Tabular Deep Learning for Algorithmic Trading: Cross-Regime Bayesian Optimisation for Equity Signal Generation

Battle Count: 97

Highly relevant to quantitative trading. The paper directly addresses algorithmic trading signal generation, portfolio construction, and performance evaluation. Key contributions include: (1) a cross-regime Bayesian optimisation framework for hyperparameter selection targeting trading performance rather than classification accuracy, (2) demonstration that ensemble combination (XGBoost + TabNet via rank aggregation) produces statistically significant CAPM alpha (0.423, p=0.011) with near-zero beta, (3) robustness analysis showing signal stability across four distinct quarterly regimes, (4) practical deployment considerations including position sizing sensitivity, transaction costs, and retraining cadence. The finding that classification metrics are poor indicators of trading performance and that portfolio-level evaluation is necessary is directly actionable for practitioners.

Moderate to high complexity. Requires: (1) Bloomberg Terminal access for sentiment and fundamental data, (2) FRED API and Google Trends data integration with chained normalisation, (3) feature engineering across 50 features in four categories with look-ahead bias prevention (lag schedules), (4) Optuna-based Bayesian optimisation with custom composite objective function across three validation folds, (5) expanding-window cross-validation with fold purging, (6) daily portfolio backtesting with transaction costs, (7) rank aggregation ensemble construction, (8) SHAP analysis with KernelExplainer, (9) non-parametric statistical testing suite. The interactive application (React + REST API + AI assistant) adds further complexity. TabNet tuning takes ~19.6 hours vs XGBoost's ~11 minutes.

The paper uses Bloomberg Terminal data requiring an institutional licence, limiting independent replication. FRED and Google Trends data are publicly available. The interactive application is deployed at a URL but the full pipeline code is not explicitly provided as a repository. Fixed random seeds are used for reproducibility within the study. The 30-trial Optuna search budget and specific hyperparameter ranges are documented. However, the reliance on licensed Bloomberg data for sentiment and fundamentals is a significant barrier to full reproduction.