Investor Learning, Earnings Signals, and Stock Returns* Peng-Chia Chiu† & Timothy D. Haight‡ December 2018 Abstract Prior studies show that investor learning about earnings-based return predictors from academic research erodes return predictability. However, the signaling power of “bottom-line” earnings has declined over time, which complicates assessments of investor learning about profitability signals underlying earnings. We show that modified earnings variables with lower susceptibility to signal weakening exhibit rates of return attenuation that are 30-64% lower than rates for bottom-line earnings variables over our sample period. Notably, return gaps between bottom-line and less susceptible variables are widest in recent years, especially within non-overlapping samples and samples with weak bottom-line signals (e.g., special items, losses, fourth fiscal quarter). Our results hold after controlling for risk factors known to predict returns, they do not appear to be attributable to ex ante earnings volatility, and they are robust to alternative sample selection criteria, sub-period partitions, and portfolio holding windows. Overall, our results suggest that while investor learning is apparent in the data, learning efforts to date have been suboptimal at exploiting profitability signals within firms’ earnings streams. Keywords: investor learning; earnings properties; market efficiency; stock returns JEL classification: G11, G12, G13, G14, M41 Data Availability: Data are publicly available from sources identified in the article. * We appreciate the helpful comments and suggestions of Brad Barber, Novia Chen, Scott Delanty, Lucile Faurel, Laurel Franzen, Eric Guest, Marinilka Kimbro, Qin Li, Alex Nekrasov, Linda Myers, Morton Pincus, Terry Shevlin, Siew Hong Teoh, Hai Tran, Mitch Warachka, and Crystal Xu; participants at the 2013 Western American Accounting Association Conference, the 2013 Haskell and White Corporate Reporting and Governance Conference, the 2014 American Accounting Association Annual Meeting, the 2015 European Accounting Association Annual Congress, and the 2015 CAAA Annual Conference; and workshop participants at Loyola Marymount University, the University of California, Irvine, National Taipei University, Singapore Management University, and Sun Yat-Sen University. † School of Accountancy, the Chinese University of Hong Kong. 12 Chak Cheung Street, Shatin, N.T., Hong Kong. Tel.: +852-3943-7835. Email: chiupc@cuhk.edu.hk. ORCID: 0000-0002-4525-8887. ‡ College of Business Administration, Loyola Marymount University. 1 LMU Drive, Los Angeles, CA, 90045. Tel.: +1 310-338-4541. Email: thaight@lmu.edu. ORCID: 0000-0002-9077-8558. Electronic copy available at: https://ssrn.com/abstract=3348325 Investor Learning, Earnings Signals, and Stock Returns Abstract Prior studies show that investor learning about earnings-based return predictors from academic research erodes return predictability. However, the signaling power of “bottom-line” earnings has declined over time, which complicates assessments of investor learning about profitability signals underlying earnings. We show that modified earnings variables with lower susceptibility to signal weakening exhibit rates of return attenuation that are 30-64% lower than rates for bottom-line earnings variables over our sample period. Notably, return gaps between bottom-line and less susceptible variables are widest in recent years, especially within non-overlapping samples and samples with weak bottom-line signals (e.g., special items, losses, fourth fiscal quarter). Our results hold after controlling for risk factors known to predict returns, they do not appear to be attributable to ex ante earnings volatility, and they are robust to alternative sample selection criteria, sub-period partitions, and portfolio holding windows. Overall, our results suggest that while investor learning is apparent in the data, learning efforts to date have been suboptimal at exploiting profitability signals within firms’ earnings streams. 1 Electronic copy available at: https://ssrn.com/abstract=3348325 There is considerable evidence that investors learn about and trade on variables that predict cross-sectional stock returns in academic research. Several recent studies report sharp postpublication declines in predictor variable returns and that declines relate to increased trading activity exploiting return predictability.4 So substantial are learning effects that some of the most well-known predictors in the accounting literature now generate insignificant returns in many sample settings. 5 While capital markets benefit when investors learn about the relevance of accounting variables for stock prices, there are indications that information underlying many of these variables has become less relevant over time. In particular, bottom-line earnings, a key input to numerous predictors in the literature, lost significant value relevance over the past several decades (Collins et al. 1997; Lev and Zarowin 1999). Although explanations vary, most studies attribute this decline to time-series changes to the properties of earnings that weaken earnings’ ability to signal information about the firm’s ongoing economic profitability. 6 Such signal weakening presents challenges for assessing the success of learning efforts at exploiting profitability signals because exploitable signals are increasingly less apparent in earnings-based predictors over time. Addressing this issue is important because earnings components that are less susceptible to signal weakening may continue to predict significant stock returns if investors fail to fully appreciate the relevance of these components for stock prices. This study investigates earnings-based signal learning by examining the time-series pricing of performance measures with lower susceptibilities to signal weakening. The intuition of this approach is straightforward. Measures less affected by earnings’ changing properties should provide relatively stable signals over time. Thus, if we observe return attenuation for variables 4 For example, McLean and Pontiff (2016) show gradual erosion in the profitability of 97 anomaly-based trading strategies (58% decline on average) in the post publication period. Other large-scale studies documenting return attenuation include Green et al. (2013) and Chordia et al. (2014). 5 Studies that link investor learning to the elimination of prominent accounting-based return predictors include Johnson and Schwartz (2001), Richardson et al. (2010), Green et al. (2011), and Milian (2015). 6 Studies that document significant changes to the properties of earnings over time include Givoly and Hayn (2000), Dichev and Tang (2008), and Bushman et al., (2016). 1 Electronic copy available at: https://ssrn.com/abstract=3348325 constructed from these measures, we can more confidently ascribe return attenuation to learning about profitability signals underlying earnings. To operationalize this intuition, we need to identify performance measures that are less susceptible to signal weakening relative to bottom-line earnings. Since signal weakening has been linked to the proliferation of one-time items and poorer matching (Collins et al., 1997; Lev and Zarowin 1999; Dichev and Tang 2008), we require less susceptible measures to exclude one-time items and exhibit more robust matching relative to bottom-line earnings. Based on these criteria, we employ the following “less susceptible” measures: gross profit, operating profit, and earnings before one-time items. Our primary analysis compares our bottom-line and less susceptible measures in the context of short-term stock price continuation. Within this context, standardized unexpected earnings (i.e., SUE) has been shown to be among the most robust return predictors in the literature (e.g., Bernard and Thomas 1990; Fama and French 2008). Nevertheless, SUE’s ubiquity made it a natural subject of investor learning, resulting in significant attenuation of its returns over the years (Johnson and Schwarz 2001; Richardson et al., 2010; Chordia et al., 2014; Milian 2015). Since SUE is constructed from changes in bottom-line earnings, we analogously construct three modified SUE variables (MSUE1, MSUE2, MSUE3) using changes in our less susceptible measures. Over a sample period spanning 1979-2017, we first show that our MSUE variables exhibit robust first-order serial correlation with increasing incremental abilities to predict one-quarterahead SUE. By contrast, SUE’s one-quarter-ahead predictive abilities have generally declined, consistent with signal weakening. These results support the notion that our MSUE variables capture relevant information about ongoing economic profitability with relatively low susceptibility to signal weakening over time. Thus, our MSUE variables appear to be suitable for assessing earningsbased signal learning. Next, we construct quarterly hedge portfolios that take long (short) positions in the highest (lowest) decile of each of our test variables and compute size-adjusted, three-month stock returns 2 Electronic copy available at: https://ssrn.com/abstract=3348325 over three subperiods (i.e., 1979-1989; 1990-1990; 2000-2017). From the first to the third subperiod, we find that SUE portfolio returns decline by 58 percent, consistent with prior studies. By contrast, MSUE portfolio returns decline at rates that are 30 percent to 64 percent smaller than the SUE rate over the same period. Moreover, we observe a widening positive return gap building over time between our MSUE portfolios and the SUE portfolio. For example, from 1979-1989, the annualized average hedge return gap between our MSUE variables and SUE (i.e., MSUE returns minus SUE returns) ranges from -1.2 percent to +0.6 percent, while, from 2000-2017, the average return gap between MSUE and SUE ranges from +2.9 percent to +5.1 percent. To better understand the return patterns generated by our test variables, we repeat our hedge portfolio analysis using (a) non-overlapping portfolios and (b) firms with weak bottom-line signals (e.g., firms with special items, losses, and fourth quarter reporting). The return patterns show even stronger contrast between the SUE and MSUE portfolios. In the case of MSUE, removing overlapping SUE portfolio firms from MSUE portfolios significantly reduces or eliminates return attenuation. Meanwhile, restricting MSUE portfolios to firms with weak bottom-line signals does not significantly erode returns, even in the most recent subperiod (2000-2017). These results suggest that SUE learning drives return attenuation of MSUE portfolios in the full sample. However, learning about SUE’s return predictability has an increasingly limited effect on MSUE portfolio performance as the degree of portfolio overlap declines over time.7 Thus, learning efforts appear to be having dwindling success at narrowing profitability-driven efficiency gaps. In the case of SUE, we find that portfolio return attenuation deepens after removing overlapping MSUE firms, while restricting SUE portfolios to firms with weak bottom-line signals eliminates SUE’s ability to generate economically and statistically significant returns in the 20002017 period. These results highlight the sensitivity of SUE portfolio returns to the strength of the signal underlying SUE. Such sensitivity is important to consider when evaluating return attenuation 7 For example, the degree of SUE-MSUE portfolio overlap decreases from an average of 61% (44% to 81% range) in the 1979-1989 subperiod to an average of 51% (39% to 65% range) in the 2000-2017 subperiod. 