We compare the performance of intrinsically explainable methods (decision trees and explainable boosting machines) with a black-box method (neural networks) with post-hoc explanations provided by SHAP, LIME and association rules, for the task of early prediction of student performance in a large-scale programming course. All three prediction methods obtained high accuracy, and all explainers were able to provide actionable insights by highlighting important features and key tasks students should complete. Decision trees provided compact rules, which were complemented by the more detailed behavioral patterns identified by the association rules. We observed high similarities between the feature ranking methods. We discuss the strengths of the methods in identifying key factors affecting student performance and providing explanations.