Combination of unsupervised discretization methods for credit risk.

Area of Science:

• Statistics
• Machine Learning
• Financial Modeling

Background:

• Robust and explainable statistical learning models are crucial for credit risk management.
• Traditional discretization methods like equal width or frequency are widely used but have limitations, potentially losing underlying patterns.

Purpose of the Study:

• To introduce innovative discretization techniques by combining traditional methods with clustering-based approaches (k-means, Gaussian mixture models).
• To propose and evaluate two novel combination strategies: Discrete Competitive Combination (DCC) and Discrete Exhaustive Combination (DEC).

Main Methods:

• Combining traditional discretization with k-means and Gaussian mixture models.
• Developing Discrete Competitive Combination (DCC) and Discrete Exhaustive Combination (DEC) strategies.
• Applying these methods to 11 credit risk datasets using logistic regression and weight of evidence transformation.

Main Results:

• Both DCC and DEC combinations demonstrated superior performance compared to individual discretization methods.
• The proposed combinations maintained computational efficiency.
• The enhanced models did not compromise the explainability of the logistic regression models.

Conclusions:

• The novel combination approaches (DCC and DEC) offer a feasible and competitive alternative to conventional discretization methods in credit risk management.
• These methods effectively capture complex patterns without hindering model interpretability or efficiency.