Introduction

Multiple explainability methods (SHAP, LIME, permutation importance) exist; which works best for volatility forecasting models? Benchmarking systematically compares explanation methods across fidelity (explanation accuracy), stability (consistency), and computational cost, guiding method selection.

Benchmark Methodology

Train volatility forecasting model (ensemble of trees, neural network). Generate predictions. Compute explanations using multiple methods. Measure fidelity: do explanations accurately represent feature contributions? Stability: do explanations change substantially with small input changes? Cost: computational time and resources.

Results and Recommendations

For tree-based volatility models: SHAP and permutation importance are most stable and fast. For neural networks: SHAP requires more computation but higher fidelity than LIME. Hybrid approach: use fast explanations (permutation importance) for online monitoring, detailed explanations (SHAP) for post-trade analysis.

Conclusion

Systematic benchmarking informs optimal selection of explainability methods for specific applications.