An Empirical Study On Stock Portfolio Management Based On Deep Reinforcement Learning

In recent years,with the continuous emergence of artificial intelligence research and development teams such as Google DeepMind,in-depth reinforcement learning has gradually become well known.At the same time,China's stock exchange market is gradually developing towards diversification,convenience and informatization,which makes the stock exchange market produce a large amount of data every day.In order to solve the defect that the traditional transaction analysis method is difficult to deal with a large amount of data,and to avoid the irrational operation that human investors lack discipline,quantitative investment with scientific,systematic and accurate has gradually become an investment method that institutional investment researchers focus on and develop.Quantitative investment is the use of computers to model financial data,mining the investment opportunities and trading signals contained in the data,and then building a dynamic and adaptive investment strategy,which is similar to indepth reinforcement learning.Therefore,it is feasible to combine the two to solve real financial problems.The application of deep reinforcement learning in the field of quantitative investment broadens the research thinking and modeling method of quantitative investment on the one hand,and promotes the development of financial transactions towards intelligence and automation on the other hand.In order to obtain the optimal investment opportunities and trading signals,this paper attempts to apply Deep Reinforcement Learning(DRL)method to the quantitative investment field to solve the problem of portfolio management in the stock market.Based on the principles and framework of A2 C,PPO and SAC three kinds of Deep Reinforcement Learning algorithms based on Actor/Critic algorithms,this paper builds a model agent,combines the historical real price and trading volume of the stock market with the technical index data of traditional investment analysis,normalizes the original data by using the method of mathematical statistics,obtains a state data matrix containing historical characteristic information of the stock market,and extracts robust and effective characteristic expressions from the state data by using the agent's internal deep neural network.Based on this feature,three reinforcement learning algorithms are used to optimize the respective strategies,and the relevant parameters of the deep neural network are optimized by combining the current reward value,so as to obtain the optimal trading strategy in an exploratory way.In the experiment,some constituent stocks of CSI 300,SSE 180 and CSI 200 Index were selected as trading asset portfolios,and tested in bull market,bear market and shock range respectively.Finally,the experimental test results are presented and analyzed from the perspective of returns and risks.The results show that the A2 C,PPO and SAC models are generally effective in the upward,downward and volatile market.Among them,the SAC model has the most obvious advantages in revenue acquisition and risk control,and is the most competent for the task of stock portfolio management.It can help investors to obtain 15%?32% of excess return while withdrawing less.Compared with the SAC model,the A2 C and PPO models are less effective,obtaining 7%?20% of excess return.This study expands the application of Deep Reinforcement Learning in the field of quantitative investment,and has strong reference significance and practical guiding value for the application of Deep Reinforcement Learning in the field of financial investment,especially in the field of stock investment.