In this study, we examine the role of jumps in forecasting crude oil volatility by proposing a hybrid framework that integrates machine learning–based jump detection with traditional econometric modeling. Jumps are detected using two alternative approaches. First, the Isolation Forest algorithm is applied directly to daily price data, from which upside and downside jumps are estimated. Second, the same algorithm is employed on realized volatility series constructed from high-frequency data, allowing for the decomposition of volatility into asymmetric jump components. These jump measures are then incorporated into an extended HAR-RV model, and the predictive performance of different specifications, machine learning-based jump detection, non-parametric jump detection, and the standard HAR-RV model without jumps, is systematically compared across short-, medium-, and long-term horizons. The findings indicate that including jumps improves volatility forecasts, with the machine learning-based framework demonstrating predictive value in long-horizon forecasting which is particularly important in contexts where high-frequency data are unavailable, thereby offering a robust and computationally efficient alternative for practitioners and researchers.

JEL classifications: C14; C53.