@

초록

We forecast monthly crude oil volatility dynamics using an interpretable machine learning framework applied to a long sample period from 1986 to 2023, encompassing multiple economic cycles, including four NBER-defined recessions, and incorporating 140 macroeconomic and uncertainty variables. Tree-based models (Random Forest and XGBoost) deliver the strongest out-of-sample forecasting performance, though their relative advantage varies across economic regimes. Uncertainty indicators, rather than oil-specific fundamentals, are the most influential predictors, indicating that oil volatility is primarily shaped by broader uncertainty conditions. Time-varying interpretability analyses reveal that the superior performance of tree-based models stems from their ability to flexibly combine multiple signals, an insight not captured by static importance scores. Our results highlight the value of dynamic interpretability in understanding regime-dependent machine learning behavior and offer guidance for volatility modeling.

키워드