Simulation and estimation of loss given default

Stefan Hlawatsch, Sebastian Ostrowski

Abstract

ABSTRACT

We aim to develop an adequate estimation model for loss given default that incorporates the empirically observed bimodality and bounded nature of the distribution. To this end, we introduce an adjusted expectation-maximization algorithm to estimate the parameters of a univariate mixture distribution consisting of two beta distributions. Furthermore, we analyze our derived estimation model using estimation models proposed in the literature on synthesized loan portfolios. The simulated loan portfolios consist of possibly loss-influencing parameters that are merged with loss given default observations using a quasi-random approach. Our results show that our proposed model exhibits more accurate loss given default estimators than the benchmark models for different simulated data sets comprising obligor-specific parameters with either high predictive power or low predictive power for the loss given default.