In current supply chain operations, the transactions among suppliers and original equipment manufacturers (OEMs) are sometimes inefficient and unreliable due to limited information exchange and lack of knowledge about the supplier capabilities. For the OEMs, majority of downstream operations are sequential, requiring the availabilities of all the parts on time to ensure successful executions of production schedules. Therefore, accurate prediction of the delivery times of purchase orders (POs) is critical to satisfying these requirements. However, such prediction is challenging due to the suppliers’ distributed locations, time-varying capabilities and capacities, and unexpected changes in raw materials procurements. We address some of these challenges by developing supervised machine learning models in the form of Random Forests and Quantile Regression Forests that are trained on historical PO transactional data. Further, given the fact that many predictors are categorical variables, we apply a dimension reduction method to identify the most influential category levels. Results on real-world OEM data show effective performance with substantially lower prediction errors than supplier-provided delivery time estimates.