Occasional Papers No. 5 June 26
Early Warning Systems on Currency Crises Authors: Irina Racaru (National Bank of Romania. Financial Stability Department) Mihai Copaciu (National Bank of Romania. Financial Stability Department) Ion Lapteacru (Montesquieu University Bordeaux IV) Abstract The paper estimates a model for early warning system on currency crises, based on a sample of emerging countries, including Romania. This analysis is justified by the current context of the Romanian market characterised by gradual capital account liberalisation focus of the Central Bank on financial price stability. The factors we have identified with an important impact on the probability of currency crises are: M2/reserves, currency overvaluation. For Romania, the probability of a currency crisis for the next 12 months, estimated for June 25, is quite low, of 4%, but higher than the value calculated for December 24, of 1.7%. JEL classification: F31, F32 Keywords: currency crises, early warning systems, noise-to-signal
ISSN 1583-3836 N o t e The opinions expressed in this paper are those of the authors do not necessarily represent the views of the National Bank of Romania, nor do they engage it in any way. Technical co-ordination of the Working Papers was carried out by the Research Publications Department. Reproduction of the publication is forbidden. Data may be used only by indicating the source.
Contents 1. NON-TECHNICAL SUMMARY... 5 2. LITERATURE REVIEW... 6 3. EARLY WARNING SYSTEMS... 1 3.1. Signal-based approach... 1 3.2. Limited dependent variable multinomial logit model... 13 3.2.1. Exchange rate Market Pressure indicator... 15 3.2.2. Currency Crisis Indicator... 16 3.2.3. The results of the econometric model... 17 3.3. The simulation of crisis probability... 2 4. EARLY WARNING SYSTEM THE CASE OF ROMANIA... 22 4.1. Currency crises definition... 22 4.2. Signal-based approach... 23 4.3. The Multinomial Logit Model... 24 5. CONCLUSIONS... 26 REFERENCES... 27 APPENDIX 1... 3 APPENDIX 2... 33 APPENDIX 3... 4
1. NON-TECHNICAL SUMMARY The crises that have affected several economies, many of them emergent, during the last decade, generated the ex-post appearance of different models trying to capture the features of crises, with no great efficiency in preventing the future ones. However, they are a good starting point in the construction of systems that seek to anticipate the crises. Two early warning methodologies will be used throughout the paper. This analysis is justified by the current context of the Romanian market characterized by gradual capital account liberalization focus of the Central Bank on financial price stability. The first approach used is the one developed by Kaminsky, Lizondo Reinhart (1998) that seeks factors which signal crises without creating false alarms, thus minimizing the noise to signal ratio. The paper uses data from 26 countries for a maximum time span of 1 years, between 1994 24 1. The factors with a high degree of efficiency whose persistency is more likely in periods anticipating crises were M2/reserves, overvaluation, short term debt/total debt, private debt/gdp, total debt/ GDP, current account deficit/ GDP. Considering the drawbacks of the above method, discussed in the paper, our second approach estimates the direct probability of a future crisis. Thus, we restricted the number of countries to those with complete data to use a multinomial logit model (with 3 values for the dependent variable: for normal period, 1 pre-crisis, 2 post-crisis; due to a different behaviour of variables in these different periods). The methodology of Bussiere Fratzscher (22) was used. The results indicate that the variables with significant impact on the predictability of crises are: overvaluation (calculated as deviation of the real effective exchange rate from a ar trend), the growth rate of the non-government credit as percentage of GDP, the current account balance/gdp, M2/reserves the export growth rate. Besides that, we found, a fact otherwise common in the literature, no direct evidence of a connection between the capital flows (foreign direct investments portfolio investments) the probability of a financial crisis. Capital account controls prove themselves useful in times of crisis, but the lack of such settlements is not necessarily an element that could provoke a crisis 2, on the contrary, such measures may also be a disadvantage (like limiting for instance the economic growth 3 or inducing little discip on the financial markets), Arvai Vincze (2). For Romania, based on the multinomial logit model, the estimated probability of crisis was 1.7% as for December 24, a level close to that observed in other countries from Central Eastern Europe 4. As for the estimated crisis probability at June 25, based on an out-ofsample forecast, it is 4%. The slight increase of this probability is mostly due to an 1 We determined the time span for each country function of the (in)existence of data. 2 Although it can be argued that they are not working in such times. 3 By limiting the access to additional financial sources to those already existing in the economy. 4 Bulgaria 2.1%, Czech Republic 2.6%, Croatia 2.7%, Pol 2.2%, Russia 1.3%, Slovakia.6%, Slovenia 1.3%, Turkey 2.1%, Hungary 1.%. 5
accelerating growth of non-government credit current account deficit (both measured as percentage of GDP),, to the slow growth of exports. For Romania, according to this model, a deviation of real effective exchange rate (REER) above the trend of 18% or an increase of 2.5 of the M2/reserves rate (that is more than doubling it) determines the increase of the probability over the threshold level by 1, or 7.7 percentage points, respectively. The growth of current account deficit/gdp or of the non-government credit/gdp did bring only minor changes in the crisis probability. The results of this analysis have to be used prudently as it is based on the assumption of all the other factors being kept constant. Despite the model limitations, it is important to see that for Romania the analysis shows a slow growth of external vulnerability between January-June 25. An important observation that has to be made from the beginning is that these models are based on general characteristics derived from the set of countries used in the panel does not count for specific factors, even if there is a high possibility of having such particular aspects that determine the crises for every of these countries. Another thing worth mentioning is that we used variables with a high frequency while data for some of these indicators was available only on a low one; that is why they had to be adjusted using statistical methods. This might have induced some noise in the data. The paper is structured as follows: Section 2, reviews the literature on currency crises, section 3 presents the construction of the early warning system using the signal based approach multinomial logit model, while section 4 analyses the results for Romania, Section 5 presents the conclusions. 2. LITERATURE REVIEW Preventing systemic crises ( especially currency ones) generated debates on academic policy makers areas in the last decade, especially due to the crises that took place in the European Monetary System (1992), Mexico (1994), countries from South East Asia in 1997 (Thail, Malaysia, Indonesia, Philippines South Korea) or in Russia (1998). Usually the economic literature mentions three types of financial crises: currency, banking debt crises. But in fact it s hard to find pure crises. A special attention in the literature is given to the concept of twin crises currency banking crises. Some good examples of theses are the crises from Asia (1997), Russia (1998) Turkey (2). Other mixed forms are represented by currency fiscal crises: Brazil (1999), or currency debt crises: Mexico (1994), Argentina (21). Thus it is hard to provide a clear definition of a currency crisis be realized in the above mentioned context. However a currency crises could be defined approximately as the loss of confidence in the national currency, resulting in an increased dem for foreign currencies 6
