Abstract
We study the predictability of risk-adjusted mutual fund performance using machine learning models on a global sample of actively managed equity funds. Predictability is highly heterogeneous: concentrated in the U.S. and Europe, and largely absent elsewhere. Globally trained models outperform locally trained ones, highlighting the value of international data. A value-weighted long-only strategy delivers an annualized out-of-sample FFC6 alpha of 2.6% after costs. We link this heterogeneity to two channels: First, larger industries with high passive market share sharpen the distinction between skilled and unskilled managers. Second, well-calibrated factor models provide higher information quality by producing more reliable skill signals.

This publication has 48 references indexed in Scilit: