Do dividends convey information about future earnings? Charles Ham Assistant Professor Washington University in St. Louis

Do dividends convey information about future earnings? Charles Ham Assistant Professor Washington University in St. Louis cham@wustl.edu Zachary Kaplan Assistant Professor Washington University in St. Louis zrkaplan@wustl.edu Mark Leary Associate Professor Washington University in St. Louis leary@wustl.edu

Do dividends convey information about future earnings? Abstract Yes. We find that dividend changes predict future unexpected earnings changes in each of the next four quarters. These earnings impacts are persistent, leading to higher than expected earnings levels for at least three years after the dividend change. These results are robust to various measures of expected earnings, including analyst forecasts and a flexible function of past earnings and returns. Second, dividend announcements convey this earnings information to investors. Market reactions to dividend changes are positively related to future earnings changes, and analysts revise their forecasts after the dividend change to completely correct their previous errors. We further find the information content of dividends is larger for dividend cuts, relative to increases, and when information asymmetry is higher. Our study differs from prior research, which finds either no evidence of information content or only short-horizon information content, in that we use quarterly data to delineate earnings announced before and after the dividend change.

1. Introduction In the absence of any market frictions, dividend payouts have no effect on valuation (Modigliani and Miller 1961). In practice, though, it is well documented that market valuations respond sharply to announcements of dividend changes. 1 The two leading theoretical explanations for the positive association between dividend changes and stock prices are the information content hypothesis, which argues that dividend changes convey information about future earnings (Miller and Rock 1985; Bhattacharya 1979), and the free cash flow hypothesis (Jensen 1986) in which payouts constrain managers from over-investing or consuming perquisites. Empirical studies have largely cast doubt on the information content hypothesis, leading many observers to emphasize agency or behavioral explanations. The most damaging evidence against the information content hypothesis has been provided by studies documenting that dividend changes, while correlated with current or past earnings, do not predict future earnings changes. 2 This is surprising given that the vast majority of CFOs agree that dividend decisions convey information to investors (Brav et al. 2005). In this paper, we re-examine the information content of dividend changes by defining more precisely the timing of dividend declarations relative to subsequent earnings announcements. In particular, we use an event window approach where we compare earnings announced prior to the dividend declaration to those announced after the declaration. This approach assumes that any earnings announced after the dividend change were (at least partially) unknown to investors at the time of the dividend declaration. The event window approach differs from the fiscal year approach predominantly used in the prior literature, which groups earnings and dividend changes into fiscal years, and then examines changes across fiscal years. The fiscal year approach has a 1 See, for example, Pettit (1972), Aharony and Swary (1980), and Grullon et al. (2002). 2 See, for example, Benartzi, et al. (1997) and Grullon et al. (2005). 2

tendency to categorize earnings announced after the dividend change but before the end of the fiscal year as current or past earnings realizations (Watts 1973; Benartzi et al. 1997). We document that this timing difference has a substantial impact on one s conclusions about whether dividend changes contain information content about future earnings. Convincingly testing the proposition that dividends have information content requires properly controlling for the earnings the market would expect in the absence of the dividend change. Because of the challenges inherent in measuring earnings expectations, we test the information content hypothesis in three settings, each with different assumptions about pre-dividend declaration earnings expectations. We first use a regression approach that controls for expected earnings changes with a set of linear and non-linear controls for past levels and changes in earnings (Fama and French 2000; Grullon et al. 2005) as well as past returns (Ball and Brown 1968). Our results suggest that earnings are higher than would be expected in the absence of a dividend change for each quarter of the first year after the dividend change. We further show that this higher earnings level persists for at least three years following the dividend change. These results are unaffected by controlling for an interaction between past earnings changes and dividends, suggesting that dividends do not solely convey information about the persistence of past earnings changes. We also find that the information content of a dividend change depends on its direction. Dividend cuts predict changes in earnings that are 2-3 times larger than that of dividend increases, consistent with extant evidence that the market reactions to dividend cuts are larger than those to dividend increases (e.g., Grullon et al. 2002). We also examine whether dividend changes have information content for revenue or gross profit, income statement line items that are less affected by timing choices such as the immediate 3

