Predicting student performance and dropout risk is a critical challenge for higher education institutions aiming to improve academic success and retention. This study applies supervised machine learning techniques to over 20 years of historical data from the Universidad Adventista de Centroamérica (UNADECA), combining academic records and demographic features to forecast two key outcomes: academic performance (pass/fail) and dropout status (enrolled/discontinued). Five classical algorithms—Logistic Regression, Decision Tree, Random Forest, Naive Bayes, and Support Vector Machine—were evaluated, along with advanced models including Gradient Boosted Trees and Artificial Neural Networks. Random Forest and GBT achieved the highest performance, with F1-scores exceeding 0.96 for academic performance and 0.86 for dropout prediction. Importantly, the study demonstrates that using only first-year academic data yields comparably high predictive accuracy, validating the feasibility of early intervention systems. Feature importance analyses also revealed consistent alignment between high-risk predictors and historically challenging courses, offering actionable insights for targeted academic support. While the algorithms used are not novel, the study contributes applied value by testing standard methods on an underexplored, longitudinal dataset from a Central American university context. Ethical considerations, including the fair use of demographic variables and interpretability, are addressed to support responsible implementation. The results provide a replicable foundation for institutions seeking to enhance retention efforts through data-informed decision-making.