Student dropout is a significant issue in higher education, affecting students and academic institutions. This study presents a prototype predictive model to identify students at risk of dropping out at the Universidad de las Fuerzas Armadas ESPE, based exclusively on their academic performance. Using historical data from a specific program as training data, records were classified into three categories: graduated, studying, and dropout. A cleaning and transformation process was applied to the data, resulting in a training dataset suitable for use across university programs. We adopt a strict semester-based temporal validation (train \(\le \) 2021–S1, validate on 2022–S1, and out-of-time test on 2023–S1) to avoid information leakage and better emulate deployment. Data were cleaned and transformed into program-agnostic snapshots; SMOTE was applied only within training splits. We compared multiclass probabilistic classifiers—Random Forest, Neural Networks, and Gradient Boosting—with hyperparameter tuning. The performance of these models was evaluated using metrics such as precision, recall, and F1-score. A probability threshold was also determined to identify and reduce the students with the highest dropout risk. The results demonstrate that using the Random Forest technique, the model achieved an F1 score for the target dropout class of 0.864. The model was tested with accurate data and effectively identified students at risk, enabling the future implementation of an early warning system. The prototype was tested with varying thresholds and consistently identified students at risk. Lower thresholds captured more cases but increased false positives, while higher thresholds reduced detection. These results highlight the need to adapt the threshold to institutional capacity, ensuring effective and reliable early interventions.