Predicting Systemic Banking Crises using Extreme Gradient Boosting

Abstract

Considering the great ability of decision trees techniques to extract useful information from large databases and to handle heterogeneous variables, this paper applies Extreme Gradient Boosting for the prediction of systemic banking crises. To this end, prediction models have been constructed for different regions and the whole world. The results obtained show that Extreme Gradient Boosting overcomes the predictive power of existing models in the previous literature and provides more explanatory information on the causes that produce systemic banking crises, being the demand for deposits, the level of domestic credit and banking assets some of the most significant variables.