Systematic Noise and News-Driven Return Reversals

Systematic Noise and News-Driven Return Reversals Eric C. So Stanford University Graduate School of Business Sean Wang * University of North Carolina Abstract October 2011 We provide systematic evidence of stock return reversals for trades placed in anticipation of an impending news release. Using quarterly earnings announcements as the information event, we find that pre-announcement returns are robust positive predictors of unexpected earnings, but robust negative predictors of announcement window returns. Selling firms with large positive preannouncement returns and buying firms with large negative pre-announcement returns earns abnormal returns of 108 basis points over the three-day announcement period. Returns are positive in 74 of 79 quarters from 1990-2009 and strongest during periods of bullish investor sentiment. Cross-sectional analyses shows that this return pattern is most pronounced among glitter firms and firms with poor information environments. Taken together, our results are consistent with systematic noise trades influencing pre-announcement prices that are subsequently reversed with the revelation of news. JEL Classifications: G10, G11, G14 * Corresponding author: Sean Wang, sean_wang@kenan-flagler.unc.edu. We thank Robert Bushman, Jennifer Conrad, Joseph Engelberg, Diego Garcia, Chotibhak Jotikasthira, Wayne Landsman, Charles Lee, and UNC-Chapel Hill seminar participants for their helpful comments and suggestions. All remaining errors are our own. Please do not cite without permission.

1. Introduction Since Ball and Brown (1968) and Fama et al. (1969), a substantial literature spanning multiple decades and academic disciplines documents the reaction of stock prices to unexpected information. The positive relation between unexpected earnings news and excess returns is one of the most welldocumented empirical regularities in capital markets research (Beaver, Clarke and Wright, 1979; Beaver, Lambert and Morse, 1980; Lev, 1989). Given this relation, market participants are incentivized to trade in anticipation of these announcements. A related stream of research documents increased trading volume in the pre-announcement period from both informed and uninformed noise traders. These studies show that informed agents tend to trade profitably in advance of upcoming news (Affleck-Graves, Jennings and Mendenhall, 1995; Bhattacharya et al., 2001; Christophe, Ferri and Angel, 2002; Christophe, Ferri and Hsieh, 2008), whereas uninformed traders tend to trade unprofitably. Dey and Radhakrishna (2006) and Taylor (2010) show significant increases in individual investor trading volume just prior to an earnings announcement, with Taylor (2010) and Linnainmaa (2010) documenting that these investors fail to profit from preannouncement trades. If earnings announcements result in the arbitrage of mispricing from uninformed trades placed during the pre-announcement period, then announcement period returns will consist of two components: (1) a reversal component that unravels the price movement emanating from noise trades and (2) a component associated with the earnings news that has not been preempted by the informed agents. We provide evidence in support of these conjectures. We document predictable return patterns around earnings announcements that are consistent with both informed and uninformed traders significantly influencing pre-announcement prices. Specifically, our results demonstrate that pre-announcement returns foreshadow earnings surprises, consistent with these returns reflecting informed trade. However, we also demonstrate that pre-announcement returns 1

tend to reverse during earnings announcements, consistent with pre-announcement returns reflecting the actions of systematic noise traders. Our finding that pre-announcement period returns positively predict earnings surprises and negatively predicts announcement period returns is particularly paradoxical given the commonly assumed equivalence between these two measures. The first half of our paper is dedicated to documenting this paradox. First, we demonstrate that pre-announcement returns (PAR) from days t-5 to t-3 prior to an earnings announcement reflect informed trading regarding impending earnings news. High (low) PAR is predictive of positive (negative) earnings announcement news measured as either standardized unexpected earnings or analyst-based surprises, even after controlling for other predictive variables established in prior research, including the prior earnings surprise (Bernard and Thomas, 1990), changes in firms financial condition (Piotroski, 2000), and the reversal of accruals (Sloan, 1996). Next, we demonstrate that a significant portion of pre-announcement returns predictably reverses during earnings announcements, resulting in low PAR firms significantly outperforming high PAR firms. A hedge strategy that shorts firms in the highest PAR quintile portfolio and holds firms in the lowest PAR quintile earns four-factor risk-adjusted abnormal returns (Carhart, 1997) in excess of one percent during a three-day earnings announcement window. Returns from this Pre-Earnings Announcement Reversal hedge strategy, henceforth PEAR, are not only economically meaningful, 2 but are also stable across the time-series. Using data from 183,228 firm-quarters spanning 1990-2009, PEAR yields positive hedge returns in all but five quarters across this twenty-year time span. Regarding robustness of the strategy, although most anomalies appear to increase in magnitude among smaller, illiquid firms with impoverished information environments (Fama, 1998), the predictability of our return reversal remains both 2 Fama and French (2008) dissect various anomalies within the literature. They find the most pervasive anomalies to be momentum, net stock issuances, and accruals. As a comparison, monthly hedge returns from these strategies over the 1963-2005 time period return 0.74%, 0.54%, and 0.54%, respectively 2

statistically and economically significant for large-cap firms. To mitigate concerns that PEAR returns reflect compensation for risk, we show that our results are robust to standard factor adjustments and risk controls in cross-sectional regressions. Furthermore, because our strategy involves placing trades against the flow of order demand, our evidence of return predictability is unlikely to be driven by illiquidity. Additionally, we show that our findings are robust to the use of predicted earnings announcement dates using the methodology from Cohen et al. (2007). In the second half of the paper, we outline and test our hypothesis that predictable return reversals during earnings announcements emanate from systematic noise trades placed during the pre-announcement period. To test this hypothesis, we exploit intertemporal variation in investor sentiment. Kumar and Lee (2006) show that retail investor trades become more highly correlated during periods of high sentiment, consistent with sentiment acting as a proxy for the influence of systematic noise trading on prices. We use an investor sentiment measure from Baker and Wurgler (2006) as a proxy for the impact of uninformed agents on the PEAR strategy returns, and find that the correlation between the sentiment index and quarterly strategy returns is 0.36. We find that when sentiment is high, PAR becomes less predictive of earnings while simultaneously becoming more predictive of announcement-window returns. The magnitude of the announcement return reversal is twice as large in high sentiment regimes relative to low investor sentiment regimes. Finally, we examine firms which are most prone to being affected by noise traders limited attention under high sentiment periods (Barber and Odean, 2006), and find that when sentiment is bullish, the magnitude of PEAR strategy returns strongest for glamour stocks and stocks with abnormally large levels of trading volume. Together, the evidence is consistent with trades during the preannouncement period being derived from a blend of informed and systematic noise traders who act 3

in anticipation of earnings news, and higher levels of sentiment resulting in a stronger influence of systematic noise on the PAR signal. 3 We survey the existing literature and find PEAR to be distinct from other previously documented short-term anomalies. For example, although Jegadeesh (1990), Lehmann (1990), and Gutierrez and Kelley (2008) document negative autocorrelation for weekly returns, we compare the magnitude of the stock price reversals for the same firm during non-announcing weeks with the magnitude of the PEAR strategy, and find that reversals are five times larger during the earnings announcement period. This suggests that a significant portion of the market reaction reflects the reversal of noise trades placed in the pre-announcement period. Other studies that focus on earnings announcements reversals include Trueman, Wong and Zhang (2003) and Aboody, Lehavy and Trueman (2010). These studies show reversals for a small sample of firms before and after the earnings announcement for high-tech Internet firms during the NASDAQ bubble, and for firms in the 99 th percentile of high momentum stock returns over a 52-week period, respectively. 4 However, whereas we document a robust relation between PAR and earnings surprises, Trueman, Wong and Zhang (2003, p. 250) note that neither the reported earnings surprise nor the reported revenue surprise is significantly associated with the price run-up in advance of the earnings announcement. Moreover, these studies only show evidence of negative reversals following positive run-ups, implying that short-selling constraints could pose execution issues, while the PEAR strategy generates abnormal returns on both sides of the hedge portfolio even after controlling for prior stock return momentum. 3 The heightened impact of systematic noise traders on stock prices when sentiment is high could result from either (1) a higher proportion of noise traders relative to informed traders, i.e., individual investors become more interested in making stock market trades when investor sentiment is bullish, or (2) a higher correlation of noise traders actions, which magnifies the impact on prices of noise traders actions. 4 Another distinguishing feature between these studies and ours is that whereas the PEAR strategy predicts reversals at the earnings event date, Trueman et al. (2003) and Aboody et al. (2010) calculate the reversal in the post earnings announcement period. 4

