Research On Stock Correlation Prediction Based On Spatial-temporal Graph Convolutional Neural Network

In the stock market,different companies may be influenced by the same macroeconomic,industry trends,policy changes,and other factors,leading to a phenomenon of simultaneous rise and fall in their stock prices.Research has shown that there is a correlation between stocks,especially within the industry sector,indicating that stocks have potential correlation characteristics.Stock correlation refers to the degree of correlation between stocks,which reflects the trend of changes between different stocks in the market.Usually,Pearson correlation coefficient,Spearman rank correlation coefficient,Kendall Tau correlation coefficient,etc.are used for correlation analysis.In practical applications,people usually use these correlation information for combinatorial optimization and risk management,which can also provide some reference for research such as stock price forecasting.With the complexity and globalization of modern financial markets,the interrelationships between various financial assets and markets are becoming increasingly close.How to accurately and efficiently predict stock correlations has become a highly concerned research direction in the financial field.(1)This article proposes a differential spatial-temporal graph convolutional network model(DST-GCN),which uses a spatial-temporal graph convolutional network to model and analyze the relationship between industry sectors and stocks.Local spatial-temporal correlation sequences of stocks are constructed,and a differential solver is used in the convolutional layer of the differential spatialtemporal network model to differentially process the hidden output state features of the convolutional layer to obtain more precise time series information.(2)This article proposes a differential spatial-temporal graph convolutional model(DAST-GCN)that integrates attention from spatial-temporal information.The encoder and decoder structures are used to construct the temporal and spatial embedding information of index component stocks,which is input into the model along with the historical correlation features of stocks.The attention coefficients of stock nodes in the spatial-temporal dimension are calculated in the convolutional layer of the differential spatial-temporal graph,and the two types of attention are fused using a gated fusion mechanism,Use an attention output conversion layer to generate future representations,input to the decoder to generate predicted values.In the experimental analysis of this article,we selected 9 sectors in the Chinese A-share market,including finance,manufacturing,real estate,and transportation,as well as 15 representative indices as the research objects.Then,we screened 72 sector stocks and 75 index component stocks with sufficient historical data and good liquidity,and constructed corresponding stock relationship graph network structures for correlation prediction experiments.On the premise of ensuring sufficient data,we compared the effectiveness of our model with four baseline models and concluded that our model has a significant improvement effect in predicting stock correlation in the short term(1 day,3 days),medium term(19 days),and long term(60 days).Overall,in the face of complex stock markets,the model proposed in this article has shown good performance in predicting stock correlation.Predicting stock correlation not only helps investors better understand the performance and risks in the stock market,but also helps them formulate hedging strategies more effectively,thereby better diversifying investment and risk management,and maximizing investment benefits,It can also provide reference for research on stock price prediction,and predicting stock prices based on the correlation characteristics between stocks can improve the accuracy of investment decisions.Therefore,the method proposed in this article has strong research significance and practical value.