expensing of period costs and asset write-downs (Donnelson et al. 2011; Novy-Marx 2013). We find that dividend changes also have a positive and persistent association with revenue and gross profit, mitigating concern that our main results are caused by accounting choices that cause accounting income to deviate from economic income. As an extension of the regression approach, we conduct a matching analysis where we compare dividend decrease and increase firms to similar firms that leave their dividends unchanged. We continue to find dividend changes predict future earnings changes and that these differences persist for several years. To reconcile our results with the prior literature, we re-estimate our regressions calculating earnings changes over fiscal years, which leads some earnings realized after the dividend change to be included in the prior fiscal year s earnings. We find dividends no longer predict future earnings changes, consistent with prior research (Watts 1973; Gonedes 1978; Benartzi et al. 1997; Grullon et al. 2005). In our second approach, we study the impact of dividend change announcements on analyst earnings forecasts. This approach has two advantages. First, analyst forecasts provide an alternative measure of expected earnings, which incorporates soft information available to market participants and has been shown to provide accurate forecasts of future earnings (Brown et al. 1987). Second, it enables us to test whether analysts (arguably those investors who study corporate earnings most closely) properly infer the earnings information in dividend change announcements. We show three main results. First, dividend changes are significantly positively related to forecast errors for forecasts made prior to the dividend change. That is, for firms that increase (cut) their dividend, future earnings realizations are higher (lower) than analysts expected before the dividend change. Second, following the dividend announcement, analysts revise their earnings 4

forecasts in the direction of the dividend change. Dividend cuts are associated with forecast revisions that are 3.5 times larger than those for dividend increases. These results suggest analysts view the dividend change as conveying information relevant to forecasting earnings. 3 Third, these revisions completely reverse the errors in pre-dividend forecasts. We find errors from forecasts issued after dividend declarations have no association with the prior dividend change. Interpreting pre-announcement forecasts as a measure of earnings expectations, these results reinforce those in the regression approach by showing that dividend increases (decreases) are associated with increases (decreases) in future earnings relative to expectations. They further highlight that analysts (correctly) interpret dividend changes as containing information about future earnings. In our third set of tests, we examine the link between the market reaction to dividend changes and future earnings. If markets are reasonably efficient, earnings expectations prior to the dividend announcement should already be impounded in the stock price. In this case, any change in expectations should be reflected in the announcement return. We document a significant positive correlation between the cumulative abnormal return in the three-day window surrounding a dividend change announcement and earnings changes over the subsequent four quarters. This provides further evidence that investors comprehend the information content in dividend changes. To isolate the information content of the dividend change itself from other earnings news, we make two comparisons. First, the relation between announcement-window returns and future earnings changes is more than twice as large when firms change their dividends than when the announced dividend is the same as the previous quarter. Second, among dividend changing firms, 3 Unlike prior research (Carroll, 1995; Yoon and Starks, 1995), we include only those forecasts made between the prior earnings announcement and the dividend declaration or between the dividend declaration and the next earnings announcement to ensure that the forecast revisions are not impacted by any new earnings announcements. 5

returns during the dividend announcement window are much more highly correlated with future earnings changes than are returns in the week prior or the week after the declaration date. If dividends convey information about future earnings that is known to managers but not investors, we would expect the predictability of dividend changes for earnings changes to be more pronounced when there is more information asymmetry between insiders and outside investors (Bhattacharya 1979; Miller and Rock 1985). We therefore investigate whether the information content of dividends varies cross-sectionally with proxies for the level of information asymmetry between managers and investors. We first find that dividend declarations contain more information about future earnings when they occur later in the quarter. Just after the earnings announcement, managers should have a smaller information advantage (Verrecchia 1982; Diamond 1985), and thus investors should update less in response to dividend news. Second, dividends have less information content for firms that change their dividend frequently. If dividend changes convey the managers private information, we expect that frequently changing the dividend in the past will reduce information asymmetry, leading subsequent dividend changes to convey less information. Third, we use six proxies for information uncertainty in the extant literature (Zhang 2006): (i) return volatility, (ii) cash flow volatility, (iii) analyst forecast volatility, (iv) analyst coverage, (v) firm size, and (vi) firm age. With the exception of firm age, in each case we find dividend changes by firms with greater information uncertainty have more information content. We note that while these results are consistent with dividend signaling theories, we have not shown that managers consciously bear deadweight costs in order to signal. In fact, firms release dividend declarations more frequently in the beginning of the quarter than at the end, 6

suggesting that in some ways firms time the release of dividend information to limit its information content. These results have several implications. First, they help our understanding of the market response to dividend changes. If dividend changes do not convey information about future earnings, the price reaction to dividend changes is somewhat puzzling. Prior studies have suggested that dividend increases indicate either a reduction in risk (Grullon et al. 2002) or an increase in the fraction of earnings that will be paid out to investors due to a reduction in the agency cost of free cash flow. While our evidence does not rule out these alternative explanations, we document clear predictability of dividend changes and their associated market reactions for future earnings changes. Second, our results suggest that investors understand the earnings information contained in dividend announcements both the direction and the magnitude of announcement returns and analyst forecast revisions are correlated with future changes in earnings. Our conclusion stands in contrast with the current consensus that there is little empirical support for the information content hypothesis or signaling theories of dividends (see, for example, reviews by Allen and Michaely (2002) and DeAngelo et al. (2009)). This difference between our conclusions and those of much of the prior literature derives from our approach to defining future earnings changes. First, by utilizing an event window approach, we ensure that all earnings announcements made after a dividend announcement are considered future earnings. Indeed, to the extent that managers have private information about earnings, this could pertain to any earnings realizations that have not yet been announced. Second, we calculate future earnings changes relative to earnings in the year or quarters prior to the dividend change, while controlling for the earnings changes that would be expected given pre-dividend earnings patterns. Benartzi et al. (1997) point 7

