Banking and Currency Crises: Differential Diagnostics for Developed Countries

• Published date: 01 January 2017
• Author: Marek Rusnák,Mark Joy,Bořek Vašíček,Kateřina Šmídková
• Date: 01 January 2017
• DOI: http://doi.org/10.1002/ijfe.1570

BANKING AND CURRENCY CRISES: DIFFERENTIAL DIAGNOSTICS

FOR DEVELOPED COUNTRIES

MARK JOY

a

, MAREK RUSNÁK

b

, KATEŘINA ŠMÍDKOVÁ

‡

and BOŘEK VAŠÍČEK

3,4,

*

,†

a

International Directorate, Bank of England, Threadneedle St, London EC2R 8AH, UK

b

Financial Stability Department, Czech National Bank, Na Příkopě28, 11503 Prague 1, Czech Republic

c

Economic Research Department, Czech National Bank, Na Příkopě28, 11503 Prague 1, Czech Republic

d

European Commission, DG Economic and Financial Affairs, Rue de la Loi 170, 1049 Brussels, Belgium

ABSTRACT

We identify a set of ‘rules of thumb’that characterize economic, financial and structural conditions preceding the onset of

banking and currency crises in 36 advanced economies over 1970–2010. We use the classification and regression tree

methodology and its random forest extension, which permits the detection of key variables driving binary crisis outcomes,

allows for interactions among key variables and determines critical tipping points. We distinguish between basic country

conditions, country structural characteristics and international developments. We find that crises are more varied than they

are similar. For banking crises, we find that low net interest rate spreads in the banking sector and a shallow, or inverted, yield

curve is their most important forerunners in the short term. In the longer term, it is high house price inflation. For currency crises,

high domestic short-term rates coupled with overvalued exchange rates are the most powerful short-term predictors. We find that

both country structural characteristics and international developments are relevant banking-crisis predictors. Currency crises,

however, seem to be driven more by country idiosyncratic, short-term developments. We find that some variables, such as

the domestic credit gap, provide important unconditional signals, but it is difficult to use them as conditional signals and, more

importantly, to find relevant threshold values. Copyright © 2016 John Wiley & Sons, Ltd.

Received 17 July 2015; Revised 17 June 2016; Accepted 27 September 2016

JEL CODE: C14; E44; F37; F47; G01

KEY WORDS: Banking crises; currency crises; binary classification tree; early warning indicators

1. INTRODUCTION

Until recently, most of the empirical research on predictors or determinants of financial crises has focused on

emerging market economies. The US subprime crisis and the euro area debt crisis have awakened interest in

systemic approaches to the early identification of crises in advanced countries. Compared with emerging market

economies, which are often characterized by economic and financial market volatility in the run-up to crises, the

pre-crisis conditions in advanced countries are often much smoother, making the identification of robust early

warning signals more challenging. This is further complicated by the fact that there is substantial disagreement over

the dating of crisis periods. Babecký et al. (2014) find this for advanced economies, while Arteta and Eichengreen

(2000) show that for emerging-market economies, dating banking crises can be just as problematic. On the other

hand, advanced countries are arguably more homogeneous than emerging market economies in terms of their

economic characteristics, which may improve the reliability of crisis signals. Yet, higher homogeneity does not

*Correspondence to: Bořek Vašíček, European Commission, DG Economic and Financial Affairs, Rue de la Loi 170, 1049 Brussels, Belgium.

†

E-mail: borek.vasicek@gmail.com

‡

In memoriam.

We devote this paper to the memory of Kateřina Šmídková, who started working on it but did not live to see its completion. The opinions

expressed in this paper are solely those of the authors and do not reflect the views of the Bank of England, the Czech National Bank and the

European Commission.

Copyright © 2016 John Wiley & Sons, Ltd.

INTERNATIONAL JOURNAL OF FINANCE & ECONOMICS

Int. J. Fin. Econ. 22:44–67 (2017)

Published online 7 November 2016 in Wiley Online Library

(wileyonlinelibrary.com). DOI: 10.1002/ijfe.1570

imply necessarily that all advanced country crises will be alike. This represents additional challenges for the early

warning literature given that it aims to identify common drivers of different periods of economic and financial

turmoil. Indeed, regression-based early warning models are based on the strong assumption that the marginal

contribution of each indicator to the probability of crisis does not depend on the value of the indicator for all

countries, over all time periods.

