Dividend Changes Do Not Signal Changes in Future - Faculty

Gustavo GrullonRice University

Roni MichaelyCornell University and the Interdisciplinary Center

Shlomo BenartziUniversity of California at Los Angeles

Richard H. ThalerUniversity of Chicago and the National Bureauof Economic Research

Dividend Changes Do Not SignalChanges in Future Profitability*

I. Introduction

One of the most important issues in corporate fi-nance is whether dividend changes contain infor-mation about future earnings and profitability.Although dividend signaling theories imply thatdividend increases signal better prospects (e.g.,Bhattacharya 1979; John and Williams 1985;andMiller and Rock 1985), many empirical stud-ies have failed to support this idea. Studies byWatts (1973), Gonedes (1978), Penman (1983),DeAngelo, DeAngelo, and Skinner (1996), Benartzi,Michaely, and Thaler (1997), and Grullon,Michaely,and Swaminathan (2002) find little or no evidencethat dividend changes predict abnormal increasesin earnings.1 Similarly, evidence based on sur-veying and interviewing hundreds of financial

(Journal of Business, 2005, vol. 78, no. 5)B 2005 by The University of Chicago. All rights reserved.0021-9398/2005/7805-0002$10.00

1659

* We thank John Kling, Ivo Welch, William Weld, JamesWeston, Monique Weston, and especially an anonymous refereefor many helpful comments and suggestions. Any errors are ourown.1. For a comprehensive literature review on dividend sig-

naling theories, see Allen and Michaely (2003).

One of the most im-portant predictions of thedividend-signaling hy-pothesis is that dividendchanges are positivelycorrelated with futurechanges in profitabilityand earnings. Contrary tothis prediction, we showthat, after controllingfor the well-knownnonlinear patterns in thebehavior of earnings,dividend changescontain no informationabout future earningschanges. We also showthat dividend changesare negatively correlatedwith future changes inprofitability (return onassets). Finally, weinvestigate whether in-cludingdividend changesimproves out-of-sampleearnings forecasts. Wefind that models that in-clude dividend changesdo not outperform thosethat do not includedividend changes.

executives indicates that managers reject the notion that dividends areused as a costly signaling device (seeBrav,Graham,Harvey, andMichaely2003).However, in a recent paper, Nissim and Ziv (2001) consider a par-

ticular model of earnings expectations and find a positive associationbetween current dividend changes and future earnings changes. Nissimand Ziv argue that previous studies failed to uncover the true relationbetween dividends and future earnings because researchers used thewrong model to control for the expected changes in earnings. Specifi-cally, they report that, when using a regression analysis that controlsfor a particular (linear) form of mean reversion in earnings, dividendchanges are positively correlated with future earnings changes.The findings in Nissim and Ziv (2001) are surprising, since past re-

searchers who have used control variables in a regression context findeither the opposite relation (Penman 1983), no relation, or a very weakrelation (Benartzi et al. 1997). Recognizing the potential nonlinearitiesin the relation between dividends and earnings, many of the prior in-vestigators used methods other than regression analysis and found re-sults opposite to the ones in Nissim and Ziv. For example, DeAngeloet al. (1996) analyzed dividend policy during times where firm earningsunexpectedly decline and found that dividend changes contain virtuallyno information about future changes in earnings. Benartzi et al. useda matched-sample approach, in which dividend changing firms arematched to non–dividend changing firms based on their attributes suchas market capitalization, industry, and past earnings performance—thusexplicitly controlling for the earnings pattern and mean reversion—andfound no evidence of positive abnormal earnings changes after dividendincreases. Similar results were obtained by Grullon et al. (2002). In fact,even a simple comparison of the evolution of earnings for dividendchanging firms (e.g., Grullon et al. 2002) yields results opposite to thosedrawn by Nissim and Ziv (2001). These consistent findings acrossstudies and methodologies make the Nissim and Ziv results surprising.Given the importance of this issue to corporate finance in general and

to the question of why firms pay dividends in particular, we revisitthe issue here. One of the main challenges in any study examining therelation between dividend changes and future earnings is choosing theappropriate method of estimating expected earnings. Since each modelof expected earnings defines the portion of total realized earnings con-sidered unexpected, it is important that researchers choose an earningsmodel that captures the features of the earning process. If, for example,earnings are assumed to follow a randomwalk, then any change in earn-ings from one year to another is unexpected. If earnings follow a randomwalk with a drift, then unexpected earnings are defined to be those thatdiffer from the prior earnings grown at the expected growth rate, and soforth. In this paper, we show that Nissim and Ziv’s assumption of linear

1660 Journal of Business

mean reversion in earnings is inappropriate. The reason for this is thatthe mean reversion process of earnings is highly nonlinear (see Brooksand Buckmaster 1976; Elgers and Lo 1994; and Fama and French 2000.)Since assuming linearity when the true functional form is nonlinear hasthe same consequences as leaving out relevant independent variables, it ispossible that the positive correlation documented inNissim andZiv (2001)between dividend changes and future earnings changes is spurious.We address this issue by using a model of unexpected earnings that

explicitly controls for the nonlinear patterns in the behavior of earnings.We correct for the nonlinear evolution in earnings using the modifiedpartial adjustment model proposed by Fama and French (2000). Thismodel assumes that the rate of mean reversion and the coefficient of auto-correlation are highly nonlinear. We use this approach because Fama andFrench (2000) empirically show that it explains the evolution of earningsmuch better than a model with a uniform rate of mean reversion.We show that, after controlling for the nonlinear patterns in the be-

havior of earnings, the relation between dividend changes and futureearnings disappears. The relation between dividend increases and earn-ing increases in Year 1 is positive and significant in only 5 out of 35 years(interestingly, this is the same number years in which the relation betweenthese two variables is negative and significant). The results indicate thatearning increases do not follow dividend increases in any systematic way.We also find that dividend changes are negatively correlated with futurechanges in profitability (return on assets, return on cash-adjusted assets,and return on sales). As a robustness check, we examine the relationbetween future earnings levels and changes in dividends and find similarresults: dividends changes are very poor indicators of both earnings andprofitability levels. Overall, we find no evidence supporting the idea thatdividend increases signal better prospects for firm profitability.The empirical tests we discussed so far depend on the assumption that

all the parameters of the model are known at the time of the forecast, anassumption likely to be violated if the parameters of the model changeover time. To this end, we perform an out-of-sample test of the power ofdividend changes to predict future earnings changes. We investigatewhether investors can improve the precision of their real-time earningsforecasts by using dividend changes, given all other available informa-tion. We therefore examine the forecasting ability of dividend changes,explicitly accounting for the parameter uncertainty faced by investorswho have access to only historical data. Our results indicate that, in-dependent of the model of earnings expectations we use, models thatinclude dividend changes do not outperform those that do not includedividend changes. In fact, the evidence indicates that investors are betteroff not using dividend changes in their forecasting models.This article is organized as follows. Section II describes the sam-

ple selection procedure, defines the variables, and provides summary

1661Dividend Changes Do Not Signal Changes in Future Profitability

statistics. Section III investigates the relation between dividend changesand future earnings. Section IVexamines the relation between dividendchanges and profitability. Section V investigates the relation betweendividend changes and levels in earnings and profitability. Section VIexamines the out-of-sample forecasting ability of dividend changes. Fi-nally, Section VII concludes the article with a discussion of the impli-cations of our findings.

II. Sample Selection

Using the Center for Research in Securities Prices (CRSP) monthlyevent file, we identify all the dividend announcements made between1963 and 1997 by firms listed on the New York (NYSE) and American(AMEX) stock exchanges.2 To be included in the final sample, a divi-dend announcement must satisfy the following criteria:

1. The firm’s financial data are available on the CRSP/Compustat mergeddatabase at the year of the announcement.

2. The firm is not a financial institution (SIC codes 6000–6999).3

3. The company paid a quarterly taxable cash dividend in U.S. dollars(code No. 1232 in the CRSP file) in the current and previous quarter.

4. Other distribution events, such as stock splits, stock dividends, mergers,and the like, were not declared between the declaration of the previousdividend and 4 days after the declaration of the current dividend.

