Early Warning Systems for Financial Crises in Indonesia:
Signal Extraction Approach and Multivariate Logit Model
University of Groningen
MSc. Economics Thesis
Eisha M. Rachbini
S1939602
Supervisor: J.P.A.M. Jacobs
ii Abstract
This paper develops early warning systems for financial crises in Indonesia. It aims at determining the leading indicators for the case of a single country by using two approaches, i.e. signal extraction and multivariate logit model. The former is applied to extract several informative indicators related to the occurrence of financial crises. This approach offers flexibility in selecting the indicators by considering their relationship to the crisis occurrence. Corresponding to this approach, the latter confirms which indicators are useful as crisis signals. It provides the robustness evaluation in the relationship between the leading indicators and the likelihood of crisis occurrence. The study uses quarterly data for Indonesia in the period of 1984 until 2006. To evaluate the out-of-sample performance of multivariate logit model, the data is extended for the period of 2007 until 2009. It is found that real GDP growth, changes in stock prices and the inflation rate are the leading indicators for financial crises in Indonesia, according to both approaches. In addition, several indicators such as the national savings rate, the growth of industrial production, M2 growth, M2 multiplier, M1 growth, excess real M1 balances, the differences between lending and deposit rate, the ratio of domestic credit to GDP, commercial bank deposits, ratio of M2 to international reserves, import growth and growth of world oil price are also informative in predicting financial crises.
1 1 Introduction
The collapse of a financial system due to crises attacks has brought large recovery costs to the relevant countries. The major costs of financial crises are observed in lost output and the fiscal consequences in securing the financial sector (Lestano, 2006). The most recent financial crisis which has had a huge impact on the global economy originated from the US financial crises in 2007-2008. Those crises events brought big losses to the US domestic real economy and world economy. In many cases, a crisis hitting an economy may trigger a shock in the financial markets and also spread over to other economies; the depth determines its exposure and contagion effects. Considering that financial crises are contagious and incur high recovery costs, anticipating the crisis events is necessary for policymakers so that they can set some policy actions to prevent the onset of a financial crisis or soften its effects.
The importance of financial crisis prevention led several notably researchers to develop an early warning system as a monitoring tool. Edison (2003) defines an early warning system as a mechanism for predicting crises as well as determining whether a country entered a vulnerable condition. He adds that an early warning system consists of a precise definition of which crisis is predicted. Lestano, Jacobs, and Kuper (2003) distinguish three types of financial crises: currency crises, banking crises and debt crises. Reinhart and Rogoff (2009) also categorize varieties of crises into crises defined by a quantitative threshold and by events. The former can be directly defined by quantitative measures, while the latter needs a qualitative analysis to define the crises. In their definition, a currency crisis belongs to the class of crises defined by a quantitative threshold, while banking crises and debt crises are defined by events. There are two mechanisms for predicting crises which are applied in some studies in developing early warning systems, i.e. the signal extraction approach and the multivariate logit model. Issues regarding to the definition of crises, the observations as well as the purpose of the studies feature in both approaches, see Section 2.
2 Indonesia. The government had to compensate the high costs in securing the financial sector and thus such large fiscal deficits were inevitable. Moreover, among the Asian countries Indonesia experienced the most drastic loss in output which declined by 13% in 1998 (Lestano, 2006). With regard to these bad experiences Indonesia needs a mechanism which can predict the occurrence of financial crisis so that the policymakers can appropriately set policies to prevent the onset of crisis.
The aims of this study are (i) to develop an early warning system for a single economy, i.e. Indonesia, since there are only a few of studies on early warning systems for a single economy, and (ii) to identify variables that can be considered as indicators of financial crises by using two methods in predicting the financial crises, the signal extraction approach and a multivariate logit model. A single country is studied to capture the characteristics of the economy and the types of financial crises which hit it. The first method, the signal extraction approach, is applied to select several variables which are useful as leading indicators for financial crises; the multivariate logit model determines which indicators contribute to the probability of financial crises. The estimations are based on a set of quarterly time series from 1980 until 2006 for the first approach and from 1984 until 2006 for the second approach. In addition, the data is extended until 2009 for evaluating the out-of-sample performance of the multivariate logit model.
The results of this paper confirm that several indicators are useful as leading indicators for financial crises in the case of Indonesia, such as real GDP growth, change in stock prices and the inflation rate. These indicators came out significantly in both approaches. In addition, there are also some indicators which may give information on the onset of financial crises, such as the national savings rate, the growth of industrial production, M2 growth, M2 multiplier, M1 growth, excess real M1 balances, the differences between lending and deposit rate, the ratio of domestic credit to GDP, commercial bank deposits, ratio of M2 to international reserves, import growth and growth of world oil price. These results of the signal extraction and the multivariate logit approach can be considered as supporting evidence that some leading indicators are informative in predicting the financial crisis for the case of Indonesia.
3 provides descriptions of the methods adopted to predict the financial crises occurrences. In addition, the methodology in dating the financial crises is also explained in this section. Section 4 lists the data used in this paper, data transformations and sources. The results for signal extraction approach and multivariate logit model are discussed in Section 5. Finally, Section 6 concludes.
2 Literature Review
The increasing number of crises in the 1990s as a result of financial liberalization in many economies triggered studies on early warning systems. The studies differ in terms of type of crises, the sample of countries, single country or multiple countries data, explanatory variables, as well as econometric tools and methods In this section, I will elaborate some of the literature on early warning systems.
2.1 Definition of crises
Lestano et al. (2003) distinguish three different types of financial crises: (i) currency crises, (ii) banking crises, (iii) debt crises. However, this paper confined to the banking crises and currency crises due to the data availability and the severe impacts of both crises in Indonesia.
4 Theoretically, currency crises are categorized into first generation, second generation, and third generation. Abiad (1999) mentioned that those three crises generations explain through which mechanism a currency crisis occurs. He refers to Krugman (1979)’s model which elaborates that the role of unsustainable government policies is the causes of the first generation crises. The main problem here is that the pegged exchange rate regime is fundamentally incompatible with the government budget deficit imposed. To this extent, policies to maintain the exchange rate after a shock result in a decline in international reserves. These unsustainable policies trigger speculative attacks to domestic currency. In the second generation, there was no evidence of lack of international reserves, deficit monetization, and high growth in domestic credit, but the speculative attacks hit the currency under a pegged regime. One of the examples of a second generation currency crisis is the European Monetary System (ERM) crisis in 1992-1993. Finally, the third generation crisis was caused by contagion. The characteristic of this type of crisis is that it spreads over from one country to other countries within the region, or the countries which have trade or financial linkages.
Banking crises are define on the basis of qualitative analyses and events judgment, rather than based on quantitative definitions (Reinhart and Rogoff, 2009). Lestano et al. (2003) for example state that banking crises occur when there are high crisis recovery costs or there is unsoundness in the banking system, such as bank runs, and the collapse of financial firms. Moreover, Dermirgüç-Kunt and Detragiache (1998) explain that a banking crisis occurs when at least one of the following indications happens: (i) the ratio of non-performing assets to total assets of the banking system is higher than 10 percent; (ii) the cost of securing in banking system exceeds 2 percent of GDP; (iii) a large scale nationalization of banks; and (iv) bank runs which trigger the government intervention. In line with those characteristics of banking crises, Kaminsky and Reinhart (1999) identify a banking crisis when bank runs lead government to intervene in securing the financial system, i.e. bank bail outs, bank closures, or takeovers.
