Heterogeneities among credit risk parameter distributions: the modality defines the best estimation method

Abstract

Comparative studies investigating the estimation accuracy of statistical methods often arrive at different conclusions. Therefore, it remains unclear which method is best suited for a particular estimation task. While this problem exists in many areas of predictive analytics, it has particular relevance in the banking sector owing to regulatory requirements regarding transparency and quality of estimation methods. For the estimation of the relevant credit risk parameter loss given default (LGD), we find that the different results can be attributed to the modality type of the respective LGD distribution. Specifically, we use cluster analysis to identify heterogeneities among the LGD distributions of loan portfolios of 16 European countries with 32,851 defaulted loans. The analysis leads to three clusters, whose distributions essentially differ in their modality type. For each modality type, we empirically determine the accuracy of 20 estimation methods, including traditional regression and advanced machine learning. We show that the specific modality type is crucial for the best method. The results are not limited to the banking sector, because the present distribution type-dependent recommendation for method selection, which is based on cluster analysis, can also be applied to parameter estimation problems in all areas of predictive analytics.