Currency Crises in Central and Eastern Europe: Does an Early Warning System work?

(1)

Currency Crises in Central and Eastern Europe:

Does an Early Warning System work?

Tomasz J. Pieśla, s1281453

Master Thesis General Economics, Rijksuniversiteit Groningen Supervisor: Jan P.A.M. Jacobs

October 2009

Abstract

In this paper I develop an Early Warning System for eight countries in Central and Eastern Europe, using a combination of a logit model and factor analysis, which is based on the approach used by Jacobs, Kuper and Lestano (2008). For the period 1992:1-2008:12 I find twenty currency crisis observations spread over the entire sample, with each country having experienced at least one crisis. Furthermore, I find that some indicators, such as growth of foreign exchange reserves, world oil prices, ratio of fiscal balance to GDP and excess real money balances seem to function as early warnings for some countries. However, the disappointing out-of-sample performance of the model suggests the good in-sample performance should be interpreted with caution.

Keywords: currency crises, early warning system, factor analysis, logistic regression.

(2)

1

1. Introduction

From the most recent countries that joined the European Union (EU), eight1 remain to adopt the euro as their currency. These countries, for now, have their own currencies that are either free floating against the euro, or in some form are pegged to the euro or a basket of currencies.

These developing economies are, like any other (developed) economy, at risk of being hit by a financial crisis, independent from whether the source of the financial crisis is domestic or foreign. In the financial crisis literature three types of crises can be distinguished, i.e. Currency Crises (CC), Banking Crises (BC) and Debt Crises (DC). The EU8 are, of course, susceptible to all three types of crises. However, these countries are especially vulnerable to the first type of crises because they have their own currency2. I will focus this paper on the currency crises only. To my knowledge there have been no studies for this region on currency crises, except for Stavárek (2007) and Vanneste, Van Poeck and Veiner (2007).

Considering the financial cost of such a crisis, it would be preferable to be able to somehow predict these potential currency crises. The EU8 are planning to enter the Exchange Rate Mechanism (II) or are already in it. One would like to avoid situations similar to the ERM-II crisis in 1993, where fluctuation bands had to be widened, and the UK had to abandon the ERM-II entirely.

First, I will determine the number of currency crises in the period 1992:1-2008:12, using an adjusted currency crises dating method developed by Kaminsky, Lizondo and Reinhart (1998). Second, I will develop an Early Warning System for these eight countries. The set-up of my EWS is based on Jacobs, Kuper and Lestano (2008). It is a combination of a logit model and factor analysis, to reduce the large number of indicators. Finally, I will test the performance of the estimated model both in-sample and out-of-sample for the periods 1992:1-2007:12 and 2008:1-2008:12 respectively.

I find a total of 20 currency crisis periods in the entire sample of eight countries. Furthermore, important early warning indicators for Central Eastern European Countries (CEECs) are: 1) OECD GDP growth; 2) growth of world oil prices; 3) GPD per capita growth; 4) growth of industrial production; 5) inflation rate; 6) ratio of fiscal balance to GDP; 7) excess real money balances; 8) growth of foreign exchange reserves; 9) domestic real interest rate. The logit model performs quite well in-sample, but out-of-sample the performance is not good. Due to the

1

These countries are: Bulgaria, Czech Republic, Estonia, Hungary, Latvia, Lithuania, Poland and Romania. In the remainder of this paper these counties will be referred to as EU8.

(3)

2 bad out-of-sample performance the results and in-sample performance should be interpreted with caution.

The remainder of this paper is organized as follows; Section 2 describes the existing literature on dating currency crises and Early Warning Systems, Section 3 describes the methodology, Section 4 describes the data, Section 5 gives the results and Section 6 summarizes and concludes.

2. Literature review

The theoretical and empirical literature on currency crises has developed tremendously in the past three decades. Ever since Krugman (1979) introduced his model on balance-of-payments crises, there has been growing interest in the topic. Krugman assumes that a country pegs its exchange rate solely through direct intervention in the foreign exchange market. A government can defend the peg until the borrowing limit is reached – through depletion of reserves – or as long as the cost of inflation of the depreciating currency is acceptable. Krugman argues that when it becomes known that a certain government is no longer willing or able to defend its currency, there is a sudden speculative attack against that currency that leads to almost full depletion of the reserves. He concludes that balance-of-payments crises are a consequence of maximizing behaviour of investors.

(4)

3 The extensive research has led to several definitions to what exactly a currency crisis is. Furthermore, there are a number of methods to determine a currency crisis. And last but not least, there are also different so-called Early Warning Systems, such as signalling approach and limited dependent regression models. Therefore, this section will be further divided into three parts. The first part will cover the different currency crises identification methods, the second part will discuss the existing EWS models and part three will describe currency crises studies in Central and Eastern Europe.

2.1. Currency crises dating methods

The literature on dating methods of currency crises can, in general, be divided into two categories, event studies and studies that focus on whether or not a certain threshold is exceeded of the so-called exchange market pressure index (EMPI). Examples of the former are events such as abandoning a peg, devaluations, suspension of convertibility, exchange rate regime changes and currency crises in neighbouring countries. Glick and Rose (1998) for example, study contagion effects for large set of developing and industrial countries. Other examples of event studies are of Bussière and Mulder (2000) and Sachs, Tornell and Velasco (1996).

The exchange market pressure index method is based on the monetary model of exchange market pressure of Girton and Roper (1977). This model formed the basis of the EMPI of Eichengreen, Rose and Wyplosz3 (1995), who defined the EMPI as a weighted average of changes in the nominal exchange rate, the ratio of gross international reserves to the monetary base (M1) and nominal interest rates, where all variables are measured relative to a reference country with a strong currency that functions as an anchor for other countries.4 The EMPI is weighted, so that the conditional volatilities of the three components are equal. Otherwise, the index would be dominated by the percentage changes of the reserve changes, which have the highest volatility of the components. According to the definition of ERW a currency crisis occurs when the index is two standard deviations above the mean of the index.5 The advantage of this method is that it both captures speculative attacks that lead to a devaluation – through the changes in exchange rate – as well as speculative attacks that are defended successfully, through changes in reserves and interest rates. Pozo and Amuedo-Dorantes (2003) criticize the use of standard deviation as a threshold, because the EMPI is assumed to be normally distributed. They test for

3_{From this point Eichengreen, Rose and Wyplosz will be referred to as ERW.} 4

ERW use Germany as their reference country. However, depending on the region under investigation one could use either the US or the Euro Area as reference countries.

(5)

4 normality of the EMPI and reject it for their whole sample and also for three regional subsamples. They develop an alternative method of dating currency crises – which is more sensitive than that of ERW- based on extreme value theory.

Currency crises can be persistent. Because the EMPI is measured for each month in the sample it can exhibit currency crises in subsequent months. It is fair to assume that when two crises occur in two months in a row, the second crisis observation is the same crisis from the month before. The same holds for currency crisis observation a few months further down the line, where the crisis observation can be interpreted as an aftermath of the original crisis a few months earlier. Therefore, ERW set an exclusion window of twelve months to avoid this double counting problem.

A far more straightforward definition of a currency crisis is proposed by Frankel and Rose (1996). A ‘currency crash’ is defined as a nominal depreciation of the currency of at least 25% that is also at least a 10 % increase in the rate of depreciation. The first threshold is set arbitrarily. The second threshold is also set arbitrarily; however, it ensures that for countries with high inflation – and therefore high expected rates of depreciation – these depreciations are not consistently taken into account as a crash. The ‘exclusion window’ is set to 3 years to avoid double counting. They also exclude unsuccessful speculative attacks from their measure, because they argue that gross international reserves are a noisy measure for exchange market pressure for almost all countries and especially for developing countries. Frankel and Rose (1996) also exclude the interest rate differential, because most countries in their sample lack market-determined short-term interest rates with long histories.6 In addition, they argue that the ‘standard’ instruments for defending speculative attacks, i.e. interest rates and reserves, do not apply in developing countries.