5 the result of banking system distress. By using data, they analyze that a crisis period is accompanied by a decline in the total deposits larger than 5 percent and a ratio of liquidity support to total deposits of at least 5 percent. The former measure indicates whether a distress in the banking sector coincides with bank runs, while the latter measure denotes an extensive intervention from the authorities in the response of distress in banking sector.
2.2 The causes of financial crises
This section addresses the causes of financial crises by focusing on two types of crisis, currency crisis and banking crisis. There are four possible causes of the currency crises (Sugandi, 2004): (i) excessive monetary policy and domestic credit, (ii) the deterioration of current account, (iii) financial system fragility, and (iv) the degree of openness in a country. Prior to the Asian crisis, most of the countries in that region peg their exchange rate to the foreign currency, for instance the US Dollar. In line with Krugman (1979)’s model, Kaminsky and Reinhart (1999) explain that expansive monetary policy, such as excessive domestic credit expansion, contributes to the currency attack under fixed exchange rate regime. The monetary authority stabilizes its domestic exchange rate by using its foreign reserves. In such stabilization policy, the authority experiences losses in the international reserves. In the end, the authorities cannot avoid the devaluation of the domestic currency due to further persistent loss in the international reserves.
Trade imbalance is also a source of currency crises in Asian countries. When exports exceed imports, the domestic currency appreciates against the foreign currency. Real appreciation in domestic currency leads to deterioration in trade balance, which may trigger a speculative attack due to the lack of international reserves as a buffer to the changes in exchange rate. Corresponding to previous analysis, Kaminsky, Lizondo, and Reinhart (1998) explain that domestic credit expansion induces the demand for traded goods and, thus, an increase in exports. Therefore, expansionary credit policy and fiscal policies also become factors determining imbalances in current account.
6 may provoke speculative attacks on the domestic currency (Corsetti, Pesenti, and Roubini, 1999). Moreover, the higher the degree of financial liberalization, the more likely a currency crisis will spread over to other countries. This contagion effect that may spread across the integrated financial market as a consequence of financial liberalization is considered as a factor that explains the financial fragility and vulnerability.
Apart from financial liberalization, trade liberalization accounts for the factors that determine the degree of openness of a country. Eichengreen et al. (1996) mentioned that currency crises may occur through the trade channel. If some countries are strongly interrelated through international trade linkages, one country’s real depreciation might influence other countries’ exports competitiveness and, thus, their trade balances. As mentioned above, the trade imbalances would become the source of currency crises.
7 2.3 Previous empirical studies on early warning systems
Previous empirical studies on the early warning systems to predict financial crises vary in terms of the types of crises, the approaches, as well as the sample of data and period. Most of studies concentrate on a certain kind of crisis, for instance currency crises (Eichengreen et al. 1996; Berg, Borensztein, and Pattillo, 2004; Jacobs, Kuper, and Lestano, 2008; Kamin, Schindler, and Samuel, 2001; Kaminsky, et al., 1998; Abiad, 1999), and banking crises (Hardy and Pazarbaşioğlu, 1999; Demirgüç-Kunt and Detragiache, 1998; Wong, Wong, and Leung, 2007). However, there are also some studies which incorporate several crises, i.e. the twin crises, banking crises and currency crises (Kaminsky and Reinhart, 1999; Kaminsky, 1999; Dooley, 2000), and banking crises, currency crises and debt crises (Lestano et al., 2003).
Regarding to the methodology used in the literature of early warning systems, there are several methods that are used in predicting financial crises: (i) the signal extraction approach, known as the leading indicators model, (ii) the multivariate logit method, and (iii) other methods, like Markov switching models (Abiad, 1999).
The leading indicator model was pioneered by Kaminsky et al. (1998). They constructed the individual leading indicators for currency crisis by using the signal extraction approach, which monitors the movement of some economic indicators in predicting the occurrence of crisis. If an indicator deviates from a certain threshold value, it necessarily gives a signal that a crisis occurs within the following 24 months. Kaminsky and Reinhart (1999) and Kaminsky (1999) extend this method to construct a composite leading indicator. This composite indicator is constructed from the individuals leading indicator which have good performance in predicting the onset of crisis. An abnormal movement in the value of this composite index from its threshold signals the financial crisis is about to occur.
8 be monitored as the crises signal. Frankel and Rose (1996), and Jacobs et al. (2008) estimate the probability of currency crises by using the multivariate logit model. However, Jacobs et al. (2008) add factor analysis as a tool to reduce number of explanatory variables included in multivariate logit model to solve the multicollinearity problem.
The signal extraction approach and multivariate logit model have advantages and disadvantages in predicting the financial crises. Abiad (1999) mentions that the results of the signal extraction approach are easy to interpret for policy makers, as leading indicators, individually or in the form of an index, deviate from its threshold or not. Therefore, this method is capable of giving some information to the policy makers about which variables should be monitored to prevent the crisis. However, Edison (2003) mentions that this method has no standard statistical test to evaluate the robust relationship between the indicators and the probability of crises. He instead applies sensitivity analysis, for regional differences and samples, to evaluate the robustness of the model.
The multivariate logit model has the advantage of easy interpretation of significant leading indicators. In addition, this method does not take into account variables that have already captured other variables’ information, meaning that high correlated indicators are usually not included jointly as the explanatory variables. Abiad (1999) adds that running the data by using the statistic software packages, for instance, Eviews and STATA, makes the use of this method easier. Despite those advantages, there are also some disadvantages regarding to the use of this method. Abiad (1999) elaborates those disadvantages by referring to Kaminsky et al. (1998). This method cannot distinguish which significant indicators perform better in predicting the financial crisis. Moreover, one should be careful in interpreting the performance of indicators related to the probability of a crisis since it is cannot be evaluated by looking at the degree of marginal effect of the changes in indicators to the probability of crisis.
9 the crisis, the duration of the crisis and the variables that are related to the occurrence and the end of the crisis are analyzed altogether in this method. However, this model is very limited to be used in constructing the early warning system due to its complicated technical and computational difficulties
The results of early warning systems in some previous studies are mixed in terms of the significant indicators, and the performance of models, depend on the method adopted (Edison, 2003). Similarly, Berg et al. (2004) argue that the early warning systems are not accurate enough to be applied as the only method to prevent the financial crisis in practice. However, they add that the early warning system contributes as a tool to anticipate the financial crises if the use of this mechanism is accompanied by qualitative and quantitative analysis about the economic vulnerability.
3 Methodology
To determine the leading indicators for predicting the occurrence of financial crisis in Indonesia, I apply the signal extraction approach or the leading indicator model and the logit multivariate method.
