Machine Learning from the Best: Predicting the Holdings of Top Mutual Funds

Abstract

I show that machine learning models, by exploiting the nonlinearities and interactions in stock characteristics, can better predict the stocks owned by top-performing mutual fund managers than suggested by their most recent holdings or a linear model. Previous ownership by mutual funds and the market cap and volume of the stock are identified as the most important predictors. The predictions also prove useful in separating stocks based on their future return potential. Shorting stocks predicted to be disliked by top managers provides higher returns than can be explained by common factors, either systematically or over time.

Keywords: Machine learning, mutual funds, stock-picking, classification, active management

JEL Classification: C38, C52, G10, G11, G12, G17, G23