Predictive analytics has the potential to improve student aid, tailor interventions, and enhance institutional outcomes, but remains a persisting issue in educational analytics due to the accurate forecasting challenge. Educational data is deeply rooted in behavioral, psychological, and demographic characteristics which make predictive analytics lack accuracy due to the data’s complexity and multidimensionality. This study seeks to solve this problem by building a framework of a predictive model that leverages Machine Learning (ML) and utilizes Deep Learning (DL) techniques in a sequential organization to ensure accuracy and reliability respectively. More specifically the methodology consists of data collection and cleaning followed by dimensionality reduction with Principal Component Analysis (PCA) to obtain new variables and expand existing ones through Recursive Feature Elimination (RFE). The data set is divided into training and testing subsets, with the Random Forest (RF) and Deep Neural Network (DNN) models trained on them, respectively, followed by comprehensive hyper parameter tuning evaluations. Evaluating the results focuses on the system’s accuracy, recall, precision, and F1 score. Demonstrated findings show that DNN significantly outperformed RF in the selected model, DNN attained an accuracy of 97.5% and sustained strong performance on the other metrics. This research not only shows the superiority of DL techniques over older approaches with complex layers of student data but also contributes to the educational data mining discipline with a robust framework for early identification of at-risk students.