Evaluating the accrual-fixation hypothesis as an explanation for the accrual anomaly

Transcription

1 Evaluating the accrual-fixation hypothesis as an explanation for the accrual anomaly Tzachi Zach * Olin School of Business Washington University in St. Louis St. Louis, MO Tel: (314) zach@olin.wustl.edu First version: August 2004 This version (1.1): September 2004 Abstract. Sloan (1996) s result of return predictability based on accrual information has generated a large stream of literature. Sloan s (1996) explanation for the phenomenon is that investors fixate on accounting accruals without taking into account their tendency to reverse. In this study, I directly examine this explanation. I use three properties of the accrual fixation hypothesis reversals, symmetry and overreaction to generate testable empirical predictions. My results are not consistent with the accrual-fixation hypothesis. First, I find that extreme accrual firms tend to remain in extreme decile in two consecutive years. Second, I find differences in return behavior across extreme high and low accrual decile. Finally, I do not find any evidence of overreaction to accrual information. * This paper is based on parts of my dissertation completed in the University of Rochester. I would like to thank the members of my dissertation committee: Bill Schwert, Jerry Warner, Ross Watts (chair) and Jerry Zimmerman for their valuable comments and suggestions. I also benefited greatly from the insights of an anonymous referee, Philip Joos, S.P. Kothari, Andy Leone, Thomas Lys (the editor) Michela Verardo and Charles Wasley I also thank the Deloitte and Touche Foundation for generous financial support. All errors are my own.

2 1 1. Introduction Sloan (1996) and a number of subsequent studies present evidence that a trading strategy based on publicly available operating accruals information earns abnormal returns of approximately 10% in the year following its initiation. The strategy is based on firms in the extreme high and low deciles of accruals crosssectional distribution. This empirical regularity has been termed the accrual anomaly. Sloan (1996) also shows that accruals and cash flows exhibit differential persistence with respect to future earnings. Sloan attributes the return predictability to the market s inability to take the differential persistence into account and to then process accrual information correctly. More specifically, Sloan (1996) argues that stock prices act as if investors fixate on earnings, failing to distinguish fully between the differential properties of the accrual and cash flow components of earnings. This study is the first to directly examine Sloan s (1996) explanation for the causes of the accrual anomaly. I label Sloan s conjecture that investors naively fixate on accruals as the accrual-fixation hypothesis. According to this hypothesis, return predictability arises because of mispricing driven by investors misinterpreting accounting information per se. Investors misprocessing of accounting information is the essence of the accrual-fixation hypothesis. Under the accrual-fixation hypothesis, and based on the results of Mishkin tests of rational pricing, market participants treat accruals and cash flows similarly, not taking into account the transitory nature of accruals (i.e. accruals tendency to

3 2 reverse). As a result, firms with extreme accruals are mispriced which results in future abnormal returns. Since abnormal returns appear in extreme accrual deciles, I expect that firms in these deciles will exhibit reversal patterns similar to those of accruals. Under this scenario accrual extremeness is expected to be transitory. Further, if aggressive accounting is related to accrual-fixation then accruals for a given firm are not likely to remain extreme in the following period because of disciplinary mechanisms, such as auditor oversight. Thus, the accrual-fixation hypothesis predicts that extreme accrual deciles are likely to be composed of different firms across periods. Another characteristic of the accrual-fixation hypothesis is symmetry. That is, the hypothesized accrual misprocessing is assumed to be similar in both high and low accrual firms. Indeed, Figure 1 in Sloan (1996) depicts reversal patterns in both extreme deciles. Both of these reversal patterns are hypothesized to not be understood by investors. This is supported by the regression models in Sloan (1996) which pool together all observations, including those in high and low deciles, thereby implying that the hypothesized information misprocessing is similar in both extreme deciles. Investors presumably do not appropriately adjust for the differential persistence of accruals and cash flows in both the low and high accrual deciles. Finally, the accrual-fixation hypothesis implies a certain behavior by market participants. In particular, the hypothesis and the empirical results are consistent with investors overreacting to accrual information. At the publication of financial statements containing high (low) accruals, investors incorrectly extrapolate the

4 3 accrual performance into the future and are surprised when the projections do not materialize. For my analysis, I use the above three features of the accrual-fixation hypothesis - reversals, symmetry and overreaction - to generate testable empirical predictions. I then perform several empirical tests to evaluate whether the data is consistent with these predictions. I examine the time-series pattern of extreme accrual firms to see whether these firms tend to leave the extreme accrual deciles in the following period or whether they are sticky. 1 In other words, do these firms habitually reside in extreme deciles in successive years? The accrual-fixation hypothesis predicts that extreme accrual firms will leave the extreme accrual portfolios in the following year. I also report characteristics associated with extreme accrual firms to determine whether these characteristics are different for high and low accrual decile firms. That is, I investigate whether the accrual anomaly behaves in a similar or different fashion across the low and high accrual deciles. Under the accrual-fixation hypothesis investors fixation is assumed to be identical for high and low accrual deciles and thus the behavior of the anomaly in both deciles is expected to be similar. Finally, I examine the relation between returns around earnings announcements in successive years to unveil any evidence of overreaction to accrual information. The overreaction implied by the accrual-fixation hypothesis would lead 1 Throughout this paper, sticky refers to firms that belong to an extreme accrual decile (high or low) in two consecutive years.