out that after the first several quarters following a dividend change, earnings do not continue to grow at a faster rate for firms that increase their dividend. However, we show that the new earnings levels that result from faster growth in the first four quarters after the dividend change persist for at least three years after the change. At the same time, our results suggest that dividend changes do not simply reflect the persistence of past earnings changes (Koch and Sun 2004; Skinner and Soltes 2011). Rather, dividend changes predict earnings changes in the quarters following the announcement that have a persistent impact on future earnings levels. The remainder of the paper is organized as follows. The next section reviews the related literature and discusses testable hypotheses for the information content view of dividends. Section 3 describes our data and sample selection. In section 4, we present results of our three approaches to testing for the information content of dividends. Section 5 documents the cross-sectional relation between the information content of dividend announcements and proxies for information asymmetry. We offer concluding remarks in Section 6. 2. Literature review and hypothesis development In this section, we first briefly review the theoretical literature on dividend signaling and the information content view of dividends. We then review the prior empirical literature testing whether dividend declarations predict future earnings changes. To place our findings in context, we classify studies as either event window studies those which classify all earnings realizations after the dividend declaration as future earnings or fiscal year studies those which classify earnings realizations as future if they occur the fiscal year after the dividend declaration. Most studies adopting an event window methodology find evidence consistent with dividends having 8

short-horizon information content about future earnings, while the majority of studies using a fiscal year approach find evidence inconsistent with the information content hypothesis. 2.1. Theory and hypotheses The seminal dividend irrelevance model of Miller and Modigliani (1961) assumes, among other things, that managers and outside investors have the same information with respect to investment policy and the value of future cash flows. However, differences in information between insiders and investors are likely to be prevalent in financial markets in practice. Indeed, Miller and Modigliani (1961) discuss the possibility that changes in the dividend level will be interpreted by investors as reflecting a change in managers views of the firm s future growth prospects. They suggest this as a means of reconciling the observed price reaction to dividend changes with their irrelevance proposition. 4,5 As discussed by Miller and Rock (1985), this information content can arise simply through the sources and uses of cash identity. That is, if the firm s investment policy is known or inferred by investors, then the dividend, net of any capital raised, allows investors to back out the (unobserved) earnings. However, Miller and Rock also point out that this version of the information content hypothesis is only sustainable in equilibrium under restrictive assumptions. In particular, if managers have any incentive to boost the current stock price, rather than solely maximizing fundamental value, they might lower investment or take other measures to increase the current 4 This relaxation of perfect markets in itself does not undermine Miller and Modigliani s dividend irrelevance proposition, since in this case it is the information about earnings, rather than the dividend per se that is relevant for valuation. 5 We note that other (non-mutually exclusive) explanations have been offered for the price reaction to dividend changes. Following Easterbrook (1984) and Jensen (1986), higher dividends may reduce the free cash flow subject to managerial discretion, thereby increasing the fraction of future earnings captured by investors. Alternatively, Grullon et al. (2002) suggest that dividend increases reflect a reduction in risk, and therefore a lower discount rate, as firms mature. Given our focus on the earnings information content of dividends, we refer the reader to excellent reviews by Allen and Michaely (2003), DeAngelo et al. (2009) and Kalay and Lemmon (2011) for fuller treatments of these alternate views. 9

dividend. This possibility for manipulation, in turn, would undermine the information content of dividends if investors are rational. One way in which information content can be restored in equilibrium is if increasing the dividend is costly enough to discourage manipulation by firms whose future earnings prospects don t warrant the increase. This idea has been formalized in a number of dividend signaling models, which differ primarily in the assumed cost of paying a high dividend. Bhattacharya (1979) presents a model in which committing to a high dividend exposes firms to a higher likelihood of having to raise costly external financing in the future. In Miller and Rock (1985), the cost of increasing the dividend is a departure from optimal investment policy, while John and Williams (1985) focus on the tax cost of dividends relative to capital gains. However, in all of these models, the dividend announcement allows investors to infer managers private information about current and future profitability. Dividend signaling models, or the information content hypothesis more generally, have several testable implications. First, if dividend decisions are a function of managers private information about current and future earnings, dividend increases (decreases) should be associated with subsequent increases (decreases) in earnings realizations. Second, if investors recognize the earnings news reflected in dividend announcements, dividend changes should be greeted by price changes in the same direction. Related, investors should update their expectations about future earnings following announced dividend changes. 2.2. Empirical tests of the information content of dividends There is a lengthy literature testing whether dividends have information content, commonly defined as information about unexpected future earnings changes. While a few studies support the information content view of dividends (Ofer and Siegel 1987; Aharony and Dotan 1994; Yoon and 10

