When Low Beats High: Riding the Sales Seasonality Premium

Transcription

1 When Low Beats High: Riding the Sales Seasonality Premium Gustavo Grullon Rice University Yamil Kaba Rice University Alexander Núñez Lehman College May 23, 2018

2 When Low Beats High: Riding the Sales Seasonality Premium Abstract We demonstrate that sorting stocks on sales seasonality predicts future abnormal returns. A longshort strategy of buying low-sales-season stocks and shorting high-sales-season stocks generates an annual alpha of 8.4%. Further, this strategy has become stronger over time, generating an annual alpha of approximately 15% over the last decade. This seasonal effect predicts future stock returns in cross-sectional regressions, and is independent of previously documented seasonal anomalies. Moreover, the alphas from this trading strategy cannot be explained by differences in stock market liquidity, systematic risk, asymmetric information, or financing decisions. Further tests indicate that this phenomenon may be driven partially by seasonal fluctuations in the level of investor attention. 1

3 I. Introduction While some firms have relatively steady sales throughout the year, others display highly seasonal sales patterns that are unavoidable and mainly reflect the dynamics of the firms product markets. For example, the popular restaurant chain Denny s has relatively stable sales throughout the year, and generates approximately 25% of its sales each quarter. On the other hand, H&R Block, a tax service provider, generates 68% of its sales during tax season and only 7% during its low season. For most firms, these seasonal fluctuations in sales are highly predictable and well known in advance; therefore, demonstrating that this publicly available information can forecast future stock returns would pose a significant challenge to the efficient markets hypothesis. Using a simple measure of sales seasonality, we find that low-sales-season firms tend to significantly outperform high-sales-season firms. A long-short strategy of buying stocks at the beginning of the given firms low season, and shorting them at the beginning of their high season, generates an annual alpha of 8.4% with a t-statistic of This long-short spread remains basically unaffected even after controlling for Fama and French s (2015) five factors. We also find that these alphas come primarily from the long leg of the portfolio. In value-weighted portfolios, the lowest-sales-season decile has an annual excess return of approximately 10.4%, while the highest-sales-season decile has an annual excess return of approximately 2%. Unexpectedly, the seasonality premium documented in our results appears to have strengthened over time. Over the periods and , the annual average alpha was approximately 5.4%; over the period , the alpha increased to 9.4%; and during the period , it grew to 15%. Moreover, our results are stronger among firms that experience extreme swings in quarterly sales. Specifically, we find that among firms more likely to fluctuate between the highest and 2

4 lowest seasonality deciles, the annual alpha of our trading strategy equals 15.2%. Furthermore, we find no evidence that the predictive power of sales seasonality has already been captured by other well-known factors. A factor based on the sales-seasonality strategy has no correlation with any of the Fama-French (2015) five factors or momentum. More significantly, because the salesseasonality strategy has a comparatively low standard deviation, it has the highest Sharpe ratio of any of the main factors identified in the asset-pricing literature. Using Fama-Macbeth (1973) regressions, we show that sales seasonality is an important determinant of the cross-section of average stock returns. Even after controlling for other factors, the coefficient of sales seasonality is negative and statistically significant. To investigate how sales seasonality interacts with other firm characteristics, we estimate alphas for independently double-sorted portfolios: three noteworthy results emerge from this analysis. First, we find that the average abnormal returns of the low-season portfolios are higher among large firms, growth firms, profitable firms, and high-investment firms. This evidence indicates that neither small nor distressed firms are the source of the abnormal alphas. Second, while the average abnormal returns of the low-season portfolios are significantly higher among high-momentum stocks, those of the high-season portfolios are significantly lower among lowmomentum stocks. Thus the long leg of the momentum strategy performs better during low-sales seasons while the short leg of the momentum strategy performs better during high-sales seasons. Buying high-momentum stocks just before a low-sales season, and shorting low-momentum stocks just before a high-sales season, generates an annual alpha of approximately 18.34%. Finally, we find evidence that the sales-seasonality anomaly interacts with the level of operating accruals. The average abnormal returns of the low-season portfolios are significantly higher among stocks with low operating accruals, while the abnormal returns of the low-season portfolios are significantly 3

5 lower among stocks with high operating accruals. Overall, our evidence asserts that this salesseasonality trading strategy amplifies the effect of other existing asset-pricing anomalies. We also show that our empirical results are independent of previously documented seasonal anomalies. Chang et al. (2016) construct a seasonality measure based on earnings announcements and find that high earnings seasonality stocks outperform low earnings seasonality stocks during the earnings announcement month. Although Chang et al. also examine a seasonal effect in a fundamental variable, our paper differs in two crucial areas. First, their analysis focuses on the month of the earnings announcements following high- or low-earnings seasons, whereas ours focuses on the quarters in which firms sales are low or high relative to their total annual sales. Second, the investment strategies are diametrically different. While the earningsseasonality strategy goes long on high-season stocks and short on low-season stocks, our strategy does the opposite. Therefore, the sales-seasonality anomaly captures a different asset-pricing effect from the one documented in Chang et al. Consistent with this distinction, we find that after removing all earnings-announcement firm-month observations from our sample, the salesseasonality premium remains economically and statistically significant. Overall, our findings show that both the sales and earnings anomalies coexist in the cross-section of stocks returns. In another related paper, Heston and Sadka (2008) find that firms tend to have comparatively high (or low) abnormal returns every year in the same month. The main difference between their paper and ours is their use of a return-based measure of seasonality; in contrast, we use a seasonality measure determined mainly by exogenous product market forces, which are often beyond firms control. This distinction is important because Keloharju, Linnainmaa, and Nyberg (2016) show that return seasonalities appear to be driven primarily by seasonalities in systematic factors. Our measure of sales seasonality is demonstrably independent of the return seasonality 4

