Research on dynamic portfolio optimization and risk control based on DRL

Dynamic investment ensemble methods that integrate deep reinforcement learning (DRL) often suffer from performance lag caused by rolling window hysteresis. To address this, this study proposes a DRL-based dynamic portfolio optimization and risk control framework. The method employs a two-layer nested ensemble of three agents—advantage actor-critic (A2C), deep deterministic policy gradient (DDPG), and proximal policy optimization (PPO). In the first layer, stage models are selected by optimizing the Sharpe ratio; in the second layer, the three DRL algorithms are integrated via weighted voting, with multiple model snapshots collected during training for ensemble prediction. Experiments on the Shanghai Stock Exchange 50 Investment Index, the Dow Jones Index, and their constituent stocks compare performance against a passive holding index strategy and a mean–variance portfolio strategy. Parameter tuning yields optimal configurations for each model, including 5 time steps, 0.005 entropy coefficient, and 0.0007 learning rate for A2C; 0.01entropy coefficient, 2048 time steps, 0.00025 learning rate, and batch size of 128 for PPO; and 10,000 replay buffer size, 0.0005 learning rate, and batch size of 64 for DDPG. Results demonstrate that the proposed method delivers strong practical performance and improved effectiveness in dynamic portfolio management.