Financial adaptive graph attention network for cross-stock forecasting

Zhipeng Jiang, Hua Zou, Dengyi Zhang, Qian Cai, Yu Li, Xiaolei Luo

Abstract: In recent years, with the explosion of financial data, the financial research community has increasingly shifted its focus from single-stock to cross-stock forecasting. This shift enhances model performance by leveraging data correlations across multiple stocks and addressing single-stock limitations, through methods such as information fusion and knowledge graph techniques. Specifically Graph Attention Networks (GATs), that excel in precisely integrating multi-source data. However, in many practical scenarios, the growing complexity of stock data not only overloads the training process but also fails to improve GAT model performance, which may be attributed to the limited depth of GNNs or their oversmoothing characteristics. Therefore, constraining the information flow and optimizing the nonlinear mapping may be meaningful. To address these challenges in cross-stock forecasting, we propose a novel graph structure: the Directed Graph of Stock Aggregation (DGSA), which is embedded within the classic GAT model, and the activation function SGFRELU—Shifted GELU Followed by ReLU. In this way, we form the Financial Adaptive Graph Attention Network (FinAdaptGAT). Specifically, in DGSA, information from auxiliary stocks flows through a virtual node to the primary node. The virtual node acts as a valve, enabling the primary node to exert full control over the information flow while limiting the influence of auxiliary stocks. In SGFRELU, GELU is shifted to align its minimum with the origin, and values in the negative interval are zeroed like ReLU, resulting in a combination of GELU’s smoothness and ReLU’s monotonicity, thus providing higher accuracy than both while maintaining ReLU-like speed. As a result, our approach outperforms state-of-the-art models, achieving 62.98% accuracy on the S&P 500 index.

External IDs:dblp:journals/mlc/JiangZZCLL25