Information in the Tails of the Distribution of Analysts’ Quarterly Earnings Forecasts

Abstract

Investors generally measure earnings announcement news on the basis of the difference between actual earnings and two salient benchmarks: earnings in the same quarter the previous year and a consensus drawn from a distribution of forecasts by financial analysts. We evaluate the implications of a third salient benchmark: the most optimistic forecast when actual earnings exceed the consensus and the most pessimistic forecast when the consensus exceeds actual earnings. We find that considering the information in these tails of the distribution of analysts’ earnings forecasts enhances the profitability of post–earnings announcement drift strategies.

Editor’s note: This article was reviewed and accepted by Robert Litterman, executive editor at the time the article was submitted.

We appreciate helpful comments from Henk Berkman, Charles Corrado, David Emanuel, Nam Tran, and seminar participants at the University of Auckland, Deakin University, Louisiana State University, Monash University, and the 2009 Annual Meetings and 2010 Western Region Meetings of the American Accounting Association. Philip Shane gratefully acknowledges financial support from the Frank Wood Accounting Faculty Research Fund at the Raymond A. Mason School of Business, College of William & Mary.

Notes

¹ Media reports of companies’ quarterly earnings generally include GAAP and non-GAAP measures of actual earnings for both the current quarter and the same quarter of the previous year, along with a consensus of analyst forecasts taken from a data aggregator, such as Thomson Reuters. Data aggregators maintain or contribute to freely and easily accessible websites with the high and low estimates in the distribution of forecasts from which the consensus emerges (e.g., Reuters, Yahoo! Finance). To the best of our knowledge, our study is the first to rigorously analyze the information content of quarterly earnings forecasts in the tails of the distribution.

Note that we use the terms earnings news, earnings surprise, unexpected earnings, and forecast error interchangeably in this article.

² In our study, we identified an approach that improves on the consensus forecast in judging the information content of earnings announcements. This approach is consistent with Beckers, Steliaros, and Thomson (2004), who found that active portfolio managers and especially buy-side analysts seek to improve on consensus forecasts in identifying investment opportunities.

³ Consistent with prior literature, we defined earnings surprise (based on a rolling seasonal random walk) as the difference between actual earnings for the current fiscal quarter and actual earnings for the same fiscal quarter of the previous fiscal year.

⁴ Several studies have documented that transaction costs are the main impediment to arbitraging PEAD (see, e.g., Ng, Rusticus, and Verdi 2008; Chordia, Goyal, Sadka, Sadka, and Shivakumar 2009) and that information risk mutes the stock price response to earnings information (Yan and Zhao 2011), leading to a more pronounced PEAD. In our study, using the best estimate of transaction costs (Chung and Zhang 2014) and usual proxies for information risk, such as analyst forecast dispersion and stock return volatility (Zhang 2006), we documented that PEAD, especially PEAD based on multiple unexpected earnings proxies, is not fully explained by transaction costs or information risk. These findings are consistent with Battalio and Mendenhall (2011) and Yan and Zhao (2011), who documented that PEAD-based trading strategies are highly profitable after accounting for transaction costs and information risk.

⁵ Researchers analyzing minimum and maximum values have found many effective applications. In a capital market context, George and Hwang (2004) demonstrated the profitability of momentum trading strategies timed to buy (sell) stocks when they reach the new 52-week high (low) point. Although investors had already been using that information to inform their trading decisions, George and Hwang provided the first large-sample scientific evidence of the importance of that widely available statistic. In a similar vein, our study provides the first large-sample scientific evidence of the importance of earnings news based on broadly available high- and low-earnings forecasts.

⁶ Although in our examples we ignored the scalar for brevity, in our empirical tests, we scaled all variables as described in Equation 1$$\begin{array}{l}{\text{SUETAIL}}_{it}=\left(E_{it}-{\text{MAX}}_{it}\right)/P_{it}\times \left(\text{POST}\right)\\ +\left(E_{it}-{\text{MIN}}_{it}\right)/P_{it}\times \left(1–\text{POST}\right),\end{array}$$ (1) .

⁷ Zhou and Shon (2013) documented that more than 40% of earnings surprises are met by opposite-direction stock price reactions, suggesting that consensus forecasts are not necessarily good proxies for the “true” market expectations. Our example illustrates, at least partially, why the market may react in the opposite direction of a positive earnings surprise based on the consensus forecast, emphasizing the need to go beyond the usual assumption of consensus forecasts as the best proxies for market expectations.

⁸ Our results are not unduly influenced by any particular industry. The sample includes 16 SIC code–defined industries, with the smallest (largest) representation from agriculture (financial services) at 0.17% (17.78%). More importantly, no single industry is unevenly split between the top and bottom deciles of SUECF or SUETAIL. For example, 17.46% (18.11%) of the observations in the top (bottom) decile of the SUETAIL distribution are from the financial services industry.

⁹ For our assessment of the robustness of our main results to the inclusion of this and other control variables, see Appendix A.

¹⁰ Brandt et al. (2008) derived EAR in the same way that we derived CAR (–1,1).

¹¹ Another way to look at how the tail forecast informs a momentum trader is to consider what happens if the investor/analyst ignores the information in SUETAIL and takes long (short) positions in stocks with large (small) SUECF and small (large) SUETAIL. In that case, the hedge portfolio return is a statistically insignificant 0.83% – 0.53% = 0.3%. Thus, large positive (or negative) consensus forecast errors do not justify a trading strategy that ignores the information in the tail of the distribution from which the consensus emerges.

¹² Holding SUETAIL constant and consistent with Brandt et al. (2008), we found significant returns to a trading strategy that takes long (short) positions in stocks whose SUECF and EAR are both in the top (bottom) quintile of their respective distribution. These quarterly returns are 1.7%, 3.1%, and 3.4%, respectively, when SUETAIL is in the bottom, middle, and top quintiles of its distribution—or 2.7%, on average. The hedge return jumps to 5.4% with a trading strategy that takes long (short) positions in stocks whose three earnings surprise measures are in their respective top (bottom) quintile.

¹³ In an alternative specification, we added control variables used in Mendenhall’s (2004) PEAD determinant model. Specifically, our results are robust to controls for the percentage of institutional ownership and the number of analyst forecasts.