Kaminsky, Lizondo and Reinhart7 (1998) define a currency crisis in a similar way to ERW. However, they drop the reference country and the percentage changes in interest rates, so that the EMPI consists of two components opposed to ERW’s three. To avoid mixing times of hyperinflation with currency crises KLR divide their sample into ‘tranquil’ times and hyperinflation8 times, for which two separate indexes are created with a unique mean and variance. A currency crisis occurs when the index is three standard deviations above the mean of each subsample index.

Lestano and Jacobs (2007), compare the methods described above for the Asian crisis.

6

This tends to be true for a lot of developing countries. 7_{Henceforth abbreviated as KLR.}

(6)

5 They define the EMPI according to ERW and KLR and compare the number of crisis observations using ad hoc and extreme value-based thresholds, following the normal distribution critique posed by Pozo and Amuedo-Dormantes (2003). They evaluate their results using the IMF chronology of the Asian crisis in 1997-1999. Lestano and Jacobs (2007) find that the ERW index is less sensitive to different threshold than the KLR index. Based on the assessment of the IMF chronology, they opt for the KLR index with an extreme value threshold as the ‘best’ currency crisis dating method, because it generates the most correct crisis events. However, it also produces the most false warning signals.

2.2. Early Warning System Models

The main goal of a EWS is to be able to predict a currency crisis ahead of time and perhaps to take some measures in order to prevent it, if possible. Furthermore, with an EWS model it is possible to determine which economic indicators are driving forces behind a currency crisis. In the EWS literature three models can be distinguished.

The signal approach is a bivariate method. The behaviour of each variable is compared in periods in advance of the crisis with periods of tranquillity. If the behaviour between these two periods is different, then extreme values of these variables can be considered as a warning signal. The value is ‘extreme’ when it exceeds a certain threshold. This threshold can be uniform for the entire sample or it can be country specific. For example, KLR define the threshold in a relation to percentiles of the distribution of the indicator. A possible set of country-specific thresholds for the growth rate of exports for example, would be the set of growth rates (one per country) that would leave 10 percent9 of the observations above the threshold for each country. Using this definition, the 10 percent can be uniform for all countries, but the growth rates will differ per country. Given the warning signals of individual variables, a composite leading indicator can be constructed as a weighted average of the individual signals. Furthermore, each variable, including the dependent variable or crisis indicator, is transformed into a dummy, i.e. it either sends a signal or it does not. Vlaar (2000) points out that – in case of the crisis indicator – when using a dummy that either is or is not above a certain threshold, one misses information about the severity of the crisis. In addition, he argues that in the case of the indicators – which are treated the same way – this approach is inefficient, because their might be a relationship between the extremity of the indicator and the severity of the crisis. Especially when constructing a composite leading indicator this inefficiency could lead to inferior results. The weights of the signals in the leading indicator form another problem in the signal approach if the signals are correlated. However, this

(7)

6 method is very useful in finding the economic indicators that cause currency crises, without being particularly interested in the probabilities of the crises occurring. Kaminsky, Lizondo and Reinhart (1998) use the signal approach for a set of 15 developing and 5 industrial countries in the period of 1970:1-1995:12. Their currency crisis dating method is described in the previous subsection. They compare levels in 24 months preceding the crisis with values in tranquil periods for 15 variables, based on economic priors and data availability. They find that the variables with the most explanatory power are: 1) the real exchange rate deviation from trend; 2) the occurrence of a banking crisis; 3) the export growth rate; 4) the stock price index growth rate; 5) M2/reserves growth rate; 6) output growth rate; 7) excess M1 balances; 8) growth of international reserves; 9) M2 multiplier growth; and 10) the growth rate of domestic credit to GDP ratio.

Limited dependent regression models are the second category in the EWS. They are better known as logit or probit models. In this approach the crisis indicator takes on values of zero and one, as in the signal approach. The explanatory variables however, are no longer transformed into dummy variables. They are grouped together in a linear fashion. The predicted outcome of the model is always between zero and one, which is ensured by the logit and probit functions. The fact that the outcome can be interpreted as the probability of a crisis occurring is one of the major advantages of the model over the signal approach. Another advantage is that because all variables are pooled, it is easy to measure the additional information of a newly added variable on the whole model. A drawback of this method is that the single influence of a variable is difficult to measure. Frankel and Rose (1996) use a bivariate probit model to estimate the probability of a currency crash for a set of 100 developing countries in the period 1972-1992. They find that currency crashes tend to occur when 1) output growth is low; 2) the growth of domestic credit is high; 3) the level of foreign interest rates is high; 4) Foreign Direct Investment inflows dry up; 5) reserves are low; and 6) the real exchange rate shows overvaluation. Berg and Pattillo (1999) also use a probit model for almost the same dataset used by KLR.10 They compare their method to that of KLR and find that the probit-based model generally gives better predictions than the signal model. In addition to KLR they report that large current account deficits and a high ratio of M2 to reserves are important risk factors. Jacobs, Kuper and Lestano (2008) use a multivariate logit model for six Asian counties to predict currency crises. They reduce the number of variables by combining them in a linear fashion into factors. Using the combination of factor analysis and logit modelling allows them to draw the conclusion that some indicators of currency crises do work, at

10

(8)

7 least for their sample of six Asian countries. This contradicts earlier research of the IMF (2002) and Edison (2003) who find that EWS have generally a rather mixed and weak performance. Jacobs, Kuper and Lestano (2008) find that in general 1) growth rates of money, i.e. M1 and M2; 2) GDP per capita; 3) national savings; and 4) imports correlate with currency crises. They also find country specific indicators such as growth rates of commercial bank deposits, foreign exchange reserves, exports, domestic real interest rates, terms of trade, and world oil price changes.

The third approach focuses on predicting the severity of the crises and not so much on predicting the timing. The main question that is to be answered in this type of model is given the occurrence of a currency crisis somewhere in the world, which countries are going to be hit hardest. To measure the severity of a crisis, a crisis index is constructed – for example based on nominal exchange rate fluctuations and international reserves – for each country during the entire period of financial stress. The differences between countries in the magnitude of the crisis index are subsequently explained by variables representing the economic situation at the onset of the crisis. Sachs, Tornell and Velasco (1996) use this framework to examine the severity of currency crises during the Mexican crisis in the end of 1994. Their main result is that only countries that were already vulnerable were hit by the Mexican crisis. Countries seem to be vulnerable when the real exchange rate is overvalued, banks are weak and international reserves are low. Berg and Pattillo (1998) test this approach by using Asian crises data and find that adding a few countries to the sample changes the coefficients of the variables significantly, probably due to small sample problems. They claim that the applicability of the model as a predictive device is doubtful, because of the large effects of small changes in the definitions of the explanatory variables. Bussière and Mulder (1999) confirm the problems mentioned above when investigating the Russian crisis in 1998 as a consequence of the 1994 Mexican crisis and the 1997 Asian crisis. Glick and Rose (1998) investigate the effects of contagion using a sample of both developing and industrial countries for five currency crises. They find strong evidence for trade linkages to be explanatory for regional patterns of currency crises for the five periods under investigation. Domestic variables do not seem to explain cross-country incidence of speculative attacks. However, they find no causality in the effects of trade on crises, because it is the only variable that is included in the regression.

