Predicting Banking Crises with Artificial Neural Networks: The Role of Nonlinearity and Heterogeneity

• Date: 01 January 2018
• Author: Kim Ristolainen
• DOI: http://doi.org/10.1111/sjoe.12216
• Published date: 01 January 2018

©The editors of The Scandinavian Journal of Economics 2016.

Scand. J. of Economics 120(1), 31–62, 2018

DOI: 10.1111/sjoe.12216

Predicting Banking Crises withArtificial

Neural Networks:The Role of Nonlinearity

and HeterogeneityÅ

Kim Ristolainen†

University of Turku, FI-20014 Turku, Finland

kim.ristolainen@utu.fi

Abstract

Studies of the early warning systems (EWSs) for banking crises usually rely on linear

classifiers, estimated with international datasets. I construct an EWS based on an artificial

neural network (ANN) model, and I also account for regional heterogeneity in order

to improve the generalization ability of EWS models. All of the banking crises in

my test set are then predictable at a 24-month horizon, using information from earlier

crises. For some countries, estimation with a regional dataset significantly improves the

predictions. The ANN outperforms the usual logit regression, assessed by the area under

the receiver operating characteristics curve.

Keywords: Artificial neural networks; banking crises; early warning system; financial stability

JEL classification:C52; C53; E58; G21

I. Introduction

The financial crises of emerging economies in the 1980s provoked further

studies on early warning systems (EWSs) for currency crises. These

predictive models are assumed to give a warning signal of an upcoming

financial crisis in some given time window, so that the policymakers can

take precautionary measures in good time. The EWSs are constructed using

various econometric and statistical models that are estimated with quarterly

or annual datasets and use information from crisis indicators, which are

usually taken from economic theory. I show that, with a monthly dataset,

the assumption of a linear decision boundary in the classification problem

used by the conventional EWS models is too strict and a significant

improvement (often almost perfect classification) can be achieved by using

artificial neural network (ANN) models. These models can approximate any

nonlinear decision boundary if there is a sufficient amount of complexity

*I am grateful to Paavo Okko, Heikki Kauppi, Matti Viren, Nadine Chlass, several seminar

participants, and both referees for helpful comments during this research. I also thank Joonas

Ollonqvist for technical support.

†Also affiliated with Bank of Finland, Helsinki.

32 Predicting banking crises with artificial neural networks

in the model. Owing to a lack of all the possible variables to describe the

possible heterogeneity between the countries used to estimate the model,

I also use regional datasets in addition to a global one. I show that for

some regions this procedure also improves the classification results and

allows almost perfect prediction of the banking crises in the 1990s, when

given the information from earlier crisis episodes.

The two pioneering papers in the field use different approaches to

construct these early warning models. Kaminsky et al. (1998) use a non-

parametric univariate signals approach, in which the idea is to compare

the behavior of economic fundamentals during normal periods of the

economy and in pre-crisis periods. In practice, a threshold for the crisis

signal is calculated for each indicator by minimizing the noise-to-signal

ratio.1For example, inflation over 10 percent would give a signal of

an upcoming crisis in some given time window. This approach was

extended by Kaminsky (1998), who constructed composite indicators from

the individual indicators using four different approaches. One approach

was that signals from individual indicators were summed together after

first multiplying them with the inverse of the indicator’s noise-to-signal

ratio; this gave the best results in currency crisis prediction. Berg and

Pattillo (1999) use a probit model for the same purpose. The probit/logit

model has the advantage of taking into account the correlation between

the indicators and it can aggregate the information into a composite

index more satisfactorily. They found that the probit model has superior

performance in predicting the currency crises of the 1990s, when compared

with the non-parametric Kaminsky–Lizondo–Reinhart model composite

indicator. Following the publication of these two papers, the body of

literature has expanded to include different types of financial crises and

more complex/advanced models.2Manasse et al. (2003) used a

classification and regression tree analysis to construct an EWS for debt

crises. Nag and Mitra (1999) used a dynamic ANN to predict currency

crises. Frank and Schmied (2003) constructed a similar analysis with a

different dataset. The last papers found ANN to be a superior predictor

when compared with the logit model. Fioramanti (2008) uses an ANN to

predict sovereign debt crises with a dataset of 46 emerging countries. He

concludes that the ANN outperforms the traditional probit EWS in crisis

prediction.

Although the costs of banking crises are compelling,3the body of EWS

literature on banking crises is rather brief compared with that for currency

1The ratio of correctly called crisis and incorrectly called crisis periods.

2For a comprehensive reviewof the body of literature on ANN-based EWS, see Sarlin (2012).

3Average bailout costs 10 percent of GDP (Caprio and Klingebiel, 1996) and the average

estimated cumulative output losses are 5.6 percent of GDP (Hoggarth et al., 2002).

©The editors of The Scandinavian Journal of Economics 2016.

K. Ristolainen 33

crises. Demirguc-Kunt and Detragiache (1998) have written the leading

paper on the banking crises in the body of EWS literature. They use a

multivariate logit model with a dataset of 77 countries. Kaminsky and

Reinhart (1999) used the signals approach to study the occurrence of twin

crises4with a dataset of 22 emerging countries. Davis and Karim (2008)

compared these methods with an updated dataset and concluded that the

signals approach might be better for country-specific banking crisis EWSs,

and the multivariate logit for a global EWS. To my knowledge, ANNs

have not been used in constructing an EWS for banking crises. This

paper tries to fill this gap in the body of literature and builds an EWS

with two different approaches to predict banking crises. The authors of

recent EWS studies (Davis et al., 2011) have argued that EWSs should

be built for each individual region, because of the regional heterogeneity

of the indicator variables signalling the crises. This is the reason why,

in this paper, I use regional datasets in addition to a global one to

study whether the homogeneity of the countries included in the estimation

improves the classification and generalization ability of the EWS models.

The monthly dataset from January 1970 to June 2003 includes a total of

18 countries, which makes the generalization of the results tedious.5An

important aspect in this research was the use of higher-frequency data

than annual or quarterly, which is why I adopted the monthly dataset

assembled by Kaminsky (2006). Although many of the variables could

be updated with more time observations and to include more countries,

I cannot update the dependent banking crisis variable that Kaminsky has

assembled by using information from financial newspapers and articles in

economic journals. This is one of the main reasons why I have not updated

more observations as I want to continue using only one author’s crisis

variable. I do not feel that it is feasible to extend this variable with some

other author’s banking crisis variable, because authors use different crisis

definitions and often make subjective decisions concerning the starting

dates for some specific crisis episode.

In this paper, my primary objective is to examine whether the out-

of-sample prediction of banking crises could have been possible with an

ANN model given only the information available before the beginning of

4The occurrence of banking and currency crises at the same time.

5The dataset was updated in 2005 to include Ecuador and Paraquay. This updated dataset

has a total of 22 countries, but I had to exclude Bolivia, Ecuador, Paraguay, and Uruguay

due to missing observations in key variables for many decades. For the same reasons that I

did not update more time observations mentioned in the text later on, I continued to use this

country set. However, many published papers have the same problem concerning the number

of countries in the analysis, and a similar set of countries has been used in three pioneering

papers in the body of EWS literature, Kaminsky et al. (1998), Berg and Pattillo (1999), and

Kaminsky and Reinhart (1999).

©The editors of The Scandinavian Journal of Economics 2016.