Skip to Content

Darla Moore School of Business


Moore School accounting professor finds underlying language cues in analysts’ reports that might signal buy quality

May 26, 2017

Call it reading between the lines. Recent research by Darla Moore School of Business professor Mark Cecchini— and coauthors Marcus Caylor of Georgia State University and Jennifer Winchel of University of Virginia — shows that certain words in a financial analyst report could imply that a company will do better in the market than one lacking those words. There isn’t some magical combination of words that mark an incredible buy, but using natural language processing tools, he found that better stock picks could potentially be made by basing buying on certain combinations of words.

To support his hypothesis that there is subliminal information in what analyst reports say, Cecchini took buy-rated analyst reports and examined the buys a year after they were rated to see if they were actually good buys. Then he split the good from the bad to search for notable differences between what was written about each.

The natural processing technique takes a portion of the good ones and a portion of the bad ones and generates a “dictionary” of differences in language. Cecchini then took that dictionary and applied it to a whole new sample to see if it provided any insights. By taking that sample portfolio and “trading” on it, he determined that money could be made using this method based on the positive returns that resulted.

Although natural language processing and other statistical tools he used are not an exact science, these results do make a valid argument for someone to read the analysts’ reports in addition to looking at the hard numbers to make the best buy possible. Considering the theory that analyst report ratings are optimistically biased, Cecchini’s findings suggest that reading the summaries themselves could give someone a better idea of what is actually going on within a company.

Cecchini began doing this type of research with his dissertation on machine learning methods for fraud and bankruptcy detection. In that research, he used natural language processing to see if he could find clues to a company going bankrupt or committing fraud.

In the research he is doing now, Cecchini is continuing his study of language structure in an effort to understand how things get written. In an experiment he and his coauthors Darin Freeburg and Joel Owens have created, participants write a corporate disclosure based on information given to them about recent results. Participants are put into various economic conditions, some more stressful than others. The researchers have found that when under more stress, the writers work harder to find appropriate words and thus write more differently than the ones in low stress conditions, who tend to write more similarly.

By Madeleine Vath