Risk-Sensitive Option Market Making with Arbitrage-Free eSSVI Surfaces: A Constrained RL and Stochastic Control Bridge

Battle Count: 51

Directly addresses option market making, one of the core activities in quantitative trading. Integrates pricing consistency (arbitrage-free IV surfaces), execution control (intensity-based fills, spread management), hedging (delta-hedging intensity), and risk management (CVaR tail shaping) into a unified RL framework. The eSSVI surface embedding ensures the agent's quotes are internally consistent, which is critical for any derivatives market maker. The white-box interpretability of five control heads with analytic sensitivities makes it directly deployable in practice. The constrained MDP formulation with learnable dual multipliers provides a principled framework for balancing revenue and risk. Highly relevant for derivatives desks, market-making firms, and quantitative risk management.

Very high complexity due to: (1) differentiable eSSVI layer with butterfly/calendar surrogate penalties requiring careful numerical implementation; (2) intensity-based execution model with Poisson sampling and logistic response functions; (3) CVaR estimation via Rockafellar-Uryasev program with mixed pathwise/likelihood-ratio gradients; (4) two-stage training (supervised warm-start + PPO with structural annealing); (5) state-dependent dual action as learnable Lagrange multiplier; (6) five-dimensional continuous action space with squashing maps; (7) Heston stochastic volatility simulation; (8) multiple theoretical components (wing-growth bounds, grid-consistency proofs, policy-gradient validity). Requires expertise in both quantitative finance (option pricing, IV surfaces, Greeks) and deep RL (PPO, GAE, constrained optimization). The paper provides pseudocode and hyperparameters but the full pipeline is non-trivial to implement correctly.

The paper releases experiment logs, figures, and scripts for the full pipeline (artifacts/run_log.csv, artifacts/step_log.csv, artifacts/artifacts.npz, settings.json). All figures are automatically generated from artifacts. Hyperparameters are fully specified in Appendix B. However, results are simulation-only (Heston fallback), not live-data backtests. The ABIDES-ready framework is described but not enabled in reported runs. Seed is fixed (seed=0) with 8 episodes. Theoretical proofs are provided in full in appendices.