INVESTOR MISPERCEPTIONS OF BALANCE SHEET INFORMATION: NET OPERATING ASSETS AND THE SUSTAINABILITY OF FINANCIAL PERFORMANCE. David Hirshleifer*

Transcription

1 INVESTOR MISPERCEPTIONS OF BALANCE SHEET INFORMATION: NET OPERATING ASSETS AND THE SUSTAINABILITY OF FINANCIAL PERFORMANCE David Hirshleifer* Kewei Hou* Siew Hong Teoh* Yinglei Zhang* *Fisher College of Business, Ohio State University. Hirshleifer: Hou: Teoh: Zhang: September 2003 We thank the editor, S.P. Kothari; an anonymous referee; David Aboody, Jing Liu, Stephen Penman, Doug Schroeder, and Accounting colloquium participants at the Ohio State University for very helpful comments.

2 INVESTOR MISPERCEPTIONS OF BALANCE SHEET INFORMATION: NET OPERATING ASSETS AND THE SUSTAINABILITY OF FINANCIAL PERFORMANCE This paper examines whether market prices fully reflect the information contained in the level of net operating assets about the long-term sustainability of the firm s financial performance. In our sample, the ratio of net operating assets to beginning total assets (NOA) is a strong negative predictor for at least three years of future stock returns. Predictability is robust with respect to the use of a characteristic portfolio benchmark matching procedure, Fama-French 3 and 4-factor benchmarks, and Fama-MacBeth crosssectional monthly regressions using an extensive set of controls for pricing anomalies in the finance literature (market value, the book-to-market ratio, and one-month, 12-month, and 3-year past stock returns). Furthermore, the sustainability effect remains after controlling for current and past period operating accruals and the most recent change in net operating assets, and provides higher and statistically more significant hedge profits than the operating accruals anomaly. Controlling for NOA weakens but does not entirely subsume the operating accruals anomaly. A Fama-MacBeth Mishkin test suggests that investors fail to take into account the adverse information contained in NOA about future earnings.

3 1. Introduction Previous research offers evidence that investors are overoptimistic about the implications of firm s operating accruals for subsequent earnings performance (see Sloan (1996) and an extensive subsequent literature 1 ). Operating accruals are, of course, a flow measure. Two interesting questions raised by this finding are whether investor overoptimism extends to nonoperating as well as operating accruals; and whether investors misinterpret stock information from the balance sheet as well as flow measures in assessing whether the firm s performance is sustainable. With regard to the first question about the effects of different types of accruals, Richardson, Sloan, Soliman and Tuna (2002) find that investors misinterpret the information contained in non-operating as well as operating accruals, leading to predictability of stock returns. Fairfield, Whisenant and Yohn (2003) suggest that the operating accruals anomaly may be part of a general anomaly in which investors misinterpret change in net operating assets as a whole rather than operating accruals specifically. They find that the two components of the annual change in net operating assets (operating accruals and non-operating accruals) are equally strong negative predictors of one-year-ahead stock returns. In contrast, Richardson et al (2002) decompose accruals into finer components that vary in reliability as indicators of future earnings, and conclude that investors fail to adjust for the fact that less reliable accrual components lead to less 1 Some of the papers extending Sloan (1996) consider alternative measures of accruals and components of accruals; see e.g., Chan, Chen, Jegadeesh, and Lakonishok (2001), Collins and Hribar (2002), Richardson, Sloan, Soliman, and Tuna (2002), and Thomas and Zhang (2002). Another branch of this literature provides evidence that managers take advantage of investor naivete about accruals in order to manage investor perceptions and auditor, and analyst behavior; see, e.g., Teoh, Welch, and Wong (1998a, b), Rangan (1998), Bradshaw, Richardson, and Sloan (2000), Xie (2001), Ali, Hwang, and Trombley (2002), and Teoh and Wong (2002). In this paper we do not address whether firms deliberately manage investor perceptions, merely whether investors interpret available accounting information optimally. 1

4 earnings persistence. With regard to the second issue on investor perceptions of balance-sheet/stock information, Barton and Simko (2003) suggest that the level of net operating assets act to constrain how much firms can manage earnings upward. They provide intriguing evidence from the time period that net operating assets indicate the extent to which the firm is likely to meet analyst forecasts. Barton and Simko s perspective suggests one reason why the level of net operating assets may contain information about future firm accounting performance. It is also possible that low net operating assets constrain a firm s ability to manage earnings downward (in order to take a big bath or create `rainy day reserves; see DeFond (2003)). A different reason why the level of net operating assets can be informative about the sustainability of firm performance, which does not rely on earnings management, is that there may be a tendency for exceptional past performance to revert towards the mean. On the upside, such reversion could be a consequence of diminishing marginal returns on new investments (see Fairfield, Whisenant and Yohn (2003)); alternatively, superior profits may attract profit-reducing entry by competitors, and low profits may encourage exit and discourage competition. More generally, the regression phenomenon in statistics suggests that either exceptionally high or exceptionally low performance in one year will tend to revert toward the mean in subsequent years. Furthermore, if a firm s operating assets have been increasing over a period of years through investment or operating accruals without a commensurate rise in net cash flows, it may be harder to sustain that apparent profitability in the future. In his textbook on valuation, Penman (2002) decomposes earnings into the sum of change in the firm s net operating assets and free cash flow. Thus, net operating assets reflects a cumulative stock of accruals over time that are either 2

