INTRODUCTION
Problem statement
The 1990s witnessed numerous significant financial crises, including those in Europe (1992-1993), Mexico (1994-1995), Asia (1997-1998), Brazil (1999), Turkey (2001), Argentina (2002), and the global economic downturn in 2008-2009 These crises profoundly impacted economies, politics, and societies, leading to heightened economic uncertainty characterized by high inflation, slow growth, elevated unemployment, and increased poverty In Argentina, for instance, the economy suffered a 20% loss in GDP growth, accompanied by a corresponding decline in real wages Policymakers faced immense pressure to implement new strategies for economic recovery, as the crises incurred substantial costs Consequently, there was a surge in empirical studies aimed at developing early warning systems (EWS) to predict future crises.
This thesis primarily examines currency crises, aligning with the focus of most early warning system (EWS) models in previous empirical research While financial crises can manifest as currency, banking, or debt crises, the EWS model specifically targets the dynamics and implications of currency crises.
Early warning systems (EWS) for currency crises were initially developed by Krugman in 1979 and later refined by Flood and Garber in 1984, highlighting reserve loss as a key predictive indicator Obstfeld introduced an alternative model in 1994 and 1996, attributing currency crises to speculative expectations, although it lacked a temporal aspect, making it ineffective in predicting the timing of crises The aftermath of the 1997 Asian crisis spurred the creation of new models, with Kaminsky and Reinhart in 1999 focusing on twin crises that encompass both banking and currency crises.
Currency crises frequently arise alongside banking crises, leading to a dual economic crisis Research typically employs macroeconomic and financial indicators to forecast these currency crises, including metrics such as foreign reserves, export and import levels, real interest rates, real exchange rates, the M2 to reserves ratio, M2 multiplier, and the current account deficit or surplus relative to GDP, as well as short-term debt in relation to reserves (Kaminsky et al., 1998; Frankel and Rose).
In recent years, economists have increasingly focused on institutional factors such as bureaucratic quality, government stability, and effectiveness, as well as elements like voice and accountability, rule of law, democracy, elections, and control of corruption, to predict the likelihood of imminent crises (Berg and Pattilo, 1996; Block, 2003; Shimpalee and Breuer, 2006; Leblang and Satyanath, 2008).
Several methods have been proposed for selecting potential indicators, with the logit model being the most popular and suitable, as demonstrated by Frankel and Rose (1996) and Berg and Pattillo (1999) Another notable approach is the signal method introduced by Kaminsky et al (1998), which has been utilized by researchers such as Edison (2003) and Bruggemann and Linne (2000) Additionally, alternative methods include cross-country regression models suggested by Sachs et al (1996), the Ordinary Least Squares (OLS) method applied by Tornell (1999) and Brussiere and Mulder (1999), and the Markov-switching method utilized by Martinez-Peria.
(1999), Abiad (2003), and Artificial Neural networks (ANN) method applied by Nag and Mitra (1999).
Many early warning system (EWS) models primarily focus on identifying statistically and economically significant indicators for predicting currency crises, yet their predictive capabilities remain largely untested To address this issue, selecting an optimal cut-off threshold is crucial for evaluating EWS models and minimizing crisis risk A lower cut-off point may increase the detection of crises but also raises the likelihood of false alarms (type 2 errors), while a higher cut-off reduces the number of detected crises and decreases the chance of missing signals (type 1 errors) Kaminsky et al (1998) introduced the noise-signal-ratio (NSR) method to determine the optimal threshold that minimizes the ratio of false signals to accurate predictions.
In 1999, researchers demonstrated that their Early Warning System (EWS) models, utilizing the Quadratic Probability Score (QPS) and Log Probability Score (LPS), outperformed those of Kaminsky et al (1998) Building on the work of Damirguc-Kunt and Detragiache (1999), Bussiere and Fratcher (2002) developed a loss function aimed at aiding policymakers in predicting currency crises, emphasizing that optimal cut-off thresholds and predictive periods should align with the degree of risk aversion More recently, Candelon et al (2012) provided a comprehensive overview of methods for determining the most effective cut-off points, concluding that the Credit Scoring approach and Accuracy measure surpass the cut-off point method and noise-to-signal ratio proposed by Kaminsky et al (1998) in evaluating EWS forecast performance.
This thesis aims to enhance Early Warning System (EWS) models by incorporating seven macroeconomic variables identified in previous research, including studies by Kaminsky et al (1998), Berg and Pattillo (1999), and Bussiere and Fratzscher (2002), along with five institutional indicators from Shimpalee and Breuer (2006) Utilizing a logit approach, the study seeks to predict the likelihood of currency crises in emerging markets To assess the effectiveness of these EWS models, the thesis will implement the Credit-scoring method as outlined by Candelon et al (2012), marking a significant advancement in evaluating EWS performance in forecasting currency crises.
Research objectives
Based on the crucial things mentioned above, some objectives of this thesis are to be identified:
- Identifying the crucial indicators that contribute to the EWS models to predict the currency crises in emerging markets.
- Identifying the optimal cut-off point to maximize the correctly prediction of currency crises.
