What Does Deep Hedging Actually Learn? Delta Corrections, Regime Fragility, and Symbolic Distillation

Battle Count: 79

Highly relevant to quantitative trading, particularly for options desks and risk management teams. The paper provides actionable insights into what a deep RL hedging agent actually learns (a systematic delta haircut), when it fails (regime fragility in 2022 and 2023), and how to make it auditable via symbolic distillation. The walk-forward empirical design on real S&P 500 options data, the two-stage bootstrap inference, and the comparison against a daily-updated Black-Scholes benchmark make the findings directly applicable to live hedging operations. The symbolic formulas can be traded out of sample, bridging the gap between black-box RL and transparent hedge rules.

Moderate to high complexity. Requires: (1) TD3 implementation with actor-critic architecture, experience replay, twin critics, delayed updates, and target networks; (2) careful walk-forward training/validation/testing split respecting calendar time; (3) episode construction from raw option chains with cleaning and filtering; (4) two-stage bootstrap inference accounting for both time-series and cross-sectional dependence; (5) symbolic regression with multiple formula families, kernel smoothing, and parsimonious selection rules; (6) Black-Scholes benchmark recomputation at each rebalancing date. The code is publicly available, reducing implementation burden, but understanding the economic interpretation and regime diagnostics requires domain expertise.

Code and exact implementation details are publicly available on GitHub. The paper follows the empirical setup of Mikkilä and Kanniainen (2023) closely. Walk-forward design, state vector, network architecture, training procedure, and bootstrap inference are all described in detail. However, OptionsDX data is commercial and not freely available. Three random-seed runs are reported for robustness. The symbolic regression selection rule and distillation procedure are fully specified.