BED-RL: Bagging-Based Encoder–Decoder Reinforcement Learning for Dynamic Portfolio Management

Portfolio management in modern financial markets faces numerous challenges, primarily due to the complexity of market dynamics, high volatility, and noisy data environments. Traditional methods struggle to adapt to these evolving conditions, requiring more sophisticated approaches. In this paper, we propose a Bagging-based Encoder-Decoder Reinforcement Learning (BED-RL) framework, designed to address the limitations of the current State-of-the-Art (SOTA) model, EarnMore, such as its susceptibility to overfitting and lack of diversity in predictions. Our novelty lies in explicitly promoting diversity via adaptive masking of stock features and parallel encoder–decoder branches, with two aggregation variants that reduce overfitting and improve generalization under regime shifts. It leverages multiple encoder-decoder modules trained on different data subsets, forming an ensemble that significantly enhances model robustness and risk-adjusted performance. Compared to the SOTA model, our proposed method achieves significant improvements across both datasets. On the S&P 500, it improves the annualized return from 97.17% to 110.54% and the Sharpe ratio from 2.032 to 2.350. On the DJ30, it boosts the annualized return from 54.11% to 73.06% and the Sharpe ratio from 1.457 to 1.762. Despite using multiple branches, inference remains under 2.1 seconds per portfolio decision with 4 modules, enabling daily or hourly rebalancing in production, and branches are inherently parallelizable for scalable deployment. These gains together with explicitly diversified representations and low-latency, parallelizable inference bridge ensemble theory and real-world portfolio management, supporting robust and adaptive deployment.