of Signal Extraction Approach and Panel Logit Model

Transcription

1 Leading Indicators of Currency Crises The Integration of Signal Extraction Approach and Panel Log Model Ta-Cheng Chang Department of International Business SooChow Universy Taipei, Taiwan Tel: Jia-Ying Li Department of International Business SooChow Universy Oct, 2002

2 Leading Indicators of Currency Crises The Integration of Signal Extraction Approach and Panel Log Model Abstract Recent currency crises such as the crisis in Europe, the peso crisis in Latin America, and the financial crisis in Asian have drawn worldwide attention about constructing an early warning system. The main purpose of this paper is that we integrate the signal extraction approach and panel log model of quantative analysis to see whether the forecasting abily can be improved. We follow three stages for our empirical study to construct a composion probabily indicator for each country. In the first stage, we find some indicators according to the suggestion of related lerature, and then we employ signal extraction approach to obtain some appropriate indicators. In the second stage, we use these selected indicators to calculate the probabilies of the monthly occurrence of currency crises by panel log model. Finally, we compute the optimal threshold of the composion probabily of the happening of currency crisis for each country. The results show that the predict abily of our integration is in average better than tradional method. In average, our early warning system is better than signal extraction approach. Besides, we calculate threshold probabily of individual country and provide these probabily values for predict currency crises. We also take Taiwan for example to verify the effectiveness of our compose indicator. There is about 70% correct prediction, and our compose probabily indicators are proven to be useful.

3 1. Introduction Recent currency crises have drawn worldwide attention not only because they happened frequently but also because they have been experienced by well-developed economies. There were several currency crises that occurred in the 1990 s: the crisis in Europe, the peso crisis in Latin America, and the financial crisis in Asian. On July , the monetary authories of Thailand failed to maintain the pegged exchange rate and then changed to floating exchange rate regime. This change led to sharp devaluation of the Thai bath, and the Asian currency crisis took place hereafter. The most serious suation in this currency crisis happened in Thailand, Malaysia, Indonesia, Philippines, and South Korea. These economies suffered many difficulties, such as sharp drop of stock price, decline in economic growth, and increase in unemployment rate, etc. Because a burst of a currency crisis will bring about crical problems in many aspects, is necessary to find out the determinants of a currency crisis and s solution before comes up. Therefore, we are interested in whether currency crises are predictable events wh early warning signals. There are also many empirical studies on the currency crisis, and according to Hawkins and Klau (2000), these studies can be classified into three major types; however, we focus on two of these types: 1 First, the signal extraction approach is to evaluate the usefulness of many different variables in signaling a pending or potential crisis. Threshold values are chosen for each indicator so as to strike a balance 1 There is still another type of these studies called qualative comparisons, graphically comparing economic fundamentals immediately preceding a financial crisis wh those in normal times or in a control group of countries which did not suffer a crisis. For example, Eichengreen, Rose, and Wyplosz (1995), Frankel and Rose (1996).

4 between the risk of adopting many false signals and the risk of missing the crisis altogether. The noise-to-signal ratio is used to decide which indicators are better. These selected indicators are useful when the noise-to-signal ratio is less than one, and the smaller, the better. The advantage of signal extraction approach is that the importance of individual indicator can be ranked and the abily of avoidance of false signals is considered at the same time. Kaminsky, Lizondo, and Reinhart (1998, hereafter KLR), Kaminsky and Reinhart (1999, hereafter KR), and Edison (2000) used this approach. Second, econometric approach eher use regressions to explain some measures of exchange rate pressure or log or prob models to test whether indicators are associated wh a higher probabily of a currency crisis. Eichengreen et al. (1995), Frankel and Rose (1996), and Sachs et al. (1996) were examples of this approach. According to KLR, the method that estimates the one-step-ahead probabily of currency crisis has the advantages of summarizing information about the likelihood of a crisis in one useful number, the probabily of a crisis. Also, this approach considers all the variables simultaneously. Signal extraction approach and log model are implemented separately in leratures and they both have some disadvantages. The signal extraction approach provides a metric for ranking the indicators according to their abily to accurately predict crises and avoid false signals. However, the causes of currency crises are usually complicated, and we should not consider indicators individually. Furthermore, the performances of individual indicator are volatile. According to KLR, there are also some limations of log model. First, the nonlinear nature of log model makes difficult to assess the marginal contribution of an indicator to the probabily of a crisis. Second, the log model does not provide a transparent reading of where the macroeconomic problems are, that is, the impact of an individual

