Research On Stock Pool Construction In Shanghai And Shenzhen Stock Markets Based On Deep Learning

According to Markowitz's classical portfolio theory,in order to achieve the optimal risk portfolio with the lowest standard deviation given the same expected return,analysts have to make accurate predictions of the expected return of each stock,the variance of expected return and the covariance between the expected return of each two stocks.However,with the increase of the number of listed companies and the total amount of financial data,especially unstructured data,which can be used for analysis.It is not possible and economic to analyse and predict all the stocks.This paper mainly aims at the above problems in the stock pool construction process at the beginning of portfolio construction.Starting from Markowitz's portfolio theory,instead of directly predicting a large number of covariances to be evaluated or offering investment suggestion,in order to minimize the non-systematic risk of the portfolio,this paper explores the feasibility of a Deep Learning stock pool construction method,in order to reduce the workload and improve efficiency.The research content does not involve the allocation of stock weight in portfolio,the selection of trading opportunity and the subsequent management of portfolio.Firstly,we extract features from the K-line graph of stock by Convolutional Neural Network AutoEncoder,that is,convert the K-line graph of stock into a vector,which represents the trend of the stock in the selected period of time.By reducing the dimension and clustering the transformed vectors,we can classify the stocks with similar price trend in the selected period into the same category,and the stocks with different trend into different categories,which can realize the purpose of risk diversification.Finally,the self-attention mechanism model is used to select a stock within each category,and the correlation coefficient matrix of the selected stock is calculated to verify the risk diversification ability of the stock pool construction method.Finally,in the 5 randomly selected time periods,by presenting the coefficient matrix and some profit indicators to evaluate the effectiveness of the constructed stock pool.The experimental results show that the proposed stock pool construction method can achieve good risk diversification effect in most time periods,and the reasons for failure in the last time period will be further discussed in the text.