In summary, although both academic researchers and practitioners generally assume earnings surprises and stock returns go hand in hand, our study is the first to our knowledge to document a variable, i.e., pre-announcement returns, which serves both as a proxy for informed trading that predicts earnings surprises and a link to predictable and anomalous patterns of return reversals that appear to be driven by systematic noise traders. Our results contribute to the literature regarding announcement anomalies, return reversals and earnings informativeness, and should also be of interest to practitioners interested in predicting returns around earnings announcement dates. The remainder of this study is organized as follows. In Section 2, we briefly discuss sample characteristics of our firms, and the methodology by which we measure the key variables of interest. In Section 3, we document reversals of pre-earnings announcement returns when information is revealed at the earnings announcement. Section 4 describes the reasoning underlying a noise-trader based explanation for the reversals and provides evidence in support of this reasoning. In doing so, we investigate and discuss alternative explanations or mechanisms for the reversals. Section 5 discusses a number of alternative tests, and provides robustness tests regarding the implementability of the PEAR strategy. Section 6 concludes. 2. Sample Selection and Variable Measurement Our analyses examine the link between pre-announcement returns, PAR, i.e., stock returns just prior to quarterly earnings announcements, and news at the earnings announcement date. As a consequence, it is important to correctly identify the earnings announcement date. Because of problems prior research has identified with Compustat earnings announcement dates, we follow Dellavigna and Pollet (2009) and compare I/B/E/S and Compustat announcement dates and assume that the earlier date is the announcement date. 5 We also use the I/B/E/S time stamp to 5 Dellavigna and Pollet (2009) compares this assumed date with the newswire time stamp and reports that the assumed date is correct more than 95% of the time. 5

determine whether the announcement occurred after the market close. We adjust the announcement date one trading day forward for announcements that occur after the market close. We require firms to have earnings announcement dates on I/B/E/S, and require six-months of daily returns prior to each earnings announcement to calculate return momentum. We also eliminate firms with a stock price below $1 to avoid potential biases resulting from the bid-ask bounce (Conrad and Kaul, 1993). 6 Our final sample consists of 183,228 firm-quarters spanning 1990-2009. Throughout the analysis, we define PAR as the market-adjusted return over a three-day window from the start of day t-5 to the end of day t-3, where t denotes the earnings announcement date. To control for distributional changes in PAR across our sample period, we assign firms to quintiles within each calendar quarter. To avoid look-ahead bias, we form PAR quintiles using the realized distribution from the prior calendar quarter. When PAR is used within a multivariate framework, we rescale it between 0 and 1 so that the magnitude of the coefficients can be easily interpreted as hedge portfolio returns. We measure three primary variables at the earnings announcement date. The first, standardized unexpected earnings, SUE, is a measure of earnings surprise based on a seasonal random walk and is defined as the realized EPS minus EPS from four quarters prior, divided by its standard deviation over the prior eight quarters. The second is the analyst-based earnings surprise, SURPRISE, defined as actual earnings minus the consensus estimate immediately prior to the announcement, scaled by beginning of quarter price. Finally, RET, equals the cumulative marketadjusted return during the earnings announcement window from days t-1 to t+1. Table 1 presents sample summary statistics. Panel A reveals that although the mean values of SUE and SURPRISE are negative, their median values are positive. This is consistent with prior research documenting managers desire to meet or beat the prior year s earnings (Burgstahler and 6 In untabulated results, we find qualitatively identical evidence of return prediction when removing firms with a stock price below $5. 6

Dichev, 1997; Degeorge, Patel and Zeckhauser, 1999). Panel B shows that although PAR is positively correlated with SUE ( = 0.024) and SURPRISE ( = 0.015), which is consistent with prior research (Beaver, Lambert and Morse, 1980; Foster, Olsen and Shevlin, 1984), it is negatively correlated with RET ( = -0.054). The combination of these marginal correlations is particularly interesting given the myriad of prior studies that document a strongly positive correlation between unexpected news and stock returns ( SUE,RET = 0.101, SURPRISE,RET = 0.118). We further investigate these relations in the following section. 3. Predictability of the PAR signal: 3.1. Earnings Surprises To the extent that pre-announcement returns reflect informed trades, all else equal, higher values of PAR should portend higher values of SUE and SURPRISE. We test this conjecture by estimating the regression of our proxies for earnings news on PAR and controls for firmcharacteristics that prior research has established as predictors of earnings surprise. Specifically, we estimate the following regression: {SUE i,q, SURPRISE i,q }= λ 0 + λ 1 PAR q + λ 2 LBM q + λ 3 SIZE q + λ 4 MOMEN q + λ 5 ACC q + λ 6 FScore q + λ 7 PastSURPRISE q + q, (1) where SIZE and LBM are the log of market capitalization and log of one plus the book-to-market ratio, respectively. MOMEN is the cumulative market-adjusted return over the sixty trading days ending on t-10. ACC is defined as the total accruals scaled by total assets. Total accruals are calculated as the change in current assets [Compustat item ACT] plus the change in debt in current liabilities [Compustat item DCL] minus the change in cash and short-term investments [Compustat item CHE] and minus the change in current liabilities [Compustat item CLI]. PastSURPRISE equals SURPRISE corresponding to prior quarter s earnings announcement. FScore is the firm s 7

fundamental analysis score from the prior quarter, as calculated in Piotroski (2000). Throughout the analysis, reported t-statistics are based on two-way cluster robust standard errors, clustered by firm and quarter to control for time-series and cross-sectional correlation in the error terms (Gow, Ormazabal and Taylor, 2010). Table 2 presents the findings from various specifications of equation (1) with and without control variables. The results in columns (1) through (3) highlight the predictive ability of PAR for SUEs, consistent with the positive correlation documented in Table 1. In particular, the coefficient on PAR ranges from 0.178 to 0.161 with t-statistics in excess of eight. In columns (4) through (6), we also report regression results where the dependent variable is SURPRISE and find that PAR is also predictive of SURPRISE, with t-statistics in excess of 6. Unscaling SURPRISE by multiplying by stock price indicates that high PAR firms have analyst-based earnings surprises that are approximately 2.93 cents per share higher than low PAR firms. Across both measures of earnings news, firm-characteristics and earnings prediction variables fail to subsume the predictive power of PAR for the impending earnings news. Given similar findings across SUE and SURPRISE, the remaining analysis focuses on SUE as the main measure of earnings news. 3.2. Earnings Announcement Returns Figure 1 plots market-adjusted returns calculated during the three-day announcement window. The graph indicates a monotonic relation across the PAR quintiles, and illustrates that holding stocks in quintile 1 (low PAR) and shorting stocks in quintile 5 (high PAR) earns a three-day announcement hedge return of 1.083%. Panel A of Table 3 presents mean pre-announcement returns and mean announcement and post-announcement returns over a variety of investment horizons across PAR quintiles. Results document that the PAR reversal is concentrated during the announcement period. We find that over three-quarters of the reversal occurs from day t-1 to t+1 8

and little evidence of the reversal extending beyond the announcement period. This finding appears consistent with the market correcting trades based on erroneous expectations about future cash flows that were placed during the pre-announcement period. We next investigate the magnitude of the hedge returns after controlling for Fama-French and momentum risk factors (Carhart, 1997), as well as other well-established predictors of earnings announcement returns based on prior anomalies. Panels B and C of Table 3 show that the PAR reversal remains robust and statistically significant in the presence of these control variables. The findings in Panel B indicate that PEAR earns risk-adjusted abnormal returns in excess of 1.01% after controlling for Fama-French factors and momentum. Similarly, the findings in Panel C indicate excess returns associated with PAR exceed 0.93% after controlling for book-to-market, size, and momentum, as well as proxies for expected mispricing related to prior accruals, financial performance, and earnings surprises. Most notably, column (4) of Panel C indicates that the observed return reversals are not subsumed by this full array of controls, and that a PEAR strategy still earns excess returns of 0.96% during the announcement period. Because prior literature provides evidence of short-term negative auto-correlation in returns (Jegadeesh, 1990; Lehmann, 1990; and Gutierrez and Kelley, 2008), it is possible the PEAR result simply captures this previously documented effect. To test whether this is the case, we compare the magnitude of PAR reversals with non-earnings announcement return reversals by executing the PEAR strategy around pseudo earnings announcement dates. Specifically, we calculate pseudo announcement dates by subtracting a random number of trading days from the actual announcement date. The randomly selected numbers are drawn from a uniform distribution spanning 10 to 40. For example, if the random draw for a firm is 15, the firm's pseudo date is set to 15 trading days prior to the firm's actual announcement date. We set the support of the uniform 9

distribution from 10 to 40 to ensure that sufficient separation between the pseudo and actual announcement date. Table 4 reports the results from repeating the analyses in Table 3 using pseudo earnings announcement dates. The Table 4 findings distinguish PEAR from these more general reversal anomalies in terms of both the magnitude and the timing of the return reversals. First, the reversal around pseudo-announcement date, k, is less than one-fifth of the magnitude that we observe around actual earnings announcements (0.21% vs. 1.08%). Second, whereas the large majority of the return reversal in Table 3 occur during the earnings announcement window, i.e., days (-1,+1), this is no longer the case in Table 4, as nearly 60% of the return reversal calculated from day k-1 to k+5 comes during the days (+2, +5) window. Overall, these results suggest that the mechanism of the PEAR phenomenon may be driven by earnings news that unravels erroneous cash flow expectations impounded into prices during the pre-announcement period. 4. Explaining PEAR The aforementioned paradox regarding PEAR in Section 3 documents pre-announcement returns being positively predictive of earnings surprise, but negatively predictive of announcement returns, where firms in low (high) PAR quintiles earn positive (negative) abnormal risk-adjusted returns. Although the prior result requires the presence of informed agents who trade on the forthcoming earnings news, the latter can be consistent with two explanations. The first explanation, the systematic noise hypothesis, is consistent with traders in the pre-announcement period comprising both informed and systematic noise traders. In this scenario, informed trades placed just prior to the earnings announcement result in SUE predictability, and systematic noise trades placed within the same period create spreads within the cross-section of pre-announcement returns that are subsequently reversed at the announcement date. In the second explanation, the overshooting 10