- Suggesting some policy implications to prevent future currency crisis occurrence.
Research questions
To reach the research objectives, the following questions should be answered.
- What are crucial indicators that will be used in the EWS models to forecast the probability of currency crisis occurrence in emerging markets?
- Which is the optimal cut-off threshold of the EWS models using to predict the currency crises in emerging markets?
The scope of the thesis
This thesis analyzes monthly data from five emerging Asian markets—Indonesia, Malaysia, Philippines, Thailand, and Turkey—covering the period from January 1992 to March 2011 The focus on emerging markets is justified by two key factors: previous studies have successfully identified indicators leading to currency crises following the 1990s financial crises and have developed early warning system (EWS) models for predicting such events Additionally, these countries share similar characteristics as emerging markets The dataset begins in January 1992 to encompass all crises from the 1990s and concludes in March 2011 due to limitations in available monthly economic data from the IFS CD-ROM.
The structure of the thesis
This thesis is structured into several key chapters: Chapter 2 offers a literature review on currency crises and early warning system (EWS) models, while Chapter 3 details the research data and methodology Chapter 4 presents the findings from the logit regression model and credit-scoring approach, including robustness tests and comparisons with prior studies Finally, Chapter 5 concludes with recommendations based on the research outcomes.
LITERATURE REVIEWS
Definition of currency crisis
To effectively develop Early Warning System (EWS) models, it is crucial to clearly define a currency crisis A currency crisis typically occurs when speculative attacks on a currency result in significant exchange rate depreciation, compelling the government to either raise real interest rates or deplete foreign reserves to stabilize the currency.
Currency crises must be assessed and converted into measurable value Eichengreen, Rose, and Wyplosz (1995) significantly advanced this field by developing a method to gauge currency pressure and identify crises They built upon the monetary model by Girton and Roper (1977) to create the exchange market pressure (EMP) index, which incorporates the nominal exchange rate, foreign exchange, and interest rates A currency is considered to be under pressure when the EMP index surpasses a specific threshold.
The methodology proposed by Eichengreen et al (1995) faced criticism from other researchers, prompting the development of alternative approaches Frankel and Rose (1996) revised the currency crash index by excluding foreign reserves and interest rates, defining currency crises as instances where depreciation exceeded 25% and the rate of depreciation increased by at least 10% However, Berg and Pattillo (1999) contended that this definition could yield false positives, as some countries routinely experience fluctuations greater than 25% in exchange rates without facing significant economic issues.
Kaminsky et al (1998, 1999) employed a methodology similar to that of Eichengreen et al (1995), but they excluded the interest rate from their analysis, citing central bank control in certain countries Additionally, while Eichengreen et al defined a threshold for economic monitoring when the Early Warning Indicator (EMP) surpasses 1.5 standard deviations from the mean, Kaminsky et al opted for a more stringent threshold of 3 standard deviations.
Bussiere and Fratzscher (2002) employed a methodology akin to that of Eichengreen et al (1995), but with a focus on "real" variables Their approach incorporates the real exchange rate and interest rate to address variations in inflation rates across different countries and over time.
This thesis, thus, will follow the newest method which is based on Bussiere andFratzscher (2002) to date the monthly of currency crisis occurrence.
Theoretical literatures of currency crises
Based on histological evidence of many currency crises occurred in the past, the theoretical literatures of the currency crises are summarized in four generation models as below:
2.2.1 First generation models of currency crises
The first-generation model was built by Krugman (1979) and enhanced by Flood
In the aftermath of the crises prior to the 1990s, Krugman (1979) highlighted that the government could stabilize the foreign exchange rate through various methods, including open-market operations and direct interventions in foreign assets His model identifies two asset types: domestic and foreign currency, indicating that maintaining a fixed exchange rate necessitates the depletion of foreign reserves Consequently, the government faces a trade-off: to avert domestic currency depreciation, it must exhaust foreign reserves, while to prevent appreciation, it must increase the money supply, leading to inflation High inflation alters speculators’ asset portfolios, raising the share of foreign reserves and lowering domestic reserves To uphold the pegged exchange rate while financing budget deficits, the government must sell its reserves, prompting speculators to rapidly sell domestic currency This accelerated depletion of reserves ultimately leaves the government unable to defend the fixed exchange rate, resulting in a currency crisis.
First-generation models identify balance of payment issues as primary causes of financial crises, effectively explaining the crises faced by Latin American countries in the 1980s, which stemmed from fundamental macroeconomic problems like fiscal deficits, excessive monetary supply, and high inflation Key indicators used to predict these crises during this period include losses in international reserves, increases in money supply, rising international interest rates, growth in budget and current account deficits, expansion of domestic credit, and fluctuations in exchange rates.
On the other hand, there were many crises occur in European countries in 1992-
In 1993, unexplained phenomena emerged that challenged first-generation currency crisis models, despite the prevailing belief that macroeconomic fundamentals were applicable This led to the recognition that existing models required enhancement and adaptation to current economic conditions, resulting in the development of second-generation models to better address these challenges.
