Intelligent Stock Forecasting by Iterative Global-Local Fusion

Jiahao Qin, Bihao You, Feng Liu

Abstract: Predicting future trends in financial time series data is a challenging task due to issues such as missing data, noise influence, and the heterogeneity gap between different data modalities. In this study, we propose a Global-Local Multimodal Guidance Fusion for Intelligent Stock Forecasting (GLMG) model to address these challenges. The GLMG model incorporates several novel components, including a wavelet transform-based approach for multi-scale feature extraction from image modality data, a Residual Modality Completion Module for handling missing data, a Global-Local Fusion Unit for balanced integration of cross-modal information, and a Stepwise Fusion module for reducing computational complexity and improving interpretability. We conduct extensive experiments on real-world stock market datasets to evaluate the performance of the GLMG model. The results demonstrate that our proposed approach outperforms methods, achieving an accuracy of 64.23% for stock risk early warning and 65.42% for stock movement forecasting, surpassing the best baseline by 6.30% and 6.28% respectively. The GLMG model also exhibits superior profitability, yielding a 16.23% profit while maintaining a low maximum drawdown of 7.96% in simulated trading scenarios.