Starks 1995; Nissim and Ziv 2001), most large sample empirical studies argue dividend changes contain little or no information about future earnings (Watts 1973; Gonedes 1978; Penman 1983; Lang and Litzenberg 1989; DeAngelo et al. 1996; Benartzi et al. 1997; Grullon et al. 2002; Grullon et al. 2005). 6 Commentary in recent review papers on payout policy suggests the current consensus is that there is little empirical support for the information content of dividends. For example, DeAngelo et al. (2009) state Knowledge of the fact that a firm has increased its dividend generally does little to improve forecasts of future earnings over and above what outsiders can infer from current earnings (p. 99) and Researchers have struggled to find evidence that dividend increases are reliable signals of future earnings increases (p. 185). Allen and Michaely (2003) opine similarly the overall accumulated evidence does not support the assertion that dividend changes convey information about future earnings (p. 73). Our review of the empirical literature suggests one research design choice has a dramatic influence on the probability a study will confirm or contradict the information content hypothesis. The pivotal research design choice is whether the study computes earnings changes using an event window approach or whether the study computes earnings changes over the fiscal year. In the event window methodology, earnings (or earnings expectations) immediately following the dividend declaration are compared to earnings prior to the dividend declaration. In the fiscal year methodology, dividend changes are aggregated over a fiscal year. These studies then compare earnings changes in the fiscal year following the dividend declaration to earnings changes in the year in which the firm declared the dividend change. The majority of studies employing the fiscal year approach do not support the information content hypothesis. Specifically, seven studies find 6 We exclude studies from our review that use a small subset of dividend paying stocks, such as Brickley (1983), which studies earnings changes for thirty-five firms that change their dividend. We also exclude studies examining dividend omissions and dividend initiations. 11

no information content (Watts 1973; Gonedes 1978; Penman 1983; DeAngelo et al. 1996; Benartzi et al. 1997; Grullon et al. 2002; Grullon et al. 2005), while only two support information content (Aharony and Dotan 1994; Nissim and Ziv 2001). Eight of these studies follow the fiscal year methodology and seven of those find no information content. 7 Perhaps the most comprehensive of these studies is Benartzi et al. (1997). Using both a regression and matched-sample approach, the authors show that dividend changes are highly correlated with earnings in the current or past fiscal years. However, dividend increases are uncorrelated with earnings growth in the subsequent fiscal years, while dividend cuts are actually followed by earnings increases. Nissim and Ziv (2001) argue that controlling for mean reversion in earnings produces results more consistent with the information content of dividends. They perform similar tests as Benartzi et al. (1997), but add lagged return on equity and lagged earnings changes to control for changes in earnings predicted by financial statement variables. However, Grullon et al. (2005) argue these findings are highly sensitive to the manner of controlling for mean reversion and demonstrate that controlling for non-linearity in mean reversion restores the conclusions of Benartzi et al (1997). The only study in this set, of which we are aware, with results that support the information content hypothesis and have not been challenged is Aharony and Dotan (1994). Not coincidentally, this study uses an event window methodology. However, this study shows positive information content for only two quarters after the dividend change and negative information content in the fourth quarter. It is unclear whether evidence that the association between dividend changes and earnings changes varies with horizon should be interpreted as consistent or inconsistent with information content. We expand on the methodology in the study by including 7 Several of these studies consider dividend changes in the first quarter of the subsequent fiscal year as part of the prior fiscal year s earnings. 12

