From News to Returns: A Granger-Causal Hypergraph Transformer on the Sphere

Battle Count: 82

Highly relevant to quantitative trading. The paper directly addresses asset return prediction, top-asset ranking (NDCG@10), and market regime classification (bull/bear) — all core tasks in quantitative trading. The interpretable causal attention pathways (e.g., Fed rate hike → finance sentiment → energy sentiment → SPY → JPM) provide actionable signal attribution. The model demonstrates robustness during volatile periods (COVID-19 crash, Fed hikes) which is critical for trading systems. The event-aligned forecasts with confidence intervals support risk-aware position sizing. However, the paper does not directly address transaction costs, slippage, or portfolio-level optimization.

High implementation complexity due to the integration of multiple advanced components: (1) multivariate Granger causality testing with FDR correction, (2) directed hypergraph construction from causal tests, (3) Riemannian manifold embedding with projection operators, (4) geodesic angular attention masking in transformers, (5) manifold-constrained gradient updates, and (6) sliding-window hypergraph reconfiguration. Requires expertise in differential geometry, causal inference, graph neural networks, and transformer architectures. The paper provides sufficient architectural detail for reimplementation but lacks code.

The paper provides detailed hyperparameters (2-layer transformer, 64 hidden units, 4 heads, lambda=10, Adam lr=1e-4, batch size 32), specifies the data sources (Yahoo Finance, Twitter Financial News Sentiment dataset on HuggingFace), and describes the implementation in PyTorch on NVIDIA RTX 2080 Ti. However, no GitHub repository or code link is provided. The Granger causality construction procedure and FDR-adjusted p<0.01 threshold are specified. Reproducibility is moderate without code release.