3.1 Dating financial crises
Before applying the two methods employed in this paper, dating financial crises is necessary in the sample period. To date currency crises, I calculate an exchange rate market pressure index (EMPI) to identify the occurrence of currency crises in the sample period. This method was pioneered by Eichengreen et al. (1996). EMPI in the Eichengreen, Rose and Wyplosz method, hereafter ERW method, is calculated from weighted changes in the nominal exchange rate, the ratio of international reserves to money supply (M1), and nominal interest rate as:
, (1)
where is the exchange rate market pressure index for Indonesia in period t, is the nominal exchange rate for the domestic currency against a foreign currency, i.e. Rupiah/US Dollar in period t,
10 international reserves to money supply (M1) in period t, is the reference country’s ratio of international reserves to money supply (M1) in period t, and US is chosen to be the reference country. Therefore, the term of
is the difference between the domestic country and the
reference country’s relative change in the ratio of international reserves to the money supply (M1). In addition, and are the domestic country and reference country’s nominal interest rate in period
t, while is the nominal interest rate differential. The weights attached to the variables are
the inverse of these following variables: (i) the relative change in exchange rate standard deviation, (ii) the difference between Indonesia and US relative change international reserves to M1 ratio standard deviation, (iii) the interest rate differential standard deviation, .
Kaminsky et al. (1998) apply another definition of EMPI which is less sophisticated. This method, KLR method, took into account a country’s changes in nominal exchange rate, international reserves, and nominal interest rate, and dropped the variables related to the reference country which are used in the ERW method.
. (2)
This paper prefers to use the EMPI formula calculated by Lestano et al. (2003) which is a KLR modification method (LJK, 2003) as follows:
, (3)
where is the exchange rate market pressure index for Indonesia in period t, is the relative
change in country’s nominal exchange rate in period t, is the relative change in international reserves in period t, is the change in the nominal interest rate in period t. The standard deviation of nominal exchange rate change, international reserves change, and nominal interest rate change are given by , , .
To determine the occurrence of currency crises, a threshold must be calculated. When exceeds the threshold value, the currency crisis exists in the period t.
11 Kaminsky et al. (1998) formed this threshold from the mean of plus three standard deviations. By looking at the currency crisis identification below, they set the parameter of equal to three. Lestano and Jacobs (2007) employ sensitivity analysis in their dating currency crises study by setting several parameters for their EMPI threshold. These parameters range from 1.5 to 3. The higher the threshold values, the lower the number of crises produced by the dating crises method.
3.2 Signal extraction approach
The signal extraction approach can be used to predict financial crisis by monitoring individual leading indicators or composite leading indicators. Kaminsky and Reinhart (1999) employ this method to predict banking crises and currency crises by setting a threshold for each individual leading indicator and monitoring the development of its level. When an indicator level exceeds a threshold, it signals a crisis in the next several months, which depends on the window adopted in the method. Kaminsky and Reinhart (1999) set a priori a 24-month window for currency crisis and a 12-month window before and after the beginning for banking crisis. For instance, if any leading indicator sends a signal of crisis, a currency crisis will occur within the following 24 months and/or a banking crisis is about to appear within the next 12 months. In this case, the indicator accurately sends signals of crisis.
Kaminsky et al. (1998) apply the procedure to predict the currency crises by: (i) setting the optimal threshold for each leading indicator, (ii) examining the forecasting ability of each leading indicator, (iii) selecting the best indicator in terms of its performance in predicting crises.
12 i.e. the number of unpredicted crises. Secondly, an indicator possibly sends a false alarm when an actual crisis does not occur as it is predicted, known as Type II error. This type of error is denoted by B in the table below.
Table 1. The possible outcomes of leading indicator signals in predicting crisis Observation
Crisis No Crisis
Signal Crisis A B
No Crisis C D
In a previous EWS study of banking crises Davis and Karim (2008) referred to Kaminsky and Reinhart (1999) when they set a threshold such that the noise to signal ratio (NTSR) is minimized. This rule is adopted to overcome the tradeoff between Type I and type II error when selecting the optimal indicator threshold. For instance, a higher threshold which may detect more crises will accommodate a higher probability of Type II error. Conversely, a lower threshold may result in a higher Type I error. According to Kaminsky and Reinhart (1999), the noise to signal ratio (NTSR) is given as follows:
, (5) The lower the noise to signal ratio, the more accurate a leading indicator predicts a crisis. As mentioned previously the large numbers in A and D imply that a leading indicator correctly predicts whether or not a crisis occurs. Therefore, an effective leading indicator in predicting a crisis has a lower noise to signal ratio.
13
, (6)
while the probability of a crisis given that signal was issued is calculated as follows:
, (7)
This measure shows the fraction of crises correctly called to a total number of crises signals issued by the indicator. The higher this measure, the higher the probability a crisis signal is precisely followed by an actual crisis.
This paper will use the same procedure as Kaminsky et al. (1998) to predict currency crises in Indonesia. Constructing the signal extraction approach consists of the following steps: (i) dating the crisis occurrence, (ii) crisis indicators identification, and (iii) selection of the best indicator by assessing the forecasting performance. First, I use the Exchange rate Market Pressure Index (EMPI) to date the currency crisis by borrowing the EMPI modification (Lestano et al., 2003), as described earlier in the previous section. I date the banking crises period based on the definitions of banking crises by Kaminsky and Reinhart (1999).
At the second step, a threshold must be set to identify whether an indicator sends a crisis signal. As mentioned above, Kaminsky and Reinhart (1999) suggested that an optimal threshold should minimize the noise to signal ratio (NTSR). A threshold is set at a percentile of the distribution of the normalized indicator values. Specifically, the threshold is set at the lowest or the highest percentile which minimizes the noise to signal ratio (Kaminsky, et al., 1998). When a normalized value of the indicator in period t exceeds the threshold, it means that the indicator sends a crisis signals. This signal should be marked by one if an actual crisis occurs within the window. Conversely, if a crisis does not occur within the window, an indicator signal will be marked by zero.
14 the percentage of crises correctly called ≥ 70%, (ii) the noise to signal ratio ≤ 100%, and (iii) the probability of a crisis given that a signal was issued ≥ 50%.
In previous studies, the signal extraction approach is extended to construct a composite leading indicator. Davis and Karim (2008) applied a composite leading indicator based upon Kaminsky (1999) for predicting banking crises. They used a combination of some indicators as a composite leading indicator from the best individual leading indicators in their study. Likewise, Sugandi (2004) applied the same procedure to construct a composite leading indicator for currency crises. However, these studies have slightly different criteria to choose the components of the composite leading indicator. Kaminsky (1999) and Davis and Karim (2008) select the best indicators based on the value of NTSR, while Sugandi (2004) chooses the best indicators based on the three performance measures as discussed above. Specifically, Kaminsky (1999) selects the indicators which have a value of NTSR below unity, whereas Davis and Karim (2008) limit the threshold of their best indicator to have the value of NTSR which is not higher than 0.5.
In the construction of a composite leading indicator in this paper, firstly, I screen the individual indicators based on the three performance measures, i.e. the noise to signal ratio (NTSR), the percentage of crises correctly called, and the probability of a crisis given that signal was issued. I adopt the formula of composite leading indicator by Kaminsky (1999) as follows:
, (8)
where the composite leading indicator, C, is the sum of weighted signal for variable j in period t. is defined as the signal of the individual leading indicator j in period t, while is the weight for each individual indicator calculated as the inverse of the noise to signal ratio (NTSR).
