SHAPLEY-BASED STRUCTURAL ANALYSIS OF NEURAL CALIBRATION FOR STOCHASTIC VOLATILITY MODELS

Battle Count: 50

Highly relevant for quantitative researchers and practitioners involved in volatility modeling and option pricing. The findings offer a practical method to reduce computational costs in calibration by identifying redundant input features, directly impacting the efficiency of trading systems that rely on real-time volatility surface calibration.

Implementing the neural calibration pipeline is moderately complex due to the need for accurate option pricing engines (COS method, fractional Adams method) and specific preprocessing (ZCA whitening). The XAI analysis (SHAP/nuSHAP) adds another layer of complexity, particularly the sample-based approximation for nuSHAP.

The paper provides detailed experimental setups, including parameter ranges, grid discretizations, network architectures, and preprocessing steps (ZCA whitening). However, the code repository is not explicitly linked in the text, and the synthetic data generation relies on specific numerical methods (COS, fractional Adams) which require careful implementation.