Research On The Optimization Method Of Financial Market Portfolio Based On Reinforcement Learning

The financial market is an important part of the global economy,and its changes have certain inherent regularity,and are also affected by various constraints,such as market,policy,epidemic diseases,scientific and technological development and many other factors.In the financial market,investors are most concerned about the stock market.However,when China established the securities market,all kinds of speculation will lead to false prosperity of the stock market,and even induce investors to judge irrational investment operations,and regulators will also take wrong regulatory measures,which will make the whole stock market vulnerable and dangerous.Therefore,how to accurately predict the trend of the financial market will be of great reference value to relevant practitioners and researchers.Traditional financial time series research mainly focuses on linear trend fitting based on statistical knowledge.However,the trend of financial market is not limited to simple mathematical function curve,but is comprehensively influenced by various factors in all walks of life.The development of deep learning began to make up for the diversified feature extraction that statistical theory could not complete,and improved the efficiency of the model.Today,information exchange is getting faster and faster,and investors' emotions are slowly influencing the trend of the stock market,while emotional characteristics are not related to the trend prediction of financial time series.Moreover,in the current financial market trend forecasting technology,most researchers focus on forecasting the trend of the financial market,focusing on its theoretical research,ignoring its practical significance.In this paper,based on deep learning,a model for forecasting the trend of financial market in small sample environment and a model for analyzing the trend of financial market to simulate the trading of stock market are constructed.The main work done is as follows:Firstly,a text emotion extraction method based on LSTM and Word2 vec is proposed.Through the crawler technology,we get the text comment information from the stock comment section of Oriental Fortune.com,sort out the data set that can be used in the experiment,and then combine the word2 vec algorithm model to get the word vector representation of the stock comment text information.Then,we use LSTM deep learning model to classify the obtained word vector into three categories of emotion(positive,neutral and negative),and calculate the emotional weight of each stock comment.Then,we cumulatively calculate the emotional value of every day in the obtained text data,so as to get more features for the financial market trend prediction.Secondly,in order to better tap the features of financial data,a financial market trend prediction method based on emotional analysis technology in small sample environment is proposed.By using the combination mechanism of convolution LSTM and LSTM,combined with emotional factors mined from stock evaluation information,the trend of stock price is predicted in small sample environment,and before the next LSTM,a dropout is added to avoid over-fitting.The prediction results show that the prediction effect of the proposed method is better than that of the traditional method in small sample data.Thirdly,a trend analysis method of financial market based on reinforcement learning is proposed to analyze the portfolio of different stocks.In this method,Markowitz portfolio theory is used,and the closing price,total number of shares and market value of stocks are used as the characteristic inputs to determine the portfolio weight distribution.Then,the mode of combining DDPG with custom experience box is introduced.DDPG is responsible for finding the time point suitable for buying and selling in the stock price trend,and combining with the custom experience box,the operation certainty and share of buying and selling are confirmed twice,and the stock market is simulated.Finally,the final income is counted and compared with the traditional baseline method and other reinforcement learning methods.The results show that the proposed method is superior to the traditional baseline method and other reinforcement learning methods in simulating stock market returns.