consequently to devaluation/depreciation of the national currency, loss in reserves or barriers on capital flows (Mills Omarova, 24). There are three generations of models on financial currency crises 5. The first generation models were introduced by Krugman (1979) developed afterwards by Flood Garber (1984). According to this type of models, in a fixed exchange rate environment, an expansion of domestic credit in excess of money dem will lead to a gradual but persistent loss of international reserves, finally to a speculative attack on the national currency. The agents realize that they will incur loses if they keep the national currency consequently they sell it when the shadow exchange rate 6 is equal to the fixed one. Due to this attack, the reserves are lost the monetary authorities are forced to abon the parity. According to this type of models, the crisis periods are described as periods when the international reserves persistently decrease, while the domestic credit grows more rapidly than the domestic money dem. If the excessive money supply reflects the need for public deficit financing, then high fiscal deficits credit to the government sector serve as indicators of the crisis. Agenor, Bhari Flood (1992) suggest that according to this type of models, external sector related variables such as the real exchange rate, could be used as indicators of a crisis. For example, worsening of trade current account balance could be the result of an expansionary fiscal /or credit policies, leading to a higher dem for imported goods part of the goods that were previously exported. Moreover, the relative prices of nontradable goods increase, since the expansionary policies lead to an increased dem. All these would lead to a real appreciation of the currency. Thus, the movements in the real exchange rate can be considered as indicators of a potential crisis. More recent models, suggest that the authorities abon the parity not only because of the decreasing reserves but also due to the evolution of other variables. For example, Ozkan Sutherl (1995) show that authorities have an objective function that depends positively on keeping the nominal exchange rate fixed negatively on the deviations of output from a target level. Thus, in the case of a fixed exchange rate, an increase of the external interest rates will put pressure on the internal ones. The latter, when increased, will influence negatively the output level, raising the cost of keeping the parity. According to this approach, the evolution of interest rates, both internal external, plus the movements in output can be used as indicators when assessing the probability of a crisis. Going further, other variables that affect the objective function of the authorities can be useful in the above mentioned process. A high level of interest rates leads to an increased financing cost for the central administration. Thus, an high public debt level will induce a bias towards aboning the peg. Moreover, high interest rates can bring about vulnerabilities in the banking sector, authorities would prefer to devalue than to risk for a banking crisis. Indicators such as the level of non-performing loans, the credit granted by the central bank to the commercial banks, the decrease in deposits could signal financial crises. 5 The opinions are not uniform. Krugman for example (1998) considers only two generations of models, the pure speculative one being included in the last type. 6 The exchange rate that would prevail if the regime was a floating one. 7
This first generation of models describes fairly well the crises that took place in Latin American countries such as Argentina, Brazil Chile, in the 8s 9s. Fixed exchange rates were introduced as part of the stabilization programs in order to reduce the inflation control the budgetary deficit. In most of the cases, the crisis ended in one of the above mentioned cases: the fixed exchange rate was aboned following the rapid increase in the domestic credit due the significant decrease in the foreign reserves of the central bank. Self fulfilling currency crises are the main characteristic of second generation models (Obstfeld, 1986). This type of models is described by the existence of multiple equilibrium, so that the economy can move between any of these, without significant changes in the real variables. For example, expectations that exchange rates would collapse, lead to higher interest rates. This generates additional costs for policy makers who can decide to abon the exchange rate peg, validating the initial expectations. Thus, the outcome does not necessarily imply the existence of negative evolution of the real sector. A sudden alteration of expectations is sufficient for aboning the peg moving to another equilibrium with a floating exchange rate. The remaining problem is detecting the causes of a sudden change in expectations to what extent these causes are correlated with the existence of problems in the real sector. This approach implies the idea that prediction of currency crises is very difficult due to the uncertainty of a connection between real variables the emergence of a crisis. In the absence of perfect information, real variables are important for the emergence of a crisis when they fall below a given threshold (Morris, Shin, 1998). Second generation models explain a series of crises, one of them being the one within the Exchange Rate Mechanism at the beginning of the 9s. Then, the unification of Germany generated a shock in dem which led to a rise in interest rates in other ERM member states. As the European job markets are very rigid, the consequence was an increase of unemployment. The consequence was creating the expectation that countries with high levels of unemployment are likely to be exposed to an attack on their currencies. These expectations further increased interest rates the authorities costs of maintaining the fixed exchange rate generated a series of speculative attacks, forcing the authorities to abon their peg. The third generation of models combines the first two, introducing in the analysis a series of microeconomic elements as banking sector information. The development of such models was stimulated by the Asian crisis. South Eastern Asian countries didn t have extremely expansionist monetary or fiscal policies in that period. Inflation unemployment rates were low. The problems were at the banking corporate level the revealing of these issues triggered the crisis. In particular, moral hazard issues, liquidity crunch contagion have been the essential drivers of the crisis. First, the moral hazard results from a close connection between the financial institutions the governments from that region, which induced the belief of implicit government guarantees. This led to over-indebtedness (external especially) suboptimal investment decisions. Thus, even a low shock for financial assets generated deterioration of bank s 8
portfolios. The authorities were forced to intervene through inflationary generating mechanisms in order to save the banking sector. The latter aspect led to a balance of payments crises (Krugman, 1998). According to this generation of models, the prices of assets could help in detecting a future currency crisis. Secondly, the liquidity crunch process starts from lower confidence in the banking system, leading to higher withdrawals consequently lower liquidity levels. Thus a banking crisis takes place followed by a currency one. According to this approach, the main factors that could affect the liquidity level of the banking system thus, can serve as leading indicators of a crisis, are: short term external debt of the banking sector long term capital inflows (Chang Velasco, 21). Finally, the contagion refers to the spread of the crisis on other economies. This can be the result of a devaluation due to external competitiveness problems or due to a common lender effect (e.g. an investor suffering looses in one country will withdraw the invested capital also from other countries not necessarily with similar characteristics). Furthermore, there is the possibility of herding behaviour among investors. Nowadays, among academics it is well-known that in reality currency crises are a mixture of important elements which belong to all three types of models above mentioned. However, as it was mentioned before, in some crises, certain aspects seem more important than others. Having these theoretical descriptions, it is important to check empirically if variables of the real or of the banking sector can be used as crisis indicators. On the other h, if their explanatory power is weak, one should conclude that those crisis occurred because of selffulfilling expectations herding behaviour of agents. Kaminsky, Lizondo Reinhart (1998) analyzed 28 empirical studies concerning currency crisis they divided 46 