5. There were no ex-distribution dates between the ex-distribution datesof the previous and current dividends.

Following Benartzi et al. (1997), we match the dividend announce-ments made during fiscal year t to the earnings in fiscal year t.4 We thendefine the annual dividend change as the annualized rate of quarterlydividend changes:

RDDIVt ¼ ð1þ DDIVt;1Þð1þ DDIVt;2Þð1þ DDIVt;3Þð1þ DDIVt;4Þ � 1;

ð1Þ

2. We restrict the sample to the period 1963–97 to compare our results to the ones inNissim and Ziv (2000). However, the results do not change if we extend the sample to 1999.3. We also replicate the analysis excluding utilities and the results are qualitatively the same.4. Using a different approach, Nissim and Ziv (2001) label the dividend payment paid in

the first quarter of year t + 1 as a year t dividend. This approach is inappropriate, becauseit artificially strengthens the relation between dividend changes in year t and the earningschanges in year t + 1. The reason for this is that the dividend changes in the first quarter ofyear t + 1 already contain partial information about the earnings changes in year t + 1 (seeBenartzi et al. 1997). However, even when including the dividend change in the first quarterof year t + 1 in the year t dividend, the positive relation between dividend changes andearnings in year t + 1 is significant in only about 34% of the years. Further, there is nochange in the relation between dividend changes and earnings in year t + 2, nor there is anychange in the relation for other measures of profitability such as ROA, not even for year t + 1.


where DDIVt;q is the dividend change (in fiscal year t) from quarter q � 1to q, scaled by the dividend payment in quarter q�1. The resulting sam-ple contains 2,778 firms and 14,235 dividend increases, 974 dividend de-creases, and 23,334 no-change events.Table 1 provides preliminary statistics on the percent dividend change,

the market value of equity, the market-to-book ratio, the return on equity(ROE), and the return on assets (ROA) for dividend decreasing firms(panel A), dividend increasing firms (panel B), and firms with no changein their dividend payments (panel C). The average (median) decrease individends is 45.5% (48%), compared with an average (median) increasein dividends of nearly 17.9% (12.5%). These results are consistent withprior empirical studies (e.g.,Michaely, Thaler, andWomack 1995) show-ing that dividend cuts are less common than dividend increases and moreextreme in magnitude. Otherwise, the rest of the table shows much whatmight be expected. Firms that increase dividends are larger, and havebeen more profitable recently than firms that either cut their dividends orleave them unchanged.

TABLE 1 Summary Statistics

Mean SD 5% 50% 95% N

A. Dividend Decreases

RDDIV �45.5% 17.0% �76.0% �48.0% �18.2% 974MV (millionsof $) 591.2 1,839.6 6.6 87.2 2,737.6 946

M/B 1.19 1.22 .35 .90 2.64 906ROE 7.1% 15.7% �11.7% 7.0% 23.1% 920ROA 12.2% 7.7% 2.6% 11.2% 24.5% 954

B. Dividend Increases

RDDIV 17.9% 20.0% 3.3% 12.5% 50.0% 14,235MV (millionsof $) 1,595.5 5,518.2 13.5 280.0 6,267.3 14,122

M/B 1.94 1.77 .55 1.48 4.72 13,697ROE 15.3% 8.6% 6.5% 14.4% 26.1% 13,743ROA 18.0% 7.2% 8.8% 16.7% 31.4% 14,093

C. No Changes

RDDIV 0 0 0 0 0 23,334MV (millionsof $) 868.7 3,525.3 8.1 127.2 3,434.5 22.998

M/B 1.76 1.73 .47 1.32 4.31 21,658ROE 12.2% 11.4% .24% 11.9% 25.2% 21,778ROA 16.0% 7.7% 5.8% 14.8% 29.8% 23,029

Note.—This table reports the firm characteristics for the sample firms. RDDIV is the annual per-centage change in the cash dividend payment. MV is the market value of equity. M/B is the marketvalue of equity relative to the book value of equity. ROE is equal to the earnings before extraordinaryitems scaled by the book value of equity. ROA is equal to the operating income before depreciationscaled by total assets. The values of all financial variables are determined at the beginning of the yearof the announcement. To reduce the effect of outliers, RDDIV, M/B, ROE, and ROA have beenWinsorized at the 0.1% and the 99.9% of the empirical distribution.


III. The Relation between Dividend Changes

and Future Earnings Changes

A. Linear Model of Earnings Expectations

To establish a baseline, we first examine the relation between dividendchanges and future earnings changes using a linear model of earningsexpectations. We begin our analysis by veryfing that the Nissim and Ziv(2000) results hold in our sample. Specifically, we estimate the follow-ing regression model that allows for asymmetric reactions to dividendsincreases and decreases and controls for uniform mean reversion andmomentum in earnings:

ðEt � Et�1Þ=B�1 ¼ b0 þ b1PDPC0 � RDDIV0 þ b1NDNC0 � RDDIV0

þ b2ROEt�1 þ b3ðE0 � E�1Þ=B�1 þ "t; ð2Þ

where Et is earnings before extraordinary items in year t (year 0 is theevent year),B�1 is the book value of equity at the end of year –1, RDDIV0 isthe annual percentage change in the cash dividend payment in year 0, DPC(DNC) is a dummy variable that takes the value of 1 for positive (negative)dividend changes and 0 otherwise, and ROEt�1 is equal to earnings beforeextraordinary items in year t � 1 scaled by the book value of equity at theend of year t � 1. Note that this model assumes that the relation betweenfuture earnings changes and past earnings levels and changes is linear.To reduce the problems associated with residual cross-correlation,

we use the Fama-MacBeth (1973) procedure to estimate the coefficientsof the regression model. In the first stage, we estimate cross-sectionalregression coefficients each year using all the observations in that year.In the second stage, we compute time-series means of the cross-sectionalregression coefficients. The standard deviations for these averages are es-timated using theHansen-Hodrick (1980) standard error correctionmethod.Table 2 reports the results from equation (2). Consistent with the evi-

dence in Nissim and Ziv (2000), we find that dividend changes in year 0are positively correlatedwith future earnings changes in years 1 and 2; thatis, t = 1 and t = 2. PanelA shows that the coefficient for positive dividendchanges, b1P, is equal to 0.027 when t = 1 and to 0.017 when t = 2. Bothcoefficients are significantly different from zero at standard confidencelevels. The evidence in panel A also indicates that dividend decreases arenot related to future changes in earnings. Overall, these results indicatethat, when the model of earnings expectations is linear, dividend changesconvey some information about future changes in earnings.5

5. Trying to replicate Nissim and Ziv, we use their empirical specification. However, itshould be noted that this specification suffers from a look-ahead bias when t = 2, because ituses as control variable the ROE in Year 1, which is unknown in Year 0. If we use the ROEin Year 0 instead of the ROE in Year 1 as a control variable, the relation between the earn-ings changes in Year 2 and dividend changes disappears.


TABLE 2 Regressions of Raw Earnings Changes on Dividend Changes

ðET � ET�1Þ=B�1 ¼ B0 þ B1PDPC0 � RDDIV0 � b1NDNC0 � RDDIV0

þ b2ROEt�1 þ b3ðE0 � E�1Þ=B�1 þ "t

A. Time-Series Means of the Cross-Sectional Regression Coefficients

Year b0 b1P b1N b2 b3 Average Adj. R2

t = 1 Mean .026a .027a .08 �.144a .079c 11.73%Wald statistic 32.75 12.49 .32 33.37 3.36

t = 2 Mean .021a .017b .024 �.089a �.018 9.89%Wald statistic 24.40 5.08 2.06 12.33 0.36

