Signaling Crises: How to Get Good Out-of-Sample Performance Out of the Early Warning System

In past years, the most common approaches for deriving early-warning models belong to the family of binary-choice methods, which have been coupled with a separate loss function to optimize model signals based on policymakers preferences. The evidence in this paper shows that early-warning models should not be used in this traditional way, as the optimization of thresholds produces an in-sample overfit at the expense of out-of-sample performance. Instead of ex-post threshold optimization based upon a loss function, policymakers' preferences should rather be directly included as weights in the estimation function. Doing this strongly improves the out-of-sample performance of early-warning systems.