Columns 5 through 8 report the dividend
cuts statistics. Columns 1 and 2, and Columns 5 and 6 report the statistics for the excess revision measures after
dividend announcements within a 40-trading day window around the nearest earnings announcement. Columns 3
and 4, and Columns 7 and 8 contain the statistics for the excess revision measures after dividend announcements
outside such a 40-trading day window around the nearest earnings announcement. The one-month measures are
reported in columns 1, 3, 5, and 7, and the two-month measures are reported in columns 2, 4, 6, and 8. The numbers
in the parentheses under the headings indicate subsample sizes.
Dividend fucreases Dividend Cuts
(2,674) (192)
Same month Not in same month Same month Not in same month
with earnings with earnings with earnings with earnings
announcement announcement announcement announcement
(1,763) (911) (128) (64)
1 month 2 month 1 month 2 month 1 month 2 month 1 month 2 month
measure measure measure measure measure measure measure measure
(1) (2) (3) (4) (5) (6) (7) (8)
Dollar Mean 0.001 -0.004 -0.009 -0.002 -0.180 -0.153 0.013 -0.145
Revision Median 0.000 -0.001 -0.000 0.000 -0.112 -0.095 0.020 -0.055
t-stat 0.21 -0.52 -2.13 -0.26 -2.77 -1.94 0.34 -1.79
Wilcoxon Z 0.79 -1.16 -2.41 -1.19 -4.41 -2.98 1.45 -1.87
Percentage Mean 0.002 0.005 -0.005 0.010 -0.14 -0.380 -0.047 -0.356
revision Median 0.00 0.00 -0.00 -0.10 -11.20 -8.82 0.00 -2.30
t-stat 0.87 1.67 -2.17 2.48 -2.53 -4.00 -0.57 -1.21
Wilcoxon Z 2.09 2.21 -2.43 3.12 -3.72 -4.73 -1.31 -1.88
80
The t and Z statistics for the mean consensus excess forecast revision (Table 13)
are both significant in the case of Dividend Cuts-Contemporaneous with an Earn ings Announcement ( Columns 5 and 6) , indicating that analysts significantly revise
downward their forecast revisions after the announcement of a dividend cut cor
roborated with a contemporaneous earnings announcement. The only other excess
revision measure which is significantly different than zero ( as indicated by both the
t and Z statistics) is the mean two-month percentage revision after a dividend cut
non-contemporaneous with an earnings announcement (Column 8) .
The two-month dollar excess revision non-contemporaneous with an earnmgs
announcement (Column 8, Panel A) is very near significant . Nevertheless, although
this measure is computed over two months, thus incorporating more information
into the forecast revision, it still fails to be statistically significant.
The excess revision after dividend increases is not significant, almost universally.
The Z statistic of the one-month contemporaneous with an earnings announcement
percentage revision ( Column 1) is the only instance of significance, but it is an
isolated case which does not alter the thrust of the findings. In addition, the 2-
month contemporaneous with an earnings announcement dollar revision Z statistic
( Column 2) is significantly negative, indicating that the revision after a dividend
increase is lower than the secular revision, a finding which confirms the previous
findings.
The median consensus analysts ' earnings forecast (Table 14) is likewise signif
icantly revised after the announcement of a dividend change, consistent with the
predictions of the dividend signaling hypothesis , primarily when a dividend cut is
announced contemporaneously with an earnings announcement. Table 14 presents
the same qualitative results as Table 13. The only combination of dividend change
direction and announcement timing which consistently yields excess forecast revi
sion measures different than zero is the Dividend Cut-Contemporaneous with an
Earnings Announcement (Columns 5 and 6).
The two-month percentage revision after a dividend increase not -contemporaneous
with an earnings announcement is significant (Column 4) and of the right sign. The
81
one-month rev1s10n measures after non-contemporaneous dividend increases ( Col
umn 3) are significantly negative, indicating a significant downward revision after
a dividend increase. The Z statistics of the percentage revision after both types
of timing with respect to the nearest earnings announcement ( Columns 1 and 2)
are significantly positive. Nevertheless , the inconcistency of the significance tests
of the dividend increase revisions ( with the exception of the non-contemporaneous
percentage revision) across measures and statistics leads me to discard them as noise.
The main thrust of the evidence remains in Columns 5 and 6. Analysts, in
my sample, revise their year-end earnings forecasts significantly differently than in
previous months only after announcements of dividend cuts which were made in the
same calendar month with a quarterly earnings announcement.
4.3.2 Discussion
The bulk of the evidence presented in this section indicate that analysts in my
sample, on average, do not revise their year-end earnings forecasts significanly (i.e.
beyond the expected, or secular, revision) in response to the announcement of a
dividend change per se. The only exception to this general trend are announcements
of dividend cuts that occur within 20 trading days around the nearest earnings
announcement , which seem to generate significant forecast revisions consistent with
the predictions of the Modigliani and Miller dividend signalling hypothesis. These
findings further refine the evidence presented in section 4.2 where it was shown
that dividend announcements carry informational content with respect to analysts'
earnings forecasts only after dividend cuts.
Furthermore, these findings underline the importance of examining the informa
tional impact of dividend changes on analysts' earnings forecasts in the presence of
other information sources, such as earnings announcements . The analysis presented
in this section indicates that the significance of the excess forecast revision after
dividend cuts ( documented in section 4.2) is likely to have been driven by the be
havior of the subsample of 128 events which were contemporaneous with an earnings
82
announcement . Dividend increases , by most measures I use in this study, do not
generate consistently significant revisions in analysts' earnings forecasts, whether
they are contemporaneous with an earnings announcement or not.
4.4 The regression analysis of the excess forecast
rev1s1on
In this section I present the statistical analysis of the mean and the median consensus
excess forecast revision after the announcement of a dividend change. I examine the
possible statistical relationship between the forecast revision and dividend-related
and earnings-related explanatory variables, in linear regression models. The ex planatory variables are proxies for the potential information signals of the dividend
announcement and the contemporaneous quarterly earnings announcement .
The analysis presented in this section with respect to the excess forecast revision
was also performed using the unadjusted forecast revision, for completeness. The
unadjusted forecast revision is not my variable of interest because it does not take
into account the intra-year secular revision activity, but it has been used by previous
researchers (see Lang and Litzenberger [4 7], and Yoon and Starks [73]) . All the
