A machine learning-based early warning system for systemic banking crises

Econometricians construct panel logit-based early warning systems (EWSs) as the primary predictive tool to prevent incoming systemic banking crises. Considering the actual scenario of systemic banking crises, we argue that changes in economic indicators under the crisis may impact the information extraction of EWSs based on logistic regression. According to the potential limitations of the conventional EWS and properties of the machine learning algorithm, we assume that an 'experts voting EWS' framework can better fit characteristics of data of systemic banking crisis. Indeed, among other machine learning classifiers tested in this paper, random forest classifier simulating experts voting process is the most efficient classifier showing relatively high generalization above 80% area under the receiver operating characteristic curve on constructing the EWS. In contrast to the conventional system, an image of evidence shows that the experts voting EWS synthesizing multivariate information may be suitable for providing systemic banking systemic crises alerts in varied contexts.