Predicting Stock Returns Using Natural Language Processing of Earnings Calls

Abstract

It has been well established in the current literature that stock price movements surrounding earnings announcement events exhibit predictable characteristics. These events involve the release of both quantitative and qualitative information regarding the underlying companies’ businesses, and a wide body of work has shown that such signals can be exploited to anticipate the direction of future equity returns. However, the existing literature does not extensively utilize a crucial dataset: the transcripts of the conference calls accompanying earnings announcements. Using natural language processing techniques on these transcripts, this thesis generates signals conveying information released during earnings announcements (above and beyond information gleaned from direct reading of the transcripts) and outlines an algorithm that uses these signals for return prediction and portfolio selection. Fama-French factor analyses of the resulting portfolios suggest that the transcript signals outperform human interpretation of earnings signals, as well as signals identified in the existing literature, and that this out-performance stems from incrementally greater levels of qualitative information captured by these new indicators.