Loss given default modeling: a comparative analysis

ABSTRACT In this study we investigated several of the most popular loss given default (LGD) models (least-squares method, Tobit, three-tiered Tobit, beta regression, inflated beta regression, censored gamma regression) in order to compare their performance. We show that for a given input data set the quality of the model calibration depends mainly on the proper choice (and availability) of explanatory variables (model factors), but not on the fitting model. Model factors were chosen based on the amplitude of their correlation with historical LGDs of the calibration data set. Numerical values of nonquantitative parameters (industry, ranking, type of collateral) were introduced as their LGD average. We show that different debt instruments depend on different sets of model factors (fromthree factors for revolving credit or for subordinated bonds to eight factors for senior secured bonds). Calibration of LGD models using distressed business cycle periods provide better fit than data from total available time span. Calibration algorithms and details of their realization using the R statistical package are presented.We demonstrate how LGD models can be used for stress testing. The results of this study can be of use to risk managers concerned with compliance with the Basel Accord.