2.2.2 Second generation currency crisis theoretical model
The second generation models were built after crises in European countries in 1992-
In 1993, Obstfeld (1994, 1996) argued that crises in European countries could not be solely explained by Krugman’s model, despite their strong macroeconomic fundamentals Factors such as high interest rates and rising unemployment prompted governments to respond to the crises between 1992 and 1993 This led to questions about the government's willingness to maintain the foreign exchange rate through borrowing international reserves and exploring other policy options Ultimately, it raised the critical question: "Why does a government decide to abandon a pegged exchange rate?" Obstfeld highlighted the trade-offs inherent in various government policies.
To effectively manage currency devaluation, policymakers must weigh the costs against the benefits; if costs outweigh benefits, they may choose to abandon a pegged exchange rate This theory posits that government policies are influenced by market expectations, which in turn are shaped by those same policies This bidirectional causality contributes to the presence of multiple equilibriums in the economy.
Investor and speculator expectations regarding the stability of a pegged foreign exchange rate significantly influence government policy decisions When the public anticipates a depreciation of the domestic currency, it reflects a general pessimism towards government policies, leading to herd behavior among investors who may convert domestic currency into foreign currency This collective action can ultimately trigger the very depreciation that investors fear Second-generation economic models illustrate how such crises can become self-fulfilling due to the presence of multiple equilibriums.
The crises experienced by European countries in 1992-1993 can be effectively explained by second-generation models, which highlight that these issues stemmed not from weak macroeconomic fundamentals, but rather from self-fulfilling expectations among speculators and the presence of multiple equilibriums.
Models indicate that factors influencing a government's choice to maintain or abandon a pegged exchange rate may provide insights into the potential for a crisis In econometric analysis, these factors include unemployment levels, inflation rates, the structure and amount of debt, and the stability of the financial sector.
However, the second-generation models could not detect the time of crisis occurring Moreover, the reality crises that happened in 1997-1998 led to develop the new generation crisis model.
2.2.3 Third generation currency crisis theoretical model
The third generation models emerged in the aftermath of the 1997-1998 Asian financial crises, which Krugman (1999) attributed to moral hazard and asset bubbles He noted that these crises stemmed from financial institutions, protected by implicit guarantees, engaging in risky lending practices These institutions often favored high-risk projects with the potential for significant returns, despite their low probability of success This moral hazard led to inflated asset prices, particularly in real estate, as financial firms invested heavily in fixed supply assets The expectation of rising real estate prices fueled further investment, creating a cycle where increased demand drove prices even higher, ultimately resulting in the formation of an asset bubble Financial institutions continued to support these investments, perceiving them as "too safe."
When bubble prices surge excessively, a boom occurs until investors recognize that the actual market value differs from their expectations This realization prompts a sell-off as investors anticipate further price declines, leading to a rapid increase in supply that outstrips demand Consequently, prices plummet, and banks respond by reassessing the value of collateral for loans, resulting in a halt to new lending and efforts to recover existing loans.
Radelet and Sachs (1998) attributed the Asian crises to a loss of investor confidence, stemming from doubts about the adequacy of foreign reserves to cover short-term debt In a typical banking scenario, a sudden surge in substantial withdrawals can destabilize a bank, even when its assets initially exceed liabilities As cash flow becomes insufficient to meet withdrawal demands, banks face liquidity issues, leading them to halt new lending and refuse to rollover existing loans Consequently, brokers liquidate stocks to exchange for foreign currency, exacerbating the crisis The absence of effective resolution mechanisms for both enterprise and bank debts contributes to the troubles faced by both viable and struggling businesses, deepening the financial turmoil.
Kaminsky and Reinhart (1999) highlight that early warning systems for "twin crises"—which encompass both banking and currency crises—are crucial for understanding economic vulnerabilities They identify domestic bad loans and short-term foreign debts as key factors leading to banking crises, often occurring prior to currency crises The onset of a currency crisis can exacerbate existing banking issues, ultimately resulting in a dual economic crisis.
The third generation models reveal that prior to the crises, Asian countries exhibited strong macroeconomic indicators, including high GDP growth, low unemployment, controlled inflation and government spending, low budget deficits, significant foreign investment inflows, and political and economic stability However, these positive signs masked underlying issues within the banking system, such as bad debts linked to short-term foreign currency loans Additionally, the phenomenon of crony capitalism, characterized by close ties between the government and financial institutions, created moral hazards through implicit government guarantees This environment contributed to soaring asset and stock prices, ultimately leading to the Asian financial crises.
Third generation models suggest some variables such as: real interest rate, lending or deposit rate growth, domestic credit growth, M2/reserves, bank deposit, bank cash/ bank asset, non-performing loan.
2.2.4 “Fourth generation” currency crisis theoretical model
In addition to the three generations of currency crisis models, there are emerging approaches to explore the causes of these crises The aftermath of the 1997-98 Asian Financial Crisis, highlighted by events in Russia (1998), Turkey (2000-2001), and Argentina (2001-2002), has sparked interest in identifying new causal factors and linkages beyond those established in the existing theoretical frameworks This exploration may be referred to as the "fourth generation theoretical model" of currency crises.