5 recognized revenues that are not yet received as cash (in excess of investment needs), or are costs incurred but not yet recognized as expenses. Thus, the unsustainability of growth of high net operating assets firms is broadly in the spirit of the perspective for flows of Sloan (1996) that operating accruals are a less persistent earnings component than are cash flows. However, net operating assets reflect the cumulative effects not only of operating accruals, but of nonoperating accruals as well. To the extent that net operating assets are associated with past accruals, and if investors fail to process properly the information about earnings persistence contained in past accruals, high net operating assets can be associated with low future stock returns. If investors are functionally fixated on earnings (Hand (1990)), possibly owing to limited attention and processing power (see, e.g., Hirshleifer and Teoh (forthcoming)), then they may form their forecasts of future earnings based upon past earnings, but not based upon the breakdown of earnings between accruals and cash flow components. If so, then they will fail to adjust fully for the adverse implications of accruals for future profitability. The extent of the misvaluation should therefore depend on some stock measure of cumulative accruals in previous years. An alternative argument is that imperfectly rational investors may overextrapolate growth in financial flow variables such as sales or earnings. Investors will then overvalue firms with high net operating assets to the extent that such firms have experienced growth in financial flow measures in preceding years. For example, if investors fail to discount fully for the regression phenomenon in statistics, possibly owing to use of the representativeness heuristic (Kahneman and Tversky (1972)), they will tend to overextrapolate past earnings or sales growth (see, e.g., Lakonishok, Shleifer and Vishny (1994)). Thus, arguments based upon either limited attention or upon representativeness imply a 3

6 negative relationship between net operating assets and subsequent future abnormal stock returns. We develop these hypotheses in greater depth in Section 2. In this paper we test whether market prices fully reflect the information contained in the level of net operating assets of the firm about the long-run sustainability of firms financial performance. We document a new stock return anomaly relating to the level of net operating assets scaled by beginning total assets (hereafter referred to as NOA). We find that NOA is a strong and robust negative predictor of future stock returns for at least three years after balance sheet information is released. 2 A trading strategy based upon buying the lowest NOA decile and selling short the highest NOA decile is profitable in 36 out of the 38 years in the sample, and averages equally-weighted monthly abnormal returns of 1.2 %, 0.8% and 0.5% in the one, two and three years respectively after the release of the balance sheet information. In these strategies, for each month abnormal returns for each firm are obtained by subtracting its benchmark portfolio that has been matched for firm size, book-to-market, and past 12-month return. Then, either equal weighted or valueweighted mean returns are calculated for each NOA/Accruals decile. The sustainability effect also remains strong using a Fama-MacBeth-like cross-sectional methodology that includes the above controls as well as past one-month and 12-month returns. The coefficient on NOA is highly statistically significant, indicating that the sustainability effect is independent of three well-known financial anomalies, the monthly contrarian effect (Jegadeesh (1990)), the momentum effect (Jegadeesh and Titman 1993) and the long-run winner/loser effect (DeBondt 2 A related literature considers financial ratios from the balance sheet to predict future returns; see e.g. Ou and Penman (1989, 1990), Holthausen and Larker (1992), Lev and Thiagarajan (1993), Abarbanell and Bushee (1997), and Piotroski (2000). A distinction in this paper is that we use a simple aggregate balance sheet measure, net operating assets, without imposing a particular weighting scheme upon various financial ratios from the balance sheet. 4

7 and Thaler (1985)). Also, since book-to-market and past returns are measures of past and prospective growth, these controls indicate that the findings are not merely a risk premium effect associated with the firm s growth rate. The evidence from the negative relationship between NOA and subsequent returns suggests that investors do not optimally use the information contained in NOA to assess the sustainability of performance. A Mishkin test along the lines of Sloan (1996) using NOA as an additional forecasting variable is similarly consistent with investor overoptimism about the earnings of high NOA firms. To examine whether the sustainability effect is incremental to Sloan s (1996) operating accruals effect, we first verify the presence of the operating accruals effect through the year We apply the same battery of test procedures and controls and used for the sustainability effect, which together provide a stronger set of robustness checks than exist in the current literature on the operating accruals effect. We verify that the operating accruals effect remains strong in the 1990s, lasts for 2 to 3 years of subsequent returns, and is robust with respect to well-known financial market anomalies. We find that the monthly abnormal return spread between low and high operating accruals deciles is 0.69% per month, as compared with 1.23% per month for NOA. Thus, the NOA spread is more than 78% larger in absolute value than the operating accruals spread in year t+1, a differential that grows to 152% in year t+3. In the Fama-MacBeth cross-sectional regressions that include both the operating accruals and NOA variables in addition to the previously mentioned controls, we find that both operating accruals and NOA remain highly statistically significant. Thus, neither anomaly subsumes the other in this analysis, but the inclusion of NOA weakens the accruals effect more than including accruals weakens the NOA effect. In Mishkin tests that include both NOA and accruals, investor 5