6 documented in Heston and Sadka (2008). Regardless of the definition we use to measure return seasonality, removing firms from the top and bottom return-seasonality deciles and repeating our analysis, we find slightly stronger results for our value-weighted portfolios. Furthermore, contrary to the results in Keloharju, Linnainmaa, and Nyberg (2016), we find the sales-seasonality strategy has a lower standard deviation than other well-known factors, and does not appear to be driven by a time-varying exposure to systematic risk. Overall, the sales-seasonality anomaly evidently has different characteristics from those based on return seasonalities. Recent studies argue that the discovery of new factors could be the outcome of data mining. For example, Harvey (2017) shows that if a stock market anomaly is relatively rare, standard statistical tests can produce many false positives (type-two errors). Hou, Xue, and Zhang (2017) use this premise to examine over 400 anomalies; they find that many anomalies are in fact a subset of other common anomalies. Furthermore, they show that many anomalies are not robust to the use of value-weighted portfolios, the elimination of microcaps, and/or the use of t-statistics above 3 to determine statistical significance. However, our empirical results are verifiably robust to all of these methodological refinements. Our results remain robust, for instance, after the use of either NYSE breakpoints or value-weighted portfolios, due to the sales-seasonality effect being primarily driven by large firms. Additionally, we establish that the positive alphas generated by the sales-seasonality strategy do not result from cross-sectional differences in stock market liquidity, liquidity shocks, proxies for asymmetric information (e.g., analysts forecast dispersion or idiosyncratic volatility), or financing decisions. We also run time-series regressions interacting market returns with seasonal dummies, and find no evidence of seasonal variations in market betas. To further investigate whether the sales-seasonality premium is a compensation for systematic risk, we 5

7 examine the performance of our investment strategy during economic expansions and recessions and find that the seasonality premium is positive across the entire business cycle. Finally, contrary to the claim that the sales-seasonality premium is a compensation for tail risk, we find that lowseason stocks significantly outperformed high-season stocks during the recent financial crisis. The annual alpha of the sales-seasonality strategy during the period was equal to 17.58%. In general, reconciling our results with a risk-based explanation is difficult, as recent evidence indicates firms are more susceptible to systematic shocks during their high-sales seasons (Almeida, Carvalho, and Kim 2017). Therefore, given that average returns tend to be lower when firms are more sensitive to external shocks, the claim that the sales-seasonality premium is a compensation for time-varying systematic risk is unlikely. We do find weak evidence for the seasonal-sales effect being partially related to a seasonal-investment effect. Specifically, our results indicate that firms are likely to invest more in the period preceding their high-sales seasons. This pattern suggests the sales-seasonality premium may be related to the asset-growth anomaly documented by Cooper, Gulen, and Schill (2008). We test for this possibility by examining the return correlation between portfolios based on quarterly asset-growth changes, and portfolios based on sales seasonality. This analysis finds that although the returns of low-sales-season firms are uncorrelated with those of low-asset growth firms, the returns of high-sales-season firms are positively correlated with those of high-asset growth firms (ρ=0.38). Despite these findings implying that a seasonal asset-growth anomaly may partially explain the sales-seasonality anomaly, this explanation cannot be entirely satisfactory, as a large portion of the premium documented in this paper originate in low-sales-season firms. Finally, we investigate whether our empirical findings are driven by seasonal fluctuations in the level of investor attention. If information about firms fundamentals is more valuable during 6

8 high-sales seasons than low-sales seasons, then investors may shun low-sales-season firms, since the marginal cost of acquiring information during these periods potentially exceeds its marginal benefit. To test this hypothesis, we use the post-earnings announcement drift as a proxy for the level of investor attention (see, e.g., Hirshleifer, Lim, and Teoh 2009). If higher levels of investor attention make stock prices more efficient, then the post-earnings drift should be stronger during low seasons than during high seasons. Consistent with this prediction, we find that the postearnings drift is significantly stronger during low-sales seasons. The low (high) season portfolio generates a post-earnings drift of 1.66% (0.62%) over the 45 days immediately following the announcement. To ascertain whether this effect is stronger conditional on the direction of the announcement surprise, we split the sample into positive and negative earnings surprises based on the cumulative abnormal returns around the earnings announcement ( 1 to 1). Low (high)-sales season stocks with positive earnings surprises generate a post-earnings drift of 1.65% (0.43%). On the other hand, low (high)-sales season stocks with negative earnings surprises generate a postearnings drift of 1.22% (0.34%). This evidence argues for investors paying less attention to stocks during those quarters in which firms generate the smallest fraction of their sales. As additional confirmation for the predictions of the investor inattention hypothesis, we find that sales seasonality is positively correlated with the total number of SEC filings downloaded from EDGAR ( before the earnings announcement. More important, using sales seasonality as an instrument for the number of downloads, we find that the magnitude of the post-earnings drift is negatively correlated with the instrumented variable. This evidence indicates that the stronger post-earnings drift during low-sales seasons is partially driven by investor inattention. One possible interpretation of our results entails the sales-seasonality 7

