Asymmetric Loss-Oriented Cost Sensitive Learning Model for Credit Risk Assessment

Abstract

Credit risk assessment (CRA) is a hot topic in both academia and financial circles. Existing CRA methods can encompass both experience-based and data-driven approaches. The data-driven approaches are mainstream in recent years. Various data-driven CRA techniques are constructed with machine learning techneques. Since the non-default consumers are far more than the default consumers, CRA is typically a class imbalanced problem. Indeed, the misclassification of default consumers usually encounters heavier loss than the contrary case. An asymmetric loss-oriented cost sensitive learning model named CSLINEX-SVM is proposed by using the asymmetry of LINEX loss to assign more costs to the minority class. In credit risk assessment, the default consumers (majority class) deserve more penalty than the non-default consumers (majority class). Thus, we apply the Nesterov accelerated gradient algorithm to solve CSLINEX-SVM for credit risk assessment. The experiments verify the effectiveness of CSLINEX-SVM.