Fundamental Analysis Via Machine Learning

Abstract

We examine the efficacy of machine learning in one of the most important tasks in fundamental analysis, forecasting corporate earnings. We find that machine learning models, especially those accommodating nonlinearities, generate significantly more accurate and informative forecasts than a host of state-of-the-art earnings prediction models in the extant literature. Further analysis suggests that machine learning models uncover economically sensible relationships between historical financial information and future earnings, and the new information uncovered by machine learning models is of considerable economic significance. The new information component is significantly associated with both future stock returns and analyst forecast errors, with stocks in the quintiles with the most favorable new information outperforming those in the least favorable quintiles by approximately 70 bps per month. The overall results suggest that limiting to linear relationships and aggregated accounting numbers substantially understates the decision usefulness of financial statement information to investors.

Keywords: machine learning, earnings forecasts, fundamental analysis, equity valuation, market efficiency

JEL Classification: G10, G11, G14, G17, M40, M41