Research On Financial Investment Based On Deep Reinforcement Learning

With the continuous development of the times,the financial market has continued to grow,and all kinds of financial investment methods and investment products have gradually emerged and developed.More and more enterprises and individual investors have joined the financial investment and tried to obtain through financial investment pure income.As part of the financial investment market,China’s stock market has continued to expand.In today’s rapid economic development,the increasing number of listed companies and frequent stock transactions have prompted a large amount of transaction data for the stock market every day.In the face of a large amount of stock data,investors will often face high-risk and low returns based on personal wisdom.With the rapid development of artificial intelligence,researchers have begun to use artificial intelligence algorithms to establish financial trading strategies to optimize investment decisions.One of the representative methods is to strengthen learning methods.However,the current enhanced learning method usually models different stocks separately,ignoring the relationship information between stocks.In addition,existing reinforcement learning studies mostly apply one or two reinforcement learning algorithms to the financial market and compare their performance on a single trading task,and these studies mainly focus on foreign stock markets,with little research on domestic financial markets.Based on this,this article proposes a financial investment research based on deep reinforcement learning,and summarizes the main research content into the following two aspects:On the one hand,this article considers the industry sector relationship and fund relationship of stocks,while noting that multiple stocks can belong to the same industry or fund,so the relationship between different stocks is a group relationship,not a paired relationship.Therefore,this article uses a hypergraph structure to model the relationships between stocks,and designs a dual channel hypergraph attention network to capture these two relationships.When aggregating information in each stock hypergraph,attention mechanisms are introduced to distinguish the importance of different stock neighbor nodes.This paper obtains two independent representations of the same stock from two channels,and introduces comparative learning to maximize the mutual information between stocks in the same industry and fund,and improve their own representation learning ability.Furthermore,this article takes the representations of different stocks fused as the environmental states of the strategy function in reinforcement learning,and uses the strategy gradient algorithm to train the model.This paper uses cumulative return,daily Sharpe ratio and daily volatility as evaluation indicators.The experimental results show that the portfolio model proposed in this paper based on contrastive hypergraph reinforcement learning has advantages in both profitability and risk resistance.On the other hand,in the domestic financial market,this article systematically verifies the effectiveness of value function based and strategy gradient based reinforcement learning algorithms in three investment tasks: single stock trading,multiple stock trading,and investment portfolio.The market state of a single stock trading task and multiple stock trading tasks is composed of stock closing prices and financial indicators.The market state of the investment portfolio task is composed of the covariance of the price change rate of the closing price and financial indicators,and the market state is input to the reinforcement learning agent.In three financial tasks,through the interaction between agents and financial markets,daily stock trading decisions are obtained,and asset allocation and analysis are achieved through trading decisions.After executing the buying and selling decision,receive rewards and continuously adjust the strategy to the optimal level based on the rewards,achieving maximum returns and timely risk avoidance.At the same time,the algorithm is deeply compared and analyzed by observing the backtesting results on the cumulative rate of return,sharpe ratio,maximum rollback and other evaluation indicators.