Analysis of Prediction Model for Repayment Ability of Loan Funds Using XGBoost and Neural Network in Technical Internship Training Programs in Japan

EasyChair Preprint 15866

Abstract

This study compares XGBoost and Multi-Layer Perceptron (MLP) models in predicting the delayed repayment of financial aid provided to technical internship training in Japan. The preprocessing stage includes handling missing data, normalization, and feature selection using a correlation threshold of 0.06, where features with an absolute correlation below this value are removed. The XGBoost Default, XGBoost with GridSearchCV, and MLP models are evaluated using accuracy, precision, recall, F1-score, and ROC-AUC metrics. The results show that XGBoost Default achieves the highest accuracy (95%) with precision of 95%, but a recall of only 83%, whereas XGBoost with GridSearchCV improves recall to 84% with a slight accuracy reduction (94%). The MLP model performs the worst (accuracy 92%, recall 74%), indicating difficulties in detecting delays. With a ROC-AUC of 91% for XGBoost compared to 86% for MLP, XGBoost proves to be the superior model. This model can assist training institutions providing financial aid programs in more accurately selecting participants, reducing repayment delays, and improving financial management efficiency.

Keyphrases: Late Payment Prediction, Loan Repayment, Multi Layer Perceptron, XGBoost, machine learning

@booklet{EasyChair:15866,
  author    = {Mohammad Roffi Suhendry and Gerry Firmansyah and Nenden Siti Fatonah and Agung Mulyo Widodo},
  title     = {Analysis of Prediction Model for Repayment Ability of Loan Funds Using XGBoost and Neural Network in Technical Internship Training Programs in Japan},
  howpublished = {EasyChair Preprint 15866},
  year      = {EasyChair, 2025}}