hypothesis, traders in the PAR period have knowledge regarding the sign of the earnings news, but consistently overestimate the magnitude of the news on stock prices, thereby resulting in consistent reversals when the information is revealed. In the following two sections we present results from analyses designed to distinguish between these two explanations. 4.1 Systematic Noise: Theory and Evidence Figure A: Schematic of Systematic Noise Explanation We first investigate the systematic noise explanation. This explanation is illustrated by the diagram in Figure A relating to stock price movements for a given firm surrounding its earnings announcement with varying levels of systematic noise. Under this hypothesis, traders in the preannouncement period constitute both informed and systematic noise traders. Absent systematic noise and restrictions on informed trading, as the blue (center) line illustrates, trades placed during 11

the pre-announcement period perfectly preempt the effects of SUE on stock prices, resulting in no further price movements when the news is released. 7 When systematic noise traders are introduced in a low systematic noise trading regime, as the purple (middle) set of lines indicate, in the preannouncement period stock price moves away from fundamental value, i.e., the price implied by the information contained in the earnings announcement, although the direction is unpredictable. When earnings are announced, price adjusts to fundamental value. When the influence of systematic noise becomes larger, perhaps from the increased correlation of uninformed traders, e.g. in a high systematic noise regime, as the green (outer-most) set of lines indicate, stock price moves even further away from fundamental value during the pre-announcement period, and hence the magnitude of the reversal is greater than in the low noise regime. Prior literature shows that higher levels of investor sentiment are associated with stronger influences of noise trading on stock prices (Baker and Wurgler, 2006; Kumar and Lee, 2006). Thus, we use investor sentiment to test the systematic noise explanation. Specifically, we examine the effects of investor sentiment on the magnitude of the returns from PEAR hedge strategy and the degree to which PAR remains predictive of SUE. If higher levels of investor sentiment result in more correlated trades among uninformed investors, then higher levels of sentiment should result in wider pre-announcement return spreads. However, because these trades are uninformative, adding only noise to the PAR signal, PAR should become less predictive of SUE under these conditions. When news is revealed, the market unravels the mispricing created by the noise trades, resulting in price reversals. Because we expect pre-announcement spreads to be wider when sentiment is bullish, the magnitude of the return reversal should be larger at the earnings announcement date. 7 Bhattacharya et al (2000) examines earnings announcements in a country absent of insider trading prohibitions, and finds earnings announcement days to have no abnormal price or volume reactions when compared to nonannouncement days. 12

Figure 2 plots mean quarterly hedge returns from 1990 through 2009. The graph illustrates that the PEAR strategy remains remarkably robust over time, yielding positive returns for all but five quarters. We begin our investigation of the systematic noise explanation, as discussed in the preceding paragraph, by calculating the correlation between the hedge returns in plotted in Figure 2, and investor sentiment. We use three different proxies of sentiment: two developed by Baker and Wurgler (2006, 2007), henceforth BW one of which is orthogonalized to macroeconomic business cycle indicators, and the University of Michigan Consumer Sentiment index. Table 5, Panel A, displays correlations between the quarterly PEAR hedge returns plotted in Figure 2, and the various sentiment indices. Pearson correlation coefficients are large, positive, and highly significant, ranging from 0.35 to 0.48. Unsurprisingly, correlation coefficients among the three indices are also high, ranging between 0.40 and 0.94. For brevity, we conduct the remainder of our analyses in Section IV with the orthogonalized BW index, SENT, in accordance with prior literature. 8 Panels B and C of Table 5 examine the relation between the PAR hedge for announcement returns and earnings surprise under varying levels of investor sentiment. Panel B presents univariate results, and documents that the magnitude of the price reversal from the PAR hedge doubles when moving from low to high levels of investor sentiment, i.e., from 0.751 to 1.519. Conversely, the average spread in SUEs across extreme PAR quintiles decreases from 0.282 in low sentiment regimes to 0.204 in high sentiment regimes, indicating that PAR is significantly less predictive of SUE when sentiment is high. Panel C verifies these results within a multivariate framework by regressing RET and SUE on the interaction of PAR and SENT and controls for size, book-tomarket, and momentum. The multivariate analyses confirm the results in Panel A, with PAR*SENT coefficients of -0.590 and -0.082 for the RET and SUE regressions, with both coefficients significant at the 5% level. 8 Untabulated findings based on tests using the other two alternative sentiment indices result in inferences that are qualitatively similar to those based on tabulated findings. 13

Taken together, the results in Panels B and C confirm that the magnitude of the PEAR strategy returns become larger, and PAR as a signal of the impending earnings surprise becomes weaker as investor sentiment increases. These findings are consistent with the reduction of the signal-to-noise ratio of the PAR signal when sentiment is high, i.e., an increase in the influence of noise traders relative to informed agents. If this conjecture is correct, then the magnitude of the spreads during the pre-announcement period between high and low PAR quintiles will be positively correlated with the level of investor sentiment. Table 5, Panel D, presents mean PAR spreads by quintile for various levels of investor sentiment by partitioning SENT into low, medium, and high terciles. Consistent with predictions, return spreads between low and high PAR quintiles monotonically increase with sentiment. The increase in PAR spreads between low and high sentiment regimes is 35.4%, i.e., 11.9% to 16.2%. Baker and Wurgler (2006) find that firms with the highest levels of valuation uncertainty are most influenced by these waves of sentiment. This suggests that the effect of investor sentiment on PAR should be amplified for firms with high levels of information uncertainty (IU). We examine whether this is the case by regressing the absolute value of PAR on a proxy for IU, SENT, and the interaction between IU and SENT. We create our proxy for information uncertainty by taking the average quintile rank of SIZE, VLTY, COV, DISP, and INST, after multiplying SIZE, COV, and INST by -1. VLTY is return volatility over the prior six months. COV (DISP) is the number of analysts covering a firm (dispersion in analysts earnings forecasts) immediately prior to the firm s quarterly earnings announcement. INST is the percentage of total shares outstanding held by institutional investors in the prior calendar quarter. Results presented in Table 5, Panel E, are consistent with systematic noise having greater effects on high IU/hard-to-value firms, with the interaction coefficient, 1.416, being significantly positive. 14

These findings support the systematic noise hypothesis, whereby sentiment driven trades during the pre-announcement period widen the gap between the high and low PAR quintiles without providing any additional information about the impending earnings surprise. When information is revealed at the earnings announcement, market forces unravel the speculation-based noise trades as prices converge towards intrinsic value, thereby resulting in sharper reversals under higher levels of investor sentiment. 4.2 Overshooting: Theory and Evidence Figure B: Schematic of Overshooting Explanation Figure B provides an alternate mechanism which would lead to results consistent with those previously discussed in Section 3 showing that pre-announcement returns are positively predictive of SUE, but negatively predictive of announcement returns. Prior literature in behavioral finance 15

shows stock price overreactions can arise from psychological biases (Daniel, Hirshleifer, and Subrahmanyam, 1998; Barberis, Shleifer and Vishny; 1998) or market frictions (Hong and Stein, 1999), yielding price reversals at the earnings announcement date, and return distributions consistent with Figure 1. There are key differences between the systematic noise and overshooting mechanisms. If PEAR occurs as a result of overshooting, prices will systematically overshoot their expected intrinsic values at the announcement date, consistently leading to negative hedge returns for firms in high PAR quintiles, and positive hedge returns for firms in low PAR quintiles. Conversely, if PEAR is driven by systematic noise, hedge returns do not necessarily occur because prices are pushed too far, but rather because stock returns driven by noise trades unravel when information is revealed. Figure C illustrates an example of this for the case of positive SUE and informed traders facing constraints that prevent them from fully preempting the surprise at the earnings announcement. Figure C: Schematic of Noise Trading Positive SUE, News is partially preempted 16

