Investor Attention, Divergence of Opinions, and Stock Returns

Abstract

Using a direct measure of investor attention generated from the Securities and Exchange Commission’s EDGAR (Electronic Data Gathering, Analysis, and Retrieval) log files, the authors revisit the stock return predictability of the divergence of opinions in the presence of varying degree of investor attention and information acquisition. They document a positive relationship between the divergence of opinions and future stock returns, consistent with the risk hypothesis, as opposed to the overvaluation hypothesis. More importantly, the authors find that the predictive power of divergence of opinions is more pronounced in stocks with lower investor attention. They further document the construction and profitability of divergence of opinions portfolios augmented with investor attention. A portfolio that goes long on stocks with low investor attention and the highest divergence of opinions and short on stocks with low attention and the lowest divergence of opinions generates a Fama-French 5-factor monthly alpha of 1.14%.

Keywords:

• Divergence of opinions
• Stock returns
• Investor attention
• SEC EDGAR

JEL Classification:

• G10
• G11
• G14
• G40

Notes

1 In an earlier study, Barron et al. (1998) argued that dispersion in analysts’ forecasts is a poor proxy for divergence of opinion since it is affected by uncertainty in analysts’ earnings forecasts. Doukas, Kim, and Pantzalis (2006) showed that the results in Diether, Malloy, and Scherbina (2002) are reversed when uncertainty in analyst’s earnings forecasts is controlled for.

2 A detailed description of the SEC EDGAR access log dataset is available at: https://www.sec.gov/dera/data/edgar-log-file-data-set.html.

3 Michaely et al. (2016) argued that the result is driven by selection bias. Firms that make announcements on Fridays experience reduced market response on any weekday and have common unobserved characteristics across announcement types. After correcting for selection bias, there is no evidence that investors pay less attention to announcements made on Fridays.

4 For an incomplete list of studies using the SEC’s EDGAR access log files, see Lee et al. (2015), Drake, Roulstone, and Thornock (2015), Loughran and McDonald (2017), Ryans (2018), Gao, Wang, and Yan (2019), and Li et al. (2019), among others.

5 Loughran and McDonald (2017) pointed out that the SEC EDAGR server log files had corrupted files during September 2005 to May 2006. We purge this period out of our sample when constructing the investor attention measure.

6 The quintile sorting results are available upon request.

7 We conduct quintile sorting for robustness checks. The results are very similar and slightly weaker for quintile sorting. For instance, the Fama-French 5-factor alpha for quintile sorting is 79 basis points per month with a t-statistic of 4.45. These results are available upon request.

8 We choose to sort sample stocks into 2 groups based on investor attention because this would produce more stocks in each portfolio than if we sort them into 10 deciles. The results are stronger if we sort them into 10 deciles based on investor attention. The results are qualitatively the same if we sort by using AttDRT or AttnR.

9 Detailed results using quintile sorting are available upon request.

10 We thank an anonymous referee for bringing up this thought-provoking point.