8 misperceptions about NOA are more severe than the misperceptions pertaining to operating accruals, which are insignificant. Using the overall sample, NOA remains significant in Fama-MacBeth regressions through year t+3, whereas operating accruals are no longer significant in year t+3. The more persistent predictive power of NOA may derive partly from the persistence of NOA itself, and partly because the NOA effect is stronger even for one year ahead returns. Since the operating accruals anomaly persists for up to 2-3 years, it is useful to verify whether the sustainability effect is incremental to the cumulative effect of past operating accruals. When we regress returns on standard controls, NOA in the most recent fiscal year, and the sum of the past three years of operating accruals, NOA remains strongly significant, more so than the summed accruals variable. This indicates that NOA provides information beyond the cumulative effect of past operating accruals about investor misperceptions of future performance. Fairfield, Whisenant and Yohn (2003) document that the change in NOA predicts oneyear ahead returns. We therefore also examine the extent to which the sustainability effect derives from the level versus the most recent change in NOA. We regress monthly returns in year t+1 on standard controls, NOA, operating accruals, and the most recent change in net operating assets scaled by beginning total assets. We find that only the operating accruals and NOA variables are statistically significant. Thus, even after controlling for both current-period operating and non-operating accruals, investors do not make appropriate use of the information in the level of NOA for evaluating the long-run sustainability of performance. In general, in multiple regressions on correlated variables, if the linearity assumption is misspecified, one of the variables can spuriously gain explanatory power at the expense of the other. To test more carefully for whether or not operating accruals and the sustainability effect 6

9 are distinct anomalies, we perform two-way sequential sorts, and examine the abnormal returns relative to benchmark portfolio returns that match for size, book-to-market ratio and past momentum. Controlling for the level of NOA does not completely eliminate the accruals effect, but the return differences between high and low accruals quintiles are often insignificant, and seldom exceed marginal significance. Furthermore, the accruals effect is not monotonic across accruals quintiles. Controlling for operating accruals, we find that the return spreads between low NOA and high NOA firms are positive and highly significant for all operating accrual quintiles, so the sustainability effect is not subsumed by the operating accruals anomaly. Furthermore, for each accrual quintile, the NOA effect is monotonic across the NOA quintiles. We also perform a combination quintile/regression method that is able to include more controls than the two-way sort, and which focuses on the sensitivity of returns to NOA or accruals, rather than the combined effect of sensitivity and variability of the characteristic. In separate Fama-MacBeth accruals regressions for each Accruals quintile, we find that the NOA effect is very strongly significant in all of the Accrual quintiles except for the largest. In contrast, the operating accruals effect is significant in three of the five NOA quintiles. Finally, the relative dominance of the NOA effect over the accrual effect is particularly marked in the last 5 years. We find that the operating accruals effect was exceptionally strong in calendar year 1999, a year during which there was a strong run-up in prices of stocks, especially those traded on NASDAQ. The operating accruals effect is several times as strong during 1999 as the average effect in all other years, but is minimal or reversed from The sustainability effect is also strongest in 1999, and is strong in 1998 and 2000 as well. Both the operating accruals and sustainability effects are robust to the exclusion of

10 The remainder of the paper is structured as follows. Section 2 develops the hypotheses. Section 3 describes the sample selection and variable measurement. Section 4 presents evidence on the robustness of the accruals anomaly to a wider set of controls and test methods and to a more recent period up to year Section 5 examines the predictive power of NOA to establish that sustainability effect is not subsumed by the operating accruals effect, the latest change in NOA, and the sum of past operating accruals. Section 6 concludes the paper. 2. Hypotheses We consider two alternative hypotheses. The first, the efficient markets hypothesis, predicts zero abnormal returns subsequent to either high or low levels of NOA. It also predicts that in a Mishkin test, the forecasts of future earnings implicit in stock market prices are unbiased. The alternative, the sustainability hypothesis, rests on the premise that NOA is a cumulative measure of past and recent strong financial performance. Thus, high NOA is associated with high past growth in earnings or other performance flow variable. A further premise is that high NOA firms will have difficulty maintaining the high rate of growth they have achieved in the past. This difficulty may simply arise from the regression phenomenon; firms with unusually high growth should tend to revert to a lower long-term mean. Alternatively, high NOA firms may have achieved their strong performance in part through unusually high recent accruals. Since operating accruals are less persistent than cash flows (Sloan 1996), high recent accruals will also make it hard to sustain the same high rate of earnings growth. These premises of the sustainability hypothesis, that high NOA is associated with high past growth in earnings, and that the rate of growth is slower after the NOA conditioning date 8