In this circumstance, there exists some level of noise trading which does not result in the overshooting of prices during the PAR period, but still creates a positive hedge return because the positive reaction from high PAR is far smaller than the positive reaction from low PAR, arising from the effect of noise traders trading against the direction of the informed traders. Table 6, Panel A, which reports double-sorted announcement returns using PAR and SUE as partitioning variables, reveals that announcement returns monotonically increase across SUE quintiles, but monotonically decrease across PAR quintiles. More importantly for distinguishing between the overshooting and systematic noise explanations, results indicate that when SUE is above (below) the median, high (low) PAR firms have positive (negative) announcement returns. For example, returns for high PAR firms across SUE quintiles 4 and 5 are 0.53 and 1.06%, respectively, while returns for low PAR firms across SUE quintiles 1 and 2 are -1.02% and -0.31%, respectively. Such results provide strong evidence that that the PEAR strategy returns do not reflect overreactions during the pre-announcement period. The largest absolute magnitudes of announcement returns occur when PAR and SUE are most discordant, providing further evidence that pricing errors driven by non-informed traders appear to be responsible for these reversals. Firms in the lowest quintile of PAR and highest quintile of SUE have announcement returns of 2.60%. In contrast, firms in the highest quintile of PAR and lowest quintile of SUE have announcement returns of -2.13%. The absence of reversals among Low PAR/Low SUE and High PAR/High SUE portfolios suggests that systematic noise trading is a critical ingredient in driving return reversals and is generally inconsistent with the overshooting hypothesis. Delving deeper into the Panel A results, Panel B contains regression results of announcement window returns, RET(-1,+1), after decomposing PAR into two components: FPAR, the component explained by SUE, and NPAR, the component orthogonal to SUE. In each quarter, 17

we run cross-sectional regressions of PAR on SUE, where RPAR is the regression residual and FPAR is the fitted value. Column (1) demonstrates that RPAR negatively predicts announcement returns consistent with the earnings announcement acting as a synchronization signal for arbitrageurs to unravel noise trades placed in the pre-announcement period. Compared to the PAR coefficient in Table 3, Panel C, the coefficient on RPAR is significantly larger, in both economic and statistical terms. The RPAR coefficient increases by more than 20 basis points and the t-statistic increases to 10, consistent with RPAR capturing the component of pre-announcement returns attributable to noise trade. Column (2) demonstrates that NPAR positively predicts announcement returns. Interestingly, FPAR also predicts announcement returns incremental to SUE, consistent with preannouncement informed trading containing information incremental to SUE news. These results are indicative of systematic noise driving PAR away from intrinsic value, whereas the overshooting hypothesis implies that PAR would consistently move towards and beyond the post-announcement intrinsic value. Taken together, the findings in Tables 5 and 6 provide support for the systematic noise explanation, and are inconsistent with the overshooting hypothesis. 4.3 Risk-Based Explanation Finally, we consider the possibility that PEAR can be explained with a risk-based argument. As we note above, the strategy of holding firms in low PAR quintiles and shorting firms in high PAR quintiles just prior to earnings announcements earns positive abnormal hedge returns during the announcement period in 74 of 79 sample quarters, with returns remaining positive and significant even after controlling for Fama-French and Carhart factors. In addition, for the loss quarters that comprise 6% of the sample quarters, the average magnitude of hedge returns, -0.27%, 18

is less than one-fourth of the size of the average return earned when the strategy yields positive returns, 1.17%. Similarly, Panel C of Table 6 documents the probability of a firm remaining in the same PAR quintile over subsequent quarters. Results of a chi-squared test of the null of 20.0 percent per cell cannot be rejected at the ten percent level, which implies that PAR is unlikely to be a proxy for a static firm-level characteristic correlated with risk. Nonetheless, it is possible that idiosyncratic earnings announcement risk is a contributing factor to the profitability of the PEAR strategy returns (Cohen et al 2007; Savor and Wilson, 2010). In this circumstance, firms in the lower PAR quintile would need to have significantly higher levels of information risk relative to firms in higher PAR quintiles, resulting in undiversified investors placing an additional discount on the expected future cash flows until earnings are revealed. Prior literature (Jiang, Lee and Zhang, 2005; Zhang, 2006) suggests information risk is higher for smaller firms, firms with higher levels of return volatility, lower levels of institutional holdings and analyst coverage, and higher levels of analyst dispersion. To assess whether firms in the lower PAR quintile have significantly higher levels of information risk relative to firms in higher PAR quintiles, Panel D of Table 6 SIZE, VLTY, COV, DISP, and INST across PAR quintiles. None of these variables has monotonic relations across PAR quintiles. Instead, VLTY and STDEV appear as U-shapes, and SIZE, COV, and INST graphically appear as inverted U-shapes. These results appear inconsistent with the possibility that firms in lower PAR quintiles could have higher levels of information risk than firms in high PAR quintiles. Unless there exists an extremely large, unrealized left tail in the distribution of our hedge returns, it is difficult to conceive of an explanation whereby information risk acts as the primary driver of our documented PEAR hedge returns. 19

5. Robustness and Alternative Specifications 5.1. Additional Evidence on the Role of Sentiment In Section 4, we provide evidence that systematic noise traders influence pre-announcement prices, resulting in a negative relation between PAR and announcement-window returns. This interpretation yields an additional testable implication that we explore below. Multiple studies within the literature on individual investor behavior conclude that individuals have limited attention and are therefore more likely to trade more salient, attention-grabbing stocks (e.g. Barber and Odean, 2008; Da, Engelberg and Gao, 2011). Under the assumption that systematic noise traders and individual investors are prone to similar trading behavior, we predict that earnings announcement-window return reversals are most pronounced among glamour firms because they are more likely to attract systematic noise traders. To test this prediction, we estimate the following regression: RET(-1,+1) i,q = λ 0 + λ 1 PAR i,q + λ 2 SENT q + λ 3 PAR i,q *SENT q + λ 4 PAR i,q *SENT q (2) *HighMTB i,q + λ 5 HighMTB i,q + λ 6 SIZE i,q + λ 7 MOMEN i,q + i,q, Following prior literature, we use a firm s market-to-book ratio as a proxy for its value/glamour classification. We include HighMTB i,q, an indicator variable that equals one when the firm is in the highest quintile of market-to-book for given calendar quarter, and zero otherwise, as a proxy for high attention-grabbing stocks. The regression summary statistics for equation (2) in Table 7, Panel A, document a significantly negative interaction effect between PAR i,q and SENT q, which is consistent with earnings announcement reversals being most pronounced when investor sentiment is high. Similarly, the coefficient on the three-way interaction term between PAR i,q, SENT q, and HighMTB i,q is negative and significant, which indicates that the return reversal at earnings announcements is most 20

pronounced among glamour firms. Column (2) contains results of estimating equation (2) when SUE is the dependent variable. The three-way interaction term is negative but insignificant. Panel B contains from estimation of equation (2) in which we replace HighMTB with an indicator variable indicating that the firm has high share turnover, TO. Barber and Odean (2008) use share turnover as a measure of the extent to which a stock is more likely to grab the attention of individual, or uninformed investors. It is defined as daily equity trading volume scaled by total shares outstanding as reported in CRSP, averaged over the 60-days ending on t-10. HighTO is a dummy variable that equals one for firms in the highest quintile of TO. The three-way interaction term between PAR, SENT, and HighTO is negative and significant indicating that return reversals at earnings announcements are most pronounced among high volume/glamour firms. Moreover, column (2) documents that the three-way interaction effect is also negative and significant at the 10% level when predicting SUE. This result is consistent with PAR being less informative of earnings news among attention grabbing stocks when sentiment is high. Both of these results are consistent with noise traders playing a more pronounced role in the formation of pre-announcement prices among attention-grabbing firms and provide additional support for the systematic noise interpretation of our main findings. 5.2. Implementing the PEAR Strategy The ultimate relevance of our findings for investment decisions hinges upon investors ability to implement the PEAR strategy in practice. One potential source of concern is that the implementation of the PEAR strategy requires that investors know earnings announcement dates two-days in advance of the actual announcement. Although firms regularly announce earnings in a predictable fashion, they sometimes deviate from their historical patterns leading to uncertainty regarding the announcement date. We address this concern in two ways. First, Table 8 presents 21

expected-announcement-window returns across PAR quintiles. We calculate expected announcement dates using the methodology outlined in Cohen et al. (2007). Similar to the design of our main tests, we calculate pre-announcement returns from days k-3 to k-5 and announcementwindow returns from k-1 to k+1, where k denotes the expected earnings announcement date. Table 8 presents expected-announcement-window returns across PAR quintiles. Columns (1) through (4) present findings for all earnings announcements pooled together, followed by those that are early, on time, and late relative to the expected announcement date. Early, on time, and late announcements are those that relate to firms that announce earnings prior to day k-1, during the interval (k-1, k+1), and after day k+1. The findings in Table 8 demonstrate that the PEAR strategy is robust to the use of expected announcement dates. This is not surprising given that many firms in our sample have announcement dates that are timed consistently across quarters. The pooled results in column (1) demonstrate that the average hedge return is 75.7 basis points, approximately 25 basis points less than the results documented in Table 2. We also partition the sample into groups based on the timing of expected earnings announcement dates relative to actual announcement date to better understand the difference in hedge returns across Tables 2 and column (1) of Table 8. Note that the magnitude of the hedge returns for on-time announcements in column (3) is similar to the magnitude of those documented in Table 2. This suggests that the reduction in PEAR strategy returns when using expected announcement dates is attributable to early and late announcing firms. Late announcing firms account for most of this effect, with the PEAR strategy earning an average return of 33 basis points. The smaller reversal for late announcers is consistent with the market being able to infer the content of the earnings announcement when the announcement date is delayed. More broadly, these findings are consistent with noise traders being less likely to influence pre-announcement prices when the timing of the announcement date is uncertain. 22

