基于深度强化学习的股票量化投资研究  

Research of Stock Quantitative Trading Algorithm Based on Deep Reinforcement Learning

作　　者：宋飞 田辰磊 冯传威 Song Fei;Tian Chenlei;Feng Chuanwei(School of Information Management,Nanjing University,Nanjing 210023,China;College of Science,Nanjing Forestry University,Nanjing 210037,China;School of Mathematics,Nanjing University,Nanjing 210093,China)

机构地区：[1]南京大学信息管理学院,江苏南京210023 [2]南京林业大学理学院,江苏南京210037 [3]南京大学数学学院,江苏南京210093

出　　处：《南京师大学报(自然科学版)》2025年第1期85-92,共8页Journal of Nanjing Normal University(Natural Science Edition)

基　　金：国家自然科学基金项目(12201303)。

摘　　要：针对股票量化投资,将深度强化学习中的Deep Q-learning(DQN)模型应用于算法交易,构建端到端的算法交易系统.首先,利用股票技术分析指标设计股票交易环境,从时间尺度扩充特征集;其次,定义智能体交易的奖励函数和动作空间;然后,设计Q网络结构,将支持向量机和极致梯度提升法学习股票历史数据的涨跌信号加入强化学习中;最后,将算法交易系统应用于中国股票市场,并选择招商银行和泰和科技两支股票以及其余4支股票进行验证,从收益率、夏普比率和最大回撤率三方面评价投资绩效,结果表明该算法系统在收益率上有显著提升的同时,最大回撤率有所降低,模型的抗风险能力较高.This paper focuses on quantitative stock investment and applies the Deep Q-learning(DQN)model from deep reinforcement learning to algorithmic trading,constructing an end-to-end algorithmic trading system.Firstly,the system utilizes stock technology analysis index to design a stock trading environment,expanding the feature set from a time scale;secondly,it defines the reward function and action space for intelligent agent transactions;then,it designs Q-network structure and incorporate Support Vector Machine and eXtreme Gradient Boosting method to learn the rise and fall signals of stock historical data into reinforcement learning;finally,the algorithmic trading system is applied to the Chinese stock market and China Merchants Bank and Taihe Technology,as well as the remaining four stocks are selected for validation.The investment performance is evaluated from three aspects:return rate,sharpe ratio,and maximum drawdown rate.The results are shown that the algorithmic system significantly improved the return rate while reducing the maximum drawdown rate,indicating that the model has a high risk resistance ability.

关 键 词：量化投资 DQN 强化学习 算法交易 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]