5 variable is less easily detected. Whin this approach, is difficult to judge which of the variables is out of line, making less than ideally sued for the purpose of surveillance and preemptive action. Third, we think that the annual data used in log model makes the early warning system not sensive enough. According to World Economic Outlook of IMF (1998), the characteristics of a good early warning system is as follows: First, a set of indicators could be identified that could detect future crises sufficiently early and wh a high degree of certainty, while not giving false signals. Second, the estimated coefficients of indicators are statistically significant. Signal extraction approach and log model are not sufficient for above condions. Although some researches such as Kaminsky (1998) and Edison (2000) constructed a single compose leading indicator as a weighted-sum of the indicators, where each indicator is weighted by the inverse of s noise-to-signal ratio, the deficiencies mentioned above still exist. Many researches are interested in constructing a useful compose leading indicators. The main purpose of this paper is that we integrate the signal extraction approach and panel log model of quantative analysis to see whether the forecasting abily can be improved. The compose indicator that we construct considers all the advantages of these two methods. We follow three stages for our empirical study to construct a composion probabily indicator for each country. In the first stage, we find some indicators according to the suggestion of related lerature, and then we employ signal extraction approach to extract six good indicators. In the second stage, we use these selected indicators to calculate the probabilies of the monthly occurrence of currency crises by panel log model. Finally, we compute the optimal threshold of the composion indicator of the happening of currency crisis for each country. The results show that the predict abily of our integration is in average better than tradional method. The results are also que stable of different combinations of indicators.

6 This paper is organized as follows. In section 2, we review theoretical and empirical lerature. In section 3, we describe the data and methodology. Section 4 is the empirical results, and we will explain the empirical process. Section 5 is the conclusion and suggestion to further studies. 2. Lerature Review Sachs et al. (1996) analyze the period immediately after the crash of the Mexican peso in December They find that low international reserves relative to broad money, real exchange rate appreciation and a weak banking system explain about 70 percent of the variation of their crisis index, a compose measure of the change in reserves and the nominal depreciation. Nevertheless, their sample is far from random in term of eher time or country choice. In addion, they do not distinguish between attacks which are unwarranted by fundamentals but are triggered by macroeconomic similary. Kaminsky, Lizondo, and Reinhart (1998) use monthly data for 15 developing and 5 industrial countries from 1970 to Their definion of crisis is based on the idea of ERW, but because the interest rate data are not available for all countries, they use an exchange market pressure index that consists of changes in the nominal exchange rate and in the international reserves. A crisis is called if the index exceeds s mean by more than three standard deviations for that country. They use the signal extraction approach to observe the evolution of fifteen macroeconomic indicators before the crisis. They conclude that the weakness of the fundamental economics is the cause of currency crisis, including overvaluation of the real exchange rate, reserve loss, high ratio of broad money to reserves, and the slowdown in economic activy (industrial output, export).

7 3. Data and Methodology 3.1 Crisis Window, Sample Countries, and Time Periods Our sample countries are composed of 25 industrial and non-industrial countries. The monthly data of are used in this paper. Our empirical data are come from International Financial Statistics database of International Monetary Fund, World Development Indicators database of World Bank, AREMOS database, and Datastream database. Since we need to decide a crisis window to make the indicators have abilies to foretell crises. We use 24 months in this paper. 3.2 The Definion of a Crisis Following Sachs, Tornell and Velasco (1996), and Kaminsky and Reinhart (1996) we identify crises by looking at an index of exchange market pressure (EMP) defined as a weighted average of percentage changes in the nominal exchange rate and (the negative of) percentage changes in international reserves. Since the volatilies of reserves and exchange rates are different, the weights are chosen so as to prevent any one of the series from dominating the index by dividing the standard deviation of s change rate. Therefore, the exchange market pressure index is as follows: EMP t e t t = (1) σ e R σ R Where e t and R t denote the national currency per foreign currency (U.S. dollars) and the international reserves at the time t, respectively. and Rt denote the et change rate of the foreign exchange rate and the international reserves at the time t, respectively. σ e and σ R are the standard deviations of and Rt, respectively. et

8 From equation (1), we know that both deprecation of the exchange rate and (or) decline in the international reserves raise the EMP. The intuion behind the construction of the index is that when a currency is under a speculative attack, the monetary authories can respond to the attack by devaluing the currency, running down international reserves or raising interest rates. Because interest rate data of most emerging markets are neher available nor reliable, the index is usually defined excluding interest rates. The advantage of the weighted index is that associates crises wh both successful and unsuccessful speculative attacks. A successful attack will cause the depreciation of foreign exchange rate, and the reserves will loss if the authories try to defend the attacks. In practice, a crisis is identified whenever the EMP exceeds certain threshold value. A crisis is defined as an event that EMP is more than n standard deviations above the mean: 2 Crisis = 1 if EMP > µ + n σ 0 otherwise t EMP EMP (2) 3.3 Signal Extraction Approach We use signal extraction approach to monor several indicators that tend to exhib unusual behavior before a crisis. If a variable behaves differently before a crisis, an extreme value for this variable provides a warning signal indicating that a currency crisis may occur whin a given period. This is the most prominent model proposed by KLR. The question what value should be considered extreme is determined as follows: if X exceeds a certain threshold level, we call issuing a signal, then S = 1. If X does not exceed the threshold level, is issuing no signals; t 2 Eichengreen, Rose and Wyplosz (1995), choose n=1.5. Glick and Hutchison (1999) set n=2. Edison

