From Prediction to Action: A Portfolio Framework for Student Performance Improvement Plans in Higher Education

Abstract

Higher education institutions increasingly seek student-success systems that move beyond descriptive monitoring and generate operationally usable improvement plans. The study de-velops and tests a portfolio-oriented analytics framework which uses multiclass academic outcome prediction to create different student performance improvement plans. The study uses the UCI Predict Students’ Dropout and Academic Success dataset which contains data about 4424 students to assess three models: multinomial logistic regression, random forest, and XGBoost for predicting student dropout, continued enrollment, and graduation. The random forest model achieves its best performance with a cross-validated macro-F1 score of 0.702 and a hold-out accuracy of 0.748 which includes a macro-AUC score of 0.889. The analysis of predictors shows that the model performance depends mostly on semester-level approved units and curricular-unit grades and academic evaluations and tuition status and admission-related variables. The paper develops a priority score with an action-mapping system which places students into five different managerial plans: financial continuity, aca-demic recovery, progression coaching, stabilization, and acceleration. The resulting portfolio yields distinct empirical profiles; for example, the financial continuity and academic recovery groups each contain dropout rates above 94%, whereas the acceleration portfolio exhibits a graduate rate above 95%. The study makes two main contributions. The first part demon-strates that public higher-education records can support reproducible multiclass prediction with institutionally interpretable drivers. The second part provides an analytically based planning framework which assists organizations in developing their prediction models into intervention design, resource allocation and prioritization efforts.