Machine Learning and Fund Characteristics Help to Select Mutual Funds with Positive Alpha

Abstract

Nonlinear machine-learning methods select tradable long-only portfolios of mutual funds that earn significant out-of-sample alphas of 2.3% per year net of all costs. In contrast, linear methods deliver insignificant alphas. Machine learning unveils important interactions between fund activeness and past performance--to earn positive alpha, investors should choose more active funds conditional on their having good past performance, but less active funds conditional on poor past performance. Our findings demonstrate that investors can benefit from active management, but only if they have access to the predictions of sophisticated methods that capture complexity in the relation between fund characteristics and performance.

Keywords: Mutual-fund performance; performance predictability; active management; elastic net; random forests; gradient boosting.

JEL Classification: G23, G11, G17