Hierarchical Graph Learning for Calendar Spread Strategies in Commodity Futures Markets

Battle Count: 90

Highly relevant to quantitative trading. Proposes a novel learning-based approach for calendar spread strategies in commodity futures, a widely-used but previously unaddressed strategy in ML literature. Provides theoretical guarantees for CS strategy efficacy, introduces a practical prediction-to-position mapping method, and demonstrates strong empirical performance (96% IR improvement over benchmarks, 75% over S&P 500). Directly applicable to statistical arbitrage in futures markets with near-zero correlation to equities.

High complexity due to: (1) bi-level hierarchical graph construction with cross-level edges, (2) graph lifting method requiring virtual TTM grid alignment and linear interpolation, (3) four-step bi-level convolution (elevating, cross-commodity propagation, lowering, intra-commodity propagation), (4) dynamic trading universe with node birth/death from contract maturities, (5) two-way ANOVA-style feature centering, (6) probability integral transform for targets and features, (7) CS position projection via commodity-wise demeaning. Requires GNN framework (PyTorch Geometric) and careful handling of irregular maturity grids.

The paper provides detailed mathematical formulations, hyperparameter grids, and experimental settings. Data is sourced from LSEG Datastream via WRDS (commercial database). No GitHub repository is mentioned. The architecture is fully described with equations, but implementation code is not provided. Statistical tests and data processing steps are documented in appendices.