Research on Credit Default Prediction Model Based on TabNet-Stacking.

Area of Science:

• Machine Learning
• Financial Technology
• Data Science

Background:

• Traditional credit default prediction models struggle with the demands of financial technology.
• Experience-based and single-network models lack the sophistication for current financial landscapes.

Purpose of the Study:

• To propose an improved credit default prediction model using TabNeT-Stacking.
• To enhance the accuracy and performance of credit default prediction in financial technology.

Main Methods:

• Developed an improved TabNet structure using PyTorch.
• Optimized feature selection with a multi-population genetic algorithm and hyperparameters with particle swarm optimization.
• Employed Stacking ensemble learning with an improved TabNet for feature extraction, XGBoost, LightGBM, CatBoost, KNN, and SVM as base learners, and XGBoost as the meta-learner.

Main Results:

• The proposed TabNeT-Stacking model significantly outperforms original models.
• Demonstrated superior performance in accuracy, precision, recall, F1 score, and Area Under the Curve (AUC).

Conclusions:

• The TabNeT-Stacking model offers a robust and effective solution for credit default prediction.
• The integration of deep learning and ensemble methods provides a significant advancement in financial risk assessment.