We next examine the profitability of the PEAR strategy among large market capitalization firms. We do so by restricting our sample to firms in the top two NYSE Size deciles of each calendar year, resulting in a sample of 66,051 quarterly earnings announcements spanning 1990-2009. In untabulated results, we find that the average PEAR strategy return exceeds 100 basis points among the sample of large firms, which is similar to the PEAR hedge return for the full sample documented in Table 3, Panel A. This analysis yields two main benefits. First, firm size is positively related to media and analyst coverage, which mitigates concerns regarding the predictability of earnings announcement dates. Second, firm size is negatively correlated with transaction costs, mitigating concerns that the profitability of the PEAR strategy is attributable to market frictions. A related consideration for investors is that the PEAR strategy is designed as a zeroinvestment strategy, where investors sell-short high PAR firms and use the proceeds of the short sale to purchase low PAR firms. Because earnings announcement dates occur throughout the year, investors may be concerned that non-synchronous earnings announcements hinder their ability to implement a zero-investment strategy. To address this concern, we conduct two additional tests. First, we limit the sample to calendar months with at least 500 earnings announcements. The cutoff at 500 firms is arbitrary but significantly increases the probability that there are sufficient firms in the extreme PAR quintiles to implement a zero-investment strategy. This approach again yields an average PEAR strategy return of 104 basis points (t-statistic = 9.988). Second, we limit our sample to firm-quarters whose earnings announcement occurs on a calendar date where there is at least one firm in both the high and low PAR quintile. This ensures that on any given announcement date in our sample, an investor can short a high-par stock and use the proceeds to buy a low-par stock on the same day. Untabulated findings reveal that the PEAR hedge strategy for this restricted sample yields an average return of 98.6 basis points (tstatistic=7.985). 23

Collectively, the results of Section 5 provide additional corroborating evidence of our main findings and mitigate concerns that PEAR strategy returns are attributable to uncertainty regarding earnings announcement dates, transaction costs, and/or capital constraints. 6. Conclusion We document a pre-earnings announcement reversal (PEAR) paradox, wherein stock returns over the three-day period just prior to the earnings announcement date positively predict earnings surprises, but negatively predict the stock returns during the three-day earnings announcement period. We also provide evidence that is consistent with informed traders reversing the influence of uninformed noise trades during the pre-announcement period. We create a portfolio comprising a long position of firms in the lowest pre-announcement return quintile and a short position of firms in the highest pre-announcement return quintile, and find that this strategy yields anomalous returns in excess of one percent during the three-day earnings announcement window. This paradox is consistent with informed agents who trade in the pre-announcement period being responsible for the ability of pre-announcement returns to predict earnings surprise, and uninformed agents systematically trading in a manner that yields price movements unrelated to expected future cash flows. The effects of these noise trades are subsequently unraveled when earnings news is announced, thereby creating the observed return reversals. Consistent with this explanation, we use investor sentiment to measure of the impact of systematic noise on asset prices and find that the return reversals at earnings announcements are nearly twice as large in higher versus lower sentiment periods. The PEAR anomaly that we document is not an isolated situation that is constrained to small or illiquid firms, but instead appears to be a systematic violation of weak-form market efficiency whereby prices deviate from fundamental value across all firms, resulting in economically 24

large and predictable price reversals over extremely short holding periods. Regarding prior literature on short-run reversals, we use pseudo-announcement returns to distinguish the PEAR anomaly from more general short-run negative autocorrelations, and find our reversals to be nearly five times larger in magnitude. Regarding a look-ahead bias of the firms announcement date, we test the sensitivity of the PEAR strategy using firms with predictable earnings announcement dates. Hedge returns using this subset of firms are qualitatively similar when compared to the full sample. Finally, concerning the possibility that differences in the level of information risk across firms in low and high PAR quintiles may be responsible for our return reversals, we examine an array of variables that act as proxies for information risk, and are unable to find any evidence in support of this risk-based explanation. Overall, our results should be of interest to both money managers and academics. We provide further insight to the existing finance literature on how noise traders contribute to price reversals, while simultaneously documenting the participation of informed agents just prior to the revelation of news. While we view earnings announcements as an ideal venue to observe these noisetrading driven reversals due to their frequency, consistency across firms, and level of predictability, the phenomenon which we describe need not be limited to such an event. For example, future research in this area could potentially document the impact of noise traders around other events such as stock split, dividend, and acquisition announcements. 25

References Aboody, D., R. Lehavy, and B. Trueman, 2010, Limited attention and the earnings announcement returns of past stock market winners, Review of Accounting Studies 15, 317 344. Affleck-Graves, J., Jennings, R. and R. Mendenhall, 1995, Evidence of Informed Trading Prior to Earnings Announcements. NYU Working Paper No. FIN-94-001. Atiase, R., 1985, Predisclosure Information, Firm Capitalization, and Security Price Behavior Around Earnings Announcements, Journal of Accounting Research 23, 21-36. Bagnoli, M., M.D. Beneish, and S.G. Watts, 1999, Whisper forecasts of quarterly earnings per share, Journal of Accounting and Economics 28, 27 50. Baker, M. and J. Wurgler, 2006, Investor Sentiment and the Cross-Section of Stock Returns, Journal of Finance 61, 1645 1680. Ball, R. and P. Brown, 1968, An empirical evaluation of accounting income numbers, Journal of Accounting Research 159 178. Barber, B.M. and T. Odean, 2008, All that glitters: The effect of attention and news on the buying behavior of individual and institutional investors, Review of Financial Studies 21, 785. Barberis, N., A. Shleifer, and R. Vishny, 1998, A model of investor sentiment, Journal of Financial Economics 49, 307 343. Beaver, W., R. Clark and W. Wright, 1979. The association between unsystematic security returns and the magnitude of earnings forecast errors, Journal of Accounting Research 79, 316-340. Beaver, W. and R. Lambert, and D. Morse, 1980, The information content of security prices, Journal of Accounting and Economics 2, 3 28. 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. Bhattacharya, U., H. Daouk, B. Jorgenson, and C.H. Kehr, 2000, When an event is not an event: the curious case of an emerging market, Journal of Financial Economics 55, 69 101. Burgstahler, D. and I. Dichev, 1997, Earnings management to avoid earnings decreases and losses, Journal of Accounting and Economics 24, 99 126. Carhart, M.M., 1997, On persistence in mutual fund performance, Journal of Finance 52, 57 82. Christophe, S., M. Ferri, and J. Hsieh, 2010, Informed trading before analyst downgrades: Evidence from short sellers, Journal of Financial Economics, 95 85-106 Cohen, D.A., A. Dey, T.Z. Lys, and S.V. Sunder, 2007, Earnings announcement premia and the 26

limits to arbitrage, Journal of Accounting and Economics 43, 153 180. Conrad, J. and G. Kaul, 1993, Long-term market overreaction or biases in computed returns?, Journal of Finance 48, 39 63. Da, Z., J. Engelberg, and P. Gao, 2011, In search of attention, Forthcoming Journal of Finance. Dey, M. and B. Radhakrishna, 2007, Who Trades Around Earnings Announcements? Evidence from TORQ Data. Journal of Business Finance & Accounting, 34, 269 291. Daniel, K., D. Hirshleifer, and A. Subrahmanyam, 1998, Investor psychology and security market under-and overreactions, Journal of Finance 53, 1839 1885. Degeorge, F., J. Patel, and R. Zeckhauser, 1999, Earnings management to exceed thresholds, The Journal of Business 72, 1 33. DellaVigna, S. and J.M. Pollet, 2009, Investor inattention and Friday earnings announcements, Journal of Finance 64, 709 749. El-Gazzar, S., 1998, Predisclosure Information and Institutional Ownership: A Cross-Sectional Examination of Market Revaluations during Earnings Announcement Periods, The Accounting Review 73, 119-129. Fama, E.F., 1998, Market efficiency, long-term returns, and behavioral finance, Journal of Financial Economics 49, 283 306. Fama, E.F., L. Fisher, M.C. Jensen, and R. Roll, 1969, The adjustment of stock prices to new information, International Economic Review 10, 1 21. Fama, E.F. and K.R. French, 2008, Dissecting anomalies, Journal of Finance 63, 1653 1678. Foster, G., C. Olsen, and T. Shevlin, 1984, Earnings releases, anomalies, and the behavior of security returns, The Accounting Review 59, 574 603. Gow, I. D., G. Ormazabal, T. J. Taylor, 2010, Correcting for cross-sectional and time-series dependence in accounting research, The Accounting Review, 85, 483-512. Gutierrez Jr, R.C. and E.K. Kelley, 2008, The Long-Lasting Momentum in Weekly Returns, Journal of Finance 63, 415 447. Hong, H. and J. Stein, 1999, A Unified Theory of underreaction, momentum trading and overreaction in financial markets, Journal of Finance 54, 65 91. Jegadeesh, N. and S. Titman, 1993, Returns to buying winners and selling losers: Implications for stock market efficiency, Journal of Finance 48, 65 91. Jiang, G., C.M.C. Lee, and Y. Zhang, 2005, Information uncertainty and expected returns, Review of 27

