Comparing logit-based early warning systems: Does the duration of systemic banking crises matter?

This paper compares the performance of binomial and multinomial logit models in the context of building early warning systems (EWS) for systemic banking crises. We test the hypothesis that the predictive performance of binomial logit models is hampered by what we define as the crisis duration bias, arising from the decision to either treat crisis years after the onset of a crisis as non-crisis years or remove them altogether from the sample. In line with our hypothesis, results from a large sample of world economies suggest that i) the multinomial logit outperforms the binomial logit model in predicting systemic banking crises, and ii) the longer the average duration of the crisis in the sample, the larger the improvement