5 4 to a prediction of a negative relation between earnings announcement returns in two successive years. Summary of results. In the time-series analysis, I find that extreme accrual firms are habitual extremes. That is, about 25% of firms in portfolios of extreme high (low) accruals have been in the same portfolios in the previous year. This frequency is much higher than a frequency of 10%, which assumes independence across years and ignores the tendency of accruals to reverse. Taking into account accrual reversal, which is an important element in the accrual-fixation hypothesis, will result in an expected frequency that is even lower than 10%. Such stickiness is observed across various measures of accruals, including working capital accruals (measured based on the balance-sheet or the statement of cash flows) and investing accruals. This evidence is inconsistent with the accrual-fixation hypothesis. Abnormal returns to an accrual strategy are, to a large extent, driven by these sticky firms. In NYSE\AMEX firms, sticky high accruals are drivers of negative future abnormal returns in the high accrual decile. In non-nyse\amex firms, sticky low accruals are drivers of positive future abnormal returns in the low accrual decile. Such a difference is not consistent with the symmetry under the accrual-fixation hypothesis. I also report evidence on characteristics of sticky firms. Sticky high accrual firms are associated with extreme past and future growth. On the other hand, sticky low accrual firms have a higher probability of bankruptcy.

6 5 In my tests of the relation between successive earnings announcement returns, I do not find evidence consistent with overreaction. Instead, I find some mild evidence of underreaction. This evidence is also inconsistent with the accrual fixation hypothesis. Collectively, my evidence suggests that there exist one or more firm characteristics, that (i) are correlated with accruals, (ii) are not likely to change dramatically over time, perhaps because they are closely tied to the underlying economic environment in which a firm operates and (iii) are different between high and low accrual firms. An alternative explanation to the accrual anomaly arises if such characteristic(s) that lead to extreme accruals are related to future abnormal returns. My analysis highlights the importance of recognizing that extreme accruals can be generated not only by accounting manipulations but also by the underlying economic environment in which a firm operates. The rest of this paper is organized as follows. Section 2 provides a brief overview of the relevant literature and a discussion of the three properties of the accrual-fixation hypothesis used here. The data are described in Section 3. Results based on the reversal and symmetry properties are reported in section 4. Results based on overreaction are reported in section 5. Section 6 provides conclusions.

7 6 2. The accrual fixation hypothesis 2.1 Literature background Sloan (1996) reports two results. First, he shows that abnormal returns are predictable based on current accrual information. Second, he shows that accruals and cash flows exhibit differential persistence with respect to future earnings. To link the return predictability result with the persistence result, Sloan (1996) performs Mishkin (1983) tests. They reveal that the persistence coefficients of accruals and cash flows that are implied in market prices are different from the true coefficients obtained from a predictive regression of future earnings on current accruals and cash flows. More specifically, the market assigns a higher persistence to the accrual component of earnings than the one implied by the predictive regression and, as a result, over (under)-prices firms whose earnings contain high (low) accrual components. The conclusion reached is that return predictability is due to investors processing accounting information in a particular, yet incorrect, way. The conclusion is essentially Sloan s (1996) original interpretation for the accrual anomaly, which has become the most widely accepted explanation for the phenomenon. I call this explanation the accrual-fixation hypothesis. According to the accrual-fixation hypothesis, return predictability is due to mispricing driven by investors mis-interpreting accrual information per se. When pricing securities investors do not consider the future reversals of accruals. Sloan views future reversals to be a result of aggressive or bad accounting that originally inflated accruals. Investors, for some reason, do not process the information correctly leading to

8 7 mispricing. It is important to note that the accrual-fixation hypothesis originated from the results of the Mishkin tests. In that sense, the hypothesis is descriptive in nature and does not, for example, stem from predictions of an information misprocessing theory. The literature has taken several paths to investigate further the accrual anomaly. First, several studies have looked for cross-sectional variation in the accrual strategy s returns across accrual components (e.g Xie, 2001; Thomas and Zhang, 2002; Richardson et al. 2003). The results indicate that the mispricing is not only associated with abnormal accruals and inventory accruals but also with long-term investing accruals. Second, some studies investigate the relation between the accrual anomaly and other known anomalies (e.g. Collins and Hribar, 2002; Zach, 2003). These studies conclude that the accrual anomaly is distinct from the post-earningsannouncement drift and anomalies associated with corporate events. However, Desai et al. (2004) argue that the accrual anomaly is related to the value-glamor anomaly. Third, recent studies investigate the implementability of an accrual-based trading strategy (Lev and Nissim, 2004; Mashruwala et al., 2004; Bushee and Raedy, 2003). Both Lev and Nissim (2004) and Mashruala et al. (2004) argue that implementing an accrual strategy is too costly for arbitrageurs. As a result, predictability remains observable. None of the above studies directly examines the accrual-fixation hypothesis. In fact, most of them accept the accrual-fixation hypothesis to be the cause of return predictability. Several studies attempt to examine the accrual-fixation hypothesis