11 than before, are directly testable. The sustainability hypothesis further posits that investors do not fully take into account the unsustainability of high growth after high NOA, or low growth after low NOA. Two alternative psychological factors can motivate this proposition: limited attention, and the representativeness heuristic. Previous literature has proposed each of these factors as possible explanations for other capital markets anomalies. 3 Owing to limited attention and processing power, people tend to focus on a few salient stimuli to the exclusion of others (see e.g., Fiske and Taylor (1991)). A reliance by investors upon earnings rather than analyzing each of its separate components may be a case of this. This involves a kind of functional fixation upon earnings. Limits on processing power may be a source of the naïve reliance of investors in the experimental laboratory on simple categories (such as earnings), as discussed in the survey of Libby, Bloomfield, and Nelson (forthcoming). Thus, limited attention suggests that investors may not distinguish sufficiently between more persistent and less persistent components of recent financial performance. If high operating accruals tend to indicate a lack of persistence of earnings, then the failure of investors to condition separately on accruals and cash flows will cause them to overestimate the future earnings prospects of high-accrual firms. Suppose that higher NOA is associated with cumulative higher past operating accruals. Thus, inattentive investors will tend to overvalue high NOA firms and undervalue low NOA firms. In the long run, when this misvaluation is corrected, high NOA firms are on average predicted to earn negative abnormal stock returns and 3 Limited attention and processing power has been offered as a possible explanation for the operating accruals anomaly, and abnormal stock performance associated with pro forma earnings announcements and with disclosed but unreported employee option compensation (Hirshleifer and Teoh, forthcoming). The representativeness heuristic has been proposed as an explanation for long-run stock return reversals, and for long-term stock performance subsequent to upward trends in sales growth or earnings (Lakonishok, Shleifer and Vishny (1994)). 9

12 low NOA firms positive abnormal returns. Thus, there is a negative relation between NOA and subsequent returns. More generally, NOA contains information about future expenses and future revenues associated with past cash transactions. For example, operating assets (e.g. fixed assets) contain costs the firm has incurred that are waiting to be expensed (e.g. as depreciation) in the income statement in future years. Similarly, operating liabilities contain cash inflows (e.g. deferred revenues) that are waiting to be recognized as revenues on the income statement only in future years. These accretions tend to reduce the future earnings of firms with high NOA and increase the future earnings of firms with low NOA. Furthermore, NOA also contains information about revenues (e.g. accounts receivables) and expenses (accrued liabilities) that are already recognized in the income statement but are awaiting fulfillment of the associated cash transactions. To the extent that the future cash fulfillments fall short of recognized amounts, NOA can contain information about future reversals in earnings. For example, a high NOA firm resulting from unusually high net accounts receivables will experience a decline in future earnings when the excessive past recognized revenues are reversed. The unusual high past recognized revenues may be the result of earnings management, but earnings management is not necessary for this interpretation of the sustainability hypothesis. In sum, investors with limited attention may sometimes fail to take into account the above implications of NOA for future earnings. Such investors will tend to overestimate the earnings prospects of high NOA firms, and underestimate the prospects of low NOA firms. Thus, high NOA will predict lower future abnormal stock returns. A second pathway to the same prediction relies on the representativeness heuristic (see 10

13 Kahneman and Tversky (1972)). In the representative heuristic, individuals recognize a pattern in the data that resembles, or is `representative of, the pattern that would be expected to occur in some underlying state of the world. For example, an investor may recognize a pattern of growing earnings which resembles the circumstance that the firm is a nascent industry leader--- the next Microsoft. According to the representative heuristic, investors overweight the probability of the state of the world that the pattern resembles. To the extent that firms with high NOA have had growth in earnings or other measures of financial performance in preceding years, imperfectly rational investors may overextrapolate such growth, causing overvaluation. For example, the representativeness heuristic implies that investors will fail to discount fully for the regression phenomenon in statistics. Experimental evidence from the psychology of judgment confirms that individuals do not adequately adjust for the regression phenomenon; see, e.g., Gilovich (1991). This suggests that investors will tend to be surprised by the subsequent mean reversion of firms whose earnings have recently been growing or shrinking rapidly. Thus, more favorable trends in earnings or other financial variables may be followed by abnormally lower returns. 4 However, in this paper we condition upon NOA, which acts as only an indirect conditioning upon past financial trends. Even if (as turns out to be the case) conditioning on high NOA 4 An ongoing branch of the capital markets literature explores whether investors overextrapolate long-term trends in sales or earnings growth. If they do, subsequent to such trends there should be a correction, leading to abnormal stock returns. Lakonishok, Shleifer and Vishny (1994) report that long term sales growth is a negative predictor of stock returns. DeChow and Sloan (1997) find that this effect is not monotonic, and provide evidence suggesting that it does not derive from naïve extrapolation of trends in sales or earnings. Daniel and Titman (2003) provide evidence that the sales growth effect derives from new issues rather than from growth in either sales or from other `tangible information. Lee and Swaminathan (2002), however, conclude that long-term earnings growth predicts negative subsequent returns. Chan, Frankel and Kothari (2003) find that repeated sequences of increases or decreases in earnings do not lead to significant subsequent abnormal returns. Bloomfield and Hales (2002) find that in the experimental laboratory traders do overextrapolate sequences of earnings increases. 11

