Predicting Mutual Fund Performance using CAPM and Machine Learning Algorithms

Abstract

This paper examines whether combining the traditional Capital Asset Pricing Model (CAPM) framework with modern machine learning (ML) algorithms improves the prediction of mutual fund performance and the identification of funds likely to generate positive risk-adjusted returns. We compare CAPM-based ex-post performance measures (e.g., Jensen’s alpha) with ML models (random forests, gradient boosting, neural networks, and ensemble stacking) using a multi-country sample of equity mutual funds between 2010–2023. Machine learning models exploit fund characteristics, flows, and past performance to improve out-of-sample predictions and the construction of tradable portfolios that aim to earn positive alpha net of fees. Our results show ML approaches generate superior out-of-sample classification of top decile funds and modest net alphas in long-short portfolios, while CAPM remains a robust baseline for risk-adjustment and interpretability. We discuss implementation caveats, look-ahead bias, survivorship bias, transaction costs, and feature stability and propose a hybrid workflow that uses CAPM measures as targets / features in ML models for better interpretability and regulatory compliance.