B. Annual Cross-Sectional Regression Coefficients of Dividend Changes

t = 1 t = 2

Year b1P tðb1PÞ b1N tðb1N Þ b1P tðb1PÞ b1N tðb1N Þ

1963 .054 1.504 �.173 �2.073 .045 1.141 .059 .7761964 .013 .566 �.041 �.483 .019 .745 �.079 �.8321965 .008 .564 .055 .552 �.027 �1.204 .013 .0871966 �.007 �.224 .026 .505 �.002 �.067 �.057 �.9321967 �.073 �2.303 �.007 �.081 �.009 �.242 .030 .2731968 .012 .551 �.105 �2.222 .008 .301 .046 .7801969 .088 2.881 �.012 �.339 .058 1.447 .045 1.0181970 .057 1.847 .003 .199 .047 1.613 .005 .3411971 �.010 �.784 �.020 �1.479 .025 1.381 .015 .8241972 .014 .753 .004 .146 .072 2.054 .068 1.1861973 �.010 �.596 .061 1.223 �.041 �2.321 .064 1.2471974 .007 .600 .065 1.741 .020 1.657 .089 2.2851975 �.031 �2.607 .023 .883 .010 .750 .050 1.6261976 .010 .935 .095 1.843 .016 1.260 �.027 �.3951977 .005 .530 �.052 �1.539 �.013 �.827 .004 .0881978 .024 1.577 .191 2.718 .012 .631 .011 .1281979 .062 3.535 .028 .487 �.025 �1.259 �.020 �.3121980 .040 1.566 �.022 �.689 .035 .951 .064 1.3391981 .030 1.040 �.013 �.249 .015 .511 �.113 �2.0311982 .020 .697 �.033 �1.305 .059 1.580 .036 1.0951983 .002 .052 �.042 �1.095 .060 1.371 .062 1.1561984 .102 3.278 �.182 �2.739 .078 2.629 .367 5.8131985 .087 2.180 .054 .989 .079 1.605 .111 1.5991986 .081 1.912 �.007 �.134 .071 1.293 .065 1.2021987 .093 1.999 .073 .994 .052 1.085 �.092 �1.1811988 .027 1.358 .095 1.095 �.015 �.606 �.067 �.6411989 �.046 �1.354 .208 2.988 �.017 �.440 .041 .4611990 .035 .980 �.001 �.024 .021 .581 �.123 �2.3521991 .104 2.489 �.004 �.109 .064 1.401 .019 .4931992 .060 1.441 .039 .896 �.008 �.183 �.023 �.4581993 �.038 �1.489 �.007 �.139 �.007 �.236 �.033 �.5561994 .054 1.802 .082 .982 �.037 �1.044 .106 1.0591995 �.002 �.060 �.081 �1.517 .045 .951 �.138 �1.8811996 .067 1.818 �.083 �1.194 �.128 �2.228 .299 2.6111997 .021 .375 .063 .736 .009 .163 �.067 �.800

Note.—This table reports estimates of regressions relating raw earnings changes to dividendchanges. Et is the earnings before extraordinary items in year t (year 0 is the event year). B�1 is thebook value of equity at the end of year �1. RDDIV0 is the annual percentage change in the cashdividend payment in year 0. DPC (DNC) is a dummy variable that takes the value of 1 for dividendincreases (decreases) and 0 otherwise. ROEt�1 is equal to the earnings before extraordinary items inyear t = 1 scaled by the book value of equity at the end of year t = 1. We use the Fama-MacBeth(1973) procedure to estimate the regression coefficients. In the first stage, we estimate cross-sectionalregression coefficients each year using all the observations in that year. In the second stage, we computetime-series means of the cross-sectional regression coefficients. The standard deviations for these av-erages are estimated using the Hansen-Hodrick (1980) standard error correction method. To reduce theeffect of outliers, all the variables have been Winsorized at the 0.1% and the 99.9% of the empirical dis-tribution. The average adj. R2 is the average (adjusted) R2 of the cross-sectional regressions. In panel B,positive and significant coefficients (at least at the 10% level) are highlighted in bold. a, b, and c denotesignificantly different from zero at the 1%, 5%, and 10% level, respectively.

To gauge the reliability of dividend changes as predictors of futureearnings changes and to see whether the relation between dividendchanges and earnings changes varies through time in a systematic pat-tern, we also report the annual cross-sectional regression coefficients ofdividend changes. The results from this analysis are reported in panel Bof table 2. (In this panel, the coefficients that are positive and significantat least at the 10% level are highlighted in bold.) Note that, even whenthe model of earnings expectations is linear, the coefficient for positivedividend changes is significant only in about 29% of the years when t =1 and in only about 9% of the years (i.e., in only 3 out of the 35 years inthe sample) when t = 2. That is, in most of the years in our sample,current changes in dividends are not a reliable signal of future earningchanges (either 1 or 2 years ahead) in the same direction.

B. Nonlinear Model of Earnings Expectations

Nissim and Ziv (2001) argue that some previous studies examining therelation between earnings and dividends omit relevant variables in theirregression analyses correlated with the dividend changes. To this end,Nissim and Ziv include the return on equity and past changes in earningsto control for the mean reversion and autocorrelation (e.g., momentum)in earnings. However, such regressions assume that the rate of meanreversion and the level of autocorrelation are uniform across all ob-servations (see equation [2]).This is an important issue, because empirical evidence indicates that

the mean reversion process of earnings and the level of autocorrelationare highly nonlinear. For example, large changes in earnings revert fasterthan small changes, and negative changes revert faster than positivechanges (see Brooks and Buckmater 1976; Elgers and Lo 1994; and Famaand French 2000.) Since assuming linearity when the true functional formis nonlinear has the same consequences as leaving out relevant indepen-dent variables, the coefficients of the regressions in Nissim and Ziv (and inequation [2]) are likely to be biased.6

There are at least two alternative methods for controlling for thenonlinearities in the earnings process. The first is the matched-sampleapproach.7 For each firm that changes its dividend, select an otherwisesimilar firm that experienced a similar level of earnings and similar his-torical pattern in earnings. Then, examine how the difference in operatingperformance between the sample firm and the matching firm (abnormalperformance) is related to the dividend change. This method was used by

6. Kennedy (1998) explains this econometric issue as follows: ‘‘The properties of theOLS estimator applied to a situation in which the true functional form is nonlinear can beanalyzed in terms of omitted relevant variables. A nonlinear function can be restated, via aTaylor series expansion, as a polynomial. Estimating a linear function is in effect omittingthe higher-order terms of this polynomial.’’7. See Barber and Lyon (1996) for a general discussion of this issue.


Benartzi et al. (1997) and Grullon et al. (2002), and neither paper foundevidence that dividend changes predict future changes in earnings. How-ever, these results may be specific to the matched-sample methodologyand may not hold in a more general specification of the relation betweencurrent changes in dividends and future changes in earnings.The second method of controlling for the nonlinearities in the earn-

ings process is regression analysis. One advantage of regression anal-ysis over the matching sample approach is that it enables the researcherto explicitly model the behavior of future earnings. Furthermore, it al-lows the researcher to control for more factors and take advantage of theinformation contained in the cross-section of earnings. To this end, weuse regression analysis that accounts for the no-linear patterns in earn-ings to examine this issue.Specifically, we use the modified partial adjustment model suggested

by Fama and French (2000) as a control for the nonlinearities in therelation between future earnings changes and lagged earnings levels andchanges. The model is the following:

ðEt � Et�1Þ=B�1 ¼ b0 þ b1PDPC0 � RDDIV0 þ b1NDPN0 � RDDIV0

þ ðg1þ g2NDFED0þ g3NDFED0� DFE0

þ g4PDFED0�DFE0Þ�DFE0 þ ðl1 þ l2NCED0

þ l3NCED0 � CE0 þ l4PCED0� CE0Þ � CE0 þ "t;

ð3Þ

where DFE0 is equal to ROE0 � E½ROE0�, where E½ROE0� is the fit-ted value from the cross-sectional regression of ROE0 on the logarithm oftotal assets in year�1, the logarithm of the market-to-book ratio of equityin year �1, and ROE�1. CE0 is equal to ðE0 � E�1Þ=B�1, NDFED0

(PDFED0) is a dummyvariable that takes the value of 1 if DFE0 is negative(positive) and 0 otherwise, and NCED0 (PCED0) is a dummy variable thattakes the value of 1 if CE0 is negative (positive) and 0 otherwise. Asdiscussed in Fama and French (2000), the dummy variables and squaredterms are designed to pick up the documented nonlinearities in the meanreversion and autocorrelation of earnings. Specifically, these variablesare meant to capture the fact that large changes in earnings revert fasterthan small changes and negative changes revert faster than positivechanges. This particular behavior of earnings has been documented inBrooks and Buckmaster (1976) and Elgers and Lo (1994).In table 3, we show the re-estimated coefficients of the regression

models using the Fama and French (2000) methods. Unlike the resultsreported in table 2 (where a linear model is used), we find no evidencethat the magnitude of dividend changes contains information aboutfuture earnings. Panel A of table 3 shows that, for the first year follow-ing the dividend change, the coefficient for positive dividend changes,


TABLE 3 Regressions of Raw Earnings Changes on Dividend Changes Using the Fama and French Approach to Predict Expected Earnings

ðEt � Et�1Þ=B�1 ¼ b0 þ b1PDPC0 � RDDIV0 � b1NDNC0 � RDDIV0 þ ðg1 þ g2NDFED0 þ g3NDFED0 � DFE0 þ g4PDFED0

� DFE0Þ � DFE0 þ ðl1 þ l2NCED0 þ l3NCED0 � CE0 þ l4PCED0 � CE0Þ � CE0 þ "t

A. Time-Series Means of the Cross-Sectional Regression Coefficients

Year b0 b1P b1N g1 g2 g3 g4 l1 l2 l3 l4 Average Adj. R2

t = 1 Mean �.003 .008 .009 �.173b �.282b �.849 �1.160 .452a �.217c 1.626 �.507 21.01%Wald statistic 1.50 1.45 .43 5.22 4.36 1.85 2.56 24.64 3.27 1.25 2.14

t = 2 Mean .011a .004 �.001 �.210 �.001 �.336 �1.078 .163c �.107 .120 .050 11.40%Wald statistic 20.46 .29 .00 2.31 .00 .21 .39 3.75 .55 .02 .00


t = 1 t = 2


1963 .037 1.470 �.051 �.882 .069 1.888 .081 .9581964 �.014 �.761 �.025 �.354 �.004 �.202 �.093 �1.1021965 .003 .194 .057 .580 �.030 –1.328 .007 .0461966 .006 .229 .088 1.996 �.017 �.489 �.056 �.9491967 �.071 �2.391 �.014 �.170 �.017 �.454 .027 .2561968 .005 .226 �.129 �2.761 .012 .445 .020 .3351969 .069 2.356 .008 .236 .022 .551 �.028 �.6371970 .027 .897 �.004 �.247 .042 1.464 .010 .6651971 �.015 –1.068 �.001 �.057 �.009 �.460 .014 .7471972 .014 .817 �.018 �.685 .100 2.800 .067 1.1791973 �.006 �.331 .028 .563 �.009 �.468 .022 .4091974 �.015 �1.321 .009 .260 .029 2.443 .070 1.8211975 �.043 �3.551 .036 1.352 .017 1.194 .037 1.1381976 .008 .823 .056 1.215 .019 1.538 �.003 �.0391977 �.009 �.921 �.101 �2.973 �.007 �.476 .006 .1141978 .016 1.079 .159 2.356 .012 .619 �.056 �.6321979 .032 1.998 �.064 �1.174 �.025 �1.197 �.067 �.976

1668

JournalofBusin

ess

TABLE 3 (Continued )


t = 1 t = 2


1980 .031 1.278 �.027 �.876 .043 1.187 .054 1.1531981 .025 .863 �.004 �.072 �.018 �.581 �.165 �2.8741982 �.072 �2.439 �.029 �1.153 .037 .950 .042 1.2121983 �.038 –1.274 �.088 �2.324 .056 1.302 .052 .9551984 .086 2.989 �.130 �2.062 .046 1.458 .404 5.9281985 .038 1.015 .067 1.320 .024 .492 .050 .7411986 .038 .872 �.095 �1.825 .017 .342 .012 .2011987 .074 1.727 .078 1.141 .017 .328 �.118 �1.4051988 .007 .372 .152 1.826 �.013 �.551 �.170 �1.6801989 �.066 �2.088 .251 3.796 �.035 �.892 �.083 �.9611990 .019 .586 �.018 �.376 �.002 �.065 �.122 �2.2151991 .040 1.070 �.014 �.444 .006 .124 �.003 �.0711992 .032 .838 .031 .760 �.036 �.794 �.028 �0.5551993 �.054 �2.156 .048 .968 �.009 �.318 �.038 �.6381994 .017 .565 .086 1.045 �.069 �1.969 .091 .9361995 �.013 �.359 �.069 –1.355 0.046 .968 �.186 �2.5531996 .068 1.917 �.017 �.247 �.167 �2.854 .230 1.9281997 .009 .176 .067 .786 �.005 �.080 �.102 �1.082

Note.—This table reports estimates of regressions relating raw earnings changes to dividend changes. Et is the earnings before extraordinary items in year t (year 0 is the eventyear). B�1 is the book value of equity at the end of year �1. RDDIV0 is the annual percentage change in the cash dividend payment in year 0. DPC (DNC) is a dummy variable thattakes the value of 1 for dividend increases (decreases) and 0 otherwise. ROEt is equal to the earnings before extraordinary items in year t scaled by the book value of equity at theend of year t. DFE0 is equal to ROE0 � E[ROE0], where E[ROE0] is the fitted value from the cross-sectional regression of ROE0 on the logarithm of total assets in year �1, thelogarithm of the market-to-book ratio of equity in year �1, and ROE�1. CE0 is equal to ðE0 � E�1Þ=B�1. NDFED0 is a dummy variable that takes the value of 1 if DFE0 is negativeand 0 otherwise. PDFED0 is a dummy variable that takes the value of 1 if DFE0 is positive and 0 otherwise. NCED0 is a dummy variable that takes the value of 1 if CE0 is negativeand 0 otherwise. PCED0 is a dummy variable that takes the value of 1 if CE0 is positive and 0 otherwise. We use the Fama-MacBeth (1973) procedure to estimate the regressioncoefficients. In the first stage, we estimate cross-sectional regression coefficients each year using all the observations in that year. In the second stage, we compute time-seriesmeans of the cross-sectional regression coefficients. The standard deviations for these averages are estimated using the Hansen-Hodrick (1980) standard error correction method.To reduce the effect of outliers, all the variables, except the log of total assets, have been Winsorized at the 0.1% and the 99.9% of the empirical distribution. The average adj. R2 isthe average (adjusted) R2 of the cross-sectional regressions. In panel B, positive and significant coefficients (at least at the 10% level) are highlighted in bold. a, b, and c denotesignificantly different from zero at the 1%, 5%, and 10% level, respectively.

1669

Divid

endChanges

DoNotSignalChanges

inFuture

Profitability

b1P , is just 0.008 (it was 0.027 using the linear model). This coefficientis neither economically nor statistically significantly different from zero.Further, similar to the results in table 2, we find no evidence that dividenddecreases are related to future changes in earnings.When comparing the results in table 3 to those in table 2, note that the

nonlinear model explains a larger fraction of the cross-sectional varia-tion in earnings changes than the linear model. Specifically, we find thatthe average adjustedR2 increases from 11.73% to 21.01%when t = 1 andfrom 9.89% to 11.40% when t = 2. Further, the evidence in table 3indicates that the behavior of profitability is highly nonlinear. Consistentwith the findings in Fama and French (2000), this evidence indicatesthat the linear model in table 2 misses important information about thebehavior of earnings that seems to be correlated with dividend changes.Panel B of table 3 shows that, when earnings changes in Year 1 are the

dependent variable, the coefficient for the positive dividend changes issignificant only in about 14% of the years when t = 1 and only in about9% of the years when t = 2. In fact, the number of negative andsignificant coefficients for positive dividend changes when t = 1 is thesame as the number of positive and significant coefficients. Thus, al-lowing for the empirically documented nonlinearities in the mean re-version process leads to the conclusion that changes in dividends are notuseful in predicting future earnings changes. Similar results emerge fordividend decreases.One explanation for these results is that dividend changes act as

a surrogate for the nonlinearity in earnings under a uniform meanreverting model. This explanation is consistent with the empiricalregularity that firms tend to change their dividend policy only whenearnings changes have been substantial. Firms increase their dividendsafter they have done well for a long period of time and they cut divi-dends after a long period of poor performance, which they expect tocontinue (see Brav et. al. 2003). Therefore, dividend changes are likelyto be correlated with the nonlinear component of earnings changesthat is missing from equation (2). When we model the earning processmore precisely and incorporate the nonlinear portion, dividends indeedcontain no information about future earnings, supporting the idea thatdividends are proxies for the nonlinear patterns in the evolution ofearnings.