evidence reported next , with a few minor exceptions which do not change the essence
of the findings, are also established in the case of the unadjusted forecast revision.
Those results are available upon request.
In addition, the regression analysis presented next was applied to the excess
forecast revision measured using Brous' method for the assessment of the secular
revision. The results are not discussed in in the body of the paper, but they are
presented in Tables A.5-A.8 of the Appendix. The conclusions reported below are
identical to the patterns established from analyzing the forecast revision using Brous'
method.
83
Lastly, all the models and findings presented below were also fitted using per
centage figures as opposed to the dollar figures I report here. In other words, the
dependent variable was the percentage forecast revision, and the explanatory vari
ables were the percentage dividend change and the percentage year-to-year quarterly
income change. Again, the findings do not change in any material way, and they are
available upon request.
Prior studies of the effect of announcements of dividend changes on analysts'
earnings forecasts, and in particular Denis, Denis and Sarin [26] (i.e. the only
paper which controls for the secular revision activity in the measure of the forecast
revision) only examine the statistical significance of the mean of a measure of the
forecast revision after the dividend announcement.
2
The regression models presented here either corroborate or clarify the findings of
the univariate analysis of the previous sections. The bulk of the evidence indicates
that the magnitude and the sign of the excess forecast revision is correlated only with
dividend cuts which were announced contemporaneously with the nearest earnings
announcement. Tables 15 and 17 present the ordinary least squares analysis of the
one-month mean (Table 15) and median (Table 17) consensus excess forecast revision
after the announcement of a dividend change. Tables 16 and 18 present the same
analysis for the two-month mean (Table 16) and median (Table 18) consensus excess
forecast revision. All models presented have been adjusted for heteroskedasticity
using White's correction method.
The first independent variable (Variable B in Tables 15 through 18) is the dollar
amount of the dividend change, defined as the current, new dividend level minus
the previous quarter's dividend, adjusted for stock splits. This is a variable for
the hypothesized information signal in the dividend change decision. The second
variable is the change in Earnings per Share Before Extraordinary Items ( Quarterly
2
As discussed in Chapter 1, Ofer and Siegel [63] regress the forecast error in each of the six
months around the dividend announcement month on a unique independent variable which accounts
for the unexpectedness of the dividend change. That analysis is not directly comparable with what
I report in this section because the dependent variable I use is the excess forecast revision, as
opposed to the forecast error that they use.
84
Table 4.7: 01S analysis of the excess mean consensus forecast rev1s10n around the
dividend month: one-month revision measure
This table presents the O1S regression analysis of the one-month measure of the dollar excess forecast revision
around the month of the dividend announcement. The sample is comprised of 2,866 announcements of a dividend
change of more than 10% or less than -10%, between 1979 and 1990. The dependent variable in all the models is
the difference between the change in the dollar mean consensus analysts' forecast of year end earnings, measured
from the month of the announcement of the dividend change to the month following the month of the dividend
change, and the average one-month forecast revision in the period preceding the dividend announcement. This is
the earnings forecast 's excess revision after the announcement of the dividend change. The first independent variable
(Variable B) is the dollar amount of the dividend change (i.e. new dividend minus previous quarter's dividend).
Variable C is the change in the quarterly Earnings per Share before Extraordinary Items (COMPUSTAT Item 8),
measured in dollars. Variable D is a dummy which takes the value 1 if the firm had positive earnings in the year
of the dividend change, and O if it had losses. Variable E is a dummy which takes the value 1 if the dividend was
increased, and O if it was cut. Variable F is a dummy which takes the value 1 if the announcement was of a dividend
cut and it occured within a 20-trading day window around the nearest earnings announcement, and O otherwise.
Variable G is a dummy which takes the value 1 if the announcement was of a dividend cut and it occured outside a
20-trading day window around the nearest earnings announcement, and O otherwise. Variable H is a dummy which
takes the value 1 if the announcement was of a dividend increase and it occured within a 20-trading day window
around the nearest earnings announcement, and O otherwise. The numbers in parentheses are the t-statistics of the
zero coefficient null hypothesis. The adjusted R
2
is reported in the last row.
Variables
I
Model I II III IV V VI VII
A. Intercept -0.011 -0.010 -0.101 -0.004 -0.051 -0.018 -0.005
(-2.52) (-3.72) (-2.04) (-0.90) (-1.68) (-1.34) (-0.64)
B. Dollar Dividend Change 0.002 -0.002 -0.005 -0.012 -0.011
(0.27) (-0.98) (-0. 76) (-2.75) (-1.69)
C. Dollar Quarterly EPS Change 0.37 0.41 0.32 0.31
(6.11) (5.84) (4.09) (4.22)
D. Quarterly Profit or Loss Dummy -0.120 -0.037
(-6.12) (-1.48)
E. Dividend Up or Down Dummy -0.420
(-4.77)
F. 1 = Cut and same month -0.161 -0.150
0 = All other combinations (-8.23) (-6.15)
1 = Cut and not same month 0.009 0.016
0 = All other combinations (0.83) (0.57)
H. 1 = Increase and same month 0.012 0.009
0 = All other combinations (0.97) (0.98)
I. Adjusted R
2
-0.00 0.03 0.03 0.02 0.03 0.05 0.04
J. F-statistic 0.07 40.18 23.94 18.74 38.24 24.18 17.53
85
Table 4.8: 01S analysis of the excess mean consensus forecast rev1s10n around the
dividend month: two-month revision measure
This table presents the O1S regression analysis of the two-month measure of the dollar excess forecast revision
around the month of the dividend announcement. The sample is comprised of 2,866 announcements of a dividend
change of more than 10% or less than -10%, between 1979 and 1990. The dependent variable in all the models is
the difference between the change in the dollar mean consensus analysts' forecast of year end earnings, measured
from the month of the announcement of the dividend change to the month following the month of the dividend
change, and the average one-month forecast revision in the period preceding the dividend announcement. This is
the earnings forecast 's excess revision after the announcement of the dividend change. The first independent variable
(Variable B) is the dollar amount of the dividend change (i.e. new dividend minus previous quarter's dividend).
Variable C is the change in the quarterly Earnings per Share before Extraordinary Items (COMPUSTAT Item 8),
measured in dollars. Variable D is a dummy which takes the value 1 if the firm had positive earnings in the year
of the dividend change, and O if it had losses. Variable E is a dummy which takes the value 1 if the dividend was
