Multinomial logit as an early warning model for predicting banking crises

Since the Global Financial Crisis in 2007, a domino of banking distress episodes took place around the world, both at individual and systemic level. The bailout cost of these banking crises was extreme, especially in USA and EU countries, which were mostly affected. In recent years, there are many empirical studies, which attempt to develop a new mechanism of Early Warnings. As a result, more integrated and accurate systems are developed. Concerning the selection of the appropriate method, multivariate logit analysis is arguably one of the most popular models for EWSs. An important methodological issue, regarding the application of binary logit models, is how to deal with the observations following the onset of a banking crisis. One can either treat years after crisis as tranquil periods or remove them from the sample. In this paper, a multinomial logit model is applied to include all the crisis observations in the sample, treating them as a different outcome. Subsequently, this paper compares the predictive ability of binomial and multinomial logit models as EWSs for systemic banking crises in Eurozone, and, contrary to previous results, multinomial logit model does not seem to perform well to signal systemic banking crises.