Loss given default estimation: a two-stage model with classification tree-based boosting and support vector logistic regression

Yuta Tanoue and Satoshi Yamashita

Abstract

The Basel Accords allow banks to estimate credit risk. Accordingly, more attention has been dedicated recently to the analysis of loss given default (LGD) and the development of an LGD estimation model. In this study, using a data set composed of five Japanese regional banks, we propose an LGD estimation model using a two- stage model, classification tree-based boosting and support vector regression (SVR). We compare the proposed model’s predictive performance with existing models by performing cross-validation and out-of-time validation. As a result, we find that incorporating nonlinearity into the LGD estimation model by classification and SVR improves its predictive performance. Further, we confirm that the boosting method improves the model’s predictive performance.