Accounting Studies 10, 185 221. Kumar, A. and C.M.C Lee, 2006, Retail investor sentiment and return comovements, Journal of Finance 61, 2451-2486. Lehmann, B., 1990, Fads, martingales, and market efficiency. National Bureau of Economic Research Cambridge, Mass., USA. Linnainmaa, J. 2010, Do Limit Orders Alter Inferences about Investor Performance and Behavior? Journal of Finance 65, 1473-1506. Livnat, J. and R.R. Mendenhall, 2006, Comparing the post earnings announcement drift for surprises calculated from analyst and time series forecasts, Journal of Accounting Research 44, 177 205. 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. Sloan, R.G., 1996, Do stock prices fully reflect information in accruals and cash flows about future earnings?, The Accounting Review 71, 289 315. Taylor, Daniel J., 2010, Individual Investors and Corporate Earnings. Available at SSRN: http://ssrn.com/abstract=1707030 Trueman, B., M.H. Wong, and X.J. Zhang, 2003, Anomalous stock returns around internet firms earnings announcements, Journal of Accounting and Economics 34, 249 271. Zhang, X.F., 2006, Information uncertainty and stock returns, Journal of Finance 61, 105 137. 28

Figure 1: Announcement-Window Returns across PAR Quintiles The figure plots the time-series average announcement window market-adjusted return across PAR quintiles. PAR is calculated from t-5 to t-3 as the cumulative market-adjusted return, where day t denotes the earnings announcement date. Announcement-window returns are calculated from t-1 to t+1. The average return to each PAR quintile is calculated on a quarterly basis and subsequently averaged over all quarters in the 1990-2009 sample window. The sample consists of 183,228 earnings announcements beginning in the second quarter of 1990 and ending in the fourth quarter of 2009. To avoid look ahead bias, quintiles are formed using breakpoints from the prior calendar quarter. 29

Figure 2: Quarterly Performance of PAR Hedge Strategy The figure plots the average announcement window hedge return for each calendar quarter in the sample. The sample consists of 183,228 earnings announcements beginning in the second quarter of 1990 and ending in the fourth quarter of 2009. The hedge strategy involves buying (selling) firms in the lowest (highest) pre-announcement returns (PAR) during the firm s three-day earnings announcement window, denoted by t-1 to t+1, where day t is the earnings announcement. PAR is calculated from t-5 to t-3 as the cumulative market-adjusted return. To avoid look ahead bias, quintiles are formed using breakpoints from the prior calendar quarter. 30

Table 1: Descriptive Statistics Panel A presents descriptive statistics of the main variables used throughout the paper. PAR is the pre-earningsannouncement return calculated as the cumulative market-adjusted return from t-5 to t-3, where day t denotes the earnings announcement date. SURPRISE equals the actual EPS number reported in IBES minus the last consensus forecast available immediately prior to the announcement, and scaled by price. SUE is the standardized unexpected earnings, calculated as realized EPS minus EPS from four-quarters ago, divided by its standard deviation over the prior 8 quarters. RET(-1,+1) is the market-adjusted earnings announcement return from t-1 to t+1. SIZE and LBM are the log of market capitalization and log of one plus the book-to-market ratio, respectively. MOMEN equals the firm s marketadjusted return over the six months ending on t-10. Price is the beginning quarter price as reported in CRSP. Panel B contains Pearson (Spearman) correlations above (below) the main diagonal. Panel A: Sample Characteristics (Obs=183,228) Mean STD P25 Median P75 PAR 0.178 5.982-2.372-0.097 2.282 SURPRISE -0.247 2.495-0.111 0.012 0.148 SUE -0.082 2.061-0.585 0.051 0.636 RET(-1,+1) 0.211 8.552-3.815 0.079 4.185 LMC 6.302 1.756 5.038 6.180 7.397 LBM 0.403 0.270 0.240 0.374 0.532 MOMEN 0.025 0.404-0.179-0.016 0.158 PRICE 24.380 25.714 9.675 18.813 32.500 Panel B: Pearson (Spearman) Correlations Above (Below) Diagonal PAR SURPRISE SUE RET(-1,+1) LMC LBM MOMEN PAR 1.000 0.015 0.024-0.054-0.006-0.015 0.006 SURPRISE 0.050 1.000 0.242 0.118 0.101 0.101-0.146 SUE 0.037 0.307 1.000 0.101 0.139 0.044-0.128 RET(-1,+1) -0.046 0.276 0.143 1.000 0.014 0.009 0.008 LMC 0.007 0.159 0.215 0.019 1.000 0.002-0.051 LBM 0.012 0.079 0.054 0.029 0.085 1.000-0.317 MOMEN -0.006-0.032-0.132 0.002-0.069-0.335 1.000 31

Table 2: Regression Results of Earnings Surprises This table presents regression results where SUE and SURPRISE are the dependent variables. SUE is the standardized unexpected earnings, calculated as realized EPS minus EPS from four-quarters ago, divided by its standard deviation over the prior 8 quarters. SURPRISE equals the actual EPS number reported in IBES minus the last consensus forecast available immediately prior to the announcement, and scaled by price. PAR is the pre-earnings-announcement return calculated as the cumulative market-adjusted return from t-5 to t-3, where day t denotes the earnings announcement date. LBM and SIZE are the log of one plus the book-to-market ratio and log of market capitalization, respectively. MOMEN equals the firm s market-adjusted return over the six months ending on t-10. ACC is total accruals scaled by prior quarter total assets. FScore is the summary score of changes in firm s financial condition as detailed in Piotroski (2000), which ranges in value from 0 to 9. PastSURPRISE equals the SURPRISE from the prior calendar quarter. All control variables, except for FScore, are assigned to quintiles where the highest (lowest) values are assigned to quintile 1 (0). t-statistics, shown in parentheses, are based on two-way cluster robust standard errors, clustered by firm and quarter. ***, **, and * indicate significance at the 1, 5, and 10% level, respectively. Dep. Variable: SUE SURPRISE (1) (2) (3) (4) (5) (6) Intercept -0.170*** -0.343*** -0.990*** -0.337*** -0.719*** -1.316*** (-3.01) (-8.48) (-15.77) (-5.51) (-9.06) (-7.54) PAR 0.178*** 0.166*** 0.161*** 0.179*** 0.164*** 0.156*** (8.08) (8.29) (8.32) (6.60) (6.45) (6.45) LBM -0.568*** -0.542*** -0.425*** -0.397*** (-5.68) (-5.49) (-3.43) (-3.34) SIZE 0.020-0.033 0.502*** 0.434*** (0.49) (-0.93) (10.26) (10.58) MOMEN 0.912*** 0.778*** 0.707*** 0.550*** (11.03) (9.47) (6.01) (5.42) ACC 0.122** 0.031 (2.43) (0.79) FScore 0.514*** 0.648*** (19.99) (9.67) PastSURPRISE 0.073*** 0.063*** (20.36) (4.99) R-square 0.001 0.038 0.050 0.001 0.024 0.035 32

Table 3: Prediction of Announcement Window Returns Panel A presents the time-series average returns by quintiles of PAR. PAR is the pre-earnings-announcement return calculated as the cumulative market-adjusted return from t-5 to t-3, where day t denotes the earnings announcement date. To avoid look ahead bias, quintiles are formed using breakpoints from the prior calendar quarter. RET(X,Y) equals the cumulative market-adjusted return from X days relative to the announcement until Y days relative to the announcement date. The average return to each PAR quintile is calculated each calendar quarter and subsequently averaged over the 1990-2009 sample window. Panel B presents the Fama-French alphas obtained as the intercept from estimating the following equation for each PAR quintile: R i,ea -R f,ea = α + β 1 (R m,ea -R f,ea ) + β 2 HML EA + β 3 SMB EA + β 4 UMD EA + ε i,ea where R i,ea is the return of firm i during the three-day earnings announcement window from t-1 to t+1. Similarly R f,ea is the risk free rate, R m,ea -R f,ea equals the excess return on the market, HML EA equals the return on the high-minus-low book-to-market strategy, SMB EA equals the hedge return on the small-minus-big strategy, and UMD EA equals the hedge return on the high-minus-low momentum strategy. All factors are obtained from Ken French s website. Panel C presents the results from regressions where the dependent variable equals the three-day market-adjusted announcementwindow return. LBM and SIZE are the log of one plus the book-to-market ratio and log of market capitalization, respectively. MOMEN equals the firm s market-adjusted return over the six months ending on t-10. ACC is total accruals scaled by prior quarter total assets. FScore is the summary score of changes in firm s financial condition as detailed in Piotroski (2000) ranging in value from 0 to 9. PastSURPRISE equals the SURPRISE from the prior calendar quarter. All control variables, except for FScore, are assigned to quintiles where the highest (lowest) values are assigned to quintile 1 (0). t-statistics, shown in parentheses, are based on two-way cluster robust standard errors, clustered by firm and quarter. ***, **, and * indicate significance at the 1, 5, and 10% level, respectively. Panel A: Time-Series Averages Across Quintiles of Pre-Announcement Returns (PAR) PAR RET(-1,+1) RET(+2,+5) RET(+2,+60) MOMEN 1 (Low PAR) -6.458 0.670 0.171 1.319 0.019 2-1.948 0.386 0.078 0.729 0.028 3-0.072 0.302 0.074 0.775 0.024 4 1.891 0.133-0.018 0.857 0.029 5 (High PAR) 7.626-0.413-0.174 1.317 0.032 Low-High -14.084 1.083 0.345 0.003-0.013 Low-High P-Value 0.000 0.000 0.000 0.995 0.097 Panel B: Fama-French Alphas during Announcement Windows by Quintiles of PAR 1 (Low) 2 3 4 5 (High) Low-High CAPM Alpha 0.5572 0.3254 0.2881 0.0846-0.4603 1.0175 (10.72) (7.90) (7.43) (2.09) -(9.38) Three-Factor Alpha 0.5690 0.3328 0.3005 0.1061-0.4591 1.0281 (11.03) (8.10) (7.76) (2.62) -(9.39) Four Factor Alpha 0.5684 0.3241 0.2891 0.1035-0.4792 1.0476 (11.03) (7.89) (7.47) (2.56) -(9.82) 33