15 3.3 Multivariate logit approach
The second method applied is the multivariate logit approach. This approach estimates a binary dependent variable by running a nonlinear probability model based on the S-shaped logistic function. The logit model is given as follows:
, (9) where is the probability that an event, Z, occurs is equal to one, is the cumulative logistic probability function, X is a set of explanatory variables, while α and β are the parameters. These parameters are estimated by maximizing the log likelihood function of the crisis occurrence probability. A transformation of this model into a logarithm function is specified as follows:
(10)
16 3.3.1 Performance of the model
There are two evaluations which will be taken in this paper. First, I will evaluate the model’s accuracy of calling crisis by looking at the percentage of correctly crisis called, as it is also applied in the first approach before. Demirgüç-Kunt and Detragiache (1998), and Davis and Karim (2008) set the threshold for the estimated probability of crisis occurrence so that they can determine whether a crisis is correctly predicted. They set the threshold, the cut-off level, to be equal to 0.5. If the estimated probability of crisis exceeds the threshold of 0.5 and the observation shows a crisis actually occurs, Z=1, it is clearly stated that the model has a good crisis predictive ability. This outcome is shown by P (1, 1) in Table 2. The matrix below displays the possible prediction of logit model, similar to the possible outcome of crisis prediction in the first approach.
Table 2. The possible estimation and observation of multivariate logit approach Observation
Crisis (Z=1) No Crisis (Z=0)
Estimated Probability High P(1,1) P(1,0)
Low P(0,1) = 1 – P(1,1) P(0,0) =1 – P (1,0)
Another way to evaluate the performance of logit multivariate approach is to calculate the quadratic probability score (QPS) and the log probability score (LPS), proposed by Berg and Pattillo (1999) and Jacobs et al. (2008).
, (11)
, (12) where and are the time series of T predicted crises occurrences probability and that of
the T actual observed crises. In this model evaluation, I test the performance of the model both in-sample and out-of-in-sample.
17 predict the correct signal. In this case, the logit model indicates low estimated crises probability when an actual crisis occurs in the observation or, the other way around, the estimated probability of crisis is high when the observation shows there is no crisis event, see Table 2. On the other hand, the interpretation of LPS is parallel to that of QPS. Lower scores give the best prediction ability of the model. As the LPS goes to infinity, the predictive ability of the model becomes poor.
4 Data
In conducting the first approach I use the indicators based upon Lestano et al. (2003), while the indicators used in the second approach are based on the best indicators in the first approach. There are some modifications in the set of indicators used in this paper, which is slightly different from the indicators in the study of Lestano et al. (2003): (i) I use 22 out of 24 indicators due to the available data, the terms of trade and the ratio of bank reserves to assets are not included; (ii) I include real GDP, instead of GDP per capita, since this paper does not use multi countries data as the observations, thus, real GDP is appropriately included as a common measure for a single economy’s domestic economic performance; (iii) I use the ratio of external debt to GDP, instead of the ratio of public debt to GDP, due to the availability data. This proxy is also one of debt indicator which shows the healthiness of government budget balance. The selected indicators are grouped into external factors, financial factors, domestic factors, and global factors. For more details, the list of indicators and sources used in this paper and their configuration are displayed in Table A of the Appendix.
18 The data are extracted from the International Financial Statistics (IFS) of the IMF, the Bank of Indonesia, and the Indonesian Ministry of Finance. The sample period ranges from 1980:1 until 2006:4 for the signal extraction approach and from 1984:2 until 2006:4 for the multivariate logit model. However, I extend the data period up to the last quarter of 2009 to evaluate the out-of-sample performance of the multivariate logit model.
5 Results
5.1 Dating financial crises
19 Figure 1. Distribution of financial crises in Indonesia (1980-2009)
Note. The Y-axis shows the crisis dummies: 1. Currency crises (KLR modification method) 2. Banking crises (dated by events) 3. Twin crises
5.1.1 Currency crises
Indonesia has implemented three different regimes in the exchange rate policy during last thirty years. Indonesia pegged its currency to USD until the year of 1978. This regime changed into a managed floating exchange rate regime in November 1978. Free floating exchange rate regime was imposed since 14 August 1997. The development of Indonesia’s domestic currency in the period 1980 to 2009 is displayed in Table 3. This table highlights the important points in times in which Rupiah was under pressure against the US Dollar.
0 1 2 3 4 19 80 19 81 19 82 19 83 19 84 19 85 19 86 19 87 19 88 19 89 19 90 19 91 19 92 19 93 19 94 19 95 19 96 19 97 19 98 19 99 20 00 20 01 20 02 20 03 20 04 20 05 20 06 20 07 20 08 20 09 Banking crises Currency crises Twin crises
Distribution of financial crises in Indonesia (1980-2009)
20 Table 3. Chronology of exchange rate pressure in Indonesia
Date Event Chronology
1983:1 Devaluation Rupiah from Rp. 702.50 to Rp. 970 per USD (March).
In early 1983 a pressure in global recession and a sharp decline in the world oil price affected domestic economic conditions.
Deficit in current account and balance of payment were larger. Economic growth was slowing down, so that devaluation was taken to overcome this problem.
1986:3 Devaluation Rupiah from Rp. 1,134 to Rp. 1,644 per USD (September).
Further decline in the world oil price reached USD 10 per barrel. It affected a decrease in the value oil commodity export which led to deficits in the balance of payment
1997:3 Exchange rate regime changed to free floating exchange rate (August)
The beginning of the financial crisis in Indonesia. Speculative attacks on Rupiah and contagious effect from depreciation in Thai Baht. Afterwards, the nominal value of Rupiah was heavily fluctuated. 1997:4 Rupiah became more
volatile. (November)
A closing of 16 banks impacted on further deterioration in Rupiah. During November to December 1997, it fluctuated from Rp. 3,250 to Rp. 4,000 per USD,
1998:1 Sharp decline in Rupiah (January)
Within three weeks in early of January, Rupiah fell from Rp. 4,850 to Rp. 13,650 per USD.
1998:2 Further deterioration in nominal exchange rate (June)
Rupiah reached its lowest level of Rp. 16000 per USD.
2001:2 Depreciation of Rupiah (April)
Domestic political condition put a pressure in the level of nominal exchange rate. Rupiah was
depreciated to Rp. 11,600 per USD, compared to its average nominal value in year 2000, i.e. Rp. 8,438 per USD.
2008:3 Depreciation in Rupiah; reached level of Rp. 12,150 per USD (November)
The global financial crisis triggered capital outflow from Indonesia which put pressure on the domestic exchange rate
2009:1 The lowest level of Rupiah in 2009 reached Rp. 12,020 per USD (March)
Further deterioration in the global financial system had an impact on the pressure of nominal exchange rate
21 Figures 2 and 3 show the dating for the occurrence of currency crises in Indonesia by using EMPI in the KLR method and the KLR modification method. In the identification of currency crisis, I simulated two thresholds, i.e. by using parameter β=2 and β =3. One with parameter β =3 was originally applied in the KLR method. Meanwhile, the parameter β=2 is used in the modification of KLR method, following Lestano et al.(2003).