9 8 indirectly by looking at the behavior of analysts and institutional investors (e.g. Bradshaw et al., 2001; Ali et al., 2000; Collins et al, 2003). These investors presumably are less subject to any processing biases implied by the accrual-fixation hypothesis. The studies seek to exploit cross-sectional variation in the presumed processing bias to generate predictions about the nature of the anomaly across groups of investors. The evidence is mixed. While Ali et al. (2000) do not find evidence consistent with naïve fixation, Collins et al. (2003) show that returns to an accrual strategy vary with the level of institutional (or sophisticated) investors in a manner consistent with naïve fixation. Bradshaw et al. (2001) also find evidence that analysts are subject to the same misprocessing bias as the market. In this paper I take a more direct and proactive approach in investigating the validity of the accrual-fixation hypothesis. In the following section, I generate testable empirical predictions to evaluate the accrual-fixation hypothesis based on its characteristics. 2.2 Properties of the accrual-fixation hypothesis Reversals. Several features of the accrual-fixation hypothesis are noteworthy. Under the accrual-fixation hypothesis, and stemming from the Mishkin tests results, market participants treat accruals and cash flows similarly ignoring the transitory nature of accruals (i.e. accruals tendency to reverse). As a result, firms with extreme high (low) accruals are over- (under-) priced and future abnormal returns are generated. Since abnormal returns appear in extreme accrual deciles and because abnormal returns under the accrual-fixation hypothesis are driven by accrual

10 9 reversals, I expect that firms in extreme deciles will exhibit reversal patterns similar to those of accruals. Thus, under this scenario accrual extremeness is expected to be transitory. Further, if aggressive accounting is related to accrual-fixation then, for a given firm, accruals are not likely to remain extreme in the following period because of disciplinary mechanisms such as auditor oversight. Thus, my first testable prediction is that under the accrual-fixation hypothesis, extreme accrual deciles are likely to be composed of different firms across periods. To test my first prediction, I examine the time-series properties of extreme accrual firms. I look at whether extreme accrual firms tend to leave the extreme accrual deciles in the following period as implied by the accrual-fixation hypothesis. Alternatively, a finding that these firms reside in extreme deciles in successive years would be inconsistent with the accrual-fixation hypothesis. I view such extreme accrual firms as being sticky (i.e. belong to an extreme decile, High or low, in two consecutive years). Symmetry. A second feature of the accrual-fixation hypothesis is its symmetry. Under the accrual-fixation hypothesis there is no difference in reversals between high and low accrual firms. Figure 1 in Sloan (1996) depicts reversal patterns in both extreme deciles. Also, the regression models in Sloan (1996) pool together all observations, including those in high and low deciles, implying that the hypothesized information misprocessing is similar in both extreme deciles. Because the hypothesized accrual misprocessing is similar in both groups, abnormal returns are expected to behave similarly across them. In both deciles, investors presumably do

11 10 not appropriately adjust for the differential persistence of accruals and cash flows. With respect to symmetry, I investigate the behavior of future abnormal returns across the low and high portfolios to detect any differences between them. In addition, I evaluate firm characteristics that may be different between high and low accrual firms. Overreaction. The accrual-fixation hypothesis is consistent with investors overreacting to accrual information. The return reversals that are experienced by extreme accrual firms are viewed as corrections to the initial overreaction. Several studies in the accounting literature perceive the accrual anomaly to be a result of overreaction. For example, Collins and Hribar (2000) say that the market appears to overreact to earnings that contain a large accrual component (emphasis in the original). Similarly, Desai et al. (2004) write: We posit that the (accrual anomaly) represents overreaction to past accounting data..(i)nvestors extrapolate past earnings into the future and are surprised when earnings announced subsequently are lower or higher than expected due to accrual reversals. Kothari (2001) wonders why does the market overreact to accruals in annual earnings (as in Sloan, 1996), but underreact to quarterly earnings information as seen from the post-earningsannouncement drift? In addition to the evidence of return predictability, Sloan (1996) also provides evidence about the timing of the reversals. He argues that investors correct their past mistakes at the time of future earnings realizations, i.e. around future earnings announcements.

12 11 Testing whether the accrual anomaly is consistent with overreaction is informative about the validity of the accrual-fixation hypothesis. The accrual-fixation hypothesis combined with Sloan s earnings announcements evidence predict the timing of the reversals. Therefore, I evaluate whether there is a correlation between the announcement returns around the revelation of extreme accruals (in year t) and the announcement returns around earnings announcements in the following year (t+1). The accrual-fixation hypothesis predicts that there will be a negative correlation between announcement returns in two successive years. 3. Data 3.1 Sample selection and variable measurement For my main analysis, I use a sample of NYSE/AMEX firms during the period. In some cases, where information from the statement of cash flows is required for the accrual calculation, I use the sample period. Also, for tests that require earnings announcement returns, I start my sample in 1974, the year in which an increasing number of firms have announcement dates available in COMPUSTAT. Accounting variables. Accounting variables are drawn from the COMPUSTAT primary, secondary and tertiary files as of First, I use differences between successive balance sheet accounts according to the following formula (COMPUSTAT item numbers in parentheses, stands for annual changes in the corresponding items):