increased, and O if it was cut. Variable F is a dummy which takes the value 1 if the announcement was of a dividend
cut and it occured within a 20-trading day window around the nearest earnings announcement, and O otherwise.
Variable G is a dummy which takes the value 1 if the announcement was of a dividend cut and it occured outside a
20-trading day window around the nearest earnings announcement, and O otherwise. Variable H is a dummy which
takes the value 1 if the announcement was of a dividend increase and it occured within a 20-trading day window
around the nearest earnings announcement, and O otherwise. The numbers in parentheses are the t-statistics of the
zero coefficient null hypothesis. The adjusted R
2
is reported in the last row.
Variables
I
Model I II III IV V VI VII
A. Intercept -0.018 -0.087 -0.007 -0.006 -0.012 -0.002 0.001
(-3. 78) (-5.21) (-3.26) (-1.1 7) (-1.09) (-0.24) (0.11)
B. Dollar Dividend Change 0.009 0.002 -0.002 -0.011 -0.009
(1.36) (0.51) (-0.30) (-1.19) (-1.23)
C. Dollar Quarterly EPS Change 0.39 0.31 0.32 0.37
(5.72) ( 4.99) (4.11) ( 4.4 7)
D. Quarterly Profit or Loss Dummy -0.214 -0.138
(-9.94) (-5.01)
E. Dividend Up or Down Dummy -0.092
(7.72)
F. 1 = Cut and same month -0.201 -0.131
0 = All other cases (-8.25) (-4.87)
1 = Cut and not same month -0.088 -0.027
0 = All other cases (-1.67) (-0.85)
H. 1 = Increase and same month -0.002 -0.003
0 = All other cases (-0.23) (-0.29)
I. Adjusted R
2
0.00 0.02 0.02 0.04 0.04 0.05 0.05
J. F-statistic 1.84 29.37 13.68 50.33 46.53 23.87 25.33
86
Table 4.9: 01S analysis of the excess median consensus forecast revision around the
dividend month: one-month revision measure
This table presents the OLS regression analysis of the one-month measure of the dollar excess median consensus
forecast revision around the month of the dividend announcement. The sample is comprised of 2,866 announcements
of a dividend change of more than 10% or less than -10%, between 1979 and 1990. The dependent variable in all
the models is the difference between the change in the dollar mean consensus analysts' forecast of year end earnings,
measured from the month of the announcement of the dividend change to the month following the month of the
dividend change, and the average one-month forecast revision in the period preceding the dividend announcement.
This is the earnings forecast 's excess revision after the announcement of the dividend change. The first independent
variable (Variable B) is the dollar amount of the dividend change (i.e. new dividend minus previous quarter's
dividend). Variable C is the change in the quarterly Earnings per Share Before Extraordinary Items (COMPUSTAT
Item 8), measured in dollars. Variable D is a dummy which takes the value 1 if the firm had positive earnings in the
year of the dividend change, and O if it had losses. Variable Eis a dummy which takes the value 1 if the dividend was
increased, and O if it was cut. Variable F is a dummy which takes the value 1 if the announcement was of a dividend
cut and it occured within a 20-trading day window around the nearest earnings announcement, and O otherwise.
Variable G is a dummy which takes the value 1 if the announcement was of a dividend cut and it occured outside a
20-trading day window around the nearest earnings announcement, and O otherwise. Variable H is a dummy which
takes the value 1 if the announcement was of a dividend increase and it occured within a 20-trading day window
around the nearest earnings announcement, and O otherwise. The numbers in parentheses are the t-statistics of the
zero coefficient null hypothesis. The adjusted R
2
is reported in the last row.
Variables
I
Model I II III IV V VI VII
A. Intercept -0.010 -0.017 -0.048 -0.003 -0.012 -0.0ll -0.003
(-1.89) (-3.ll) (-2.74) (-0.62) (-1.58) (-0.91) (-0.35)
B. Dollar Dividend Change -0.004 -0.012 -0.0ll -0.022 -0.018
(-0.56) (-1.06) (-1.35) (-2.96) (-2.20)
C. Dollar Quarterly EPS Change 0.26 0.21 0.17 0.20
(6.32) (6.83) (5.98) (5.43)
D. Quarterly Profit or Loss Dummy -0.ll 7 -0.022
(-4. 76) (-0.69)
E. Dividend Up or Down Dummy -0.083
(-3.76)
F. 1 = Cut and same month -0.157 -0.172
0 = All other combinations (-6.43) (-5.63)
1 = Cut and not same month 0.018 0.016
0 = All other combinations (0.29) (0.44)
H. 1 = Increase and same month 0.005 0.008
0 = All other combinations (1.03) (0.76)
I. Adjusted R
2
-0.00 0.04 0.04 0.01 0.03 0.04 0.02
J. F-statistic 0.31 58.74 34.48 ll.49 40.44 19.36 12.77
87
Table 4.10: 01S analysis of the excess mean consensus forecast rev1s10n around the
dividend month: two-month revision measure
This table presents the OLS regression analysis of the two-month measure of the dollar excess median consensus
forecast revision around the month of the dividend announcement. The sample is comprised of 2,866 announcements
of a dividend change of more than 10% or less than -10%, between 1979 and 1990. The dependent variable in all
the models is the difference between the change in the dollar mean consensus analysts' forecast of year end earnings,
measured from the month of the announcement of the dividend change to the month following the month of the
dividend change, and the average one-month forecast revision in the period preceding the dividend announcement.
This is the earnings forecast 's excess revision after the announcement of the dividend change. The first independent
variable (Variable B) is the dollar amount of the dividend change (i.e. new dividend minus previous quarter's
dividend). Variable C is the change in the quarterly Earnings per Share Before Extraordinary Items (COMPUSTAT
Item 8), measured in dollars. Variable D is a dummy which takes the value 1 if the firm had positive earnings in the
year of the dividend change, and O if it had losses. Variable Eis a dummy which takes the value 1 if the dividend was
increased, and O if it was cut. Variable F is a dummy which takes the value 1 if the announcement was of a dividend
cut and it occured within a 20-trading day window around the nearest earnings announcement, and O otherwise.
Variable G is a dummy which takes the value 1 if the announcement was of a dividend cut and it occured outside a
20-trading day window around the nearest earnings announcement, and O otherwise. Variable H is a dummy which
takes the value 1 if the announcement was of a dividend increase and it occured within a 20-trading day window
around the nearest earnings announcement, and O otherwise. The numbers in parentheses are the t-statistics of the
zero coefficient null hypothesis. The adjusted R
2
is reported in the last row.
Variables
I
Model I II III IV V VI VII
A. Intercept -0.010 -0.01 7 -0.037 -0.004 -0.008 0.000 0.003
(-1.876) (-2.81) (-2.74) (-0.72) (-1.01) (O.Dl) (0.32)
B. Dollar Dividend Change -0.003 -0.009 -0.008 -0.056 -0.015
(-0.40) (-1.41) (-1.10) (-2.77) (-1.61)
C. Dollar Quarterly EPS Change 0.35 0.39 0.35 0.27
(5.33) (5.74) ( 4.41) (4.86)
D. Quarterly Profit or Loss Dummy -0.098 -0.033
(-4.22) (-1.13)
E. Dividend Up or Down Dummy -0.121
(-4.73)
F. 1 = Cut and same month -0.083 -0.084
0 = All other cases (-3.55) (-2.90)
1 = Cut and not same month -0.098 -0.096
0 = All other cases (-3.27) (-3.15)