9 premium serving as a compensation to investors to hold temporarily neglected stocks, as in Merton (1987). The remainder of the paper is as follows: In Section II we describe the data and methodology used in our analysis. In Section III we present our main empirical results and document the sales-seasonality anomaly. In Section IV we examine how the sales-seasonality effect interacts with other firm characteristics. Section V examines several viable explanations for the sales-seasonality anomaly. Section VI concludes the paper. Section II. Data and Methodology Our main sample consists of all non-financial U.S. stocks on the Center for Research in Security Prices (CRSP) files listed on the NYSE, AMEX, and NASDAQ exchanges. 1 Our sample excludes stocks without CRSP shares codes 10 or 11 (common equity). To construct our accounting variables, we merged these data with the Compustat annual and quarterly fundamental files. All yearly variables are formed at the end of June in year t, using accounting information from Compustat at the end of fiscal year t-1. This selection process generates a final sample of 1,482,860 firm-month observations and 13,617 firms over the period 1970 to To measure sales seasonality, we construct a quarterly variable (SEA) equal to the sales in quarter q of year t scaled by the annual sales in year t: SEA qt = SALES qt /ANNUALSALES t Before calculating this measure, we remove any observation with negative quarterly sales or non-positive annual sales. We also require firms to have non-missing sales data for all four quarters during a particular year. When the sum of the quarterly sales from the Compustat quarterly files does not equal the total sales from the annual files, we include only firm-year 1 All our main results are unaffected by including financials (firms with SIC codes between 6000 and 6999). 8

10 observations in which the sum of the quarterly sales is within 95% and 105% of the total annual sales. 2 To mitigate any potential impact of outliers, we calculate AVGSEAqt, which is equal to the average of SEAqt over years t-2 and t-3. 3 We then use AVGSEAqt in year t-2 to predict SEAqt in year t to ensure that the seasonality data is available to investors at the time of the portfolio formation. Interpreting this seasonality measure is essentially straightforward: For firms with steady sales throughout the year, SEAqt would be approximately equal to 25% in all quarters. In contrast, firms with extremely seasonal sales would have a SEAqt above 25% during their high seasons and SEAqt below 25% during their low seasons. Table 1 reports summary statistics for portfolios sorted on our measure of sales seasonality. Note that the firms in the lowest and highest deciles have similar characteristics. This finding is, however, reasonable since the firms in those portfolios are likely to be the same firms. For example, about 80% of the low-season firms end up in the high-season portfolio in the same calendar year. Moreover, Figure 1 shows that firms in the lowest-seasonality decile generate on average 17.4% of their sales during their low seasons, while those in the highest-seasonality decile generate, on average, 33.8% of their sales during their high seasons. INSERT TABLE 1 and FIGURE 1 HERE To assess the persistence of SEAqt over time, Table 2 reports the fraction of firms in the lowest- and highest-seasonality portfolios at time t that remain in the same portfolios at time t+1, t+3, t+5, t+10, and t+20. The evidence shows that 64.90% (66.68%) of the firms in the low (high) sales seasonality decile stay in the same portfolios after one year. Even two decades following the portfolios formation, the firms in the lowest (highest) sales seasonality bucket have a 31.46% 2 All our main results are unaffected by requiring that the sum of the quarterly sales must be equal to the total annual sales. 3 Our results are insensitive to the number of years we use to calculate the average. 9

11 (35.12%) probability of remaining in their initial bucket. Table 2 also shows the comparatively small probability of a firm moving from the lowest to the highest decile, or vice versa, during the same quarter of the year. For example, the probability of a firm moving from the lowest (highest) decile at time t to the highest (lowest) decile at time t+1 is less than 3%. This persistence in sales seasonality is reasonable given that seasonal variations in sales reflect the dynamics of firms product markets, which are mainly out of managers control. INSERT TABLE 2 HERE Section III. The Sales-Seasonality Premium To examine the predictive power of sales seasonality, we allocate stocks at the beginning of each month into deciles based on the level of their ex-ante measure of sales seasonality, AVGSEAqt in year t-2. The portfolios are held for one month and then rebalanced at the beginning of the next month. Following the methodological recommendations in Hou, Xue, and Zhang (2017), we measure the performance of these portfolios using value-weighted (VW) returns. 4 We use four alternative asset-pricing models to control for the effect of prominent factors: (i) CAPM; (ii) Fama-French (1993) three-factor model; (iii) Carhart (1997) four-factor model; and (iv) Fama- French (2015) five-factor model. Table 3 reports the alphas and factor loadings from our trading strategy. Table 3 Panel A shows that the firms in the lowest decile of sales seasonality earn an average annual excess return of 10.44% (t-stat = 3.28), whereas the firms in the highest decile earn an annual average excess return of approximately 2.04% (t-stat = 0.67). The last column of Table 3 shows the annual spread between the lowest and the highest decile is both statistically and economically significant (8.4% with a t-stat of 4.61). The results in Table 3 Panels B, C, D, and E illustrate low-season firms 4 We find our results to be robust to the use of equally-weighted returns. However, the use of equally-weighted returns generates smaller alphas, indicating that our main results are driven by larger firms. 10