Figure 2. Dating currency crises 1980-2009 by using KLR method
Note: (i) B3 stands for parameter β =3, while B2 stands for parameter β=2,
(ii) KLR stands for EMPI by using KLR method
Figure 3. Dating currency crises 1990-2009 by using KLR modification method (LJK, 2003)
Note: (i) Threshold2 stands for parameter β =2, while Threshold3 stands for parameter β=3, (ii) KLR_M stands for EMPI by using KLR
modification method
From table 4, the KLR modification method (LJK, 2003) detects more events of currency crises in Indonesia than KLR method does. One can see that the output of KLR modification method (LJK, 2003) which uses parameter β=2 fits better with the chronology of domestic exchange rate pressure in Table 3. Thus, I employ these crises period in the multivariate logit model. As noted earlier, the five periods of currency crises correspond to two currency crises events, devaluation of Rupiah in 1983, and the Asian currency crisis in 1998.
22 Table 4. Distribution of currency crises in Indonesia (1980 – 2009)
Method KLR method KLR Modification
(LJK, 2003) Threshold β=2 β =3 β=2 β =3 Date 1983:Q1 1983:Q1 1983:Q1 1983:Q1 1998:Q1 1998:Q1 1998:Q2 1998:Q2 1998:Q2 1998:Q3 1998:Q3 1998:Q3 1998:Q4 Total 4 1 5 3 Share of total observation 3,33% 0,83% 4,16% 2,5% 5.1.2 Banking Crises
The development of banking sector in Indonesia can be categorized into four periods, i.e. pre-deregulation period (prior to 1983), pre-deregulation period (1983-1997), banking crisis period (1997-1999), and post-crisis period (from 1999 onwards).
Prior to 1983, the regulation in banking system was rather tight. The government limited flows of credit loans from banks in order to control the high inflation rate. Higher interest rate was set to tackle this problem. Until the oil boom period in 1970s the government still implemented tight regulation in the banking system. The inward policy imposed by government was aimed at stimulating the economic growth from oil and mining revenues and to protect the economy by packages of regulations.
In the early 1980s, the world oil crisis had a larger impact on Indonesia’s macroeconomic conditions. A sharp decline in the world oil price caused a decrease in economic growth and balance of payments deficits. To overcome this problem, the government launched deregulation policies in the banking sector to strengthen the role of banking system in the economy. These banking policies were imposed in 1983 and 1988 to induce economic growth.
23 and triggered a bank run in 1992, while the other two cases did not lead to any serious distress in the banking system in that period.
The vulnerability of the banking system can be especially apparent when the banking system collapsed in 1997. In this banking crisis period many insolvent banks were liquidated and taken over by government. The collapse in the banking sector incurred a high recovery cost which was burdened to government budget. Table 5 shows the banking distress in Indonesia from 1980 to 2009.
24 Table 5. Chronology of banking distress in Indonesia
Date Event Chronology
1992:Q4 Closing of Bank Summa (November)
The default of Bank Summa triggered a bank run.
1997:Q4 Government closed 16 banks (November)
This closing was accompanied by the implementation of IMF assistance program.
Liquidity support from Bank of Indonesia costs 5% of GDP (December)
The closing of 16 banks triggered a massive bank run. Lack of liquidity in the banking system forced the Bank of Indonesia to give liquidity support.
1998:Q1 Blanket guarantee to depositors in private and state owned bank (January)
This emergency assistance aimed at protecting depositors and creditors of domestic banks
Establishment of IBRA (Indonesia Bank Restructuring Agency) in January
It aimed at controlling and taking over insolvent banks. In the early of the operational period, February 1998, it controlled 44 banks which had 40% of deposits in the banking system
1998:Q2 Closing of Seven banks (April) Each liquidated bank has amount of liabilities more than 500% of its capital
IBRA took over one of the biggest private commercial bank, i.e. Bank Central Asia (BCA).
Merger between four banks. (May)
BCA comprises 12% of banking sector liabilities. There were massive withdrawals from this bank that forced Bank of Indonesia provided Rp. 30 trillion liquidity to this bank 1998:Q3 Closing of three private banks
IBRA took over three private banks
Merger between four banks (August)
Recovery policies were continued by closing more insolvent banks.
1998:Q4 Merger between four state owned banks became a single entity, Bank Mandiri (October)
Bank Mandiri becomes the largest state-owned bank which has 30% share of banking sector assets.
1999:Q1 Closing of 38 private banks (March) Liquidation of insolvent bank was subject to further recovery strategy of banking system 2000:Q3 Merger between Bank Danamon with
eight private banks. (June)
This merger downsized the numbers of bank to strengthen the foundation of banking system
2000:Q4 Closing of two private banks (October)
- 2002:Q3 Merger between five banks
(September)
- 2008:Q3 Bank Century was bailed out
(November)
To prevent systemic banking distress due to global financial shock, Bank Indonesia bailed out Bank Century and guarantee the depositor’ account which has less than Rp. 2 million. The total of this bail out reached Rp. 6.7 trillion.
25 5.2 Signal extraction approach
5.2.1 Individual leading indicators
In constructing the individual leading indicator I follow the method constructed by Kaminsky et al. (1998), and Kaminsky and Reinhart (1999). Determining the threshold is one of the important stages in this signal approach. In the KLR (1998), the threshold of each individual leading indicator is set by either the lowest or the highest percentile of standardized value of the indicator which minimizes the noise to signal ratio. Following Kaminsky et al. (1998), I set the low percentile threshold for the following variables: real exchange rate, export growth, foreign reserves growth, commercial bank deposit, change in stock prices, and real GDP growth. Kaminsky et al.(1998), Berg and Patillo (1999), Edison (2003) include industrial production growth as the proxy for the real output growth. I also set the low percentile for the ratio of fiscal balance to GDP as Eichengreen and Arteta (2000) denote that budget deficits are closely related to financial crisis. Kamin, Schindler, and Samuel (2001) suggest that a decrease in ratio of current account to GDP increases the probability of crisis occurrence. This indicator’s threshold is also set as the low percentile of its distribution.
26 the probability of crisis occurrence, I also set the high percentile threshold for growth of world oil price.
Table 6 shows the performance of individual leading indicator for financial crises in Indonesia. Column (3) displays the threshold of each individual leading indicator which minimizes the noise to signal ratio (NTSR), shown in column (4). The lower the value of the noise to signal ratio, the more informative is a crisis occurrence signal of an individual indicator gives. All individual leading indicators show a good performance in terms of NTSR, since they have a NTSR value below unity. Each indicator in this set of variables gives an informative signal to the occurrence of crisis.
In addition, column (5) shows the percentage of crises correctly called which indicates the individual leading indicator’s ability to accurately predict the crises occurrence. Only a few indicators have a good performance in predicting the financial crises correctly, real GDP growth, the change in stock prices, inflation, the national saving rate, growth of industrial production, M2 money multiplier, M2 growth, the spread between lending and deposit rate, and import growth. These indicators have the ability to predict financial crises accurately with the probability above 50%.
