Ensemble methods have received a great deal of attention in the past years in several disciplines. One reason for their popularity is their ability to model complex relationships in large volumes of data, providing performance improvements compared to traditional methods. In this article, we implement and assess ensemble methods’ performance on a critical predictive modeling problem in marketing: predicting cross-buying behavior. The best performing model, a random forest, manages to identify 73.3 % of the cross-buyers in the holdout data while maintaining an accuracy of 72.5 %. Despite its superior performance, researchers and practitioners frequently mention the difficulty in interpreting a random forest model’s results as a substantial barrier to its implementation. We address this problem by demonstrating the usage of interpretability methods to: (i) outline the most influential variables in the model; (ii) investigate the average size and direction of their marginal effects; (iii) investigate the heterogeneity of their marginal effects; and (iv) understand predictions for individual customers. This approach enables researchers and practitioners to leverage the superior performance of ensemble methods to support data-driven decisions without sacrificing the interpretability of their results.