Please use this identifier to cite or link to this item: https://hdl.handle.net/10419/233315 
Year of Publication: 
2015
Series/Report no.: 
Discussion paper No. 99
Publisher: 
Aboa Centre for Economics (ACE), Turku
Abstract: 
The early warning system literature on banking crises has often relied on linear classifiers such as the logit model, which are usually estimated with large datasets of multiple regions of countries. We construct an EWS based on an artificial neural network model with monthly data from the Scandinavian countries to tackle the poor generalization ability of the usual models that might be due to regional heterogeneity of the countries and a nonlinear decision boundary of the classification problem. We show that the Finnish and Swedish banking crises in 1991 were quite predictable with an artificial neural network model when information from earlier crises in Denmark and Norway was used. We also use cross validation in the model selection process to get the optimal amount of complexity to the models. Finally the area under the ROC-curve is used as the model assessment criteria and in this framework we show that the artificial neural network outperforms the logit regression in banking crises prediction.
Subjects: 
Early Warning System
Banking Crises
Scandinavia
Neural Networks
Validation
JEL: 
G21
C45
C52
Document Type: 
Working Paper

Files in This Item:
File
Size





Items in EconStor are protected by copyright, with all rights reserved, unless otherwise indicated.