extensive controls for pre-dividend declaration earnings and returns, which allows us to isolate the unexpected information content in the dividend change and delivers consistent results across the earnings horizon. We also make several additional contributions. First, we highlight the source of discrepancy between their findings and the bulk of the related literature. Second, we demonstrate how dividend announcements affect investor expectations by linking the information content of dividends to analyst forecast revisions and announcement period returns. Third, we use crosssectional tests to document that the information content of dividends is more pronounced in settings of greater information asymmetry. In contrast to studies examining information content using actual earnings changes where the fiscal year approach is the norm, the three studies of which we are aware using analyst forecasts all use an event study methodology where they compare expectations for the same time period before and after the dividend. Two of the three studies find significant information content (Ofer and Siegel 1987; Yoon and Starks 1995) while one does not (Lang and Litzenberg 1989). However, all these prior studies use summary files, which offer only approximate information about the timing of revisions. A limitation of such a research design is that these studies cannot rule out the possibility that the information causing the revision was either (i) a concurrent earnings release, or (ii) information released before the dividend declaration (Allen and Michaely 2003). By using the I/B/E/S detail file, we are able to ensure that we compare only forecasts made after the previous earnings release but before the dividend change to forecasts made between the dividend change and the next earnings release. Further, we remove the impact of analyst-specific biases by only including forecasts made by the same analyst before and after the dividend change. 3. Sample selection and descriptive statistics 13

We obtain data on dividend declarations from the CRSP events database. We first select all ordinary quarterly dividend declarations (distribution code 1232) over the period 1972 2014 for which the firm made a previous quarterly dividend declaration in the past 180 days. 8 This allows us to compute the percentage dividend change. We limit the sample to: (i) firms listed on the NYSE, AMEX, or Nasdaq exchanges, (ii) ordinary common stocks (i.e., those with share code 10 or 11), and (iii) non-financial firms (we exclude firms with a four digit SIC beginning with six). We also exclude: (i) dividend declarations for which the firm declared a distribution other than a quarterly dividend between the declaration dates of the current and prior quarterly dividends, to focus our analysis on the information content of quarterly dividends (Benartzi et al. 1997; Nissim and Ziv 2001), and (ii) firms that split their shares between the month of the prior dividend declaration and the month of the current dividend declaration, as splits affect the information content of dividend changes (Nayak and Prabhala 2001). We require data on CRSP to compute dividend declaration announcement returns and past returns. We obtain earnings data from the CRSP/Compustat Merged database and require twelve consecutive quarters of earnings to calculate seasonal earnings changes around the dividend declaration (i.e., four earnings changes before and after the declaration). We winsorize all continuous variables at the top and bottom one percent to mitigate the influence of outliers, except the percentage dividend change for which we set all dividend increases larger than 200% to 200%. 9 8 The first year earnings announcements were available on Compustat is 1972. 9 Several dividend increase observations are extremely large in percentage terms, so to mitigate their influence we winsorize the percentage dividend change at +200%. We do not winsorize dividend decreases because they are bounded at -100% and dividend decrease observations comprise just over 1% of the sample. We winsorize all variables involving earnings at the top and bottom one percent for two reasons: (i) the distribution of changes in earnings values is highly kurtotic and skewed so extreme values account for much of the variance in earnings changes (Grammery and Gerakos 2014), and (ii) large changes in accounting income have little relation with economic income (Freemen and Tse 1992). All standard errors are clustered by year of the dividend declaration. 14

Table 1 presents descriptive statistics for our sample. As shown in Panel A, 85.0% of dividend declarations maintain the prior dividend level, while 14.0% (1.0%) increase (decrease) the dividend. Although dividend decreases are less frequent, they tend to have a larger percentage effect on the dividend. The average decrease reduces the dividend by 49.8% (Panel C) while the average increase raises the dividend by 19.6% (Panel D). The average decrease has announcement window returns of -2.6% while the average increase has announcement window returns of only 0.8%, suggesting a greater reaction to dividend decreases. Declarations that change the dividend tend to be preceded by returns of the same sign as the dividend change, suggesting at least some of the information affecting the decision to change the dividend was released to the market before the dividend declaration. Examining earnings changes, we find positive (negative) earnings changes for firms that increase (decrease) the dividend in the four quarters before and after the dividend declaration (except four quarters ahead for dividend decreases). The goal of our empirical tests is to identify the portion of the post-dividend declaration earnings change that is unexpected. 4. Do dividend changes predict future earnings changes? In this section, we test whether dividend changes have information content about future earnings via three distinct approaches. First, we estimate the relation between dividend changes and future earnings changes in a regression approach where we include variables suggested by the extant literature to control for expected earnings changes in the absence of any dividend change. We also augment this approach with a matched sample approach that compares earnings changes for dividend changers to non-dividend changers with similar characteristics. Second, we use analyst forecasts of earnings to examine whether dividend changes predict forecast revisions and 15