variables into 1 categories: (1) capital account; (2) debt profile; (3) current account; (4) international variables; (5) financial liberalization indicators; (6) other financial variables; (7) real sector; (8) fiscal sector; (9) political variables; (1) institutional factors. The authors characterized the explanatory power of the variables by examining a number of studies in which those variables had a relevant power in explaining crises. The following factors were selected: foreign reserves, real exchange rate, growth of nongovernment credit, credit to public sector inflation. Other indicators with a relatively good explanatory power of crises were: commercial balance, exports performance, growth of money supply, M2/reserves, growth of real GDP fiscal deficit. The influence of political institutional variables is hard to estimate since they were not included in many studies. The variables associated with external debt profile current account balance have had bad performances. As a possible explanation one could argue that the information provided by these indicators was already incorporated in the other ones. A more recent study of currency crises is that of Chui (22) which analyses a wider spectrum of crises, compared with Kaminsky, Lizondo Reinhart (1998) which had included only the events before 1997. Along with the old indicators, some new ones emerged, namely those connected to the weaknesses of the banking sector its relation with the public sector, which proved to be an important feature of the Asian crisis. 9
3. EARLY WARNING SYSTEMS Currency crises may have many patterns causes as we already shown in the previous chapter. The necessity to predict systemic crises caused a new instrument to be used, called early warning system, to estimate the probability of a crisis in a given horizon. It can be used for currency, banking or fiscal crises. The method consists in the analysis of financial economic indicators to study the vulnerability of the balance of payments or the unsustainable level of a currency. There are three main methods to build an early warning system: signal-based approach. It consists in monitoring a set of indicators. If the level of one indicator is greater than a calculated threshold, a signal is issued. The indicators may be computed (exchange market pressure, banking system stability, external position) or used from already reported financial economic indicators (GDP growth, fiscal deficit, capital market indicators, sovereign spread indicators) (Kaminsky, Lizondo Reinhart, 1998); limited dependent variable approach. It consists in regressing a logit/probit model in which the dependent variable is the crisis indicator (computed based on the exchange market pressure indicator) the independent variables are economic financial indicators (mostly the same as those used in signal-based approach). This model has the advantage of measuring the direct effect of each variable on the crisis probability, other things unchanged (Frankel Rose, 1996 or Bussiere Fratzscher, 22); severity of crisis indicators approach. It consists in finding the countries that will be most hit by a crisis that affects one of the countries from the same region. The method is based on a crisis indicator calculated over a stress period on the international markets the differences of this index between countries being explained by the economic condition of the crisis analysed (Sachs, Tornell Velasco, 1996). In this paper the first two methods will be used, where the first one being used only as an intermediary step to the second approach. 3.1. Signal-based approach The signal-based models have been considerably developed in several articles by Kaminsky, Reinhart Lizondo (1996, 1998), it is their methodology that will be used in the present paper. We should add that in order to calculate some of the indicators data used refer to more countries than those used for the second model 7. 7 For the signal-based approach we used data for the following countries: Argentina, Brazil, Bulgaria, the Czech Republic, Chile, Columbia, South Korea, Croatia, Estonia, Philippines, Indonesia, Latvia, Lithuania, Malaysia, Mexico, Pol, Romania, Russia, Slovakia, Slovenia, Thail, Turkey, Hungary, Venezuela. 1
The crisis in this context is defined as the period when the exchange market pressure indicator (EMP) exceeds the average value plus two stard deviations. The EMP, as we describe it in Section 3.2.1, represents a weighted average of three factors: growth of real interest rate, growth of real exchange rate the growth of international reserves. An indicator generates a signal when its value exceeds the predetermined percentile 8 (we choose the 15 th 85 th respectively, depending on the indicator). It should be noted that the threshold value is different from country to country, though the percentile is the same. The signalling period was set at 12 months. Thus, if the chosen factor sends a signal 12 months at most before the crisis, we consider this a valid signal, if the signal is not followed by a crisis, it is a false signal. The analysis can be summarised as in table 1. Table 1. Analysis of the signal according to identified crisis periods Crisis within 12 months No crisis within 12 months Signal issued A B No signal issued C D Table 2 presents the results obtained by using the signal-based approach. In column 3 we have the percentage of good signals in total signals that could have been issued. As already mentioned by Kaminsky (1998), a 1% percentage would mean that there should be a crisis signal every month during the 12 months before each crisis. We note that the ratio for foreign direct investments/gdp (FDI/GDP) shows the most correct crisis signals followed by the ratio for foreign direct investments/total debt (FDI/total debt) that for exports/gdp (exports/gdp). Nevertheless we have to be careful when interpreting these signals. In column 4 we have the number of false alarms as percentage of all bad signals that could have been transmitted. Thus, it can be noticed that the variables mentioned before generate a large number of false alarms. Variables which generate the smallest amount of false signals are: private debt/gdp, total debt/gdp M2/reserves. To describe the ability of an indicator to issue at the same time good bad signals, Kaminsky et al. (1998) suggest the use of what they call the noise-to-signal ratio, that is the fraction of the ratio of bad signals issued to total periods in which they could have been issued (B/(B+D)) over ratio of good signals that could have been issued (A/(A+C)). Column 5 in table 2 presents this ratio for the data used in this paper. For all indicators that generate rom signals as well as for a data panel large enough, the law of large numbers implies that the noise to signal ratio is 1. Thus, those indicators with noise-to-signal ratio over 1 have a very small crisis signalling performance. These are: FDI/GDP, public debt/gdp, exports/ 8 For example, for the credit growth we choose the 85th percentile, while for current account balance/gdp the 15th one. 11
GDP, real GDP growth, trade openness 9, inflation rate. The value of this indicator is unusual high for variables like real GDP growth inflation rate, variables which, according to Kaminsky et al. (1998), have a relatively high accuracy in forecasting crises. Indicators that have the ratio of correct crisis signalling false signals avoidance below 1 are: M2/reserves, overvaluation, short term debt/total debt, private debt/gdp, total debt/gdp, current account balance/gdp, portfolio investments/gdp, short-term debt/exports, exports growth, governmental balance, non-government credit/gdp. Table 2. Performance of indicators using signal-based approach M2/Reserves 1 2 Overvaluation Signals issued Good signals as percentage of possible good signals Bad signals as percentage of possible bad signals Noise- tosignal ratio (NS) Conditional probability (crisis/signal) Unconditional crisis probability Signal persistency (crisis vs. normal period) 1 2 3 4 5 6 7 8 A/(A+C) B/(B+D) (B/(B+D))/ (A/(A+C)) A/(A+B) (A+C)/ (A+B+C+D) Another element discussed by Kaminsky et al. (1998) is the difference between the conditional (column 6) unconditional (column 7) probability of crisis. If the indicator analysed has a high accuracy in crisis anticipation, the conditional probability should be somewhat higher than the unconditional one. This holds for the first 6 indicators (in bold in 1/ NS 527.44.12.28.2.7 3.55 44.45.14.31.17.6 3.25 3 Short-term debt/total debt 487.38.14.36.16.6 2.75 4 Private debt/gdp 455.22.1.47.11.6 2.14 5 Total debt/gdp 456.22.11.48.11.6 2.7 6 Current account 548.35.18.53.1.6 1.9 balance/gdp 7 Portfolio investments/gdp 46.24.15.62.9.6 1.61 8 Short-term debt/exports 481.22.14.62.1.6 1.61 9 Exports growth 489.19.14.71.9.6 1.4 1 Budget balance 4.25.18.73.9.7 1.37 11 Non-government credit /GDP 473.17.14.84.7.6 1.19 12 FDI/GDP 2,153.77.85 1.1.5.5.91 13 Public debt/gdp 445.12.13 1.12.5.6.89 14 FDI/Total debt 2,318.76.86 1.14.5.6.88 15 Exports/GDP 2,457.6.8 1.32.5.6.76 16 Real GDP growth 455.1.17 1.66.4.6.6 17 Trade Openness 722.4.24 5.6.1.6.18 9 Calculated as ratio of the sum of exports imports in GDP 12
