Modeling Loss Given Default

We investigate the puzzle in the literature that various parametric loss given default (LGD) statistical models perform similarly by comparing their performance in a simulation framework. We find that, even using the full set of explanatory variables from the assumed data generating process, these models still show similar poor performance in terms of predictive accuracy and rank ordering when mean predictions and squared error loss functions are used. Therefore, the findings in the literature that predictive accuracy and rank ordering cluster in a very narrow range across different parametric models are robust. We argue, however, that predicted distributions as well as the models' ability to accurately capture marginal effects are also important performance metrics for capital models and stress testing. We find that the sophisticated parametric models that are specifically designed to address the bi-modal distributions of LGD outperform the less sophisticated models by a large margin in terms of predicted distributions. Also, we find that stress testing poses a challenge to all LGD models because of limited data and relevant explanatory variable availability, and that model selection criteria based on goodness of fit may not serve the stress testing purpose well. Finally, the evidence here suggests that we do not need to use the most sophisticated parametric methods to model LGD