H. 1 = Increase and same month -0.003 -0.003
0 = All other cases (-0.78) (-0.26)
I. Adjusted R
2
-0.00 0.03 0.02 0.01 0.02 0.03 0.01
J. F-statistic 0.16 25.65 11.94 8.99 18.84 11.03 6.18
88
COMPUSTAT Item 8) between the latest earnings announcement (i .e. before the
dividend announcement) and the corresponding announcement of four quarters ago.
This variable is measured in U.S. dollars, per COMPUSTAT.
The third explanatory variable is a dummy which takes the value 1 if the latest
(i.e. before the dividend announcement) quarterly earnings announced were negative
(i.e. losses), and the value O if earnings were positive. The fourth independent
variable is also a dummy which takes the value 1 if the announcement was of a
dividend cut of more than -10%, and O if the announcement was of a dividend
increase of more than 10%.
The last three independent variables are interaction dummies. Variable F takes
the value 1 if the dividend was cut and the announcement was within a 40-trading
day window around the nearest earnings announcement , and O otherwise. Variable
G takes the value 1 if the dividend was cut and the announcement was outside a
40-trading day window around the nearest earnings announcement , and O other
wise. Finally, Variable H takes the value 1 if the dividend was increased and the
announcement was within a 40-trading day window around the nearest earnings
announcement , and O otherwise. In the models where the interaction dummies are
included the regression intercept captures the effect of a dividend increase announced
outside a 40-trading day window around the nearest earnings announcement.
4.4.1 Analysis of the one-month excess forecast revision
The models presented in Tables 15 and 17 ( as well as those of Tables 16 and 18,
which are discussed in the following section) refine and corroborate the findings of
the univariate analysis presented in the previous sections. In particular, the only
dividend signal-related variable which is a significant explanatory variable of the
forecast revision, across the board, is a dummy variable which takes the value 1 when
the dividend was cut, and the announcement occurred within 40 trading days around
the nearest earnings announcements , and O otherwise. In addition, the magnitude of
the dividend change is not a significant explanatory variable of the forecast revision.
89
These patterns corroborate the findings of the previous sections, where the bulk of
the evidence pointed toward dividend cuts announced near quarterly earnings as
being the only dividend-related signals incorporated in analysts ' earnings forecasts,
in my sample.
The models discussed are successful in documenting the relationship between the
explanatory variables and the various measures of the forecast revision. With the
exception of the regressions that include the dividend change variable as the only
independent variable (for all combinations of the measures of the forecast revision,
i.e. throughout Tables 15-18), the F-statistics are highly significant, ranging from
9 to 50, indicating significant model explanatory power. The low predictive power
of the models (indicated by R
2
's that range from 1 % to 5% across the four tables)
implies either the existence of significant amount of noise in the forecast data, or
that earnings forecast revisions are affected by a variety of information sources which
are not captured by just the dividend- and earnings-related independent variables
included in the models presented in this section.
The coefficient of the dollar dividend change variable is significant only in one of
the regression models of the one-month measure of the mean consensus excess fore
cast revision (Model VI, Table 15). Nevertheless, the negative sign of the coefficient
estimate is inconsistent with the predictions of the dividend signaling hypothesis as
it implies that the magnitude of the excess forecast revision is inversely related to the
magnitude of the dividend change. In the case of the one-month median consensus
excess forecast revision (Table 17), the dollar dividend change is significant in the
last two models (VI and VII) , but the signs of the coefficient estimates are again
negative and, therefore, contradictory to the predictions of the ICD hypothesis.
The Dividend Up-Down dummy variable (Variable E) is significant in Model V,
m both Tables 15 and 17. But, again, its negative sign indicates that dividend
increases generate less favorable forecast revisions, and vice versa. Clearly, neither
the magnitude nor the direction of the dividend change by themselves , in my sample,
are significant explanatory variables of the excess forecast revision after a dividend
change announcement .
90
In contrast, the earnings per share change variable (Variable C) enters signifi
cantly, and with the correct sign, in all four models where it is used, for both the
mean and the median consensus forecast revision. The coefficients of the EPS change
variable are economically significant .
The last two models in Tables 15 and 17 (Models VI and VII) are the most
informative, because they capture the joint effects of the earnings and dividend
announcement signals on the excess forecast revision after the announcement of a
dividend change. The first interaction dummy ( which is equal to 1 if the dividend
was cut and announced within a 40-trading day window around the nearest earnings
announcement , and O otherwise) is highly significant in both models. The only
other variable which remains significant in these multivariate regression models is
the quarterly earnings per share change (Model VI) .
The significance of the Dividend Cut- Same Month dummy (Variable F) in the
regression models of the mean and the median consensus excess forecast revision is
consistent with the main thrust of the evidence presented in Tables 13 and 14. It
was documented there that the excess forecast revision is significantly different from
zero almost exclusively in the case of dividend cuts which were announced within a
40-trading day window around the nearest earnings announcement .
Nevertheless, the findings of this section may indicate that the documented signif
icant forecast revision after dividend cuts which were announced contemporaneously
with an earnings announcement may not be caused by the potential information
content of the dividend announcement. Instead, they may largely be due to the
information content of the earnings announcement . It is possible that the experi mental design has not fully adjusted for the earnings announcement effect, and the
forecast revision after the dividend announcement is reflecting part of the effect of
the earnings announcement .
This issue can only be investigated with the use of an historical earnings forecast
database which would include individual forecasts , as opposed to consensus forecasts,
and also would cite the exact release date for each forecast . Since I do not have access
to such a database this is an issue arising from my research which I point out as
91
a potential shortcoming, and a subject for future investigation. Lastly, the same
qualification applies to the findings of the next section.
4.4.2 Analysis of the two-n1onth excess forecast rev1s1on
Tables 16 and 18, which present the results of fitting the models of the previous
