STN-TGAT: New Graph Attention Model Improves Top-K Portfolio Construction

A new paper on arXiv introduces STN-TGAT, a model designed to tackle stock ranking and portfolio construction under realistic investment constraints. The approach jointly models temporal dynamics and cross-sectional dependencies by combining a temporal Transformer for sequential patterns with a Graph Attention Network (GAT) for inter-stock relationships. A key innovation is the use of an NMI (Normalized Mutual Information)-based prior graph paired with a learnable soft-threshold sparsification mechanism. This method enhances structural robustness by filtering out noisy correlations while preserving informative connections, making the model more resilient to market noise.

STN-TGAT's portfolio formation process mimics real-world trading by incorporating Top-5 selection within the Top-50 S&P 500 constituents, explicit weight allocation, and transaction cost adjustments. Empirical results on real-world data show that STN-TGAT consistently outperforms benchmark models in both predictive accuracy and investment profitability as measured by portfolio returns. The findings suggest that combining decision-aligned training with adaptive relational modeling provides a coherent and practically effective framework for data-driven portfolio construction, bridging the gap between academic research and real-world trading.