(Non) consistency of the Beta Kernel Estimator for Recovery Rate Distribution

In this paper, we explain why a nonparametric approach based on a betakernel [Renault, Scaillet (2004)] will lead to significant bias when appliedto recovery rate distributions. This is due to a specific feature of thesedistributions, which admit strictly positive weights at 100 % correspondingto full recovery (and also at 0 % corresponding to total loss). Moreover, fordistributions without point mass at 0% and 100%, the beta kernel approachfeatures significant bias in finite sample. In large sample the method isconsistent, but other competing approaches presented in the paper providemore accurate results.