Credit risk predictions with Bayesian model averaging

Model uncertainty remains a challenge to researchers in different applications. When many competing models are available for estimation, and without enough guidance from theory, model averaging represents an alternative to model selection. Despite model averaging approaches have been present in statistics for many years, only recently they are starting to receive attention in applications. The Bayesian Model Averaging (BMA) approach sometimes can be difficult in terms of applicability, mainly because of the following reasons: firstly two types of priors need to be elicited and secondly the number of models under consideration in the model space is often huge, so that the computational aspects can be prohibitive. In this paper we show how Bayesian model averaging can be usefully employed to obtain a well calibrated model, in terms of predictive accuracy for credit risk problems. In this paper we shall investigate how BMA performs in comparison with classical and Bayesian (single) selected models using two real credit risk databases.