NOTE: non-parametric oversampling technique for explainable credit scoring.

Area of Science:

• Computational Finance
• Machine Learning
• Data Science

Background:

• Credit scoring models are vital for financial institutions to manage borrower risk and ensure profitability.
• Machine learning enhances credit scoring accuracy, but imbalanced datasets and non-linear data pose significant challenges.
• Existing methods like Synthetic Minority Oversampling TEchnique (SMOTE) struggle with high-dimensional, non-linear data and can introduce noise.

Purpose of the Study:

• To address limitations in current oversampling techniques for imbalanced credit scoring datasets.
• To develop a novel approach for extracting non-linear latent features and improving model explainability.
• To introduce the Non-parametric Oversampling Technique for Explainable credit scoring (NOTE) as a superior alternative.

Main Methods:

• Developed the Non-parametric Oversampling Technique for Explainable credit scoring (NOTE), a unified approach.
• Integrated a Non-parametric Stacked Autoencoder (NSA) for capturing non-linear latent features.
• Utilized Conditional Wasserstein GANs (cWGANs) for minority class oversampling and incorporated an explainability-focused classification process.

Main Results:

• The NOTE method demonstrated superior performance compared to state-of-the-art oversampling techniques.
• NOTE significantly improved classification accuracy and model stability on non-linear and imbalanced credit scoring datasets.
• The proposed technique enhanced the explainability of credit scoring model outcomes.

Conclusions:

• The NOTE approach effectively handles complex, imbalanced credit scoring data.
• NOTE offers a promising solution for improving both the predictive power and interpretability of credit scoring models.
• This research advances the application of machine learning in financial risk assessment by providing a more robust and transparent method.