13 12 Balance-sheet accruals: current assets(#4)- cash(#1)- current liabilities(#5)+ debt in current liabilities(#34)-depreciation(#14)) Second, following Bradshaw et al. (2001), I calculate accruals directly from the statement of cash flows, according to the following formula: Cash-flow-statement accruals: Earnings before extraordinary items from Cash-Flow Statement (#123) Cash flows from operations (#308). Finally, in some analyses I use four additional measures of accruals, as in Richardson et al. (2003). All of these measures are calculated from the balance sheet. Working capital accruals: current assets(#4)- cash(#1)- current liabilities(#5)+ debt in current liabilities(#34)) Investing accruals: total assets (#6)- current assets(#4) - Long-term Investments (#32)- { total liabilities(#181)- current liabilities(#5)-long term debt(#9)} Financing accruals: short-term investments(#193)+ Long-term Investments(#32)-{long term debt(#9)+ debt in current liabilities(#34)+preferred stock(#130)} Total accruals: The sum of the above working capital accruals, investing accruals and financing accruals. All accounting variables are deflated by average total assets (at time t and time t-1). Stock returns. I obtain stock return data from the Center for Research in Security Prices (CRSP) tapes as of I choose an equal-weighted, size-adjusted buy-and-hold abnormal return as my reference normal return benchmark. I use NYSE cutoffs to assign firms to 10 non-equal-sized portfolios at the end of fiscal year. Abnormal-returns calculation is done after compounding monthly returns of both firms raw returns and benchmark portfolio returns. Benchmark portfolios are

14 13 implicitly rebalanced every month and their composition may change after the ranking period. Accrual strategy. To apply the accrual strategy, at the end of each year firms are ranked into deciles based on the magnitude of their deflated accruals. Once the rankings are determined, abnormal returns for each accrual decile are calculated by averaging the abnormal returns of all firms in a particular accrual decile. Like other studies in the literature, I start the calculation period four full months after the end of the fiscal year. This assures that all the information in the financial statements is available to implement the strategy. 3.2 Descriptive statistics Table 1 reports medians and means of several variables of interest by deciles of accruals calculated from the balance sheet and deflated by average total assets. In general, the descriptive statistics are similar to those reported in previous studies. First, the negative correlation between accruals and cash flows is immediately apparent. Cash flows from operations decrease monotonically as we move from the low-accrual decile (median of 0.22) to the high-accrual decile (median of 0.02). Second, it is evident that firms with extreme accruals, those that occupy the top and bottom accrual deciles, are smaller in terms of market capitalization. A similar pattern is also present for total assets or total revenues (not reported). All these variables exhibit an inverted U-shaped pattern with respect to the accrual deciles. Moreover, the median size of low accrual firms is smaller than that of high-

15 14 accrual firms (p-value smaller than 1%), but the mean size of low accrual firms is larger than the mean size of high accrual firms. Third, high-accrual firms have lower book-to-market ratios. Starting with decile 8, there is a decrease in the median book-to-market ratio. It falls from around 0.55 to 0.51 in decile 8, to 0.48 in decile 9 and to 0.45 in the high accrual decile. The difference in median book-to-market ratios between the low and high accrual deciles is significant at conventional levels. Fourth, firm performance measured both in terms of contemporaneous stock performance and in terms of accounting rates of return and growth, is increasing monotonically with accruals. For example, the median raw return in the fiscal year for which accruals are measured increases from 3.2% in the low-accrual decile to 16.6% in the high-accrual decile. The median return on assets starts at 5.5% in the low accrual decile and increases to 10.5% in the high accrual decile. The median growth in sales is only 1.9% for the low-accrual decile, but rises to 23.4% for the highaccrual decile. This difference may be slightly misleading because of the larger preponderance of mergers in the high accrual deciles and of divestitures in the low accrual deciles. Controlling for these factors, however, the median sales growth remains different across the extreme accrual deciles (1.4% for the low decile versus 20.2% for the high decile). Fifth, compared to firms in the middle accrual deciles, forecasts of long-term growth are slightly larger for low accrual firms (13.5% vs. 12%), whereas they are much larger for high accrual firms (18% vs. 12%).

16 15 Finally, I report summary statistics on two common measure of firms degree of solvency. The first is Altman s Z (1968) that measures a company s financial strength. Altman s Z is negatively correlated with the probability of bankruptcy. The second measure is Ohlson s O developed by Ohlson (1980). It is positively correlated with the probability of bankruptcy. Table 1 reveals that the ex-ante probability of bankruptcy is monotonically decreasing as we move from the low to the high accrual deciles. That is, low accrual firms have a higher bankruptcy probability than high accrual firms. The median Z-score of low accrual firms is 2.7 and it reaches 3.9 in the high accrual decile. The O measure is for the low accrual firms and it reaches in the high accrual decile. All the patterns mentioned above (size, book-to-market, performance, bankruptcy risk) are present in the unreported summary statistics of non- NYSE/AMEX firms. 4. Reversals and symmetry: Results In section 4.1 I test whether extreme accrual firms are habitual extremes (i.e. sticky) or whether reversal patterns of firms across deciles are observed. In section 4.2 I investigate whether there is a relation between stickiness and future abnormal returns. Characteristics that distinguish sticky extreme accruals from other extreme accrual firms are examined in section 4.3.