12 significantly outperforming high-season firms even after controlling for prominent factors. For example, when we control for risk using the Fama-French (2015) five-factor model (see Table 3 Panel E), the lowest decile generates an annual alpha of 6.84% (t-stat = 4.90) while the highest decile generates an annual alpha of 1.68% (t-stat = 1.12). Notably, the annual spread between the lowest and the highest decile is unaffected by the inclusion of the five factors (8.4% with a t- stat of 4.52). INSERT TABLE 3 HERE Hou, Xue, and Zhang (2017) argue that calculating breakpoints based on the NYSE- AMEX-NASDAQ universe generally underrepresents large firms in the extreme deciles of portfolio sorts. By using only NYSE firms to determine the breakpoints, they show this bias towards microcaps in the extreme portfolios to be less severe. To address any concerns regarding our trading strategy being partially driven by stocks that are relatively difficult to arbitrage, we repeat our tests using NYSE breakpoints. Table 4 reports the results from this alternative methodological approach. Consistent with our previous evidence, we find the spread between the lowest- and highest-seasonality deciles remains economically and statistically significant. For example, the Fama-French five-factor annual alpha for the long-short portfolio is equal to 5.40% (t-stat = 3.29). INSERT TABLE 4 HERE We also perform a test to determine whether the volatility of quarterly sales plays a role in the sales-seasonality premium. Despite an obvious mechanical relation between seasonality and sales volatility, sorting stocks exclusively by the level of sales seasonality may conceal the potential effect of sales volatility on the sales-seasonality premium. As illustration, posit two firms, both of which generate 40% of their sales during their high season. Then suppose the first 11

13 firm generates 20% of its sales during its low season, and the second generates only 1%. If we sort stocks using AVGSEAqt, both firms are likely to appear in the highest/lowest seasonality portfolio. Nevertheless, the firm with the largest dispersion in quarterly sales may generate a seasonality premium different from that of the firm with the lowest dispersion. To disentangle the level effect from the volatility effect, we form five portfolios based on AVGSEAqt, and then divide each of these portfolios into five new portfolios based on SEARANGE qt, which is equal to the difference of the maximum and minimum values of AVGSEA q over the last 4 quarters: 5 SEARANGE qt = max (AVGSEA q n) min (AVGSEA q n). n=0 to 3 n=0 to 3 Table 5 reports the alphas for these double sorts with their corresponding t-statistics. Table 5 Panel A illustrates firms with more extreme swings in quarterly sales earning higher returns within the lowest seasonality quintile. Firms in their lowest sales season and lowest SEARANGE quintile earn an average Fama-French three-factor annual alpha of 0.90% (t-stat = 0.60), while those in the highest SEARANGE quintile earn 4.18% (t-stat = 2.50). Significantly, we find the opposite effect among firms in the highest sales-season quintile, where the firms in the lowest SEARANGE quintile earn an average annual alpha of 1.16% (t-stat = 0.91), and those in the highest SEARANGE quintile earn approximately 10.52% (t-stat = 5.49). The last column of Table 5 Panel A exhibits that within high SEARANGE firms, a trading strategy of buying lowsales-season stocks and shorting high-sales-season stocks produces an annual alpha of 14.7% (tstat = 6.20). Table 5 Panel B shows this spread is slightly larger when we use the Fama-French five-factor model (alpha = 15.2% with a t-stat = 6.17). 6 INSERT TABLE 5 HERE 5 We find similar results if we use the standard deviation of quarterly sales. 6 Untabulated results. 12

14 To ensure our results hold over different time periods, Table 6 reports alphas from the sales seasonality strategy by decades. Table 6 Panel A shows that over the period , the annual Fama-French five-factor risk-adjusted alpha derived from buying low-sales season stocks and shorting high-sales season stocks is roughly 5.4% (t-stat = 1.64). On the other hand, over the period , the same investment strategy generates an annual alpha of 15% (t-stat = 3.94). Table 6 Panel B shows that within the high SEARANGE firms, the profitability of buying lowseason firms and shorting high-season firms also increased over time. The annualized riskadjusted alpha derived from this strategy increased from 8.72% (t-stat = 2.56) over the period , to 16.68% (t-stat = 4.53) over the period INSERT TABLE 6 HERE To further examine the predictive power of sales seasonality, we run Fama and MacBeth (1973) regressions of monthly stock returns on the following variables: (i) our measure of sales seasonality; (ii) book-to-market; (iii) the log of the market value of equity; (iv) investment scaled by assets; (v) gross profits scaled by assets; and (vi) momentum. Consistent with our previous findings using portfolio sorts, Table 7 demonstrates sales seasonality is negatively correlated with futures stock returns. Note that the coefficient of sales seasonality becomes more statistically significant after controlling for the other firm characteristics. INSERT TABLE 7 HERE Section IV. Double sorts with known factors To rule out the possibility that sales seasonality is a subset of an existing asset-pricing anomaly, we estimate the alphas of this trading strategy using double-sorted portfolios, using a wide array of firm characteristics for this analysis. Specifically, portfolios are formed by sorting on size, book-to-market, gross profitability, investment-to-assets, momentum, and operating 13

