Narrating Minimal Data: Rethinking Cohort-Based GPA Prediction in Low-Resource Higher Education Contexts

Berdinata Massang (Institut Agama Kristen Negeri Manado)
Rolty Glendy Wowiling (Institut Agama Kristen Negeri Manado)
Allin Junikhah (Universitas Islam Negeri Maulana Malik Ibrahim Malang)
Firmanians Romula Tuerah (Institut Agama Kristen Negeri Manado)
Andrew Nathanael Ratag (Institut Agama Kristen Negeri Manado)
Febri Kurnia Manoppo (Hoseo University)

International Journal of Educational Narratives
Vol. 4 No. 1 (2026)

Background. Student performance prediction has become a major topic in educational data mining and learning analytics. However, most previous studies rely on high-dimensional datasets such as attendance records, course-level grades, and learning management system logs, which are often unavailable in institutions with limited digital infrastructure. Purpose. This study aims to evaluate the feasibility of predicting student academic performance using minimal institutional data and to establish a practical baseline for machine learning implementation in low-resource higher education contexts. Rather than maximizing predictive accuracy, this research examines the lower boundary of predictive capability when only simple academic variables are available. Method. A quantitative descriptive–predictive design was applied to 355 student records from the Christian Religious Education Study Program at IAKN Manado, Indonesia. GPA values were categorized into four classes (Poor, Fair, Good, and Very Good). The dataset was split into 75% training and 25% testing subsets, and class imbalance was addressed using SMOTE. Four models were evaluated: Dummy Classifier, Decision Tree, Random Forest, and Neural Network (MLP). Performance was assessed using accuracy and 5-fold cross-validation. Results. The Dummy Classifier achieved an accuracy of 15.73%, establishing a realistic baseline under balanced class conditions. Decision Tree and Random Forest produced the highest accuracy at 46.06%, while the Neural Network achieved 40.44%. However, cross-validation results remained lower, indicating limited generalization and possible overfitting under minimal-feature conditions. Conclusion. This study shows that simple institutional data can still provide non-trivial predictive signals, but predictive performance remains moderate. The main contribution of this study lies in positioning minimal-data prediction as a baseline methodological framework for institutions with constrained academic datasets, rather than as a high-accuracy predictive solution.