17 Are extreme-accrual firms sticky? Recall from section 2.2 that the accrual-fixation hypothesis predicts that extreme accrual firms will tend to move away from the extreme accrual portfolios in the following year. Table 2 reports transition matrices containing frequencies of firms that occupy decile i in year t-1 and decile j in year t. I use several accrual measures in my analysis. First, panels A and B present the traditional definition of working capital accruals as in Sloan (1996) and many subsequent studies. I measure these working capital accruals using both the balance-sheet approach and the cash-flow-statement approach. Second, following Richardson et al. (2003), panel C uses a more comprehensive measure of total accruals that includes working capital accruals, investing accruals and financing accruals. The first component of panel C s total accruals, working capital accruals, is separately reported in panel D. Unlike panels A and B, working capital accruals in panel D do not include depreciation and amortization. The reason is that depreciation and amortization represent reversals of investing accruals and are not directly related to working capital accruals (see Richardson et al., 2003). The second and third components of total accruals, investing and financing accruals, are used in panels E and F. All the accruals in panels C through F are calculated using the balance-sheet approach. The results in Table 2 are for NYSE\AMEX firms. Similar (unreported) results apply for non-nyse\amex firms. Beginning with the traditional measure of working capital accruals, the first two panels of Table 2 indicate that there is a strong tendency for extreme accrual

18 17 firms in year t-1 to remain in the extreme accrual deciles in year t. For example, 26.4% of firms in the bottom balance-sheet accrual decile remain in that decile in the following year. Similarly, 27.1% of firms in the top accrual decile in year t-1 remain in that same decile in year t. Based on chi-squared tests, these frequencies are significantly higher than a frequency of 10%, which assumes independence across years and ignores the tendency of accruals to reverse. 2 Taking accrual reversals into account yields an expected frequency that is even lower than 10%. In fact, inventory accruals, which are part of working capital accruals, are likely to be the most transitory, reinforcing this lower expected frequency. However, the null hypothesis of expected frequency under accrual reversals is not readily and easily specified. As Hribar and Collins (2002) argue, measuring accruals using the cash-flow approach is preferable in the context of the accrual anomaly. Thus, it is worth noting that when accruals are measured using this approach, they exhibit an even stronger stickiness. Panel B shows that in the bottom (top) accrual deciles, 31.7% (30.7%) of firms remain in their respective accrual deciles in the subsequent year compared to 26.4% (27.1%) in panel A, wherein accruals are measured using the balance-sheet approach. Part of this increased frequency, compared to balance-sheet accruals, is because many firms are in the bottom (top) balance-sheet accrual decile as a result of divestitures (mergers) which are presumably transitory events (Zach, 2003). The results are similar in panels C through F. Total accruals in panel C appear sticky, with 21.1% of low accrual firms remaining in that decile in the following year, 2 P-values from chi-squared tests were lower than 0.01.

19 18 and an even higher frequency of 27.5% remained in the high accrual deciles in the following year. Working capital accruals in panel D display a similar pattern. The difference between these accruals and those in panels A and B is the exclusion of depreciation and amortization from the calculation, which makes accruals less sticky. In addition, there is a stronger tendency for working accruals to reverse to the opposite extreme than in panels A or B. Of the low accruals in year t-1, 15.1% of firms switch to the extreme high accrual decile in the following year. Similarly, 16.7% of high accrual firms switch to the extreme low accrual decile in year t. Investing accruals in panel E are even stickier than working capital accruals, with 24.2% and 27.2% of firms remaining in the extreme low and high accrual deciles, respectively, in the following year. Financing accruals, in panel F, are the only ones with modest stickiness. In addition they also exhibit some switching between extreme high and low deciles, similar to working capital accruals. In summary, Table 2 shows that extreme accrual firms, regardless of the type of accruals, are sticky. Recall that, under the accrual-fixation hypothesis, extreme accrual firms are likely to be transitory. Thus, the results reported in Table 2 are not consistent with this hypothesis. 4.2 Extreme accrual stickiness and abnormal returns Given the documented evidence of accrual stickiness in section 4.1, I now explore whether future abnormal returns to extreme accrual firms vary depending on

20 19 whether the extreme accrual firm in year t is sticky. In other words, does the path leading to extreme deciles matter in predicting future abnormal returns? The data for this analysis appear in Table 3. Panel A reports mean annual book-to-market and sizeadjusted abnormal returns. Panel B reports abnormal returns that are only adjusted for size. Abnormal returns are reported separately for firms stratified based on their accrual decile assignments in both year t (the year immediately preceding the computation of future abnormal returns) and year t-1. Firms that belong to extreme deciles are reported separately while firms that belong to the middle eight deciles are combined in the middle category. My analysis focuses on the traditional measures of working capital accruals, calculated using the cash-flow-statement approach and corresponding to panel B in Table 2. Similar (unreported) results emerge for working capital accruals calculated using the balance-sheet approach. High accruals. I first discuss firms with high accruals in year t. Past studies (e.g. Sloan, 1996) show that these firms earn negative future abnormal stock returns. Table 3, however, shows that the negative performance is limited to the set of sticky firms, whose accruals have been extremely high in both year t and year t-1. In the NYSE\AMEX sample, these firms generated an abnormal return of -8.8% compared to -4.2% of most other high accrual firms. Recall that these firms constitute about 30% of the high accrual decile. Results are similar when only size is used in calculating abnormal returns. In this case the returns to sticky high-accrual firms are - 7.6% compared to -1.7% for non-sticky firms.