2.3. Currency crises in Central and Eastern Europe

(9)

8 on early warning systems of currency crises in this region is scarce. However, Stavárek (2007) and Vanneste, Van Poeck and Veiner (2007) describe the implications of the exchange market pressure and exchange rate regimes in the new EU member states. Vanneste et al. (2007) use the EMPI defined by ERW and look for the number of crisis events during different exchanger rate regimes. They find that the so-called bipolar view of exchange rate arrangements and currency crises, where an adjustable peg is more vulnerable to currency crises than credible fixed peg or free floating regimes, seems to hold for the CEECs in the period 1990-2003. They argue that these countries should not enter the ERM II scheme before their fundamentals are in the ‘safe’ region so that they are a less candidate for speculation.

Stavárek (2007) uses a model-dependent and a model-independent approach to calculate the EMPI for four CEECs in the period 1993-2006. He finds that the two models are not compatible and provide opposite results. The relationship between EMPI and exchange rate regime is extremely sensitive to selection of EMPI estimation method. Stavárek (2007) does not find confirmation that the concerns that the unavoidable shift in the exchange rate regime towards a quasi-fixed ERM II could evoke EMPI to grow to excessive levels.

3. Methodology

Considering the aforementioned methods of dating currency crises and their performance in earlier studies it seems appropriate, for determining the EMPI, to use the composition proposed by KLR and include the nominal interest rates. The ERW method seems less appropriate, because of the reference country. It is fair to say that for the EU8 the reference countries would be Germany in the 90s and the Euro Area from 1999. This ‘break’ in reference country, in my opinion, provides sufficient reason to drop ERW’s method entirely. Furthermore, to define a currency crisis, I will use a threshold of two standard deviations above the mean of the index.11 I make no distinction between periods of hyperinflation and low inflation.

The EMPI can be described by the following formula:

t i i e t i t i r e t i t i t i i r r e e EMPI _, , , , , , + ∆ ∆ − ∆ =

σ

, (1)

(10)

9 where EMPIi,t is the exchange rate market pressure index of country i in period t, ei,t is the number of LCU of country i per ECU/EUR in period t, (∆ei,t/ei,t) is the relative monthly change of the nominal exchange rate, σe is the standard deviation of the relative change of the exchange rate; ri,t is the ratio of gross international foreign reserves to monetary base of country i in period t, (∆ri,t/ri,t) is the relative change of the ratio of gross international foreign reserves, σr is the standard deviation of that relative change; ∆ii,t is the change in nominal interest rate of country i in period t, and σi is the standard deviation of the change in nominal interest rate.

A crisis occurs when the EMPI is two standard deviations above its mean, which translates into the following:

(

)

   − > = , 0 2 / 1 _, Otherwise EMPI if

Crisis it

µ

EMPIi

σ

EMPIi

(2)

where EMPIi,t is again the exchange market pressure index of country i in period t, µEMPIiis the mean of the EMPI of country i and

i

EMPI

σ

is the standard deviation of the EMPI of country i. As mentioned before, an exclusion window of twelve months is applied. This can decrease the number of crisis periods and has important implications for the balance between the ones and zeros in the sample. Berg and Pattillo (1998) for example not only define the moment of crisis as ‘one’, but also the 23 months prior to the crisis. One major advantage of this method is that it improves the balance between the ones and zeros in the series. Furthermore, this approach implies that a crisis can be ‘spotted’ two years in advance. Since the number of crises is limited in my sample, I will also use the same approach as Berg and Pattillo (1998) and compare eventually what effects both approaches have on the EWS model.12

3.1. Factor analysis

The large number of economic indicators makes it impossible to do a meaningful regression when including all of the indicators at the same time, because of the multicollinearity between the indicators. Factor analysis provides a practical solution for this problem. Factor analysis is used, in general, for two reasons, i.e. 1) to reduce the number of variables and 2) to detect structure in the relationship between variables, or to classify variables. I will use factor analysis for data reduction. The main idea is combining two or more correlated variables and let these variables be represented by a factor (regression line) that maximizes the variance. Because

(11)

10 one factor does not cover the entire variance, more factors are extracted until all the variance is explained. These factors have the role of explanatory variables in the logit model.

There are several methods to determine the number of factors one wants to use, because of the declining addition to the total variance with each additional factor. The one most often used is the so-called Kaiser (1960) criterion, which includes all factors with an eigenvalue above one. Another method is the Cattell (1966) scree test. This is a test based on a graphical representation of the factors – on the horizontal axis – and the eigenvalues, which are on the vertical axis. The number of factors is determined by the point in the graph where the smooth decrease of the eigenvalues levels out. Although the Kaiser criterion is the most widely used method, I will use both criteria to determine the number of factors.

I will apply two ways of grouping the indicators, i.e. one where all economic indicators are included to determine a number of factors; and a second one where the economic indicators are grouped into external, domestic and global indicators, similar to the grouping in the data description. The advantage of the former is that the factors are uncorrelated, as opposed to the latter. The factors are determined on both ungrouped and grouped indicators for each country separately and also for a stacked panel data.

3.2. Logit model

There are two most widely used types of limited dependent regression models, i.e. the so-called logit and probit model. Both have a binary dependent variable – in this case the currency crisis variable – which has a value of one in case of a crisis and zero otherwise. Although both models provide the probability of a currency crisis, there are differences. For example, the probit model is based on the normal probability density function, whereas the logit model has an S-shaped logistic function. The latter ensures that the probabilities are constrained to an outcome between zero and one. The logit model is specified as follows:

(

)

( )

_Z ₍ _X₎ e e Z F Z P ₋ ₋_{α +}_β + = + = = = 1 1 1 1 1 , (3)

where P(.) is the probability of Z = 1, i.e. a currency crisis occurring, F(Z) is the cumulative logistic probability function, X is the set of regressors13 and α and β are parameters. Following from the above model specification the regression equation takes the following from:

(12)

11 X Z P P Ln = =

α

+

β

     − 1 . (4)

3.3. Performance of the model

The outcomes of the model represent probabilities of a currency crisis. A high probability suggests a crisis and a low probability signal a tranquil period. Kaminsky, Lizondo and Reinhart (1998) use a signal-to-noise ratio to test the performance of their model. They start with specifying the possible outcomes of the model, i.e. there is a signal of a crisis and the crisis occurs (P(1,1)); there is a signal and no crisis occurs (P(1,0)); there is no signal and a crisis occurs (P(0,1)); and there is no signal and no crisis occurs (P(0,0)). The first and last are correct predictions and the other two are wrong predictions. The signal-to-noise ratio is determined by the number of correct predictions over the number of wrong predictions. A ratio above ‘1’ indicates that at least half of the crises are predicted correctly. Table 1 gives the four possibilities.

Table 1: The probabilities of right and wrong crisis predictions

Crisis (Z = 1) No crisis (Z = 0)

high P(1,1) P(1,0)

Estimated probability

low P(1,0) = 1- P(1,1) P(0,0) = 1- P(1,0)

Diebold and Rudebusch (1989) propose a different method of measuring the goodness of fit of a forecasting model. These are the so-called quadratic probability score (QPS) and log probability score (LPS). Both methods measure the closeness, on average, of predicted probabilities and observed realisations, as measured by a zero-one dummy variable. The QPS works as follows: first, there is a time series of T probability forecasts

{ }

P_t T_t₌₁, where Ptis the probability of a currency crisis occurring at time t; second there is a time series

{ }

Zt T_t=1of the observed realisations, where Zt equals one if there is a currency crisis and zero otherwise. The QPS equation takes the following form:

(13)

12 The QPS ranges from 0 to 2, where zero indicates perfect accuracy of the model. Moreover, the QPS has the desirable property of being strictly proper, meaning that it achieves a strict minimum under truthful revelation of probabilities by the forecaster. In addition, it is the unique proper scoring rule that is a function only of the discrepancy between realisations and assessed proabilities.