IV. Do Changes in Dividends Forecast Profitability?

Dividend-signaling theory does not indicate precisely which firm per-formance metric (e.g., future income or future profitability) should beused. Therefore, in addition to earnings, an alternative firm performancevariable that is widely used is profitability, as measured by the returnon assets. Indeed, Fama and French (2000) use the same methods to


forecast changes in both earnings and profitability (return on assets).ROA is defined as the operating income before depreciation (EBITDA)scaled by the book value of total assets. This measure of operatingperformance is preferable to ROE (or other scaled-earnings variables)in several dimensions. First, ROE is sensitive to changes in capitalstructure while ROA is not (since ROA is measured using EBITDA andnot net income). Second, the ROA is not affected by factors such asspecial items (i.e., unusual and nonrecurring items reported beforetaxes), accounting for minority interest, and income taxes that usuallyobscure the ROE. Indeed, using simulation analysis Barber and Lyon(1996) show that ROA is the best available measure to detect abnormaloperating performance under most circumstances.With this in mind, we replicate all the previous analyses replacing the

change in earnings with the change in ROA as the dependent variable.These results are shown in table 4. In Panel A of table 4, we rerun thelinear regression model (parallel to table 2). For consistency, we replacethe ROE on the right-hand side with ROA. Panel A shows that therelation between positive and negative dividend changes and ROAchanges is indistinguishable from zero, at both the 1-and 2-year hori-zons. Panel B of table 4 shows that, when using the nonlinear earningsmodel, the relation between positive dividend changes and changes infuture profitability is negative and significant when t = 1. Moreover,panel B shows that the relation between negative dividend changes andchanges in future profitability has the wrong sign and is highly signif-icant when t = 1.Overall, the results in this section indicate that firm profitability is not

positively associated with past changes in dividends. In fact, contrary tothe predictions of the signaling hypothesis, the evidence seems to in-dicate that dividend changes are negatively correlated with future changesin ROA.8

V. The Relation between Dividend Changes and Future Earnings Levels

In addition to the regressions of changes in earnings on dividendchanges, Nissim and Ziv (2001) also find that dividend changes arepositively correlated with future earnings levels. Although the specifi-cation using earnings levels provides an alternative way to examine the

8. For robustness, we also replicate our analysis using the return on cash-adjusted assets,the return on sales, and the cash-flow return on assets (see Barber and Lyon 1996 for a detaileddescription of these measures). We find that, after controlling for the nonlinear patterns in theearnings process, dividends are not positively correlated with future changes in performance.In fact, we find that dividend changes are negatively correlated with the return on cash-adjustedassets and the return on sales. We also use analysts’ earnings forecast as a proxy for expectedearnings. The results indicate that dividends are not positively correlated with future earnings,consistent with our previous results.


TABLE 4 Regressions of Changes in ROA on Dividend Changes

A. Time-Series Means of the Cross-Sectional Regression Coefficients from the Linear ModelROAt � ROAt�1 ¼ b0 þ b1PDPC0 � RDDIV0 þ b1NDNC0 � RDDIV0 þ b2ROAt�1 þ b3ðROA0 � ROA�1Þ þ "t

Year b0 b1P b1N b2 b3 Average Adj. R2

t = 1 Mean .019a �.001 �.008 �.146a �.017 10.76%Wald statistic 156.95 .02 2.19 547.25 .85

t = 2 Mean .019a �.001 .004 �.144a �.085a 11.63%Wald statistic 135.46 .06 .33 327.31 26.03

B. Time-Series Means of the Cross-Sectional Regression Coefficients from the Nonlinear ModelROAt � ROAt�1 ¼ b0 þ b1PDPC0 � RDDIV0 þ b1NDPN0 � RDDIV0 þ ðg1 þ g2NDFED0 þ g3NDFED0 � DFE0 þ g4PDFED0 � DFE0Þ � DFE0

þ ðl1 þ l2NCED0 þ l3NCED0 � CE0 þ l4PCED0 � CE0Þ � CE0 þ "t

Year b0 b1P b1N g1 g2 g3 g4 l1 l2 l3 l4 Average Adj. R2

t = 1 Mean �.003a �.007b �.014a �.582a �.101 3.147a �.625 .535a .033 �2.076b �.160 14.15%Wald statistic 8.64 5.63 6.87 40.58 .55 9.60 .36 30.20 .06 5.59 .02

t = 2 Mean .000 �.004 �.004 �.511a .064 .434 �1.559c .294a .044 .007 .992 10.31%Wald statistic .01 2.41 .35 72.32 .61 .39 3.07 18.74 .19 .00 1.21

Note.—This table reports estimates of regressions relating changes in ROA to dividend changes. ROAt is equal to the operating income before depreciation in year t scaled bytotal assets at the end of year t. RDDIV0 is the annual percentage change in the cash dividend payment in year 0. DPC (DNC) is a dummy variable that takes the value of 1 fordividend increases (decreases) and 0 otherwise. DFE0 is equal to ROA0 � E[ROA0], where E[ROA0] is the fitted value from the cross-sectional regression of ROA0 on thelogarithm of total assets in year �1, the logarithm of the market-to-book ratio of equity in year �1, and ROA�1. CE0 is equal to ROA0 � ROA–1. NDFED0 is a dummy variablethat takes the value of 1 if DFE0 is negative and 0 otherwise. PDFED0 is a dummy variable that takes the value of 1 if DFE0 is positive and 0 otherwise. NCED0 is a dummyvariable that takes the value of 1 if CE0 is negative and 0 otherwise. PCED0 is a dummy variable that takes the value of 1 if CE0 is positive and 0 otherwise. We use the Fama-MacBeth (1973) procedure to estimate the regression coefficients. In the first stage, we estimate cross-sectional regression coefficients each year using all the observations in thatyear. In the second stage, we compute time-series means of the cross-sectional regression coefficients. The standard deviations for these averages are estimated using the Hansen-Hodrick (1980) standard error correction method. To reduce the effect of outliers, all the variables, except the log of total assets, have been Winsorized at the 0.1% and the 99.9%of the empirical distribution. The average adj. R2 is the average (adjusted) R2 of the cross-sectional regressions. a, b, and c denote significantly different from zero at the 1%, 5%,and 10% level, respectively.

1672

JournalofBusin

ess

relation between earnings and dividend changes, it has several limi-tations.9 Nevertheless, we replicate the analyses in Sections III and IVusing earnings levels instead of earnings changes to ensure that ourresults are robust to alternative specifications. Specifically, we estimatethe following linear and nonlinear regression models:

ROEt ¼ b0 þ b1PDPC0 � RDDIV0 þ b1NDNC0 � RDDIV0 þ b2ROEt�1

þ b3ðROE0 � ROE�1Þ þ b4MB�1 þ b5SIZE�1 þ "t ð4Þ

and

ROEt ¼ b0 þ b1PDPC0 � RDDIV0 þ b1NDNC0 � RDDIV0

þ ðg1 þ g2NDFED0 þ g3NDFED0�ROE0 þ g4PDFED0

� ROE0Þ�ROE0þðl1 þ l2NCED0 þ l3NCED0 � CE0

þ l4PCED0 � CE0Þ � CE0 þ j1MB�1 þ j2SIZE�1 þ "t; ð5Þ

where the variables have been previously defined. Note that, in theseregressions, we relate ROE (the scaled level of earnings at time t ðEt=BtÞ)to changes in dividends and other control variables. Further, note that, inequation (5), we use the ROE in Year 0 instead of DFE, as we do inequation (3), because the dependent variable in equation (5) is deter-mined primarily by the ROE in Year 0 (e.g., ROE1 ¼ ROE0 þ DROE1).The results from this analysis are reported in table 5. Consistent with

the evidence in Nissim and Ziv (2001), panel A shows that the level ofROE in Year 1 and 2 are positively correlated with the positive changesin dividends. Note that the coefficient for positive dividend changes isequal to 0.021 when t = 1 and to 0.025 when t = 2. Both coefficientsare significantly different from zero at standard confidence levels. Thisfurther shows that the results in Nissim and Ziv hold in our sample.However, consistent with the evidence in Section III, panel B showsthat, after controlling for the nonlinear patterns in earnings, dividendschanges do not contain information about the future level of ROE. Onceagain, the evidence in this section does not support the predictions of thesignaling hypothesis.10

9. Conceptually, if a change in dividend is supposed to contain ‘‘information’’ aboutearnings, then, by definition, that information should be about changes because the currentlevel of earnings is already known. Thus, it is not clear what we learn by using earnings levelsinstead of earnings changes. More important, empirical evidence suggests that changes inprofitability or earnings tend to have better statistical properties than levels. For example,Barber and Lyon (1996) found that test statistics using the change in a firm’s operating per-formance yield more powerful test statistics than those based on the level of a firm’s operatingperformance.10. For robustness, we also replicate all of the previous analyses replacing ROE with

ROA as the dependent variable. These results show that dividend increases are uncorrelatedwith the future level of ROA, independent of the earnings model we use. Interestingly, wefind that dividend decreases are negatively correlated with the level of ROA in Year 1. Thatis, firms that reduce their dividends tend to do better in the future than other firms.