27 Table 6. The performance of individual leading indicators
External Sector (1) Code (2) Threshold (3) NTSR (4) % crisis correct (5) P(crisis|signal) (6) Real exchange rate RER < 0.050 0.08 27.27 60.00 Ratio M2/foreign reserves MFR > 0.950 0.09 36.36 57.14
Export growth EXG < 0.060 0.10 36.36 50.00
growth of international reserves IFR < 0.080 0.11 45.45 50.00 Ratio of current account to GDP CAY < 0.125 0.18 45.45 38.46
Import growth IMG > 0.080 0.24 90.91 32.26
Financial Sector (1) Code (2) Threshold (3) NTSR (4) %crisis correct (5) P(crisis|signal) (6) Ratio of domestic credit to GDP DCY > 0.975 0.02 45.45 83.33 Commercial Bank Deposits CBD < 0.050 0.03 36.36 80.00
M2 Growth M2G > 0.950 0.03 54.55 75.00
Excess real M1 balance ERM > 0.950 0.05 45.45 71.43
M1 growth M1G > 0.950 0.07 45.45 62.50
Domestic real interest rate RIR > 0.900 0.11 36.36 50.00 Lending and deposit rate spread LDS > 0.975 0.11 54.55 50.00 M2 Money multiplier MMM > 0.900 0.12 60.00 50.00 Public Sector (1) Code (2) Threshold (3) NTSR (4) %crisis correct (5) P(crisis|signal) (6) GDP growth GDP < 0.100 0.01 100.00 91.67
Change stock price CSP < 0.150 0.01 90.00 90.00 National saving rate NSR > 0.900 0.01 72.73 88.89 Growth of industrial production GIP < 0.050 0.01 72.73 88.89
Inflation rate INF > 0.950 0.02 54.55 85.71
Ratio Fiscal Balance to GDP FBY < 0.010 0.11 9.09 50.00 Ratio of External Debt EBY > 0.900 0.20 45.45 35.71
Global Economy (1) Code (2) Threshold (3) NTSR (4) %crisis correct (5) P(crisis|signal) (6) Growth of world oil price WOP > 0.975 0.06 18.18 66.67
US interest rate USI < 0.100 0.15 27.27 42.86
28 To sum up, the best individual leading indicators which satisfy all three measures, i.e. the lowest NTSR, the percentage of crises correctly called, and the probability of a crisis given that signal was issued, are real GDP growth, the change in stock prices, the national savings rate, and the growth in industrial production. Moreover, several indicators such as import growth, the spread between lending and deposit rate, M2 growth, M2 money multiplier, the ratio of domestic credit to GDP, commercial bank deposits, excess real M1 balance, M1 growth, real exchange rate, ratio of M2 to foreign reserves, and world oil price growth are also informative as financial crises, since they provide satisfactory signals prior to crisis.
5.2.2 Composite leading indicators
29 Table 7. The components of composite leading indicators
Index 1 Index 2 Index 3
Real GDP growth Real GDP growth Real GDP growth
Change in Stock Price Change in Stock Price Change in Stock Price National Saving Rate National Saving Rate National Saving rate Growth in Industrial Production Growth in Industrial Production Growth in Industrial Production
Inflation Rate Inflation Rate
M2 Growth M2 Growth
M2 Multiplier Excess real M1 balance Lending and deposits rate spread M1 Growth
Import Growth Ratio of Domestic Credit to GDP
Commercial Bank Deposits Real Exchange Rate Ratio of M2 to Foreign
Reserves World Oil Price
Notes: (i) Index 1 consists of the best indicators according to three performance measures, (ii) Index 2 consists of the best indicators in terms of the percentage of crises correctly called, and (iii) Index 3 consists of the indicators which have the lowest NTSR.
The performance of these three composite indices is shown in Table 8. One can see that the performances of three indices in terms of NTSR are quite good, but Index 1 and Index 2 have slightly higher NTSR value than Index 3. Similarly, Index 1 and Index 2 outperform Index 3 in terms of the percentage of crises correctly called. Finally, in terms of probability of a crisis given that signal was issued Index 2 performs better than Index 1 and Index 3.
Table 8. The performance of composite leading indicators Performance of
indicator Index 1 Index 2 Index 3
NTSR 0.139 0.139 0.141
% crisis correct 80 80 70
P(crisis|signal) 44.44 47.06 46.67
Threshold > 0.875 > 0.875 > 0.875
30 stock prices (CSP), national savings rate (NSR), and growth of industrial production (GIP), are the indicators that give a signal for the financial crises in the case of Indonesia.
Growth of real GDP is one of the domestic sector variables which measure macroeconomic stability. The development of GDP genuinely reflects the performance of an economy, and therefore the probability of a crisis. When real GDP growth is declining sharply it may indicate a financial crisis is about to occur. According to Kaminsky (2000) a slowdown in the growth of output often marks the occurrences of banking and currency crises. In emerging countries real GDP growth is a good indicator for signaling crises as this indicator generally demonstrates the healthiness of macroeconomic condition (Frankel and Rose, 1998).
Change in stock prices (CSP) is also one of the best indicators according to the signal extraction approach. In an emerging market, particularly Indonesia, where the financial market is less developed, the volatility of stock prices determines the behavior of traders in the stock market. The fact that the share of foreign traders dominates in the Indonesia financial market leads to a financial system fragility. The reason is that this factor may trigger massive capital outflow when the stock prices decline sharply and, therefore, a decrease in the value in exchange rate against foreign currency should compensate this. A massive capital outflow due to deterioration in stock prices may occur as a reaction to traders’ losses. To support this finding, Kaminsky (1999) suggests that a decline in asset price may be a symptom of both currency crisis and banking crisis. Moreover, signal extraction study by Kaminsky and Reinhart (1999) highlights the interrelation between the stock prices decline and the output slowdown as they are closely related to the domestic financial problem.
31 were not decreasing prior to the Asian crisis (Corsetti et al., 1999). In this case, the Asian countries experienced high saving rates prior to the crisis. The data shows that the level of the saving rate is around 30% in average in 1990s prior to crisis. Thus, the results of the signal extraction approach that suggest an increase in national saving rate as the indicator of financial crisis is parallel to the fact that saving rates was high prior to the Asian crisis.
Apart from these four best indicators, there are several indicators that may send the informative signal prior to the occurrence of crisis, such as the inflation rate, M2 growth, M2 multiplier, the spread between lending and deposits rates, import growth, excess real M1 growth, M1 growth, the ratio of M2 to foreign reserves, ratio of domestic credit to GDP, commercial bank deposits, the real exchange rate, and the growth of world oil price. The first two indicators are the leading indicators which are informative in signaling crisis, since they appear both in Index 2 and Index 3. Prior to the Asian crisis in 1997, an excessive monetary policy intervention at the domestic exchange rate could be observed in an increase monetary base, and, therefore, M2 growth. Consequently, an increase in the money supply puts pressure on the inflation rate, as they are positively related.
In addition, the development in money supply as the monetary policy expands can also be analysed by looking at the increase in M1 growth, M2 multiplier, and excess real M1 balances, since these indicators are related to the variables of money supply. The ratio of M2 to foreign reserves is also an informative crises signal. This indicator is included as the component of Index 3. The intervention against devaluation in exchange rate is costly since its foreign reserves decrease as the expenses of monetary expansion. This impact can be observed as the ratio of M2 to foreign reserves increases. Moreover, real exchange rate overvaluation and the growth of oil price are also informative indicators according to Index 3. Real exchange rate overvaluation is a symptom of currency crises (Kaminsky, 1999). As a global sector indicator, a rise in the world of oil price growth contributes to signal the occurrence of crisis, since this indicator is related to an economic recession.