21 20 Similar (unreported) results are obtained for the non-nyse\amex firms. Zach (2003) shows that the returns to high accrual firms in this subsample are much higher than in the NYSE\AMEX sample. Sticky accrual firms generate BM-sizeadjusted returns of -17.3% compared to -11.8% for non-sticky firms. Low accruals. In past studies, low accrual firms have been shown to generate positive future abnormal returns. Focusing on NYSE\AMEX firms, we see that sticky low accrual firms earn a return of 6.9% compared to 8.4% for non-sticky accrual firms. Thus, in the case of low accrual firms, stickiness does not seem to matter, at least in the sample of NYSE\AMEX firms. In the non-nyse\amex group the low sticky accruals earn 17.2% compared to -0.1% for the non-sticky firms. The difference is also present if returns are only size-adjusted (15% compared to -2.4%). Thus, in the case of low accruals, stickiness matters in the non-nyse\amex. Summary. Results in Table 3 show that in the NYSE\AMEX sample, firms belonging to the high accrual decile in two consecutive years earn lower returns then high-accrual firms that were not part of the extreme decile in the previous year. In the case of low accruals, this result holds only for non-nyse\amex firms. The sticky low-accrual firms earn higher returns than the non-sticky firms. Thus, I conclude that the accrual anomaly is more prominent for firms that belong to extreme accrual deciles in two consecutive years. In addition, Table 3 reports differences in the behavior of abnormal returns across exchanges.

22 21 Related to the evidence in Table 3, it is worth recalling the evidence in Sloan (1996) regarding future abnormal returns beyond year t+1. Sloan (1996) documents that abnormal returns to extreme accrual portfolios arise not only in the first year following portfolio formation but also in the second and, to a lesser extent, the third year. The evidence in this section suggests that the returns arising in the second year are attributable, to some extent, to sticky accrual firms. 4.3 Characteristics of sticky extreme-accrual firms I next evaluate whether sticky accruals, which I have shown are largely responsible for the returns to an accrual strategy, differ from other extreme accrual firms. I examine various characteristics of sticky accruals at two points in time. In section 4.3.1, I examine characteristics of sticky accruals at the end of year t, immediately preceding the initiation of an accrual strategy. This can help distinguish firm characteristics that are associated with future abnormal returns. In section 4.3.2, I examine the characteristics at the end of year t-1. This analysis sheds light on the determinants of sticky accruals Characteristics at time t Table 4 reports medians of several variables of interest, stratified by accrual deciles in both year t and year t-1. The variables are reported as of the end of year t, immediately preceding the calculation of future abnormal returns to an accrual strategy. I also report p-values of Wilcoxon differences-in-medians tests. These tests are conducted within the accrual assignments in year t. In the row of low accruals in year t, I report the p-value of the difference between sticky accruals (low-low) and

23 22 non-sticky accruals in the middle group (middle-low). 3 In the row of high accruals in year t, I report the p-value of the difference between sticky accruals (high-high) and non-sticky accruals in the middle group (high-middle). In the following discussion, I focus attention on NYSE\AMEX firms in the left part of Table 4. I mention the results for the non-nyse\amex sample whenever they differ. Beginning with market capitalization, there does not seem to be a statistically significant difference between sticky high or low accruals and their middle counterparts in the NYSE\AMEX sample (p-values of for low accruals and for high accruals). The right part of table 4 shows that non-nyse\amex sticky high (low) accruals are larger (smaller) than their middle counterparts. Statistically significant differences are apparent in the rest of the variables. Previous studies have shown that high accrual firms exhibit higher sales growth than low accrual firms. The results for my sample of firms confirm this prior result. However, Table 4 shows that among year t s high accrual firms, sales growth is the highest for sticky firms, 31.9%, compared to 22.3% for non-sticky firms. This pattern also holds, although with much lower growth rates, for the sticky low accrual sample where growth is 5.9% compared to 1.8% for the non-sticky low accruals. Table 4 also shows that analysts forecasts of future long-term growth are higher for both high and low sticky accruals. These are 20% for the sticky high 3 When discussing cells of accruals at t-1 and t, I first refer to year t-1 s accrual decile. For example, middle-low cell refers to middle accrual deciles in year t-1 that switched to the low accrual decile in year t.

24 23 accruals compared to 16% for non-sticky high accruals. Forecast growth for the low sticky accruals firms is 15%, compared to 12.5% in the non-sticky subsample. Another indication of the differences in growth prospects between sticky and non-sticky firms is evident in their lower book-to-market ratios. While sticky high accruals have a median BM ratio of 0.508, their middle counterparts BM ratio is Low sticky firms BM ratio is compare to of middle accruals. In Table 4, I also report abnormal accruals calculated using the modified Jones model. 4 The differences between sticky accruals and other accruals are only evident in the case of high accruals (0.116 vs ). These differences are not apparent for the sample of low accruals in the NYSE\AMEX sample. There is, however, weak evidence that abnormal accruals are different between the sticky low accruals and non-sticky low accruals in the non-nyse\amex sample (-0.14 compared to with a p-value of 0.085). Next, I report the Z-score from a bankruptcy prediction model developed by Altman (1968). A higher Z-score indicates a lower probability of bankruptcy. In general, high accrual firms have higher Z-scores. Within the high accrual firms, sticky firms have an even higher Z-score (4.14) than non-sticky firms (3.89). An opposite picture emerges in the low accrual deciles. Sticky low accruals have lower Z-scores (2.63) than non-sticky firms (3.00) indicating a higher risk of bankruptcy for these sticky firms. 4 Abnormal accruals using the Jones model yield similar results.

