Research And Implementation Of A General Model For Asset Portfolio Management Based On Deep Reinforcement Learning

According to the latest statistics from the World Bank Organization on the scale of the global financial market in 2022,China has become the second largest financial market in the world after the United States with a scale of more than 12 trillion(USD).The prosperity and development of Chinese financial market has attracted more and more investors who wish to earn capital gains through quantitative trading.However,investing through quantitative trading has three difficulties for ordinary investors: the first,facing with the many different investment theories,the complex models of quantitative trading,investors are difficult to adopt effective methods;then,data acquisition and preprocessing is error-prone,model establishment,development and debugging are arduous;the last,model training is time-consuming and lacks standard model evaluation.Without professional knowledge and proficient skills,ordinary investors are unable to implement effective quantitative trading strategies.Therefore,on the basis of analyzing the characteristics of the existing mainstream asset portfolio management technology,our paper designs and implements a general model of asset portfolio management based on deep reinforcement learning,which is friendly to ordinary quantitative trading and investment enthusiasts.The model mainly has the following three advantages:(1)Wide applicability.On the one hand,our model can support trading windows of different time granularities,such as minutes,days,and weeks in the stock market;on the other hand,this model can simulate the trading environment of multiple stock markets,including the Dow Jones Index(DJIA),the Shanghai Stock Exchange 50(SSE50)and CSI 300,etc.;(2)High integrity.Our model provides the mainstream DRL algorithm(DDPG,PPO)implementation and complete asset portfolio management capabilities,including various dataset preprocessing,multi-market environment simulating,well-designed reward function,as well as standard evaluation index calculation;(3)Strong extensibility.Adopting hierarchical structure organization and modular design,our model is divided into asset portfolio service,dataset service,model service and environment service.Beyond that,our model retains a complete user extension interface for implementing different reinforcement learning algorithms,preprocessing different market datasets,and simulating different market environments.In addition,we designed a series of comparative experiments to evaluate the practicability of the model.In-depth experiments on the stock markets of China and the United States reveal that our model gains a cumulative return of more than 30%,and the practicality is obvious.After the above work has been done,we start the optimization on the existing models in three directions,data enhancement,exploration enhancement and policy network enhancement.Comparing with the benchmark experiments,the feasibility of the model optimization scheme is verified.