15 accruals. Next we independently sort on our measure of sales-seasonality SEA, and assign them to quintiles based on their NYSE breakpoints. The results for these portfolios are shown in Table 8. For all cases, we report the average monthly value-weighted returns using Fama-French threeand five-factor models. 7 INSERT TABLE 8 HERE IV.A Double sorts on size In this subsection we examine how the sales-seasonality anomaly performs across different portfolios sorted on firm size. This analysis is significant, because if the sales-seasonality premium is driven mainly by small firms, then limits to arbitrage may prevent investors from exploiting this trading strategy. Sorting on size reveals that the sales seasonality premium is close to zero among the smallest firms. Specifically, Table 8 Panel A shows that the Fama-French five-factor model generates insignificant alphas for both low- and high-sales-season portfolios in the lowest size quintile. While the sales-seasonality effect is absent among the smallest firms, the evidence shows that the magnitude of the sales-seasonality premium is stronger among larger firms. High-season firms generate lower alphas in the size quintiles 2, 3, and 4, and low-season firms generate higher alphas in size quintiles 3, 4, and 5. Overall, we find no evidence for the sales-seasonality premium being a small-firm effect. IV.B Double sorts on book-to-market Table 8 Panel B presents the results based on book-to-market sorts. The value-weighted portfolio results indicate low-season firms generating higher alphas in the lowest book-to-market quintiles, while the underperformance of the high-season firms is strongest in the mid book-to- 7 In the case of momentum, we report alphas using Fama-French three-factor + momentum model. 14

16 market quintile. The low-sales-season portfolio generates a positive and significant Fama-French five-factor annual alpha of 4.68% (t-stat = 3.91) in the lowest book-to-market quintile, and the high-sales-season portfolio generates negative and significant annual alphas of 2.88% (t-stat = 2.31) and 5.04% (t-stat = 3.52) in the book-to-market quintiles 2 and 3, respectively. Significantly, we find no sales-seasonality premium in the highest book-to-market quintile, which indicates the sales-seasonality premium is unrelated to the value premium. In fact, our findings demonstrate the sales-seasonality premium is much stronger among growth firms. IV.C Double sorts on profitability Recent evidence posits profitability as an important determinant of the cross-section of stock returns. For example, using gross profits-to-assets as a measure of profitability, Novy-Marx (2013) finds that more (less) profitable firms tend to earn higher (lower) future returns. To investigate whether the sales-seasonality premium interacts with the profitability premium, we report alphas for portfolios independently double-sorted on profitability and sales seasonality. The results from this analysis, reported in Table 8 Panel C, reveal low-sales season alphas to be positive in profitability quintiles 1 and 4, and strongest in quintile 5. 8 The Fama-French five-factors annual alphas for the low-sales-season firms are equal to 2.76% (t-stat=1.99) in the first quintile, 3.72% (t-stat=2.18) in the fourth quintile, and 5.52% (t-stat=3.87) in the fifth quintile. Evidence also supports the annual alphas of the high-sales-season firms as being more negative among less profitable firms. The Fama-French five-factors annual alpha for the high-season firms is 4.03% (t-stat= 2.64) in the second profitability quintile. IV.D Double sorts on investments 8 For robustness, we also use operating profitability (Fama French (2015)) and cash-based operating profitability (Ball, Gerakos, Linnainmaa, and Nikolaev (2016)) as proxies for profitability, and find similar results. 15

17 Copper, Gulen, and Schill (2008) provide evidence that firms with high asset growth (aggressive firms) underperform firms with low asset growth (conservative firms). The assetgrowth premium may be related to the sales-seasonality premium because firms appear to invest just before the beginning of a high-sales season. Table 8 Panel D reports the alphas for portfolios independently double sorted on asset growth and sales seasonality. This analysis shows that the low-sales-season effect is slightly stronger among aggressive firms. The Fama-French five-factors annual alphas for low-season firms are equal to 2.96% (t-stat=1.90) in quintile 1 and 2.64%% (tstat=1.88) in quintile 2, while the annual alphas are equal to 4.34% (t-stat=2.89) in quintile 4 and 4.37%% (t-stat=3.39) in quintile 5. However, although this analysis also concludes that conservative firms will likely underperform aggressive firms during high-sales seasons, the alphas of the high-sales-season firms, despite being negative, are statistically insignificant across all investment quintiles. Overall, our findings show that aggressive firms outperform other firms during their low-sales seasons. IV.E Double sorts on momentum We also explore the interaction between momentum and sales seasonality. In this analysis, we add the momentum factor to the Fama-French three-factor model. The results are shown in Table 8 Panel E. The low-sales season effect is present across all momentum quintiles. However, analogous to the investment findings, high-season firms have negative and insignificant alphas across all momentum quintiles. When we use the Fama-French five-factor model, the alpha of the low-season portfolio is significantly higher in the highest-momentum quintile, while the alpha of the high-season portfolios is significantly lower in the lowest-momentum quintile. Buying highmomentum stocks just before a low-sales season, and simultaneously shorting low-momentum stocks just before a high-sales season, generates an annual alpha of approximately 18.34%. 16