25 24 Summary. Table 4 shows that within the high accrual decile, sticky accruals firms are different from non-sticky accruals firms on several dimensions. First, firms with high accruals are growth firms in general, but sticky high accruals possess even higher past sales growth and expected future growth. Second, there is some evidence that abnormal accruals are higher for sticky accruals firms, compared to non-sticky accruals firms. This is consistent with the bias in calculating abnormal accruals being more prominent for firms with higher growth. Third, sticky-high accruals firms are less likely to go into bankruptcy, as evident by their higher Z-scores. On the other hand, in the case of low accruals, sticky-low accrual firms are more likely to go into bankruptcy based on their Z-scores. There is some weak evidence of higher past sales growth and future growth for sticky low accrual firms. Abnormal accruals do not differ between sticky and non-sticky accruals. This can be related to the lack of difference between the returns to an accrual strategy between low sticky and non-sticky accruals. Overall the evidence in this section suggests that there exist some underlying characteristics common to sticky high accrual firms (expected future growth) and sticky low accrual firms (bankruptcy risk). These characteristics are not likely to be transitory, contributing to stickiness. Also, the fact that these are different characteristics is not consistent with the symmetric story underlying the accrualfixation hypothesis. Recall from Table 3 that return predictability is also different between the low and high accrual deciles.

26 Characteristics at time t-1 Table 5 reports medians of several variables of interest, stratified by accrual deciles in both year t and year t-1. The variables are reported as of the end of year t-1. I also report p-values of Wilcoxon differences-in-medians tests. Unlike the tests in Table 4, these tests are conducted within the accrual assignments in year t-1. For the column of low accruals in year t-1, I report the p-value of the difference between the sticky accruals (low-low) and the non-sticky accruals in the middle group (lowmiddle). For the column of high accruals in year t-1, I report the p-value of the difference between the sticky accruals (high-high) and the non-sticky accruals in the middle group (middle-high). Beginning with high-accrual firms in year t-1, Table 5 shows that sticky firms had higher sales growth rates in year t-1 (32.3% vs. 23.6%), had higher growth forecasts (20% vs. 17%), but had much lower BM ratios (0.43 vs. 0.58). The characteristic that seems to be the determinant of stickiness firm is past and future growth. Results for the non-nyse-amex sample are similar. In the low-accrual group of the NYSE/AMEX sample there are no differences between sticky low accruals and non-sticky low accruals in their past sales growth, growth forecasts and BM ratios. The one characteristic that stands out and is different between sticky low accruals and non-sticky low accruals in both sides of Table 5 is bankruptcy risk as measured by the Z-scores. Sticky firms possess a median Z-score of 2.54 at time t-1, compared to a median Z-score of 3.08 in the non-sticky group.

27 26 I conclude that an important determinant of stickiness in the high-accrual group is past sales growth and forecasted future growth. In the low accrual group, stickiness is associated with higher bankruptcy risk. 4.4 Summary In this section, I find that extreme accrual firms are habitual extremes. That is, portfolios of extreme high (low) accruals are populated with firms that have been in the same portfolios in the previous year. About 25% of extreme accrual firms are sticky. The stickiness is observed across various measures of accruals, including working capital accruals (measured based on the balance-sheet or the statement of cash flows) and investing accruals. This evidence is inconsistent with the accrualfixation hypothesis. Abnormal returns to an accrual strategy are, to a large extent, driven by sticky firms. In NYSE\AMEX firms, sticky high accruals are drivers of negative future abnormal returns in the high accrual decile. In non-nyse\amex firms, sticky low accruals are drivers of positive future abnormal returns in the low accrual decile. Such differences are not consistent with the symmetry property under the accrual-fixation hypothesis. I also report evidence on certain characteristics of sticky firms. Sticky high accrual firms are associated with extreme past and future growth. On the other hand, sticky low accrual firms have a higher probability of bankruptcy. Consistent with these characteristics, untabulated results reveal that delisting frequency in low accrual deciles is higher than in high accrual deciles, especially in the non-nyse\amex

