Spatial-temporal Graph Attention Network for Forex Forecasting with Hierarchical Transformer

Kaiwen Guan, Yan Ge

Keywords: Graph Attention Networks, Transformer, Forex Forecasting

Abstract: The foreign exchange market, with its daily trading volume reaching nearly trillions of dollars, presents significant opportunities for the application of advanced predictive analytics. Traditional exchange rate forecasting methods often overlook the interdependencies between currencies and struggle with long-range data dependencies, leading to challenges in capturing the true market dynamics. To overcome these limitations, this paper introduces a novel Spatial-Temporal Graph Attention Network with Hierarchical Transformer (STGAT). Our model innovatively combines spatial graph convolutions with a dual-view temporal transformer-based mechanism, utilizing a Temporal Linearity Graph Attention Network (TLGAT) to account for currency relations in a time-sensitive manner. By integrating a linear attention mechanism for enhanced efficiency and capturing both local and global sequential data embeddings, STGAT provides a framework based on a hierarchical transformer for predicting exchange rates. We validate our approach on exchange rates of seventeen currencies over 2,092 trading days, demonstrating superior performance compared to state-of-the-art models.

Primary Area: learning on graphs and other geometries & topologies

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 10093