Abstract
The early identification of students at risk of non-progression poses a practical challenge for many universities, particularly for institutions in data-rich contexts where academic data are stored as repeated event records rather than model-ready student profiles.We developed an end-to-end machine learning workflow for academic progression-risk prediction using large-scale institutional grade data from Kazakhstan.A raw log of 732,965 grade-event records was aggregated by a log-toentity procedure to yield a student-level dataset of 20,868 learners represented by ten compact features describing central performance, variability, extrema, and academic exposure.We compared Logistic Regression, Random Forest, and XGBoost, training and testing them under a common evaluation protocol on the finalized labeled dataset.The results show that ensemble-model-based predictions outperformed the linear baseline, with XGBoost giving the strongest overall performance at 94.0% accuracy, 93.1% F1-score, and 97.2% receiver operating characteristic-area under the curve (ROC-AUC) in the baseline evaluation.Logistic Regression provided an interpretable baseline, with Random Forest forming a strong intermediate benchmark.Feature-level analysis indicates that average mark and minimum mark are the most informative predictors for progression-risk appraisal, along with variability-related measures.Beyond simple model comparison, we operationalised inference through a REST-based service that supports single and batch prediction, runtime feature completion, and automated endpoint testing.Because the target is an operational proxy rather than a directly observed progression outcome, the reported performance should be interpreted as evidence for a deployable baseline system rather than as final proof of causal or policy effectiveness.The work, therefore, describes an end-to-end educational analytics system that links academic log aggregation, predictive modeling, interpretation, and deployment in one framework.