section on the two-month measures of the mean and the median consensus excess
forecast revision, tell almost the same story. The dividend change variable is in
significant in all models, except in Model VI of Table 18. It enters this model with a
negative sign which, again, is contrary to the predictions of the ICD hypothesis, and
most likely due to the effect of the interaction dummies , and especially the ones for
the dividend cut . Here again, the models that include the dividend change variable
as the only independent variable are the only ones with no explanatory power, as
shown by the F-statistic. The predictive power of these models is also low. Lastly, in
this set of models again, the earnings per share change variable is significant across
models , as is the Profit/Loss dummy.
The Dividend Cut interaction dummies (Variables F and G) enter the median
consensus forecast revision models VI and VII (Table 18) with coefficients signif
icantly different from zero. The implication is that when the median consensus
forecast revision is measured over two months, the potential dividend reduction
announcement signal does add explanatory power to the regression model of the
magnitude of the forecast revision.
This finding warrants two comments. First, even when measured over two
months, thus capturing a theoretically stronger signal than the one-month mea
sure, the earnings forecast revision after a dividend increase is not correlated with
the dividend signal. Second, the two-month median forecast revision may be sig
nificant after both types of dividend cuts because their timing separation may not
have been successful, as explained next, because of the measurement period of the
revision.
92
In particular, the dividend announcements were qualified as "contemporaneous"
and "non-contemporaneous" with respect to the nearest earnings announcement if
they occurred within or outside a 40-trading day window around the earnings an
nouncement, respectively. When the forecast revision is measured from the month of
the dividend announcement to the following month, the 40-trading day window rule
is successful in separating the potential information effect of the dividend announce
ments based on their timing with respect to the nearest earnings announcement .
The information in earnings announcements which occurred 20 trading days before
or after the dividend announcement is likely to be incorporated in the earnings fore
cast released in the month after the dividend change, i.e. in the one-month forecast
revision .
In the two-month forecast revision case, though, the revision is measured from
the month before, to the month after the dividend announcement. Earnings an
nouncements which occurred outside of the 40-trading window around the dividend
announcement but still in the month before the earnings announcement, may be
incorporated in the earnings forecast of the month prior to the dividend announce ment month, and therefore the whole two-month revision measure . This potential
effect may be present and widespread in my sample, and it may cause the statis
tical significance of the two-month median forecast revision to be independent of
the timing of the announcement with respect to the nearest earnings announcement.
Nevertheless , it does not alter the main thrust of the evidence.
93
Chapter 5
Conclusions
This study examined the revisions of analysts ' earnings forecasts after announce
ments of dividend changes. The aim was to test the cash flow signaling hypothesis
of corporate dividend policy of Modigliani and Miller [58]. Significant analysts'
earnings forecast revisions after the announcement of a dividend change could be
interpreted as direct evidence of a link between the potential informational signal in
the dividend change and managers ' expectations of firms' earnings . In other words,
if managers use changes in dividend policy to signal their expectations of higher
future earnings then analysts should use the signal in updating their forecasts of
firms ' profitability.
I used measures of the earnings forecast revision which account for the secular
forecast revision activity which has been documented in the literature, and found
to be present in my sample. In addition, I used both the mean and the median
consensus earnings forecast from the I/B/E/S Inc. database for the calculation of
the forecast revision, measured over both the one month after, and the two months
around the dividend announcement month, for the first time in the literature.
The bulk of the evidence coming out of my study documents that announcements
of dividend changes, on average, do not cause earnings forecast revisions that are
significantly different from analysts' secular monthly forecast revision. Announce
ments of dividend changes have marginal informational impact on analysts' year-end
earnings forecasts only when the dividend is cut, and the announcement was made
within 20 trading days of the nearest quarterly earnings announcement .
94
In almost every other case, the potential signal conveyed through the dividend
policy change either does not yield significant average excess forecast revisions, or
it is not correlated with the post-dividend announcement excess forecast revision,
after controlling for the contemporaneous earnings information. In particular, the
magnitude of the dividend change is not a significant explanatory variable of the
excess forecast revision, even in regression models where it is the only independent
variable. A dummy variable for the direction of the dividend change is a significant
explanatory variable of the excess forecast revision, but it enters the regression
models with a negative sign, indicating that dividend cuts generate upward forward
revisions , and vice versa, a finding that is contrary to the predictions of the ICD
hypothesis .
The main empirical results of this study can be summarized as follows:
• The excess forecast revision, with one exception, is significant only after divi
dend cuts which were announced within a 40-trading day window around the
nearest earnings announcement. In other words , the bulk of the evidence in
dicates that announcements of dividend changes cause analysts to revise their
year-end forecasts of earnings only when the dividend is cut and the signal in
the dividend announcement is corroborated by a contemporaneous quarterly
earnings announcement. In almost every other case, the potential dividend
signal seems to be ignored by analysts .
• The excess forecast revision is not correlated with the magnitude of the div
idend change. In linear regression models of the excess forecast revision, the
coefficient of the dividend change magnitude is not statistically significant,
even in models where it is used as the only independent variable.
• In contrast, the two consistently significant variables in all regression models of
the dollar excess forecast revision are (a) the year-to-year change in quarterly
Net Income Before Extraordinary Items in the quarter of the dividend change,
and (b) a dummy variable which takes the value 1 if the dividend was cut
and its announcement was within a 40-trading day window around the nearest
95
