Further investigation of parametric loss given default modeling

Phillip Li, Min Qi, Xiaofei Zhang and Xinlei Zhao

Abstract

ABSTRACT

We conduct a comprehensive study of some parametric models that are designed to fit the unusual bounded and bimodal distribution of loss given default (LGD). We first examine a smearing estimator, a Monte Carlo estimator and a global adjustment approach to refine transformation regression models that address issues with LGD boundary values. Although these refinements only marginally improve model performance, the smearing and Monte Carlo estimators help to reduce the sensitivity of transformation regressions to the adjustment factor. We then conduct a horse race among the refined transformation methods, five parametric models that are specifically suitable for LGD modeling (two-step, inflated beta, Tobit, censored gamma and two-tiered gamma regressions), fractional response regression and standard linear regression. We find that the sophisticated parametric models do not clearly outperform the simpler ones in either predictive accuracy or rank-ordering ability, in-sample, out-of-sample or out of time. Therefore, it is important for modelers and researchers to choose the model that is appropriate for their particular data set, considering differences in model complexity, computational burden, ease of implementation and model performance.