Machine Learning Techniques for Forecasting Student Achievement in Digital Courses

The growing adoption of digital learning platforms has created an urgent need for accurate, interpretable methods to predict student performance and identify at-risk learners. While machine learning (ML) has shown potential in this domain, existing approaches often lack generalizability, real-time applicability, and transparency for educators. This study addresses these gaps by developing and evaluating multiple ML models including XGBoost, LSTM, and Random Forest on the Open University Learning Analytics Dataset (OULAD), which integrates demographic, behavioral, and academic data. Our results demonstrate that XGBoost outperforms other models, achieving 87% accuracy and 0.91 AUC-ROC, with engagement metrics and procrastination patterns emerging as key predictors. Additionally, we enhance model interpretability using SHAP values, enabling educators to understand and act on predictive insights. Unlike prior studies limited to single-institution data, our framework offers scalable and generalizable solutions for diverse online learning environments. These findings advance the field by providing actionable, data-driven tools for early intervention and academic support in digital education.