the table). But we have to say that, for the panel data we used, these probabilities are relatively small compared to those estimated by Kaminsky et al. (1998). This could be due to the usage of data from countries that did not have major crises in the analysed period which led to diminishing of conditional as well as of unconditional probability. In the last column in the above table we estimate the persistency of the signal over the 12 months period before the crisis, compared to a normal period. Thus, the signals issued by M2/reserves currency overvaluation are 3 times more persistent in periods before crisis, than in normal periods. A value over 2 is also obtained for the next 3 indicators that are different forms of debt compared to GDP. We have to explain the relatively high accuracy of total private debt/gdp, compared with the small accuracy of public debt/gdp. This can be due to a larger percentage of the private debt in total debt for the countries subject to analysis /or to a low quality of data for these indicators, explanation that is also confirmed by the econometric analysis in the next sections where these indicators prove irrelevant for estimating the probability of crisis. Berg Pattillo (1998) were the first to test the accuracy of these models (not only those developed by Kaminsky et al. 1998), but also the logit-based ones), not only in-sample, but also out-of-sample. To do this, they estimate the models using observations up to 1995 make estimates for the next 2 years. The authors use a probability threshold of 25% 5% to indicate a crisis. Then they compare the results to present data. The Kaminsky, Reinhart Lizondo (1998) model make correct estimates in 7% of the cases. Forecasting crises is anyhow of large interest, considering the fact that such a result can be due to long periods without crises correctly identified. Thus, the model estimates correctly only % of the pre-crisis periods when the threshold is of 25%. At the same time, more than half of signals are false. Besides, crises appeared in 24% of cases without any signal issued. Unlike these models, the logit/probit models have a higher accuracy of prediction. When the threshold is of 25%, the model anticipates correctly 79% of observations. 73% of pre-crisis periods were correctly estimated the ratio of false alarms is little under 5%. To surpass some of the limits of the models described above, we used, for the next part of our analysis, an LDV model, but based on a multinomial logit procedure. 3.2. Limited dependent variable multinomial logit model The model is part of those based on qualitative dependent variables, but the dependent variable in this case is not binary. The econometric instrument used is a multinomial logit model. The difference between this model a binary model is that the crisis period is divided in two: one preceding the crisis one during after the crisis. This separation allows us to avoid the post-crisis bias effect that is explained by the different evolutions of macroeconomic indicators in the two periods (Bussiere Fratzscher, 22). 13
The creation of a warning model based on a multinomial logit implies the following steps to be taken: computation of an exchange market pressure indicator: that determines the crisis periods, including not only successful attacks against a currency (i.e. forcing the central bank to abon the peg), but also the moments of external vulnerability when the measures adopted by the monetary authority or the favourable external conditions on financial markets made possible the avoidance of a currency crisis; calculation of the currency crisis indicator; construction of the crisis indicator (multinomial); estimation of the model using the multinomial logit; establishing the optimal value for the currency crisis threshold. The multinomial logit model is built based on the crisis indicator. The independent variables are chosen among those that can characterize the external, financial economic conditions that change before during a currency crisis. The selection was made taking also into account the signal-based analysis presented in section 2.1 1. Thus the main variables used are: external competitiveness indicators: currency overvaluation, current account balance, commercial trade balance, imports/exports as absolute level growth rates. We used the real effective exchange rate instead of the real exchange rate in order to analyse the external competitiveness to be analyzed evaluate economies with fix exchange rate regimes; external exposure: short-term debt/reserves, total debt/reserves, short-term debt growth rate; internal economic indicators: real GDP growth, budget deficit, inflation rate; financial indicators: non-government credit, government credit, currency multiplier, M2/GDP, bank deposits; contagion indicators: bank system contagion. The bank system contagion was calculated using the method suggested by Fratzscher (2): CB ij = d F F dj d F F di i (1) where F di represents the loans of the country d to the emerging country i, F d is the total amount of loans issued by the country d. The countries d are the developed countries while the countries i,j (i j) are the emergent ones analysed in this paper. The interpretation of this indicator is based on the common lender effect: if in the country j there is a currency crisis the country d exposure towards the first 1 To have an image of variables used in economic literature, see Lestano Koper (23) or Aziz, Caramazza Salgado (2). 14
is high, than high is also the probability that the country d will refuse to prolong the debt or even will withdraw the capital from the country i. Signalling of a crisis for certain countries of the period is done considering an optimal threshold (a probability over this threshold is interpreted as a crisis signal). Thus, the result estimated on the basis of this model can fall in one of the situations described in section 3.1 (Table 1). The choice of the optimal threshold of the period must be done based on the number of nonsignaled crises false alarms considered as optimal by the monetary authority. The following loss function 11 can be considered: L( T ) θ * P ( T ) + (1 θ ) * P ( T ) (2) CN CS where T is the probability threshold; P CN is the probability of not signalling a crisis; P CS the probability of signalling a crisis, Ө the cost of not signalling a crisis or the risk aversion degree. By increasing the time horizon the probability threshold, the number of nonsignalled crises will also rise, while the number of false alarms will decrease. 3.2.1. Exchange rate Market Pressure indicator Exchange rate Market Pressure indicator (EMP) is constructed as a weighted sum of three factors: real exchange rate growth, real interest rate growth, international reserves growth. EMP i, t = 1 σ 2 e e e i, t i, t 1 1 + σ 2 r r r i, t i, t 1 1 σ 2 res res res i, t i, t 1 (3) where σ 2 e represents the exchange rate volatility; σ 2 r interest rate volatility; σ 2 rez international reserves volatility. The reason for which the EMP is thus defined is that in case of a currency attack the monetary authority has two options: either it tries to maintain the exchange rate peg (in fixed currency regimes) by diminishing the reserves /or increasing the interest rate; or the currency is strongly devaluated. The inverse variance as weighting coefficient was used as the factors with less volatility are more important (the most important factor in determining a crisis is the change in the international reserves). In addition, using constant weights for all the countries assures the comparability of the EMP indicator between countries by that, of the crisis indicator, especially in the economies with fixed currencies (the exchange rate volatility is less in these cases, which would have determined a bigger weight of the exchange rate) (Bussiere Fratzscher, 22). Some studies (Edison, 2 or Miller Omarova, 24) do not include interest rate in the calculation of the EMP indicator, but this is explained in most cases by the lack of data for 11 Bussiere Fratzscher (22) 15
