Abstract
We employ machine learning methods to identify skill among active bond mutual fund managers. Using a comprehensive dataset of 3,021 unique U.S. bond funds from May 1995 to November 2024, we demonstrate that fund-level and family-level characteristics, particularly past performance metrics, reliably predict future bond fund performance. A prediction-weighted portfolio strategy that goes long the best-10% of funds and short the worst-10% of funds generates monthly abnormal returns of 30 basis points with an information ratio of 24.6%. The outperformance persists for up to 36 months. Holdings-based characteristics do not help to separate good from bad managers. The predictive signal is strongest for corporate and municipal bond funds, while Treasury bond funds exhibit weak performance differentiation. Incorporating equity-side information for corporate bond fund families that also offer equity mutual funds does not enhance predictive performance.