9 therefore, S = 0 : t S t 1 = 0,, if if X X t t > X X (3) Where X t is a value of certain indicator at the time t, X is a certain threshold level, S t is the signal at the time t. In order to examine the effectiveness of individual indicators, would be useful to examine the performance of each indicator by the matrix in Table 3.1. Table 3.1 Matrix for evaluate the performance of an indicator Crisis (whin k months) No crisis (whin k months) Signal was issued A B No signal was issued C D In Table 3.1, A is the number of months in which the indicator issued a good signal, B is the number of months in which the indicator issued a bad signal or noise, C is the number of months in which the indicator failed to issue a signal which would have been a good signal, and D is the number of months in which the indicator did not issue a signal that would have been a bad signal. 3.4 Log Model In the limed dependent regression models (log or prob models), the currency crisis indicator is modeled as a dummy variable. Recent leratures on currency crises tend to capture the cause of crises and the prediction for crises in terms of probabily, by using the log model. However, the number of crises is usually limed using this method. There are only a few ones and more zeros in the sample, resulting in poor estimation results. (2000) set n=2.5. KLR (1998), and Berg and Pattillo (1999) set n=3.

10 The dependent variable is represented by a binary choice variable y = 1, if the specified event happens and y = 0, if the event does not happen for country i at time t. Let P is the probabily of the happening of the event, then E ( y ) = P (1) + (1 P )(0) = P P = P( y = 1) = E( y X ) = F( X ' β) For the log model, F ( X ' β ) is the logistic distribution function. These two models constrain Pˆ to lie between 0 and 1. A country experiences a crisis if the explanatory variables exceed certain threshold, i.e. y 1 = 0,, if if y * y * > 0 0 where y * = X ' β + ε, so that * P ( y = 1) = P( y > 0) = P( ε > X ' β) = 1 F( X ' β ε ) = F( X ' β ) Because the logistic distribution function is symmetric around zero, the last equaly holds. 4. Empirical Results In this section, we discuss our empirical process, and we employ RATS5.01 to complete the empirical study. 4.1 Currency Crises Period Time Periods, Sample Countries, and Crisis Window As we mentioned in the section 3, our sample are monthly data of 8 industrial countries: Australia, Canada, Denmark, Iceland, Japan, New Zealand, Norway, and Swzerland, and 17 non-industrial countries: Chile, Indonesia, Israel, Jamaica, Korea, Malaysia, Mexico, Morocco, Nigeria, Peru, Philippines, Singapore,

11 Sri Lanka, Taiwan, Thailand, Venezuela, and Zimbabwe. 3 Because we want to investigate the features of different regions, we will classify our sample into three groups: industrial countries, non-industrial countries, and all countries. The crisis window is set at 24 months following KLR inially in this paper. Choice of Indicators We classify our 19 indicators into five categories: macroeconomic indicators, external indicators, financial instutional indicators, instutional indicator, and crisis indicators. The followings are the indicators of each category: A. Macroeconomic indicators: inflation, excess M1 balance, M2/reserves, M2 multiplier, stock price, and real interest rate. B. External indicators: real exchange rate, reserves, imports, exports, depreciation, and real interest-rate differential. C. Financial instutional indicators: bank deposs, domestic cred/gdp, foreign gross liabilies/gdp, and lending-depos rate ratio. D. Instutional indicator: financial liberalization (dummy variable) 4. E. Crisis indicators: banking crisis (CMAX) 5, and banking crisis (dummy variable). More detailed definions of all the indicators and their sources are provided in the Appendix. In order to eliminate seasonal effects, we follow KLR and define all indicators on a given month to be the percentage change in the level of the variable wh respect to s level a year earlier except real exchange rate, interest rates, and excess M1 balance. Crises Identified 3 The sample countries of KLR (1998) are Argentina, Bolivia, Brazil, Chile, Colombia, Denmark, Finland, Indonesia, Israel, Malaysia, Mexico, Norway, Peru, Philippines, Spain, Sweden, Thailand, Turkey, Uruguay, and Venezuela. 4 Refer to Demirgüç-Kunt and Detragiache (1998).

