SHAP and νSHAP analysis of neural calibration mappings for Heston and rough Heston finds short maturities and smile wings dominate parameter inference, enabling rough Heston input reduction at matched calibration accuracy

Short maturities and the wings of the implied volatility smile consistently dominate parameter inference for both Heston and rough Heston, and this dominant attribution structure remains qualitatively stable across network architectures. Prior neural calibration work focused mainly on predictive accuracy; this work instead characterizes the internal structure of the learned inverse calibration mapping, bringing an explainable-AI attribution lens to stochastic volatility calibration. Based on two complementary attribution methods, SHAP and νSHAP, compared across multilayer perceptron, highway, and softmax-parametrized highway architectures for Heston and rough Heston; the abstract reports that the dominant structure stays qualitatively stable despite differences in predictive accuracy and parameter count.

SHAP and νSHAP differ at the parameter level, indicating that distinct regions of the implied volatility surface contribute differently to parameter recovery and exposing substantial redundancy in the calibration input. By separating feature relevance into the complementary notions of sensitivity and sufficiency of feature subsets, the analysis reveals parameter-level correspondences between surface regions and parameter recovery rather than a single importance ranking. The abstract states that SHAP and νSHAP correspond to sensitivity and sufficiency of feature subsets respectively, and that their parameter-specific differences are used to explain how surface regions contribute to parameter recovery and to identify input redundancy.

Exploiting the redundancy revealed by νSHAP explanations allows a significant reduction in input dimensionality for the rough Heston model while matching calibration accuracy relative to the full implied volatility surface. It advances attribution analysis from a diagnostic tool to a practical route for feature selection in neural calibration problems. The abstract reports that νSHAP-guided dimensionality reduction for rough Heston matches the calibration accuracy of the full implied volatility surface.

The results are aimed at researchers and practitioners who use neural networks for stochastic volatility model calibration, in inverse calibration settings that take the implied volatility surface as input and target Heston and rough Heston, covering multilayer perceptron, highway, and softmax-parametrized highway architectures. Their value lies in treating attribution analysis as a structural diagnostic and feature-selection step within the calibration pipeline: the dominance of short maturities and smile wings suggests prioritizing those surface regions, while νSHAP-guided reduction offers rough Heston a path to compress inputs while preserving calibration accuracy.

The currently visible text is the abstract and does not give the magnitude of dimensionality reduction, quantitative comparisons of calibration error, the surface grid and term-structure setup, training data size, or the computational cost of the attribution methods, so it is not possible to judge under which market conditions or surface resolutions the reduction remains robust. The economic or numerical mechanisms behind the parameter-specific differences between SHAP and νSHAP, and whether the dominant attribution structure persists at longer maturities, more extreme smiles, or different noise levels, remain open questions worth watching.