The log probability score (LPS) considered by Diebold and Rudebusch (1989) is also a strictly proper scoring rule. The LPS is given by:

(

) (

)

( )

[

]

∑

= + − − − = T t t t t t P Z P Z T LPS 1 ln 1 ln 1 1 . (6)

The LPS has a range from 0 to ∞, where zero indicates perfect accuracy of the model. The LPS depends exclusively on the probability forecast of the event that actually occurred, assigning as a score the log of the assessed probability. In two event of crises, i.e. crisis (Z = 1), no crisis

(Z = 0), the LPS is a fully general scoring rule, because the probability forecast of a crisis (Pt) implicitly determines the probability forecast of a tranquil period (1 – Pt). The loss function associated with LPS differs from that corresponding to QPS, as large mistakes are penalised more heavily under LPS.

4. Data

The data are collected for eight countries in Central and Eastern Europe, i.e. Bulgaria, Czech Republic, Estonia, Hungary, Latvia, Lithuania, Poland and Romania over a period of 1992:1-2008:12. The in-sample period for the logit model is 1992:1-2007:12 and the out-of-sample period for testing the performance of the logit model is 2008:1-2008:12. Most of the data are from International Financial Statistics (IFS) of the International Monetary Fund (IMF), the Bank for International Settlements (BIS) and Eurostat. Furthermore, additional data were collected from the Central Banks’ websites. I will use the following set of economic indicators, which can be categorized into three distinct groups.14

(14)

13

External: real exchange rate (REX), export growth (EXG), import growth (IMP), ratio of the current account to GDP (CAY), the ratio of M2 to foreign exchange reserves (MFR), and growth of foreign exchange reserves (GFR).

Domestic (Financial, real and public): M1 and M2 growth (GM1 and GM2), M2 money

multiplier (MMM), the ratio of domestic credit to GDP (DCY), excess real M1 balances (ERM), domestic real interest rate (RIR), lending and deposit rate spread (LDS), commercial bank deposits (CBD), and the ratio of bank reserves to bank assets (RRA), the ratio of fiscal balance to GDP (FBY), the ratio of public debt to GDP (PBY), growth of industrial production (GIP), inflation rate (INR), GDP per capita (YPC), and growth of national savings (NSR).

Global: the growth rate of world oil prices (WOP), the US interest rate (USI), and OECD GDP growth (ICY).

The choice for these indicators is based on previous research and economic theory. The full list of indicators, their source and transformation can be found in the appendix. An important note is that the set of indicators differs slightly from the set used by Lestano et al. (2003). First of all, there are no data available for terms of trade and stock prices for the majority of the countries in the sample. Therefore, these indicators are dropped. Furthermore, seven15 indicators have a unit root for some countries, which may lead to spurious results in regression analysis. For that reason all indicators – except the real exchange rate (REX) – are transformed into twelve month percentage changes.

For some indicators monthly data were partly or not available. Therefore, the series were interpolated from annual and quarterly data where possible. As a result the dataset is not balanced for all indicators and countries.

(15)

14

5. Results

This section will be divided in three parts. First, I will present the number of crises in the sample period and the results of the factor analysis. Second, I will present the results of the logit regression. Finally, I will test the in-sample and out-of-sample performance of the model using the aforementioned QPS and LPS measures.

5.1. Crisis periods and factor analysis

Figure 1 shows the graphical representation of the Exchange Market Pressure Index of the eight CEECs. The crisis periods can be easily spotted looking at ‘peaks’ above the standard deviations threshold. For Estonia and Latvia it is also easy to see that there are several crisis periods close to each other. These subsequent crisis periods will be removed from the sample.

Figure 1: Exchange Market Pressure Index for all countries

(16)

15 The second column of table 2 presents the number of crisis periods using the Regular Approach and the third column presents number of crisis periods using the extended Berg-Pattillo Approach.

Table 2: Number of Crises in the Period 1992-2008

Country Crises (% of sample)* Crises (% of sample)**

Bulgaria 1 (0.98%) 13 (12.7%) Czech Republic 2 (0.98%) 26 (12.7%) Estonia 3 (1.96%) 39 (19.1%) Hungary 4 (1.96%) 52 (25.5%) Latvia 3 (1.96%) 39 (19.1%) Lithuania 2 (0.98%) 26 (12.7%) Poland 2 (0.98%) 16 (7.8%) Romania 3 (1.47%) 39 (19.1%) Total 20 (1.23%) 287 (15.3%)

* Number of crisis periods in the Regular Approach ** Number of crisis periods in the Berg-Pattillo Approach

For the factor analysis I use two ways of grouping the economic indicators, i.e. one where all indicators are included (ungrouped) and one where the indicators are categorized into

domestic, external and global (grouped) indicators. As mentioned before, the difference between the two ways of grouping is that for the former the factors are uncorrelated as opposed to the latter. Table 3 presents the paired correlation coefficients for factors of the grouped indicators.

Table 3: Correlation coefficients for grouped factors

Factor External 1 External 2 Domestic 1 Domestic 2 Domestic 3 Global 1

External 1 1.000 External 2 -0.002 1.000 Domestic 1 -0.112 0.379 1.000 Domestic 2 0.034 -0.125 -0.012 1.000 Domestic 3 0.240 0.228 0.023 0.017 1.000 Global 1 0.319 0.027 0.094 -0.013 0.038 1.000

(17)

16 Tables A1 and A2 (see appendix B) present the eigenvalues of factors for ungrouped and grouped indicators respectively. Tables A3 to A5 (see appendix B) present the country specific eigenvalues of factors for the ungrouped, external and domestic indicators respectively. The tables show only eigenvalues above one, which is the cut-off point of the so-called Kaiser (1960) criterion. Figure A1 (see appendix C) presents the scree plots for the ungrouped, external,

domestic and global indicators.

Figures A2 to A9 (see appendix C) present the country specific scree plots for ungrouped,

external and domestic indicators. Global indicators are not presented, because they are the same for all countries so there is no difference between country specific factors and those of a stacked panel. The scree plots are used for the aforementioned Cattell (1966) scree-test for factor extraction. Combining the Kaiser (1960) criterion with the Cattell (1966) scree-test, I extract six factors for the ungrouped indicators, two factors for the external indicators, three factors for the

domestic indicators and one factor for the global indicators. The results are similar for country specific factors.

5.2. Logit model output

Summarizing the previous section, there are two sets of crisis observations, i.e. the Regular Approach and the Berg-Pattillo Approach. Furthermore, there are two ways of grouping the indicators, i.e. ungrouped and grouped.

The logit model is estimated in two ways, i.e. one where each country is regressed separately with its own constant and regression coefficients; and another where the regression is made using stacked panel data with fixed effects, where the regression coefficients are the same for all countries. Combined with the previous section the regression results are presented in the eight logit model output estimates.

