Simple Explanation (ELI5)

Think of it like riding a bike: classic methods follow a simple rule ('look ahead and keep your balance'), while neural networks are like a beginner learning from scratch. With little data (little practice), the beginner gets confused and wobbles too often. Indeed, the classic Whalley-Wilmott method, which knows when to 'stay still', beats neural networks that keep pedaling nervously.

Key Advantages & Current Limits

🟢 Pro: Whalley-Wilmott halves transactions (8× fewer rebalances) saving $1.79 per episode, ideal for volatile assets like Bitcoin. 🔴 Con: Deep models suffer from overfitting on small datasets and lack mechanisms to avoid useless trading, ending up more expensive than expected.

Future Outlook & Takeaways

In the next 3-5 years, we'll likely see hybrid models combining the 'stay-still bands' of classical methods with deep learning flexibility, and reinforcement learning approaches with structured rewards penalizing excessive trading, reducing costs for those hedging crypto positions.

Dynamic Hedging on Bitcoin: Do Neural Networks Beat Classic Methods?

Article

100%

Bitcoin option hedging is complex: classic methods like Black-Scholes assume free, continuous rebalancing, impossible in real markets. Deep hedging trains neural networks to handle these frictions, but so far relied on simulated data. This study uses real Deribit data 2020-2024. It compares Black-Scholes, Leland, and Whalley-Wilmott with LSTM and feedforward networks. All face 5 bps transaction cost. Across 11,546 episodes, Whalley-Wilmott saves $1.79/episode vs Black-Scholes, rebalancing 8x less. Deep models don't beat classics on any metric. Even with trading penalties, they keep trading hourly. Whalley-Wilmott's edge depends on market regime. Likely causes: small dataset and no built-in 'stay still' mechanisms.

Bonus

100%

The study analyzes 11,546 Bitcoin option hedging episodes from Deribit (Sep 2023-Dec 2024). Deep models include LSTM and feedforward networks with CVaR loss and trading penalties. Whalley-Wilmott uses no-trade bands to reduce transactions. Black-Scholes delta rebalances hourly. Leland adjusts for costs. All subject to 5 bps transaction cost. Whalley-Wilmott achieves better P&L and lower tail-risk, significant only partially. Deep models show overfitting: trained on 5 years data, but validation on calm period reduces edge. Tested architectures lack explicit mechanisms to inhibit trading. Proposed solution: integrate inactivity bands or early-stopping into models. Future research could test reinforcement learning with rewards structured to address cost frictions.