An Ensemble Model for Recovery Value in Default

Terry Benzschawel

This article was first published as a chapter in Credit Modelling (2nd edition), by Risk Books.

Although models for estimating firms’ probabilities of default (PD) abound, there are very few models for estimating recovery values in default.¹¹The only models for recovery value in default are Moody’s LossCalc (Gupton and Stein, 2002), Standard & Poor’s (2008) Recovery Value Rating and Citi’s decision tree model (Benzschawel, Haroon and Wu, 2011), with only Moody’s and Citi’s being available commercially. This is surprising as recovery values in default vary widely and expected losses in default depend equally on recovery value and likelihood of default. In an attempt to address this problem, Benzschawel and Li (2013) acquired Moody’s default and recovery value database, and used the data to develop a decision tree model (DTM) for recovery value in default. They demonstrated that the DTM performed better than using industry averages at predicting recovery values on defaulted bonds. Nevertheless, it was suspected that applying more sophisticated statistical techniques could enhance model performance. To that end, Benzschawel and Su (2013) developed an ensemble system that combines