Introduction

Deep hedging—using deep learning to determine optimal hedging strategies for derivatives—generates optimal decisions but offers limited insight into reasoning. Explainability techniques reveal what features drive hedging decisions, enabling validation and improvement of learned strategies.

Feature Attribution for Deep Hedging

Use attention mechanisms: deep networks with attention layers highlight which features (market conditions, option parameters) are most relevant to hedging decisions. High attention to spot price and implied volatility is expected; unexpected high attention to other features may indicate overfitting or dataset artifacts. Visualize attention patterns to validate model logic.

Counterfactual Analysis

For a specific hedging decision (sell 10 put options), generate counterfactuals: if spot price were 2% higher, how would hedging change? If volatility 5% lower? Understand hedging sensitivity to market variables. Validate that sensitivities match theoretical expectations (Greeks).

Conclusion

Explainability techniques applied to deep hedging enable validation and improvement of learned hedging strategies.