emerging countries. Others (Berg, Borensztein Pattillo, 24) motivate the exclusion of interest rate by the fact that the exchange rate depreciation the interest rate growth are different events its use in calculating the EMP indicator would take us to a cumulative estimation of the two events. 3.2.2. Currency Crisis Indicator The Currency Crisis indicator (CC) defines the crisis period as that moment when the EMP indicator surpasses the average value twice the stard deviation 12. CC i, t, if EMPi, t > EMPi + 2σ =, otherwise 1 EMP( i) (4) Once the currency crisis period is defined, the crisis indicator can be constructed it will be used in the multinomial logit analysis. The main problem in defining this indicator is the time horizon wanted for observing the occurrence of a currency crisis. The separation into pre-crisis post-crisis can be done taking into account this predetermined time span. But all economies that have passed through such crises had different recession recovery periods. Thus, the setting of a time horizon as pre- post-crisis periods must realise a compromise between: (i) the horizon wanted by the authority for analysis crisis prevention acts, (ii) the period between the first vulnerability signs the moment of a currency crisis beginning. The time horizons most used in the economic literature are of 12, 18, 24 months. In this paper, the estimations were done using all these intervals, but we obtained the most reliable results using a 12-month span. The crisis indicator is calculated as follows: 1, if k = 1,12 CCi, t+ k = 1 CCi, t+ 1 k 1 Yi, t = 2, if k = 1,12 CCi, t+ 1 k = 1 (5), otherwise The values,1 2 have the following meanings: Y=, normal period: there was no crisis 12 months before there is no probability of a crisis in the following 12 months; Y=1, period preceding a crisis: there is a crisis announced in the following 12 months there was no crisis before; Y=2, period after a crisis: there was a crisis in the preceding 12 months. 12 The value 2 is given by the choice of a one sided interval of 95%. Edison (2) uses 2.5, while Miller Omarova (24), Kaminsky et al. (1998) use 3. We use 2 in this paper, like Bussiere Fratzscher (22). 16
3.2.3. The results of the econometric model The multinomial logit analysis was done on a set of 21 economies 13. We used only emerging countries, as in these countries the different financial problems, internal as well as external, represent a crucial element in the occurrence of a currency crisis, which is not applicable to developed countries. This is, in fact, also the conclusion of Kaminsky (23). The maximum time span used is 1994-24 14. The main crises during this period in the analysed countries are: Mexico (1994), the Czech Republic (1997), Bulgaria (1996), Asia (1997), Russia (1998), Brazil (1999) Turkey (2). The results of the econometric model are presented in Table 3. The indicators we used are (the details on the sources for statistical data are described in Appendix 1): currency overvaluation (calculated as deviation of the real effective exchange rate from a ar trend); non-government credit as percentage of GDP; current account balance as percentage of GDP; the ratio between the M2 monetary aggregate reserves; exports growth rate. The indicators were chosen based on the noise-to-signal values, as they were computed in Section 3.1 as indicators with a smaller noise-to-signal ratio have a higher accuracy in explaining crises. The first part in Table 3 presents the coefficients for the five variables used, describing the probability of a pre-crisis period as compared to the probability of a normal period. All the variables have the expected sign. The currency overvaluation, the M2/reserves ratio are both significant at 1%; the growth of non-government credit as compared to GDP, the current account balance on GDP are significant at 5%; the exports growth at 1%. A growth of the real effective exchange rate as compared to the trend, a lending boom (the growth rate of the non-government credit/gdp), a growth of M2 compared to the reserves, are all leading to a higher probability of crisis. At the same time, a high current account deficit a diminution of exports growth rate also contribute to a growth of the crisis probability. The difference between the pre post crisis period can be noticed, especially for the currency overvaluation the growth of non-government credit. Besides, the current account balance is significantly improving after the crisis. 13 For the signal-based approach we used data for the following countries (26): Brazil, Bulgaria, Czech Republic, Chile, Columbia, South Korea, Croatia, Estonia, Philippines, Latvia, Lithuania, Malaysia, Mexico, Pol, Romania, Russia, Slovakia, Slovenia, Turkey, Hungary, Venezuela. 14 The time span for each country was determined depending on the (in)existence of data. 17
The model general degree of explanation is relatively good if we take into consideration the values obtained usually with panel-type data (pseudo R2 is of.25). Yet this is not the only evaluation criterion. We have in Table 4 a more detailed analysis of the performance degree of the model that was calculated on the basis of the chosen probability threshold. Table 3. The results of the econometric estimation on the basis of the multinomial logit model for 1994-24 Multinomial logistic regression Number of obs = 29 LR chi2(1) = 653.33 Prob > chi2 =. Log likelihood = -973.47669 Pseudo R2 =.2513 --------------------------------------------------------------------------------- Y Coef. Std. Err. z P> z [95% Conf. Interval] -------------+------------------------------------------------------------------- Y=1 Overvaluation.187522.9215 11.81..97.126843 NGC/GDP.136496.69529 1.96.5.223.27277 CA/GDP -.444268.21994-2.2.28 -.8417 -.48367 M2/Reserves.688641.7771 9.74..55568.8272235 Exp growth -.82245.48127-1.71.87 -.176571.1282 Constant -5.67516.316928-18.81. -6.266413-5.83799 -------------+------------------------------------------------------------------- Y=2 Overvaluation -.659354.8891-7.42. -.833598 -.485111 NGC/GDP -.1248.43645-2.85.4 -.29891 -.3885 CA/GDP.3649.11645 2.64.8.7865.5394 M2/Reserves.995183.647752 14.4..7825613 1.36475 Exp growth -.27937.36976-7.33. -.9 -.198465 Constant -4.5922.228277-2.11. -5.37498-4.142941 ---------------------------------------------------------------------------------- (Y= is the basis group) The choice of this threshold was made on the basis of the loss function minimisation presented in Chapter 3.2, which, in the case of a neutral risk authority (θ=.5), is of 1%, as it can be seen in Figure 1. If we increase the risk aversion degree, the probability threshold decreases. The choice of risk aversion degree has to be made according to the trade-off between the cost of not signalling a crisis that of signalling a crisis (that is, of taking measures accordingly) when there is no real possibility of having a crisis. 18
Figure 1. The loss function using different pre-crisis periods different risk aversion degrees 9%.55 8%.5 7%.45 6%.4 5%.35 4% 3% 2% 1% %.1.1.1.2.3.4.5.6.7.8.7.6.5.3.25.2.15.1.5.1.15.2.25.3.35.4.45 12 months 18 months 24 months.5 The model goodness of fit is estimated for the pre-crisis period, that is for P(Y=1). In this case 9% of observations 6.5% of crises are correctly estimated, while the crisis probability within the signal is of only 37.55%. Comparatively to models estimated by the International Monetary Fund (-Developing Country Studies Division), Kaminsky-Lizondo-Reinhart (1998), the GS-WATCH model of Goldman-Sachs Credit Suisse First Boston 15, the model estimated has a very good number of correctly estimated observations, the number of false alarms in total alarms, of the probability of not signalled crises, of the crisis probability within the signal (37.55% versus 37.2%, 29.7%, 26% 6.5%, respectively). Table 4. Goodness of fit (probability threshold of 1%) Signal % of obs. correctly called: 9.28% Crisis S= S=1 Total % of crises correctly called: 6.54% Y= 1,825 148 1,973 % of false alarms of total alarms: 62.45% Y=1 58 89 147 % prob. of crisis given an alarm: 37.55% Total 1,883 237 2,389 % prob. of crisis given no alarm: 6.15% The percentage of correctly estimated crises remains relatively below the value obtained in the studies mentioned above (6.54% versus 65.1%, 59.8%, 66.2% 61.1% respectively). Despite this, the performance of the model is below those estimated by Bussiere Fratzscher (22): the number of correctly estimated crises in those is of 73.7%, the number of false alarms within total alarms: 44.1%, the crisis probability within signal: 55.9% the inclusion of countries that did not have major crises could be one of the explanations. Also, the degree of performance of the model can be seen in Appendix 2, on the basis of the graphic analysis of the probability estimated for each country. 15 See Berg, Borensztein, Pattillo (24) 19