12 According to KLR, a crisis is identified by the behavior of the foreign exchange market pressure index (EMP) defined in equation (1). Referring to equation (2), periods in which the index is above s mean by more than n standard deviations are defined as crisis. The detailed currency crisis periods are presented in the Table 4.1. [Insert Table 4.1] 4.2 Performance of the Individual Indicators The performance of each indicator is assessed on the basis of the noise-to-signal ratio, which is the ratio of bad signals to good ones, and the share of crises called correctly. Table 4.2, Table 4.3 and Table 4.4 are the information on the performance of individual indicator of industrial countries, non-industrial countries, and all countries respectively. The first column display the crical value of the percentile of each indicator where the noise-to-signal ratio is minimum, and the crical value is determined by using a grid search from 0.8 to 0.9. The second to fifth column display the numbers of A, B, C, and D in term of the matrix in Table 3.1. The sixth column shows the number of good signals issued by the indicators, expressed as a percentage of the number of months in which good signals could have been issued (A/(A+C) in terms of the matrix in Table 3.1). The higher the number in this column, the better the indicator. In industrial countries, the inflation is the indicator that issued the highest percentage of possible good signals (25.72%), while the real interest-rate differential issued the lowest percentage of possible good signals (8.20%). In non-industrial countries, the performance of the real exchange rate is the best (27.31%), while the performance of the lending-depos rate ratio is the poorest (6.61%). In all countries, the real exchange rate is the indicator that issued the highest percentage of possible good signals (26.73%), while the M2 multiplier issued the lowest percentage of possible good signals (11.25%). The seventh column 5 The definion of CMAX refer to Anne Vila (2000).

13 measures the performance of individual indicators regarding sending bad signals. It shows the number of bad signals issued by the indicator, expressed as a percentage of the number of months in which bad signals could have been issued (B/(B+D) in terms of the matrix in Table 3.1). In contrast to column six, the lower the number in this column, the better the indicator. In industrial countries, the real exchange rate is the indicator that issued the lowest percentage of possible bad signals (8.17%), while the M2 multiplier issued the highest percentage of possible bad signals (19.21%). In non-industrial countries, the performance of the real exchange rate is the best (5.01%), while the performance of the banking crisis is the poorest (14.16%). In all countries, the real exchange rate is the indicator that issued the lowest percentage of possible bad signals (6.09%), while the lending-depos rate ratio issued the highest percentage of possible bad signals (20.82%). [Insert Table 4.2, Table 4.3 and Table 4.4] The information about the indicators abily to issue good signals and to avoid bad signals can be combined into a measure of the noisiness of the indicators. The eighth column reports the noise-to-signal ratio, calculated as the ratio of false signals as a portion of months in which there is no crisis [B/(B+D)] relative to the good signal which is the proportion of months in which there is a crisis [A/(A+C)]. We can take out some indicators failing to predict a crisis by noise-to-signal ratio. Other things being equal, the lower the number in this column, the better the indicator. Ratios that are equal to or greater than one imply that the indicator is not helpful in predicting a crisis. Therefore, in industrial countries, the M2 multiplier, excess M1 balance, real interest-rate differential, exports, and lending-depos rate ratio are dropped; while in non-industrial countries, only the real interest rate and lending-depos rate ratio are dropped. In all countries, the real interest rate, real interest rate differential, and lending-depos rate ratio are dropped. Real exchange

14 rate is the best indicators in the three groups. Another way of interpreting the abily of predict a crisis is by comparing the probabily of a crisis condional on a signal from the indicator, A/(A+B) in terms of the matrix, wh the uncondional probabily of a crisis, (A+C)/(A+B+C+D) in terms of the matrix. To the extent that the indicator has useful information, the condional probabily should be higher than the uncondional probabily. The ninth column of the table is the estimates of the condional probabilies, while the tenth column is the difference between the condional and uncondional probabilies for each of the indicators. consistent. From these estimates, we can find that these two rules are actually When the noise-to-signal ratio is higher than uny, the difference between the condional and uncondional probabilies is negative. In Table 4.5, we compare our noise-to-signal ratio results wh other three researches: KLR (1998), KR (1999), and Edison (2000). Our results have ltle differences from the other three researches results. For example, real exchange rate is best indicator in all results. In contrast, lending-depos rate ratio is the worst in all results. The real interest rate in our result is worse than the other researches, but the imports, bank depos are better than theirs. [Insert Table 4.5] Change the crisis window and multiplier We change our inial sets in two ways: first, we change n from 3 to 1.5, then we change the crisis window from 24 months to 12 months. Table 4.6 shows the results under different condions. The results have ltle difference from the results of 24 months crisis window and 3 standard deviations. Therefore, we still apply the set of 24 months crisis window and 3 standard deviations to pick out useful indicators from the 19 indicators. [Insert Table 4.6]

