ExplainabilityTransparencyAccountabilityq-fin.CP
Shapley-based Structural Analysis of Neural Calibration for Stochastic Volatility Models
Shaïn Afzali, Serena Della Corte, Antonis Papapantoleon
October 2, 2026
Key Takeaways
-
Neural networks learn interpretable calibration patterns: The study found that across different neural network architectures (multilayer perceptron, highway, and softmax-parametrized highway networks), the models consistently relied on short maturities and smile wings of the implied volatility surface to infer financial parameters, suggesting the networks discover economically meaningful relationships rather than arbitrary patterns.
-
Explainability methods reveal redundancy in financial data: Using Shapley-based explanation techniques (SHAP and νSHAP), researchers discovered "substantial redundancy in the calibration input," meaning neural networks don't need all available data to make accurate predictions—a finding with implications for data efficiency and computational costs in financial modeling.
-
Different explanation methods capture complementary insights: The paper demonstrates that SHAP (measuring sensitivity) and νSHAP (measuring sufficiency) reveal different aspects of how neural networks use financial data, highlighting that a single explanation method may not fully capture how AI systems make decisions in complex domains.
-
Dimensionality reduction is possible without sacrificing accuracy: Building on their explainability analysis, the researchers showed that input data could be significantly reduced for the rough Heston model while maintaining calibration accuracy, suggesting that explanation-guided feature selection could improve AI system efficiency in financial applications.
-
Architecture differences don't fundamentally change learned logic: Despite variations in predictive accuracy and model complexity, the underlying structure of how different neural network architectures infer parameters remained qualitatively stable, suggesting the core learned relationships are robust across design choices.
Abstract
Neural network-based approaches have emerged as efficient alternatives to traditional optimization-based procedures for the calibration of stochastic volatility models. However, existing work has focused primarily on predictive accuracy, with comparatively little attention devoted to understanding the structure of the learned inverse calibration mappings. In this work, we analyze neural calibration mappings for the Heston and rough Heston models across multilayer perceptron, highway, and softmax-parametrized highway architectures, using complementary Shapley-based methods from explainable AI. Specifically, we consider SHAP and $ν$SHAP explanations, which capture distinct, complementary notions of feature relevance, corresponding to sensitivity and sufficiency of feature subsets, respectively. Short maturities and smile wings consistently dominate parameter inference, and the dominant attribution structure remains qualitatively stable across architectures despite differences in predictive accuracy and parameter count. Parameter-specific differences between SHAP and $ν$SHAP further reveal how distinct regions of the implied volatility surface contribute to parameter recovery and expose substantial redundancy in the calibration input. Building on this redundancy, we show that $ν$SHAP explanations can guide a significant reduction in input dimensionality for the rough Heston model while matching calibration accuracy relative to the full implied volatility surface. These findings demonstrate that complementary Shapley-based methods provide structural insight into learned inverse calibration mappings beyond predictive error metrics, and offer a practical route to feature selection in neural calibration problems.
Paper Info
Responsible AI Topics
ExplainabilityTransparencyAccountability
Links
Citation data from Semantic Scholar