18 IV.F Double sorts on operating accruals Existing factor models have failed to explain the alphas generated by portfolios sorted on accruals (Hou, Xue, and Zhang 2015, Fama and French 2015). Given that firms are likely to increase their accruals in preparation for a high sales season, sales seasonality may be related to the accrual anomaly. Table 8 Panel F shows the results of double-sorted portfolios based on operating accruals. As predicted by Sloan (1996), firms with high operating accruals earn lower average returns than those with low operating accruals. Low-sales-season positive alphas are present in all but the mid accruals quintile. The Fama-French five-factors annual alphas for these portfolios are 6% (t-stat=3.64), 2.4% (t-stat=1.83), 2.88% (t-stat=2.20), and 3.43% (t-stat=2.22), respectively. High-sales-season negative alphas are present on the largest accrual quintiles. The Fama-French five-factor alphas for these portfolios are 2.76% (t-stat= 2.17) in quintile 4 and 5.52% (t-stat= 3.76) in quintile 5. Significantly, although high accruals have been associated with negative alphas, low-season firms with high accruals have positive and significant alphas. IV.G Correlation with other asset pricing factors Ascertaining whether the sales-seasonality effect is related to existing asset-pricing factors is crucial. To directly investigate this issue, we estimate Pearson s correlation coefficients between existing asset-pricing factors and a factor based on the sales-seasonality strategy (SEAF). SEAF is equal to the returns of a low-season portfolio minus the returns of a high-season portfolio. Factors are formed using Fama-French convention. 9 We report correlations between the salesseasonality factor (SEAF) and the following six factors: (i) market excess return (MKTrf); (ii) size s small-minus-big (SMB); (iii) book-to-market s high-minus-low (HML); (iv) investment s conservative minus aggressive (CMA); (v) profitability s robust minus weak (RMW); and (vi) 9 In the appendix, we include detailed description of how these factors are created. 17

19 momentum s up-minus-down (UMD). Table 9 shows that the seasonality factor is uncorrelated with all the most important asset pricing factors in the literature. INSERT TABLE 9 HERE We also compare means, standard deviations, and the mean scaled by the standard deviation, which is a measure akin to a Sharpe ratio. The factor with the highest annualized mean is UMD with 8.4%, followed by a tie between RMW, CMA, and SEAF with a mean of 3.6%. When we scale the mean by the standard deviation, the factor with the highest annualized ratio is SEAF with 0.69, followed by UMD with 0.55, and CMA with Interpreting this ratio as a compensation per unit of risk indicates that SEAF has the strongest performance among all factors. Section V. Possible Sources of the Sales-Seasonality Premium V.A Ruling out other Seasonalities Distinguishing the sales seasonality from other seasonal anomalies is of paramount importance. Although our sales-seasonality measure appears related to Chang et al. s (2016) earnings seasonality, it is not. First, their analysis focuses on the month of the earnings announcements following high or low earnings seasons, while ours focuses on the quarters in which firms sales are low or high relative to their total annual sales. Second, the investment strategies themselves are diametrically different: the earnings-seasonality strategy goes long on high-season stocks and short on low-season stocks, whereas our strategy goes short on high-season stocks and long on low-season stocks. However, to ensure these strategies are unrelated, we remove all earnings announcement firm-month observations from our sample. This effectively removes Chang et al. s (2016) sample from ours. We run asset-pricing tests sorting on sales seasonality (SEA) and present these results in Table 10. To be consistent with the methodology 18

20 in Chang et al. (2016), we use NYSE, AMEX, and NASDAQ breakpoints for this analysis. 10 The lowest sales-seasonality decile has a mean annualized excess return of about 7.48% while the highest sales-seasonality decile has a mean annualized excess return of about 1.54% for valueweighted portfolios. The long-short Fama-French five factor annual alpha is equal to 5.30% with a t-stat of In addition, we replicate their investment strategy using our measure of seasonality and find insignificant excess returns and alphas. INSERT TABLE 10 HERE We also show that the sales seasonality effect is unrelated to the return seasonality documented by Heston and Sadka (2008). They find that firms tend to have relatively high (or low) abnormal returns every year in the same month. The main difference between their paper and ours is that Heston and Sadka (2008) use a return-based measure of seasonality, whereas we use a seasonality measure determined primarily by exogenous product market forces that are often beyond firms control. This distinction is significant because Keloharju, Linnainmaa, and Nyberg (2016) show that return seasonalities are driven chiefly by seasonalities in systematic factors. For robustness, we replicate Heston and Sadka s (2008) measure of seasonality and drop their highest and lowest deciles from our sample. Using the remaining firms, we then create sales seasonality deciles and calculate the spread between low-season and high-season firms. Heston and Sadka (2008) use several definitions based on different lags to find evidence of predictable behavior of returns; consequently, we use 12, 36, and 60 months to make comparisons with our measure of sales seasonality. 11 The results are shown in Table 11. In all cases, we report Fama- 10 We run the analysis using NYSE breakpoints only and find qualitatively similar results , 36, and 60 months refers to the 1-,3-,and 5-month average of the previous same month returns. 19

21 French five-factor alphas and t-stats for portfolios. 12 For convenience, Table 11 Panel A shows the alphas and t-stats for our measure of sales seasonality, as reported in Table 4 Panel D. INSERT TABLE 11 HERE We find that regardless of the particular definition of return seasonality employed, the sales-seasonality premium remains the same. Table 11 Panel B shows that when the definition of returns-seasonality is based on the firm s prior 12-month returns, our long-short portfolio generates an alpha of approximately 7.80%. When the definition of return seasonality is based on either 36 or 60 months, the long-short portfolio generates annual alphas of 8.76% and 7.44%, respectively. In all cases the t-statistics are greater than 3. Table 11 Panel C reports portfolios sorted on return-seasonality. It presents annual crosssectional autocorrelation patterns, since, as documented in Heston and Sadka (2008), stocks tend to have relatively high or low returns every year in the same calendar month. Using 12, 36, and 60 lags to define return seasonality, the long-short portfolio has an annualized Fama-French five factor alpha of 7.2% (t-stat=2.42), 10.92% (t-stat=4.00) and 11.28% (t-stat=3.71), respectively. We then repeat the test in reverse order: we first drop the lowest and highest deciles from the sales-seasonality strategy, and then sort on return seasonality. As before, we use three definitions of autocorrelation in returns, i.e., 12, 36, and 60 lagged months, to define return seasonality. Table 11 Panel D reports the result from this analysis. Significantly, the results in Heston and Sadka (2008) become slightly stronger after removing the sales-seasonality effect. In an unreported analysis, we also find that a factor based on the return-seasonality is uncorrelated with a factor based on the sales-seasonality premium. Overall, the evidence in this subsection 12 We report NYSE, AMEX, and NASDAQ breakpoints results to be consistent with Heston and Sadka (2008) and Keloharju et al. (2016). We repeat the analysis with NYSE breakpoints and find qualitatively similar results. 20