earnings announcement, and O otherwise. In other words, dividend announce
ments, in my sample, carry information content only when they concern a
dividend cut, and coincide with a contemporaneous earnings announcement.
• Using measures of the post-dividend announcement forecast rev1s1on which
do not control for the documented secular forecast revision activity results
in drawing erroneous conclusions about the true significance of the forecast
revision. In particular, measures of the forecast revision which do not control
for the secular revision activity document significant forecast revisions after
both dividend increases and cuts.
• The interval over which the measures of the mean consensus forecast revi
sion are assessed does not affect the statistical significance of the documented
revision, in most cases.
The empirical regularities established here are not consistent with the findings of
(a) Lang and Litzenberger [47] who find no statistically significant forecast revision
after a dividend announcement using a forecast revision measure which does not
account for the secular revision activity, (b) Yoon and Starks who, using the Lang
and Litzenberger measure over a considerably larger sample document statistically
significant forecast revisions across both dividend increases and cuts, ( c) Denis,
Denis , and Sarin [26] who use an excess revision measure similar to the one I use,
and find economically significant revisions across dividend increases and cuts, and
( d) Ofer and Siegel [63], who use a different methodology than these papers, and
conclude that analysts significantly revise their forecasts after the announcement of
a dividend change.
The results provide support for part of the cash flow signalling hypothesis of
dividend announcements of Modigliani and Miller [58]. In particular, announcements
of dividend changes cause professional earnings expectations to shift in the direction
of the dividend change only in the case of dividend cuts corroborated by an earnings
announcement. This finding is consistent with Lintner's observation that although
managers are reluctant to frequently change dividend policy, they are particularly
96
loath to cut dividends. As a result of this managerial disposition, dividend cuts may
cause the market to form negative expectations about current and future earnings,
which may be reflected on analysts' earnings forecasts.
These findings are also consistent with the dividend signalling model of Warther
[70] , who posits that only dividend cuts are informative signals of firms' earnings
prospects. They are also consistent with the findings of DeAngelo, DeAngelo and
Skinner [24] who showed that dividend cuts are informative of future prospects.
97
Reference List
[1] Aharony, J., and I. Swary, "Quarterly Dividend and Earnings Announcements
and Stockholders' Returns: An Empirical Analysis", Jo'Urnal of Finance, Vol.35,
No.I, March 1980, 1-12.
[2] Allen, F. and R. Michaely, "Dividend Policy" , Working Paper 14-94, Rodney L.
White Center for Financial Research, The Wharton School of the University of
Pennsylvania, May 1994.
[3] Ang, James S. , "Do Dividends Matter? A Review of Corporate Dividend The
ories and Evidence", Monograph Series in Finance and Economics, Salomon
Brothers Center for the Study of Financial Institutions, New York University,
Monograph 1987-2.
[4] Asquith, P., and D. Mullins, Jr. , "The Impact of Initiating Dividend Payments
on Shareholders' Wealth" , ]o'Urnal of B'Usiness, 56 (1983) , 77-96.
[5] Bajaj, M. , "Dividend Omissions and Forecasts of Future Earnings: Some Positive
Information on the Information Content of Dividends", Working Paper 91-19,
University of Southern California, 1991.
[6] Bajaj, M. and A. M. Vijh, "Dividend clienteles and the Information Content of
Dividends", ]o'Urnal of Financial Economics, 26, 1990, 193-219.
[7] Bajaj, M. , A.M. Vijh and R.W. Westerfield, "Ownership Structure, Agency Costs
and Dividend Policy", Working Paper, University of Southern California, Febru
ary 1993.
[8] Barefield R, and E. Comiskey, "The accuracy of analysts'forecasts of earnings
per share" , Jo'Urnal of B'Usiness Resear·ch, 3, 1975, 241-252.
[9] Bhattacharya, S. , "Imperfect Information, Dividend Policy and the 'bird in the
hand fallacy" ', Bell Jo'Urnal of Economics, 10, 207-232.
[10] Black, F., "Estimating Expected Return", Financial Analysts Jo'Urnal,
September-October 1993, 36-38.
98
[11] Brickley, J. A., "Shareholder Wealth, Information Signaling and the Specially
Designated Dividend: An Empirical Investigation", Jo'Urnal of Financial Eco
nomics, vol.15, August 1983, 187-209.
[12] Brous, P.A., "Common Stock Offerings and Earnings Expectations: A Test of
the Release of Unfavorable Information" , Jo'Urnal of Finance, 47 (1992), 1517-
1536.
[13] Brown, R., J. Durbin, and J. Evans, "Techniques for Testing the Constancy
of Regression Relationships over Time", Jo'Urnal of the Royal Statistical Society,
Series B, Vol.37, No.2, 1975.
[14] Brown, L.D., Hagerman, R.L., Griffin, P.A. , and M.E. Zmijewski, "An Evalua
tion of Alternative Proxies for the Market 's Assessment of Unexpected Earnings" ,
]o'Urnal of Acco'Unting and Economics, 9 (1987) , 159-93.
[15] Brown, L., G. Richardson, and C. Trzcinka, "Strong-Form Efficiency on the
Toronto Stock Exchange: An Examination of Analyst Price Forecasts", Working
Paper, State University of New York at Buffalo, 1988.
[16] Brown, L.D., and M.S. Roze-ff, "The Superiority of Analyst Forecasts as Mea sures of Expectations: Evidence from Earnings" , Jo'Urnal of Finance, Vol.33,
No.1, March 1978, 1-16.
[17] Brown S. , and J. Warner, "Measuring Security Price Performance", Jour·nal of
Financial Economics, 1980.
[18] Charest, G., "Dividend Information, Stock Returns and Market Efficiency - II",
Jo'Urnal of Financial Economics, 6, 297-330.
[19] Chowdry, B., and V. Nanda, "Repurchase Premia as a Reason for Dividends",
unpublished working paper, University of Southern California and UCLA, Jan
uary 1993.
[20] Crichfield, T., T. Dyckman and J. Lakonishok, "An Evaluation of Security
Analysts ' Forecasts" , The Accounting Review, Vol. LIII, No.3, July 1978, p.651-
668.
[21] Christie, W.G. , "Dividend yield and expected returns: The zero-dividend puz
zle", Jo'Urnal of Financial Economics, 28, 1990, 95-125.
[22] Cragg, J.G. , and B.G. Malkiel, "The Consensus and Accuracy of Some Pre
dictions of the Accuracy of Corporate Earnings", The ]o'Urnal of Finance, 23,
March 1968, 67-84.
99
[23] DeAngelo, H. , and L. DeAngelo, "Dividend Policy and Financial Distress: An
Empirical Investigation of Troubled NYSE Firms" , ]o'Urnal of Finance, Vol.45,
No.5, December 1990, 1415-1431.
[24] DeAngelo, H., L. DeAngelo and D. J. Skinner, "Dividends and Losses", ]o'Urnal
of Finance, December 1992.
[25] DeAngelo, H. , L. DeAngelo and D. J. Skinner, "Reversal of Fortune: Divi
dend signaling and the disappearance of sustained earnings growth" , ]o'Urnal of
Financial Economics, ( 40) 1996, 341-371.
[26] Denis, David J. , Diane K. Denis, and Atulya Sarin, "The Information Con
tent of Dividend Changes: Cash Flow Signaling, Overinvestment , and Dividend
Clienteles", unpublished, Virginia Polytechnic Institute and State University,
February 1992.
[27] Dimson, E., and P. Marsh, "An Analysis of Brokers' and Analysts' Unpublished