3 Electronic copy available at: https://ssrn.com/abstract=3348325 evidence because weakening signals mechanically drive down returns even in the absence of investor learning. Given the rising frequencies with which firms report one-time items and losses (Hayn 1995; Donelson et al., 2011; Li 2011), controlling for decaying earnings signals is increasingly crucial for assessing the efficiency implications of earnings-based learning evidence. Next, we examine temporal patterns in the return predictability of levels specifications of our performance measures. Since predictive signals and susceptibility to signal weakening are properties of our test variables’ underlying performance measures, return patterns that are driven by these properties should arise when using levels specifications of our measures. Thus, levels tests should mitigate concerns that our main results stem from the nature of our variable construction (e.g., earnings expectation models underlying SUE and MSUE are misspecified). Consistent with our earlier results, we find that levels of our less susceptible measures exhibit robust serial correlation and less return attenuation over time relative to levels of bottom-line earnings. Last, we run a battery of additional tests to confirm the robustness of our results. FamaMacBeth regressions show that the return predictability of all three MSUE variables remains significant and relatively stable over time after controlling for SUE and risk characteristics that are known to predict returns. We also find that our hedge portfolio results are not explained by ex ante earnings volatility (Cao and Narayanamoorthy 2012) and that they continue to hold under alternative sample selection criteria, subperiod partitions, and portfolio holding windows. Overall, our results suggest that while investor learning is apparent in the data, learning efforts to date have been suboptimal at closing profitability-driven efficiency gaps. In particular, significant inefficiencies remain in the pricing of earnings components with reliable serial properties (i.e., robust predictive signals). Such mispricing has become less apparent over time when studying earnings-based return predictors because evolving earning properties weaken signals that helped to generate returns in the first place. We emphasize that the goal of our study is not to identify “new” return predictor signals with our less susceptible measures. Indeed, earnings components up and down the income 4 Electronic copy available at: https://ssrn.com/abstract=3348325 statement have long been shown to have incremental signaling abilities and cross-sectional return predictability (Ou and Penman, 1989; Lev and Thiagarajan, 1993; Abarbanell and Bushee, 1997; Piotroski, 2000; Novy-Marx, 2013; Ball et al., 2015). Rather, we exploit the robust signaling abilities of our less susceptible measures to increase the power of learning tests that examine whether investors are closing efficiency gaps created by profitability signals underlying earnings. Thus, unlike most prior studies, we examine the return predictability of earnings components in a time-series context and show that despite investor learning efforts, signal-driven return predictability in earnings remains significant with little sign of going away in the near future. Furthermore, we do not seek to run a horse race among our less susceptible measures to identify the one that is least susceptible to signal weakening. Although some studies link individual income statement line items to changing earnings properties that weaken earnings signals (e.g., Donelson et al., 2011), other studies find pervasive signal weakening effects across recurring and nonrecurring line items (e.g., Srivastava 2014). Thus, it is not clear whether a line item approach would best minimize signal weakening in all cases. Our study contributes to the investor learning literature by examining whether investors have become more efficient at pricing signals underlying return predictor variables rather than just the variables themselves. This distinction is important when assessing earnings-based learning evidence because profitability signals have become less apparent in earnings over time. We show that the pricing of earnings components with lower susceptibility to signal weakening has seen little evidence of a learning effect over our sample period. Similar to Milian (2015), our findings suggest that investor learning activities do not always narrow efficiency gaps. While Milian focuses on how investor sophistication affects the pricing of SUE, we use multiple variables to show that inabilities to fully exploit signals within firms’ earnings streams impede efficiency gap closures. Our study also contributes more broadly to the literature on stock return predictability by highlighting the importance of controlling for evolving earnings properties when interpreting return attenuation of earnings-based predictor variables. While prior studies attribute attenuation to 5 Electronic copy available at: https://ssrn.com/abstract=3348325 eroding market frictions (Chordia et al., 2014), investor learning about variables with stock return predictability (McLean and Pontiff 2016) and institutional shocks to the information environment (Hung et al., 2014), our results provoke the possibility that return attenuation may also arise mechanically as a result of decaying serial correlation in accounting data. A key implication of our findings is that studies that interpret return attenuation of earnings-based predictors as evidence of efficiency gains are more likely to provide upper bound estimates of those gains (Hung et al. 2014, Chordia et al. 2014, Richardson et al. 2010).8 Moreover, our findings of corroborating evidence with levels specifications of our earnings measures suggests that evolving time series properties may have intertemporal implications for a wide variety of earnings-based return predictors (accruals, P/E ratios, book-tax differences, and many others), which we leave for future research.9,10 1. Prior Literature, Motivation, and Earnings Measures 1.1 Prior Literature Researchers have documented numerous accounting and market-based variables with excess return predictability (see Subrahmanyam 2010 and Green et al., 2013 for literature reviews). Given the potential trading gains implied by this stream of research, it is perhaps unsurprising that excess returns for many documented predictor variables attenuated over time (Richardson et al., 2010; Chordia et al., 2014; McLean and Pontiff 2016). Facilitated by eroding trading frictions (Chordia et al., 2014), return attenuation has been linked to investors learning about return predictability from academic research. For example, McLean and Pontiff (2016) report that postpublication returns decline by an average of 58% for the 97 return predictors in their sample. Within the accounting literature, learning effects appear to be so substantial that many of the most widely 8 We do not attempt to formally quantify the effect of signal weakening on return attenuation in relation to other documented drivers. We leave such analysis for future research. 9 Green et al. (2013) report that 147 of their 330 are accounting-based return predictors. 10 Bushman et al. (2016) speculate that the declining accrual/cash-flow relation over time may explain, in part or in full, the decline in the accrual anomaly in recent years (e.g., Green et al., 2011). 6 Electronic copy available at: https://ssrn.com/abstract=3348325 documented return predictors no longer yield profitable hedge returns in recent years (Johnson and Schwarz 2001; Richardson et al., 2010; Green et al., 2011; Milian 2015). During the time period over which investor learning contributed to return attenuation, the properties of earnings changed dramatically. 11 Givoly and Hayn (2000) document declining earnings levels and timelier recognition of bad news in earnings over the latter part of the twentieth century, consistent with increased conservatism in financial reporting. Over a similar timeframe, Dichev and Tang (2008) document declining earnings persistence, rising earnings volatility, and a weakening contemporaneous revenue-expense correlation, consistent with declining matching in earnings. Bushman et al., (2016) document a weakening accrual-cash flow relation over the past half-century, which they attribute to temporal increases in non-timing-related accrual recognition as proxied by one-time and non-operating items and by loss frequencies. An adverse consequence of these changing properties is that earnings gradually became less relevant for stock prices. Collins et al. (1997) report a forty-year decline in the value relevance of earnings and they show that temporal increases in the reporting of losses and large one-time items are major contributors to the decline. Lev and Zarowin (1999) also report a strong decline in the relevance of earnings, which they attribute to declining matching in earnings driven by rising levels of investments that are expensed as incurred (e.g., R&D, restructuring costs). Srivastava (2014) shows that the increasing presence of intangible-intensive firms, which tend to have higher earnings volatility, poorer matching and greater unrecorded growth options, accounts for a considerable portion of the overall decline in earnings’ value relevance. 1.2 Motivation The evidence in this section pertains to “bottom-line” earnings, which the literature typically defines as GAAP earnings before extraordinary items. Later we will argue that various subtotals of earnings (e.g., gross profit) are likely to be significantly less susceptible to the changes summarized in this section. 11 7 Electronic copy available at: https://ssrn.com/abstract=3348325 Before we discuss how earnings’ changing properties and declining relevance relate to investor learning evidence, it is important to first consider the link between accounting outputs and future stock returns. A common explanation of accounting-based return predictability is that it reflects investor mispricing of the future earnings implications of accounting outputs (Bernard and Thomas 1990; Sloan 1996). In other words, accounting outputs provide “signals” for future profitability and variables constructed from accounting outputs predict returns when investors fail to fully and immediately impound these signals into stock prices. In view of this linkage, two issues arise when assessing the success of learning efforts at exploiting profitability signals underlying earnings. First, prior evidence of earnings’ changing properties and declining relevance suggests that earnings’ signal for future profitability has likely weakened over time. This implies that return magnitudes would have declined in the absence of investor learning, thereby making it difficult to isolate and cleanly quantify the various sources of earnings-based return attenuation.12 However, the more fundamental issue presented by earnings’ signal weakening is that it makes return attenuation of earnings-based predictors weak evidence of efficiency gains achieved through signal learning. While investor learning about documented return predictors can narrow efficiency gaps, complete gap closure ultimately requires investors to detect underlying predictive signals and appropriately impound such signals into stock prices. Given earnings’ evolving nature, signal detection likely requires investors to deal with a moving target, necessitating a deeper understanding of the factors that create signals in earnings streams so that earnings information can be appropriately priced. Thus, even if investors can eliminate the return predictability of documented earnings-based predictors, significant efficiency gaps may remain if investors fail to fully track and appropriately price components of earnings that are relevant for firm value. 