22 indicates that neither earnings seasonality nor return seasonality drive our results. In the following section, we test alternative explanations for the phenomenon detailed in this paper. V.B Alternative explanations In this subsection, we perform tests to identify the potential drivers behind our main empirical findings. Specifically, we examine whether cross-sectional differences in stock market liquidity, liquidity shocks, proxies for asymmetric information, and/or financing decisions can explain the sales-seasonality premium. Table 12 reports summary statistics of (i) Amihud s illiquidity measure (Amihud 2002); (ii) liquidity shocks (Bali, Peng, Shen, and Tang 2014); (iii) share turnover (Datar, Naik, and Radcliffe, 1998) (iv) idiosyncratic volatility (Ang, Hodrick, Xing, and Zhang 2006); (v) net stock issues (Pontiff and Woodgate 2008); (vi) most recent earnings announcement market surprise (Chan, Jegadeesh, and Lakonishok 1996); (vi) analyst dispersion (Diether, Malloy, and Scherbina 2002); and (vii) asset growth (Cooper, Gulen, and Schill 2008). If low-season firms are on average less liquid than high-season firms, then the positive alpha generated by the sales seasonality strategy could be attributed to a liquidity premium (Amihud 2002). However, we do not find evidence supporting this potential explanation. Table 12 shows that although firms in the lowest seasonality decile have the highest level of illiquidity (0.194), firms in the highest-seasonality decile have the second-highest level of illiquidity (0.182). The last column of Table 12 shows that the difference in Amihud s illiquidity measure between the highest- and lowest-seasonality deciles is not statistically different from zero. This finding confirms that cross-sectional differences in the level of illiquidity cannot explain our main findings. INSERT TABLE 12 HERE 21

23 We subsequently examine whether there are any differences in the average magnitudes of liquidity shocks. Bali et al. (2014) find that following a positive (negative) shock to liquidity, a firm will earn higher (lower) future returns. While firms in the high-sales season portfolio tend to experience large negative liquidity shocks, we also find that firms in the low-sales season portfolio experience the largest negative liquidity shocks. Therefore, the phenomenon documented in Bali et al. (2014) cannot explain the sales-seasonality premium. Further, we investigate whether cross-sectional differences in share turnover can explain our main results. Evidence exists for firms with higher (lower) share turnover having lower (higher) returns due to several factors related to market frictions (Datar, Naik, and Radcliffe, 1998). However, Table 12 shows that the difference in share turnover between the highest- and lowestseasonality deciles is not statistically different from zero. We also explore the relation between corporate investment and sales seasonality. Table 12 shows that firms tend to invest more aggresively prior to high seasons than during low seasons. While the low-season fims grow at a quarterly rate of 3.2%, high-season firms grow by a rate of 6.1%. This supports the idea that firms prepare in advance for their highest sale season. To determine if a relation exists between quarterly assets growth and sales seasonality returns, we form ten portfolios in month t based on the average quarterly asset growth in quarters q-5 and q We calculate value-weighted excess returns for each portfolio, and estimate the correlation between low-season portfolio returns and low asset-growth portfolio returns, and then between high-season portfolio returns and high asset-growth portfolio returns. While the low-season portfolio has a correlation of nearly 0% with the low asset-growth portfolio, the high-season 13 Because the data for the quarter right before the investment at time t may introduce a look-ahead bias, we use the average of the quarterly asset growth rates in q-5 and q-9 to construct our investment portfolio deciles. If investment is seasonal, this lagged measure should proxy for future investment. 22

24 portfolio has a correlation of nearly 40% with the high asset-growth portfolio. Our finding of this high season/high asset-growth correlation supports the claim for seasonal firms being more likely to make their investments in the quarter prior to their high season, and, in turn, earn lower returns during the high season. Nevertheless, this explanation cannot explain the whole premium, as a large portion of the premium documented in this paper originate in low-sales-season firms. We also find no evidence for differences in analyst s forecast dispersion driving the salesseasonality premium. Low- and high-season firms display the same level of analysts forecast dispersion. Additionally, we examine whether differences in market surprises around earning announcements can explain our results. We measure the market surprise as the market adjusted cumulative abnormal return for days 1 to 1 of the earnings announced during the low- and highsales-season portfolio assignment. We find that the low-sales-season portfolio generates no surprises, while the high-sales-season portfolio generates negative surprises. Although this finding can potentially explain the high-sales-season portfolio returns, it cannot explain the low-salesseason portfolio returns. A further possible explanation we consider is idiosyncratic volatility. If firms in high-sales season experience higher idiosyncratic volatility, this phenomenon can potentially explain their low returns (Ang, Hodrick, Xing, and Zhang 2006). However, the low-sales season has the highest idiosyncratic volatility, which is inconsistent with this explanation. Additionally, we examine the differences in price pressure measures across the portfolios, but observe no patterns (Chordia and Subrahmanyam, 2004). Moreover, we find no evidence that institutional investors engage in seasonal window-dressing, which could affect stock returns. 14 It is important to highlight that many of the variables examined in this sub-section follow a U-shape or reverse U-pattern across 14 Untabulated results. 23

