Predicting Cohort-Level Qualification Non-Completion Across All NZQCF Levels: A Temporal Evaluation of Traditional and Ensemble Machine Learning Models Using 17 Years of National Data

Student dropout, where learners leave study without completing a qualification, carries personal, social and economic costs. Machine learning is increasingly used to predict dropout, but most studies rely on individual records from a single institution and evaluate models with random cross-validation, which can give an overly optimistic picture. This study examines how well traditional and ensemble machine learning models identify cohorts with high non-completion, a cohort-level proxy for dropout, across all ten levels of the New Zealand Qualifications and Credentials Framework (NZQCF). Using national Ministry of Education data for 2008–2024, we classified 680 cohorts, defined by year, qualification group, gender and student type, using a same-year non-completion ratio. Logistic regression, decision tree, random forest, gradient boosting and a heterogeneous stacking ensemble were compared with simple historical baselines, training on 2008–2019 and testing on 2020–2024. The stacking ensemble achieved a ROC-AUC of 0.899, significantly higher than logistic regression (0.833) but not significantly different from random forest (0.905) or each cohort’s historical mean (0.891). Random cross-validation overstated performance (0.958), and performance was lower within qualification groups (0.788). The analysis is limited to cohort-level data, and the ratio overstates non-completion in longer programmes. Future work should test these models on individual-level linked records using the same temporal evaluation.