28 27 sample. Also, the reason for delistings in that sample is mainly due to poor performance. The differing characteristics between low and high accrual firms are not consistent with the symmetry property under the accrual-fixation hypothesis. 5. Accrual-fixation and overreaction: Results As discussed in section 2.2, the accrual-fixation hypothesis along with the empirical evidence of return predictability and their concentration around earnings announcements are consistent with investors initially overreacting to accrual information and subsequently correcting prices after realizing their mistakes. To test whether overreaction is descriptive of the evidence, I evaluate the relation between the returns around earnings announcements of extreme accruals and the returns around earnings announcements in the following year. The accrual fixation hypothesis predicts that this correlation is negative. I argue that the directional reaction of investors to extreme accruals at time t, proxied by the abnormal returns over a short-window around the earnings announcement, can assist in identifying firms in which an overreaction is more likely to have occurred. For example, take a high accrual firm that experiences a positive earnings announcement abnormal return. This implies that investors were positively surprised by the earnings news. Since this firm s earnings are likely to reverse, there is a higher chance that investors overreacted to the accrual information for that firm, compared to a high accrual firm that experienced a negative abnormal return around its earnings announcement. A similar logic, in an opposite direction, applies in the

29 28 case of a low accrual firm experiencing a negative earnings-announcement return. Thus, an important feature of my tests is exploiting the information in the market s initial reaction to the accrual signal. Since the evidence in the accrual anomaly literature suggests that investors make mistakes in period t and subsequently correct them around earnings announcements in period t+1, I make predictions about the relation between the earnings announcements returns in subsequent periods. The table below summarizes these predictions under the accrual fixation hypothesis and investors overreaction. The table partitions firms based on both the level of their accruals and the sign of their earnings announcement abnormal returns at time t. High Accruals Low accruals Positive announcement return at t Investors overreacting to accrual information Investors that are less likely to overreact to accrual information. Predicted relation between announcement returns in t and t+1 Predicted relation between announcement returns in t and t+1 Negative Positive or zero Negative announcement return at t Investors that are less likely to overreact to accrual information. Positive or zero Investors overreacting to accrual information Negative

30 29 The partitioning yields two groups of firms. The first group consists of: (i) high-accrual firms in which investors are positively surprised at earnings announcement; (ii) low-accrual firms in which investors are negatively surprised at earnings announcement. In this group, investors are more likely to overreact to earnings because the surprise, reflected by the initial stock price reaction to earnings information, is opposite to the direction that accruals predict future earnings will follow. According to the overreaction story, at least some of the surprises in this group are unwarranted. The second group is high-accrual firms in which investors are negatively surprised and low-accrual firms in which investors are positively surprised. Since in this group investors surprises are in accordance with how accruals are expected to reverse, I argue that investors in this group are less likely to have overreacted. In this group, investors do not seem to be misled by the extreme level of accruals at the time the earnings information is made public. Therefore, a negative correlation between earnings announcements in consecutive years is less likely. If anything, an underreaction is likely to occur in this sample. 5.1 Empirical tests I use short-window returns around earnings announcements of fourth-quarter earnings of year t to proxy for the direction of investors initial reaction to accruals. I choose the fourth-quarter returns because I am interested in investors initial reaction to the announcement of annual accruals. Also, I examine the returns around earnings announcements and not around the publication of 10-K s because, under the accrual-

31 30 fixation hypothesis, the behavior implied by market participants does not distinguish between cash flows and accruals. Thus, the reaction to the overall earnings number is sufficient in describing investors behavior. To evaluate investors reaction to future earnings announcements, I separately look at the returns around earnings announcements in the four quarters of year t+1. This not only allows me to assess the existence of an association between current and subsequent earnings-announcement returns but also helps me evaluate the timing of pricing corrections. To operationalize my tests, I regress returns around future earnings announcements on fourth-quarter announcement returns of year t as follows: RE + = α + β + ε (1) q 4 ( t 1) qre ( t ) Where R(E q t ) is the earnings announcement return around the q th quarter of year t, where q equals 1,2,3 or 4. Year t refers to the fiscal year at the end of which accruals are measured and ranked. Year t+1 refers to the year in which returns on an accrual strategy are realized. Under the accrual fixation hypothesis announcement returns in year t are expected to be negatively correlated with those in year t+1. These associations are expected to be stronger for firms with extreme accruals. 5.2 Results Summary Statistics. First, I provide some descriptive statistics in Table 6. The table reports statistics for the full sample, as well as the good news and bad news

32 31 samples. Good or bad news are determined based on the sign (positive or negative) of abnormal announcement returns of fourth-quarter earnings in year t. Several observations are noteworthy. It is evident that high and low accrual firms are almost evenly split between subsamples of good and bad news. In other words, high and low accrual firms are equally likely to generate good and bad news at the announcement of earnings. Also, the average announcement returns conditional on good (bad) news exhibits a(n) (inverted) u-shaped pattern with respect to accruals. Regression results. To test my predictions about the relation between announcement returns in year t and announcement returns in t+1 I run regression (1) and report the results in table 7. The model is run separately for each accrual quintile group as well as for good news and bad news firms. The type of news is determined by the sign of abnormal announcement returns of fourth-quarter earnings in year t. Each set of columns in table 7 reports the regression results for different independent variables, the announcement returns of quarters 1 through 4 in year t+1. The overreaction story, which is consistent with the accrual fixation hypothesis, predicts a negative sign for the β coefficients. Moreover, predicted signs are stronger for extreme accrual quintiles, depending on the type of news the firm had in time t. These are depicted by three minuses based on the predictions made in section 2. On the other hand, positive coefficients are more consistent with an opposite behavior of underreaction. In general, the results in table 7 are not consistent with overreaction. In fact, they are more supportive of an underreaction story. First, recall that high accrual