Forecasts of U.K. Stock Returns" , ]o'Urnal of Finance, 39, 1984, 1257-1292.
[28] Dreman, D.N. , and M.A. Berry, "Analyst Forecasting Errors and Their Im
plications for Security Analysis", F inancial Analysts ]o'Urnal, May-June 1995,
p.30-41.
[29] Eades, K.M. , "Emprirical Evidence of Dividends as a Signal of Firm Value",
]o'Urnal of Financial and Q'Uantitative Analysis, November 1982, 471-500.
[30] Eades, K.M., P.J. Hess and E.H. Kim, "Market Rationality and Dividend An
nouncements", ]o'Urnal of Financial Economics, 1985, 14, 581-604.
[31] Easterbrook, Frank H., "Two Agency-Cost Explanations of Dividends", Amer
ican Economic R eview, Vol. 74, No. 4, September 1984.
[32] Elton, E.J.,and M.J. Gruber, "Earnings Estimates and the Accuracy of Expec
tational Data" , Managem ent Science, Vol. 18, No. 8, April 1972, 409-424.
[33] Elton, E.J., M.J. Gruber and S. Grossman, "Discrete Expectational Data and
Portfolio Performance" , ]o'Urnal of Finance, 41 , 1986, 699-714.
[34] Elton, E.J., M.J. Gruber and M.N. Gultekin, "Professional Expectations: Accu
racy and Diagnostics of Errors", ]o'Urnal of Financial and Q'Uantitative Analysis,
Vol.19, No.4, December 1984, 351-363.
[35] Fried, D., and D. Givoly, "Financial Analysts' Forecasts of Earnings: A Better
Surro()'ate for Market Expectations" ]o'Ur-nal of Acco'Unting and Economics 4
0 ' . , ' '
1982, 85-104.
100
[36] Givoly, D. , and J . Lakonishok, "The Information Content of Financial Analysts'
Forecasts of Earnings: Some Evidence on Semi-Strong Efficiency", ]o'Urnal of
Acco'Unting and Economics, 1, 1979, 165-185.
[37] Givoly, D., and J. Lakonishok, "Financial Analysts' Forecasts of Earnings:
Their Value to Investors" , ]o'Urnal of Banking and Finance, 4, 1980, 221-233.
[38] Gonedes, N., "Corporate Signaling, External Accounting, and Capital Market
Equilibrium: Evidence on Dividends, Income, and Extraordinary Items", ]o'Urnal
of Acco'Unting Research, Vol.16, No.l , Spring 1978, 26-79.
[39] Griffin, P.A., "Competitive Information in the Stock Market: An Empirical
Study of Earnings, Dividends and Analysts' Forecasts" , ]o'Urnal of Finance,
Vol.31 , No.2, May 1976, 631-650.
[40] Hayn, Carla, "The information content of losses", Jour·nal of Acco'Unting and
Economics, 20 (1995), 125-153.
[41] Rausch, D.B., and J.K. Seward, "Signaling with Dividends and Share Repur
chases: A Choice between Deterministic and Stochastic Cash Disbursements",
Review of Financial Studies, Vol.6, No.1 , 1993, 121-154.
[42] Healy, P. M. and K. G. Palepu, "Earnings Information Conveyed by Dividend
Initiations and Omissions", ]o'Urnal of Financial Economics, 21, 1988, 149-175.
[ 43] Jensen, M. , "Agency Costs of Free Cash Flow, Corporation Finance and
Takeovers" , Amer-ican Economic Review, 76, 1986, 323-329.
[44] Jensen, G.R. , and J.M. Johnson, "The Dynamics of Corporate Dividend Re
ductions" , Financial Managem ent, Vol. 24, No. 4, Winter 1995, 31-51.
[45] John, K. and J. Williams, "Dividends, dilution and taxes: A signaling equilib
rium", ]o'Urnal of Finance, 40, 1985, 1053-1070.
[46] Kane, A., Y.K. Lee, and A. Marcus, "Earnings and Dividend Announcements:
Is There a Corroboration Effect", ]o'Urnal of Finance, Vol.39, No.4, September
1984, 1091-1099.
[47] Lang, L. and R. Litzenberger, "Dividend Announcements: Cash Flow Signaling
vs . Free Cash Flow Hypothesis" , ]o'Urnal of Financial Economics, September
1989, 181-191.
[48] Leftwich, R. and M. E. Zmijewski, "Contemporaneous Announcements of Div
idends and Earnings" , Working Paper, University of Chicago, November 1991.
[49] Lintner, J. , "Distribution of Incomes of Corporations Among Dividends, Re
tained Earnings, and Taxes", American Economic Review, 46, May 1956.
101
[50] Litzenberger, R.H. and K. Ramaswamy, "The effects of dividends on common
stock prices: Tax effects or information effects" , Jo'Urnal of Finance, 37, 1982,
429-443.
[51] McDonald, C.L .. "An Empirical Examination of the Reliability of Published
Predictions of Future Earnings" , The Acco'Unting Review, July 1973, 502-510.
[52] Michaely, R., R.H. Thaler, and K.L. Womack, "Price reactions to dividend
initiations and ommissions: Overreaction or drift?" , Jo'Urnal of Finance, 50,
573-608.
[53] Miller, M.H., "The Informational Content of Dividends", in John Bosons, Rudi
ger Dornbusch, and Stanley Fischer, eds., Macroeconomics: Essays in Honor of
Franco Modigliani, MIT Press, 1980.
[54] Miller, M.H., "Behavioral Rationality in Finance: The Case of Dividends",
]o'Urnal of B'Usiness, 1986, vol.59, No.4, S451-S468.
[55] Miller, M.H. and F. Modigliani , "Dividend Policy, growth, and the valuation of
shares" , Jo'Urnal of B'Usiness, 34, No.4, October 1961 , 411-433.
[56] Miller, M.H. , and K. Rock , "Dividend Policy under Asymmetric Information",
Jo'Urnal of Finance, 40, 1985, 1031-1051.
[57] Miller, M.H., and M. Scholes, "Dividends and Taxes: Some empirical evidence",
Jo'Urnal of Political Economy, 90, 1982, 1118-1141.
[58] Modigliani, F. and M. H. Miller, "The Cost of Capital, Corporation Finance
and the Theory of Investment: A reply", American Economic Review, 49, 1959,
655-669.
[59] Morgan, LG. , "Dividends and capital asset prices", Jo'Urnal of Finance, 37,
1982, 1071-1086.
[60] O'Brien, P.C. , "Analyst 's Forecasts as Earnings Expectations", Jo'Urnal of Ac co'Unting and Economics, 10(1988), 53-83.
[61] O'Brien, P.C. , "Forecast Accuracy of Individual Analysts in Nine Industries",
]o'Urnal of Acco'Unting Research, Vol.28, No.2, Autumn 1990, 286-304.
[62] O'Brien, P.C., and R. Bhushan, "Analyst Following and Institutional Owner
ship", Jo'Urnal of Acco'Unting Resear·ch, Vol.28, 1991, 55-76.
[63] Ofer, A. R. and D. R. Siegel, "Corporate Financial Policy, Information, and
Market Expectations: An Empirical Investigation of Dividends", Jo'Urnal of Fi
nance, 1987, No.4, 889-911.
102
[64] Penman, S.H. , "The Predictive Content of Earnings Forecasts and Dividends",
Jo'Urnal of Finance, Vol.38, No.4, September 1983, 1181-1199.
[65] Pettit, R. , "Dividend Announcements, Security erformance and Capital Mar
ket Efficiency", Jo'Urnal of Finance, December 1972, 993-1007.
[66] Richards, R.M., "Analysts' Performance and the Accuracy of Corporate Earn
ings Forecasts", The ]o'Urnal of B'Usiness, Vol. 49, July 1976, 350-357.
[67] Riding, A.L. , "The Information Content of Dividends: An Other Test", Jo'Urnal
of B'Usiness Finance and Acco'Unting, 11(2), Summer 1984, 163-176.
[68] Ross, S.A., "The Determination of Financial Structure: The Incentive Sig
nalling Approach" , The Bell Jo'Urnal of Economics, Vol.8, No.I, Spring 1977.
[69] Spence, A.M. , "Job Market Signaling", Q'Uar·ter-ly Jo'Urnal of Economics, 87,
August 1973, 355-379.
[70] Warther, V. , "Dividend Smoothing: A Sleeping Dogs Explanation", Ph.D. Dis
sertation, Graduate School of Business, University of Chicago, December 1991.
[71] Watts, R., "The Information Content of Dividends", ]o'Urnal of B'Usiness, 46,
191-211.
[72] Womack, K.L. , "Do brokerage analysts' recommendations have investment
value ? , Jo'Urnal of Finance, 51 , March 1996, 137-167.
[73] Yoon, P.S., and L.T. Starks, "Signaling, Investment Opportunities, and Divi
dend Announcements", The Review of Financial Studies, Winter 1995, Vol. 8,
No. 4, 995-1018.
103

