paperthgnntransformergatfisher-zspongebasket-tradingus-equitysp500regime-detection

Fanshawe, Masih & Cameron (2025) — Forecasting Equity Correlations with Hybrid Transformer Graph Neural Network

Authors: Jack Fanshawe, Rumi Masih (corresponding), Alexander Cameron. University of Queensland Business School.

Venue: arXiv preprint 2601.04602, dated 01/07/2025.

Contribution

A Temporal-Heterogeneous GNN (Transformer temporal encoder + edge-aware GAT relational encoder) trained on 37 S&P 500 features to predict 10-day-ahead stock-stock correlation in Fisher-z residual space relative to a 30-day rolling baseline. The predicted correlation matrix feeds SPONGEsym clustering + Korniejczuk Slepaczuk 2024 Graph Clustering Stat Arb's ML-ensemble signal filter for long/short basket trading.

Headline numbers

Correlation prediction (out-of-sample 2019–2024, 269 M edges)

Metric Persistence baseline THGNN
Pearson r 0.310 0.778
Spearman \rho 0.314 0.795
MAE 0.3071 0.2302
RMSE 0.3852 0.2940

Trading (2019-04 → 2024-10, S&P 500)

Metric THGNN forward-looking S&P 500
Sharpe 1.84 0.65
Annualised return 19.20% 14.43%
Max drawdown −9.4% −33.9%
Max loss duration 82 days 512 days

Sharpe difference: one-sided Ledoit-Wolf stationary bootstrap, p = 0.0200.

Architectural lineage

The paper's contribution is the dependence-metric upgrade — swapping rolling-window correlations for ML-predicted forward correlations — rather than any new component in the trading engine.

Importance to this wiki

  • First paper in this wiki to explicitly frame correlation as a predictable state variable rather than a backward-looking summary.
  • Strongest empirical demonstration (0.31 → 0.78 Pearson on 269 M edges) of the value of replacing rolling-window estimators with forward predictions in a downstream trading task.
  • Self-honest about limitations: no slippage, no ablation vs LSTM/CNN, no hyperparameter search, single attribution family, failed replication of Korniejczuk–Slepaczuk ML-filter gains.

See also

Full summary