TABLE 5 Regressions of ROE Levels on Dividend Changes

A. Time-Series Means of the Cross-Sectional Regression Coefficients from the Linear Model

ROEt ¼ b0 þ b1PDPC0 � RDDIV0 þ b1NDNC0 � RDDIV0 þ b2ROEt�1 þ b3ðROE0 � ROE�1Þ þ b4MB�1 þ b5SIZE�1 þ "t

Year b0 b1P b1N b2 b3 b4 b5 Average Adj. R2

t = 1 Mean .000 .021a .009 .706a �.023 .019a .003a 59.05%Wald statistic .00 9.33 .22 492.53 .37 25.79 18.02

t = 2 Mean .007 .025c .048 .518a �.080b .024a .005a 44.87%Wald statistic .32 3.36 2.40 200.28 6.31 30.83 21.39

B. Time-Series Means of the Cross-Sectional Regression Coefficients from the Nonlinear Model

ROEt ¼ b0 þ b1PDPC0 � RDDIV0 þ b1NDPN0 � RDDIV0 þ ðg1 þ g2NDFED0 þ g3NDFED0 � ROE0 þ g4PDFED0 � ROE0Þ� ROE0 þ ðl1 þ l2NCED0 þ l3NCED0 � CE0 þ l4PCED0 � CE0Þ � CE0 þ j1MB�1 þ j2SIZE�1 þ "t

Year b0 b1P b1N g1 g2 g3 g4 l1 l2 l3 l4 j1 j2 Average Adj. R2

t = 1 Mean �.014b .007 .001 .957a .335 1.794 �.419a .028 .142 �.465 �.857 .011a .002a 62.64%Wald statistic 3.96 1.58 .00 265.30 .11 .95 8.78 .12 1.11 .59 2.39 21.96 11.12

t = 2 Mean �.006 .011 .030 .775a �1.454b �21.75 �.428a �.072 .319b 1.05 �.578 .015a .004a 47.33%Wald statistic .35 .56 1.06 130.88 4.13 1.00 6.52 .65 5.28 .85 1.15 15.46 17.66

Note.—This table reports estimates of regressions relating ROE levels to dividend changes. ROEt is equal to the earnings before extraordinary items in year t scaled by thebook value of equity at the end of year t. RDDIV0 is the annual percentage change in the cash dividend payment in year 0. DPC (DNC) is a dummy variable that takes the value of1 for dividend increases (decreases) and 0 otherwise. CE0 is equal to ðROE0 � ROE�1Þ. NDFED0 is a dummy variable that takes the value of 1 if ROE0 is negative and 0otherwise. PDFED0 is a dummy variable that takes the value of 1 if ROE0 is positive and 0 otherwise. NCED0 is a dummy variable that takes the value of 1 if CE0 is negative and 0otherwise. PCED0 is a dummy variable that takes the value of 1 if CE0 is positive and 0 otherwise. MB�1 is the logarithm of the market-to-book ratio of equity in year �1. SIZE�1

is the logarithm of total assets in year �1. We use the Fama-MacBeth (1973) procedure to estimate the regression coefficients. In the first stage, we estimate cross-sectionalregression coefficients each year using all the observations in that year. In the second stage, we compute time-series means of the cross-sectional regression coefficients. Thestandard deviations for these averages are estimated using the Hansen-Hodrick (1980) standard error correction method. To reduce the effect of outliers, all the variables, except thelog of total assets, have been Winsorized at the 0.1% and the 99.9% of the empirical distribution. The average adj. R2 is the average (adjusted) R2 of the cross-sectional regressions.a, b, and c denote significantly different from zero at the 1%, 5%, and 10% level, respectively.

1674

JournalofBusin

ess

VI. The Economic Significance of the Predictive Power of Dividends

The results discussed so far show that, after controlling for the nonlinearpatterns in the earnings process, dividend changes are not positivelycorrelated with future changes in earnings. However, these results fol-low the standard procedure of estimating the model using all the time-series data. Actual market participants, of course, would not be able tomake use of future years in determining the relationship between, say,dividend changes and earnings changes. To circumvent this problem,we use out-of-sample methods to evaluate the forecasting ability ofdividend changes. These out-of-sample methods are useful in deter-mining the forecasting ability of a variable and are widely used in thefinance and economics literature (see, e.g., Meese and Rogoff 1983;Keim and Stambaugh 1986; Akgiray 1989; Pesaran and Timmermann1995; and Graham 1996, among others.)The main advantage of using these methods is that they help us assess

the economic significance of dividend changes, explicitly accountingfor the fact that investors can use only historical data to estimate theparameters of the earnings model. This alternative analysis is similar tothe matched-sample approach in the sense that it uses only the infor-mation available at the time of the forecast. However, it also has theadvantage of allowing the researcher to control for many factors.To assess the economic value of dividends to predict future earnings, we

compare the out-of-sample forecasting ability of the linear (equation [2])and nonlinear (equation [3]) models to the forecasting ability of similarmodels that exclude dividend changes as an explanatory variable. If div-idends provide useful information about future earnings, then the modelthat includes dividends should systematically outperform the model thatexcludes dividends.We use the conditional predictive ability approach in Giacomini and

White (2003) to measure the forecasting ability of each model. To im-plement this approach, we calculate the out-of-sample forecast error ofeach model using the following equations:

f1 ¼ ½ðE1 � E0Þ=B�1 � gðb0

�1X0Þ� ð6Þ

and

f2 ¼ ½ðE2 � E1Þ=B�1 � gðb0

�2X0Þ�; ð7Þ

where f1 and f2 are the out-of-sample forecast errors for the 1-year aheadand 2-year ahead forecasts, respectively, Et is the actual earnings beforeextraordinary items in year t (year 0 is the event year), B�1 is the bookvalue of equity at the end of year�1, g(�) is the earningsmodel, andX0 isthe set of independent variables in year 0.


To measure the forecasting ability of each model, we calculate theout-of-sample forecast error of each model using a rolling scheme and arecursive scheme. Under the rolling scheme, the parameter of variable iin year t is equal to the value of the cross-sectional coefficient of variablei in year t � 1 when t = 1 and to the value of the cross-sectionalcoefficient of variable i in year t � 2 when t = 2. The advantage of thisscheme is that it uses the most recent estimates to forecast futureearnings. Under the recursive scheme, the parameter of variable i in yeart is equal to the average of all the annual cross-sectional coefficients ofvariable i up to year t� 1 when t = 1 and to the average of all the annualcross-sectional coefficients of variable i up to year t � 2 when t = 2.11

Note that these coefficients are estimated using only the data available atthe time of the forecast, to take into account the parameter uncertaintythat investors actually face.Using the forecast errors from equations (6) and (7), we calculate the

annual differences in squared errors:

dSE ¼ f 2DIV � f 2NODIV: ð8Þ

and the annual differences in absolute deviations:

dAD ¼j fDIVj � j fNODIVj ð9Þ

between the model that include dividends (DIV) and the model thatexcludes them (NODIV).We then calculate time-series averages of the annual differences in

squared errors (MSE) and absolute deviations (MAD). To assess the statis-tical significance of the differences inMSE andMAD, we use a bootstrap-ping methodology. This procedure is carried out following these steps:

1. We randomly select observations from the sample of annual differ-ences in squared errors (dSE) and absolute deviation (dAD), withreplacement.

2. We calculate the time-series averages of the annual differences insquared errors and absolute deviations from the bootstrap sample.

3. We repeat steps 1 and 2 10,000 times.4. Using the simulated distribution, we calculate the p-value of the test

statistics.

