Forecasting the realized volatility of oil futures: Machine learning with cross-market spillovers

This paper proposes a framework that captures cross-market risk spillovers and evaluates the predictive power of the resulting factors for forecasting oil price volatility using a range of machine learning models. Using 15-minute high-frequency trading data from oil futures and related financial assets, we measure spillover risk across continuous, jump, and mixed channels via a decomposition of realized covariance. We further demonstrate that machine learning models that integrate cross-market risk factors can substantially improve forecasting performance over the benchmark HAR model in all horizons. The relative performance of different machine learning models is also examined. In addition, combination models such as the Discounted Mean Squared Prediction Error (DMSPE) approach can deliver superior and robust forecasting performance over individual models, both from statistical and economic perspectives. Overall, we extend the literature by incorporating granular cross-market spillover signals into machine learning volatility forecasting models, yielding a better understanding and more accurate forecasts of oil price volatility. • Factors arising from cross-market risk spillovers can be used to forecast oil price volatility. • Spillover risk across continuous, jump, and mixed channels are measured via a decomposition of realized covariance. • Machine learning models that integrate cross-market risk factors can substantially improve forecasting performance. • Combination models can deliver superior forecasting performance over individual models.