A framework for early-warning modeling with an application to banks

This paper proposes a framework for deriving early-warning models with optimal out-of-sample forecasting properties and applies it to predicting distress in European banks. The main contributions of the paper are threefold. First, the paper introduces a conceptual framework to guide the process of building early-warning models, which highlights and structures the numerous complex choices that the modeler needs to make. Second, the paper proposes a flexible modeling solution to the conceptual framework that supports model selection in real-time. Specifically, our proposed solution is to combine the loss function approach to evaluate early-warning models with regularized logistic regression and cross-validation to find a model specification with optimal real-time out-of-sample forecasting properties. Third, the paper illustrates how the modeling framework can be used in analysis supporting both microand macro-prudential policy by applying it to a large dataset of EU banks and showing some examples of early-warning model visualizations.