15 According to this rule, the six indicators whose noise-to-signal ratios are minimum and lower than uny will be chosen. In industrial countries, the real exchange rate (the noise-to-signal ratio is ), imports (0.4602), reserves (0.5402), foreign gross liabilies/gdp (0.5438), depreciation (0.5815), and bank deposs (0.5918) are chosen. In non-industrial countries, the real exchange rate (0.1833), reserves (0.4072), M2/reserves (0.4253), depreciation (0.4759), excess M1 balances (0.5089), and exports (0.5876) are chosen. In all countries, the real exchange rate (0.2278), reserves (0.4434), M2/reserves (0.4777), depreciation (0.5033), excess M1 balances (0.6131), imports (0.6262) are chosen. 4.3 Regression of Panel Log Model From the signal extraction approach, we extract six good indicators from all indicators for each group. In the second stage, these indicators will be the necessary explanatory variables in the regression, and all indicators are one-month-ahead probabily of currency crisis. The dependent variable is the currency crisis, and s value is eher 0 or 1. We adopt maximum likelihood method to estimate the coefficients of log regression. The results are presented in Table 4.7, Table 4.8, and Table 4.9, and the equation 1 of each table is the base regression. In the industrial countries (Table 4.7), all indicators are not statistically significant except the reserves which being significant at 1% level. Lower reserves seem to raise the odds of crises. In the non-industrial countries (Table 4.8), the M2/reserves, depreciation, and exports are statistically significant at 5%, 1%, and 1% level individually. Loose monetary policy (higher M2), sharp depreciation, and low exports tend to increase the probabily of currency crisis. Table 4.9 is the regression results for all sample countries. The M2/reserves, reserves, and depreciation are statistically significant at 5%, 1%, and 1% level individually. Adequate monetary policy and sufficient foreign

16 reserves can decrease the probabily of currency crisis. [Insert Table 4.7, Table 4.8, Table 4.9] Robustness It is preferable to test the robustness of the results obtained by the six good indicators. In addion to the six base indicators, other indicators, whose noise-to-signal ratios are lower than 1, will be added into the based regression sequentially. The equation 10 in Table 4.7, the equation 13 in Table 4.8, and the equation 12 in Table 4.9 are the regressions that contain all indicators whose noise-to-signal ratios are lower than 1. The last equations in these three tables are the regressions that contain all indicators whose noise-to-signal ratios are lower than 1 and financial liberalization. In Table 4.7, the reserve is statistically significant at 1% almost in all regression equations. The expected sign are also consistent wh equation 1. Similarly, the M2/reserves, depreciation, and exports are robust in non-industrial countries in Table 4.8. Also, the M2/reserves, reserves, and depreciation are robust in all countries in Table 4.9. Table 4.10 is a summary of the results of signal extraction approach and log model. The inflation, reserves, depreciation, bank deposs, and foreign gross liabilies/gdp are important factors in both industrial and non-industrial countries. The M2/reserves, M2 multiplier, and excess M1 balances are related to monetary policies. These three indicators are not statistically significant in industrial countries, but the M2/reserves and M2 multiplier are significant in non-industrial countries. The independence of the monetary authories is not sufficient; therefore, the adequacy of monetary policies is an important factor of currency crises. The export is not statistically significant in industrial countries, while is significant in non-industrial countries. Non-industrial countries are short of capal, and they depend on exports deeper than industrial countries. The depreciation is statistically

17 significant both in industrial and non-industrial countries. Besides, referring to Table 4.7 and Table 4.8, we can observe that the depreciation is more important in non-industrial countries than industrial countries. The foreign exchange rate in industrial countries is more stable than non-industrial countries; therefore, the impact of the depreciation on currency crises is not so obvious in industrial countries. In Table 4.8, the coefficient of domestic cred/gdp is negative, and means that cred contraction will cause a currency crisis. Non-industrial countries inherently lack for capal, and cred contraction will cause the cost of capal of firms increase. The decrease of output will then lead to a recession. We can conclude that the country characteristics are actually essential and that is the reason why we classify our sample into three groups. [Insert Table 4.10] 4.4 Performance of Compose Indicators In the last stage, we calculate the optimal threshold of the composion probabily of the occurrence of currency crisis. The noise-to-signal ratios in the lower part of Table 4.7, Table 4.8, and Table 4.9 represent the predictive abily of each set of indicators. The lower the ratio, the better the compose of indicators. In the industrial countries (Table 4.7), the noise-to-signal ratio of equation 1 is , and that of the other eight equations are also stable, ranging from 0.45 to In the non-industrial countries (Table 4.8), the noise-to-signal ratio of equation 1 is , and the ratio is lower than those of industrial countries. The noise-to-signal ratio of the other eleven equations range from 0.3 to 0.5, and the variabily of the ratios are small. In all countries (Table 4.9), the noise-to-signal ratio of equation 1 is Interestingly, the noise-to-signal ratio of base regression of all countries is higher than which of both industrial and non-industrial countries. It might indicate