14 implies that on average past earnings growth was high, this does not mean that subsequent earnings growth will merely regress to the mean. NOA may also contain information about how persistent past growth in earnings is likely to be. Thus, if investors are naïve extrapolators, the relation between NOA and returns can potentially be stronger than the relation between past earnings trends and returns. (This further argument leans to some extent upon limited attention as well.) In summary, arguments based upon either limited attention, representativeness or both imply a negative relationship between NOA and subsequent future abnormal stock returns. 3. Sample Selection, Variable Measurement, and Data Description Starting with all NYSE-AMEX and NASDAQ firms in the intersection of the most recent (2002) COMPUSTAT and CRSP tapes, the sample period spans a maximum of 462 months from July 1964 through December All observations are required to have sufficient financial data to compute accruals, net operating assets, firm size, book-to-market ratios, and 12-month return momentum. An initial sample of 1,605,193 firm-month observations is available for the Fama-MacBeth monthly cross-sectional regressions and the characteristics portfolio-matching analyses. The different test methods further impose varying restrictions depending on the controls such as past momentum and contrarian returns and the horizon of future returns. In the Mishkin market efficiency tests, we impose a minimum four-month gap between the fiscal year end and the start of the return cumulation. Since the returns data ends in December 2002, for these tests we use annual data from fiscal year 1965 through fiscal year We have an initial 138,908 firm-year observations with sufficient returns and financial data during this period. The sample is reduced by the requirement that observations have one-year ahead earnings from COMPUSTAT for the forecasting equation in the Mishkin tests to 136,151 12

15 observations. After deleting the smallest and largest 0.5% of the observations on the financial and returns variables to avoid extreme outlier effects, the final sample contains 128,519 firmyear observations Measurement of Operating Accruals and NOA As in previous studies using operating accruals in periods prior to the enactment of SFAS #95 in 1988, we calculate operating accruals using the indirect balance sheet approach for all fiscal years in our sample. This ensures comparability of the measure through time, as well as comparability with past studies. The variable Accruals is calculated as the change in non-cash current assets less the change in current liabilities excluding the change in short-term debt and the change in taxes payable minus depreciation and amortization expense, deflated by lagged total assets, Accruals t = [( Current Assets t - Cash t )-( Current Liabilities t - Short-term Debt t - Taxes Payable t ) - Depreciation and Amortization Expense t ]/Total Assets t-1. (1) See Table 1 for the associated Compustat item numbers. Scaled net operating assets (NOA) are calculated as the difference between operating assets and operating liabilities, scaled by lagged total assets, where Operating Assets t = Total Assets t - Cash and Short-Term Investment t (2) Operating Liabilities t = Total Assets t - Short-Term Debt t - Long-Term Debt t 5 The estimation of the annual nonlinear Mishkin system is sensitive to extreme outliers in three of the 36 years in the sample period we examine. However, trimming extreme values can induce bias in tests of market efficiency (see Kothari, Sabino, and Zach, (2002)). Since we perform separate Fama-MacBeth tests where no trimming is done, our inferences about the predictability of long-run returns do not rely on trimming. The additional insight from the Mishkin tests concerns the extent to which return predictability derives from investor errors in forecasting future earnings from accruals or NOA. When we trim at 0.25% level instead of 0.5% level in the Mishkin test in Table 7, the results are similar. 13

16 - Minority Interest t - Preferred Stock t - Common Equity t. (3) When calculating operating accruals and net operating assets, if short-term debt, taxes payable, long-term debt, minority interest, or preferred stock has missing values, we treat these values as zeroes to avoid unnecessary loss of observations. Instead of scaling net operating assets by same period sales as in Barton and Simko (2003), we scale by lagged total assets (as in Fairfield et al., 2003) to achieve a comparison with the operating accrual anomaly results, and because sales is a flow variable whereas both NOA and total assets are stock variables. The change in NOA for year t is calculated as the annual change in net operating assets, scaled by lagged total assets. Finally, the further firm performance variables, Earnings and Cash Flows, are defined respectively as income from continuing operations (Compustat#178)/lagged total assets, and as Earnings - Accruals. 3.2 Measurement of Asset Pricing Control Variables Following the recommendation of Daniel, Grinblatt, Titman and Wermers (1997), we use the characteristics approach for the asset pricing control variables in predicting returns. Size is the market value of common equity (in millions of dollars) measured as the closing price at fiscal year end multiplied by the number of common shares outstanding. The book-to-market ratio is the book value of common equity divided by the market value of common equity, both measured at fiscal year end. In addition to these controls, we also include controls for one month-reversal, 12-month momentum, and three-year reversal, all measured relative to the test month t of returns. Ret(-1:- 1) is the return on the stock in month t-1. Ret(-12:-2) is the cumulative return between month t- 12 and month t-2. Finally, Ret(-36:-13) is the cumulative return between month t-36 and month t-12. Thus, the return control variables are updated each month as the Fama-MacBeth cross- 14