Three things are important in evaluating the regression output, i.e. first, the significance

(z-statistic) of the regression coefficients; second, the variance explained by the model (McFadden

R2); finally, the overall significance of the logit model (Likelihood Ratio statistic). Before discussing each table in detail it is worth mentioning that several tables can be discarded right away. For example, tables A6 and A7 (see appendix B) where the country specific logit regression model in the Regular Approach is insignificant, although there is at least one significant factor (slope coefficient) for each country for both ungrouped and grouped indicators. The LR statistic fails to exceed the critical value for all countries except for Romania.16 The LR

(18)

17

statistic tests the joint null hypothesis that all slope coefficients except the constant are zero. The overall significance of the country specific logit regression improves greatly for the Berg-Pattillo Approach (tables A8 and A9, see appendix B). However, the variance explained by the model (McFadden R2) does not differ much between the two approaches. The McFadden R2 is analogue to the R2 in a standard regression model. This ratio indicates to what extent the variance is explained by the explanatory variables, i.e. the factors. In my opinion, extending the number of crisis observations by 12 months is questionable when used solely for increasing the number of ones in the sample. This is especially true for relatively short samples, as is the case here. Based on this argument the results in tables A10 and A11 (see appendix B) – which represent the results of the stacked panel logit regression in the Berg-Pattillo Approach – can be discarded as well.

Table 4: Stacked panel logit output for factors with ungrouped indicators (Regular Approach)

Variable Coefficient z-Statistic

Factor 1 -0.14 -1.16 Factor 2 0.05 0.21 Factor 3 0.34 1.26 Factor 4 -0.29 -1.49 Factor 5 -0.36 -2.01 Factor 6 0.51 2.88 Bulgaria -5.08 -5.18 Czech Republic -4.48 -6.10 Estonia -4.25 -7.09 Hungary -4.67 -6.63 Latvia -4.56 -6.22 Lithuania -5.14 -5.03 Poland -4.61 -6.43 Romania -4.15 -8.07 McFadden R-squared 0.10 LR statistic 18.44 Observations with Z = 1 18

Note: The model is estimated with Huber-White robust standard errors. Critical values of the z-statistic at the 1% and 5% level are 2.57 and 1.96 respectively.The critical value of the likelihood ratio at 1% significance level (6 degrees of freedom) is 16.81

(19)

18 two crisis observations are out-of-sample. The LR statistic is 18.44, which rejects the null hypothesis that the slope coefficients are equal to zero at a 1% significance level. Furthermore, the explained variance (McFadden R2) is 0.10. Although usually a high (above 0.50) R2 is desired, in a logistic regression a McFadden R2 of 0.20 is often satisfactory. So clearly, in this logit regression the explained variance is too low. Table 4 shows that factor 6 has the largest slope parameter that is also significant. Although interpretation of the estimated coefficients in terms of the underlying indicators is not trivial, the eigenvector of factor 6 is informative, since factor 6 is a linear combination of the indicators with weights given by the sixth eigenvector. Table 5 presents the country specific eigenvectors.

Table 5: Loadings for the sixth factor that has the largest contribution to predicting crises probabilities

Indicator Bulgaria

Czech

Republic Estonia Hungary Latvia Lithuania Poland Romania

(20)

19 Since the weights are different for each country, I constructed a top ten of indicators with the highest positive weights, which is given in table 6.

Table 6: Country specific top ten of indicators with largest weights

Bulgaria

Czech

CAY INR FBY GIP FBY GIP USI GIP

FBY ERM GFR IMP IMP ICY ERM GM1

ICY MFR YPC GM1 NSR EXG CBD EXG

YPC GFR GM2 ICY GM1 WOP MFR ERM

GIP EXG ICY REX USI RRA NSR REX

NSR RIR CAY LDS INR RIR GFR NSR

MFR FBY ERM YPC CAY GFR IMP WOP

INR MMM RIR DCY WOP MMM GM2 RIR

MMM GIP NSR RIR PBY INR YPC PBY

EXG PBY CBD MFR YPC USI WOP ICY

The indicators that are in the top ten of four or more countries are: OECD GDP growth (ICY), GPD per capita growth (YPC), growth of industrial production (GIP), inflation rate (INR), growth of world oil prices (WOP), ratio of fiscal balance to GDP (FBY), excess real money balances (ERM), growth of foreign exchange reserves (GFR) and domestic real interest rate (RIR). Other, more country specific, dominant indicators are import growth (IMP), export growth (EXG), growth of money (GM1 and GM2), M2 money multiplier (MMM) and ratio of public debt to GDP (PBY).

(21)

20 fundamentals are weak. This supports the view of Stavárek (2007) and Vanneste et al. (2007) that CEECs should only enter the ERM II if their economic fundamentals are strong enough.

5.3. Performance of the model

The performance of the model is measured using QPS and LPS described earlier. I test the performance both in-sample and out-of-sample for the periods 1992:1-2007:12 and 2008:1-2008:12 respectively. As mentioned before, the number of crisis observations is limited in-sample. Out of sample the number is even less, since there are only two countries (Hungary and Latvia) where a currency crisis is observed. Table 7 presents the results of the performance of the logit model.

Table 7: Performance of the stacked logit model with ungrouped indicators

In sample Out-of-sample Country QPS LPS QPS LPS Bulgaria 0.014 0.044 0.000 0.010 Czech Republic 0.022 0.061 0.001 0.024 Estonia 0.035 0.087 0.003 0.037 Hungary 0.030 0.101 0.161 0.350 Latvia 0.022 0.076 0.159 0.318 Lithuania 0.012 0.038 0.001 0.025 Poland 0.019 0.054 0.000 0.012 Romania 0.032 0.069 0.002 0.026

Note: QPS and LPS stand for Quadratic Probability Score and Log Probability Score respectively

(22)

21

6. Conclusion

In this paper I apply an earlier developed EWS – by Jacobs et al. (2008) – for a set of eight countries – that are to adopt the Euro as their currency in the (near) future – in Central and Eastern Europe in the period 1992:1-2008:12. The underlying indicators are based on previous research. I use a binomial qualitative choice logit model in combination with factor analysis.

The above results suggest that for this set of countries there are some – for each country different – indicators that seem to contribute to predicting currency crises. These are: OECD GDP growth (ICY), GPD per capita growth (YPC), growth of industrial production (GIP), inflation rate (INR), growth of world oil prices (WOP), ratio of fiscal balance to GDP (FBY), excess real money balances (ERM), growth of foreign exchange reserves (GFR) and domestic real interest rate (RIR). Other, more country specific, dominant indicators are import growth (IMP), export growth (EXG), growth of money (GM1 and GM2), ratio of public debt to GDP (PBY) and M2 money multiplier (MMM). An indicator is included if it is in the top ten of indicators – in four countries or more – contributing to predicting currency crises. These results have implications on macro economic policy regarding the exchange rate. A government should keep its fiscal balance stable, together with a stable or declining public debt. These two also form the cornerstone of the growth and stability pact of the EMU regarding the admission to the Euro Area. It is obvious that countries have no influence on World oil prices and the OECD production. Therefore, it can be concluded that some of the countries in Central and Eastern Europe are very vulnerable to foreign threats. Furthermore, it is not very surprising to see growth of foreign exchange reserves (GFR) as an early warning indicator, since the size of reserves represents the ability to withstand speculative attacks.

Although the results are satisfactory with regard to establishing some early warning economic indicators, the overall performance should be interpreted with caution. The performance of the model – for countries that experienced a currency crisis in the out-of-sample period – is not good. This leads me to conclude that for the reasons described above, the good in-sample performance is not clear cut.

