Deep Hedging Fails Under Real Market Frictions: Evidence from Bitcoin Options
ORIGINAL / Deep Hedging Under Realistic Market Frictions: A Regime-Conditional Empirical Study of Dynamic Option Hedging on Bitcoin Options
This study compares classical option hedging strategies with deep hedging models using real Bitcoin options data from Deribit (2020-2024). The results show that deep hedging does not outperform classical benchmarks, while Whalley-Wilmott significantly reduces costs, with performance dependent on market conditions.
01 ABSTRACT
Based on five years of Bitcoin options data from Deribit, this study compares six hedging strategies. Over 11,546 test episodes from Sep 2023 to Dec 2024, the Whalley-Wilmott strategy significantly reduced transaction costs by $1.79 per episode (p<0.0001), with about 8 times less trading frequency, but P&L and tail risk improvements were not statistically significant. None of the deep hedging models (LSTM and feedforward) outperformed any classical benchmark, and they traded almost every hour regardless of penalty weight. In a calmer validation period, Whalley-Wilmott's P&L advantage shrunk or disappeared, but cost savings remained. The authors speculate that poor performance of deep models may be due to small training sets and lack of a mechanism to stay still.
02 KEY FINDINGS
- Whalley-Wilmott significantly reduces transaction costs by $1.79 per episode and trades about 8 times less often.
- Deep hedging models (LSTM and feedforward) fail to outperform classical strategies on real data.
- All deep models trade almost every hour; penalty weights have minimal effect.
- Whalley-Wilmott's P&L advantage diminishes in a calmer market, but cost saving persists.
- Authors suspect deep models' failure is due to small training set and lack of a stand-still mechanism.
AI GENERATED SUMMARY / DISCOVERED BY ARXIV Q-FIN