Earnings announcement return extrapolation

Beliefs and preferences are the two important ingredients for decision-making (Barberis and Thaler 2003).

A related finding is documented by Milian (2015), who examines a sample of firms with active exchange-traded options, and argues that these easy-to-arbitrage firms face a pre-earnings rise and post-earnings reversal based on excessive trading by arbitrageurs on the post-earnings-announcement drift. We offer distinct evidence of investors’ extrapolative beliefs based on individual investor trading and the cross-section of returns around earnings announcements. Our findings also highlight the importance of recency in the autocorrelation of earnings announcement returns. The stability of our findings throughout our sample period indicates that a behavioral bias contributes to the pattern above-and-beyond the arbitrage activity documented by Milian (2015). In particular, we document extrapolative trading in the early 1990s among individual investors who are unlikely to be aware of stock market anomalies like post-earnings-announcement drift.

There is significant evidence that investors avoid short selling: for example, Almazan et al. (2004) find that most mutual funds are restricted from short-selling by charter and only 2% actually do sell short.

Related work that finds similar evidence is by Berkman et al. (2012) and Aboody et al. (2018).

Similar to DellaVigna and Pollet (2009), we identify the earnings announcement date as the earlier of the IBES earnings announcement date (anndats_act) and Compustat earnings announcement date (rdq).

We impose the minimum share-price restriction at the firm-quarter level, rather than at the firm-day level. In doing so, we keep the observation count much more stable for each firm-quarter. By contrast, performing the data exclusion at firm-day level would have included some observations and excluded some observations for firms whose share price hovers around the threshold (five dollars), which could bias our estimates.

First, we define a predicted EA date based on the firm’s year-ago same-quarter EA dates (consistent with the EA date prediction of Cohen et al. 2007 and Aboody et al. 2010 and with EA month prediction of Barber et al. 2013). This predicted date is the same calendar day of the year-ago same quarter after adjusting for nontrading days. We then create an indicator (Predictable EA) recording whether the predicted EA is within 2 days of the actual EA in both the current quarter and the previous quarter.

A potential concern is that managers may be concerned that a price rise and reversal around earnings announcements could increase litigation risk. To the extent that investors could form a class and demonstrate loss causation due to (non-)disclosure by the managers, we might expect managers to try to deter EA return extrapolation. To investigate managers’ incentives, we explore the relationship between securities class action lawsuits and our measure of extrapolated returns using data collected from the Stanford Securities Class Action Clearinghouse. In untabulated findings that are available upon request, we find that litigation of this type is rare: 0.61% of firm-quarter observations in our sample are the first of a class period of any type of lawsuit, and only 0.18% are the first of a class period for lawsuits related to EAs. We also find that there is no systematic difference between firms in the top decile of our extrapolated return measure and others in the propensity to be subject to litigation. These results suggest that litigation risk does not motivate managers to mitigate EA return extrapolation. Nevertheless, investigating the potential disciplining role of managerial incentives for investors’ behavioral biases is an interesting area for future research.

Investors who hold throughout the entire 11-day window would earn positive returns. However, their returns would be even higher if they instead sold their holdings before the reversal period.

Shanthikumar (2012) finds an association between strings in earnings growth and trade imbalances surrounding earnings announcements, whereas we dissect the earnings announcement window and only find returns evidence that is consistent with extrapolative trading during the pre-announcement period. Frieder (2008) examines post-earnings-announcement trade imbalances and attributes them to overextrapolation. However, she does not show that the trade imbalances caused by consecutive positive earnings surprises are associated with negative returns. Instead, she shows that the trade imbalances are associated with negative returns after conditioning on the number of consecutive positive earnings surprises.

We opt for Wall Street Horizons following previously raised concerns about the accuracy of IBES time stamps (e.g., Bradley et al. 2014; Michaely et al. 2014, 2016; Lyle et al. 2018).

Wall Street Horizons timestamp data is available from 2006. For 2000–2005, which precedes Wall Street Horizons’s coverage, we obtain the time stamp data from IBES, although this procedure does not materially affect our conclusions.

In the internet appendix, we present estimates of portfolio returns across subperiods of our sample, consistent with the daily return tests presented in Section 3.1. In IA Table 3.1, we show that the portfolio returns are consistently positive throughout our sample period.

Data for this setting is described in detail in Appendix 2.

The regressions in this table account for market returns on a country-by-country basis by controlling for the daily value-weighted market return for the firm’s country.

We subtract the S&P return to calculate the excess return.

For tests of differences within volatility groups, we use a seemingly unrelated regression estimates with calendar-quarter clusters. The significance of the findings is similar if we instead cluster by firm.

Barber et al. (2013) provide evidence against this by noting that the pre-earnings-announcement rise in the international sample is higher for firms with lower volume.

Lee et al. (1993), Krinsky and Lee (1996), and So and Wang (2014) provide evidence that transactions costs increase before earnings announcements due to adverse selection—informed traders will want to bet on earnings announcements—and inventory risks.

Consider a simple example. Suppose that the median investor thinks the stock is worth $6. The liquidity provider may set the ask price (she will purchase) at $5.50 and the bid (she will sell) at $6.20. Assume the liquidity provider could predict demand and knows that the demand for the stock at $6.20 will be far greater than the demand to sell the stock at $5.50. This will lead to an average price above $6 and a decrease in inventory for the liquidity provider.

