In this paper we investigate oil market volatility prediction using a comprehensive data set of 205 variables spanning macroeconomic, financial, energy-related and sentiment indicators. We employ machine learning techniques for variable selection and dimension reduction, combining hard thresholding, soft thresholding and principal component analysis, evaluated through an out-of-sample time-series backtesting framework. The empirical findings reveal that financial variables dominate short-horizon forecasting, while macroeconomic and sentiment factors become progressively more important at longer horizons. A hybrid approach combining principal component analysis with preliminary variable filtering improves forecast accuracy, particularly for medium-term predictions. Support vector regression and random forest methods demonstrate strong performance when paired with appropriate feature selection techniques, suggesting the importance of capturing nonlinear relationships while maintaining robustness to outliers. These results indicate that effective oil volatility forecasting requires careful consideration of both the forecast horizon and the interaction between feature selection and machine learning methods.