Stock index prediction using global market indices: A Granger causality-based graph representation learning method

Abstract

Global stock markets are increasingly connected and may experience strongly correlated stock price fluctuations due to external shocks, cooperation, and competition. In this paper, we propose a Granger causality-based graph representation learning method to improve the accuracy of stock index prediction. In the first stage, a Granger causality test is performed using the historical data of the target stock market and other global stock markets to select the neighboring stock markets with Granger causality and get the relational weight of each neighboring stock market. Then, the temporal weight of trading days is obtained based on the constructed VAR model coefficients. Subsequently, the daily graph embedding of the target stock index is learned by graph convolution operation for the prediction of the next day's closing price. The method developed in this paper experimented on four well-known Chinese stock indices, and it outperformed the benchmark model used in this paper.