基于深度强化学习的投资组合优化研究

Research on Portfolio Optimization Based on Deep Reinforcement Learning

摘要: 本文将投资组合调仓过程建模为马尔科夫决策过程，基于Actor-Critic算法框架进行策略优化。为增强状态表征能力，本文整合三类多源异构数据(股票历史交易数据、技术指标、K线图和财经新闻标题)丰富模型的状态空间。最后，将从上述三种数据中提取到的特征进行拼接融合，形成深度强化学习算法所需的环境状态。基于这一状态，算法能够学习并优化投资组合的交易策略，以实现收益最大化和风险最小化的目标。在中国A股市场的实证研究表明：本文提出的投资组合优化策略收益显著超越了自定义价格加权指数和其他传统的静态交易策略，多空交易测试验证了其于市场下行期的稳健性。

Abstract: This paper models the portfolio rebalancing process as a Markov Decision Process (MDP) and optimizes the strategy based on the Actor-Critic algorithm framework. To enhance state representation, the study integrates three types of multi-source heterogeneous data—historical stock trading data, technical indicators, and candlestick charts with financial news headlines—to enrich the model’s state space. Finally, the features extracted from these three data sources are concatenated and fused to form the environmental state required by the deep reinforcement learning algorithm. Based on this state, the algorithm can learn and optimize the trading strategy of the portfolio to achieve the dual objectives of maximizing returns and minimizing risks. Empirical research in China’s A-share market demonstrates that the proposed portfolio optimization strategy yields significantly higher returns compared to custom price-weighted indices and other traditional static trading strategies. Long-short trading tests further verify its robustness during market downturns.