Harvesting the Volatility Risk Premium: A Learning-to-Rank Approach

Battle Count: 76

Highly relevant to quantitative trading. The paper presents a complete, deployable trading system for 0DTE SPXW options with explicit margin modeling, fee schedules, execution assumptions, and risk controls. It addresses practical implementation costs (margin, fees, bid-ask spreads) that determine whether published Sharpe ratios are realizable. The learning-to-rank formulation for cross-sectional strategy selection is directly applicable to systematic trading. The confidence gate and regime-conditional features provide actionable risk management. The walk-forward + out-of-time evaluation framework sets a rigorous standard for strategy validation. The seven sizing methods provide a comprehensive comparison of position-sizing approaches under realistic constraints.

High implementation complexity. The system requires: (1) minute-frequency CBOE options data with bid-ask quotes, (2) a ~190-feature engineering pipeline with cross-sectional and per-strategy features, (3) a five-stage per-window feature selection algorithm, (4) LightGBM LambdaRank training with Optuna hyperparameter search (50 trials per window), (5) confidence gate calibration on held-out slices, (6) seven distinct sizing methods with per-window parameter calibration, (7) Reg-T margin modeling, (8) tiered fee schedule implementation, (9) four-window expanding walk-forward with annual retraining, (10) path-aware Sortino-on-bars label construction at one-minute resolution, and (11) multiple statistical tests (PSR, DSR, Diebold-Mariano). The interplay between components (gate calibration sensitivity to feature clustering, sizing calibration inheriting in-sample optimism) adds further complexity.

The paper provides extensive methodological detail including hyperparameter search spaces, feature definitions (Appendix B), algorithm pseudocode (Algorithms 1 and 2), and complete evaluation tables. However, no code repository is mentioned. The data (CBOE minute-frequency SPXW options data) is proprietary and not publicly available. The feature engineering pipeline with ~190 features and the specific calibration procedures are well-documented but complex to replicate without access to the same data sources. The walk-forward schedule, sizing methods, and ablation studies are fully specified.