Machine learning and manager selection: evidence from South Africa

PurposeThis study investigates whether the skilled minority of active equity managers in emerging markets can be identified using a machine learning (ML) framework that incorporates a large set of performance characteristics.Design/methodology/approachThe study uses a cross-section of South African active equity managers from January 2002 to December 2021. The performance characteristics are analysed using ML models, with a particular focus on gradient boosters, and naive selection techniques such as momentum and style alpha. The out-of-sample nominal, excess and risk-adjusted returns are evaluated, and precision tests are conducted to assess the accuracy of the performance predictions.FindingsA minority of active managers exhibit skill that results in generating alpha, even after accounting for fees, and show that ML models, particularly gradient boosters, are superior at identifying non-linearities. LightGBM (LG) achieves the highest out-of-sample nominal, excess and risk-adjusted return and proves to be the most accurate predictor of performance in precision tests. Naive selection techniques, such as momentum and style alpha, outperform most ML models in forecasting emerging market active manager performance.Originality/valueThe authors contribute to the literature by demonstrating that a ML approach that incorporates a large set of performance characteristics can be used to identify skilled active equity managers in emerging markets. The findings suggest that both ML models and naive selection techniques can be used to predict performance, but the former is more accurate in predicting ex ante performance. This study has practical implications for investment practitioners and academics interested in active asset manager performance in emerging markets.