TGNS: A transformer-based graph neural network for stock trend forecasting

Advancements in computational technology and financial theory have accelerated quantitative investment research, with stock trend forecasting emerging as a critical challenge. While deep learning models have shown promise, many rely exclusively on market data, ignoring fundamental factors (e.g., financial health, industry trends) and dynamic concept-stock correlations. This paper presents TGNS, a Transformer-based Graph Neural Network framework that addresses these limitations. TGNS integrates a Concept Graph Attention (CGA) module and Stock Graph Attention (SGA) module to model time-varying concept dependencies and global stock interactions, mitigating over-smoothing in traditional graph networks. A novel Weighted Robust Loss (WRL) prioritizes high-return stocks by exponentially weighting cross-sectional returns, enhancing long-short strategy performance. Experiments on CSI100/300 and NASDAQ-100 datasets demonstrate TGNS outperforms state-of-the-art methods (e.g., PatchTST, HIST) by 12.3–21.9% in Information Coefficient (IC). Ablation studies validate the dual-attention architecture's effectiveness in capturing complex market dynamics, confirming TGNS's superiority in dynamic financial environments.