12 Formal attempts to isolate and quantify the various sources of earnings-based return attenuation (e.g., investor learning, signal weakening, and declining trading frictions) go beyond the scope of our study, so we leave such attempts to future research. 8 Electronic copy available at: https://ssrn.com/abstract=3348325 1.3 Bottom-Line Earnings and “Less Susceptible” Performance Measures We investigate earnings-based signal learning by examining the time-series pricing of performance measures with lower susceptibilities to the signal-weakening effects of changing earnings properties. The intuition underlying this approach is straightforward. Lower susceptibility to the signal-weakening effects of changing earnings properties should help to preserve a measure’s predictive signal over time. Therefore, if we observe return attenuation for these measures, we will have higher confidence that return attenuation captures learning about profitability signals underlying earnings. To provide a benchmark for our analysis and to facilitate comparisons with prior research, we examine the traditional bottom-line earnings measure, defined as earnings before extraordinary items. Studies that document changing earnings properties typically use bottom-line earnings to demonstrate their effects, so we consider bottom-line earnings to be highly susceptible to signal weakening. To identify measures with lower susceptibility to signal weakening, we consider two key sources of signal weakening from our literature review. The first source is a temporal decline in the matching of expenses to revenues in earnings (Dichev and Tang 2008; Lev and Zarowin 1999). The second source is a temporal increase in the frequency and magnitude of one-time items in earnings (Collins et al., 1997; Bushman et al., 2016). These sources suggest that less susceptible measures should (1) be less susceptible to temporal declines in matching and (2) exclude one-time items. Based on these criteria, we identify three measures that should be less susceptible to signal weakening over time. Our first measure is gross profit, which we define as net sales minus cost of goods sold. As with all of our less susceptible measures, gross profit excludes one-time items, and its lower susceptibility to declining matching follows from its exclusion of all other expenses except cost of goods sold. Although such a high degree of exclusions should mitigate mismatching effects, 9 Electronic copy available at: https://ssrn.com/abstract=3348325 these exclusions are also likely to filter out many well-matched expenses that enhance earnings’ signaling ability. Nevertheless, prior research suggests that innovations in gross profit tend to be highly informative to investors (Lipe 1986) and more recent studies suggest that gross profit provides significant incremental information about future profitability and stock returns (NovyMarx 2013, Akbas, Jiang, and Koch 2017).13 Moreover, gross profit has attracted growing interest in recent years among investment professionals (e.g., Zweig 2013), which suggests that the empirically documented signaling ability of gross profit has not significantly declined over time. Our second measure is operating profit, which we define as gross profit minus selling, general and administrative (SG&A) expense. Ball et al. (2015) use matching arguments to motivate operating profit’s suitability as a proxy for expected profitability in the context of dividend discount models of equity valuation. Consistent with the signal-enhancing role of matching, they show that operating profit outperforms both bottom-line earnings and gross profit as a proxy for expected profitability. This suggests that operating profit outperforms other performance measures in capturing “recurring” profitability, so we expect operating profit will have strong signaling power overall. Nevertheless, Srivastava (2014) finds that the temporal rise of SG&A-intensive firms contributes to declining earnings quality, so whether operating profit will provide a more robust signal over time compared to gross profit is ultimately an empirical question. Our third measure is earnings before one-time items, which we define as bottom-line earnings excluding special items (i.e., one-time items identified by Compustat). Exclusion of onetime items mirrors the pro forma approach commonly used by market participants to remove valueirrelevant components from GAAP earnings (Bradshaw and Sloan 2002; Bhattacharya et al. 2003; Loughee and Marquardt 2004). Also, Donelson et al. (2011) show that much of the temporal decline in matching is attributable to increases in large one-time items. Nevertheless, this measure includes Novy-Marx (2013, pp. 2-3) describes gross profit as “the cleanest accounting measure of true economic profitability. The farther down the income statement one goes, the more polluted profitability measures become, and the less related they are to true economic profitability.” 13 10 Electronic copy available at: https://ssrn.com/abstract=3348325 non-operating items, which are often indirectly matched to revenues and have been shown to contribute to the weakening timing role of accruals (Bushman et al., 2016). Furthermore, earnings streams of newer firm cohorts tend exhibit poorer matching among recurring items (Srivastava 2014). This suggests that earnings before one-time items may be increasingly susceptible to signal weakening given its relatively large composition of recurring expense items. Nevertheless, we include earnings before one-time items among our less susceptible measures because firms and analysts continue to use it when formulating pro forma earnings, which suggests that the market continues to recognize its superior signaling ability relative to bottom-line earnings. 2. Sample Selection, Variables, and Descriptive Statistics We draw our sample from the CRSP monthly returns database and the Compustat quarterly database for fiscal years spanning 1979 through 2017. We require firms to have stock prices exceeding $1 per share. While we examine both levels and changes specifications of our performance measures, we conduct our main analyses using changes specifications to facilitate comparisons with SUE (i.e., standardized unexpected earnings), which is among the most widely documented accounting-based return predictors.14 We define SUE as quarterly earnings per share minus expected earnings per share, scaled by the standard deviation of quarterly earnings growth over the previous eight quarters, as in Jegadeesh and Livnat (2006).15 We construct three analogues to SUE by substituting (all on a per-share basis) gross profit (MSUE1), operating profit (MSUE2), and earnings before one-time items (MSUE3) in place of earnings (see Section 1.3 for measure definitions). Since SUE is often interpreted as a “news” variable, we may interchangeably refer to SUE and our MSUE variables as “news” or “changes” variables. SUE’s return predictability, conventionally referred to as post earnings announcement drift (PEAD), dates back to the seminal work of Ball and Brown (1968) who were the first to document the phenomenon. For reviews of the PEAD literature, see Kothari (2001). 15 Expected earnings are assumed to follow a seasonal random walk with drift. The drift term is measured as the average of quarterly earnings growth over the previous eight quarters. 14 11 Electronic copy available at: https://ssrn.com/abstract=3348325 As in Thomas and Zhang (2011), our return window is the three-month period beginning in the first month of the calendar quarter that is at least three months subsequent to fiscal quarterend. This ensures that earnings information is publicly available before the holding period begins.16 We compute excess returns using ADJ_RET, which is the size-adjusted buy-and-hold three-month stock return calculated as described in Lyon et al. (1999). We also employ a set of known risk characteristics: SIZE is measured as the natural log of the market capitalization at the end of the most recent fiscal quarter for which data are available; BM is the book-to-market ratio, calculated as the book value of equity divided by the market value of equity at the end of the most recent fiscal quarter for which data are available; MOM is the buy-and-hold six month raw stock return leading up to the month prior to the return holding window. Table I provides descriptive statistics for our primary variables of interest. All variables except ADJ_RET are winsorized quarterly at the 1st and 99th percentiles. Panel A provides summary statistics for variables used in our regressions. SUE has a mean (median) of -0.366 (-0.049) and a standard deviation of 4.814. By comparison, our MSUE variables have means (medians) ranging from -0.190 to +0.051 (-0.024 to +0.144) and standard deviations ranging from 4.043 to 4.220. The lower standard deviations of our MSUE variables are consistent with theoretical implications in Dichev and Tang (2008) that link better matching in earnings to lower earnings volatility. Distribution statistics for ADJ_RET, SIZE, BM and MOM are in line with values tabulated in prior literature. Table I, Panel B presents correlations of our SUE and MSUE variables with ADJ_RET (i.e., future size-adjusted stock returns), by subperiod. We observe attenuation in return predictability for SUE and all three MSUE variables. However, from the first to the third subperiod (i.e., 19791989 to 2000-2017) attenuation for MSUE variables (correlation coefficients declining from an 16 With regard to our portfolio tests, we implement the model with the most flexible design rather than the model with the maximum return. Results are qualitatively unchanged under various holding windows and portfolio formation dates. Section 3 further discusses portfolio test considerations. 12 Electronic copy available at: https://ssrn.com/abstract=3348325 average of 0.059 to 0.027) is much milder than the attenuation for SUE (correlation coefficient declining from 0.057 to 0.011). Table I, Panel C shows, by subperiod, the degree of overlap between hedge portfolios formed on extreme deciles of SUE and each of our three MSUE variables (portfolio formation details are provided in Section 3). We note two findings. First, we observe considerable variation across our MSUE portfolios in their degree of overlap with the SUE portfolio. Second, for all three MSUE portfolios, we observe a decreasing trend over time in the degree of overlap with the SUE portfolio. For example, in the first subperiod, the degree of overlap ranges from 44.37% to 81.30%, while in the last subperiod, the range of overlap declines from 39.28% to 64.55%. These trends suggest that extreme SUE values (i.e., those determining SUE portfolio membership) are increasingly driven by earnings components that are excluded from our less susceptible measures. 