In untabulated results, we examine the behavior of other market participants—the media, company insiders, and sell-side analysts. We find evidence of extrapolative beliefs among company insiders and do not find strong evidence of extrapolative beliefs in news articles or sell-side analyst forecasts. Results are available upon request.

For the plots in Fig. 2, we restrict the sample to firm-quarters with predictable earnings announcements. We make this sample restriction for the corresponding regression tests as well. For the plots, we further restrict the sample to firm-quarters that have observations for the entire window shown in the plot to ensure that changes in the plot are not driven by changes in sample composition. In Table 4.1 in the internet appendix, we describe investor purchasing behavior in firm-quarter observations with predictable EAs.

We exclude the day of the EA from the sample because we do not know the time of day when the announcement occurred for the EAs in this sample period. This is because our sample period for these data runs from 1991 to 1996, which does not overlap with our data on precise EA timing. If the announcement occurred after the close of the market, then the EA day should be included in the 5 days before the EA. However, if the announcement occurred earlier in the day, then it should not be included. Without knowing the timing of the announcement, we do not know how to treat each EA day, so we drop these days from the sample.

Due to the inclusion of firm-quarter fixed-effects, we cannot simultaneously identify the coefficient on the top extrapolated return decile dummy.

Similar determinations of the predicted earnings announcement date have been used in the academic accounting literature (e.g., Cohen et al. 2007; Begley and Fischer 1998). In Table 4.2 of the internet appendix, in a quasi-placebo test, we show that there is little evidence of extrapolative purchases when the earnings announcement date is not predictable.

Importantly, other investors who are not retail investors might be extrapolating too. We have no indication that these individual trades are the sole drivers of the return patterns we observe. Indeed, in an untabulated analysis, we find that total market trading volume (from CRSP) follows a pattern very similar to the individual purchases: rising for top-decile firms right before the EA. This rise in trading volume is of a similar proportion to the rise in individual purchases, indicating that other market participants might extrapolate just as much as the individuals in our sample. We corroborate this inference later in the paper with evidence that institutional investors are extrapolating EA returns.

For robustness, we perform the regressions in Table 7 with alternative measures of extrapolated returns, including one that takes the simple average of earnings announcement returns over the past eight quarters. We present our findings in the internet appendix (Table 3.1).

These results differ from those of Hirshleifer et al. (2008), who use the same dataset to examine individual trades. Hirshleifer et al. (2008) check whether the seasonal difference in standardized unexpected earnings (SUE) predicts individual investor trades right before the next earnings announcement, with the goal of determining whether individuals drive post-earnings-announcement drift. While our tests do not find a relationship between earnings growth and trades right before the next announcement, Hirshleifer et al. (2008) find that firms in both the top and bottom SUE deciles have significantly greater purchases and sales at that time. They also find marginally significant evidence (p value = 0.09) that net purchases right before the next earnings announcement are higher for firms in the top SUE decile. Our results differ from theirs because our firm-quarter fixed effects control for the baseline level of trades during the firm-quarter. In other words, our test removes the trades that would have occurred even if the earnings announcement was more than a week away. This is necessary because investors likely pay more attention to firms with extreme performance. Hirshleifer et al. (2008) do not control for the level of trades throughout the quarter, meaning their estimates also include trades that do not relate to the earnings announcement. In our tests, controlling for the baseline level of trades reveals that investors do not change their pre-announcement trading based on recent earnings growth. They only do so based on the earnings announcement return.

Since we do not know when earnings is announced on the day of the EA, some of these investors will be pre-EA purchasers.

As with the other tests of individual trades in Section 4.1, we make the following sample restrictions. (1) We restrict the sample to predictable earnings announcements, and (2) we exclude the earnings announcement itself from the sample.

We assign individual-firm combinations to all days from the date of one’s first reported holding to the date of his or her last reported holding. Individuals are assigned to any firm in which they ever hold a position throughout the sample. Individuals are defined based on their account numbers. Individual-firm-quarter fixed effects are dummies based on account numbers, security numbers, and earnings announcement dates.

We set Watcher to missing, and thus drop the observation from the sample, if data is missing for any of the days in the window in which Watcher is measured. This is especially important in column (3), when the window is 63 days long. Without this restriction, we would measure Watcher with 63 days for some firms and with as few as one trading day for others.

Regrettably, Abel Noser Solutions, LLC, discontinued their program supporting the use of their data on institutional trades by academics.

Related work that finds similar/related evidence is by Berkman et al. (2012) and Aboody et al. (2018).

We exclude the day of the earnings announcement and the day after the earnings announcement for mechanical reasons. That is, the return on those 2 days will be part of our extrapolated return measure.

In an additional analysis, we run the returns and individual trades tests for a third alternative, trailing returns excluding past EA returns. Our results continue to hold (untabulated).

We exclude nontrading days. We exclude a firm’s observations for the month if the firm has zero-return days for more than 80% of the days that month. We set returns to missing if Datastream’s return index variable is below 0.01 on the current day or the previous day (Ince and Porter 2006). We set returns to missing if they are in the top or bottom 0.1% of the cross-sectional distribution within a country over time. We set the daily returns to missing if the daily return is greater than 100% on a given day or on the previous day but the compound return over the 2 days is less than 20%. We set daily returns to missing if they exceed 200%, and we exclude depositary receipts, real estate investment trusts, preferred stocks, investment funds, warrants, debt, unit trusts, and other stocks with special features.

Sometimes, Worldscope records the earnings report date in other variables: “wc05901,” “wc05902,” “wc05903,” “wc05904,” and “wc05905.”