Robust Hedging Valuation Adjustment for Deep Hedging Policies under Market Frictions
This paper introduces a robust Hedging Valuation Adjustment (HVA) as a post-training layer that aggregates the policy-dependent hedging costs, funding add-ons, and margin add-ons under a common stress distribution, enabling internally consistent reserves for policy ranking . By combining tracking-loss CVaR with explicit liquidity and market-friction adjustments, the methodology compares classical hedge bands and learned neural policies across three liquidity regimes, using a KL-based upper expectation to guard against model misspecification . Empirical results indicate that in High and Middle Liquidity environments, tuned classical gamma-wide bands typically dominate after budget relaxation, while in Low Liquidity sparse learned execution can offer additional reserve reductions under strict tracking-risk budgets . The study emphasizes that a single shared stress path should be used for HVA, funding, and margin rather than three separate benchmarks, ensuring coherent risk–cost assessment and preventing misleading conclusions from policy-specific opposite worst-case scenarios . The framework also points to extensions such as integrating balance-sheet features (netting, collateral, CVA/KVA) and considering retraining or robustness under KL stress, enabling future research on end-to-end xVA-like solutions for deep hedging policies .

