Cross-country linkages and spill-overs in early warning models for financial crises

This paper uses data on bilateral foreign exposures of domestic banking systems in order to construct early warning models for financial crises that take into account cross-country spill-overs of vulnerabilities. The empirical results show that incorporating cross-country financial linkages can improve the signalling performance of early warning models. The relative usefulness increases from 65% to 87% and the AUROC from 0.89 to 0.97 when weighted foreign variables are added to domestic variables in a multivariate logit early warning model. The findings of the paper also suggest that global variables still play a role in predicting financial crises, even when foreign variables are controlled for, which could suggest that both cross-country spill-overs and contagion are important factors for driving financial crises. A parsimonious model with nine variables that combines domestic, foreign and global variables yields an out-of-sample relative usefulness of 0.82 with Type I and Type II errors of 0.11 and 0.07.