32 excellently as a crisis signal. Indonesia’s high dependency to import goods, for especially machinery and raw materials, prior to Asian crisis promotes higher import growth which may lead to decrease the international reserves. As a consequence, it may put pressure on the exchange rate.
The spread between lending and deposits rates, the ratio of domestic credit to GDP and commercial bank deposits are indicators for the banking sector, which are also useful as crises signals. According to Kaminsky (1999), a large lending and deposits rate spread indicates a credit crunch and, thus, a slowdown output growth. An increase in the spread between lending and deposits rates reflects a decrease in loan quality, captured by high non-performing loan. As a consequence, the banking sector becomes more fragile and vulnerable. In the case of Asian countries, the collapse of the banking sector is related to the massive expansion of domestic credit prior to the crisis (Corsetti et al., 1999). A massive expansion in domestic credit was addressed to induce economic growth. However, if it is not supported by good supervision and regulation, a credit expansion leads to higher non-performing loans and a collapse in the banking sector. The unsoundness of the banking sector may trigger depositors’ panics and lead to massive withdrawals. As a result, there is a large decrease in commercial bank deposits. Those phenomena, which actually happened as the Asian banking crisis chronology, provide insight that such banking sector indicators are informative as signals prior to crisis, according to the signal extraction approach.
33 5.3 Multivariate logit model
This section elaborates the results of the logit multivariate approach which is applied to determine the significant leading indicators in predicting the financial crises. I include the explanatory variables into the regression based on the results of the signal extraction approach. The selected indicators correspond to the composite indices, since these indices consist of the best indicators according to signal extraction approach. As the result, I analyze three multivariate logit models in this section. The multivariate logit model 1, model 2 and model 3 estimate the explanatory variables according to the component indicators in Index 1, Index 2 and Index 3, respectively.
Table 9. The results of multivariate logit model 1
(1) (2) (3) (4) Constant -4.15 -0.40 -3.29 (-1.65) * (-0.38) (-1.63) GDPG -27.50 -43.44 -48.81 (-1.16) (-2.29) ** (-1.94) * CSP -3.43 -4.49 - (-1.50) (-1.95) * - NSR 15.454 - 17.79 (1.63) * - (2.03) ** GIP -3.61 - - (-0.95) - - McFadden R-squared 0.59 0.53 0.51 LR Statistic 33.21 *** 33.46 *** 32.41 *** Number of observations 91 91 91 Number of crises 11 11 11
Notes: (i) *significant at 10%, ** significant at 5%, *** significant at 1%. (ii) Numbers in brackets show z-statistics.
34 crises. In sum, the multivariate logit model 1 suggests that the growth of real GDP (GDPG), change of stock price (CSP), and national saving rate (NSR) significantly signal the financial crises.
Table 10. The results of multivariate logit model 2
(1) (2) (3) (4) (5) (6) Constant 3.11 2.55 -0.48 -4.57 -10.37 (0.42) (0.39) (-0.47) (-3.24) *** (-2.73) * GDPG -30.90 -50.19 -46.82 - - (-0.90) (-1.84) * (-2.29 ** - - CSP -3.05 -3.44 -4.76) -5.44 -6.42 (-1.07) (-1.32) (-2.17 ** (-2.46) ** (-2.60) * NSR 8.92 4.55 - - - (0.82) (0.47) - - - GIP -4.32 - - - - (-1.17) - - - - INF 4.13 - - 16.96 12.86 (0.36) - - (1.74) * (1.91) *** M2G -2.65 -1.36 - - - (-0.33) (-0.16) - - - MMM -14.33 -8.85 -7.47 -16.06 - (-1.13) (-0.84) (-1.08) (-2.23) ** - LDS -4.85 -2.69 - - 4.81 (-0.91) (-0.60) - - (1.94) *** IMG 3.52 3.33 2.20 1.62 - (0.87) (0.87) (0.87) (0.66) - McFadden R-squared 0.64 0.60 0.59 0.58 0.51 LR Statistic 40.24 *** 37.96 *** 37.25 *** 36.29 *** 32.07 *** Number of Observations 91 91 91 91 91 Number of crises 11 11 11 11 11
Note: (i) *significant at 10%, ** significant at 5%, *** significant at 1%. (ii) Numbers in brackets shows z-statistics.
35 due to the degree of significance shows that CSP is significant as the leading indicator of financial crises. By replacing GDPG with INF, column (5) shows that CSP and INF are significant in determining the occurrence of crises. MMM is significant to the probability of crises, but its estimated sign is different from the expected one. According to this outcome, MMM is not considered as the leading indicator in this case. Finally, one can see in Appendix Table C that the correlation between MMM and LDS are rather high. By including LDS in the model instead of MMM, in column (6), LDS can be considered as a significant leading indicator. From logit model 2, I conclude that growth of real GDP (GDPG), change in stock prices (CSP), the inflation rate (INF) and the spread between lending and deposit rate (LDS) are leading indicators of financial crises. Specifically, GDPG and CSP are found to be leading indicators both in model 1 and model 2.
36 Table 11. The results of multivariate logit model 3
(1) (2) (3) (4) (5) (6) Constant -2.40 -1.82 -3.15 -1.22 -3.72 (-0.68) (-0.64) (-1.18) (-1.09) (-4.54) *** GDPG -41.74 -31.77 - -31.80 *** - (-1.26) (-1.45) - (-1.72) - CSP -5.11 -4.21 -3.91 -4.54 -4.19 (-1.26) (-1.72) *** (-1.74) *** (-1.99) ** (-2.04) ** NSR 13.63 - - - - (0.83) - - - - GIP -4.46 - - - - (-0.72) - - - - INF 40.39 - 7.65 - 8.39 (1.10) - (1.15) - (1.90) * M2G -60.21 - - - - (-1.52) - - - - ERM -1.26 7.68 10.71 7.63 13.11 (-0.09) (0.76) (1.07) (1.08) (1.91) *** M1G 1.18 -1.31 0.96 - - (0.07) (-0.13) (0.10) - - DCY -8.65 -5.86 -5.79 -4.96 -5.42 (-1.29) (-1.15) (-1.18) (-1.62) (-1.82) *** CBD 57.00 0.25 1.20 - - (1.40) (0.03) (0.11) - - RER 0.00 0.00 0.00 - - (091) (0.42) (0.17) - - MFR 0.47 0.19 -0.16 - - (0.59) (0.29) (-0.27) - - WOP -1.44 0.13 -0.31 - - (-0.47) (0.05) (-0.13) - - McFadden R-squared 0.66 0.59 0.56 0.58 0.55 LR Statistic 41.55 *** 36.96 *** 35.12 *** 36.74 *** 34.86 *** Number of observations 91 91 91 91 91 Number of crises 11 11 11 11 11
Note: (i) *significant at 10%, ** significant at 5%, *** significant at 1%. (ii) Numbers in brackets shows z-statistics.
