Default Prediction in the Finance Industry Based on Ensemble Learning: Combining Machine Learning and Deep Learning

Abstract

Background: Financial institutions face significant challenges in predicting loan defaults, which directly impact the non-performing loan (NPL) rate. Incorrect predictions can lead to misinformed decisions and substantial financial losses.  Objectives: This study aims to enhance default prediction by employing advanced ensemble learning techniques in machine learning and deep learning. Methods/Approach: Instead of relying on transformation, fine-tuning, or single algorithm models, this research focuses on combining multiple models using voting and stacking techniques, particularly highlighting a stacking model combining Light Gradient Boosting Machine (LGBM) and Artificial Neural Networks (ANN). Results: The ensemble learning methods, especially the LGBM-LSTM and XGB-LSTM stacking models, showed higher precision in identifying borrowers who defaulted, while the LGBM-LSTM and XGB-LSTM voting models excelled in recall and achieved an F1-score 0.1% higher. Both the stacking and voting models attained AUC values close to 90%, indicating strong overall classification performance. Conclusions: The findings not only contribute to the fields of lending and peer-to-peer financial operations but also offer crucial insights that aid financial organizations in making well-informed decisions regarding loan processing and management.