17 sectional regressions roll forward in time. The NOA, Accruals, Size and Book-to-market variables, however, are only updated every 12 months. Beyond these controls, we also report results after additional adjustments for the CAPM, the Fama-French 3 factor model, and a 4- factor model which adds a momentum factor. 3.3 Summary Statistics of Data Characteristics Table 1 describes the mean and median values for selected characteristics of the sample by NOA deciles. Firms are ranked annually by NOA and sorted into ten portfolios. Net operating assets vary from about 28% of lagged total assets in the lowest NOA decile to 140% in the highest NOA decile. This suggests that high NOA firms are likely to have experienced recent very rapid growth. In later tests, we separate NOA into asset growth components from other components. Turning to the financial performance, Earnings increases monotonically from a mean of 0.3% for NOA Decile 1 to a mean of 13.3% for NOA decile 10. Low NOA firms have experienced recent poor earnings performance while high NOA firms have experienced recent good earnings performance. This difference in performance seems to be driven primarily by large differences in Accruals across extreme NOA deciles. Accruals also increase monotonically across NOA deciles; they are a large negative 7% for NOA decile 1 and a large positive 10.5% for NOA decile 10. Operating Cash Flows do not vary monotonically across deciles. NOA decile 10, however, has significantly lower Cash Flows than all other deciles. NOA decile 1 s Cash Flows are similar to those of NOA decile 8 and 9, and are slightly lower than the Cash Flows in deciles 2 through 7, which are quite similar to each other. The high mean Earnings for NOA decile 10 despite its extreme low level of Cash Flows is consistent with the fact that the Accruals are extremely high in NOA decile 10. Similarly, the 15

18 extreme negative accruals for NOA decile 1 contribute to the portfolio s low Earnings despite its moderate level of Cash Flows. Thus, the extreme NOA deciles also have extreme Accruals. Our hypothesis concerns investors failing to attend to accruals, broadly construed, or to mistakenly extrapolating recent performance trends. Table 1 reports short-term trends in Earnings in relation to NOA. These are the current period change in Earnings and the next period s change in Earnings. NOA decile 1 experienced the worst current decline in Earnings, and amongst the highest turnaround in Earnings in the next period. The reverse holds for NOA decile 10, which previously did well and subsequently does poorly. These results suggest that the general pattern in earnings performance relative to NOA rank involves crossover. Thus, if investors ignore the fact that NOA provides information about reversals in earnings growth, NOA will predict future abnormal returns as well as changes in earnings. Turning next to stock market characteristics, Table 1 indicates that extreme (both high and low) NOA firms have the smallest size measured by either book value of equity or market value of equity; the lowest book-to-market ratios; and the highest betas. Thus, the extreme deciles seem to be small, possibly high growth or are overvalued, and risky firms. It is therefore essential to perform careful controls for risk in measuring abnormal returns in our tests. Panel C is based on two alternative decompositions of NOA. In the first decomposition, we split NOA into three parts based on the equation: NOA t = Total Assets t /Total Assets t-1 financial assets t /Total Assets t-1 operating liabilities t /Total Assets t-1. (4) We therefore define three pieces that enter positively into NOA: Growth, defined as total assets divided by lagged total assets; Fin_Assets, defined as (-1) * financial assets (=#1 cash and 16

19 short term investment) divided by lagged total assets; and Op_Liability, defined as (-1)* operating liabilities. In the second decomposition, we split NOA based on the equation: NOA t = equity t /Total Assets t-1 financial assets t-1 /Total Assets t-1 + financial liabilities t /Total Assets t-1. (5) We define Equity as book value of equity divided by lagged total assets; and Fin_liability, as financial liabilities (# 34 short term debt+ # 9 long term debt) divided by lagged total assets. For the first decomposition, we find that except for NOA decile 1, Growth is increasing with NOA, with an especially sharp increase in NOA decile 10. Fin_Assets is monotonicially decreasing in NOA, except for NOA deciles 9 and 10. Op_Liability is monotonically decreasing in NOA, except that it rises in NOA deciles 9 and sharply in NOA decile 10, leading to a U-shaped pattern. It appears that all three components contribute substantially to variations in NOA, with an especially strong contribution coming from Growth decile 10. For the second decomposition, we see that equity is U-shaped in NOA deciles, with a marked increase in NOA decile 10. Equity is relatively flat except in deciles 9 and 10, so we would expect any effect of equity to be relatively concentrated the top quintile. Fin_liability is monotonically increasing in the NOA decile. Put Table 1 about here. Table 2 reports the correlations between NOA, the variable of interest, and the performance measures and firm characteristics. NOA is persistent; the correlation between NOA and next period NOA is positive and significant. Confirming the patterns of association between NOA and Accruals in Table 1, NOA is significantly positively correlated with Accruals. Therefore we control for Accruals in all our tests. Section 5 provides further tests to distinguish 17

20 the effects on returns of NOA versus Accruals. Also consistent with Table 1 findings, the Spearman correlation indicates that NOA is positively correlated with Earnings, and current period change in Earnings, and is negatively associated with Cash Flows and next period change in earnings. Because of outliers, the Pearson and Spearman correlations are of the opposite sign for NOA with earnings, and with current period change in earnings. After trimming the extremes at 0.5% the Pearson and Spearman correlations are similar. While Table 1 shows similar characteristics in terms of size, beta, and book-to-market for extreme levels of NOA relative to the middle deciles, the correlations indicate that NOA is negatively correlated with beta and positively correlated with firm size. The correlation with book-to-market is positive for the Spearman and negative for the Pearson tests. Put Table 2 about here. 3.4 Industry Distribution Across NOA Deciles Table 3 reports the industry distribution of our sample across NOA deciles pooled across all sample years. Each decile is widely represented by all the two-digit SIC codes. Following a standard method in the literature, industry (four digit SIC) codes are grouped into fourteen industry groups. Panel A reports the percentage of firms in each industry group for each NOA decile. Comparing across NOA deciles, the extreme NOA deciles (1 and 10) have a relatively lower presence in the Food, Textile, Chemicals, and Durables industry groups. The extreme NOA deciles also have a higher presence in the Mining and Construction, Transportation, and Computers industry groups. In addition, NOA decile 1 has a relatively high presence in the Pharmaceuticals and Financials groups, and a relatively lower presence in the Extractive, Utilities, Retail, and Services groups. NOA decile 10 has a relatively higher presence in the 18