The purpose of this paper is to identify a set of economic ‘rules of thumb’that characterize economic, financial

and structural conditions preceding the onset of financial crises in 36 advanced countries (EU countries and

non-EU OECD countries), between 1970 and 2010. We use a quarterly database of crises in advanced countries

constructed by Babecký et al. (2014), investigating banking and currency crises separately.

1

We use the classifica-

tion and regression tree (CART) methodology, advanced by Breiman et al. (1984), specifically the binary

classification tree approach. This permits the detection of key variables driving crises, allows for interaction effects

and determines critical tipping points.

2

The framework has both advantages and disadvantages compared with

other common early warning methods, namely, the discrete choice models (logits or probits). On the one hand,

it allows explicitly for the fact that not all crises are alike, and accommodates nonlinearities by including variables’

conditional thresholds. This advantage should not be understated. Unlike logit models, which do not provide

policymakers with easily actionable advice on when to act to prevent a crisis (marginal effects provide a continuum

of probabilities rather than a yes–no recommendation on action), CART is able to tell the policymaker that while

some crisis indicators may be indicative of the average propensity for crisis, the specific conditions of the

policymaker’s country are important and under these specific conditions, certain crisis indicators are more reliable

than others. CART helps us to identify, for instance, that while rapid house price inflation is a good predictor of

banking crises 2–3 years ahead, banking crises can also be consistent with sluggish house price growth, if short-

term interest rates are low and the yield curve flat (almost half the banking crises in our sample can be identified

as having these characteristics). On the other hand, CART is a non-parametric approach, so it cannot calculate

the marginal contributions of each explanatory variable or confidence intervals for the estimated thresholds. While

typical early warning analyses look for indicators that are unconditional triggers of crises, in the logic of the CART

approach, an indicator can become a trigger only when it breaches a certain threshold or when it interacts with

another indicator. This makes the comparability of CART with traditional regression approaches complicated,

although they can be complements in the policymaker’s toolkit. In addition, we use the random forest (RF)

algorithm, which is an extension of CART, to overcome some of the weaknesses of the CART approach.

Our contributions are as follows: (i) we look at advanced countries given that advanced countries were at the

epicentre of the global financial crises, whereas most studies centre their analysis on emerging market economies,

which are arguably substantially more heterogeneous with more dispersed economic developments; this might be

problematic, especially when one wants to identify a common threshold across countries; (ii) we use a quarterly

dataset, which allows us to provide more detailed crisis-dating information, especially when it comes to

distinguishing between crisis onset and crisis occurrence; (iii) we extend the common list of leading indicators

beyond domestic macro-financial variables (i.e. the core indicators providing ‘differential diagnostics’)by

including (1) domestic factors that have significant cross-country variation but vary significantly less over time

and can, as such, be seen as structural characteristics of the economy (e.g. economic openness, exchange rate

regime or financial development) so as to further deal with country heterogeneity in a more explicit way than

the common fixed-effects approach, and (2) international factors that can, in addition to having a direct effect

on the domestic economy, contribute to crises indirectly by interacting with domestic variables (e.g. commodity

prices, global GDP growth and global private credit)

3

; and finally (iv) we focus on both banking crises and

currency crises, and by doing so, we obtain an insight into common determinants.

It should be noted that our results, like the results of any early warning model, are conditioned on the country

sample, time span and predictors. We do not claim to identify causality. Our results should be considered mainly as

a structured presentation of past experience and should not, without due care, be used for predicting future crises.

Our results feature a number of interesting findings: (i) we find that a high net interest rate spread in the banking

sector (i.e. the spread between the loan and deposit rate) combined with a flat or inverted yield curve are the most

reliable indicators of banking crises 1–2 years ahead of a crisis; we interpret this as evidence in favour of the

hypothesis put forward by Adrian et al. (2010) that the term spread can be thought of as representing the marginal

profitability of bank lending, and compression of the term spread, at the peak of a banking boom, can be a causal

BANKING AND CURRENCY CRISES 45

Copyright © 2016 John Wiley & Sons, Ltd. Int. J. Fin. Econ. 22:44–67 (2017)

DOI: 10.1002/ijfe