Loan Repayment Behavior Prediction of Provident Fund Users Using a Stacking-Based Model

With the rapid development of big data and cloud computing, the analysis and prediction of user loan repayment behavior has become a research hotspot in the field of financial technology. The majority of previous studies on such issues have focused on e-commerce platforms and commercial banks, but we innovatively apply the data mining and machine learning methods to the provident fund users. In addition to the traditional two prediction classes of normal loan repayment and loan overdue, we have also added the early loan repayment class to help the provident fund center better manage asset liquidity. We proposed a model based on stacking ensemble method, and compared with six individual models: Cart decision tree, Random Forest, AdaBoost, XGBoost, CatBoost and LightGBM. After experiments, our stacking model has a good improvement under the classification performance evaluation metrics such as accuracy score, recall score, F1 score, AUC score, and kappa score. Furthermore, we also analyzed the importance of features in the process of training the loan repayment behavior prediction model. Through our stacking model and research analysis, the provident fund centers can reduce loan risks and provide better services to their customers.