Can investors apply deep learning software to expose obscure but useful management sentiment in firm SEC Form 10-K filings? In their July 2019 paper entitled “Is Positive Sentiment in Corporate Annual Reports Informative? Evidence from Deep Learning”, Mehran Azimi and Anup Agrawal apply deep learning to detect positive and negative sentiments at the sentence level in 10-Ks. They train their model using 8,000 manually evaluated sentences randomly selected from 10-Ks. They then use the trained model to assign sentiments to all sentences in each 10-K. Their overall measure of negative (positive) sentiment is number of negative (positive) sentences divided by the total number of sentences in the 10-K. They assess impact of 10-K sentiment on stock performance based on 4-factor (market, size, book-to-market, momentum) alpha during short intervals after 10-K filing. Using 10-K filings for non-utility and non-financial U.S. public firms with at least 200 words, associated daily stock prices/trading volumes and daily 4-factor alphas during January 1994 through December 2017, they find that:
- During trading days 0 through +3 relative to 10-K filing date, after controlling for quantitative information in the filing and other relevant variables, a one standard deviation increase in negative (positive) overall sentiment predicts 4-factor alpha -0.13% (0.07%). Findings are robust to controlling also for quarter-of-year and for exclusion of reports associated with an earnings announcement within two days prior to filing date.
- Both positive and negative overall sentiment indicate higher abnormal during intervals up to one month after trading day +3, suggesting that the market underreacts to positive sentiment and overreacts to negative sentiment.
- A trading strategy for firms with fiscal years ending December that is each year at the end of March long (short) stocks with either highest (lowest) fifth of positive sentiments, or lowest (highest) fifth of negative sentiments, generates insignificant 3-factor (market, size, book-to-market) alpha.
- In general, extreme sentiment stimulates abnormal trading volumes, more for negative than positive extremes.
- Positive (negative) sentiment predicts improving (degrading) fundamentals such as return on assets and operating cash flow.
In summary, evidence offers little support for belief that investors can exploit deep learning-assisted 10-K textual analysis to beat the stock market.
Cautions regarding findings include:
- Results are gross, not net. Costs of deep learning software, expert use of the software and trading/shorting to exploit findings would reduce already economically slim abnormal returns.
- The annual trading strategy tested is incommensurate with the short-term tests of abnormal returns.
- Methods are elaborate and beyond the reach of most investors, who would bear fees for delegating to a fund manager.
See results of this search for other perspectives on using textual analysis for trading signals.