The results from this analysis are reported in tables 6 and 7. Theresults indicate that models that include dividend changes do not out-perform models that exclude dividends, and in most cases models withdividends actually underperform those models without dividends. Table 6

11. We also set the parameter b�1ðb�2Þ of variable i in year t equal to the average of theparameters over the period t � 5 to t � 1 ðt � 6 to t � 2), and the results are similar.


show the annual differences in mean squared error and mean absolutedeviation between the linear model (equation [2]) and a similar modelthat excludes dividend changes as an explanatory variable. Under therolling scheme, the model that excludes dividends always dominatesthe model that includes dividends. Under the recursive method, the twomodels seem to have the same out-sample performance, except in thecase when we calculate the MSE for the change in earnings in Year 2,where the model that excludes dividends dominates the model thatincludes dividends. Table 7 examines the differences in the out-sampleperformance of the non-linear model (equation [3]) and a similar modelthat excludes dividend changes as an explanatory variable. This tableshows that the model that excludes dividend changes almost always

TABLE 6 The Out-of-Sample Ability of Dividend Changes toPredict Future Earnings Changes: Linear Model

Year Mean p-Value

A. Rolling Scheme

t ¼ 1MSEDIV �MSENODIV .000045a .0031MADDIV �MADNODIV .000271a .0003


B. Recursive Scheme

t ¼ 1MSEDIV �MSENODIV .000002 .3650MADDIV �MADNODIV .000003 .4618

t ¼ 2MSEDIV �MSENODIV .000012c .0611MADDIV �MADNODIV .000040 .2112

Note.—This table compares the out-of-sample forecasting ability of the fol-lowing model:

ðEt � Et�1Þ=B�1 ¼ b0 þ b1PDPC0 � RDDIV0 � b1NDNC0 � RDDIV0

þ b2ROEt�1 þ b3ðE0 � E�1Þ=B�1 þ "t

to the forecasting ability of a similar model that excludes dividend changes as anexplanatory variable. To measure the forecasting ability of each model, we calculatethe out-of-sample forecast error of eachmodel using a rolling scheme and a recursivescheme. Under the rolling scheme, the parameter of variable i in year t is equal to thevalue of the cross-sectional coefficient of variable i in year t� 1 when t = 1 and tothe value of the cross-sectional coefficient of variable i in year t � 2 when t = 2.Under the recursive scheme, the parameter of variable i in year t is equal to theaverage of all the annual cross-sectional coefficients of variable i up to year t � 1when t = 1 and to the average of all the annual cross-sectional coefficients of var-iable i up to year t � 2 when t = 2. Using these forecast errors, we calculate theannual differences in squared errors and absolute deviations between the model thatincludes dividend changes and the model that excludes them. MSEDIV (MADDIV) isthe average annual mean squared error (mean absolute deviation) of the model thatincludes dividend changes. MSENODIV (MADNODIV) is the average annual meansquared error (mean absolute deviation) of the model that excludes dividend changes.To assess the statistical significance of the MSE and the MAD, we use a bootstrap-ping methodology. To reduce the effect of outliers, all the variables, except the log oftotal assets, have been Winsorized at the 0.1% and the 99.9% of the empirical dis-tribution. a, b, and c denote significantly different from zero at the 1%, 5%, and 10%level, respectively.


dominates the model that includes dividend changes, except in the casewhen we calculate the MSE for the change in earnings in Year 1.For robustness, we also replicate all of the previous analyses re-

placing the change in earnings with the change in ROA as the depen-dent variable. The results from this analysis are reported in tables 8and 9. Once again, we find that, on average, the model that includesdividend changes underperforms the model that excludes dividendchanges. Interestingly, table 9 shows that, under the recursive scheme,the model that includes dividend changes outperforms the model that

TABLE 7 The Out-of-Sample Ability of Dividend Changesto Predict Future Earnings Changes:Nonlinear Model

Year Mean p-Value

A. Rolling Scheme

t ¼ 1MSEDIV �MSENODIV .000084 .1044MADDIV �MADNODIV .000385a .0000


B. Recursive Scheme


t ¼ 2MSEDIV �MSENODIV .000569b .0470MADDIV �MADNODIV .000226a .0019

Note.—This table compares the out-of-sample forecasting ability of thefollowing model:

ðEt � Et�1Þ=B�1 ¼ b0 þ b1PDPC0 � RDDIV0 þ b1NDPN0 � RDDIV0

þ ðg1 þ g2NDFED0 þ g3NDFED0 � DFE0

þ g4PDFED0 � DFE0Þ � DFE0 þ ðl1 þ l2NCED0

þ l3NCED0 � CE0 þ l4PCED0 � CE0Þ � CE0 þ "t

to the forecasting ability of a similar model that excludes dividend changes as anexplanatory variable. To measure the forecasting ability of each model, we calcu-late the out-of-sample forecast error of each model using a rolling scheme and arecursive scheme. Under the rolling scheme, the parameter of variable i in year t isequal to the value of the cross-sectional coefficient of variable i in year t� 1 when� = 1 and to the value of the cross-sectional coefficient of variable i in year t� 2when � = 2. Under the recursive scheme, the parameter of variable i in year t isequal to the average of all the annual cross-sectional coefficients of variable i upto year t � 1 when � = 1 and to the average of all the annual cross-sectionalcoefficients of variable i up to year t�2 when � = 2. Using these forecast errors,we calculate the annual differences in squared errors and absolute deviationsbetween the model that includes dividend changes and the model that excludesthem. MSEDIV (MADDIV) is the average annual mean squared error (mean ab-solute deviation) of the model that includes dividend changes. MSEMADNODIV

(MADMADNODIV) is the average annual mean squared error (mean absolute de-viation) of the model that excludes dividend changes. To assess the statisticalsignificance of the MSE and the MAD, we use a bootstrapping methodology. Toreduce the effect of outliers, all the variables, except the log of total assets, havebeenWinsorized at the 0.1% and the 99.9% of the empirical distribution. a, b, and cdenote significantly different from zero at the 1%, 5%, and 10% level, respectively.


excludes dividend changes when t = 1. This is the only situation inour analysis where dividends appear to have out-of-sample predictivepower. However, it is important to note that the coefficients of dividendchanges in the nonlinear model when we use ROA are negative (seepanel B of table 4). Thus, the predictive power of dividends comes fromthe fact that firms that increase (reduce) their dividends experience adecline (an increase) in ROA. This result is at odds with the predictionsof the signaling hypothesis.12

TABLE 8 The Out-of-Sample Ability of Dividend Changesto Predict Future ROA Changes: Linear Model

Year Mean p-Value

A. Rolling Scheme



B. Recursive Scheme

t ¼ 1MSEDIV �MSENODIV .000001 .1017MADDIV �MADNODIV �.000001 .5790

t ¼ 2MSEDIV �MSENODIV .000004a .0006MADDIV �MADNODIV .000040b .0149


ROAt � ROAt�1 ¼ b0 þ b1PDPC0 � RDDIV0 þ b1NDNC0 � RDDIV0

þ b2ROAt�1 þ b3ðROA0 � ROA�1Þ þ "t

to the forecasting ability of a similar model that excludes dividend changes as anexplanatory variable. To measure the forecasting ability of each model, wecalculate the out-of-sample forecast error of each model using a rolling schemeand a recursive scheme. Under the rolling scheme, the parameter of variable i inyear t is equal to the value of the cross-sectional coefficient of variable i in yeart� 1 when t = 1 and to the value of the cross-sectional coefficient of variable iin year t� 2 when t = 2. Under the recursive scheme, the parameter of variable iin year t is equal to the average of all the annual cross-sectional coefficients ofvariable i up to year t� 1 when t = 1 and to the average of all the annual cross-sectional coefficients of variable i up to year t� 2 when t = 2. Using theseforecast errors, we calculate the annual differences in squared errors and abso-lute deviations between the model that includes dividend changes and the modelthat excludes them. MSEDIV (MADDIV) is the average annual mean squarederror (mean absolute deviation) of the model that includes dividend changes.MSENODIV (MADNODIV) is the average annual mean squared error (mean abso-lute deviation) of the model that excludes dividend changes. To assess the statis-tical significance of theMSE and theMAD,we use a bootstrappingmethodology.To reduce the effect of outliers, all the variables, except the log of total assets,have been Winsorized at the 0.1% and the 99.9% of the empirical distribution. a,b, and c denote significantly different from zero at the 1%, 5%, and 10% level,respectively.

12. We also replicate the out-of-sample analysis using levels in ROE and ROA instead ofchanges. Our results are qualitatively the same.