In these models, weak institutions worsen problems associated with economic growth and contribute to causing currency crises.
Li and Inclan (2001) highlight that institutions influence currency crises in two significant ways Firstly, the health of national economies is closely linked to institutional effectiveness; poor institutions can lead to economic decline and increase the likelihood of currency crises, whereas strong institutions can foster positive economic fundamentals and mitigate crisis risks Secondly, institutions serve as vital sources of information for market agents regarding future economic conditions, shaping market expectations When institutions are associated with weak economies, they contribute to market instability and heightened uncertainty, leading to increased capital outflows and a greater risk of currency crises Conversely, robust institutions linked to strong economies promote stable market expectations, reduce uncertainty, and lessen capital outflows, thereby decreasing the likelihood of currency crises.
Empirical studies of currency crises
An Early Warning System (EWS) model is designed to define and predict currency crises, as outlined by Glick and Hutchison (2011) This model typically includes three key components: a method for defining the onset of currency crises, a set of explanatory indicators, and appropriate statistical methods This section will highlight significant indicators commonly utilized in EWS models for forecasting currency crises, followed by a discussion of methodologies employed in prior research Finally, the thesis will summarize findings from various empirical studies related to currency crisis prediction.
Effective operation of Early Warning Systems (EWS) models relies heavily on the careful selection of indicators, as these choices significantly enhance the accuracy of predicting currency crises.
Numerous theoretical and empirical studies, including influential works by Eichengreen et al (1996), Frankel and Rose (1996), and Kaminsky and Reinhart (1999), have identified a variety of potential variables that contribute to currency crises Key macroeconomic fundamentals highlighted include real exchange rate growth, broad money growth, domestic credit growth, and the current account surplus or deficit relative to GDP Additionally, factors such as reserve loss, export and import growth, and short-term debt relative to reserves are crucial Institutional factors like government stability, control of corruption, law and order, and both internal and external conflicts also play significant roles in the occurrence of currency crises.
There are many variables that may possible to enter the predicting model for currency crises These indicators are classified into categories as below:
Capital account: M2/foreign reserves, foreign reserves growth, gross external debt/ export and short-term debt/foreign reserves.
Current account: the growth of real exchange rate, export growth, terms of trade, import growth, and current account/GDP,
Domestic and public real sector: public debt/ GDP, change in stock price, GDP growth, an index of equity prices.
Financial sector: the growth of M1 and M2, M2 multiplier, domestic credit/GDP, domestic real interest rate.
Institutional factors play a crucial role in shaping a country's stability and governance Key elements include the level of openness, the presence of exchange controls, and the impact of government changes Political instability, whether from legal or illegal executive transfers, significantly affects the environment Furthermore, the prevalence of corruption, both external and internal conflicts, and the degree of voice and accountability are critical Lastly, the quality of regulations contributes to the overall institutional framework, influencing economic and social development.
Global economy: US interest rate, growth of world oil prices
2.3.2 Existing methods approach in EWS model of currency crisis
In the 1990s, researchers such as Kaminsky et al (1998), Kaminsky and Reinhart (1999), Berg and Pattillo (1999), Frankel and Rose (1996), Edison (2000), and Bussiere and Fratzscher (2002) focused on developing models for predicting currency crises, known as Early Warning System (EWS) models These models are significant both statistically and economically, employing two common approaches: the Signal approach and the quantitative methods outlined by Edison and others.
2003) and the Logit/Probit model (Frankel and Rose, 1996, Eichengreen et al.,
The signals approach was first built by Kaminsy et al (1998), they proposed many indicators have the unusual exhibit prior the crises In their researches, they selected
In previous theoretical and empirical studies, 15 key indicators have been identified that signal potential currency crises when they exceed predetermined thresholds Selecting the optimal threshold is essential; a lower threshold generates numerous signals, while a higher one may overlook significant warnings When a signal is triggered, it can either indicate an impending crisis, which is a reliable alert, or a false alarm if no crisis follows Conversely, if no signal is issued, it could mean the absence of a crisis, validating the absence of a warning, or it may represent a missed opportunity to alert stakeholders about an actual crisis This relationship can be effectively summarized in a decision matrix.
Crisis occurs Non-crisis occurs
- A: good signal - a number of months the indicators issue the signal, the crisis occurs in following.
- B: false alarm - a number of month the indicators the signal but the crisis does not occur (Type 2 error) or called it “noise”
- C: missing alarm - a number of month the indicators were not issued a signal precede the crisis imminent (Type 1 error).
- D: bad signal - a number of month the indicator “refrain” from the crisis occur, it did not issue the signal and the crisis also does not occur.
The predictors are only observable in cells A and D, indicating that a signal will be issued and a crisis is expected to occur within the next 24 months if A is greater than 0 and C equals 0 Conversely, if D is greater than 0 and B equals 0, no signal will be issued, and a crisis will not occur within the same timeframe.