forecast errors. Third, we infer changes in earnings expectations from market prices and examine whether stock returns on dividend declaration days contain more information about future earnings relative to comparable days during the period. 4.1. A regression approach to testing for information content In our first approach to testing for the information content in dividend changes, we regress earnings changes on the percentage dividend change (ΔDIV) and a series of control variables. Eit+n = β0 + β1 DIVit + βjcontrols + ε (1) E is the change in earnings using income before extraordinary items (IBQ) from the CRSP/Compustat merged quarterly file. All earnings changes are computed as the difference between earnings announced after the dividend change and earnings for the same period in the prior year (before the dividend change) and scaled by the market value of equity the quarter before the dividend announcement, similar to Benartzi et al. (1997). We calculate earnings changes over five different horizons: one, two, three, and four quarters ahead, as well as one year ahead which is the sum of the four quarterly changes after the dividend announcement. Refer to Figure 1 for a visual depiction of the earnings change calculations. If a dividend declaration occurs the day of an earnings announcement, we use the earnings announced at the time of the dividend change as the prior quarter s earnings. We cluster all standard errors by the year of the dividend declaration. We present the results from estimating equation (1) in Table 2. In column (1), we regress the annual change in earnings ( E(y+1)), calculated as the sum of the four quarterly earnings values following the dividend change minus the sum of the four quarterly earnings values before the dividend change, on the percentage dividend change ( DIV). We find a highly significant coefficient on the dividend change (β=0.028; t=4.2), which suggests a 50% increase in the dividend corresponds to an increase in earnings equal to 1.4% of the market value of equity of the firm. 16

One concern with our results measuring information content is that the earnings changes associated with the dividend change might have been expected in the absence of the dividend change. To address this concern, in column (2) we include the four past quarterly earnings changes and four past earnings levels. We find our coefficient estimates and significance levels are unchanged by the inclusion of these variables, suggesting the univariate finding of information content cannot be explained by past earnings changes. In column (3), we include non-linear functions of the annual earnings change and level (Fama and French 2000; Grullon et al., 2005), to more fully control for variation in expected earnings changes associated with past earnings realizations. 10 We also include five variables capturing returns over the 240 trading days before the dividend announcement, to capture information about expected earnings changes impounded into returns (Ball and Brown 1968). We continue to find a highly significant coefficient on the dividend change (β=0.024; t=4.1). Prior literature suggests at least some results are sensitive to the choice of deflator (Nissim and Ziv 2001). To examine whether our results are invariant to the choice of deflator, in columns (4) (6), we use three different deflators: the market value of equity five quarters before the dividend declaration in column (4), the book value of common equity five quarters before the dividend declaration in column (5), and the book value of common equity one quarter before the dividend declaration in column (6). We continue to find the dividend change has a significant effect on earnings changes in the year following the dividend declaration in all three specifications. 10 The past earnings level (earnings change) is the sum of the four quarterly earnings levels (earnings changes) before the dividend announcement. Specifically, we include a total of six variables, three each for the earnings change and level: (i) an interaction between the variable and an indicator equal to one if the variable is negative, (ii) an interaction between a positive indicator and the variable squared, and (iii) an interaction between a negative indicator and the variable squared. We exclude the main effect because it will be multi-colinear with our four quarterly earnings change and levels variables. We find our coefficient estimates and significance levels are unchanged by including non-linear controls for each quarterly change and level. 17

In columns (7) (10), we examine the horizon over which dividend changes correlate with earnings changes, by regressing future quarterly changes in earnings on the percentage dividend change. In each column, the dependent variable is earnings from a post-dividend declaration quarter minus earnings from the same quarter in the prior fiscal year, scaled by the market value of equity the quarter before the dividend declaration. In column (7), we examine the association between the dividend change and the earnings change one quarter ahead ( E(q+1)). We find a statistically significant coefficient on ΔDIV (β=0.008; t=4.6) that is approximately 33% of the coefficient when the dependent variable is the annual earnings change. In columns (8), (9), and (10) we examine the association between the dividend change and the earnings change two, three, and four quarters ahead, respectively. The coefficient remains significant at all horizons, but decreases monotonically with horizon: β=0.006 at two quarters ahead, β=0.004 at three quarters ahead, and β=0.003 at four quarters ahead. Our finding of a significant, positive association between dividend changes and unexpected earnings at horizons of three and four quarters contrasts with Aharony and Dotan (1994), who do not control for expected earnings changes and find negative or insignificant information content at these longer horizons. 4.2. Comparison with prior literature The prior literature has typically computed earnings changes over fiscal years, which can result in earnings that occur after the dividend declaration but before the fiscal year end being included in the prior year s earnings (i.e., the year of the dividend change). In other words, in some cases, earnings that are announced well after the dividend change are not considered future earnings. To examine whether the disagreement between our findings and those of the prior literature are attributable to computing earnings changes over the fiscal year, in Table 3, we 18