3. Empirical Results 3.1 Serial Properties of SUE and MSUE Variables Over Time We begin our analysis by examining the signaling power of SUE and our three MSUE variables over time. Since we argue that gross profit, operating profit, and earnings before one-time items should all be less susceptible to signal weakening relative to bottom-line earnings, we expect all three measures to have comparatively stable signaling power throughout our sample period. We assess signal power stability by examining temporal trends in the first-order serial correlation patterns of variables constructed from bottom-line measures (SUE) and our three less susceptible measures (MSUE1, MSUE2, MSUE3). Specifically, we estimate the following two regression models for the full sample and by subperiod using the Fama-MacBeth approach (1973): ππππΈπ,π‘+1 = πΌ0 + πΌ1 ππππΈπ,π‘ + πΌ2 πππΈπ,π‘ + ππ,π‘+1 (1) πππΈπ,π‘+1 = π½0 + π½1 ππππΈπ,π‘ + π½2 πππΈπ,π‘ + ππ,π‘+1 (2) 13 Electronic copy available at: https://ssrn.com/abstract=3348325 We assess the stability of MSUE’s signaling power by analyzing the trend in α1 over our three subperiods, and we expect α1 to not significantly decline over time. By contrast, we expect β2 to decline over time if time-series changes to earnings’ properties weaken the implications of current earnings for one-quarter-ahead earnings. We estimate multivariate (rather than univariate) regressions so that we can also analyze the trend in MSUE’s incremental predictability for next quarter’s SUE (i.e., the trend in β1). An increasing β1 coefficient would be consistent with the notion that signals that underlie our less susceptible measures provide increasing amounts of incremental information about future profitability relative to the information provided by bottom-line earnings alone. Table II provides Fama-MacBeth regression results for our estimations of equations 1 and 2.17 Regression results are tabulated across three panels (A.1, A.2, A.3) each corresponding to a specific MSUE variable (i.e., MSUE1, MSUE2, MSUE3). Within each panel are 3 sets of 2 columns. Each set corresponds to a specific subperiod (“1979-1989,” “1990-1999,” “2000-2017”) while each column within that set corresponds to a specific equation (equation 1 results are on the left, equation 2 results are on the right). Considering first MSUE’s incremental abilities to predict one-quarter-ahead MSUE (i.e., looking at the first row, left column within each subperiod), we see that MSUE coefficients increase from the first to the third subperiod by a range of 6% to 11%. These results are consistent with our less susceptible measures having stable signaling power over time. Furthermore, we note that SUE’s incremental predictability for one-quarter-ahead MSUE is considerably weaker and negative when MSUE is defined using gross and operating profit (i.e., Panels A.1 and A.2). While SUE’s incremental predictability is positive and somewhat stronger when predicting next quarter’s earnings before one-time items (Panel A.3), we see that such predictability is decreasing over time (e.g., the SUE coefficient declines by 33% from the first to the third subperiod). 17 Results (untabulated) are qualitatively unchanged using univariate specifications. 14 Electronic copy available at: https://ssrn.com/abstract=3348325 Considering next SUE’s incremental abilities to predict one-quarter-ahead SUE (i.e., looking at the second row, right column within each subperiod), we see that coefficients decline from the first to the second subperiod by a range of 22% to 33%, though the coefficients in Panels A.2 and A.3 increase somewhat from the second to the third subperiod. The uptick from subperiod two to three may be attributable to increasing classification fluidity between recurring and nonrecurring items (McVay 2006; Cready et al., 2010). Nevertheless, we observe an overall declining trend in SUE’s coefficient from the first to the third subperiod (5% to 32% decline), consistent with signal weakening in bottom-line earnings. Moreover, when we consider the overall trend in MSUE’s incremental abilities to predict next quarter’s SUE (i.e., looking at the first row, right column within each subperiod), we generally observe coefficient increases (12% in Panel A.1, 13% in Panel A.2).18 Overall, the results in Table II are consistent with our measures exhibiting varying susceptibilities to the signal-weakening effects of earnings’ changing properties. While bottom-line earnings appear to be most susceptible to signal weakening, our less susceptible measures appear to provide relatively stable signals over time. Furthermore, our less susceptible measures generally exhibit increasing incremental signaling power for future profits, which suggests their signals are increasingly important for stock prices. 3.2 Excess Return Predictability of SUE and MSUE Variables Over Time Table III reports the time-series means of future size-adjusted stock returns for hedge portfolios formed on SUE and our three MSUE variables. At the end of each calendar quarter, stocks are sorted into deciles based on the value of each sorting variable (i.e., SUE, MSUE1, MSUE2, or MSUE3). We then form zero-investment hedge portfolios for each variable by going long (short) Our finding in Panel A.3 of an uptick (downtick) from sub-period two to three in SUE’s (MSUE3’s) incremental ability to predict next quarter’s SUE likely stems from two sources: (a) increased classification fluidity between recurring and non-recurring items and (b) signal weakening among non-operating items (included in MSUE3), consistent with evidence in Bushman et al. (2016). 18 15 Electronic copy available at: https://ssrn.com/abstract=3348325 in the variable’s highest (lowest) decile stocks. Size-adjusted buy-and-hold returns for each stock are calculated over the three months subsequent to the portfolio formation date, and an equalweighted mean return is computed for each hedge portfolio. Fama-MacBeth t-statistics are computed for each portfolio based on the time-series distribution of the mean hedge portfolio returns. Looking first at the returns of the SUE portfolio, we see that the mean return decreases from 4.07% (t-stat = 7.70) in the 1979-1989 subperiod to 1.69% (t-stat = 3.12) in the 2000-2017 subperiod. This amounts to a 58% decline over time, which is in line with SUE’s rate of return attenuation over comparable periods in prior research (e.g., Richardson et al., 2010; Chordia et al., 2014). Looking next at the returns of our three MSUE portfolios, we see that mean returns range from 3.86% to 4.22% in the 1979-1989 subperiod, 3.51% to 4.23% in the 1990-1999 subperiod, and 2.41% to 2.96% in the 2000-2017 subperiod (t-stats across all variables and periods are greater than 4). When considering mean returns from the first to the third subperiod for each MSUE portfolio, we observe declines ranging from 21% (MSUE2) to 41% (MSUE3), all shallower than the SUE decline (58%). These trends indicate that the rate of return attenuation for our MSUE variables is 30% to 64% smaller than the SUE rate over our sample period [i.e., 1- (41/58) to 1 (21/58), with rounding]. Moreover, the return trends in Table III indicate a widening positive return gap building over time between our MSUE portfolios and the SUE portfolio. For example, from 1979-1989, the annualized average hedge return gap between our MSUE variables and SUE (i.e., MSUE returns minus SUE returns) ranges from -1.2 percent to +0.6 percent, while, from 2000-2017, the average return gap between MSUE and SUE ranges from +2.9 percent to +5.1 percent. Thus, while we see 16 Electronic copy available at: https://ssrn.com/abstract=3348325 return attenuation for all four of our portfolios, there appears to be some value relevant information in our less susceptible measures that is not being picked up by investors.19 To provide an alternative perspective on these portfolio trends, Figure 1 plots the value of a dollar invested in each portfolio at the beginning of the year 2000 over various time intervals.20 The overall picture that emerges from the figure is that all three MSUE portfolios increasingly outperform the SUE portfolio with no indication of reversion in the immediate future. These results, along with the results in Table III, provide an indication that investor learning may not be significantly narrowing earnings-driven efficiency gaps, as measures with robust signaling power remain significantly mispriced. 3.3 Non-Overlapping Hedge Portfolios Recall from Table I, Panel C that there is considerable overlap in firm membership across the SUE and MSUE portfolios, especially in the early part of our sample period (e.g., overlap ranges from 44% to 81% in the 1979-1989 subperiod). With the degree of portfolio overlap declining over time, two issues arise when interpreting the full sample results in Table III. First, attenuation of MSUE portfolio returns could largely be driven by SUE learning since efforts to arbitrage SUE will drive down returns for overlapping MSUE firms. While SUE learning in the overlapping case is clearly beneficial from an efficiency perspective, the declining overlap trend suggests that the influence of this source on the pricing of MSUE stocks is waning over time. Thus, if MSUE attenuation is driven by SUE learning and not by learning about robust predictive signals underlying MSUE, then MSUE-driven efficiency gaps are unlikely to close. 19 We also examine the higher moments of the portfolio returns in each subperiod (untabulated). Standard deviations increase from subperiod one to subperiod three for all four portfolios, though increases are not monotonic (standard deviations decrease from subperiod one to subperiod two for all portfolios except MSUE2). MSUE returns are positively skewed in most subperiods (MSUE3 returns are negatively skewed in subperiod two), while SUE returns are negatively skewed in subperiod two (skewness = -0.096) and subperiod three (skewness = -1.460). 20 Other return attenuation studies that use cumulative return graphs to depict temporal return patterns include Richardson et al. (2010) and Green et al. (2011). 17 Electronic copy available at: https://ssrn.com/abstract=3348325 Second, the declining presence of MSUE stocks in the SUE portfolio is likely to weaken the signaling power of the SUE portfolio, thereby providing a mechanical source of return attenuation. Although SUE learning evidence strongly suggests that learning is a significant driver of SUE’s declining returns, failing to account for potential sources of mechanical attenuation (i.e., earnings’ changing properties) can lead to overstated efficiency gains. Moreover, time-series changes in SUE’s signaling power can potentially complicate efforts to arbitrage SUE, especially among unsophisticated arbitrageurs, whose arbitrage efforts can exacerbate SUE mispricing (Milian 2015). Given these potential issues, we consider non-overlapping subsamples of our SUE and MSUE portfolio firms. To form these portfolios, we remove from each full sample portfolio firmquarters that appear in both the SUE and MSUE portfolio. Since we use three different MSUE variables, we consider three separate non-overlapping cases. Table IV repeats the hedge portfolio analysis from Table III using non-overlapping SUE and MSUE portfolios. After removing overlapping SUE firms from our MSUE portfolios, we first see that MSUE1 portfolio returns decline by 26.91% from the first to the third subperiod, which is shallower than the 37.56% decline of the full sample MSUE1 portfolio returns in Table III. Looking at the other non-overlapping MSUE portfolios, we see that MSUE2 portfolio returns decline by only 2.29% from the first to the third subperiod, while MSUE3 portfolio returns increase by 3.06% from the first to the third subperiod (full sample declines reported in Table III are 21.28% and 40.76% for the MSUE2 and MSUE3 portfolios, respectively). These results suggest that SUE learning drives much of the decline in MSUE portfolio returns in the full sample.21 However, as suggested by the prevailing return predictability of all three non-overlapping MSUE portfolios, such learning fails to fully detect robust predictive signals that are increasingly unique to MSUE over time. 21 The decline of the non-overlapping MSUE1 portfolio return in Table IV likely reflects investor learning about the return predictability of gross profit levels documented in Novy-Marx (2013). 