Table 3 [Continued]: Prediction of Announcement Window Returns Panel C: Regression Results of Announcement Window Returns (1) (2) (3) (4) Intercept 0.678*** 0.249 0.085 0.162 (7.02) (1.62) (0.46) (0.64) PAR -0.936*** -0.940*** -0.942*** -0.955*** (-8.26) (-8.33) (-8.40) (-8.50) LBM 0.385*** 0.403*** 0.424*** (3.20) (3.41) (3.56) SIZE 0.472*** 0.453*** 0.307** (3.92) (3.85) (2.51) MOMEN 0.333** 0.259* (2.31) (1.84) ACC -0.646*** (-5.92) FScore 0.058*** (2.66) PastSURPRISE 0.053 (0.57) R-square 0.002 0.002 0.002 0.003 34

Table 4: Pseudo-Announcement Window Returns This table presents the time-series average returns by quintiles of PPAR. PPAR is the pre-pseudo-earningsannouncement return calculated from k-5 to k-3 as the cumulative market-adjusted return, where day k denotes the randomly generated pseudo earnings announcement date. Pseudo announcement dates are calculated by subtracting a randomly selected number of trading days from the actual announcement date. The randomly selected numbers are drawn from a uniform distribution spanning 10 to 40. To avoid look ahead bias, quintiles are formed using breakpoints from the prior calendar quarter. RET(A,B) equals the cumulative market-adjusted return from A days relative to the pseudo announcement until B days relative to the pseudo announcement date. The average return to each PAR quintile is calculated on a quarterly basis and subsequently averaged over the 1990-2009 sample window. Panel B presents the results from regressions where the dependent variable equals the three-day market-adjusted pseudo-announcementwindow return. LBM and SIZE are the log of one plus the book-to-market ratio and log of market capitalization, respectively. MOMEN equals the firm s market-adjusted return over the six months ending on t-10. ACC is total accruals scaled by prior quarter total assets. FScore is the summary score of changes in firm s financial condition as detailed in Piotroski (2000) ranging in value from 0 to 9. PastSURPRISE equals the analyst-based earnings surprise, scaled by price, from the prior calendar quarter. All control variables, except for FScore and PPAR, are assigned to quintiles where the highest (lowest) values are assigned to quintile 1 (0). PPAR is assigned to quintiles ranging from 1 to 5. t-statistics, shown in parentheses, are based on two-way cluster robust standard errors, clustered by firm and quarter. ***, **, and * indicate significance at the 1, 5, and 10% level, respectively. Panel A: Time-Series Averages Across Quintiles of Pseudo Pre-Announcement Returns (PPAR) PSAR RET(-1,+1) RET(+2,+5) RET(+2,+60) MOMEN 1 (Low PPAR) -6.485 0.118 0.185 1.433 0.029 2-1.951 0.030 0.098 1.238 0.026 3-0.148 0.019-0.030 0.960 0.029 4 1.692-0.014-0.055 1.087 0.038 5 (High PPAR) 6.910-0.092-0.124 1.295 0.045 Low-High -13.395 0.210 0.308 0.137-0.016 Low-High P-Value 0.000 0.001 0.000 0.715 0.022 Panel B: Regression Results of Pseudo Announcement Returns (1) (2) (3) (4) Intercept 0.136*** 0.145* 0.043 0.004 (2.83) (1.93) (0.47) (0.03) PPAR -0.043*** -0.043*** -0.044*** -0.046*** (-3.88) (-3.89) (-4.01) (-4.14) LBM -0.003 0.003 0.005 (-0.18) (0.21) (0.36) SIZE -0.002-0.005-0.018 (-0.13) (-0.38) (-1.33) MOMEN 0.050** 0.037* (2.49) (1.83) ACC -0.052*** (-5.00) FScore 0.022** (2.09) PastSURPRISE 0.037*** (3.46) R-square 0.000 0.000 0.000 0.001 35

Table 5: Sentiment, Pre-Announcement Returns, and PEAR Strategy Returns Panel A contains correlations between quarterly PAR hedge returns and three measures of investor sentiment. The hedge strategy involves buying (selling) firms in the lowest (highest) pre-announcement returns during the firm s threeday earnings announcement window, denoted by t-1 to t+1, where day t is the earnings announcement. PAR is the preearnings-announcement return calculated as the cumulative market-adjusted return from t-5 to t-3, where day t denotes the earnings announcement date. SENT (RawSENT) is the investor sentiment index used in Baker and Wurgler (2006) that is (not) orthogonalized to macro factors. MichSENT is the University of Michigan consumer sentiment index. Panel B contains average PAR hedge returns across calendar months sorted into terciles of investor sentiment, SENT. Panel B also contains the average difference in SUEs across high and low PAR quintiles for each sentiment portfolio. SUE is the standardized unexpected earnings, calculated as realized EPS minus EPS from four-quarters ago, divided by its standard deviation over the prior 8 quarters. Panel C contains regression results where the dependent variables are RET(-1,+1) and SUE. RET(-1,+1) is the market-adjusted earnings announcement return from t-1 to t+1. LBM and SIZE are the log of one plus the book-to-market ratio and log of market capitalization, respectively. MOMEN equals the firm s marketadjusted return over the six months ending on t-10. Panel D contains time-series average values of PAR across PAR quintiles and terciles of SENT1. Panel E contains regression results where the dependent variable equals the absolute value of PAR. IU equals the average quintile rank of firm SIZE, VLTY, COV, DISP, and INST. SIZE is the log of market capitalization. VLTY is return volatility over the prior six months. COV (DISP) is the number of analysts covering a firm (dispersion in analysts earnings forecasts) immediately prior to the firm s quarterly earnings announcement. INST is the percentage of total shares outstanding held by institutional investors in the prior calendar quarter. t-statistics, shown in parentheses, are based on two-way cluster robust standard errors, clustered by firm and quarter. ***, **, and * indicate significance at the 1, 5, and 10% level, respectively. Panel A: Pearson (Spearman) Correlations Above (Below) Diagonal PAR Hedge SENT RawSENT MichSENT PAR Hedge 1.000 0.349 0.323 0.481 SENT 0.340 1.000 0.942 0.462 RawSENT 0.303 0.927 1.000 0.401 MichSENT 0.407 0.562 0.554 1.000 Panel B: Monthly Hedge Returns by Sentiment Terciles RET(-1,+1) SUE Mean Median Mean Median 1 (Low SENT) 0.751 0.585-0.282-0.281 2 0.910 0.973-0.201-0.209 3 (High SENT) 1.519 1.420-0.204-0.182 High-Low 0.768 0.835 0.078 0.099 Panel C: Regression Results of RET(-1,+1) and SUE Dep. Variable: RET(-1,+1) SUE (1) (2) (3) (4) Intercept -0.024-0.174-0.241*** -0.262*** (-0.13) (-0.96) (-5.37) (-5.46) PAR -0.931*** -0.629*** 0.167*** 0.209*** (-10.02) (-6.93) (10.18) (8.38) SENT 0.094 0.387-0.179*** -0.138** (0.56) (1.52) (-3.37) (-2.34) PAR*SENT -0.590** -0.082** (-2.53) (-2.18) LBM 0.394*** 0.394*** -0.379*** -0.379*** (3.58) (3.58) (-11.27) (-11.26) SIZE 0.482*** 0.482*** 0.043 0.043 (4.49) (4.48) (1.31) (1.31) MOMEN 0.410*** 0.411*** 0.842*** 0.842*** (2.89) (2.90) (22.99) (23.02) R-square 0.003 0.003 0.039 0.039 36

Table 5 [Continued]: Sentiment, Pre-Announcement Returns, and PEAR Strategy Returns Panel D: Average PAR across Monthly B&W Sentiment Partitions Low SENT Mid SENT High SENT 1 (Low PAR) -5.55-5.77-7.26 2-1.64-1.70-2.23 3-0.03-0.08-0.11 4 1.64 1.65 2.15 5 (High PAR) 6.41 6.49 8.94 Low-High -11.961-12.262-16.201 Panel E: Regression Results of PAR (1) (2) Intercept 3.436*** 3.346*** (21.40) (19.15) SENT 1.189*** 1.340*** (4.16) (4.21) IU 4.725*** 3.993*** (19.38) (11.98) IU*SENT 1.416** (2.50) R-square 0.058 0.059 37