18 that individual estimation according to region characteristic is better than altogether estimation. Compared to the performance of individual indicators 6 (Table 4.6), the noise-to-signal ratios of compose indicators are the second small ratio in industrial and non-industrial countries. It shows that the predictive abily of the compose indicators is better than individual indicator. But in the all countries, the predictive abily of compose indicators is not significantly better than individual indicator. It perhaps means that the effectiveness of mixed estimation is not very good. 4.5 The Threshold of Currency Crisis In order to predict currency crisis, we provide the threshold of the probabily of currency crisis of individual country in Table We can use the compose indicators to predict the currency crisis according to this table. For example, we want to know whether Taiwan in the next month will have currency crisis or not. We can use the data of the six indicators (real exchange rate, reserves, M2/reserves, depreciation, excess M1 balances, and exports) for non-industrial countries to calculate a compose probabily of currency crisis. If this calculated probabily is higher than 1.11%, we consider that currency crises will happen in Taiwan in the next month. It is worthy to pay attention to the threshold probabilies, which are almost very low, not higher than 10%. The reason is that the n of EMP is three standard deviations, and the identified currency crises are not very much. Consequently, there are only a few ones in the sample, compared to a huge amount of zeros, resulting in poor estimation results. [Insert Table 4.11] 4.6 Out Sample Prediction The Evidence of Taiwan We use monthly data from December 2000 to November 2001 of Taiwan to test 6 The benchmark is the window=24, and n=3.

19 and examine the performance of the compose probabily indicator during January 2001 to December The following equation is the base regression of non-industrial countries that we estimate (see Table 4.8): y t = M 2Rt EM1Bt RERt REVt 1 (2.0214) ( ) ( ) ( ) DEP t EX t 1 (5.4177) ( ) Where M 2 R represents the M2/reserves, EM1 B represents the excess M1 balances, RER represents the real exchange rate, REV represents the reserves, DEP represents the depreciation, and EX represents the exports. We use the following equation to calculate the probabily values: exp( y) P( Currency Crisis) = 1+ exp( y) The results are showed in Table The first column is the currency crisis of Taiwan that we observe. The second column is the compose probabily of the occurrence of currency crisis in the next month. The compose probabily threshold of Taiwan is (see Table 4.11). The last column is the signal issued. There are four times of false prediction, 2001/6, 2001/8, 2001/9, and 2001/10 respectively. There is about 70% correct prediction, and our compose probabily indicators are proven to be useful. [Insert Table 4.12] 5. Conclusions This paper is an extension of KLR model. We combine the signal extraction approach and the panel log model to construct a set of compose indicators for different groups to predict currency crises. In average, our early warning system is better than signal extraction approach. Besides, we calculate threshold probabily of

20 individual country and provide these probabily values for predict currency crises. In this paper, we find that region feature indeed affect the indicators chosen to predict currency crisis. In industrial countries, the real exchange rate, imports, reserves, foreign gross liabilies/gdp, depreciation, and bank deposs are chosen. In non-industrial countries, the real exchange rate, reserves, M2/reserves, depreciation, excess M1 balances, and exports are effective indicators. In all countries, the real exchange rate, reserves, M2/reserves, depreciation, excess M1 balances, and imports are useful for predicting currency crises. It is difficult to construct an early warning system to predict currency crises because of s uncertainty. Besides, data of some countries are not complete. Here are some suggestions for further researches. First, in order to refine the early warning system, addional explanatory variables should be considered, for example, variables that capture contagion effect or polical stabily. Second, the degree of a currency crisis is ignored in our model. Future researches can try to design more than one threshold to measure the severy of crises. Finally, the lag period is one period for all indicators in the log regression in our study, the optimal lag periods can be decided by time-series method. References Kaminsky G. and C. Reinhart, The Twin Crises: The Causes of Banking and Balance-of-Payments Problems, IMF Working Paper 97/79, Kaminsky G. and C. Reinhart, The Twin Crises: The Causes of Banking and Balance-of-Payments Problems, American Economic Review, Vol. 89, No. 3, June 1999, pp

21 Table 4.1 Currency crisis periods identified Country name Currency crisis periods (3 standard deviations) Financial liberalization Industrial countries Australia 1974/9,1976/11,1985/2,1986/ Canada 1976/11,1980/3,1981/7,1992/ Denmark no crises Iceland 1974/9,1975/2,1982/8,1983/ Japan 1979/ New Zealand 1975/8,1985/ , Norway 1978/11,1986/5,1991/3,1992/12,1997/ Swzerland 1981/ Non-industrial countries Chile 1971/7,1972/9,1973/5,1973/10,1974/12,1975/6,1975/12,1985/ Indonesia 1975/5,1978/11,1986/9,1998/1,1998/5-1998/ Israel 1974/11,1977/11,1983/10,1984/7,1984/9-1984/ Jamaica 1978/5,1983/11,1991/ Korea 1980/1,1997/ / , Malaysia 1997/7-1997/8,1997/ / Mexico 1976/9-1976/10,1981/9,1982/2,1982/12,1994/ Morocco 1981/1,1982/2,1983/ Nigeria 1986/10,1992/3,1992/ / Peru 1976/6,1988/9,1988/11,1990/ , Philippines 1970/2,1983/9-1983/10,1984/6,1986/2,1997/ Singapore 1970/12,1975/7,1997/12,1998/ Sri Lanka 1977/11,1998/ Taiwan 1995/8,1997/ Thailand 1997/7-1997/8,1997/ / Venezuela 1984/2,1986/12,1989/3,1994/5,1995/12,1996/ Zimbabwe 1997/11,1998/8,2000/