18 Electronic copy available at: https://ssrn.com/abstract=3348325 Turning to the non-overlapping SUE portfolios, we see that removing overlapping MSUE firms deepens SUE’s return attenuation relative to the full sample. For example, the full sample SUE return decline was 58% from subperiod one to three, while non-overlapping SUE return declines range from 71.88% to 99.85%. Moreover, non-overlapping SUE returns are much weaker in the 2000-2017 subperiod, ranging from marginally significant to insignificant using conventional cutoffs. These results suggest that the presence of MSUE firms significantly improves the performance of the SUE portfolio. Nevertheless, this presence is declining over time, highlighting the potential for mechanical attenuation to increasingly act on the performance of SUE portfolios. 3.4 Excess Return Predictability in “Weak Signal” Subsamples Next, we consider the excess return predictability of SUE and MSUE in subsamples of firms with reporting characteristics that suggest bottom-line earnings provides weak signals for future profitability. Researchers find that special items and losses—characteristics that are linked to earnings’ declining relevance—have been reported with increasing frequency and magnitude over the years (Hayn 1995; Collins et al., 1997; Donelson et al., 2011; Li 2011). Since SUE hedge portfolios are formed on firms with extreme SUE realizations, these trends suggest that firms with weak signal characteristics are increasingly likely to be selected into SUE portfolios over time. Thus, analyzing SUE portfolios formed on weak signal firms provides an additional way to highlight the potential for mechanical attenuation of SUE returns in the full sample. In the case of our MSUE portfolios, restricting membership to weak signal firms challenges the robustness of portfolio returns in two ways. First, it challenges the merits of our signal identification criteria, as we argue that measures underlying our MSUE variables should exhibit relatively robust signaling power over time. Second, it provides a setting in which investors are likely to be aware of earnings deficiencies, particularly in recent years. Thus, to the extent MSUE retains signaling power within these samples, investors should have strong incentives to learn about 19 Electronic copy available at: https://ssrn.com/abstract=3348325 signals underlying MSUE so that they may capitalize on information that escapes the notice of earnings-fixated investors. We consider three reporting characteristics that are likely to indicate weak signals in bottom-line earnings: special items, losses, and fourth quarter reporting. We suspect that special item firms have relatively weak signals in bottom-line earnings because special items tend to be both non-recurring in nature and poorly matched to current period revenues.22 Loss firms are likely to report special items with higher frequency, but losses could also arise from investments that are expensed as incurred without a temporal matching to revenues (e.g., R&D expense), which can further weaken earnings signals. Lastly, the fourth fiscal quarter likely provides weak signals for two reasons. One reason is that non-recurring charges tend to be disproportionately recognized in the fourth quarter (Kinney and Trezevant 1997; Burgstahler et al., 2002). Another reason is that the fourth quarter contains many year-end adjustments that “true up” revenue and expense items to their annual figures (e.g., adjustments to effective tax rate estimates that impact income tax expense). As these adjustments may reflect corrections of interim period estimation errors, they are not always informative about future profitability. Table V reports SUE and MSUE hedge portfolio returns for weak signal firms. In all three panels, we see that all of our MSUE portfolios continue to generate economically and statistically significant returns in the most recent subperiod (2000-2017) when bottom-line signals should be weakest and when investor learning should be highest. Thus, even in cases where the incremental benefits to extracting signals underlying MSUE should be highest, there appears to remain significant mispricing of such signals. By contrast, we find that SUE portfolios in all three weak signal cases fail to generate economically or statistically significant returns in the 2000-2017 subperiod. While we caution 22 Cready et al. (2010) and Cready et al. (2012) show that certain special item charges (e.g., restructurings) have recurring effects on firms’ earnings streams, suggesting that not all special items are irrelevant for firm value. This possibility should bias against finding return disparities across our weak signal portfolios. 20 Electronic copy available at: https://ssrn.com/abstract=3348325 readers that interpreting rates of return attenuation may be problematic in this table because weak signals are cross-sectional characteristics, we note that SUE returns are generally much lower in each subperiod relative to the full sample. This finding highlights the importance of the signaling power of SUE firms to the SUE portfolio’s performance at any given time. Thus, as firms with weak signal characteristics are becoming more commonplace in large sample settings, it is increasingly important to control for signal weakening when assessing the efficiency implications of earnings-based return attenuation evidence.23 Figure 2 plots the widening hedge return differential between the SUE and MSUE portfolios for non-overlapping subsamples (Panels A.1-A.3) and weak signal subsamples (Panels B-D) starting from 2000. These plots reinforce two key takeaway points from our subsample analysis in sections 3.3 and 3.4. First, robust predictive signals underlying our MSUE variables continue to generate significant efficiency gaps, even for firms where investors are most likely aware of deficiencies in bottom-line earnings. Second, it is increasingly important to consider the role of a variable’s signaling power on the performance of portfolios formed on that variable, as a failure to control for signal weakening can lead to overstated efficiency gains in the variable’s pricing over time. 3.5 Excess Return Predictability of Levels-Based Variables Over Time Next, we repeat our hedge portfolio analyses using quarterly levels of bottom line earnings and less susceptible measures. Since predictive signals and susceptibility to signal weakening are properties of our test variables’ underlying performance measures, return patterns that are driven by these properties should arise when using levels specifications of our measures. Thus, portfolio analysis based on levels of our measures provides a natural out-of-sample test of whether our return 23 In the full sample, we find (untabulated) that the percentage of firms reporting special items (losses) increases from 14.6% to 36.2% (24.7% to 30.9%) from sub-period one to three. Meanwhile, the percentage of firms reporting special items or losses in the fourth quarter increases from 39.3% to 60.51%. 21 Electronic copy available at: https://ssrn.com/abstract=3348325 patterns can be explained by varying susceptibilities to signal weakening. Moreover, levels tests should mitigate concerns that our main results stem from the nature of our variable construction (e.g., earnings expectation models underlying SUE and MSUE are misspecified).24 Table VI presents the results of our hedge portfolio analysis using levels-based variables constructed from bottom-line and less susceptible performance measures. Similar to our changesbased results from Table III, we observe varying degrees of return attenuation for both bottom-line and less susceptible portfolios. For example, the rate of return attenuation for the bottom-line earnings portfolio is 41.6%, while rates of return attenuation for gross profit, operating profit, and earnings before one-time items are 21.7%, 24.4%, and 36.6%, respectively (i.e., 12% to 48% smaller than the bottom-line rate). These return patterns are generally consistent with our changesbased results in Table III. One important factor to consider when interpreting the bottom-line earnings portfolio’s performance is that the return predictability of bottom-line earnings levels was documented toward the tail end of our sample period (Balakrishnan et al., 2010). Thus, learning from academic research is likely to play a lesser role in bottom-line earnings’ return attenuation relative to the changes-based SUE case in Table III. 3.6 Robustness Checks In this subsection, we test alternative explanations and perform sensitivity checks for our main hedge portfolio results. First, we run Fama-MacBeth regressions to test whether our main findings are robust to controls for known risk characteristics. Table VII presents the results of this analysis. Our dependent variable is three-month buy-and-hold raw returns, and our regressors are rank-transformations of SUE, MSUE, Size, BM, and MOM. Note that the rank-transformations restrict our regressors to vary from 0 to 1, so we can interpret coefficients as three-month buy-andhold hedge returns on the high-low portfolio for each variable. The results show that the return 24 Balakrishnan et al. (2010), Novy-Marx (2013), and Ball et al. (2015) show that levels of both bottom line and less susceptible measures predict cross-sectional excess returns over long horizons. 22 Electronic copy available at: https://ssrn.com/abstract=3348325 predictability of our three MSUE variables continues to hold even after controlling for SUE and known risk factors (annualized MSUE returns range from 5.4% to 7.7%). Moreover, MSUE coefficients exhibit either mild declines (-13.0% and -11.9% for MSUE1 and MSUE3, respectively) or increases (+9.5% for MSUE2) over time, while SUE coefficients decline in all three cases (51.2% to 66.9%), with an insignificant loading in in the 2000-2017 MSUE3 case. Next, an alternative explanation of our main hedge portfolio results is that firms selected into our MSUE portfolios tend to exhibit lower earnings volatility over time than SUE portfolio firms. Therefore, the widening excess return gap that we document merely reflects lower ex ante earning volatility for MSUE portfolio firms (Cao and Narayanamoorthy 2012). To rule out this explanation, we partition our sample into low, medium, and high volatility groups based on the variance of each firm’s bottom line earnings (scaled by average assets) over the previous eight quarters. In untabulated analysis, we find that SUE hedge returns generally diminish as ex ante earning volatility increases, consistent with the findings of Cao and Narayanamoorthy (2012). By contrast, MSUE hedge returns generally increase with increasing ex ante earnings volatility with significant returns in each tercile in each subperiod. Therefore, it is unlikely that ex ante earnings volatility differences between portfolios explain our results. Last, we perform a series of sensitivity checks for our main hedge portfolio analysis and tabulate the results in Table VIII. In Panel A, we form portfolios using a constant sample of firms (i.e., firms that have non-missing values for all test variables). Our results continue to hold using a constant sample, suggesting that differences across SUE and MSUE samples (e.g., information environment) are unlikely to drive our results. In Panel B, we re-partition our sample into four equally spaced subperiods and continue to see significant excess returns for all three MSUE variables across all subperiods, whereas SUE’s return predictability is insignificant in the 20092017 subperiod. In Panel C, we adjust the portfolio-holding period to the three days surrounding next quarter’s earnings announcement and use cumulative market-adjusted returns over that window. The results show that returns generally decline over time; however, by the post-2000 