Stock Market Forecasting with Differential Graph Transformer: A Novel Approach to Temporal and Spatial Stock Data Analysis

Abstract

Forecasting stock market trends are a difficult endeavor because of the intricate interrelations and ever-changing characteristics of financial markets. This paper presents the Differential Graph Transformer (DGT), a novel deep learning model that integrates temporal attention with differential graph attention methods to understand time-series dynamics and interstock relations. Utilizing global and local correlation matrices based on mutual information and Pearson coefficients, the DGT surpasses conventional models in forecasting stock prices. Experiments performed on the S&P500 dataset indicate that the DGT results in a 13.5% lower Root Mean Squared Error (RMSE) and a 12.2% reduction in Mean Absolute Error (MAE) when compared to baseline GRU models. Significantly, the DGT utilizing local mutual information matrices demonstrates the highest performance, validating its capacity to accurately model short-term interstock dependencies. This research highlights the capability of differential attention mechanisms in enhancing stock market predictions.