calculate earnings changes as in the prior literature earnings in the fiscal year after the dividend declaration less earnings in the fiscal year of the dividend declaration. In column (1), we regress the fiscal year earnings change ( E(fy+1)) on the dividend change. We find a marginally significant and negative coefficient on the dividend change (β=-0.009; t=-1.9), suggesting dividends have no or negative information content when computing earnings changes over fiscal years. In column (2), we add non-linear controls for past earnings changes and levels (Fama and French 2000; Grullon et al. 2005), which much of the prior literature has used in its estimation of information content. We find an insignificant, although positive coefficient on the dividend change (β=0.003; t=0.9). In column (3), we add the return controls and find a coefficient estimate of zero. In column (4), we include the earnings level and change for the past four quarters and find the coefficient on the dividend change remains insignificant. Overall, these results are consistent with the conclusions of the prior literature that when computing earnings changes over fiscal years, dividend changes appear to have minimal or no information content. 4.3. Information content of dividends beyond the subsequent year The results in columns (7) (10) of Table 2 appear to suggest the information content of dividends declines with horizon. This result is somewhat puzzling given that dividend changes tend to be persistent (Lintner 1956). We expect that if dividend changes represent a commitment to change cash outflows for a multi-year period, cash inflows should also change over a similar duration. In Panel A of Table 4, we examine whether the relation between dividend changes and future earnings changes extends beyond the four quarters after the dividend declaration. In column 19

(1), we examine the relation between the dividend change and the earnings change five to eight quarters ahead. We compute the earnings change taking the difference between the sum of the earnings announced five to eight quarters after the dividend declaration less the sum of quarterly earnings from the four quarters before the dividend declaration. We use pre-dividend declaration control variables, so we do not control for any of the earnings changes realized in the first year after the dividend change. As a result, the coefficient on ΔDIV allows us to examine the persistence of the earnings changes associated with the dividend change. We find a positive and statistically significant coefficient on the dividend change (β=0.013; t=2.6), suggesting dividends have information content for earnings for at least two years. The magnitude is 54% of the magnitude of the coefficient in column (3) of Table 2, suggesting some mean reversion in the earnings change. In column (2), we specifically test for mean reversion by computing the dependent variable as the difference between quarterly earnings recognized five to eight quarters after the dividend change and the first four quarters announced after the dividend change. We find a significantly negative coefficient estimate (β=-0.011; t=2.9), consistent with mean reversion in the earnings changes associated with the dividend changes. In columns (3) (4), we conduct similar analyses, except that we take the difference between the sum of earnings announced nine through twelve quarters after the dividend change and earnings announced the four quarters before the dividend change. Our test of information content at the three-year horizon allows us to test whether the information content of dividends diminishes further as the horizon extends. We continue to find significant information content, and the coefficient estimate is actually greater than the coefficient estimate in column (1). In column (4), we take the difference between earnings announced nine through twelve quarters after the dividend declaration and earnings announced five through eight quarters after the dividend 20

declaration. Unlike in column (2), we find no economically or statistically significant mean reversion at the three-year horizon. Overall, our results suggest that dividend changes have a positive association with future earnings over a long horizon, but the association mean reverts with most of the mean reversion occurring shortly after the dividend declaration. The strong association with earnings changes shortly after the dividend change can further explain why research designs which do not classify all earnings changes after the dividend change as future earnings, find evidence inconsistent with information content. 4.3.1. Information content of dividends for revenue and gross profit changes One potential concern with our use of accounting income to measure changes in economic income, is that accounting standards accelerate expenses into earnings and thus proxy for economic income with error. Because the accelerated expenses often proxy for investment, it is possible they are correlated with the dividend change. 11 To address this concern, we test for information content using revenues and gross profit (i.e., revenue minus cost of goods sold). Because cost of goods sold are matched explicitly to the revenues they generate, gross profit should be largely unaffected by inter-temporal variation in investment. 12 To test for information content using revenue, we estimate equation (1) replacing all earnings amounts in both dependent and independent variables with revenue amounts. Specifically, we regress the change in revenue on the level and change in revenue over each of the past four quarters while still controlling for past returns. We scale each of the revenue variables 11 Two well documented reasons why accounting income differs from economic income is: (i) the accounting system requires immediate expensing of investments such as advertising and research and development, even though the firm realizes the benefits of these expenses over a period of years (Luminita and Srivastava 2016), and (ii) the accounting system requires assets to be written down when impaired (Basu 1997). 12 Novy-Marx (2013) states that gross profits is the cleanest accounting measure of true economic profitability. 21

