Forecasting Equity Correlations with Hybrid Transformer Graph Neural Network

Battle Count: 146

Directly addresses statistical arbitrage basket trading with forward-looking correlation forecasts. Demonstrates 19.20% annualized return with Sharpe of 1.837 vs S&P 500's 14.43% and 0.647. Integrates with SPONGEsym clustering and ML signal filtering. Provides actionable insights on regime-dependent correlation dynamics, sector attention patterns, and feature importance for trading strategy design. The 10-day rebalance horizon aligns with practical trading frequencies.

Complex hybrid architecture requiring: (1) 4-layer Transformer with 8 heads processing 30x37 feature sequences, (2) dynamic daily graph construction with edge sampling (top 50, bottom 50, 75 random mid), (3) 3-layer GAT with 4 heads and edge-conditioned gating, (4) three expert MLP heads for correlation regime routing, (5) Fisher-z residual prediction with histogram matching loss, (6) SPONGEsym eigenvalue decomposition and k-means++ clustering, (7) 5-model ML ensemble with grid-searched hyperparameters. Requires GPU training with gradient accumulation. Large-scale evaluation over 269M+ edges.

Architecture details, hyperparameters, training procedure, and feature construction are described in detail. However, no code repository is provided. Data sources (CRSP, Compustat via WRDS, FRED) are commercial/subscription-based. The authors note they could not fully replicate Korniejczuk and Slepaczuk (2024) results, attributing differences to data source (WRDS vs Yahoo Finance). Training details (AdamW, cosine schedule, 75 epochs, batch size 3 with gradient accumulation of 6) are specified.