Abstract (if available)

Abstract This thesis examines the impact of announcements of dividend changes on analysts' earnings forecasts. In particular, I inquire whether security analysts revise their forecasts of annual earnings per share, after the announcement of shifts in dividend policy, significantly differently than in prior months, after controlling for other relevant information sources such as contemporaneous quarterly earnings announcements. This is a direct test of the Information Content of Dividends Hypothesis of Modigliani and Miller, who proposed that shifts in dividend policy should signal changes in managerial expectations of future earnings. ❧ I find that changes in corporate dividend policy, on average, do not generate revisions in earnings forecasts which are significantly different from the documented secular forecast revision activity in the months prior to the dividend announcement month. The only instance in which forecast revisions are statistically significantly different from their secular forecast revision activity is after dividend cuts which were announced within a calendar month of a quarterly earnings announcement. The findings of this thesis do not support the predictions of the Information Content of Dividends Hypothesis.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Share repurchases: how important is market timing?

PDF

Cash holdings and corporate diversification

PDF

Shrinkage methods for big and complex data analysis

PDF

Innovation: financial and economics considerations

PDF

Essays on revenue management with choice modeling

PDF

Theory-practice gap: MBA curricula as preparation for business practice in marketing

PDF

Personalized normative feedback applied to undergraduates with problem drinking: a comparison with psychoeducation and an examination of cognitive-affective change mechanisms via the articulated ...

Asset Metadata

Creator Konstantinides, Perikles Fotiou (author)

Core Title Dividend policy, earnings announcements, and analysts' earnings forecasts

School Marshall School of Business

Degree Doctor of Philosophy

Degree Program Business Administration

Publication Date 04/19/2019

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag corporate finance,dividend policy,information content of dividends,OAI-PMH Harvest

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor DeAngelo, Harry (committee chair), Murphy, Kevin (committee member), Papavassilopoulos, George (committee member), Weinstein, Mark (committee member)

Creator Email pfk@syracusemain.com,pfkonstant@gmail.com

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-139547

Unique identifier UC11675519

Identifier etd-Konstantin-7205.pdf (filename),usctheses-c89-139547 (legacy record id)

Legacy Identifier etd-Konstantin-7205.pdf

Dmrecord 139547

Document Type Dissertation

Format application/pdf (imt)

Rights Konstantinides, Perikles Fotiou

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA