This paper presents a comprehensive interpretability analysis of deep learning methods applied to rough volatility models, focusing on the rough Heston model. The authors train a feedforward neural network to learn the inverse mapping from implied volatility surfaces (across strikes and maturities) to the model parameters (ρ, V0, κ, θ, ν, H). They generate a synthetic dataset of implied volatilities using Fourier transform pricing methods and then apply several interpretability techniques—including LIME, Gradient*Input, DeepLIFT, LRP, and SHAP—to understand how the neural network makes its predictions.

The key contribution is demonstrating that the neural network's learned mapping is not a black box but can be interpreted in terms of financial intuition. For instance, SHAP analysis shows that the Hurst exponent (H) is the most critical parameter for explaining the at-the-money short-term skew, which aligns with known properties of rough volatility models. The study also finds that other parameters have interpretable effects on different regions of the volatility surface. By providing a framework for interpretability, the paper enhances trust and safety in using neural networks for pricing and calibration in quantitative finance, paving the way for more transparent AI-driven financial models.