Keywords: Bayesian panel models, credit risk, default probability, longitudinal models, financial loss function, default forecasting, SMEs, small and medium-sized enterprises, Germany, financial risk
Default forecasting for small-medium enterprises: does heterogeneity matter?
In this paper we discuss and compare a set of classical and Bayesian longitudinal models to predict small-medium enterprise (SME) default probability, taking unobservable firm and business sector heterogeneities, as well as time variation, into account. By using a panel data set of German SMEs, we compare a large set of models by looking at their in-sample and out-of-sample forecasts. To choose the best model, we consider both a threshold independent criteria as well as a novel financial loss function. In terms of in-sample performances, we find that Bayesian models perform much better that classical longitudinal models and pooled logit models. Similarly, the former models show significant lower loss functions compared with the latter models. Instead, the out-of-sample performances are much closer and complex models are not statistically different from a simple pooled logit model.