Multi-factor portfolio optimization: A combined random Forest–AdaBoost model with cost-sensitive learning1

Abstract

This paper proposes a machine learning-driven multi-factor investment strategy, denoted as DE-CS-RFA, which integrates the Random Forest-AdaBoost (RFA) ensemble learning model, Cost-Sensitive (CS) learning, and the Differential Evolution algorithm (DE). The model utilizes 110 heterogeneous predictive features as input characteristics, eliminating redundant features via Kendall correlation analysis to enhance computational efficiency while comprehensively capturing market information. Subsequently, the Rank-Sum Ratio comprehensive evaluation method is employed to construct the initial investment universe and to develop an investment strategy based on the model's predicted data. Empirical results demonstrate that RFA outperforms other mainstream machine learning models on multiple evaluation metrics. Moreover, the simulation trading results indicate that the DE-CS-RFA model can effectively capture the market complexity and individual investor differences, enhancing the applicability and effectiveness of the investment strategy. Interpretability analysis further reveals the key factors influencing the stock price trends in the A-share market. Finally, robustness tests confirm that the DE-CS-RFA model can adapt to diverse financial market characteristics, holding potential to promote the widespread application of multi-factor investment strategies in the A-share market.