Learn to explain the smile: An interpretable hybrid machine learning model to understand the implied volatility of CSI 300 options

We propose an interpretable hybrid machine learning framework for forecasting and explaining implied volatility surface dynamics of CSI 300 index options. Our methodology leverages machine learning to correct a theory-based baseline model. Initial predictions are derived from an analytical model, while the second stage involves a machine learning model trained on the residuals of the first stage. We construct three variants of hybrid models using XGBoost: a baseline three-feature model, a VIX-augmented four-feature model, and a five-feature model incorporating a newly developed options-implied ambiguity index. Empirical results using 2019–2025 CSI 300 options data show that the five-feature model significantly outperforms both the analytical benchmark and VIX-only model. Performance improvements are especially pronounced in market rallies and high-ambiguity regimes, where ambiguity attenuates implied volatility compression and amplifies perceptions of downside risk. We further use SHAP value analysis to demonstrate that feature effects are economically coherent and state-dependent. Our findings confirm that ambiguity is a distinct and quantitatively meaningful risk factor for explaining implied volatility dynamics in emerging market.