Table 6: Overshooting, Risk and PEAR Strategy Returns Panel A presents time-series average announcement-window returns independently sorted into quintiles of PAR and SUE. PAR is the pre-earnings-announcement return calculated as the cumulative market-adjusted return from t-5 to t-3, where day t denotes the earnings announcement date. SUE is the standardized unexpected earnings, calculated as realized EPS minus EPS from four-quarters ago, divided by its standard deviation over the prior 8 quarters. Panel B contains regression results of announcement window returns after PAR is decomposed into two components: RPAR and FPAR. We run quarterly cross-sectional regressions of PAR on SUE, where RPAR is the regression residual and FPAR is the fitted value. LBM and SIZE are the log of one plus the book-to-market ratio and log of market capitalization, respectively. MOMEN equals the firm s market-adjusted return over the six months ending on t-10. ACC is total accruals scaled by prior quarter total assets. FScore is the summary score of changes in firm s financial condition as detailed in Piotroski (2000) ranging in value from 0 to 9. PastSURPRISE equals the analyst-based earnings surprise, scaled by price, from the prior calendar quarter. All control variables, except for FScore, are assigned to quintiles where the highest (lowest) values are assigned to quintile 1 (0). t-statistics, shown in parentheses, are based on two-way cluster robust standard errors, clustered by firm and quarter. ***, **, and * indicate significance at the 1, 5, and 10% level, respectively. Panel C provides the average transition rates across quintiles of PAR in adjacent calendar quarters. Panel D contains time-series average firm characteristics across PAR quintiles. SIZE is the log of market capitalization. VLTY is return volatility over the prior six months. COV (DISP) is the number of analysts covering a firm (dispersion in analysts earnings forecasts) immediately prior to the firm s quarterly earnings announcement. All variables are assigned to quintiles where the highest (lowest) values are assigned to quintile 1 (0). Panel A: Double-Sorted Announcement-Window Returns by SUE 1 (Low SUE) 2 3 4 5 (High SUE) 1 (Low PAR) -1.019-0.317 0.778 1.948 2.603 2-1.094-0.452 0.471 1.254 1.714 3-1.045-0.595 0.415 1.084 1.517 4-1.297-0.693 0.133 0.884 1.450 5 (High PAR) -2.127-1.365-0.310 0.527 1.064 Low-High 1.108 1.048 1.089 1.421 1.539 Panel B: Regression Results of Announcement Window Returns (1) (2) (3) (4) Intercept 0.272-1.172*** -0.606*** -1.146*** (1.04) (-5.22) (-2.72) (-5.69) RPAR -1.178*** -1.105*** -1.144*** (-10.18) (-9.67) (-9.83) FPAR 1.817*** 1.771*** 1.054*** (10.10) (9.91) (7.48) LBM 0.423*** 0.422*** 0.420*** 0.677*** (3.55) (3.65) (3.54) (5.85) SIZE 0.312** 0.307*** 0.320*** 0.340*** (2.54) (2.59) (2.67) (2.97) MOMEN 0.251* 0.102 0.101-0.357** (1.78) (0.78) (0.76) (-2.54) ACC -0.641*** -0.591*** -0.600*** -0.690*** (-6.00) (-5.82) (-5.82) (-7.28) FScore 0.058*** 0.066*** 0.066*** 0.014 (2.69) (3.14) (3.13) (0.66) PastSURPRISE 0.051-0.031-0.021-0.375*** (0.55) (-0.34) (-0.23) (-3.93) SUE 3.050*** (20.48) R-square 0.004 0.007 0.009 0.023 38

Table 6 [Continued]: Overshooting, Risk and PEAR Strategy Returns Panel C: Transition Matrix of PAR Quintiles One-Quarter Ahead PAR Quintiles 1 (Low) 2 3 4 5 (High) 1 (Low PAR) 23.40 18.79 16.92 18.17 22.72 2 18.53 20.93 21.34 20.73 18.47 3 17.58 21.27 22.47 21.07 17.60 4 18.36 20.71 21.55 20.39 18.99 5 (High PAR) 22.78 18.40 17.43 18.39 23.00 Panel D: Sample Characteristics by PAR Quintiles SIZE VLTY COV DISP INST 1 (Low PAR) 5.809 0.037 1.603 0.024 0.301 2 6.376 0.029 1.715 0.023 0.329 3 6.477 0.028 1.726 0.022 0.359 4 6.400 0.029 1.724 0.022 0.328 5 (High PAR) 5.866 0.037 1.632 0.023 0.308 High-Low 0.057 0.000 0.030-0.001 0.007 39

Table 7: Return Prediction and Sentiment Panels A and B contains regression results where the dependent variables are RET(-1,+1) and SUE. RET(-1,+1) is the market-adjusted earnings announcement return from t-1 to t+1.sue is the standardized unexpected earnings, calculated as realized EPS minus EPS from four-quarters ago, divided by its standard deviation over the prior 8 quarters. LBM and SIZE are the log of one plus the book-to-market ratio and log of market capitalization, respectively. MOMEN equals the firm s market-adjusted return over the six months ending on t-10. HighMTB is a dummy variable that equals one when the firm-quarter is the lowest quintile of LBM. HighTO is a dummy variable that equals one when the firm-quarter is the highest quintile of equity share turnover, defined as total share volume scaled by shares outstanding averaged over the 60-days ending on t-10. All control variables, except for HighMTB, are assigned to quintiles where the highest (lowest) values are assigned to quintile 1 (0). t-statistics, shown in parentheses, are based on two-way cluster robust standard errors, clustered by firm and quarter. ***, **, and * indicate significance at the 1, 5, and 10% level, respectively. Panel A: Interaction Effects with Market-to-Book (MTB) RET(-1,+1) SUE Intercept 0.087-0.532*** (0.55) (-11.83) PAR -0.632*** 0.211*** (-6.97) (8.18) SENT 0.382-0.138** (1.50) (-2.33) PAR*SENT -0.451** -0.079** (-1.97) (-2.07) PAR*SENT*HighMTB -0.622*** -0.013 (-2.87) (-0.23) HighMTB -0.156* 0.137*** (-1.91) (6.34) SIZE 0.426*** 0.135*** (4.01) (4.53) MOMEN 0.403*** 0.856*** (2.86) (23.18) R-square 0.003 0.035 Panel B: Interaction Effects with Turnover (TO) RET(-1,+1) SUE Intercept 0.204-0.506*** (1.21) (-9.59) PAR -0.773*** 0.216*** (-6.50) (6.10) SENT 0.207-0.168** (0.79) (-2.04) PAR*SENT -0.177-0.068 (-0.83) (-1.15) PAR*SENT*HighTO -0.512** -0.086* (-2.11) (-1.88) HighTO -0.170-0.031 (-1.50) (-1.47) SIZE 0.381*** 0.168*** (3.47) (5.84) MOMEN 0.390*** 0.864*** (2.73) (23.39) R-square 0.003 0.034 40

Table 8: Expected Announcement Window Returns Panel A presents market-adjusted returns across portfolios of pre-announcement returns measured prior to the expected announcement date. We calculate 'expected' announcement dates using the methodology detailed in Cohen et al. (2007). Column (1) contains the pooled results when using the full sample of expected announcement dates. Column (2) contains the results for firms that announced earnings prior to day t-1, where t denotes the expected announcement date. Column (3) contains the results for firms that announcement earnings within one day of the expected announcement date. Column (4) contains the results for firms that announced earnings after day t+1. Panel B contains regression results of expected announcement window returns. LBM and SIZE are the log of one plus the book-tomarket ratio and log of market capitalization, respectively. MOMEN equals the firm s market-adjusted return over the six months ending on t-10. ACC is total accruals scaled by prior quarter total assets. FScore is the summary score of changes in firm s financial condition as detailed in Piotroski (2000) ranging in value from 0 to 9. PastSURPRISE equals the SURPRISE from the prior calendar quarter. All control variables, except for FScore, are assigned to quintiles where the highest (lowest) values are assigned to quintile 1 (0) t-statistics, shown in parentheses, are based on two-way cluster robust standard errors, clustered by firm and quarter. ***, **, and * indicate significance at the 1, 5, and 10% level, respectively. Panel A: Averages Across Quintiles of All Pre-Expected-Announcement Returns (1) (2) (3) (4) All Early On-Time Late 1 (Low RET(k-5,k-3)) 0.540 0.623 0.753 0.132 2 0.289 0.115 0.463 0.114 3 0.147-0.047 0.258 0.038 4 0.158 0.043 0.288-0.054 5 (High RET(k-5,k-3)) -0.216-0.025-0.336-0.203 Low-High 0.757 0.649 1.089 0.334 Low-High P-Value 0.000 0.000 0.001 0.011 Panel B: Regression Results of Pre-Expected-Announcement Returns (1) (2) (3) (4) Intercept 0.622*** 0.386*** 0.367** 0.391** (6.25) (2.91) (2.39) (2.22) RET(k-5,k-3) -0.151*** -0.151*** -0.151*** -0.153*** (-6.36) (-6.40) (-6.40) (-6.45) LBM 0.060** 0.060** 0.065** (2.33) (2.38) (2.56) SIZE 0.059*** 0.059*** 0.041** (3.06) (3.11) (2.21) MOMEN 0.009 0.000 (0.31) (0.01) ACC -0.088*** (-5.26) FScore 0.035** (2.39) PastSURPRISE 0.001 (0.06) R-square 0.001 0.001 0.001 0.002 41