23 Electronic copy available at: https://ssrn.com/abstract=3348325 period, SUE’s hedge return is insignificant whereas MSUE returns remain highly significant. Finally, in untabulated analyses, we experiment with varying portfolio formation dates (e.g., one month after earnings announcement) and portfolio holding windows (e.g., 3 to 12 months). We find that the tenor of our results holds under various formation date and holding window specifications. Therefore, our results cannot be fully explained by earlier dissemination of bottom line earnings news over our sample period.25 4. Conclusion In this paper, we find that variables constructed from performance measures with low susceptibility to the signal-weakening effects of changing earnings properties continue to predict significant stock returns with only mild evidence of return attenuation over time. Excess return gaps between bottom-line and less susceptible variables widen within sub-samples of nonoverlapping firms and firms with weak bottom-line signals, consistent with variation in portfolio signaling power driving return patterns. We also find that our main return patterns show up in levels specifications, reinforcing the notion that return patterns are driven by profitability signals underlying our predictor variables rather than the variables themselves. Our results are insensitive to sample selection, subperiod, and holding window choices as well as controls for risk and ex ante earnings volatility. Overall, our results suggest that while investor learning is apparent in the data, learning efforts to date have been suboptimal at closing profitability-driven efficiency gaps. In particular, significant inefficiencies remain in the pricing of earnings components with reliable serial In other untabulated analysis, we examined temporal patterns in each news variable’s earnings response coefficient (ERC) by regressing cumulative market-adjusted stock returns over the 3-day earnings announcement window on each news variable (along with conventional control variables) within each subperiod examined in our paper (i.e., 1979-1989, 1990-1999, and 2000-2017). We found that ERCs increase over our sample period for all four news variables, with all three MSUE ERCs increasing at a higher rate (57.6% increase on average from the first subperiod to the third subperiod) than the SUE ERC (32.3% increase from the first subperiod to the third subperiod). These results suggest that the stronger declines of SUE’s return predictability that we document (relative to the declines of the three MSUE variables) are not driven by stronger earnings announcement window responses to SUE over time. 25 24 Electronic copy available at: https://ssrn.com/abstract=3348325 properties (i.e., robust predictive signals). Such mispricing has become less apparent over time when studying earnings-based return predictors because evolving earning properties weaken signals that helped to generate returns in the first place. Future research should consider whether evolving earnings properties have intertemporal ramifications for other documented accounting-based return predictors, such as accruals. For example, Bushman et al. (2016) speculate that the declining accrual/cash-flow relation over time may explain, in part or in full, the decline in the accrual anomaly in recent years (e.g., Green et al., 2011). In addition, future research should examine potential explanations for why learning does not occur in the context of accounting-based return predictors. To this end, researchers might consider the applicability of learning theories advanced outside of the accounting literature, including theories of situated learning (Lave and Wenger 1991; Anderson et al., 1996) and categorical learning (Kruschke and Johansen 1999; Peng and Xiong 2006). 25 Electronic copy available at: https://ssrn.com/abstract=3348325 Appendix Variable Definitions Variable Name MSUE1 MSUE2 Definition = = Standardized unexpected gross profit, calculated as quarterly gross profit per share minus expected gross profit per share, scaled by the standard deviation of quarterly gross profit growth over the previous eight quarters. Gross profit is defined as revenue minus cost of goods sold. Expected gross profit follows a seasonal random walk with drift. The drift term is the average of quarterly gross profit growth over the previous eight quarters. Standardized unexpected operating profit, calculated as quarterly operating profit per share minus expected operating profit per share, scaled by the standard deviation of quarterly operating profit growth over the previous eight quarters. Operating profit is defined as revenue minus cost of goods sold and selling, general and administrative expense. Expected operating profit follows a seasonal random walk with drift. The drift term is the average of quarterly operating profit growth over the previous eight quarters. = Standardized unexpected earnings before one-time items, calculated as quarterly earnings per share adjusted for special items minus expected earnings per share adjusted for special items, scaled by the standard deviation of quarterly earnings growth adjusted for special items over the previous eight quarters. Expected earnings adjusted for special items follows a seasonal random walk with drift. The drift term is the average of quarterly earnings growth adjusted for special items over the previous eight quarters. SUE = Standardized unexpected earnings, calculated as quarterly earnings per share minus expected earnings per share scaled by the standard deviation of quarterly earnings growth over the previous eight quarters, as in Jegadeesh and Livnat (2006). Expected earnings are assumed to follow a seasonal random walk with drift. The drift term is the average of quarterly earnings growth over the previous eight quarters. Size = Firm size, calculated as the natural log of the market capitalization as of the end of the most recent fiscal quarter for which data are available (in millions). MSUE3 26 Electronic copy available at: https://ssrn.com/abstract=3348325 BM = Book-to-market ratio, calculated as book value of equity divided by market value of equity at the end of the most recent fiscal quarter for which data are available. MOM = The buy-and-hold six-month stock return ending one month prior to the portfolio formation date. R_MSUE1 = The decile ranking of MSUE1 based on the distribution for each calendar quarter. R_MSUE2 = The decile ranking of MSUE2 based on the distribution for each calendar quarter. R_MSUE3 = The decile ranking of MSUE3 based on the distribution for each calendar quarter. R_SUE = The decile ranking of SUE based on the distribution for each calendar quarter. R_Size = The decile ranking of Size based on the distribution for each calendar quarter. R_BM = The decile ranking of BM based on the distribution for each calendar quarter. R_MOM = The decile ranking of MOM based on the distribution for each calendar quarter. = Size-adjusted return over the three-month period beginning in the first month of the calendar quarter that is at least three months subsequent to fiscal quarter-end. The methodology to construct size-adjusted portfolios is based on Lyon et al. (1999). ADJ_RET 27 Electronic copy available at: https://ssrn.com/abstract=3348325 REFERENCES Abarbanell, J. S., and B. J. Bushee. 1997. Fundamental Analysis, Future Earnings, and Stock Prices. Journal of Accounting Research 35 (1): 1–24. Akbas, F., Jiang, C., and P. D. Koch. 2017. The trend in firm profitability and the cross-section of stock returns. The Accounting Review 92(5): 1-32. Anderson, J. R., L. M. Reder, and H. A. Simon. 1996. Situated learning and education. Educational Researcher 25 (4): 5–11. Balakrishnan, K., E. Bartov, and L. Faurel. 2010. Post loss/profit announcement drift. Journal of Accounting and Economics 50 (1): 20–41. Ball, R., and P. Brown. 1968. An Empirical Evaluation of Accounting Income Numbers. Journal of Accounting Research 6 (2): 159–178. Ball, R., J. Gerakos, and J. T. Linnainmaa. 2015. Deflating Profitability. Journal of Financial Economics 117 (2): 225-248. Bernard, V. L., and J. K. Thomas. 1990. Evidence That Stock Prices Do Not Fully Reflect the Implications of Current Earnings for Future Earnings. Journal of Accounting and Economics 13: 305–340. Bernard, V. L., and J. K. Thomas. 1989. Post-Earnings-Announcement Drift: Delayed Price Response or Risk Premium? Journal of Accounting Research 27 (3): 1–36. Bhattacharya, N., E. L. Black, T. E. Christensen, and R. D. Mergenthaler. 2004. Empirical Evidence on Recent Trends in Pro Forma Reporting. Accounting Horizons 18 (1): 27–48. Bradshaw, M. T., and R. G. Sloan. 2002. GAAP versus The Street : An Empirical Assessment of Two Alternative Definitions of Earnings. Journal of Accounting Research 40 (1): 41–66. Burgstahler, D., J. Jiambalvo, and T. Shevlin. 2002. Do Stock Prices Fully Reflect the Implications of Special Items for Future Earnings? Journal of Accounting Research 40 (3): 585–612. Bushman, R. M., A. Lerman, and X. F. Zhang. 2016. The Changing Landscape of Accrual Accounting. Journal of Accounting Research 54 (1): 41-78. Cao, S. S., and G. S. Narayanamoorthy. 2012. Earnings volatility, post-earnings announcement drift, and trading frictions. Journal of Accounting Research 50 (1): 41–74. Chordia, T., A. Subrahmanyam, and Q. Tong. 2014. Have capital market anomalies attenuated in the recent era of high liquidity and trading activity? Journal of Accounting and Economics 58 (1): 41–58. 28 Electronic copy available at: https://ssrn.com/abstract=3348325 Collins, D. W., E. L. Maydew, and I. S. Weiss. 1997. Changes in the value-relevance of earnings and book values over the past forty years. Journal of Accounting and Economics 24 (1): 39– 67. Dichev, I. D., and V. W. Tang. 2008. Matching and the changing properties of accounting earnings over the last 40 years. The Accounting Review 83 (6): 1425–1460. Donelson, D. C., R. Jennings, and J. Mclnnis. 2011. Changes over time in the revenue-expense relation: Accounting or economics? The Accounting Review 86 (3): 945–974. Fama, E. F., and K. R. French. 2008. Dissecting anomalies. Journal of Finance 63 (4): 1653– 1678. Fama, E. F., and J. D. MacBeth. 1973. Risk , Return , and Equilibrium : Empirical Tests. The Journal of Political Economy 81 (3): 607–636. Getmansky, M., P. A. Lee, and A. W. Lo. 2015. Hedge Funds: A Dynamic Industry In Transition. Annual Review of Financial Economics 7: 483-577. Givoly, D., and C. Hayn. 2000. The changing time-series properties of earnings, cash flows and accruals: Has financial reporting become more conservative? Journal of Accounting and Economics 29 (3): 287–320. Green, J., J. R. M. Hand, and M. T. Soliman. 2011. Going, Going, Gone? The Apparent Demise of the Accruals Anomaly. Management Science 57 (5): 797–816. Green, J., J. R. M. Hand, and X. F. Zhang. 2013. The Supraview of Return Predictive Signals. Review of Accounting Studies 18 (3): 692–730. Hayn, C. 1995. The information content of losses. Journal of Accounting and Economics 20 (2): 125–153. Hung, M., X. Li, and S. Wang. 2015. Post-earnings-announcement drift in global markets: Evidence from an information shock. The Review of Financial Studies 28 (4): 1242–1283. Jegadeesh, N., and J. Livnat. 2006. Revenue surprises and stock returns. Journal of Accounting and Economics 41 (1-2): 147–171. Johnson, W. B., and W. C. Schwartz Jr. 2000. Evidence that Capital Markets Learn from Academic Research : Earnings Surprises and the Persistence of Post-Announcement Drift. Working Paper. Ke, B., and S. Ramalingegowda. 2005. Do institutional investors exploit the post-earnings announcement drift? Journal of Accounting and Economics 39 (1): 25–53. Kothari, S. P. 2001. Capital markets research in accounting. Journal of Accounting and Economics 31: 105–231. 