Leveraging Tabular Variational Autoencoders for Student Loan Risk Modelling in Data-Scarce Environments

0 Article Views All time
0 Downloads All time
0 Citations Source: OpenAlex

Citations by year

No citation data available yet.

Abstract

Higher Education Loan Boards (HELBs) in Low- and Middle Income Countries (LMICs) must assess student loan repayment risk despite data scarcity, missingness, and class imbalance. Traditional models such as Logistic Regression and Random Forest often perform poorly under these conditions, limiting early risk identification. This study develops a two-stage predictive framework integrating a Tabular Variational Autoencoder (TVAE) for high quality synthetic data augmentation and a Multi-Input Multi-Output neural architecture with Multi-Task Learning (MIMO-MTL) to jointly estimate Probability of Default (PD) and Loss Given Default (LGD). TVAE outperforms CTGAN and Gaussian Copula and was selected for its ability to preserve joint feature structure. Using data from Malawi’s Higher Education Students Loans and Grants Board (HESLGB), the framework achieved an AUC of 0.96 for PD and an R² of 0.74 for LGD, improving risk detection and enabling earlier identification of high-risk borrowers. SHAP analysis enhanced interpretability by identifying key socioeconomic and academic rivers, offering a scalable, policy-relevant approach to strengthening student loan sustainability in a data scarce environment.