Research On Stock Trading Based On Deep Reinforcement Learning

During the period of stock trading,investors judge the stock market and choose trading actions through their previous experience,they are easily affected by their own sentiment in the process of making trading decisions.Facing a large number of trading data,difficulty for investors to analyze and obtain potentially favorable trading patterns.And a variety of factors can affect the stock market and cause stock price fluctuations.Stock trading based on neural network model can avoid the influence of investors sentiment during the trading process,and learn the trading strategy that can obtain more investment returns.When analyzing the stock market through a deep neural network model,how to minimize the impact of noise in stock data,analyze the temporal relationship between financial data,and obtain useful information from different data sources are still problems to be solved.This paper mainly studies stock trading based on deep reinforcement learning(DRL),and analyze stock market through multi-source data.The main work of this paper are as follows:Firstly,deep learning,reinforcement learning and DRL algorithms are introduced,which lays the foundation for the research content of subsequent chapters.Secondly,the data sources of the existing stock trading researches are relatively single,and lack the analysis on the temporal relationship of stock data.This paper based on DRL judges the stock market by analyzing the candlestick charts and historical stock trading data,and proposes a multi-source data fusion framework to implement stock trading.In this framework,different deep neural networks are adopted to obtain the temporal characteristics,the characteristics are added to implement fusion,the fused features help the agent to learn the optimal trading strategies.To avoid the problem of Q value overestimation,we adopt the improved algorithm based on DQN.From the experiments on different stock markets with different trends,it can be seen that our trading strategies can get higher returns and have better performance.Thirdly,considering financial news impact on the stock prices,financial news and stock historical trading data reflect stock market information from different aspects.When implementing stock trading based on DRL,whether the informative feature representation of the stock market can be obtained from multi-source data will affects the learning of the final trading strategy.This paper analyzes the stock news headlines and historical stock trading data,and proposes a stock trading framework based on DRL.Because the stock news headline is the generalization of the whole report,only the stock news headlines is considered in this paper.For historical stock data and technical indicators,Long Short-Term Memory(LSTM)is used to obtain the time series features,and get more useful information through the attention mechanism,and the fusion method of weighted feature addition is adopted to obtain the fused features.In the setting of reward function,sharpe ratio is added,and investment risk is taken into account as well as return.The experimental results show that the trading strategies learned by the proposed framework have better performance in different fields of stocks.