Deep reinforcement learning-driven intelligent portfolio management with green computing: Sustainable portfolio optimization and management

Abstract

Portfolio management remains a key area in quantitative trading. To address limitations in existing deep reinforcement learning (DRL)-based models, which fail to adapt trading strategies and properly utilize supervisory information, we propose a Dynamic Predictor Selection-based Deep Reinforcement Learning (DPDRL) model. The DPDRL model integrates multiple predictors to forecast stock movements and dynamically selects the most accurate predictions, optimizing investment allocation via a market environment evaluation module. Our model was evaluated using daily candlestick data from the SSE 50 and CSI 500 indices. The results show that DPDRL outperforms other models in key evaluation metrics: it achieves a 48.99 % Annualized Rate of Return (ARR), a Sharpe ratio of 2.34, an Annualized Volatility (AVoL) of 0.1390, and a Maximum Drawdown (MDD) of 8.21 %, significantly improving risk-return performance. Ablation experiments confirm the contributions of the dynamic predictor selector and market evaluation module to the model's accuracy and decision-making quality.