This study based on data from the National Assessment of Educational Progress (NAEP) in the United States, developed an online mathematics test efficiency prediction model using stacked ensemble learning. The research compared the performance of two machine learning algorithms—Random Forest and Gradient Boosting—and combined them to construct a stacked ensemble model. Through ten-fold cross-validation, the model achieved an average F1 score of 74.92% and a recall rate of 77.45%. Feature importance analysis based on SHAP (SHapley Additive exPlanations) indicated that question type played a dominant role in predicting student test efficiency, while learning behavior features had a relatively small impact. This study innovatively integrated stacked ensemble learning with explainable artificial intelligence techniques, enhancing prediction accuracy while providing reliable support for educational decision-making.