(23)

22

Appendix A: Explanatory Variables

Variable Code Defenition and source Transformation

External sector (current account)

Real exchange Rate

REX The nominal exchange rate is in local currency unit (LCU) per ECU/EUR (bilateral nominal exchange rate from EUROSTAT). The CPI is IFS line 64. The real exchange rate is the nominal exchange rate times the ratio of foreign (EA HICP) and domestic prices. REX = ePf/P, where e = nominal exchange rate, P = domestic CPI and Pf = Euro Area HICP. A decline in the real exchange rate denotes a real appreciation of the LCU.

Deviation from trend

Export growth

EXG IFS line 70 12 month

percentage change Import

growth

IMP IFS line 71 12 month

percentage change Ratio of the current account to GDP

CAY Current account (IFS line 78al) divided by nominal GDP (interpolated IFS line 99b).

12 month percentage change

External sector (capital account)

Ratio of M2 to foreign exchange reserves

MFR Ratio of M2 (IFS line 34 plus 35) and international reserves (IFS line 1L.D).M2 is converted into USD. 12 month percentage change Growth of foreign exchange reserves

GFR IFS line 1L.D. 12 month

percentage change

Domestic (financial, real and public)

M1 growth GM1 IFS line 34 12 month

percentage change

M2 growth GM2 IFS line 35 12 month

percentage change M2 money

multiplier

MMM Ratio of M2 (IFS line 34 plus 35) to base (reserve) money (IFS line 14).

(24)

23

Ecxess real M1 balances

ERM Percentage difference between M1 (IFS line 34)

deflated by CPI (IFS line 64) and estimated money demand for M1. Demand for real M1 is estimated as a function of real GDP, nominal interest rates (IFS line 60L), and a time trend. Real GDP is obtained by deflating the interpolated nominal GDP (IFS line 99B).

Based on estimated money demand equation Ratio of domestic credit to GDP

DCY Total domestic credit (IFS line 32) divided by nominal GDP (interpolated IFS line 99b).

12 month percentage change Domestic real interest rate

RIR 6 month time deposit rate (IFS line 60L) deflated by CPI (IFS line 64).

12 month percentage change Lending and deposit rate spread

LDS Lending interest rate (IFS line 60P) divided by 6 month time deposit rate (IFS line 60L).

12 month percentage change Commercial bank deposits

CBD Demand deposit (IFS line 24) plus time savings and foreign currency deposits (IFS line 25) deflated by CPI (IFS line 64).

12 month percentage change Ratio bank reserves to bank assets

RRA Bank reserves (IFS line 20) divided by bank assets (IFS line 21 plus IFS line 22a to IFS line 22f). 12 month percentage change Ratio of fiscal balance to GDP

FBY Government budget balance (IFS line 80) divided

by nominal GDP (interpolated IFS line 99b).

12 month percentage change Ratio of public debt to GDP

PBY Public and puvblicly guaranteed debt (World

Bank) divided by nominal GDP (interpolated IFS line 99b). 12 month percentage change Growth of industrial production

GIP IFS line 66. If not available, the index was obtained from the Central Bank site of that country.

12 month percentage change

Inflation rate INR IFS line 64. 12 month

percentage change GDP per

capita

YCP GDP (interpolated IFS line 99B) divided by total population (IFS line 99Z).

12 month percentage change National

savings

NSR Public (IFS line 91F) and private (IFS line 96F) consumption subtracted from GDP (interpolated IFS line 99B). 12 month percentage change Global economy Growth of world oil prices

WOP IFS line 176.AA 12 month

(25)

24

US interest rate

USI The US Treasury bill rate (IFS line 60C). 12 month

percentage change OECD GDP

growth

ICY Proxied by industrial production (IFS line 66). 12 month

percentage change

(26)

25

Appendix B: Tables

Table A1: Eigenvalues for factors with ungrouped indicators

Factor Eigenvalue Factor 1 4.847 Factor 2 2.619 Factor 3 1.822 Factor 4 1.578 Factor 5 1.410 Factor 6 1.239 h2 0.56

(27)

26 Table A2: Eigenvalues for factors with grouped indicators

Factor Eigenvalue Factor Eigenvalue Factor Eigenvalue

External 1 1.842 Domestic 1 4.205 Global 1 1.806

Domestic 4 1.126

Domestic 5 1.039

h2 0.76 0.70 1.00

(28)

27

Table A3: Eigenvalues of country specific Ungrouped Factors

Bulgaria

Czech

Republic Estonia Hungary Latvia Lithuania Poland Romania Factor Factor 1 5.310 5.390 7.377 5.894 8.205 4.838 7.163 7.454 Factor 2 4.850 4.325 6.160 4.184 4.347 4.379 3.654 3.466 Factor 3 2.509 2.815 2.855 2.581 3.233 3.236 2.927 2.360 Factor 4 1.996 1.980 1.659 1.732 1.984 2.754 2.431 2.196 Factor 5 1.625 1.748 1.196 1.585 1.209 1.637 1.621 2.091 Factor 6 1.520 1.462 1.131 1.306 1.084 1.467 1.349 1.388 h2 0.74 0.74 0.85 0.72 0.84 0.76 0.80 0.79

Note: h2 represents the cumulative sum of the variance proportion explained by each factor.

Table A4: Eigenvalues of country specific External Factors

Bulgaria

Czech

Republic Estonia Hungary Latvia Lithuania Poland Romania Factor

External 1 2.079 1.856 2.004 2.296 1.929 1.840 2.194 2.257

Externa 2 1.638 1.363 1.458 1.309 1.417 1.468 1.131 1.502

External 3 1.165 1.023 1.023 1.080 1.041

h2 0.62 0.54 0.77 0.77 0.73 0.73 0.73 0.63

Note: h2 represents the cumulative sum of the variance proportion explained by each factor.

Table A5: Eigenvalues of country specific Domestic Factors

Bulgaria

Czech

Republic Estonia Hungary Latvia Lithuania Poland Romania Factor Domestic 1 4.713 4.669 5.700 4.930 5.447 4.211 5.699 6.338 Domestic 2 2.785 3.342 3.389 2.958 3.552 3.497 2.531 1.984 Domestic 3 1.973 1.742 2.405 1.522 1.543 1.717 2.382 1.827 Domestic 4 1.259 1.267 1.362 1.198 1.489 1.210 1.114 Domestic 5 1.066 1.057 1.086 1.160 h2 0.79 0.81 0.77 0.72 0.86 0.80 0.79 0.75

(29)

Table A6: Logit output for factors with grouped indicators (Regular Approach) Bulgaria

Czech

Coefficient Z-statistic Coefficient Z-statistic Coefficient Z-statistic Coefficient Z-statistic Coefficient Z-statistic Coefficient Z-statistic Coefficient Z-statistic Coefficient Z-statistic

Constant -21.98 -2.58 -16.93 -3.12 -5.52 -5.51 -8.57 -2.33 -19.36 4.44 -33.36 -1.68 -30.51 -2.43 -12.52 -2.15 Factor 1 11.31 2.17 -2.16 -2.00 0.38 2.53 1.69 1.02 -0.71 0.65 -17.90 -1.58 -32.43 -2.13 -3.67 -1.87 Factor 2 4.65 2.56 1.22 1.83 1.67 6.67 2.77 1.51 9.34 3.02 8.64 1.28 -8.40 -2.08 -5.15 -1.75 Factor 3 2.30 1.64 -2.80 -4.46 0.61 2.41 11.70 3.63 -0.74 -2.63 Factor 4 4.22 1.84 -2.88 -2.01 1.32 1.91 -2.28 1.55 Factor 5 -1.34 -1.21 2.37 1.95 1.50 1.09 Factor 6 -8.01 -2.39 3.31 1.66 McFadden R2 0.36 0.58 0.10 0.32 0.50 0.66 0.48 0.47

Likelihood ratio statistic 3.89 12.35 1.15 9.91 5.37 7.60 5.71 13.85

Observations with Z = 1 1 2 1 3 1 1 1 3

Note: The model is estimated with Huber-White robust standard errors. Critical values of the z-statistic at the 1% and 5% level are 2.57 and 1.96 respectively.The critical values of the likelihood ratio test are 9.21, 11.34, 13.28, 15.09 and 16.81 for 2,3,4,5 and 6 degrees of freedom respectively at a 1% significance level.

