Reproduce “Tabular Deep Learning for Algorithmic Trading: Cross-Regime Bayesian Optimisation for Equity Signal Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Joshua Le Grice
A cross-regime Bayesian optimisation framework is used to evaluate tabular models for daily cross-sectional equity ranking under changing market conditions. On a held-out 2025 period, an XGBoost–TabNet rank-aggregation ensemble achieved a 51.26% return, 2.44 Sharpe ratio, and −7.76% maximum drawdown, outperforming its individual components and the other tested models, while also showing stable long–short precision across quarters.
Key insights
- Portfolio-level performance separates the models more clearly than accuracy or F1: XGBoost is the strongest standalone model, while the ensemble benefits from TabNet’s lower prediction correlation rather than its standalone returns.
- Cross-regime validation favours configurations that avoid severe failure in any one market environment, and the resulting ensemble maintains above-baseline long and short precision across all four quarters of 2025.
- Technical features dominate attribution, whereas alternative data contributes more to short-side predictions and varies substantially by model; attribution alone does not establish that these features cause the trading signal.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.