29 Electronic copy available at: https://ssrn.com/abstract=3348325 Kruschke, J. K., and M. K. Johansen. 1999. A model of probabilistic category learning.pdf. Journal of Experimental Psychology: Learning, Memory, and Cognition (25): 1083–1119. Lee, C. M. C. 2001. Market efficiency and accounting research: A discussion of “capital market research in accounting” by S.P. Kothari. Journal of Accounting and Economics 31 (1-3): 233–253. Lave, J., and E. Wenger,. 1991. Situated Learning: Legitimate peripheral participation. Cambridge, M.A: Cambridge University Press. Lev, B., and R. S. Thiagarajan. 1993. Fundamental Information Analysis. Journal of Accounting Research 31 (2): 190–215. Lev, B., and P. Zarowin. 1999. The boundaries of financial reporting and how to extend them. Journal of Accounting Research 37 (2): 353–385. Li, K. K. 2011. How Well Do Investors Understand Loss Persistence? Review of Accounting Studies 16 (3): 630–667. Lipe, R. C. 1986. The Information Contained in the Components of Earnings. Journal of Accounting Research 24 (3): 37–64. Lo, A. W. 2004. The Adaptive Markets Hypothesis. The Journal of Portfolio Management 30 (5): 15–29. Lo, A. W. 2012. Adaptive markets and the new world order. Financial Analysts Journal 68 (2): 18–29. Lougee, B. A., and C. A. Marquardt. 2004. Earnings Informativeness and Strategic Disclosure: An Empirical Examination of Pro Forma Earnings. Accounting Review 79 (3): 769–795. Lyon, J. D., B. M. Barber, and C.-L. Tsai. 1999. Improved Methods for Tests of Long-Run Abnormal Stock Returns. The Journal of Finance 54 (1): 165–201. McLean, R., and J. Pontiff. 2016. Does Academic Research Destroy Stock Return Predictability? The Journal of Finance 71 (1): 5-32. McVay, S. E. 2006. Earnings Management Using Classification Shifting : An Examination of Core Earnings and Special Items. The Accounting Review 81 (3): 501–531. Milian, J. A. 2015. Unsophisticated Arbitrageurs and Market Efficiency: Overreacting to a History of Underreaction? Journal of Accounting Research 53 (1): 175–220. Novy-Marx, R. 2013. The other side of value: The gross profitability premium. Journal of Financial Economics 108 (1): 1–28. Ou, J. A., and S. H. Penman. 1989. Financial Statement Analysis and the Prediction of Stock Returns. Journal of Accounting and Economics 11: 295–329. 30 Electronic copy available at: https://ssrn.com/abstract=3348325 Pedersen, L. H. 2015. Efficiently Inefficient: How Smart Money Invests and Market Prices Are Determined. Princeton University Press. Peng, L., and W. Xiong. 2006. Investor attention, overconfidence and category learning. Journal of Financial Economics 80 (3): 563–602. Penman, S. H., and X.-J. Zhang. 2002. Accounting Conservatism, the Quality of Earnings, and Stock Returns. The Accounting Review 77 (2): 237–264. Piotroski, J. D. 2000. Value Investing: The Use of Historical Financial Statement Information to Separate Winners from Losers. Journal of Accounting Research 38: 1–41. Richardson, S., I. Tuna, and P. Wysocki. 2010. Accounting anomalies and fundamental analysis: A review of recent research advances. Journal of Accounting and Economics 50 (2-3): 410– 454. Riedl, E. J., and S. Srinivasan. 2010. Signaling firm performance through financial statement presentation: An analysis using special items. Contemporary Accounting Research 27 (1): 289–332. Sloan, R. G. 1996. Stock Prices in Fully Reflect Information Accruals and Flows About Future Earnings? The Accounting Review 71 (3): 289–315. Srivastava, A. 2014. Why have measures of earnings quality changed over time? Journal of Accounting and Economics 57 (2-3): 196–217. Subrahmanyam, A. 2010. The cross-section of expected stock returns: What have we learnt from the past twenty-five years of research? European Financial Management 16 (1): 27–42. Thomas, J., and F. X. Zhang. 2011. Tax Expense Momentum. Journal of Accounting Research 49 (3): 791–821. Weil, R. L., K. Schipper, and J. Francis. 2012. Financial Accounting: An Introduction to Concepts, Methods and Uses. 14th ed. South-Western College Publication. Zwieg, J. Have Investors Finally Cracked the Stock-Picking Code? Wall Street Journal 2 Mar. 2013: B1. Print. 31 Electronic copy available at: https://ssrn.com/abstract=3348325 Figure 1: Comparison of cumulative hedge returns to MSUE and SUE strategies from 20002017 Figure 1 plots the value of a dollar invested in hedge portfolios formed on MSUE1 (blue line), MSUE2 (red line), MSUE3 (green line), and SUE (purple line). At the end of each calendar quarter, firms are sorted into deciles based on the value of the sorting variable (MSUE1-MSUE3, SUE). We then form zero-investment hedge portfolios by going long (short) in the highest (lowest) decile of MSUE1-MSUE3 and SUE, and we calculate size-adjusted buy-and-hold equalweighted returns for each portfolio over the three months subsequent to the portfolio formation date. MSUE1-MSUE3 and SUE are defined in the Appendix. 7 6 5 4 3 2 1 0 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 MSUE1 MSUE2 MSUE3 SUE 32 Electronic copy available at: https://ssrn.com/abstract=3348325 Figure 2: Comparison of hedge returns to MSUE and SUE strategies using non-overlapping and weak bottom-line signal subsamples from 2000-2017 Figure 2 plots the value of a dollar invested in hedge portfolios formed on MSUE1 (blue line), MSUE2 (red line), MSUE3 (green line), and SUE (purple line) using non-overlapping subsamples (Panels A.1-A.3) and three weak bottom-line signal subsamples (Panels B-D) over the 2000-2017 subperiod. The subsamples in Panels A.1-A.3 comprise extreme decile firm-quarters that do not overlap both portfolios in the panel (e.g., SUE and MSUE1 in Panel A.1). The subsample in Panel B comprises extreme decile firm-quarters that report special items (as identified by Compustat). The subsample in Panel C comprises extreme decile firm-quarters that report negative income before extraordinary items. The subsample in Panel D comprises extreme decile fourth fiscal quarter observations. At the end of each calendar quarter, firms are sorted into deciles based on the value of the sorting variable (i.e., MSUE1, MSUE2, MSUE3, and SUE). We then form zero-investment hedge portfolios by going long (short) in the highest (lowest) decile of each variable, and we calculate size-adjusted buy-and-hold equal-weighted returns for each portfolio following the methodology described in Lyon et al. (1999) over the three months subsequent to the portfolio formation date. MSUE1MSUE3 and SUE are defined in the Appendix. Panel A.1: Non-overlapping MSUE1 vs. SUE 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 MSUE1 SUE 33 Electronic copy available at: https://ssrn.com/abstract=3348325 Panel A.2: Non-overlapping MSUE2 vs. SUE 6 5 4 3 2 1 0 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 MSUE2 SUE Panel A.3: Non-overlapping MSUE3 vs. SUE 5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 MSUE3 SUE 34 Electronic copy available at: https://ssrn.com/abstract=3348325 Panel B: Special items subsample 5 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 MSUE1 MSUE2 MSUE3 SUE Panel C: Loss subsample 4.5 4 3.5 3 2.5 2 1.5 1 0.5 0 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 MSUE1 MSUE2 MSUE3 SUE 35 Electronic copy available at: https://ssrn.com/abstract=3348325 Panel D: Fourth fiscal quarter subsample 2 1.8 1.6 1.4 1.2 1 0.8 0.6 0.4 0.2 0 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 MSUE1 MSUE2 MSUE3 SUE 36 Electronic copy available at: https://ssrn.com/abstract=3348325 Table I: Descriptive statistics Table I reports univariate statistics for our main test variables over a sample period spanning 1979-2017. SUE is standardized unexpected earnings, calculated as quarterly earnings per share minus expected earnings per share scaled by the standard deviation of quarterly earnings growth over the previous eight quarters, as in Jegadeesh and Livnat (2006). Expected earnings are assumed to follow a seasonal random walk with drift. The drift term is the average of quarterly earnings growth over the previous eight quarters. MSUE1, MSUE2, and MSUE3 are defined analogously, with (1) gross profit (revenue minus cost of goods sold), (2) operating profit (revenue minus cost of goods sold and sales, general and administrative expense), and (3) earnings before one-time items (earnings before extraordinary items adjusted for special items) in place of earnings. ADJ_RET is size-adjusted returns over the three-month period beginning in the first month of the calendar quarter that is at least three months subsequent to fiscal quarter-end. The methodology to construct sizeadjusted portfolios is based on Lyon et al. (1999). SIZE is firm size, calculated as the natural log of the market capitalization as of the end of the most recent fiscal quarter for which data are available (in millions). BM is the book-tomarket ratio, calculated as book value of equity divided by market value of equity at the end of the most recent fiscal quarter for which data are available. MOM is the buy-and-hold six-month stock return ending one month prior to the portfolio formation date. Panel A: Descriptive statistics for main test variables Variable Mean Std 25th Pctl 50th Pctl 75th Pctl MSUE1 MSUE2 MSUE3 SUE ADJ_RET SIZE BM MOM 0.051 -0.112 -0.190 -0.366 0.001 5.320 0.676 0.085 4.043 4.052 4.220 4.814 0.276 2.139 0.579 0.422 -2.202 -2.229 -2.068 -1.939 -0.132 3.733 0.303 -0.148 0.144 0.026 -0.024 -0.049 -0.015 5.169 0.552 0.036 2.406 2.239 1.943 1.783 0.103 6.771 0.894 0.235 Panel B: Correlation between return predictor variables (MSUE1, MSUE2, MSUE3 and SUE) and future size-adjusted stock returns (ADJ_RET), by subperiod Variable 1979-1989 1990-1999 2000-2017 MSUE1 MSUE2 MSUE3 SUE 0.056 0.059 0.060 0.057 0.034 0.039 0.040 0.033 0.024 0.029 0.029 0.011 Panel C: Overlap between SUE and MSUE hedge portfolios, by subperiod Variables 1979-1989 1990-1999 2000-2017 MSUE1 44.37% 40.82% 39.28% MSUE2 MSUE3 56.05% 81.30% 53.24% 67.96% 48.57% 64.55% 37 Electronic copy available at: https://ssrn.com/abstract=3348325 Table II: Fama-MacBeth regressions of future MSUE or SUE on current MSUE and SUE Table II reports Fama-Macbeth multivariate regression results for estimates of one-quarter ahead MSUE or SUE on current quarter MSUE and SUE over three subperiods (1979-1989, 1990-1999, and 2000-2017). In each subperiod column, results for when the dependent variable is one-quarter-ahead MSUE are reported at left, while results for when the dependent variable is one-quarter-ahead SUE are reported at right. Panel A.1 reports results for when MSUE is defined using gross profit (i.e., MSUE1). Panel A.2 reports results for when MSUE is defined using operating profit (i.e., MSUE2). Panel A.3 reports results for when MSUE is defined using earnings before one-time items (i.e., MSUE3). MSUE1-MSUE3 and SUE are defined in the Appendix. Panel A.1 Regression of MSUE1t+1 or SUEt+1 on MSUE1t and SUEt 1979-1989 MSUE1t+1 SUEt+1 MSUE1 SUE Adj R2 1990-1999 MSUE1t+1 SUEt+1 0.456 (44.86) -0.018 0.120 (18.43) 0.246 0.460 (44.20) (-3.29) 0.197 2000-2017 MSUE1t+1 SUEt+1 -0.032 0.129 (20.69) 0.192 0.484 (54.31) -0.037 0.134 (33.87) 0.188 (19.03) (-11.33) (15.60) (-10.80) (0.07) 0.101 0.195 0.064 0.221 0.073 Panel A.2 Regression of MSUE2t+1 or SUEt+1 on MSUE2t and SUEt 1979-1989 MSUE2t+1 SUEt+1 MSUE2 SUE Adj R2 1990-1999 MSUE2t+1 SUEt+1 0.421 (35.85) -0.013 0.193 (22.68) 0.201 0.434 (36.63) (-1.69) 0.166 2000-2017 MSUE2t+1 SUEt+1 -0.028 0.227 (24.94) 0.134 0.446 (50.54) -0.018 0.218 (37.24) 0.138 (13.98) (-6.55) (12.54) (-3.24) (14.82) 0.117 0.168 0.081 0.190 0.092 Panel A.3 Regression of MSUE3t+1 or SUEt+1 