A  new method of financial multivariate time series forecasting based on complex network attention mechanism

Abstract

Due to the high degree of correlation and interaction among multiple financial series, as well as the non-stationary and nonlinear characteristics of these series, predicting financial multivariate time series accurately using traditional time series analysis methods is challenging. However, neural network models can effectively handle nonlinearity and instability, enabling relatively accurate predictions. In this paper, we propose a novel method called MTP-CVGAT(Multivariate Time-Series Prediction via Correlation and Visible Graphs Attention Network), which combines the correlation and visibility features of multivariate series. Taking into account accommodation and catering stocks of companies listed on China's Shanghai-Shenzhen A-share market as an example, our results demonstrate that the prediction performance of MTP-CVGAT surpasses that of other models including Long short-term memory neural network model(LSTM), multi-variable time series anomaly detection model based on graph attention network, and transfer learning model based on Wang X G et al.'s approach.