Predicting Early Dropout in University Entry Cohorts: Temporal Inter-Cohort Validation in a Chilean Public University

Early university dropout remains a major concern for higher education, particularly amid expanded access, diverse student profiles, and institutional efforts to promote persistence. This study evaluates predictive models of early dropout at a Chilean university using ten entry cohorts from 2014 to 2023 and an analytical sample of 22,485 students. Only enrollment-time information was used, and dropout was defined as leaving before the second academic year. The strategy combined within-cohort harmonization of admission scores, leave-one-cohort-out temporal validation, and comparison of ridge logistic regression, XGBoost, and a mixed logistic model with cohort-level random intercepts. XGBoost showed the best overall performance in Brier score, log-loss, area under the receiver operating characteristic curve (ROC-AUC), area under the precision–recall curve (PR-AUC), and risk concentration in the top 10%, although differences with ridge regression were small. Entry data provide a useful, moderate signal for early-warning systems and student-support strategies, while preserving methodological transparency and feasibility for university management.