22 Table 4.2 Performance of indicators (industrial countries) A. Macroeconomic indicators (1) (2) (3) (4) (5) Good signals as percentage of possible good signals (6) Bad signals as percentage of possible bad signals (7) Indicators percentile A B C D A/(A+C) B/(B+D) Noise/signal (8) [B/(B+D)]/ [A/(A+C)] P(crisis/signal) (9) Inflation M2/reserves M2 multiplier Stock price Real interest rate Excess M1 balances B. External indicators Real exchange rate Reserves Imports Depreciation Real interest-rate differential Exports C. Financial instutional indicators Bank deposs Domestic cred/gdp Foreign gross liabilies/gdp Lending-depos rate ratio D. Instutional indicator Finanacial liberalization E. Crisis indicator Banking crisis (CMAX) Banking crisis (D.V.) Note: Crisis window is set at 24 months, and n=3. A/(A+B)

23 Table 4.3 Performance of indicators (non-industrial countries) A. Macroeconomic indicators (1) (2) (3) (4) (5) Good signals as percentage of possible good signals (6) Bad signals as percentage of possible bad signals (7) Indicators percentile A B C D A/(A+C) B/(B+D) Noise/signal (8) [B/(B+D)]/ [A/(A+C)] P(crisis/signal) (9) Inflation M2/reserves M2 multiplier Stock price Real interest rate Excess M1 balances B. External indicators Real exchange rate Reserves Imports Depreciation Real interest-rate differential Exports C. Financial instutional indicators Bank deposs Domestic cred/gdp Foreign gross liabilies/gdp Lending-depos rate ratio D. Instutional indicators Financial liberalization E. Crisis indicator Banking crisis (CMAX) Banking crisis (D.V.) Note: Crisis window is set at 24 months, and n=3. A/(A+B)

24 Table 4.4 Performance of indicators (all countries) A. Macroeconomic indicators (1) (2) (3) (4) (5) Good signals as percentage of possible good signals (6) Bad signals as percentage of possible bad signals (7) Indicators percentile A B C D A/(A+C) B/(B+D) Noise/signal (8) [B/(B+D)]/ [A/(A+C)] P(crisis/signal) (9) Inflation M2/reserves M2 multiplier Stock price Real interest rate Excess M1 balances B. External indicators Real exchange rate Reserves Imports Depreciation Real interest-rate differential Exports C. Financial instutional indicators Bank deposs Domestic cred/gdp Foreign gross liabilies/gdp Lending-depos rate ratio D. Instutional indicator Finanacial liberalization E. Crisis indicator Banking crisis (CMAX) Banking crisis (D.V.) Note: Crisis window is set at 24 months, and n=3. A/(A+B)

25 Table 4.5 Noise-to-signal ratio results compared wh other leratures Industrial countries Nonindustrial countries All countries KLR(1998) KR(1999) Edison(2000) n=3 n=3 n=3 n=3 n=3 n=2.5 A. Macroeconomic indicators Inflation M2/reserves M2 multiplier Stock price Real interest rate Excess M1 balances B. External indicators Real exchange rate Reserves Imports Depreciation Real interest-rate differential Exports C. Financial instutional indicators Bank deposs Domestic cred/gdp Foreign gross liabilies/gdp Lending-depos rate ratio D. Instutional indicator Financial liberalization(d.v.) E. Crisis indicator Banking crisis (CMAX) Banking crisis (D.V.)

26 Table 4.6 Noise-to-signal ratio n = 1.5 n = 3 n = 1.5 n = 3 n = 1.5 n = 3 n = 1.5 n = 3 n = 1.5 n = 3 n = 1.5 n = 3 A. Macroeconomic indicators Inflation M2/reserves M2 multiplier Stock price Real interest rate Excess M1 balances B. External indicators Real exchange rate Reserves Imports Depreciation Real interest-rate differentia Exports C. Financial instutional indicators Bank deposs Domestic cred/gdp Foreign gross liabilies/gd Lending-depos rate ratio D. Instutional indicator Financial liberalization(d.v.) E. Crisis indicator Banking crisis (CMAX) Banking crisis (D.V.) Industrial countries (k=24) Industrial countries (k=12) Nonindustrial countries (k=24) Nonindustrial countries (k=12) All countries (k=24) All countries (k=12)