3.3. The simulation of crisis probability As stated before, the multinomial logit model consists in dividing the crisis period in two: pre post crisis periods in order to avoid the post crisis bias. In table 5 the average values of the indicators studied are displayed. As it can be noted, the values differ according to the period for which they are calculated. For example, the pre-crisis period is characterised by currency overvaluation, while the post-crisis by depreciation. The normal periods are not showing significant deviation of real effective exchange rate from the trend. The rate of growth of non-government credit/gdp has higher values in pre post crisis periods than in the normal ones. The pre-crisis is characterized by lending boom. Generally, the credit is a procyclic phenomenon, the pre-crisis periods are almost always a period of economic expansion. In post-crisis the average value for this indicator should be dropping fast. This will be possible if credit institutions were able to recuperate the invested funds. The slow decrease of the average value is explained by the fact that the balance sheets of companies are deteriorating rapidly the companies are, in the best case, insolvent, or in the worst case, in default, cannot pay back the loans taken in the pre-crisis period. Current account balance attains a large deficit in pre-crisis sustained by large capital inflows. The growth of monetary aggregates currency overvaluation stimulate import growth. In post-crisis, the depreciation of currency has a positive impact on exports, consequently, on current account balance. Also, M2/reserves is at a higher level in pre-crisis due to decreasing reserves or growing M2. In post-crisis, the ratio is dropping (especially due to declining reserves), remaining at a higher level compared to the normal period. Exports witness a distinct development. If, during the crisis, the growth rate is high, it drops in the pre-crisis periods, reaching the minimum values in the post-crisis period. It should be noted that the rate of growth remains positive in all three periods. Table 5. The average values for the indicators used Variables Average values All periods Normal period Pre-crisis period Post-crisis period Overvaluation 1.4.45 13.77-1.14 NGC/GDP 5.2 4.87 7.69 6.39 CA/GDP -2.26-2.5-3.63.3 M2/Reserves 2.67 2.52 4.21 3.2 Export growth 12.1 14.6 4.97.16 The coefficients of a logit/probit model cannot be interpreted as marginal effects, due to nonnormal distributions of the independent variables; therefore the marginal effects have to be calculated for a predetermined value of an explanatory variable. 2
Table 6 presents the effect on the estimated probability based on different scenarios. As reference scenario we choose the one with the variables on the average values from the normal period. In this scenario, the crisis probability for the following 12 months is very low, of only 2.2%. The value rises significantly to 27.58% when all the variables are set to the average values from the pre-crisis period. To estimate the impact of variables on the crisis probability, it was considered that all are at the reference level (the average values from the normal period) with the exception of the analysed variable that undergoes a specified change, displayed in the table. The variables with the highest impact are M2/reserves currency overvaluation, the probability doubling more its value, approaching the signalling threshold. The first one generates a probability of 7.23%, 5.21 percentage points higher than the reference level; while the second one almost 6% i.e. more than 3.9 percentage points higher. Table 6. The probability of currency crisis general scenarios Prob. change Prob. of Scenarios (percentage crises (%) points) 1 All variables are at the normal period average 2.2-2 All variables are at pre-crisis period average 27.58 +25.56 3 All variables are at the normal period average except for: a. Exchange rate +2% 2.51 +.49 b. Exchange rate +5% 3.48 +1.46 c. Exchange rate +1% 5.92 +3.9 d. M2/reserves +2.5 7.23 +5.21 e. CA/GDP -5% 2.53 +.51 f. NGC/GDP +5% 2.16 +.14 g. Export growth -15% 2.22 +.2 The analysis confirms the previous results from the noise-to-signal approach, presented in Section 3.1. in which the M2/reserves currency overvaluation are the factors with the lowest ratio of signals of a not realised crisis in total no-crisis periods on signals of a realised crisis in all pre-crisis periods. 21
4. EARLY WARNING SYSTEM THE CASE OF ROMANIA Among the processes that at this moment suggest a more careful monitoring of the Romanian economy, one could mention: the capital account liberalization while still having high nominal interest rates when compared with similar economies. These rates can stimulate capital inflows, while their rapid decrease can fuel aggregate dem have thus a negative impact on the inflation rate; the appreciation of the currency due to these inflows, including the remittances from the Romanian workers from abroad; the increase in non-government credit, especially the foreign currency denominated component. 4.1. Currency crises definition As shown in the Section 3.2.1, the identification of the currency crisis is done through the exchange market pressure index. According to its evolution, three crises have been identified for Romania, during 1994-24: 1. January 1997 the start of the liberalization of the exchange rate (initiating a managed float type of policy having as background also the start of the liberalization process for the goods with administered prices, the latter fact generating currency depreciation the increase in inflation expectations); 2. February 1999 problems related to the signing of the st-by arrangement with the, the start of the war on Yugoslavia, an overappreciated currency (around 3% in September 1998), contributed to the deterioration of inflation expectations currency depreciation; 3. November 1999 the financing of the energy imports, while facing unfavourable external conditions due to the recent at that time Russian crisis especially the bailing-in imposed by 16. The exchange market pressure index for Romania was computed using the exchange rate of the Romanian currency versus US dollar. While using the exchange rate versus the Euro would have been more appropriate for the recent evolution, its introduction the relative small variations in the two international currencies constitute arguments that support the approach used here. 16 Romania s sovereign rating in 1999 was one class above the default, namely Moody s: B3, Stard Poor s Fitch: B-. Russia Ukraine had the same rating at that moment. 22
Figure 2. Exchange market pressure index the threshold for identifying currency crises 2 1.5 Jan.97 1.5 -.5-1 -1.5-2 Feb.99 Nov.99 Jan.94 Jul.94 Jan.95 Jul.95 Jan.96 Jul.96 Jan.97 Jul.97 Jan.98 Jul.98 Jan.99 Jul.99 Jan. Jul. Jan.1 Jul.1 Jan.2 Jul.2 Jan.3 Jul.3 Jan.4 Jul.4 Jan.5 4.2. Signal-based approach The variables that signal the February 1999 crisis as they were identified earlier are: the private debt to GDP ratio, the current account deficit to GDP ratio, the increase in nongovernment credit relative to GDP, while for the November 1999 crisis they are: the M2/reserves the FDI/GDP ratios. For the 1997 crisis, the lack of some data did not allow to undertake this analysis. According to the signal-based approach, the indicators that need a careful current investigation due to their recent evolution are: overvaluation of currency as shown by the real effective exchange rate developments, the latter being over the threshold for the February-May 25 period. The evolution of the real effective exchange rate should be treated with caution since it reflects deviations from a ar trend, the latter being considered in this case a proxy for the equilibrium real exchange rate. Still when looking back, the maximum value of the so-defined overvaluation is attained during April-August 1998 17 ; current account balance to GDP ratio for December 24-February 25 period when it was over the threshold, while afterwards it came back under the threshold; debt indicators: short-term debt/long term-debt, private debt/gdp, public debt/gdp, foreign debt/gdp. While the first the last indicators were above the threshold in December 24 18, the other indicators showed positive developments; 17 The same period for which the currency was over appreciated (around 3% NBR Annual Report). 18 This was the last period for which data taken from was available (Appendix 1). However, as mentioned before, the quality of debt-related data might not be appropriate. 23