They defined “optimal” threshold at minimize the noise-to-signal ratio (NSR), that is the false signals to good signals ratio and calculate as equation below:
Kaminsky et al (1998) proposed an alternative interpretation of signal indicators by analyzing the conditional probability of a crisis [A/(A+B)] in comparison to the unconditional probability [A+C/(A+B+C+D)] For indicators to provide valuable insights, the conditional probability must exceed the unconditional probability Additionally, their research aimed at early warning for currency crises within a 24-month timeframe, ranking indicators based on their predictive ability for the initial signal of a crisis.
Zhang (2001) highlighted the usefulness of a method for evaluating abnormal indicators in predicting currency crises, noting its ease of application and interpretation However, this basic approach has significant limitations, as Glick and Hutchison (2011) pointed out It assesses each indicator in isolation, failing to analyze their interrelationships and contributions to currency crisis likelihood Additionally, Abiad (2003) criticized the signaling approach for not measuring the marginal effects of indicators on crisis probability and for overlooking multicollinearity issues Furthermore, it lacks statistical testing and comparative analysis with other methods, undermining its effectiveness.
The logit/probit approach addresses limitations of the signaling method by estimating the relationship between a binary dependent variable—indicating the occurrence of a currency crisis—and explanatory variables This method accounts for correlations among the explanatory variables, measures the marginal effects of each variable on the probability of a crisis, and tests the statistical significance of individual variables Additionally, it provides probabilities for potential future crises (Glick and Hutchison, 2011).
The methodology involves utilizing crisis dummy variables, which are binary indicators (1 or 0) as outlined in section 2.3.1 To effectively predict a crisis at its onset, the definition of a crisis is redefined within specific exclusion windows or forecasting horizons, such as 1, 3, 12, 18, or 24 months Additionally, it necessitates a collection of explanatory variables identified in previous literature that serve as potential early warning indicators.
Then, the probability of a crisis occur is calculated by the equation following
And the probability of crisis not occur is (1- P) The outcome of model is the probability of a crisis occurs in given value in the next k months (predicting time).
The logit/ probit approach were applied by many studies such as Eichengreen et al.
(1995), Frankel and Rose (1996) and Berg and Pattillo (1999).
Figure 2.1: The flowchart of developing an EWS model to predict currency crises 2.3.3 Summary of recent empirical findings
Numerous empirical studies have investigated currency crises, with this thesis summarizing research that utilizes the Early Warning System (EWS) model for predicting such events It also highlights the explanatory variables proposed in these studies.
Eichengreen, Rose, and Wyplosz (1995) defined currency crises using the EMP index, which incorporates nominal exchange rates, foreign reserves, and interest rates, identifying crises that exceed the mean plus one and a half standard deviations Despite using a similar definition, their analysis revealed variations in crisis occurrences by adjusting the standard deviation threshold to two or three Their study utilized quarterly data from 20 industrial countries spanning from 1959 to 1993 and employed the probit method to estimate the contagion of currency crises, finding significant results.
Frankel and Rose (1996) used annual data of 105 developing countries from 1971 -
In 1992, researchers identified currency crises as instances where the exchange rate depreciated by more than 25%, with a minimum of 10% change compared to prior depreciation They analyzed sixteen variables across four categories: foreign variables, macroeconomic variables, external variables, and capital inflow composition Using a probit approach to estimate early warning system (EWS) models within a three-year exclusive window, they discovered that crises were linked to low foreign direct investment (FDI) inflows, diminished foreign reserves, excessive domestic credit growth, elevated interest rates, and overvaluation of the real exchange rate.
Kaminsky et al (1998) introduced the Early Warning System (EWS) model, known as the KLR model, which utilized 15 monthly indicators from 20 countries between 1970 and 1995 to signal potential crises when specific thresholds were crossed By employing a signal approach, they estimated the probability of crises occurring within the subsequent 24 months based on the current indicators Their findings demonstrated the model's effectiveness in predicting the crises of 1997.
Conceptual framework
This thesis will analyze 12 potential early warning indicators, categorizing the explanatory variables into two groups: macroeconomic factors and institutional factors, based on a comprehensive review of theoretical and empirical literature alongside available data.
Macroeconomic factor: reserve loss, export growth, import growth, real exchange rate growth, current account surplus/GDP, short-term debt/reserve and the GDP growth.
Institution factor: government stability, corruption, law and order, external conflict and internal conflict.
This thesis is going to test relationship between currency crises occurring in the next 12 months and the explanation variables in emerging markets as the following:
RESEARCH METHODOLOGY AND DATA
The EWS model specification
To develop the Early Warning System (EWS) model, three key components must be addressed: identifying the timing of the currency crisis, selecting appropriate indicators, and determining the statistical methods to be employed.
3.1.1 Dating the currency crisis and define the dependent variable
Currency crises can be defined in various ways, including the observation of changes in exchange rates alone (Frankel and Rose, 1996), or by assessing changes in both exchange rates and international reserves (Kaminsky et al., 1998) Additionally, some definitions consider fluctuations in nominal exchange rates, international reserves, and interest rates (Eichengreen et al., 1995, 1996; Bussiere and Fratzscher, 2002), while others focus on changes in real exchange rates, international reserves, and real interest rates (Bussiere and Fratzscher, 2002).