21 Extractive and Utilities industry groups. Panel B reports the percentage of firms in each NOA decile within each industry group. Looking across NOA deciles, the extreme NOA deciles (1 and 10) have a relatively larger presence in Mining and Construction, Computers, and Financials industry groups. Low NOA deciles additionally have a larger presence among Pharmaceuticals, and Transportation, and high NOA deciles have a larger presence among Agriculture, Extractive, and Utilities industry groups. Put Table 3 about here. 3.5 Time Trends in Earnings, Accruals, Cash Flows and Returns in NOA Deciles Figure 1 describes the time series means of Earnings, Accruals, Cash Flows and annual raw buy and hold stock returns in relation to the NOA decile ranking. High NOA firms hit a peak and low NOA firms hit a trough of Earnings and Accruals levels at the conditioning year. Thus, potentially consistent with the limited attention account discussed in Section 2, higher NOA is associated with higher levels of accruals, and with lower cash flows at the conditioning date. Accruals are so high that these firms report high earnings despite their large average recent drop in cash flows. If investors fail to discount for these facts, high NOA firms may be overvalued. The upward spike in Accruals at the conditioning date for high NOA firms is quite sharp. So if operating accruals were the only source of investor misperceptions at the conditioning date, we would not expect the poor performance of high NOA firms to continue much longer than the poor performance of high operating accruals firms in general. On the other hand, as suggested in the hypothesis section, the misperceptions that limited attention can bring about with respect to NOA go beyond just operating accruals. 19

22 Potentially consistent with the representativeness heuristic/extrapolation account of Section 2, higher NOA is associated with upward trending Earnings over the previous several years. Not only is this upward trend not continued at the same rate, after the conditioning year it switches to a downward trend in Earnings. Thus, if investors naively extrapolate past earnings trends at the conditioning year, they are in for a rude surprise. In general, behavioral accounts of over-extrapolation of earnings or sales growth trends involve a failure to recognize the regression phenomenon, which should cause the growth rate to revert to a normal long-run mean. In this case, however, earnings growth does not just revert to a normal, slower rate, it turns sharply negative. This suggests that we may be able to pick up the tracks of any investor overextrapolation much more easily by conditioning on NOA than by conditioning directly on earnings trends. Average Earnings is uniformly higher for high NOA firms than for low NOA firms, which reflects the glory of their past. As a result, even though high NOA predicts a sharp drop in earnings, cross-sectionally high NOA does not necessarily predict lower future Earnings across firms. This is an empirical question that depends on the balance between the time-series and the cross-sectional effect. The annual returns of high NOA versus low NOA firms display a dramatic cross-over pattern at the event year. High NOA firms earn higher returns than low NOA firms before the event year, and lower returns after. As the event year approaches, the (non-cumulative) annual returns of high NOA firms climb to about 35% in year 1, but the returns are under 5% in year +1. Low NOA firms somewhat less markedly switch from doing poorly in year 1 to well in year +1. Even as far as 5 years after the event year, high NOA firms are averaging annual returns lower than those of low NOA firms. 20

23 4. The Operating Accruals Anomaly We wish to examine whether investor misperceptions of net operating assets generates stock return predictability above and beyond effects identified in previous literature. We begin by first verifying whether Sloan s (1996) operating accruals anomaly applies in our more recent sample, and whether the operating accruals effect is robust with respect to a wider battery of controls and test procedures than have been applied in the existing literature. 4.1 Characteristic-Adjusted Portfolio Benchmark Abnormal Returns by Accrual Deciles Tables 4A and B report the average returns of portfolios sorted on NOA and Accruals variables as defined in Section 2. Every month, stocks are ranked by NOA or Accruals, placed into deciles, and the equal-weighted and value-weighted monthly raw and characteristic adjusted returns are computed. We require at least a four-month gap between the return portfolio formation month and the fiscal year end to ensure that investors have the financial statement data prior to trading. The average raw and characteristic-adjusted returns and t-statistics on these portfolios, as well as the difference in mean returns between decile portfolio 1 (lowest ranked) and 10 (highest ranked), are reported. Adjusted returns are calculated using a characteristic-based benchmark to take into account return premia associated with size, book-to-market and momentum. We do not take a stand as to whether these known return effects derive from risk or mispricing; in either case, our use of these controls implies that any effect we find is incremental to these known effects. The benchmark portfolio is based on the matching procedure used in Daniel, Grinblatt, Titman, and Wermers (1997). All firms in our sample are first sorted each month into size quintiles, and then within each size quintile further sorted into book-to-market quintiles. Stocks are then further sorted within each of these 25 groups into quintiles based on the firm s past 12-21

