Are systemic banking crises in developed and developing countries predictable?

This paper contributes to the empirical literature on early warning systems of banking crises using a new methodology accounting for model uncertainty. We introduce new variables measuring exposure and connectivity of the domestic banking sector to international financial markets. We show that a multinomial logit model based on Bayesian Model Averaging is favored to conventional multinomial and binary models highlighting what is called by Bussiere and Fratzsher (2006) “post-crisis bias”. We show that the application of the multinomial logit model, which distinguishes between more than two states and uses Bayesian Model Averaging, is a valid way to solve this problem and leads to a substantial improvement in the ability to predict banking crises. The empirical results show that for a set of 49 developing and developed countries, the model would have correctly predicted the vast majority of crises.