Research On Stock Index Prediction Based On Representation Learning Of Stock Correlation Network

In recent years,stock index prediction has attracted much attention from academia and industry,and the stock correlation network has become a research hotspot in the complex network field.Because the stock market has the characteristics of high noise,nonlinear,non-stationary,and chaotic,the stock index time series represents its inherent complexity and provides essential investment reference for investors.Accurate prediction of the stock index is beneficial to monitor better and managing the financial market,which is highly correlated with the stock market.It can also provide practical guidance for investment decisions.Traditional stock index forecasting methods are usually based on technical analysis of the stock market,but no single technical index can consistently predict the trend accurately.In addition,there are many complex correlation features among the components of the stock index,such as lead lag,price synchronization,industry rotation,and risk contagion.Therefore,how to construct feature indices for stock index prediction from the perspective of stock correlation networks is a problem worth exploring.In addition,at different stages of the stock index operation,the impact weights of varying industry sectors on future price fluctuations of the stock index also have time-varying properties.How to capture in real-time the impact of different industry sectors on the future trend of the stock index in existing stock index prediction studies has also been rarely reported.To address the above problems,this study proposes for the first time to transform the stock index prediction problem into a graph classification problem,that is,to learn a mapping function that maps the stock correlation network to a set of labels.To capture the impact weights of different industries on future price fluctuations of the stock index at the meso-level,this paper hierarchically aggregates node attribute information of the stock correlation network based on the different industry perspectives using the differentiable pooling(DIFFPOOL)framework,attempting to extract mesoscale predictive factors of stock index price fluctuations,and reconstructs the coarse-grained graph of the stock correlation network.This feature extraction model is driven by the stock correlation network classification task and can effectively aggregate feature indicators of stock index constituents from different industry sectors.Based on the hierarchical clustering results of the stock correlation network,the inter-cluster influences are captured,and the final hierarchical representation vector is aggregated as the feature input of the differentiable classifier.The entire system can be trained end-to-end using stochastic gradient descent.This paper uses historical data of constituent stocks in the CSI 300 Index,S&P500 Index,FTSE 100 Index,and Nikkei 225 Index as the research objects.First,based on the visual graph method,the stock distance matrix is calculated,and the maximum planar filter graph algorithm is used to construct the stock correlation network.Topological indicators are extracted based on complex network theory and combined with technical indicators as feature variables.Secondly,the DIFFPOOL deep learning framework is used to learn the hierarchical representation of the stock correlation network.Finally,the differentiable pooling(DIFFPOOL)architecture and graph convolutional neural networks(GCN)are combined with multiple classification models,such as long short-term memory(LSTM),to compare different stock index prediction models.Lastly,the DIFFPOOL architecture is compared with numerous time series regression models.The results of the experiments show that the prediction accuracy can be further improved by hierarchically aggregating the node attribute information of the stock correlation network.