This thesis utilized the Exchange Market Pressure (EMPi,t) index, as developed by Bussiere and Fratzscher (2002), to detect currency crises in country i during period t The EMPi,t was determined by calculating a weighted average of three key factors: the real effective exchange rate (RER), the real interest rate (r), and foreign reserves (RES).
It presented in equation as below:
According to Bussiere and Fratzcher (2002), a currency crisis (CC) was considered occurred at country i in period t if EMPi,t is greater than its average plus two standard deviations.
This thesis aims to identify the timing of currency crises to predict their likelihood within a specific timeframe It transforms contemporary currency crisis variables into a forward-dependent variable, Yi,t, enabling a more accurate analysis of potential currency crisis occurrences.
This model aims to predict the likelihood of a crisis occurring within the next 12 months, striking a balance between timely intervention and the risk of overreacting A shorter forecast may leave policymakers with insufficient time to respond effectively, while a longer prediction could lead to unnecessary preemptive measures Therefore, a 12-month forecast serves as an optimal compromise, allowing for proactive decision-making without inciting self-sufficiency issues.
3.1.2 Explanation variables choice and hypothesis testing
The selection of explanatory variables for our Early Warning System (EWS) model is grounded in existing literature on currency crises, emphasizing the significance of indicators identified in prior studies and the availability of data Consequently, this thesis incorporates 12 potential early warning indicators, categorizing them into two distinct sets: macroeconomic factors and institutional factors.
Macroeconomic factors encompass seven key variables that significantly influence economic performance: reserve loss, export growth, import growth, real exchange rate growth, the current account to GDP ratio, the short-term debt to reserve ratio, and GDP growth.
Institution factor: includes 5 variables, there are: government stability, corruption, law and order, external conflict and internal conflict.
The analysis utilized various variables derived from extensive research, as detailed in Table A.1 (Appendix) Specifically, five macroeconomic factors were measured following the framework established by Berg and Pattillo (1999b), while two additional variables were based on the work of Kaminsky et al (1998) Furthermore, all institutional variables were measured according to the methodology proposed by Shimpalee and Breuer (2006).
A significant decline in foreign reserves serves as a reliable indicator of potential currency devaluation, often preceding a currency collapse that follows efforts to maintain a pegged exchange rate This substantial reduction in reserves increases the likelihood of a currency crisis, highlighting the economic challenges faced by a country Consequently, the depletion of foreign reserves emerges as a critical warning sign of impending currency instability, as noted by Kaminsky et al (1998) and Berg and Pattillo.
Research by Glick and Moreno (1999), Bussiere and Fratzscher (2002), Edison (2003), and Tuan (2009) demonstrates a significant influence of this indicator on currency crises, with an anticipated positive correlation.
H1: Reserve loss has positive relationship with probability of predicting currency crises
Numerous studies, including those by Kaminsky et al (1998) and Edison (2003), indicate that a decline in export growth may be linked to an overvalued exchange rate An overvalued currency allows domestic consumers to purchase more foreign currency, making exported goods more expensive for foreign buyers and subsequently reducing demand This decline in demand can lead to increased unemployment Additionally, exchange rate instability can provoke speculative attacks on the currency, suggesting that decreasing export growth may serve as a negative leading indicator for potential currency crises.
H2: Export growth has negative relationship with probability of predicting currency crises.
A decrease in import growth, contrary to the rise in export growth, may result from a devalued exchange rate, which makes foreign goods more expensive and domestic exports cheaper This situation leads to a decline in demand for imports, ultimately lowering living standards Additionally, high levels of imports can contribute to a current account deficit, increasing the risk of a currency crisis (Kaminsky et al., 1998; Berg and Pattillo, 1999; Edison, 2003) Consequently, the expected sign is positive.
H3: Import growth has positive relationship with probability of predicting currency crises.
According to the Kaminsky et al (1998); Berg and Pattillo (1999); Glick and Moreno (1999), Kamin et al (2001); Bussiere and Fratzscher (2002), Edison
In a pegged currency system, high inflation can lead to an increase in real exchange rate growth, potentially exerting pressure on exports When the real exchange rate appreciates, it may result in currency devaluation and a heightened risk of currency crises Consequently, the expected coefficient in this model is positive.
H4: Real exchange rate growth has positive relationship with probability of predicting currency crises.
3.1.2.5 Current account surplus/GDP ratio
A current account surplus occurs when a country's exports exceed its imports, leading to greater capital inflows than outflows This surplus, expressed as a ratio to GDP, can help mitigate currency devaluation and reduce the risk of currency crises Researchers such as Berg and Pattillo (1999), Kamin et al (2001), and Bussiere Fratzscher (2002) have utilized this indicator to analyze economic stability.
(2009) with the expected sign of this indicator is negative.
H5: Current account surplus/GDP has negative relationship with probability of predicting currency crises.
Berg and Pattillo (1999b) indicate that a higher ratio of short-term foreign debts to reserves signals caution to external lenders and investors, leading to a reduction in foreign currency supply and increasing the risk of currency crises This finding is also significant in the model developed by Bussiere and Fratzscher.