25 the sales-seasonality portfolios, and the low- and high-sales-season portfolios tend to be on the extremes. As mentioned earlier, this is not surprising because the firms in the extreme portfolios of the sales-seasonality strategy are likely to be the same firms at different points in time. Finally, we examine whether differences in systematic risk across seasons can explain our results. We run monthly regressions of excess returns as the dependent variable on the market return, high- and low-season dummies, and interaction terms between the dummies and the market return. We then run time-series regressions for each firm, and average the coefficients across all firms in a seasonal portfolio. If a systematic risk premium for low-season firms does exist, we should observe a positive and significant coefficient on the interaction term between the lowseason dummy and the market excess return. On the other hand, if a systematic risk discount exists for high-season firms, we should observe a negative and significant coefficient on the interaction term between the high season dummy and the market excess return. Table 13 reports the results from this analysis. Contrary to the predictions of the time-varying beta hypothesis, we find that both interaction terms are statistically insignificant and close to zero. INSERT TABLE 13 HERE V.C Earnings announcements and investor inattention If learning about the information regarding fundamentals of a firm is more valuable during high seasons than low seasons, then investors may shun low-season firms, since the marginal cost of acquiring information during these periods exceeds its marginal benefit. Markets may therefore become less efficient during low seasons. To test this hypothesis, we construct proxies for inattention and examine whether the level of inattention is related to sales seasonality. Specifically, using the post-earnings drift as a proxy for investor inattention (Hirshleifer, Lim, and Teoh 2009), we examine whether this drift is higher during low seasons. 24

26 We begin by examining the post-earnings announcement market-adjusted average cumulative abnormal returns (CAR) across all sales-seasonality deciles. Figure 2 shows the performance from day t+2 until day t+45, where t is the day of the earnings announcement. Following the earnings announcements, the high-season decile generates the lowest CARs, while the low-season decile generates the highest CARs. This finding is consistent with the argument that attention around the earnings announcement will be higher during the high season and lower during the low season. On average, low-season firms have a steeper rise in their CARs, reaching almost 1.66% on day 45. In contrast, high-season firms generate CARs of 0.6% over the same period. To examine whether any differences occur between positive and negative surprises postannouncement drifts, we split the data in two groups: one, firms with positive market reactions around the day of the announcement, and two, firms with negative market reactions. In order to do this, we compute the cumulative abnormal return between t 1 and t+1, and then recompute the performance from day t+2 until day t+45. Figures 3 and 4 present the results from this analysis. When the earnings announcement generates a positive market reaction, the difference between low- and high-sales-season portfolios is greater than the difference occurring when the announcement generates a negative market reaction. The low-season portfolio has a CAR of 1.65%, while the high-season portfolio has a CAR of 0.46%. 15 Moreover, when the earnings announcement generates a negative market reaction, the low-season portfolio has a CAR of 1.22%, while the high-season portfolio has a CAR of 0.34%. 15 Although previous studies find that negative market surprises are associated with negative post-earnings announcement drift, we find that firms 4 years or older do not exhibit this pattern. Since our measure of sales seasonality requires at least 4 years of data to be implemented, the negative market surprise post-earnings announcement drift is not negative. 25

27 To further investigate the predictions of the investor-inattention hypothesis (Hirshleifer, Lim, and Teoh 2009), we formally examine the relation between the level of investor attention during earnings announcements and the subsequent drift depicted in Figure 2. Specifically, we use two-stage least square regression analysis to estimate the effect of sales seasonality on the level of attention investors give to firms before the earnings announcement, and then use the instrumented attention measure to examine its effect on the post-earnings drift. We use the number of file downloads from the EDGAR Server Log data posted during the days before the earnings announcement as a proxy for the level of investor attention. 16 This dataset, which contains information on internet search traffic for EDGAR filings through SEC.gov, is produced by the Securities and Exchange Commission s (SEC) Division of Economic and Risk Analysis; the dataset contains thousands of files with more than four billion documented server requests for the period For each request for a specific firm, a typical server will log the client s IP address, timestamp of the request, and page requested. The Central Index Key (CIK) is the identification number used to identify corporations. The raw Server Log data from the SEC includes observations that are not useful for our analysis (e.g., robot downloads, and missing or incorrect variables). Loughran and McDonald (2017) condensed the data after identifying the non-robot total count of forms downloaded and filtering the data for missing or incorrect variables. 17 We use their condensed version of the dataset that contains the date of the entries, CIK of the firm, non-robot total counts, and the form and filing date corresponding for the form downloaded. We then merge the dataset via the CIK, and for each earnings announcement event, we calculate two attention proxies with different horizons: ATTE1 is the average number of non-robot views of any document between the beginning of the quarter 16 For more information, visit 17 For more information on data filtering, visit 26