Early-warning models for systemic banking crises: the effect of model and estimation uncertainty

• Applicant: Professor Dr. Gregor von Schweinitz
• Subject Area: Economic Policy, Applied Economics ; Statistics and Econometrics
• Term: from 2016 to 2019
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 334853253

Early warning models for costly economic crises - such as the series of crises we just experienced - are used to guide several areas of economic policy such as macroprudential regulation or the macroeconomic imbalances procedure of the European Commission. The usability strongly relies on the economic and statistical significance of the models, that is, their informativeness with regard to the future occurrence of crises, and the precision of their warnings.Early warning models in the literature have often fallen short of at least one of these two goals. Most of the literature simply disregards statistical significance. Policy recommendations are derived from point estimates of probabilities without recurrence to the uncertainty of probability estimates. This naive approach is particularly problematic due to the rare and clustered nature of crises, which implies that it may take a long time until estimation (and conceptual) errors are uncovered. Even if not formally accounted for, statistical significance is often the reason for large and heterogeneous datasets. These, however, come at the disadvantage of sometimes strongly reduced economic significance.This project is going to account for and improve on statistical and economic significance in three ways. In Module 1, we explore a country-specific weighted early-warning model, where observation weights describe the similarity to observations from the country under review. This model enforces a rather homogeneous sample without losing vital information. Thus, it should improve on both goals at the same time. In Module 2, we extend Bayesian models beyond capturing model and estimation uncertainty to the derivation of Bayesian policy recommendations. Bayesian models are robust with respect to selection of explanatory variables. This is sorely needed in early warning models, as the number of potential early warning indicators is very high, while the number of explained events is often very low. Thus, Bayesian methods increase the performance of the early warning model while properly accounting for estimation uncertainty. In Module 3, we give economic meaning to the derivation of policy recommendations from early-warning probabilities. Currently, the probability threshold used to split probabilities into signals (high probabilities) and no signals (low probabilities) is selected based on a cost parameter which is not economically deduced. Instead, we base the selection of the threshold on an estimation of the costs and benefits of regulation and systemic banking crises from a growth regression. Thus, we increase the economic significance of policy recommendations, as they will be directly related to the output loss that can be expected in case of a false alarm or a missed event. The individual models in the three modules can be easily combined.The research in this project will thus (a) correctly capture the statistical significance of recommendations and (b) improve their economic significance.

• DFG Programme: Research Grants
• International Connection: Austria, Finland
• Cooperation Partners: Professor Dr. Jesus Crespo-Cuaresma; Professor Dr. Reint Gropp, Ph.D.; Professor Peter Sarlin, Ph.D.