Uncertainty-Aware Deep Hedging

Battle Count: 104

Directly addresses derivative hedging under realistic market frictions (transaction costs, stochastic volatility). The uncertainty-aware blending framework provides a principled method for deploying neural hedging strategies in production risk management systems. The finding that ensemble advantage comes from transaction cost savings rather than price prediction is highly relevant for trading system design. The regime-dependent inversion of uncertainty signals has direct implications for when to trust ML-based hedges. The CVaR improvements of 35-80 bps over Black-Scholes delta are economically meaningful for institutional hedging desks.

Uses the existing pfhedge library for the core deep hedging framework, reducing implementation burden. However, requires training 5 independent LSTM ensembles, computing per-time-step uncertainty, and optimising blending parameters via differentiable P&L computation. The architecture is relatively simple (2-layer LSTM, 32 units), but the full pipeline (simulation, training, ensemble inference, blending optimisation, bootstrap evaluation) involves multiple stages. Computational cost scales with ensemble size and number of MC paths (20,000 per epoch, 500 epochs, 5 members).

The paper uses the open-source pfhedge library (Imaki et al., 2023) for implementation. Detailed hyperparameters are provided (5 ensemble members, 2-layer LSTM with 32 hidden units, 500 epochs, 20,000 MC paths per epoch, Adam optimizer, 5 basis point transaction costs). Three Heston calibrations are fully specified. Paired bootstrap with 5,000 resamples and three random seeds are used for statistical validation. However, no explicit GitHub repository is provided for the paper's specific code, and the exact random seeds for ensemble training are not listed.