An Early Warning System for Systemic Banking Crises: A Robust Model Specification

Using a panel dataset of 27 developed economies, estimated quarterly from 1980-2016, we develop a flexible systemic banking crisis early warning system (EWS). Evidence is provided that fitted multivariate logit probabilities, estimated recursively against documented crises, yield more informative crisis signals than any single macroeconomic, credit aggregate or asset price variable does independently. When the model robustness techniques of Young and Holsteen (2017) are applied, even stronger crisis signals are generated. Deciding which variables to include in the model is determined by adopting a signals-based approach to each prospective indicator, with the most informative yet robust variables identified in terms of their performance according to noise-to-signal ratios, weighted noise-to-signal ratios and an Alessi and Detken (2011) “usefulness” measure. The latter takes policy-makers’ preferences for false versus missed signals into account. The approach ensures a parsimonious yet effective EWS yielding forwardlooking indicators that outperform all raw input indicators in crisis-signaling terms.