Nonlinear and Semi-Parametric Modelling of Personal Loan Credit Scoring

Abstract

This paper describes an empirical investigation into the predictive ability of four credit scoring models as applied to US personal loans. The models tested include the Logit model (LM), the divergence – a discriminant – method (DVM), neural networks (NN), and the generalized additive model (GAM). The first three methods have been widely applied in the literature, whilst GAM, a semi-parametric method, has also become popular recently. These models were applied to a dataset consisting of 2,429 consumer loans with 306 defaults during the period 2007-2013, and eight loan-specific characteristics. The performance of the four models was assessed in terms of receiver operating characteristics (ROC) and the area under the curve (AUC). Notwithstanding its simplicity, the DVM outperformed the LM and NN. The GAM had superior classification ability in the testing sample due to its flexibility. Furthermore, the underlying relationship between one of the key risk characteristics, debt/equity ratio, and the response variable was found to be nonlinear which was only revealed by the semi-parametric GAM – a noteworthy outcome of this study. By contrast, the effect of debt/equity ratio on the probability of default (PD) was found to be insignificant when using the parametric LM. These findings have implications for banks and other lending institutions in reducing losses due to loan defaults, as well as improving credit risk analysis, integral to lending institutions for implementing Basel rules on regulatory capital allocation.