Deep Hedging Under Realistic Market Frictions: A Regime-Conditional Empirical Study of Dynamic Option Hedging on Bitcoin Options

Battle Count: 87

Directly addresses a core quantitative trading problem: dynamic option hedging under realistic transaction costs. The paper provides a fair empirical comparison between deep learning-based hedging and classical cost-aware methods on real market data, with actionable findings about when and why classical no-trade band strategies outperform neural policies. Highly relevant for derivatives desks, crypto options market makers, and quantitative risk managers evaluating whether to adopt deep hedging approaches.

The neural network architectures are relatively simple (32-unit LSTM, 2-layer 32-unit feedforward with 4-step lookback). Training uses standard Adam optimizer with CVaR loss. The classical benchmarks (BS delta, Leland, Whalley-Wilmott) have closed-form solutions. The main complexity lies in the data pipeline: processing tick-level options data, applying multiple cleaning filters, constructing episodic hedging simulations, and implementing block bootstrap statistical testing. The methodology is well-specified but requires careful implementation of the hedging simulation loop and regime classification.

The paper uses freely available Tardis.dev historical data (first-of-month snapshots), provides detailed methodology including exact equations, hyperparameters (learning rate 1e-3, weight decay 1e-5, batch size 256, patience 10), architecture specifications (32 hidden units, 2 hidden layers, 4-step lookback), and statistical testing procedures (5000 block bootstrap resamples). However, no code repository is explicitly linked, and the specific data processing pipeline details (cleaning filters, feature engineering) would need to be reconstructed. The episodic structure and regime classification methodology are well-described.