Book

Predicting which students will succeed—and which will struggle—has never been more urgent, yet the machine-learning literature on academic performance prediction remains fragmented and difficult to translate into practice. This systematic literature review of 40 peer-reviewed studies published between 2018 and 2025 shows that ensemble methods, especially Random Forest, XGBoost and LightGBM, dominate the field and routinely achieve classification accuracies of 85–94 percent, while deep-learning models such as LSTM and Bi-LSTM can perform even better when sequential engagement data are available. Prior academic record and LMS-derived behavioural features consistently carry the greatest predictive weight; demographic variables alone rarely justify deployment. Beneath these headline results, however, lie serious methodological weaknesses: most models are trained on single-institution data, accuracy is often reported on class-imbalanced sets without correction, external validation is almost non-existent, and questions of explainability, early prediction and ethical governance receive scant attention. The review maps six concrete research gaps and argues that progress on them is essential before predictive systems can be deployed equitably and responsibly at scale—making this an essential synthesis for researchers, institutional leaders and policymakers alike.