Modeling Analysts’ Recommendations via Bayesian Machine Learning

Abstract

We apply state-of-the-art Bayesian machine learning to test whether we can extract valuable information from analysts’ recommendations of stock performance. We use a probabilistic model for independent Bayesian classifier combination that has been successfully applied in both the physical and biological sciences. The technique is ideally suited for the particular problem where any individual analyst only focuses on a handful of the thousands of companies and it allows for dynamic inference as we track the performance of the analysts through time. The results suggest this technique holds promise in extracting information that can be deployed in active investment management.

Keywords: Variational Bayes, VB, IBCC, Machine Learning, Data Science, Analysts’ Forecasts, IBES, Analysts’ Recommendations, Forecast Fatigue, Forecast Combination, Dirichlet Distributed Prior, Galaxy Zoo

JEL Classification: G11, G14, G17, C11, C58, M41