24 month returns, skipping the most recent month (e.g., cumulative return from t-12 to t-2). Stocks are weighted both equally and according to their market capitalizations within each of these 125 groups. The equal-weighted benchmarks are employed against equal-weighted portfolios, and the value-weighted benchmarks are employed against value-weighted portfolios. To form a size, book-to-market, and momentum-hedged return for any stock, we simply subtract the return of the benchmark portfolio to which that stock belongs from the return of the stock. The expected value of this return is zero if size, book-to-market, and past year return are the only attributes that affect the cross-section of expected stock returns. Table 4 confirms that the operating accruals effect remains strong in the 1990s beyond the years examined by Sloan (1996), lasts for up to 3 years of subsequent returns, and is robust with respect to some well-known financial market anomalies. We find that in the year following the reported operating accruals, the monthly abnormal equally-weighted returns spread between high and low Accruals deciles is -0.69% per month (t = -7.32). In year t+2, the effect is about one half as strong, -0.33% per month (t = -4.12), and remains significant in year 3, -0.21% per month (t=-2.22). Put Table 4 about here. The abnormal returns are, in absolute value, considerably larger and more significant for the highest Accruals decile (-0.55%, -0.34%, and -0.20%, all significant, in years t+1, t+2 and t+3 respectively) than for the lowest Accruals decile (0.14% and significant in year t+1; 0.00%, and 0.02%, insignificant in years t+2 and t+3 respectively). As a result, an investor who wants to avoid short positions would find it hard to profit from the operating accruals anomaly. The main opportunity appears in year t+1, in which there are significantly positive abnormal returns associated with the four lowest ranking portfolios. In the subsequent 2 years, the lowest ranking 22

25 decile portfolio does not attain significantly positive abnormal returns. Table 4 Panels A and B also describe hedge profits using CAPM, Fama-French 3-factor, and 4-factor α s. The Fama-French 3-factor model contains the market portfolio and two factormimicking portfolios associated with the size effect (SMB) and the book-to-market effect (HML). The 4-factor model includes a momentum factor-mimicking portfolio as well. We take the characteristics-adjusted spread between decile 10 and decile 1, adjust for the premium for beta, and adjust a second time for the premia associated with size, book/market, and momentum using the factor model regressions. The results using these alternatives are broadly similar. Results are weaker using value weights than equal weights, the only significant portfolio return being for the highest accruals portfolio. Using value-weighted returns, there are no profits opportunities associated with long positions in low accrual firms. Figure 2 Panel A graphs the profits from the hedge strategy of going long in Accrual decile 1 (lowest) and shorting Accrual decile 10 (highest), broken down by calendar year by summing the monthly abnormal equally weighted portfolio benchmark returns. 6 This strategy is profitable in the great majority of years (34 out of 38). The operating accruals effect was particularly strong in 1999, but the finding of an effect is robust with respect to the removal of Figure 3 Panel A provides a similar picture for value-weighted returns, in which the strategy profits in 29 out of 38 years. 6 Zach (2002) also shows using a portfolio characteristic matching procedure that the accruals effect remains after controlling for firm size, the book-to-market ratio, and the past 12-month returns up to 1998 for annual returns. He shows that the size of the effect is robust with respect to whether buy and hold or cumulative abnormal returns are used to measure the abnormal annual returns, and 80% of the excess returns remain after mergers and acquisition firms are eliminated. 23

26 4.2 Fama-MacBeth Monthly Cross-Sectional Regression Method An advantage of the characteristics-based portfolio matching procedure is that it does not rely on the linearity assumptions of regression analysis. However, regression analysis lends itself to inclusion of a greater number of control variables. As a further robustness check, Table 5 describes the relation between Accruals and the cross-section of average returns using an extension of the Fama and MacBeth (1973) method first introduced in Teoh, Welch, and Wong (1998a,b). Every month from July, 1966 through December, 2002, the firm s stock return is regressed on ln(size) where size is defined as the firm s market capitalization; ln(book-tomarket); the previous month s return on the stock, denoted ret(-1:-1); the previous year s return on the stock from month t-12 to t-2, denoted ret (-12:-2); the return on the stock starting from month t-36 to t-13, denoted ret(-36:-13); and Accruals lagged either one, two or three years. The time-series average of the monthly coefficient estimates and their associated time-series t- statistics (in italics) are reported. Put Table 5 about here. By using this Fama-MacBeth-like approach, we are able to control not only for momentum, size, and book-to-market, but also for the short-term one-month contrarian effect, and (using returns from month -36 through -13) the long-run winner/loser effect. Panels A, B, and C respectively describe the effects of Accruals on returns one year, two year and three years in the future. In all the panels, the standard control variables are significant. Model 2 in each of Panels A, B and C, indicates that operating accruals are highly significantly negatively related to cross-sectional stock returns for years t+1 and t+2, and are marginally significant for year t+3. The t-statistics on Accruals in Model 2 are -6.33, and in Panels A, B and C respectively. 24