Abstract
Ensemble algorithms can learn complex nonlinear relationships in large datasets resulting in higher predictive accuracies than the conventional methods. Practitioners and regulators have shown substantial hesitance in adopting them in credit risk management because of their need for explainablity. Using five ensemble learning techniques and a one-dimensional convolutional neural network, we assess indicators to predict asset quality deterioration in a consumer loan dataset using the SHAP framework to achieve explainablity of the models' ranking of features significance. We implemented a novel model-agnostic aggregate ranking method to rank the importance of the overall features from each model in predicting NPLs.
Original language | English |
---|---|
Article number | 104084 |
Journal | Finance Research Letters |
Volume | 56 |
Early online date | 10 Jun 2023 |
DOIs | |
Publication status | Published - Sept 2023 |
Keywords
- Credit risk
- Ensemble methods
- Explainable artificial intelligence
- Non-performing loans
ASJC Scopus subject areas
- Finance