InfoGAT: A Superior Stock Recommendation Framework Combining Informer Time-Series Modeling with Graph Attention Networks

The growing significance of equities in portfolio allocation has intensified the demand for advanced predictive models capable of handling noisy, high-dimensional, and highly nonlinear financial data. Traditional approaches often fall short in jointly modeling the structural dependencies among stocks and the long-term temporal dynamics inherent in financial markets. To address these challenges, we propose InfoGAT, a novel multi-scale prediction framework that integrates Informer-based time-series modeling with Graph Attention Networks (GAT).InfoGAT incorporates three key innovations. First, the GAT module dynamically encodes inter-stock structural relationships, capturing both strong intra-industry correlations and weak cross-industry dependencies. Second, the Informer model, with its probabilistic sparse attention mechanism, efficiently learns long-term market dynamics and periodic fluctuations. Third, a multi-scale graph reconstruction with hierarchical regularization enables dynamic correction, mitigating over-concentration, preserving prediction diversity, and preventing degradation during inference.An end-to-end multi-task learning architecture further allows InfoGAT to simultaneously forecast future returns and price trends, while a Top-K selection mechanism translates predictions into actionable investment strategies. Extensive experiments on the CSI-300 and S&P 500 datasets demonstrate that InfoGAT consistently outperforms four state-of-the-art baselines in terms of annualized return, Sharpe ratio, and information coefficient, highlighting its strong cross-market generalization and practical relevance for quantitative investment.