Do certain types of news about specific stocks reliably predict risk-adjusted returns of those stocks? In their March 2017 paper entitled “Using Natural Language Processing Techniques for Stock Return Predictions”, Ming Li Chew, Sahil Puri, Arsh Sood and Adam Wearne investigate relationships between financial news headlines and stock returns stripped of non-news risks. They use natural language processing to classify corporate events by firm, illustrating via five types: dividend declaration; oversold conditions; receipt of approval; signing an agreement; and, hiring an advisor. They isolate each type by segmenting headlines into 10, 20, 50 or 100 clusters of similar headlines. They then form portfolios for the most relevant clusters that are long (short) stocks for which events have occurred (same-industry stocks for which there are no events), with positions weighted to eliminate exposures to market, size and value factors. Outputs include factor-adjusted cumulative and daily average returns. They focus on stocks in the S&P 500 as it evolves and divide the sample into 2006-2014 to identify event clusters in-sample and 2015-2016 to test cluster portfolio performance out-of-sample. Using 60,949 active voice financial news headlines that relate to specific S&P 500 firms and associated daily/quarterly stock price and firm characteristics data during 2006 through early 2017, they find that:
- S&P 500 headlines concentrate in recent years, with a quarterly intra-year pattern associated with earnings announcements. The 10 stocks with the most headlines comprise 27% of the sample. However, after removing the top three (AAPL, MSFT and FB), there is no relationship between logarithm of market capitalization and headline count.
- In general, as reasonably expected, using more headline clusters to isolate events strengthens/clarifies effects.
- Four of five illustrative event types exhibit predictive power for pure future stock returns out-of-sample:
- Oversold conditions predict positive future cluster portfolio returns, but positive performance concentrates in the first year of the two-year out-of-sample test period.
- Receipt of some kind of approval predicts positive cluster portfolio future returns, but positive performance concentrates strongly in the last six months of the two-year out-of-sample test period.
- Signing a deal or agreement predicts negative cluster portfolio future returns, but performance is volatile.
- Hiring an advisor predicts negative cluster portfolio future returns, with the decline fairly steady over the two-year out-of-sample test.
- There are a few “events” that exhibit some predictive power for pure future stock returns out-of-sample, but cluster keywords have no obvious financial interpretation.
In summary, evidence suggest that machine processing of news headlines about specific stocks may be useful in identifying the types of firm events that move stock price.
Cautions regarding findings include:
- Cluster portfolio performance results are gross, not net. Accounting for trading frictions and, when applicable, shorting costs would reduce reported returns.
- The out-of-sample test period is short in terms of variety of market conditions. Investors may respond differently to specific types of news during different market conditions. Inconsistent performance during the out-of-sample test period for several event types, as noted above, amplifies this concern.
- As illustrated in the paper, some events are so infrequent that the cluster portfolio is sometimes in cash.
- Parsing into more and more clusters reduces the number of headlines in the clusters, increasing the role of luck in (and the volatility of) cluster portfolio returns.
- The process outlined is beyond the capabilities of most investors, who would bear fees for delegating to experts/fund managers.
