A1490
Title: Distributional reinforcement learning for tail risk control in weekly options sell-side strategies
Authors: ChiFang Chao - National Chengchi University (Taiwan) [presenting]
Ming-Hua Hsieh - National Chengchi University (Taiwan)
Mu-En Wu - National Taipei University of Technology (Taiwan)
Abstract: Options sell-side stop-loss is conventionally treated as a prediction problem, where fixed price thresholds trigger position exit. The decision is reframed as one of tail risk control and an IQN-CVaR framework is proposed that learns the full return distribution via Implicit Quantile Networks and selects exit actions by minimizing Conditional Value-at-Risk. A single parameter alpha controls risk sensitivity at inference without retraining. The framework is evaluated on 8,923 out-of-the-money TXO weekly option episodes spanning 2019 to 2025, using six-fold walk-forward validation and a 2025 hold-out test. At alpha = 0.50, CVaR at the 25\% level improves by 41\% relative to a hold-to-stop baseline in walk-forward validation, and by 71\% on a 2025 hold-out test. A standard DQN with the same training budget achieves only 13\% improvement in walk-forward and 2\% in hold-out. The distributional agent exhibits risk-sensitive behavior that adapts to the asymmetric tail risk structure of call versus put positions. Exit timing varies smoothly with alpha, enabling practitioners to calibrate the trade-off between tail risk reduction and premium income retention.