Overall, our results indicate that models that include dividendchanges systematically underperform models that exclude dividends.This suggests that, after accounting for the fact that investors can useonly historical data to estimate the parameters of the earnings model,dividends changes are not reliable predictors of future earnings. One po-tential reason for this result is that the coefficients of dividend increasesand decreases are unstable over time, to the extent that the inclusion ofthese variables in the earnings model generates only noise. The evidence

TABLE 9 The Out-of-Sample Ability of Dividend Changesin Predicting Future ROA Changes:Nonlinear Model

Year Mean p-Value

A. Rolling Scheme


t ¼ 2MSEDIV �MSENODIV .00002b .0146MADDIV �MADNODIV .00012a .0080

B. Recursive Scheme

t ¼ 1MSEDIV �MSENODIV �.00001a .0000MADDIV �MADNODIV �.00006a .0000



ROAt � ROAt�1 ¼ b0 þ b1PDPC0 � RDDIV0 þ b1NDPN0 � RDDIV0

þ ðg1 þ g2NDFED0 þ g3NDFED0 � DFE0

þ g4PDFED0 � DFE0Þ � DFE0 þ ðl1 þ l2NCED0

þ l3NCED0 � CE0 þ l4PCED0 � CE0Þ � CE0 þ "t

to the forecasting ability of a similar model that excludes dividend changes asan explanatory variable. To measure the forecasting ability of each model, wecalculate the out-of-sample forecast error of each model using a rolling schemeand a recursive scheme. Under the rolling scheme, the parameter of variable iin year t is equal to the value of the cross-sectional coefficient of variable i inyear t � 1 when t = 1 and to the value of the cross-sectional coefficient ofvariable i in year t � 2 when t = 2. Under the recursive scheme, the parameterof variable i in year t is equal to the average of all the annual cross-sectionalcoefficients of variable i up to year t � 1 when t = 1 and to the average of allthe annual cross-sectional coefficients of variable i up to year t � 2 when t =2. Using these forecast errors, we calculate the annual differences in squarederrors and absolute deviations between the model that includes dividend changesand the model that excludes them. MSEDIV (MADDIV) is the average annualmean squared error (mean absolute deviation) of the model that includes divi-dend changes. MSENODIV (MADNODIV) is the average annual mean squared er-ror (mean absolute deviation) of the model that excludes dividend changes. Toassess the statistical significance of the MSE and the MAD, we use a bootstrap-ping methodology. To reduce the effect of outliers, all the variables, except thelog of total assets, have been Winsorized at the 0.1% and the 99.9% of the em-pirical distribution. a, b, and c denote significantly different from zero at the1%, 5%, and 10% level, respectively.


in panels B of tables 2 and 3 seems to support this argument. Note that, inthose tables, the coefficients of positive and negative dividend changestend to vary significantly over time.

VII. Conclusions

Since the influential papers of Miller and Modigliani (1961) andWatts (1973), economists have been looking without success for evi-dence that changes in dividends contain information about future changesin earnings. Using various empirical methods, many researchers havebeen unable to find a reliable link between dividend changes and fu-ture changes in earnings or profitability. Using a linear model of earn-ings expectations, however, a recent paper by Nissim and Ziv (2001)finds that dividend changes are positively correlated with future earningschanges.In this paper, we show that dividend changes are uncorrelated with

future earnings changes when one controls for the well-known non-linearities in the earnings process. This result underscores the impor-tance of controlling for nonlinearities in the earnings process whenexamining the performance of a firm following a corporate event. Thus,even when researchers find a weak link between dividend changes andfuture earnings changes (in only about 29% of the years, as we show intable 2), the association is not reliable, and can be attributed to incorrectmodeling of the earnings process.We also find that, regardless of the model of earnings expectations,

models that include dividend changes do not outperform those that donot include dividend changes in out-of-sample tests. Some of our resultseven suggest that investors may be better off not using dividend changeswhen they forecast earnings changes.Given the evidence presented here and in the other papers we cite, it is

sensible to conclude that changes in dividends are not useful in pre-dicting future changes in earnings. It is possible to find a weak asso-ciation between dividend changes and future earnings, but only with anincorrectly specified model. Using several different estimation methodsand various measures of profitability, we find that the association be-tween dividend changes and future profitability is not consistent withthe predictions of the signaling hypothesis. We cannot rule out thatdividend increases signal something, but that something is neither ab-normal increases in future earnings nor abnormal increases in futureprofitability. Perhaps, the motives for paying dividends, and the marketreaction to it, lie elsewhere.13

13. For example, recent evidence suggests that dividend changes contain informationabout unexpected changes in systematic risk (Grullon et al. 2002).


References

Akgiray, V. 1989. Conditional heteroscedasticity in time series of stock returns: Evidenceand forecasts. Journal of Business 62 (January): 55–80.

Allen, F., and R. Michaely. 2003. Payout policy. In Handbook of the economics of finance,vol. 1A, ed. G. M. Constantinides, M. Harris, and R. Stulz. Amsterdam: North-Holland.

Barber, B. M., and J. D. Lyon. 1996. Detecting abnormal operating performance: Theempirical power and specification of test statistics. Journal of Financial Economics 41(July): 359–99.

Bhattacharya, S. 1979. Imperfect information, dividend policy, and ‘‘the bird in the hand’’fallacy. Bell Journal of Economics 10 (Spring): 259–70.

Benartzi, S., R. Michaely, and R. H. Thaler. 1997. Do changes in dividends signal the futureor the past? Journal of Finance 52 (July): 1007–34.

Brav, A., J. R. Graham, C. R. Harvey, and R. Michaely. 2003. Payout policy in the 21stcentury. Working paper. National Bureau of Economic Research, Cambridge, MA.

Brooks, L. D., and D. A. Buckmaster. 1976. Further evidence on the time series properties ofaccounting income. Journal of Finance 31 (December): 1359–73.

DeAngelo, H., L. DeAngelo, and D. J. Skinner. 1996. Reversal of fortune: Dividend sig-naling and the disappearance of sustained earnings growth. Journal of Financial Eco-nomics 40 (March): 341–71.

Elgers, P.T., and M.H. Lo. 1994. Reductions in analysts’ annual earnings forecast errorsusing information in prior earnings and security returns. Journal of Accounting Research32 (Autumn): 290–303.

Fama, E. F., and K. R. French. 2000. Forecasting profitability and earnings. Journal ofBusiness 73 (April): 161–75.

Fama, E. F., and J. D. MacBeth. 1973. Risk, return, and equilibrium: Empirical tests.Journal of Political Economy 81 (May–June): 607–36.

Giacomini, R., and H. White. 2003. Tests of conditional predictive ability. Working paper,University of California, San Diego.

Gonedes, N.J. 1978. Corporate signaling, external accounting, and capital market equilib-rium: Evidence on dividends, income, and extraordinary items. Journal of AccountingResearch 16 (Spring): 26–79.

Graham, J. R. 1996. Is a group of economists better than one? than none? Journal ofBusiness 69 (April): 193–232.

Grullon, G., R. Michaely, and B. Swaminathan. 2002. Are dividend changes a sign of firmmaturity? Journal of Business 75 (July): 387–424.

Hansen, L. P., and R. J. Hodrick. 1980. Forward exchange rates as optimal predictors of futurespot rates: An econometric analysis. Journal of Political Economy 88 (October): 829–853.

John, K., and J. Williams. 1985. Dividends, dilution, and taxes: A signaling equilibrium.Journal of Finance 40 (September): 1053–70.

Keim, D. B., and R. F. Stambaugh. 1986. Predicting returns in the stock and bond markets.Journal of Financial Economics 17 (December): 357–90.

Kennedy, P. 1998. A guide to econometrics. 4th ed. Cambridge, MA: MIT Press.Meese, R. A., and K. Rogoff. 1983. Empirical exchange rate models of the seventies: Do

they fit out of sample? Journal of International Economics 14 (February): 3–24.Michaely, R., R. H. Thaler, and K. L. Womack. 1995. Price reactions to dividend initiations

and omissions: Overreaction or drift? Journal of Finance 50 (June): 573–608.Miller, M. H., and F. Modigliani. 1961. Dividend policy, growth, and the valuation of shares.

Journal of Business 34 (October): 411–33.Miller, M. H., and K. Rock. 1985. Dividend policy under asymmetric information. Journal

of Finance 40 (September): 1031–51.Nissim, D., and A. Ziv. 2001. Dividend changes and future profitability. Journal of Finance

56 (December): 2111–33.Penman, S.H. 1983. The predictive content of earnings forecasts and dividends. Journal of

Finance 38 (September): 1181–99.Pesaran, M. H., and A. Timmermann. 1995. Predictability of stock returns: Robustness and

economic significance. Journal of Finance 50 (September): 1201–28.Watts, R. 1973. The information content of dividends. Journal of Business 46 (April): 191–

211.


Dividend Changes Do Not Signal Changes in Future - Faculty

Documents

Transcript of Dividend Changes Do Not Signal Changes in Future - Faculty