(2002) and Leblang and Satyanath (2008) Accordingly, the expected sign is positive.
H6: Short-term debt/reserve has positive relationship with probability of predicting currency crises.
How to choose the optimal cut-off threshold
The primary objective of any Early Warning System (EWS) model is to accurately distinguish between crisis and non-crisis periods This distinction enables policymakers to implement appropriate strategies during both stable and turbulent times In this section, we will propose a method to evaluate the effectiveness of the EWS by determining the optimal cut-off threshold.
The optimal cut-off threshold is essential for distinguishing between crisis and tranquil periods in an Early Warning System (EWS) model By adjusting the parameters at time t, we can calculate the probability of a crisis occurring within a specified future period (t + k) using a defined equation If this probability exceeds the cut-off point, it indicates a crisis period; if it falls below, it signifies a tranquil period Type I error occurs when a crisis happens despite the probability being below the cut-off, resulting in a missed alarm, while Type II error arises when the probability is above the cut-off, but no crisis occurs, leading to a false alarm A higher cut-off point reduces false alarms but increases missed alarms, whereas a lower cut-off point does the opposite Thus, selecting the optimal threshold involves balancing Type I and Type II errors to minimize both missed and false alarms.
Determining the optimal cut-off point is crucial in early warning system (EWS) models; however, existing methods often appear arbitrary (Candelon et al., 2012) While numerous studies have examined EWS model specifications and the influence of various explanatory variables, there is a notable lack of empirical research focused on evaluating EWS performance, particularly in selecting the most effective cut-off point.
Currently, there are two prevalent methods for selecting cut-off thresholds in crisis signal detection The first method involves setting various cut-off levels, such as 0.5, 0.25, or 0.1, and evaluating the resulting crisis signals to determine the most effective option, often guided by loss functions and qualitative assessments for policy makers (Bussiere and Fratzscher, 2002) The second method, proposed by Kaminsky et al (1998), focuses on minimizing the noise-to-signal ratio (NSR) to establish the optimal cut-off threshold Unlike the first method, which does not differentiate between Type I and Type II errors, the NSR ratio specifically addresses the balance between false alarms and accurate signals, primarily considering Type II errors while neglecting Type I errors.
Candelon et al (2012) introduced two methods for determining the absolute optimal cut-off by addressing two types of errors The first method, based on accuracy measures by Lambert and Lipkovich (2008), aims to maximize the correct identification of crises while minimizing false alarms and missed detections, effectively optimizing the Youden index The second method, derived from the credit-scoring approach of the Basel Committee on Banking Supervision (2005), focuses on minimizing the absolute difference between correctly predicting crisis occurrences and accurately identifying non-crisis situations.
We let three methods discuss more on the future researches In this thesis, we will describe the Credit-scoring approach in details as below.
Credit scoring, as outlined by the Basel Committee on Banking Supervision in 2005, incorporates the concepts of sensitivity and specificity Sensitivity (Se) refers to the ratio of correctly identified crises by an Early Warning System (EWS) model to the total number of crises observed In contrast, specificity (Sp) measures the proportion of accurately identified non-crisis events that the EWS model does not signal, relative to the total number of non-crises The optimal cut-off is determined by the formula: c = min | Se – Sp |.
We have the matrix based on Kaminsky et al (1998) as below:
Number of False alarm – Type 2 error
Number of Missing alarm – Type 1 error
- A is number of good signal of correctly crisis,
- B is number of false alarm, when having signal but no crisis occur
- C is number of missing alarm, when having no signal but crisis occur
- D is number of no signal and no crisis occur
The sensitivity and specificity were calculated as following
The optimal cut-off point in this method balances sensitivity and specificity, as illustrated in Figure 3.2 The sensitivity curve slopes downward, indicating that a higher cut-off leads to fewer crisis signals and a lower percentage of correctly detected crises Conversely, the specificity curve slopes upward, showing that a higher cut-off results in more non-crisis signals and an increased percentage of accurately identified non-crises While the shapes of these curves depend on the specific model of the Early Warning System (EWS), the optimal cut-off is consistently found at the intersection of both curves, as depicted in Figure 3.2.
Figure 3.2: The optimal cut-off identification
Data collection
This study utilizes panel data spanning from January 1992 to March 2011, focusing on five emerging Asian markets—Indonesia, Malaysia, the Philippines, Thailand, and Turkey—along with four emerging Latin American markets: Argentina, Brazil, Colombia, and Mexico, to evaluate out-of-sample performance.
The economic variable data utilized in this analysis primarily comes from the International Financial Statistics (IFS, CD-ROM, 2011), supplemented by information from the World Bank (WB) Additionally, all institutional variables are sourced from the International Country Risk Guide (ICRG, 2012) While most data is collected on a monthly basis, any available annual or quarterly data has been converted to a monthly format for consistency.
The lists of the data, sources and period time are presented in Table A.2 (Appendix)
Estimation strategy and statistical tests of the model
To achieve the research objective this strategy could be classified in following steps:
The model was estimated to identify key indicators for predicting currency crises by incorporating all potential variables Statistical tests conducted included multicollinearity assessment, specification error testing, and goodness-of-fit evaluation following the logit regression analysis.
