Reducing model risk in early warning systems for banking crises in the euro area

We aim at detecting periods preceding banking crises in the euro area through an early warning system (EWS) and propose a new method to deal with model uncertainty. In a first step, we select a set of macro-financial risk indicators for their signaling ability among a large number of candidates over the period spanning from 1985:q1 to 2009:q4. Then, we run all the possible logit models including four of these indicators as explanatory variables in order to assess the pre-crises probabilities at each time. We retain two sets of models: a small one only including models with all coefficients significant and with the expected signs, and a large set, obtained by relaxing the selection criteria. In a second step, we calculate the weighted average of the pre-crisis probabilities estimated by the models belonging to the two selected sets. The weight given to each model is proportional to its usefulness at identifying pre-crises periods either at the panel or the country-level. The simulations performed both over and out of sample show that aggregating more models yields better results than relying on any single model or only a few of them, as model uncertainty is reduced. Performance is also enhanced by aggregating models’ results with country-specific weights relatively to common panel-weightings.