Impact of segmentation on the performance measures of LGD models

Thomas, Lyn and Bijak, Katarzyna (2015) Impact of segmentation on the performance measures of LGD models. Credit Scoring and Credit Control XIV, Edinburgh, United Kingdom. 25 - 27 Aug 2015.

Record type: Conference or Workshop Item (Other)

Abstract

Loss Given Default (LGD) is difficult to model due to its distribution shape that usually features a high peak at zero and a possible spike at one. As a result, LGD models often perform rather weakly, and their performance measures can take embarrassingly poor values. Normally the performance of LGD models is assessed at the individual loan level. However, the Basel Accords require LGD estimates at the portfolio segment level. This research looks at how the measures vary depending on the number of loans in each segment, including the case where each segment has only one loan. It shows that in such a case some measures produce very poor results, while still demonstrating that the model performs well when the portfolio is divided into fewer but larger segments. Hence, for some measures the segmentation leads to a significant improvement in the resulting values. Nevertheless, there are also measures for which the segmentation makes no or little difference. The research undertakes both analytical and empirical analysis of how the measures applied at the segment level vary as a portfolio is divided into different numbers of segments. The empirical analysis is based on a portfolio of over 10,000 defaulted personal loans belonging to a large UK bank.

This record has no associated files available for download.