27 Table 4.7 Multivariable regression- Industrial countries Variable A. Macroeconomic indicators Inflation (1.8984)* ( ) M2/reserves (0.6881) ( ) M2 multiplier Stock price ( ) ( ) Real interest rate (0.3779) (2.5345)*** Excess M1 balances B. External indicators Real exchange rate ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) Reserves ( )*** ( )*** ( ) ( )*** ( )*** ( )*** ( )*** ( )*** ( )*** ( ) Imports ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) Depreciation ( ) ( ) (0.0281) ( )** ( ) ( ) ( ) ( ) ( ) ( )* Real interest-rate differential Exports C. Financial instutional indicators Bank deposs ( ) ( ) ( ) ( ) ( ) ( )** ( ) ( ) ( ) ( ) Domestic cred/gdp (2.7904)*** (1.3866) Foreign gross liabilies/gdp (1.4721) (1.3207) (1.4774) (2.0057)** (1.3729) (0.3667) (1.5974) (1.4571) (1.3269) (0.5666) Lending-depos rate ratio D. Instutional indicator Financial liberalization(d.v.) ( ) E. Crisis indicator Banking crisis (CMAX) (0.3125) (0.5400) Banking crisis (D.V.) ( ) Noise-to-signal ratio Percentile Usable Observations Cases Correct Log Likelihood Pseudo-R** Note: * represent a significant level of 10, ** represent a significant level of 5, *** represent a significant level of 1, and the number of parentheses is t-value

28 Table 4.8 Multivariable regression- Nonindustrial countries Variable A. Macroeconomic indicators Inflation (2.1516)** (0.3032) M2/reserves (2.0214)** (1.3842) (2.4259)*** ( ) (2.1090)** (1.1189) (2.4023)*** (2.3034)** (1.9751)** (2.0297)** (1.3913) (1.5065) ( ) M2 multiplier ( )** ( ) Stock price ( ) (0.0241) Real interest rate Excess M1 balances ( ) ( ) ( ) (0.1936) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) B. External indicators Real exchange rate ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) Reserves ( ) ( )* ( ) ( )*** ( ) ( )* ( ) ( ) ( ) ( ) ( ) ( ) ( )*** Imports ( ) (0.9929) Depreciation (5.4177)*** (0.4156) (5.5131)*** (3.5106)*** (5.2039)*** (3.1226)*** (4.5644)*** (5.4279)*** (5.9430)*** (5.3303)*** (4.9088)*** (5.4264)*** (1.5411) Real interest-rate differential (0.0776) ( ) Exports ( )*** ( )** ( )*** ( )* ( )** ( ) ( )*** ( )*** ( )*** ( )*** ( )*** ( )*** ( )*** C. Financial instutional indicators Bank deposs ( )*** ( ) Domestic cred/gdp ( )* ( ) Foreign gross liabilies/gdp ( )*** (0.0954) Lending-depos rate ratio D. Instutional indicator Financial liberalization(d.v.) (0.1581) E. Crisis indicator Banking crisis (CMAX) (2.2255)** (1.1142) Banking crisis (D.V.) ( ) Noise-to-signal ratio Percentile Usable Observations Cases Correct Log Likelihood Pseudo-R** Note: * represent a significant level of 10%, ** represent a significant level of 5%, *** represent a significant level of 1%, and the number of parentheses is t-value

29 Table 4.9 Multivariable regression- All countries Variable A. Macroeconomic indicators Inflation (1.5604) (1.9052)* M2/reserves ( )** (1.3096) (2.5008)*** ( ) (1.8580)* (2.5578)*** (2.3278)** (2.2060)** (2.0686)** (1.6801)* (1.7274)* ( ) ( ) M2 multiplier ( )** ( ) ( ) Stock price Real interest rate ( ) ( ) ( ) Excess M1 balances ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) B. External indicators Real exchange rate (0.9565) (1.8097)* (0.9283) (0.1047) (0.9202) (1.0379) (0.5898) (0.4512) (0.9709) (0.9044) (0.9217) (0.4693) (0.5907) Reserves ( )*** ( )*** ( )*** ( )*** ( )** ( )** ( )*** ( )** ( )*** ( )*** ( )*** ( )*** ( )*** Imports ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) ( ) (0.1671) (0.2265) Depreciation Real interest-rate differential (5.2142)*** ( ) (5.2842)*** (2.8242)*** (5.3003)*** (4.2134)*** (4.7815)*** (5.3743)*** (5.1266)*** (4.8775)*** (5.2319)*** ( ) ( ) Exports C. Financial instutional indicators ( )*** ( )*** ( )*** Bank deposs ( )*** ( ) ( ) Domestic cred/gdp ( ) ( ) ( ) Foreign gross liabilies/gdp Lending-depos rate ratio ( )** (0.2104) (0.1474) D. Instutional indicator Financial liberalization(d.v.) E. Crisis indicator ( ) (1.7879)* Banking crisis (CMAX) Banking crisis (D.V.) ( ) (1.3147) (1.6857)* (1.7508)* Noise-to-signal ratio Percentile Usable Observations Cases Correct Log Likelihood Pseudo-R** Note: * represent a significant level of 10%, ** represent a significant level of 5%, *** represent a significant level of 1%, and the number of parentheses is t-value