<![CDATA[Early identification of unstable academic development can support timely advising, yet models evaluated only with random splits may not transfer across cohorts. This study examines whether data available after semester-2 grades are finalized can classify two K-Means-derived GPA trajectory labels—stable-improving and fluctuating—observed in semesters 3–7. A retrospective dataset of 1,362 anonymized students was analyzed. Labels were generated from standardized semester 3–7 GPA profiles, while predictors were limited to admission attributes and semester 1–2 records. Train-only preprocessing, RFECV, validation-set threshold tuning, calibration analysis, and GaussianNB, logistic regression, Random Forest, and SVM-RBF were evaluated using a stratified random test and a 2024/2025 administrative-cohort holdout. Random Forest achieved 0.831 balanced accuracy and 0.916 ROC-AUC on the random test. In the holdout, where fluctuating-label prevalence increased from 25.5% to 69.3%, Random Forest had the highest threshold-based balanced accuracy (0.581), whereas SVM-RBF had the highest ROC-AUC (0.763). Drift was substantial in student type, admission semester, status, and study-program composition. In this dataset, random splitting substantially overestimated cross-cohort performance. The observed degradation represents a moderate generalization challenge; these cluster-derived labels require cohort-aware recalibration and educational validation, and predictions should be used only as a human-supervised screening aid.]]>
