Comparing Model Selection Criteria to Distinguish Truncated Operational Risk Models

Abstract

In this paper three information criteria are employed to assess the truncated operational risk models. The performances of the three information criteria on distinguishing the models are compared. The competing models are constructed using Champernowne, Frechet, Lognormal, Lomax, Paralogistic, and Weibull distributions, respectively. Simulation studies are conducted before a case study. In the case study, certain distributional models conform to the external fraud type of risk data in retail banking of Chinese banks. However, those models are difficult to distinguish using standard information criteria such as Akaike Information Criterion and Bayesian Information Criterion. We have found no single information criterion is absolutely more effective than others in the simulation studies. But the information complexity based ICOMP criterion says a little bit more if AIC and/or BIC cannot kick the Lognormal model out of the pool of competing models.