Learning from Trees : A Mixed Approach to Building Early Warning Systems for Systemic Banking Crises

Banking crises can be extremely costly. The early detection of vulnerabilities can help prevent or mitigate those costs. We develop an early warning model of systemic banking crises that combines regression tree technology with a statistical algorithm (CRAGGING) to improve its accuracy and overcome the drawbacks of previously used models. Our model has a large set of desirable features. It provides endogenously-determined critical thresholds for a set of useful indicators, presented in the intuitive form of a decision tree structure. Our framework takes into account the conditional relations between various indicators when setting early warning thresholds. This facilitates the production of accurate early warning signals as compared to the signals from a logit model and from a standard regression tree. Our model also suggests that high credit aggregates, both in terms of volume and as compared to a long-term trend, as well as low market risk perception, are amongst the most important indicators for predicting the build-up of vulnerabilities in the banking sector