by the market value of equity the quarter before the dividend declaration. In column (1) of Table 4 Panel B, we calculate the dependent variable as the sum of revenue for the four quarters after the dividend declaration less the sum of revenue for the four quarters before the dividend declaration. We find a statistically significant positive coefficient estimate on the dividend change (β=0.084, t=2.7), consistent with dividend changes conveying information about future sales. In columns (2)-(5), we break out the annual change into quarterly changes and find positive and statistically significant coefficient estimates one quarter (β=0.016, t=2.7), two quarters (β=0.021, t=2.5), three quarters (β=0.022, t=2.5), and four quarters (β=0.024, t=2.6) ahead. It is noteworthy that the information content of revenues actually increases across the horizon. In columns (6)-(10), we estimate equation (1) replacing all earnings variables with gross profit variables. In column (6), when estimating information content at the annual horizon, we find a statistically significant positive coefficient estimate on the dividend change (β=0.026, t=3.1). In columns (7)-(10), we find positive and statistically significant coefficient estimates on the dividend change one quarter (β=0.008, t=3.4), two quarters (β=0.007, t=2.9), three quarters (β=0.006, t=2.8), and four quarters (β=0.006, t=2.6) ahead. Similar to our results for revenue, we observe little attenuation in information content over the first year after the dividend change. 4.4 Information content of dividend decreases vs. increases Our finding in section 4.1 that dividend changes have information content suggests that a change in expected future cash flows can explain the positive association between dividend changes and returns. Empirically, negative dividend changes have a larger effect on returns than positive dividend changes. If variation in information content causes the association between dividend changes and returns, we would expect dividend decreases to have more information content about future earnings than dividend increases. 22

In Table 5, we examine whether the relation between future earnings and dividend changes varies with the sign of the dividend change via a modified version of equation (1). Eit+n = β0 + β1 DIVit + β2 DIVit*I[ DIV<0]it + βjcontrols + ε (2) Where I[ DIV<0] is an indicator variable equal to one if DIV is negative, zero otherwise. We include both the level and change of the past four quarterly earnings changes and controls for past returns in all models. In column (1), when using annual earnings changes as the dependent variable, we find a significantly positive coefficient on both DIV (β=0.016; t=3.2) and the interaction DIV*I[ DIV<0] (β=0.047; t=5.3). In columns (2)-(5), we examine how the information content of positive and negative dividend changes varies with the horizon over which we compute earnings changes. Consistent with the results in Table 2, the earnings information content of both dividend decreases and increases decline with the horizon. However, the dividend decrease has an association with earnings changes at least double that of dividend increases. The dividend decrease has a significantly larger association with earnings changes than increases for the first three quarters. Overall, the results indicate that both positive and negative dividend changes have associations with future earnings consistent with the sign of their announcement window returns. Further, negative dividend changes have a larger impact on future earnings than positive changes, consistent with the asymmetric reaction to dividend news. 4.5. Matched sample analysis The regression results presented thus far use a series of lagged earnings realizations and return realizations to control for expected earnings changes. However, if the effect of past earnings changes on future earnings changes varies with the size of the firm, its industry, and/or over time, these interactions could lead us to find dividend changes predict earnings changes when the 23

predictability arises because of these heterogeneous effects. To address this possibility, in Table 6, we match dividend change firms to similar dividend paying firms that do not change their dividend. Specifically, we estimate a propensity score model of the probability the firm will change the dividend as a function of the past four quarterly dividend changes and levels. I[ DIV 0]it = β0 + β1eiq-1 + β2eiq-2 + β3eiq-3 + β4eiq-4 + β5 Eiq-1 + β6 Eiq-2 + β7 Eiq-3 + β8 Eiq-4 + ε (3) We estimate the model separately for dividend increases (I[ DIV>0]) and decreases (I[ DIV<0]). We then match each dividend increase or decrease firm to a firm that did not change the dividend. We choose the firm with the closest propensity score within the same dividend declaration year and industry (two digit SIC). Matching is performed with replacement and we impose a caliper distance of 0.01 (Shipman et al. 2016). Panel A of Table 6 and Figure 2 report the level and change in earnings from four quarters before, to four quarters after, the dividend declaration for dividend decrease firms and the matched firms that did not change the dividend. Panel A demonstrates that both dividend decreasing firms and matched non-changers exhibit declining performance in the four quarters prior to the dividend declaration, though the differences between the dividend changers and non-changers are insignificant. However, the firms that decrease their dividend exhibit significantly worse performance the quarter after the dividend declaration. The significantly worse performance persists for each of the next four quarters, again illustrating persistent information content. In Panel B of Table 6 and Figure 2, we repeat the analysis for firms that increase the dividend. Opposite to Panel A, both the dividend increasing firms and the matched non-changers exhibit improving performance via positive earnings changes in the four quarters before the dividend declaration. The differences in earnings changes during the pre-period are again 24