Cubic Ridge Regression for Machine Learning Credit Risk Scores

This technical report presents a novel approach to scaling machine learning credit risk scores using cubic Ridge regression. The document outlines the challenges faced by financial institutions in developing credit risk models, particularly the need to scale raw score values to match existing score ranges. It explains the common practice of using a two-step linear transformation for scaling and highlights the lack of literature on high decimal precision analytics in credit scoring. The authors introduce their method, which combines third order polynomial Ridge regression with established techniques to enhance score fidelity. Empirical results indicate that this approach leads to a significant increase in unique score values, improving risk differentiation for business users. The report details four key benefits of the proposed framework, including better matching of odds at base score values and preservation of points to double the odds. The findings are supported by data from a proprietary dataset, demonstrating the effectiveness of the new scaling method.