TAIL-SAFEHEDGING: EXPLAINABLE RISK-SENSITIVE REINFORCEMENT LEARNING WITH A WHITE-BOX CBF–QP SAFETY LAYER IN ARBITRAGE-FREE MARKETS

Battle Count: 50

Directly addresses derivatives hedging (SPX-VIX), tail-risk management, safe RL with hard constraints, market microstructure-aware execution, and explainable AI for institutional trading. The framework is designed for deployability in regulated financial environments with model risk governance. Highly relevant to quantitative risk management, algorithmic hedging, and safe AI deployment in trading. The only limitation for direct trading application is the synthetic-only evaluation.

Combines multiple advanced components: IQN distributional critic, CVaR optimization, PPO with KL/entropy regularization, PID-controlled temperature-tilted quantile sampling, discrete-time CBF constraints, ellipsoidal NTB, box/rate limits, sign-consistency gates, convex QP solving (OSQP), SSVI/Dupire/VIX market generation, ABIDES/MockLOB execution simulation, and a full telemetry/governance pipeline. Requires expertise in RL, convex optimization, financial mathematics, control theory, and software engineering. The theoretical guarantees add further complexity for verification.

The paper provides detailed implementation notes (Appendix B), hyperparameter ranges, statistical protocols (paired bootstrap, FDR adjustment), and a reproducibility checklist. However, all experiments are conducted in synthetic environments (SSVI→Dupire→VIX with ABIDES/MockLOB), no real-data replay is performed, and no GitHub repository or code is explicitly provided. The synthetic nature limits external validation. The paper acknowledges this as deliberate to isolate methodological contributions.