【摘要】 针对投资组合管理问题，设计一种基于深度强化学习TD3(Twin Delayed Deep Deterministic policy gradient algorithm)双延迟确定性策略梯度算法的投资组合框架，投资者通过观察股票的因子信息做出决策以达到终期收益最大。因子选择上采用LGBM方法选取有效因子，模型训练过程通过数据增强的方法加强对环境的探索能力。选取两组股票做为风险资产，TD3策略在测试时期的年化收益均超过60%，夏普比率均超过2，综合来看TD3策略收益、风险控制、稳定性方面都要显著优于其他对照组（等权重、沪深300指数和DDPG策略），表明该策略在风险与收益的综合指标下有效。更多还原

【Abstract】 Aiming at the problem of portfolio management,this paper designs a portfolio framework based on the deep reinforcement learning TD3 algorithm. Investors make decisions by observing the factor information of stocks to maximize the final return. In factor selection,the LGBM method is used to select effective factors,and the model training process strengthens the exploration ability of the environment through the method of data reinforcement. Two sets of stocks are selected as risk assets. The annualized returns of the TD3 strategy during the test period are both more than 60%,and the Sharpe ratio is more than 2. Taken together,the TD3 strategy’s return,risk control,and stability are significantly better than other control groups,such as equal-weight,CSI300 index and DDPG strategy,indicating that the strategy is effective under the comprehensive indicators of risk and return.更多还原