HGAIT: heterogeneous graph attention with inverted transformers for correlation-aware stock return prediction

초록

This study introduces HGAIT, a novel predictive framework that substantially enhances stock return prediction by synergistically integrating Transformer-based architectures with heterogeneous graph attention networks. Building upon recent advances like inverted Transformer, HGAIT employs channel-independent GRU structures instead of traditional MLP-based embeddings, effectively preserving intrinsic temporal inductive biases characteristic of financial data. The framework explicitly models intricate inter-variable interactions through dedicated variable attention mechanisms, capturing essential nonlinear dependencies among financial indicators. Additionally, a heterogeneous graph attention layer dynamically constructs asset neighborhoods based on positive and negative correlations, comprehensively integrating structural asset interrelationships crucial for accurate return prediction. Empirical validations using comprehensive U.S. market-wide data demonstrate HGAIT's superior predictive capabilities. Notably, the model significantly outperformed benchmark methods across multiple metrics, particularly excelling in ranking-based indicators such as Rank Information Coefficient and Rank Information Coefficient Information Ratio. Extensive portfolio back-testing further confirmed its practical effectiveness, with HGAIT achieving remarkably higher Sharpe and Sortino ratios alongside the lowest maximum drawdowns, highlighting its exceptional risk-adjusted returns and robust downside risk management. Sub-period analyses across diverse market regimes, including stable, transitional, and highly volatile periods, further validated its predictive stability, emphasizing HGAIT's robustness in adapting to dynamic financial environments. The generalizability of these findings across comprehensive market data confirms HGAIT's broad applicability, making it a powerful and reliable tool for real-world financial decision-making and portfolio management.

키워드