Loan Repayment Prediction Using XGBoost and Neural Network in Japan's Technical Internship Training

Abstract

Delayed repayment of financial aid among participants in Japan’s Technical Internship Training Program presents challenges for training institutions in managing funds efficiently. To address this issue, this study aims to compare the performance of two machine learning models: Extreme Gradient Boosting (XGBoost) and Multi-Layer Perceptron (MLP) in predicting the likelihood of delayed loan repayments. The research begins with data preprocessing, including handling missing values, normalization, and feature selection based on a correlation threshold of 0.06, where features with absolute correlation values below this threshold are excluded. Three models are tested: XGBoost Default, XGBoost optimized using GridSearchCV, and MLP. These models are evaluated using performance metrics such as accuracy, precision, recall, F1-score, and ROC-AUC. The XGBoost Default model achieves the highest accuracy at 95% and precision of 95%, although its recall is slightly lower at 83%. Tuning XGBoost improves recall to 84%, albeit with a marginal reduction in accuracy to 94%. In contrast, the MLP model demonstrates the lowest performance, with an accuracy of 92% and recall of 74%, indicating limitations in identifying delayed repayments. XGBoost also outperforms MLP in terms of ROC-AUC, scoring 91% compared to MLP’s 86%. These findings suggest that XGBoost is the more effective model for this predictive task. The results have practical implications for training institutions, enabling better participant selection, reducing repayment delays, and supporting more effective financial aid management.

GS Cited Analysis