the growth of non-government credit relative to GDP which was over the threshold only in January 25, while in the second part of the year the growth rate increased after some months of contraction to a much lower level (1.3% in March 25); the budget balance to GDP ratio which was over the threshold in January February 25, coming under it afterwards. The indicators that show a consolidation process during December 24 June 25 are: M2/reserves due mostly to the increase in the reserves; capital account indicators: FDI/total debt, FDI/GDP due to the increase in FDI (following privatization). As also shown in Section 3.1. the signal-based approach, though allows for the usage of a larger set of variables (potentially allowing to detect more signals), it does not offer a crisis probability consequently it does not allow for a quantifying effect on the probability. Thus, the next section captures the analysis of the Romanian economy using the multinomial logit approach. 4.3. The Multinomial Logit Model The probability of crisis according to this model is presented in Figure 3. As it can be observed, the probability was above the threshold during February-November 1998 (mainly due to the over appreciation of the currency). After this period, the probability has decreased substantially, from June 2 has been below 5%, while in the last two months of 24 it increased slightly. The probability of crisis during January-December 25, as it is estimated in December 24 is really low, namely approximately 1.7%. Analyzing the economies in Central Eastern Europe, one could observe that the estimated probabilities for 25, as estimated in December 24, are similar with those obtained in the Romanian case, namely 19 : Bulgaria 2.1%, Czech Republic 2.6%, Croatia 2.7%, Pol 2.2%, Russia 1.3%, Slovakia.6%, Slovenia 1.3%, Turkey 2.1%, Hungary 1.%. Furthermore, according to an out-of-sample forecast the crisis probability for the July 25- June 26, as estimated in June 25, is just 4%. The slight increase in the probability from the level estimated in December 24 is due mainly to the upward evolution of the nongovernment credit overvaluation of the currency, with a minor contribution of the current account deficit (the decrease in the growth rate of exports). 19 The graphs of the estimated probabilities for each country are shown in Appendix 2. 24
Figure 3. The probability of crisis for Romania according to the multinomial logit model.35.3.25.2.15.1.5. Feb.98 May Dec.98 May 5 Mar.97 Jul.97 Nov.97 Mar.98 Jul.98 Nov.98 Mar.99 Jul.99 Nov.99 Mar. Jul. Nov. Mar.1 Jul.1 Nov.1 Mar.2 Jul.2 Nov.2 Mar.3 Jul.3 Nov.3 Mar.4 Jul.4 Nov.4 Mar.5 Using these forecasts, different scenarios were constructed as presented in Table 7. As one can notice, the probability is over the threshold for scenarios 4c 4d overvaluation of the currency with another 1%, compared with June 25 situation the increase of the M2/reserves ratio with 2.5 (more than the double of the ratio). Thus, an excessive overvaluation of the currency an important decrease in the reserves (or excessive monetization) represent the major factors in the development of a possible currency crisis. The increase of the current account deficit/gdp ratio by 5 percentage points (namely reaching a value of 16%) or the acceleration of the increase of the non-government credit has a minor impact on the probability of crisis. Table 7. The probability of currency crisis scenarios for the case of Romania Scenarios Prob. of Prob. change crises (%) (percentage points) 1 All variables are at the normal period average 1.49 2 All variables are at pre-crisis period average 5.2 +3.71 3 All variables are at June 25 level 4 +2.51 4 All variables are at June 25 level except for: a. Exchange rate +2% 4.93 +.93 b. Exchange rate +5% 6.72 +2.72 c. Exchange rate +1% 11.5 +6.5 d. M2/Reserve +2.5 17.74 +13.74 e. CA/GDP -5% 4.95 +.95 f. Credit growth/gdp + 5% 4.27 +.27 g. Export growth -15% 4.48 +.48 However, one should be careful when interpreting these results since the analysis is done in a caeteris paribus context the simultaneous effects that could occur are not considered (an increase in the overvaluation of the currency is accompanied for example by a deterioration of the current account). 25
5. CONCLUSIONS The present study tested empirically two early warning systems on currency crises using a sample of emerging economies. The main methodology used (Bussiere Fratzscher 22), more comprehensive than the approach initiated by Kaminsky, Lizondo Reinhart (1998), identifies as the main indicators of currency crises the overvaluation of the currency (computed as the deviation of the real effective exchange rate from a ar trend), the increase of the non-government credit to GDP ratio, the current account deficit to GDP ratio, M2/reserves the growth rate of exports. From the indicators mentioned above, the M2/reserves overvaluation of the national currency have, caeteris paribus, the highest impact on crisis probability (although in reality the determinant factors can simultaneously interact). The performance of the multinomial logit model when compared with the similar models realized by the (Developing Country Studies Division), Kaminsky-Lizondo-Reinhart (1998), Goldman-Sachs (GS-WATCH) Credit Suisse First Boston, is generally better but weaker when compared with the model developed by Bussiere Fratzscher (22). One of the possible explanations for the latter fact could be the inclusion in the sample of some countries that did not face major crises throughout the period analyzed. As for Romania, the analysis was necessary considering the current mix of policies applied. The results show, that as of December 24, the estimated crisis probability was only 1.7%, a level comparable with the one registered in the case of other Central Eastern European Countries. Furthermore, the forecast of the probability of crisis within the next 12 months, as of June 25 was 4%, the main factors contributing to its slight increase in the six month period being the acceleration of the non-government credit growth, the increase in the current account deficit (both divided by GDP) the decrease in the growth rate of exports. The paper presents some common factors for the investigated sample impacting on the probability of crisis. However, one should take into account that some country specific aspects could represent important factors in determining currency crises. 26
REFERENCES Agenor, P. R., Bhari, J. Flood, R., 1992 Árvai, Z., Vincze, J., 2 Aziz, J., Caramazza F., Salgado R., 2 Berg, A., Borensztein E., Milesi-Ferretti G. M., Pattillo C., 1999. Berg, A., Borensztein E., Milesi-Ferretti G. M., Pattillo C., 24 Bussiere, M., Fratzscher, M., 22 Calvo, G., 1998 Calvo, G., Reinhart, C., 2 Calvo, G., Izquierdo, A., Talvi E., 22 Chang, R., Velasco, A., 21 Speculative Attacks Models of Balance of Payments Crises, Staff Papers, Vol. 39 Financial Crises in Transition Countries: Models Facts, Central Bank of Hungary, WP 2/6 Currency Crises: In search of Common Elements, International Monetary Fund, Working Paper 67 Anticipating Balance of Payment Crises, The Role of Early Warning Systems, International Monetary Fund, Occasional Paper 186 Assessing Early Warning Systems: How Have They Worked in Practice?, International Monetary Fund, Working Paper 52, March, 24 Towards a New Early Warning System of Financial Crises, European Central Bank, Working Paper 145, May, 22 Capital Flows Capital-Market Crises: The Simple Economics of Sudden Stops, Journal of Applied Economics, November,1998 When Capital Inflows Come to a Sudden Stop: Consequences Policy Options, in P, Kenen A, Swoboda (eds), Key Issues in Reform of the International Monetary Financial System (Washington, DC: International Monetary Fund) Sudden Stops, the Real Exchange Rate, Fiscal Sustainability: Argentina's Lessons, Inter-American Development Bank A model of Financial Crises in Emerging Markets, Quarterly Journal of Economics, Vol. 116 Chui, M., 22 Leading Indicators of Balance-of-Payments Crises: a Partial Review, Working Paper 171, Bank of Engl 27