- After having the EWS model, this thesis applied the credit-scoring method that discussed above to find out the optimal cut-off point for EWS model.
- Then, to evaluate the predictive ability of our EWS model, this thesis applied this EWS model to predict the currency crises in Asian crises in 1997, Turkey crises in 1994 and 2001.
- Finally, this thesis test out-of-sample in Latin American area to exam whether these methods can be applied to the other regions.
RESEARCH RESULTS
The descriptive statistic of the sample
Table 4.1 reveals that emerging countries in Asia experienced an average reserve loss of -17.19%, with values ranging from 53.33% to -374.54% During stable periods, the average reserve loss is -18.89%, while in times of crisis, it drops to -2.81% This suggests that a reserve loss near -2.81% could serve as an indicator of an impending crisis.
Besides, mean of export growth is 13.14%, it approximate with mean of export growth in the tranquil time (12.26%), and it should be considered when the export growth is reach around 20.62%.
Observing import growth can be challenging, as the average growth rate remains consistent across different periods, including stable and crisis times, at approximately 12.50%.
The real exchange rate growth exhibits a significant disparity, ranging from -44.87% to 837.01%, with Indonesia and Turkey showing the largest fluctuations Indonesia's exchange rate growth varies between -44.87% and 837.01%, while Turkey's ranges from -16.18% to 615.88% In comparison, Malaysia's growth spans from -13.95% to 86.99%, the Philippines from -16.63% to 73.65%, and Thailand from -31.76% to 126.96% Due to these substantial gaps, the average exchange rate values differ markedly among these countries Indonesia has an average of approximately 32%, Turkey's average is notably high at 119%, while the other countries maintain lower averages, ranging from 1.7% to 7%.
The average current account to GDP ratio is approximately 1% during stable periods, dropping to around 0.77% during crises The limited fluctuation between the highest and lowest values throughout different periods suggests that this indicator requires close monitoring Additionally, since the current account ratio is influenced by GDP growth, it is essential to keep an eye on it, especially during times of low economic growth.
The Short-term debt/reserve ratio fluctuates between a minimum of 9.05% and a maximum of 236.2% during stable periods, while the minimum during crises is notably lower at 32.28% It is advisable to maintain the Short-term debt/reserve ratio within the range of 9.05% to 32.28%, as values exceeding 32.28% warrant close monitoring until they approach the average crisis level of 120.48%.
The disparity between the highest and lowest GDP growth rates remains consistent during both stable and crisis periods Throughout all time frames, the average GDP growth is approximately 4.5% in stable conditions, while it plummets to nearly 1% during crises Consequently, when GDP growth falls to around 3%, 2%, 1%, or lower, it serves as a clear indicator of an economic crisis.
Institutional variables show a slight increase during crises, with government stability averaging between 8.2% and 8.8% and internal conflict ranging from 8.5% to 9.6% In contrast, external conflict remains low, averaging around 3.4% to 3.8% across all periods Additionally, corruption and law indices show average values of approximately 2.5% to 3.0% and 3.5% to 4.2%, respectively Consequently, it is challenging to identify clear signals of crisis through institutional variables, necessitating careful observation and analysis of any changes.
Table 4.2 indicates the absence of multicollinearity among the independent variables, with a mean VIF of 1.79 and individual VIF values ranging from 1.14 to 3.04 Additionally, Table 4.3 illustrates the correlation among the explanatory variables, confirming that there is no significant correlation between them.
Table 4.1: The summary of sample used in the regressions
A - The summary of sample used in the regressions in whole time 1992-2011
Variables Obs Mean Std Dev Min Max
B - The summary of variables in tranquil time from 1992-2011
Variables Obs Mean Std Dev Min Max
C - The summary of variables in crisis time from 1992-2011
Variables Obs Mean Std Dev Min Max
Table 4.2: The multicollinearity between independent variables
Table 4.3: The correlation between independent variables
RESERVE EXPORT IMPORT RER CAGDP STDRES GDP GOVERN-
Empirical results
Table 4.4 presents the results of a binary logit model analyzing data from five Asian countries, utilizing a total of 1,155 observations The model's likelihood ratio (LR) statistic is 569.71, which is highly significant with a p-value of less than 0.001, indicating that the null hypothesis—asserting that the coefficients of all regressors are simultaneously zero—can be rejected Consequently, it can be concluded that the model is significant overall, and all twelve variables included are essential to the analysis.
The thesis employs the Hosmer-Lemeshow goodness-of-fit test, calculated using the Pearson chi-square from the observed and predicted frequency contingency A successful goodness-of-fit test, indicated by a large p-value, suggests a strong model fit As shown in Table 4.4, the Hosmer-Lemeshow chi-square statistic is 2.49 with a p-value of 0.9623, confirming that the model adequately fits the data.
The specification error test results presented in Table 4.5 indicate that the _hatsq variable is insignificant (p = 0.760), suggesting that the linktest is not significant Therefore, we can conclude that the model does not exhibit specification error, supported by the significance of the _hat variable.
(p