Intertemporal Forecasts of Defaulted Bond Recoveries and Portfolio Losses

Variation in the composition of the defaulted debt pool and credit conditions at the time of default generate time variation in the distribution of recoveries on defaulted debt, and the related distribution of losses on portfolios of credit sensitive debt. We quantify the importance of accounting for such time variation in out-of-sample comparisons of alternative approaches to forecasting recoveries or losses given default (LGD) on defaulted bonds. Using simulations of losses on defaultable bond portfolios, we show that conditional mixture models improve forecasts of expected credit losses through capturing time variation in the recovery/LGD distribution. However, the best forecasts of instrument or firm-level recovery/LGD do not necessarily provide the best forecasts of portfolio-level losses, as the latter depend on the association between errors in the default and recovery/LGD forecasts. Our systematic comparisons of cross-sectional and intertemporal forecasting performance are enabled by a fast maximum-likelihood approach to estimating conditional mixtures of distributions.