<p>Crude oil prices play a critical role in the real economy and environmental sustainability. Given the complexity of forecasting crude oil prices, this paper proposes a novel model that integrates multisource predictors, factor screening, and forecast combination. First, the multisource predictors encompass factors from eight distinct categories. Second, the predictors are refined sequentially using the mutual information method followed by six feature selection techniques. Third, the selected predictors are paired with powerful machine learning models to forecast crude oil prices. These individual forecasts are then aggregated using the mean combination method. This study conducts a comprehensive empirical analysis employing the sliding window technique, using data from February 1986 to March 2023. The results demonstrate that the proposed model considerably outperforms alternative models, underscoring the effectiveness of combining the above three components. Furthermore, the out-of-sample forecasts and factor screening results are examined across different sub-periods. Finally, a series of robustness checks confirms the strong predictive performance of the proposed model.</p>