A Stock Prediction Model Based On Community Detection And Graph Attention Networks

Stock forecasting aims to predict the future trend of stocks to help investors make good investment decisions,and using statistical models to predict stocks has become a hot research content in the stock market.Stock price fluctuations are not only related to their own state,but also closely linked to the state changes of other stocks.Previous research on stock forecasting mainly learns inter-stock interactions through stock industry relationships or supply chain relationships,without considering the correlations presented by stock price fluctuations.This paper portrays stock price fluctuation through stock yield fluctuation,based on the existing stock forecasting model that learns the interactions among stocks through graph attention networks jointly with stock industry relationships,this paper proposes to construct a new stock relationship using the overlapping community structure of stock networks to mine the correlations of stock yield fluctuations to replace the industry relationships in the stock forecasting model,and based on the historical trading data of 134 stocks in the CSI 300 in 2020 and 2021,we compare and analyze The stock forecasting effects when using the industry relationship and the new stock relationship are compared.The results show that stock relationships based on correlations of stock yield fluctuations enhance the stock forecasting model’s ability to learn the interactions among stocks relative to industry relationships.When the stocks ranked in the top 1,5,10,15,and 20 forecast returns are evaluated separately,the new stock relationship achieves the largest improvement of more than 10% relative to the industry relationship in both the average inverse ranking and the average ranking hit ratio model evaluation metrics.Meanwhile,the prediction effect fluctuates significantly with the change of community structure,and there is a significant difference in the overlapping community structure of the network detected by different overlapping community detection algorithms,and the more stable the overlapping community detection algorithm is,the less fluctuation in the prediction effect.This study combines the correlation of stock yield fluctuations with stock prediction models through the overlapping community structure of stock networks,providing a new stock relationship for learning the interactions between stocks.Ranking stocks in terms of returns can provide investors with an objective comparison of future trends of stocks and help them make reasonable investment decisions.