Penalized maximum likelihood estimation of logit-based early warning systems

Panel logit models have proved to be simple and effective tools to build early warning systems (ews) for financial crises. But because crises are rare events, the estimation of ews does not usually account for country-specific fixed effects, so as to avoid losing all the information relative to countries that never face a crisis. I propose using a penalized maximum likelihood estimator for fixed-effects logit-based ews where all the observations are retained. I show that including country effects, while preserving the entire sample, improves the predictive performance of ews, both in simulation and out of sample, with respect to the pooled, random-effects and standard fixed-effects models.