MTIDHHGAN: A Multimodal Temporal Information-Driven Hierarchical Heterogeneous Graph Attention Network for Stock Movement Prediction

The continuous change in stock market dynamics as well as stochastic patterns has transformed stock movement prediction into an enduringly complex scientific challenge. Graph neural networks became ubiquitous tools for modeling stock relations across the industry during the past few years. Despite their proven effectiveness the stock market dynamics exceed the capacity of present-day graph neural networks to process complex system evolution and numerous interconnected patterns. The proposed framework introduces MTIDHHGAN which stands for Multimodal Temporal Information-Driven Hierarchical Heterogeneous Graph Attention Network. This model represents its main strength through its ability to unite and embed two different kinds of multimodal data types (stock prices and news information) at multiple temporal scales. These multimodal embeddings are then used to make predictions about stock movements through a hierarchical heterogeneous graph neural network that models the relationships between companies and their associated executives. Experimental testing of MTIDHHGAN was executed on two stock market datasets named CSI100 and CSI300. MTIDHHGAN proves superior to prevailing baseline models using test results from both datasets. The prediction results strongly demonstrate that the MTIDHHGAN framework achieves optimal success when used for stock movement predictions.