Table A7: Logit output for factors with grouped indicators (Regular Approach) Bulgaria

Czech

Constant -6.08 -5.48 -5.26 -5.26 - - -6.56 -4.68 -4.72 -4.51 -9.35 -6.65 -10.85 -8.51 -17.86 -2.44 external 1 -0.09 -0.07 -0.31 -0.31 - - 2.60 2.51 -5.66 -4.37 -0.66 -1.60 2.06 5.08 -7.71 -2.56 external 2 -0.68 -0.56 0.09 0.04 - - -1.39 -2.33 -14.94 -4.21 6.47 4.64 8.08 6.39 -1.64 -0.94 domestic 1 1.97 5.51 -0.12 -0.82 - - -0.12 -0.09 2.32 3.31 3.38 3.36 4.00 5.40 -20.88 -1.82 domestic 2 1.34 3.19 -1.61 -4.96 - - -0.14 -0.24 4.40 4.53 -11.97 -1.15 domestic 3 -1.12 -3.34 - - 1.89 2.81 4.28 1.54 global 1 0.19 0.38 -0.16 -0.39 - - 0.83 0.56 -2.47 -2.18 1.57 3.61 -1.16 -2.82 -3.06 -0.98 McFadden R2 0.22 0.10 - 0.22 0.31 0.28 0.42 0.58

Likelihood ratio statistic 2.43 2.13 - 6.69 3.36 3.25 4.96 17.14

Observations with Z = 1 1 2 3 1 1 1 3

(30)

29 Table A8: Logit output for factors with ungrouped indicators (BP Approch)

Bulgaria

Czech

Constant -69.14 -2.46 -2.95 -6.23 -2.59 -6.94 -2.19 -6.67 -23.70 -2.94 -8.65 -2.72 -10.69 -3.02 -1.56 -7.91 Factor 1 70.75 2.53 1.39 2.97 0.12 0.30 0.83 4.05 14.97 2.41 1.61 1.61 -5.14 -1.67 -1.18 -3.70 Factor 2 13.02 2.74 -0.04 -0.20 2.64 3.93 0.05 0.19 7.84 2.76 -2.71 -2.09 2.02 1.59 -1.10 -5.54 Factor 3 6.54 2.58 -1.02 -2.79 1.72 5.41 9.15 2.71 -6.28 -2.16 0.64 1.26 0.12 0.57 Factor 4 -1.41 -1.72 -0.59 -2.14 0.06 0.23 1.69 0.46 2.15 1.47 1.27 2.49 Factor 5 -13.31 -2.83 -1.07 -5.01 0.79 2.11 -3.14 -2.41 7.49 2.78 Factor 6 -0.92 -3.97 -1.05 -2.98 -1.52 -1.26 -2.44 -1.49 McFadden R2 0.91 0.36 0.28 0.43 0.87 0.70 0.55 0.32

Table A9: Logit output for factors with grouped indicators (BP Approach) Bulgaria

Czech

Constant -10.61 -3.89 -2.08 -7.14 -11.95 -4.28 -2.10 -7.25 -14.81 -2.38 -10.65 -4.49 -10.46 -3.31 -1.26 -4.07 external 1 18.37 4.14 -0.75 -2.02 -1.38 -1.71 -0.87 -2.74 -4.94 -1.68 2.15 1.69 -2.02 -1.08 -1.14 -3.37 external 2 11.21 2.89 -0.06 -0.08 -1.17 -0.75 0.08 0.36 -43.63 -2.37 6.85 3.39 -0.88 -1.20 -1.44 -3.15 domestic 1 3.43 4.59 -0.32 -1.61 2.78 3.99 -0.76 -1.27 0.49 0.26 1.36 0.96 -5.76 -3.09 -2.42 -3.52 domestic 2 0.55 1.35 -0.06 -0.27 6.00 3.18 -1.85 -5.12 21.59 2.77 -4.58 -3.54 -6.94 -3.41 0.02 0.03 domestic 3 -0.20 -0.97 -8.09 -2.94 0.18 0.76 1.72 1.18 global 1 3.44 1.68 -1.07 -3.82 0.35 0.17 -0.22 -0.91 -14.29 -2.31 7.61 4.34 4.50 2.66 -0.45 -1.51 McFadden R2 0.74 0.23 0.70 0.24 0.89 0.85 0.43 0.43

(31)

30

Table A10: Stacked panel logit output for factors including all indicators (BP Approach)

Factor 1 -0.16 -3.11 Factor 2 0.08 1.30 Factor 3 0.37 6.12 Factor 4 -0.16 -2.22 Factor 5 -0.26 -4.50 Factor 6 0.10 0.50 Bulgaria -1.70 -7.98 Czech Republic -1.67 -7.99 Estonia -1.22 -6.62 Hungary -1.45 -7.58 Latvia -1.65 -7.20 Lithuania -2.38 -8.57 Poland -2.47 -9.54 Romania -1.13 -7.01 McFadden R-squared 0.07 LR statistic 90.99 Observations with Z = 1 226

Note: The model is estimated with Huber-White robust standard errors. Critical values of the z-statistic at the 1% and 5% level are 2.57 and 1.96 respectively.The critical value of the likelihood ratio at 1% significance level (6 degrees of freedom) is 16.81

Table A11: Stacked panel logit output for factors with grouped indicators (BP Approach)

External 1 0.14 1.38 External 2 -0.38 -3.57 Domestic 1 0.40 3.84 Domestic 2 -0.43 -5.19 Domestic 3 0.46 4.74 Global 1 0.08 0.67 Bulgaria -1.51 -7.69 Czech Republic -1.47 -6.63 Estonia -1.34 -7.30 Hungary -1.24 -5.94 Latvia -1.68 -7.01 Lithuania -2.40 -8.95 Poland -2.37 -9.14 Romania -1.29 -7.80 McFadden R-squared 0.07 LR statistic 88.95 Observations with Z = 1 226

(32)

31

Table A12: Stacked panel logit output for factors with grouped indicators (Regular Approach)

External 1 -0.25 -0.75 Exteran 2 -0.16 -0.38 Domestic 1 -0.23 -0.92 Domestic 2 0.47 3.09 Domestic 3 -0.12 -0.40 Global 1 0.53 1.69 Bulgaria -4.94 -5.75 Czech Republic -4.76 -6.14 Estonia -4.28 -8.15 Hungary -4.50 -6.70 Latvia -4.70 -6.53 Lithuania -5.41 -5.05 Poland -4.67 -6.10 Romania -3.80 -5.87 McFadden R-squared 0.07 LR statistic 13.91 Observations with Z = 1 18

(33)

32

Appendix C: Figures

Figure A1: Scree Plots of stacked indicators (Grouped and Ungrouped)

