Latin American Oil Export Destination Choice: A Machine Learning Approach

We implement machine learning techniques to predict the destination for Latin American crude oil exports. Utilizing a unique dataset of micro-level crude oil shipment data, derived from the Automatic Identification System (AIS) for ship tracking, we investigate the micro- and macro-level determinants of the destination choice. We use decision tree, Random Forests and boosted trees techniques in training a model to predict the export destinations which can help to identify seller/buyer groups with similar oil trade requirements. The results show that while macro data, such as regional oil price differences and crack spreads, impacts the crude oil flow, micro level information about the oil shipment are key attributes in the destination prediction. Our research has practical implications, particularly with regards to prediction of oil transportation demand, spatial price arbitrage and short-term forecasting of regional crack spreads.