Robust Hedging Valuation Adjustment for Deep Hedging Policies under Market Frictions

Battle Count: 73

Highly relevant to quantitative trading. The paper directly addresses the practical question of whether a trading desk can afford to run a deep hedging policy by computing a single internally consistent reserve (HVA + FVA + MVA) under robust stress. It provides a framework for comparing classical and learned hedging policies under transaction costs, liquidity frictions, and tracking-risk budgets. The results have direct implications for hedging policy selection, trading cost management, and risk budgeting in derivatives desks.

Moderate to high complexity. Requires: (1) Merton jump-diffusion simulation with Markov liquidity chain, (2) Neural network training with CVaR and cost penalties via backpropagation through trading paths, (3) Robust optimization via KL divergence dual formulation with exponential tilting, (4) Funding and margin add-on computation with pathwise allocation, (5) Validation/test split selection protocol, (6) Multiple candidate policy families (classical bands, gamma-wide, learned NTB, sparse execution). The KL reweighting and common-stress aggregation add mathematical sophistication. However, the neural network architecture is relatively simple (3-layer feedforward) and the simulation framework is standard.

The paper provides detailed parameter specifications (Table 1, 2), candidate grids for band widths and thresholds, network architecture details (3 hidden layers, 64 SiLU units, AdamW optimizer settings), training hyperparameters (250 epochs, 16384 paths/epoch, 64 trading dates), and evaluation protocol (5 validation blocks, 20 test blocks, 5000 paths each). KL radii are specified as external inputs. However, no code repository is provided, and the specific random seeds for training are only partially disclosed. The Merton jump-diffusion simulation and Markov liquidity chain are fully specified.