37 crises. The McFadden R-squared has values above 0.5, meaning that the variance of the dependant variable, i.e. the occurrence of crises, is explained well by the variance of the explanatory variables.
To sum up, growth of real GDP (GDPG) and change in stock price (CSP) are significant as financial crises signals in multivariate logit model 1, model 2 and model 3. In addition, the inflation rate (INF) is found to be a leading indicator in logit model 2 and logit model 3. Based on those models, there are also other indicators that may give informative signal to the occurrence of financial crises, such as the national saving rate (NSR), the spread between lending and deposit rate (LDS), and excess real money balances (ERM). Finally, excessive public savings (NSR) and a large excess real money balance (ERM) positively signal the occurrence of crises.
According to Davis and Karim (2008), real GDP growth is one of the factors which explain the likelihood of banking crisis. In addition, Lestano et al. (2003) argue that a slowdown of output increases the probability of crisis. They found that a decline in GDP per capita is significantly related to the likelihood of financial crises in their multivariate logit model. They added that inflation is useful as a leading indicator in both banking crisis and currency crisis. As explained earlier, Kaminsky et al. (1998) found changes in stock prices (CSP) to give an informative signal prior to currency crisis.
38 5.3.1 Performance of the multivariate logit model
To evaluate the performance of the model, I use a common measure that has been applied in the signal extraction approach too, i.e. the percentage of crises correctly called. Following Berg and Pattillo (1999) and Davis and Karim (2008), I set the threshold or cut-off level of 0.5. If the predicted probability of crisis exceeds the threshold and the actual crisis occurs, the crisis is predicted correctly (see Table 2). In other words, the logit model accurately predicts the correct crisis. In Table 12, the percentage of crises correctly called ranges from 60% to 80% for all regressions in the logit models. This indicates that the multivariate logit models predict crisis quite well, since more than half of the financial crises are accurately predicted by a model.
In addition, if the predicted probability is below or equal to cut-off level, and there is no actual crisis occurs, the logit model accurately predicts the tranquil period (see Table 2). The percentage of no crisis correctly called (% no crisis correct) in Table 12 shows a perfect accuracy of the model predicting the tranquil period for all regressions. Overall, the estimated logit models perform excellent in terms of its ability to determine crises and tranquil period. Table 12 demonstrates this performance by showing the percentage of total correct predictions above 95% for each regression.
39 Table 12. The performance of multivariate logit models
Model 1 Model 2 Model 3
(1) (2) (3) (1) (2) (3) (4) (5) (1) (2) (3) (4) (5) % Crisis Correct* 70.0 60.0 60.0 80.0 70.0 70.0 70.0 60.0 80.0 70.0 60.0 70.0 70.0 % No crisis correct* 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 100.0 % Total correct* 96.7 95.6 95.6 97.8 96.7 96.7 96.7 95.6 97.8 96.7 95.6 96.7 96.7 *) Probability cut-off = 0.5.
Table 13. The QPS and LPS in-sample and out-of-sample performance of the models
Model 1 Model 2 Model 3
(1) (2) (3) (1) (2) (3) (4) (5) (1) (2) (3) (4) (5) In-sample QPS 0.060 0.084 0.076 0.056 0.066 0.069 0.069 0.081 0.056 0.068 0.071 0.066 0.070
40 6 Conclusion
The early warning system study in this paper focuses on determining leading indicators for financial crises in Indonesia by using two different approaches, signal extraction approach and multivariate logit approach. The results of the signal extraction approach suggest that real GDP growth (GDPG), changes in stock prices (CSP), national savings rate (NSP), and industrial production growth (GIP) can be considered as leading indicators for financial crises. Those indicators perform well in terms of three measures: the percentage of crises correctly called, the noise to signal ratio, and the probability of a crisis given that signal was issued. Besides, there are several indicators which are useful as the indicators in predicting the financial crises, such as the inflation rate, M2 growth, M2 multiplier, excess M1 balance, the spread between lending and deposit rates, the ratio of domestic credit to GDP, commercial bank deposits, the ratio of M2 to reserves, the real exchange rate, import growth and the world oil price.
In the second method, the multivariate logit model, I estimated the likelihood of crisis occurrence based on the best indicators corresponding to the results of the signal extraction approach. Some indicators are confirmed as leading indicators, such as real GDP growth (GDPG), changes in stock prices (CSP), national savings rate (NSR), inflation rate (INF), lending to deposit rate spread (LDS), and excess real M1 balances. The multivariate logit model shows satisfactory performance in the QPS and LPS both in-sample and out-of-sample model.
41 monitoring several informative leading indicators. Thus, precautionary policy actions can be properly designed to anticipate financial crisis.
The signal extraction approach has the advantage of the flexibility in selecting the indicators regarding the occurrence of financial crises. The indicators are chosen based on theory and previous studies. Kaminsky et al. (1998) also highlight that this method is a useful tool for the financial system early warning system. Due to its practical use, this approach is recommended as the early warning system for policymakers in predicting the financial crisis occurrence. Nevertheless, the signal extraction approach also has a drawback in the absence of standard statistical tests to evaluate the robustness in the relationship between the indicators and the likelihood of crisis occurrence.
42 References
Abiad, A. G. (1999). “Early warning system for currency crises: a Markov-Switching approach with application to Southeast Asia”, available from http://www.ssc.upenn.edu/~abiad/paper1213.pdf [accessed 10 March 2010].
Bank of Indonesia (2001). Economic report on Indonesia. Bank of Indonesia, Jakarta. Bank of Indonesia (2008). Economic report on Indonesia. Bank of Indonesia, Jakarta. Bank of Indonesia (2009). Economic report on Indonesia. Bank of Indonesia, Jakarta.
Bank of Indonesia (2010). Krisis global dan penyelamatan sistem perbankan Indonesia. Bank of Indonesia, Jakarta.
Bank of Indonesia. “History of Bank Indonesia: monetary period from 1966 to 1983”, available from
http://www.bi.go.id/NR/rdonlyres/8236D48A-1175-43A9-B521-2CFD974AD49E/1285/MicrosoftWordHistoryofMonetaryPeriod19661983.pdf [accessed 25 April 2010].
Bank of Indonesia. “History of Bank Indonesia: monetary period from 1983 to 1997”, available from
http://www.bi.go.id/NR/rdonlyres/8236D48A-1175-43A9-B521-2CFD974AD49E/1286/MicrosoftWordHistoryofMonetaryPeriod19831997.pdf [accessed 25 April 2010].
Bank of Indonesia. “History of Bank Indonesia: monetary period from 1997 to 1999”, available from
http://www.bi.go.id/NR/rdonlyres/8236D48A-1175-43A9-B521-2CFD974AD49E/1287/MicrosoftWordHistoryofMonetaryPeriod19971999.pdf [accessed 25 April 2010].
Berg, A. and C. Pattillo (1999). “Predicting currency crises: the indicator approach and an alternative”, Journal of International Money and Finance, 18, 561-586.
Berg, A., E. Borensztein, and C. Pattillo (2004). “Assessing early warning systems: how have they worked in practice?”, IMF Working Papers, 52, International Monetary Fund, Washington, D.C.