on MSUE3t and SUEt 1979-1989 MSUE3t+1 SUEt+1 MSUE3 SUE 2 Adj R 1990-1999 MSUE3t+1 SUEt+1 0.211 (25.84) 0.131 0.201 (29.45) 0.149 0.236 (25.41) (17.21) 0.108 2000-2017 MSUE3t+1 SUEt+1 0.097 0.233 (27.30) 0.112 0.235 (28.16) 0.088 0.172 (26.57) 0.142 (13.83) (20.50) (14.50) (24.82) (16.82) 0.104 0.097 0.078 0.092 0.076 38 Electronic copy available at: https://ssrn.com/abstract=3348325 Table III: Future size-adjusted stock returns for hedge portfolios formed on MSUE1, MSUE2, MSUE3 and SUE Table III reports three-month buy-and-hold size-adjusted stock returns three-month buy-and-hold size-adjusted stock returns to hedge portfolios formed on extreme deciles of MSUE1, MSUE2, MSUE3 and SUE over three subperiods (1979-1989, 1990-1999, 2000-2017). At the end of each calendar quarter, firms are sorted into deciles based on the value of the sorting variable (MSUE1-MSUE3 or SUE). We then form a zero-investment hedge portfolio for each variable by taking a long (short) position in the variable’s highest (lowest) decile stocks. Size-adjusted buy-and-hold returns are calculated over the three months subsequent to the portfolio formation date and an equal-weighted mean return is computed for each portfolio. We compute Fama-MacBeth t-statistics within each subperiod based on the timeseries distribution of the mean hedge portfolio returns. The % Change column reports the percentage change in the mean portfolio return from the 1979-1989 subperiod to the 2000-2017 subperiod. MSUE1-MSUE3 and SUE are defined in the Appendix. Decile 1979-1989 1990-1999 2000-2017 % Change MSUE1 3.86% 3.51% 2.41% -37.56% (6.96) (7.81) (4.78) 3.76% 4.23% 2.96% (6.98) (7.13) (5.19) 4.22% 4.13% 2.50% (8.50) (9.39) (5.84) 4.07% 3.68% 1.69% (7.70) (8.01) (3.12) MSUE2 MSUE3 SUE -21.28% -40.76% -58.48% 39 Electronic copy available at: https://ssrn.com/abstract=3348325 Table IV: Future size-adjusted stock returns for hedge portfolios formed on MSUE1MSUE3 and SUE using non-overlapping subsamples Table IV reports three-month buy-and-hold size-adjusted stock returns to hedge portfolios formed on extreme deciles of MSUE1, MSUE2, MSUE3 and SUE after removing overlapping firms from each portfolio. Overlapping firms are those with identical calendar quarter extreme decile rankings for MSUE and SUE. Panel A.1 reports results for when MSUE is defined using gross profit (i.e., MSUE1). Panel A.2 reports results for when MSUE is defined using operating profit (i.e., MSUE2). Panel A.3 reports results for when MSUE is defined using earnings before one-time items (i.e., MSUE3). At the end of each calendar quarter, firms are sorted into deciles based on the value of the sorting variable (MSUE1-MSUE3 or SUE). We then form a zero-investment hedge portfolio for each variable by taking a long (short) position in the variable’s highest (lowest) decile stocks. After removing overlapping firm-quarters from each portfolio, we calculate size-adjusted buy-and-hold returns over the three months subsequent to the portfolio formation date and an equalweighted mean return is computed for each portfolio. We compute Fama-MacBeth t-statistics within each subperiod based on the time-series distribution of the mean hedge portfolio returns. The % Change column reports the percentage change in the mean portfolio return from the 1979-1989 subperiod to the 2000-2017 subperiod. MSUE1-MSUE3 and SUE are defined in the Appendix. Panel A.1 MSUE1 and SUE MSUE1 SUE 1979-1989 1990-1999 2000-2017 % Change 2.75% (6.09) 3.41% (7.30) 2.19% (5.05) 2.74% (6.10) 2.01% (4.56) 0.96% (1.96) -26.91% 1979-1989 1990-1999 2000-2017 % Change 2.62% (4.58) 3.78% (6.94) 2.74% (5.26) 2.68% (7.03) 2.56% (4.86) 0.78% (1.65) -2.29% 1979-1989 1990-1999 2000-2017 % Change 2.29% (3.56) 1.62% (2.21) 2.91% (5.78) 1.75% (3.44) 2.36% (5.38) 0.00% (0.40) 3.06% -71.88% Panel A.2 MSUE2 and SUE MSUE2 SUE -79.50% Panel A.3 MSUE3 and SUE MSUE3 SUE -99.85% 40 Electronic copy available at: https://ssrn.com/abstract=3348325 Table V: Future size-adjusted stock returns for hedge portfolios formed on MSUE1-MSUE3 and SUE using subsamples with weak bottom-line earnings signals Table V reports three-month buy-and-hold size-adjusted stock returns to hedge portfolios formed on extreme deciles of MSUE1, MSUE2, MSUE3 and SUE using subsamples of firms with weak bottom-line earnings signaling characteristics. Panel A uses samples of firms reporting special items. Panel B uses samples of firms in which earnings before extraordinary items is below zero. Panel C uses samples of firms reporting results for their fourth fiscal quarter. See Table III for details on how we perform hedge portfolio analysis. The % Change column reports the percentage change in the mean portfolio return from the 1979-1989 subperiod to the 2000-2017 subperiod. MSUE1-MSUE3 and SUE are defined in the Appendix. Panel A: Special item subsamples MSUE1 MSUE2 MSUE3 SUE 1979-1989 1990-1999 2000-2017 % Change 2.63% (2.31) 3.94% (2.89) 2.75% (2.79) 2.82% (3.58) 3.79% (4.59) 3.47% (3.88) 3.57% (5.24) 1.77% (2.97) 2.08% (3.65) 2.38% (3.44) 2.22% (4.46) 0.91% (1.55) -20.91% 1979-1989 1990-1999 2000-2017 % Change 4.45% (4.04) 4.97% (3.24) 1.50% (1.28) 0.41% (0.33) 3.63% (3.21) 3.25% (3.31) 4.03% (5.31) 2.63% (2.77) 1.58% (2.08) 2.71% (3.36) 3.01% (4.24) 1.36% (1.05) -64.49% 1979-1989 1990-1999 2000-2017 % Change 4.58% (3.85) 5.04% (5.07) 4.19% (7.92) 3.88% (5.75) 2.99% (4.11) 3.05% (2.00) 4.24% (5.86) 3.10% (4.33) 3.76% (2.53) 3.56% (1.96) 2.78% (2.24) 0.89% (0.49) -17.90% -39.59% -19.27% -67.62% Panel B: Loss subsamples MSUE1 MSUE2 MSUE3 SUE -45.47% 100.67% 228.50% Panel C: Fourth quarter subsamples MSUE1 MSUE2 MSUE3 SUE -29.37% -33.65% -77.09% 41 Electronic copy available at: https://ssrn.com/abstract=3348325 Table VI: Future size-adjusted stock returns for portfolios formed on levels of earnings (IB) and less susceptible measures (MIB1, MIB2, and MIB3) Table VI reports three-month buy-and-hold size-adjusted stock returns to hedge portfolios formed on extreme deciles of quarterly gross profit (MIB1), operating profit (MIB2), earnings before one-time items (MIB3) and earnings before extraordinary items (IB) over three subperiods (1979-1989, 1990-1999, 2000-2017). At the end of each calendar quarter, firms are sorted into deciles based on the value of the sorting variable (MIB1-MIB3 or IB). We then form a zeroinvestment hedge portfolio for each variable by taking a long (short) position in the variable’s highest (lowest) decile stocks. Size-adjusted buy-and-hold returns are calculated over the three months subsequent to the portfolio formation date and an equal-weighted mean return is computed for each portfolio. We compute Fama-MacBeth t-statistics within each subperiod based on the time-series distribution of the mean hedge portfolio returns. The % Change column reports the percentage change in the mean portfolio return from the 1979-1989 subperiod to the 2000-2017 subperiod. MIB1 is quarterly gross profit, defined as revenue minus cost of goods sold. MIB2 is quarterly operating profit, defined as revenue minus cost of goods sold and sales, general and administrative expense. MIB3 is earnings before one-time items, defined as income before extraordinary items adjusted for special items. IB is quarterly bottom-line earnings, defined as income before extraordinary items. Decile MIB1 MIB2 MIB3 IB 1979-1989 4.84% (7.18) 6.31% (7.95) 6.25% (7.72) 6.08% (7.57) 1990-1999 5.27% (6.13) 5.17% (3.63) 5.10% (3.56) 4.57% (3.29) 2000-2017 3.79% (4.69) 4.77% (3.96) 3.96% (3.22) 3.55% (2.93) % Change -21.69% -24.41% -36.64% -41.61% 42 Electronic copy available at: https://ssrn.com/abstract=3348325 Table VII: Fama-MacBeth regressions of future stock returns on rank-transformed MSUE, SUE and controls for known risk factors Table VII reports results for Fama-MacBeth regressions of three-month buy-and-hold raw stock returns on decile rank transformations of MSUE1, MSUE2, MSUE3, SUE and control variables capturing known risk factors over three subperiods (1979-1989, 1990-1999, 2000-2017). The “R_” prefix denotes decile rank transformations performed each calendar quarter for our variables of interest. The % Change column reports the percentage change in the coefficients on R_MSUE1, R_MSUE2, R_MSUE3, and R_SUE from the 1979-1989 subperiod to the 2000-2017 subperiod. SUE is standardized unexpected earnings, calculated as quarterly earnings per share minus expected earnings per share scaled by the standard deviation of quarterly earnings growth over the previous eight quarters, as in Jegadeesh and Livnat (2006). Expected earnings are assumed to follow a seasonal random walk with drift. The drift term is the average of quarterly earnings growth over the previous eight quarters. MSUE1, MSUE2, and MSUE3 are defined analogously, with (1) gross profit (revenue minus cost of goods sold), (2) operating profit (revenue minus cost of goods sold and sales, general and administrative expense), and (3) earnings before one-time items (earnings before extraordinary items adjusted for special items) in place of earnings. All other variables are defined in Appendix. 1979-1989 R_MSUE1 0.016 0.013 (4.71) (4.95) (3.14) 0.017 (4.46) R_MSUE3 R_SIZE R_BM R_MOM Adj R2 2000-2017 0.015 R_MSUE2 R_SUE 1989-1999 0.016 (3.98) % Change -12.99% 0.018 (3.85) 9.52% 0.022 0.022 0.019 (6.94) (6.48) (5.44) 0.020 0.016 0.011 0.022 0.020 0.016 0.010 0.005 0.004 (5.40) (4.06) (2.37) (6.44) (5.53) (5.55) (3.36) (1.67) (1.26) 0.000 -0.002 0.000 -0.004 -0.004 -0.002 -0.003 -0.003 -0.001 (-0.03) (-0.14) (0.03) (-0.25) (-0.25) (-0.12) (-0.21) (-0.26) (-0.11) 0.002 0.019 0.023 0.012 0.012 0.014 0.023 0.021 0.024 (1.85) (1.72) (2.08) (0.69) (0.73) (0.80) (2.27) (2.25) (2.30) 0.019 0.021 0.021 0.027 0.031 0.027 0.006 0.007 0.007 (1.94) (2.17) (2.25) (2.44) (2.71) (2.45) (0.52) (0.62) (0.59) 0.047 0.044 0.045 0.029 0.026 0.029 0.027 0.026 0.027 43 Electronic copy available at: https://ssrn.com/abstract=3348325 -11.87% -51.24% -66.27% -66.94% Table VIII: Other robustness tests Table VIII reports three-month buy-and-hold size-adjusted stock returns to hedge portfolios formed on extreme deciles of MSUE1, MSUE2, MSUE3 and SUE under various specifications. Panel A uses a constant sample of firms (i.e., firms that have non-missing values for SUE, MSUE1, MSUE2, and MSUE3). Panel B employs four equally spaced subperiods. Panel C uses three-day cumulative market-adjusted stock returns centered over the subsequent quarter’s earnings announcement. See Table III for details on how we perform hedge portfolio analysis. The % Change column reports the percentage change in the mean portfolio return from the 1979-1989 subperiod to the 2000-2017 subperiod. MSUE1MSUE3 and SUE are defined in the Appendix. Panel A: Non-missing values for all variables Decile MSUE1 MSUE2 MSUE3 SUE 1979-1989 3.97% (6.31) 3.76% (6.95) 4.05% (6.99) 3.64% (6.19) 1990-1999 4.11% (7.71) 4.25% (7.10) 4.55% (8.72) 3.71% (7.15) 2000-2017 2.84% (5.03) 2.99% (5.22) 2.63% (6.00) 1.76% (3.05) % Decline -28.46% -20.48% -35.06% -51.65% Panel B: Four equally spaced subperiods MSUE1 MSUE2 MSUE3 SUE 1979-1988 1989-1998 1999-2008 2009-2017 % Change 3.67% (6.19) 3.50% (6.27) 4.13% (7.61) 4.02% (6.97) 3.80% (8.46) 4.58% (7.93) 4.17% (9.77) 3.80% (8.79) 3.19% (4.33) 3.84% (4.37) 3.37% (4.88) 2.64% (3.74) 1.60% (2.72) 1.96% (3.24) 1.80% (4.61) 0.08% (1.05) -56.40% -44.00% -56.42% -98.05% Panel C: Three-day returns around next earnings announcement Decile MSUE1 MSUE2 MSUE3 SUE 1979-1989 0.77% (4.65) 0.81% (3.61) 0.69% (4.80) 0.45% (3.41) 1990-1999 0.56% (4.44) 0.40% (3.15) 0.59% (4.81) 0.15% (0.98) 2000-2017 0.50% (4.41) 0.64% (4.94) 0.45% (3.51) 0.03% (0.20) % Decline -34.68% -20.74% -34.64% -93.80% 44 Electronic copy available at: https://ssrn.com/abstract=3348325