Research
Predicting Student Fee Default: A Comparative Model Study
IndabaX Ghana, Data Science Summit
A note on this one
The abstract below is the version submitted for the poster, and it describes a logistic regression model. Since then the project has grown into a comparative study across four classifiers, and gradient boosting, not logistic regression, turned out to give the best recall on the default class. The write up has not caught up with the work yet. I am revising it, and I would rather show you the gap than quietly paper over it.
Poster
Poster coming soon
Abstract
Fee default is a critical financial sustainability challenge for private schools in sub Saharan Africa, where tuition fees are the primary revenue stream. This study develops and deploys a logistic regression model to predict student fee default at a private basic school in Ghana's Volta Region, using 2,280 term level administrative payment records spanning three academic years (2023/24 to 2025/26). Fourteen predictor variables were identified from routine school management records, including payment compliance indicators, first payment behaviour, and student level characteristics. The model achieved 89% accuracy and a ROC-AUC of 0.942, with compliance with the final payment policy emerging as the strongest protective predictor. To operationalise the findings, an interactive web based dashboard was built using Streamlit, backed by a PostgreSQL database, and containerised with Docker, enabling real time default risk monitoring by school administrators. All student records were anonymised prior to deployment in adherence to institutional data protection obligations, demonstrating that ethical AI principles can be applied in resource constrained educational settings. This work shows that standard administrative data, when responsibly modelled, can transform reactive fee collection into a proactive early warning system accessible to under resourced private schools across Ghana and West Africa.
Abstract (PDF)
Key finding
Gradient boosting lifts default class recall to 0.70 on the test set and 0.64 under cross validation, the highest of all four classifiers, and the metric that matters most for early intervention.