Edison, H. J., 2 Eichengreen, B., Rose, A., Wyplosz, C., 1997 Frankel, J. A., Rose, A.K., 1996 Fratzscher, M., 2 Kaminsky, G.L., Lizondo S., Reinhart C.M., 1998 Kaminsky, G.L., Reinhart C.M., 1999 Kaminsky, G.L., 23 Krugman, P., 1979 Krugman, P., ed. 1998 Lestano, J.J., Koper, 23 Mills, C., Omarova, E., 24 Morris, S., Shin, S.H., 1998 Obstfeld, M., 1994 Do Indicators of Financial Crises Work? An Evaluation of An Early Warning System, International Finance Discussion Papers, Federal Reserve Board of Governors, no. 675, July, 2 Contagious Currency Crisis, http://www.haas.berkeley.edu/~arose/ Currency Crashes in Emerging Markets: An Empirical Treatment, Journal of International Economics, Vol. 41 (November), pp. 35-66 On Currency Crises Contagion, Institute for International Economics, December, 2 Leading indicators of currency crisis, International Monetary Fund Staff Papers 45/1 The Twin Crises: The Causes of Banking Balance-of- Payments Problems, American Economic Review; 89(3), pp. 473-5 Varieties of Currency Crises, National Bureau of Economic Research, Working Paper 1193, http://www.nber.org/papers/w1193 A Model of Balance-of-Payment Crises, Journal of Money, Credit, Banking, 11(3), pp. 311-25 Currency Crises, Chicago, University of Chicago Press, 2 Indicators of Financial Crises Do Work! An Early- Warning System for Six Asian Countries, December, 23 Predicting Currency Crises a practical application for risk managers, Business Economics, The National Association for Business Economists, Gale Group Unique equilibrium in a model of self-fulfilling currency attacks, American Economic Review, Vol. 88 The Logic of Currency Crises, Cahiers Economiques et Monétaries (Banque de France), no. 43, pp. 189-213 28
Obstfeld, M., 1986 Ozkan, G. F., Sutherl, A., 1995 Sachs, J., Tornell A., Velasco A., 1996 Rational Self-Fulfilling Balance-of-Payments Crises, American Economic Review, Vol. 76 Policy Measures to Avoid a Currency Crisis, Economic Journal, Vol. 15 Financial Crises in Emerging Markets: The Lessons from 1995, Brookings Papers on Economic Activity: 1, Brookings Institution, pp. 147-215 29
Appendix 1 The indicators used in the model estimation data sources The main data sources are: International Financial Statistics from the International Monetary Fund (); Bank of International Settlements (); the common data base of International Monetary Fund, World Bank, OECD, Bank of International Settlements, Economist Intelligence Unit (EIU); EUROSTAT; National Bank of Romania (NBR), other central banks. The data used were monthly except for gross domestic product, current account debt indicators that were interpolated from quarterly data. Argentina Brazil Bulgaria Czech Republic Chile Columbia South Korea Croatia Estonia Exchange rate (avg.) Exchange rate (end of period.) REER EURO STAT REC REC REC REC EURO STAT REC REC Interbank interest rate 6B 6B 6B 6B 6P 6P 6B 6B 6B CPI 64 64 64 64 64 64 64 64 64 Foreign reserves M1 M2 +35 +35 +35 +35 +35 +35 +35 +35 +35 Nongovern ment credit (NGC) GDP Budget balance 8 8 8 8 8 8 8 Export 7 7 EIU 7 7 7 7 7 7 Import 71 71 EIU 71 71 71 71 71 71 Current account 78 78 78 78 78 78 78 78 78 Transf. FDI Portfolio investments Debt (short term, private, public) 3
Philippines Hong Kong Indonesia Latvia Lithuania Malaysia Mexico Exchange rate (avg.) Exchange rate (end of period.) REER REC EURO STAT EURO STAT EURO STAT REC EURO STAT Interba nk interest rate 6B 6B 6B 6B 6B 6B 6B CPI 64 64 64 64 64 64 64 Foreign reserves M1 M2 +35 +35 +35 +35 +35 +35 +35 Nongovern ment credit (NGC) GDP Budget balance 8 8 EIU 8 Export 7 7 7 7 7 7 7 Import 71 71 71 71 71 71 71 Current account 78 78 78 78 78 78 78 Transf, FDI Portfolio investments Debt (short term, private, public) Pol REC 6B 64 +35 8 7 71 78 Romania Russia Singapore Slovakia Slovenia REC REC REC REC EURO STAT NBR 6B 6B 6P 6B 64 64 64 64 64 +35 +35 +35 +35 +35 NBR EIU NBR 8 8 8 7 7 7 7 7 71 71 71 71 71 78 78 78 78 78 31
Thail Turkey Hungary Venezuela Exchange rate (avg.) Exchange rate (end of period.) REER EURO STAT REC REC Interba nk interest rate 6B 6B 6P 6P CPI 64 64 64 64 Foreign reserves M1 M2 +35 +35 +35 +35 Nongovern ment credit (NGC) GDP Budget balance 8 8 8 Export 7 7 7 7 Import 71 71 71 71 Current account 78 78 78 78 Transf, FDI Portfolio investments Debt (short term, private, public) 32
The graphs for crisis probability within 12 months based on the multinomial logit model on 21 countries for 1994-24 Appendix 2 Brazil.8.7.6.5.4.3.2 Mar.96 Apr. Mar.2.1 Mar.95 Nov.95 Jul.96 Mar.97 Nov.97 Jul.98 Mar.99 Nov.99 Jul. Mar.1 Nov.1 Oct.4 Jul.2 Mar.3 Nov.3 Jul.4 Bulgaria.5.4.3.2.1 Jan.96 Jun.96 Nov.96 Apr.97 Sep.97 Feb.98 Jul.98 Dec.98 May.99 Oct.99 Mar. Aug. Jan.1 Jun.1 Nov.1 Apr.2 Sep.2 Feb.3 Jul.3 Dec.3 May.4 Oct.4 Czech Republic.1 Feb.97 Mar.95 Sep.95 Mar.96 Sep.96 Mar.97 Sep.97 Mar.98 Sep.98 Mar.99 Sep.99 Mar. Sep. Mar.1 Sep.1 Mar.2 Sep.2 Mar.3 Sep.3 Mar.4 Sep.4 33
Chile.2 Nov.97 Jun.98.1 Mar.96 Nov.96 Jul.97 Mar.98 Nov.98 Jul.99 Mar. Nov. Jul.1 Mar.2 Nov.2 Jul.3 Mar.4 Nov.4 Columbia.3.25.2.15 Nov.96 Mar.97 Apr.98 Mar.99.1.5 Mar.96 Nov.96 Jul.97 Mar.98 Nov.98 Jul.99 Mar. Nov. Jul.1 Mar.2 Nov.2 Jul.3 Mar.4 Nov.4 South Korea.4.35.3.25.2.15.1.5 Mar.95 Nov.95 Jul.96 Mar.97 Nov.97 Jul.98 Mar.99 Nov.99 Jul. Mar.1 Nov.1 Jul.2 Mar.3 Nov.3 Jul.4
Croatia.4.35.3.25.2.15.1 Nov.97.5 Mar.97 Nov.97 Jul.98 Mar.99 Nov.99 Jul. Mar.1 Nov.1 Jul.2 Mar.3 Nov.3 Jul.4 Estonia.2.18.16.14.12.1.8.6.4.2 Mar.95 Nov.95 Jul.96 Mar.97 Nov.97 Jul.98 Mar.99 Nov.99 Jul. Mar.1 Nov.1 Jul.2 Mar.3 Nov.3 Jul.4 Philippines.7.6.5.4.3.2.1 Sep.95 Dec.97 Mar.95 Nov.95 Jul.96 Mar.97 Nov.97 Jul.98 Mar.99 Nov.99 Jul. Mar.1 Nov.1 Jul.2 Mar.3 Nov.3 Jul.4 35
Latvia.2.18.16.14.12.1.8.6.4.2 Apr.95 Dec.95 Aug.96 Apr.97 Dec.97 Aug.98 Apr.99 Dec.99 Aug. Apr.1 Dec.1 Aug.2 Apr.3 Dec.3 Aug.4 Lithuania.2.1 Mar.95 Nov.95 Jul.96 Mar.97 Nov.97 Jul.98 Mar.99 Nov.99 Jul. Mar.1 Nov.1 Jul.2 Mar.3 Nov.3 Jul.4 Malaysia.3.25.2.15 Sep.96.1.5 Mar.95 Nov.95 Jul.96 Mar.97 Nov.97 Jul.98 Mar.99 Nov.99 Jul. Mar.1 Nov.1 Jul.2 Mar.3 Nov.3 Jul.4 36
Mexico.7.6.5.4.3.2.1 Ian.96 Apr.99 Feb. Mar.2 Mar.95 Nov.95 Jul.96 Mar.97 Nov.97 Jul.98 Mar.99 Nov.99 Jul. Mar.1 Nov.1 Jul.2 Mar.3 Nov.3 Jul.4 Pol.3.25.2.15.1 Apr.1 Nov.1.5 Mar.95 Nov.95 Jul.96 Mar.97 Nov.97 Jul.98 Mar.99 Nov.99 Jul. Mar.1 Nov.1 Jul.2 Mar.3 Nov.3 Jul.4 Romania.3.25.2.15.1 Feb.98 May.5 Mar.97 Sep.97 Mar.98 Sep.98 Mar.99 Sep.99 Mar. Sep. Mar.1 Sep.1 Mar.2 Sep.2 Mar.3 Sep.3 Mar.4 Sep.4 37
Russia.8.7.6.5.4.3.2.1 Aug.95 Jul.95 Jan.96 Jul.96 Jan.97 Jul.97 Jan.98 Jul.98 Jan.99 Jul.99 Jan. Jul. Jan.1 Jul.1 Jan.2 Jul.2 Jan.3 Jul.3 Jan.4 Jul.4 Slovakia.6.5 Jan.97.4.3.2.1 May 97 Dec.97 Mar.95 Nov.95 Jul.96 Mar.97 Nov.97 Jul.98 Mar.99 Nov.99 Jul. Mar.1 Nov.1 Jul.2 Mar.3 Nov.3 Slovenia.2.18.16.14.12.1.8.6.4.2 Mar.95 Nov.95 Jul.96 Mar.97 Nov.97 Jul.98 Mar.99 Nov.99 Jul. Mar.1 Nov.1 Jul.2 Mar.3 Nov.3 Jul.4 38
Turkey.4.35.3.25.2.15.1 Apr..5 Mar.95 Sep.95 Mar.96 Sep.96 Mar.97 Sep.97 Mar.98 Sep.98 Mar.99 Sep.99 Mar. Sep. Mar.1 Sep.1 Mar.2 Sep.2 Mar.3 Sep.3 Mar.4 Sep.4 Hungary.1.9.8.7.6.5.4.3.2.1 Mar.95 Nov.95 Jul.96 Mar.97 Nov.97 Jul.98 Mar.99 Nov.99 Jul. Mar.1 Nov.1 Jul.2 Mar.3 Nov.3 Jul.4 Venezuela.8.7.6.5.4.3.2 Apr.98 Apr..1 Nov.98 Mar.97 Nov.97 Jul.98 Mar.99 Nov.99 Jul. Mar.1 Nov.1 Jul.2 Mar.3 Nov.3 Jul.4 39
The graphs of the indicators for Romania used in the noise-to-signal approach M2/Reserves Appendix 3 6 5 4 3 2 1 Jan.94 Oct.94 Jul.95 Apr.96 Jan.97 Oct.97 Jul.98 Apr.99 Jan. Oct. Jul.1 Apr.2 Jan.3 Oct.3 Jul.4 Apr.5 Currency overvaluation 3 2 1-1 -2-3 Jan.94 Oct.94 Jul.95 Apr.96 Jan.97 Oct.97 Jul.98 Apr.99 Jan. Oct. Jul.1 Apr.2 Jan.3 Oct.3 Jul.4 Apr.5 4
Current account/gdp 2-2 -4-6 -8-1 -12 Mar.97 Dec.97 Sep.98 Jun.99 Mar. Dec. Sep.1 Jun.2 Mar.3 Dec.3 Sep.4 Export growth (over the same month of previous year) 12 1 8 6 4 2-2 Jan.94 Jul.94 Jan.95 Jul.95 Jan.96 Jul.96 Jan.97 Jul.97 Jan.98 Jul.98 Jan.99 Jul.99 Jan. Jul. Jan.1 Jul.1 Jan.2 Jul.2 Jan.3 Jul.3 Jan.4 Jul.4 Jan.5 41
Non-government credit/gdp growth rate 25 2 15 1 5-5 -1-15 -2-25 -3 Mar.97 Dec.97 Sep.98 Jun.99 Mar. Dec. Sep.1 Jun.2 Mar.3 Dec.3 Sep.4 42