Selecting Mutual Funds from the Stocks They Hold: A Machine Learning Approach

Abstract

We select mutual funds in real time by combining individual fund holdings and a large number (94) of stock characteristics to compute fund-level characteristics on the basis of the stocks they hold. We show that, first, the majority of funds are largely exposed---both positively and negatively---to approximately 40-50 characteristics. Second, fund performance is non-linearly related to fund characteristics and there are significant degrees of interaction between different fund characteristics and fund performance. Third, when we predict fund performance, these non-linearities and interactions prove important as machine learning methods such as Boosted Regression Trees (BRT) outperform significantly standard linear frameworks and the BRT-generated forecasts encompass the ones generated by the predictors of mutual fund performance that have been proposed in the literature so far. Fourth, while in our setting BRT outperform the LASSO, elastic nets, random forests, and neural networks with 1 through 5 hidden layers, these other machine learning methods deliver good performance and they all outperform ordinary least squares models. Finally, while we detect signi cant predictability using machine learning methods, the fund characteristics that matter the most in predicting fund returns and the functional relation between fund characteristics and fund performance are time-varying.