0 1 2 3 4 5 2 4 6 8 10 12 14 16 18 20 22 24

Scree Plot: Ungrouped Indicators

E ig e n v a lu e Number of Factors 0.0 0.4 0.8 1.2 1.6 2.0 1 2 3 4 5 6

Scree Plot: External Indicators

E ig e n v a lu e Number of Factors 0 1 2 3 4 5 2 4 6 8 10 12 14

Scree Plot: Domestic Indicators

E ig e n v a lu e Number of Factors 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0 1 2 3

Scree Plot of the Global Indicators

(34)

33

Figure A2: Scree Plots for Bulgaria

0 1 2 3 4 5 6 2 4 6 8 10 12 14 16 18 20 22 24

Scree Plot: ungrouped indicators

Factors E ig e n v a lu e 0.0 0.4 0.8 1.2 1.6 2.0 2.4 1 2 3 4 5 6

Scree Plot: external indicators

Factors E ig e n v a lu e 0 1 2 3 4 5 2 4 6 8 10 12 14

Scree Plot for domestic indicators

(35)

34

Figure A3: Scree Plots for the Czech Republic

0 1 2 3 4 5 6 2 4 6 8 10 12 14 16 18 20 22 24

E ig e n v a lu e Factors 0.0 0.4 0.8 1.2 1.6 2.0 1 2 3 4 5 6

0 1 2 3 4 5 2 4 6 8 10 12 14

(36)

35

Figure A4: Scree Plots for Estonia

0 1 2 3 4 5 6 7 8 2 4 6 8 10 12 14 16 18 20 22 24

E ig e n v a lu e Factors 0.0 0.4 0.8 1.2 1.6 2.0 2.4 1 2 3 4 5 6

E ig e n v a lu e Factors 0 1 2 3 4 5 6 2 4 6 8 10 12 14

(37)

36

Figure A5: Scree Plots for Hungary

0 1 2 3 4 5 6 2 4 6 8 10 12 14 16 18 20 22 24

Factors E ig e n v a lu e 0 1 2 3 4 5 2 4 6 8 10 12 14

(38)

37

Figure A6: Scree Plots for Latvia

0 2 4 6 8 10 2 4 6 8 10 12 14 16 18 20 22 24

(39)

38

Figure A7: Scree Plots for Lithuania

0 1 2 3 4 5 2 4 6 8 10 12 14 16 18 20 22 24

E ig e n v a lu e Factors 0 1 2 3 4 5 2 4 6 8 10 12 14

Scree Plot: Domestic Indicator

(40)

39

Figure A8: Scree Plots for Poland

0 1 2 3 4 5 6 7 8 2 4 6 8 10 12 14 16 18 20 22 24

(41)

40

Figure A9: Scree Plots for Romania

0 1 2 3 4 5 6 7 8 2 4 6 8 10 12 14 16 18 20 22 24

Factors E ig e n v a lu e 0.0 0.4 0.8 1.2 1.6 2.0 2.4 1 2 3 4 5 6

E ig e n v a lu e Factors 0 1 2 3 4 5 6 7 2 4 6 8 10 12 14

(42)

41

References

Berg, A and C. Pattillo (1998), ‘Are Currency Crises Predictable? A Test’, IMF Working Paper, 98-154.

Berg, A. and C. Pattillo (1999), ‘Predicting Currency Crises: The Indicators Approach and an Alternative’, Journal of International Money and Finance, 18, 561-586. Bussière, M. and C. Mulder (2000), ‘Political Instability and Economic Vulnerability’,

International Journal of Finance and Economics, 5(4), 309-330.

Cattell, R. B. (1966), ‘The scree test for the number of factors’, Multivariate Behavioural Research, 1, 629-637.

Diebold, F. X. and G. D. Rudebusch (1989), ‘Scoring the Leading Indicators’, Journal of Business, 62(3), 369-391.

Edison, H. J. (2003), ‘Do Indicators of Financial Crises Work? An Evaluation of an Early Warning System’, International Journal of Finance and Economics, 8(1), 11-53. Eichengreen, B., A. K. Rose and C. Wyplosz (1995), ‘Exchange Rate Mayhem: the

Antecedents and Aftermath of Speculative Attacks’, Economic Policy, 21, 251-312.

Frankel, J. A. and A. K. Rose (1996), ‘Currency Crashes in Emerging Markets: an Empirical Treatment’, Journal of International Economics, 41(3-4), 351-366. Girton, L. and D. Roper (1977), ‘A Monetary Model of Exchange Market Pressure

Applied to the Post war Canadian Experience’, American Economic Review, 67(4), 537-548.

Glick, R. and A. K. Rose (1999), ‘Contagion and Trade: Why are Crises Regional?’, Journal of International Money and Finance, 18(4), 603-617.

Haile, F. D. and S. Pozo (2006), ‘Exchange Rate Regimes and Currency Crises: An Evaluation Using Extreme Value Theory’, Mimeo, Department of Economics, Western Michigan University, Kalamazoo, Michigan.

(43)

42 Jacobs, J. P. A. M., G. H. Kuper and Lestano (2004), ‘Financial Crisis Identification: A

Survey’, In Mardi Dungey and Demosthenes Tambakis, editors, Identifying International Financial Contagion: Progress and Challenges, Oxford University Press, Oxford, 86-110.

Jacobs, J. P. A. M., G. H. Kuper and Lestano (2008), ‘Currency Crises in Asia: A Multivariate Logit Approach’, In S.-J. Kim and M. McKenzie (eds.) Asia Pacific Financial Markets: Integration, Innovation and Challenges, International Finance Review, Vol. 8, Elsevier, 157-173.

Kaiser, H. F. (1960), ‘The Application of Electronic Computers to Factor Analysis’, Educational and Psychological Measurement, 20, 141-151.

Kaminsky, G. L., S. Lizondo and C. M. Reinhart (1998), ‘Leading Indicators of Currency Crises’, IMF Staff Papers, 45(1), 1-48.

Kaminsky, G. L. (2006), ‘Currency Crises: are they all the same?’, Journal of International Money and Finance, 25, 503-527.

Krugman, P. (1979), ‘A Model of Balance-of-Payments Crises’, Journal of Money, Credit and Banking, 11(3), 311-325.

Kwack, S. Y. (2000), ‘An Empirical Analysis of the Factors Determining the Financial Crisis in Asia’, Journal of Asian Economics, 11(2), 195-206.

Lestano, J. P. A. M. Jacobs and G. H. Kuper (2003), ‘Indicators of Financial Crises Do Work! An Early Warning System for Six Asian Countries’, CCSO Working Paper 13, Department of Economics, University of Groningen, the Netherlands. Pozo, S. and C. Amuedo-Dorantes (2003), ‘Statistical Distributions and the Identification

of Currency Crises’, Journal of International Money and Finance, 22(4), 591- 609.

Rose, A. K. and M. M. Spiegel (2009), ‘Cross-country Causes and Consequences of the 2008 Crisis: Early Warning’, NBER Working Paper 15357, National Bureau of Economic Research, Cambridge, MA.

(44)

43 Stavárek, D. (2007), ‘Comparative Analysis of the Exchange Market Pressure in Central

European Countries with the Euro zone Membership Perspective’, MPRA Working Paper No. 3906.

Vanneste, J., A. Van Poeck and M. Veiner (2007), ‘Exchange Rate Regimes and

Exchange Market Pressure in the New EU Member States’, Journal of Common Market Studies, 45(2), 459-485.

Vlaar, P. J. G. (2000), ‘Currency Crises Models for Emerging Markets’, DNB Staff Report 45, De Nederlandsche Bank, the Netherlands.