Can Machines Understand Human Decisions? Dissecting Stock Forecasting Skill

Abstract

Human decisions are important but difficult to understand or predict. This paper uses machine learning models, which are adept at capturing nonlinear and complex relations, to analyze analysts’ forecasts and determine their skill. Machine-identified skilled analysts persistently outperform expert-picked star analysts. Machines rely on nonlinear interactions of analyst characteristics, such as past skill and efforts, to make predictions, in contrast with human experts, who lean more on relation-based information such as brokerage size. The puzzle of post analyst-revision drifts can be explained by our model in that such drifts are concentrated in machine-picked skilled analysts. Our approach also allows the formation of a “smart” analyst consensus that aggregates the forecasts of machine-picked skilled analysts. Investment strategies based on revisions of machine-identified skilled analysts and the smart analyst consensus both generate significant abnormal returns. Overall, we propose an interpretable machine learning framework that can be used to analyze and predict human decisions. We also provide a new, improved way to obtain the wisdom of the crowd applicable to other settings such as online forums, political opinions, and macroeconomic outlooks.