The Global Financial Crisis (GFC) has affected many regions including Latin America. This paper focuses on currency crises in Argentina and Brazil, the two largest economies in South America, and with a wide experience with currency crises. We estimate an Early Warning System, consisting of a static factor model and a multinomial ordered logit model, with monthly data for 1990-2007. Ex ante forecasts for 2008-2009 produce an increased probability of currency crises in the fall of 2008. Our model outcomes confirm that elements from earlier crises are useful to predict the currency crises during the GFC.
La crisis financiera mundial (GFC por sus signos en inglés) del 2007-2009 ha afectado a muchas regiones, incluyendo América Latina. Este documento se centra en las crisis cambiarias en Argentina y Brasil, las dos economías más grandes de Sudamérica, y con una amplia experiencia con crisis cambiarias. Estimamos un sistema de alerta temprana, que consiste en un modelo estático de factores y un modelo logit multinomial ordenado, con datos mensuales de 1990 a 2007. Predicciones ex ante para 2008-2009 producen un aumento de la probabilidad de crisis cambiarias en el otoño de 2008. Nuestros resultados confirman que elementos de crisis anteriores contienen información relevante para pronosticar las crisis cambiarias durante la crisis financiera mundial.
In the midst of the Global Financial Crisis, the fall of Lehman Brothers in September 2008 causes a global panic that affects many emerging economies including the two largest economies in South America: Brazil and Argentina. The currencies of several emerging economies depreciate sharply versus the US dollar (see
The Global Financial Crisis and its effects on other countries and currencies have been studied extensively, with different findings.
We approach the currency crises from an Early Warning System (EWS) perspective. We design a parametric model that issues signals or warnings well ahead of a potential currency crisis. The dozens of EWSs that have been developed differ widely in the definition of a currency crisis, the period of estimation, data frequency, the countries included, the inclusion of indicators, the forecast horizon, and the statistical or econometric methods used. For extensive overviews, see
The majority of the EWSs is estimated for a panel consisting a large number of emerging economies. With its rich history of financial crises (
Argentina’s long history of currency crises and other financial crises is analyzed in several studies.
We investigate if currency crises from the period 1990-2007 contain information that is useful to predict the currency crises during the GFC (2008-2009) for Argentina and Brazil. We focus on the period since the early 1990s, because of data availability and because of the significant break with the periods of hyperinflation and the remains of inward-looking economic policies. The debt restructuring (Brady plans) and successful reforms to eliminate hyperinflation in Argentina (1991) and Brazil (1994) mark the start of this period. We model the probability of a currency crisis in an ordered logit model to account for the severity of currency crises. We distinguish three classes of currency crises (mild, deep and very deep crises) and tranquil periods (periods in which no currency crisis took place). We estimate the model and analyze the outcomes for each country independently. We include a large number of possible crisis indicators, based on the three generations of currency crisis models. Our choice to include a wide range of variables instead of preselecting explanatory variables is inspired by
We find that our model predicts the currency crisis in 2008-2009 in Argentina and Brazil reasonably well, in particular for Brazil. We interpret this result as follows. Currency crises that have occurred in the period up to 2007 contain information that is useful for predicting the currency crises in the GFC. This is in contrast to the common opinion that the GFC is not comparable with earlier crises. We see similarities and differences in the run-up to currency crises between the periods 1991-2007 and 20082009. Similarities consist of the sudden stop in capital flows, drops in commodity prices and a slowdown in international trade, which can be considered external causes (
This paper is structured as follows. Section “Methodology” discusses our method. The data and brief descriptions of the Argentinian and Brazilian crises are presented in Section “Data and background”. Section “Empirical results” presents the empirical results and the analysis of out-of-sample performance, followed by the conclusion.
We first apply factor analysis to extract the factors from a large set of indicators, then use the estimated factors as regressors in the ordered logit model, with a crisis dummy as dependent variable, and lastly compute ex ante forecasts. Before we turn to these models, we first discuss crisis dating.
We use the speculative pressure approach, inspired by
As is common in EWSs of currency crises, we assign the same dummy variable value for the run-up period to the crisis.
We use the Static Factor model to avoid pre-selecting indicators, because we want to capture several crisis mechanisms. In factor models an observable set of
One of the issues in factor analysis is the determination of the optimal number of factors. Various procedures have been proposed, such as the Bayesian Information Criterium, the Kaiser Criterium and Cattell’s scree test. With the large dimensional factor models of recent years many studies have proposed solutions and consistent estimators for the number of factors using different factor models and distributional assumptions. Here we employ the criterion of
Using factor models comes at a cost. Determining the economic relevance of factors and interpreting the factors in a meaningful way is problematic. Many indicators enter in more than one factor, so focusing on a single factor only partially explains the full impact of an indicator on the probability of a crisis, and may even lead to counterintuitive results. In this paper we abstain from interpretation of the explanatory variables.
We use the ordered logit model to be able to distinguish in crisis severity. Severe crises have a different run-up than milder crises. Our dependent variable can take four values: no crisis, mild crisis, deep crisis, and very deep crisis. We employ an ordered choice model, which extends the binary choice model, allowing for a natural ordering in the outcomes
Most institutional variables that we use have low variation. When there is limited overlap in the values of (a set of) explanatory variables and the outcomes of the dependent variable, the regressions yield large estimates and standard errors. This problem is called quasi-complete separation (
The models are estimated using data up to and including 2007, and we test the out-ofsample performance of the estimated models for the period 2008M1-2009M12. We forecast the probabilities of a mild, deep and very deep crisis with our ordered logit model. We use realized monthly data for the indicators for the years 2008 and 2009, and extrapolate the factors without re-estimating the loadings of the static factor model.
We use the quadratic probability score (
The
Our sample starts in the early 1990s, after the effects of spillovers of the 1980s Latin American debt crisis have faded out and when hyperinflation has been overcome. The analysis for Argentina starts after the introduction of the Convertibility Plan (April 1991) and for Brazil after the introduction of the Real Plan (July 1994), which both can be regarded as a break with the hyperinflation periods.
We estimate the Exchange Market Pressure Index (EMPI) based on the period up to December 2007, and extend this to December 2009. For the weights we use the standard deviations from the period up to and including December 2007.
Both Argentina and Brazil have experienced several currency crises in the period 1991, resp. 1994 to 2007. Graphs of the currency crises (employing a 12 months run-up) are plotted in
Argentina experiences a currency crisis in 1995, which is related to the “tequila crisis” in Mexico. Argentina’s banking system suffers from large deposit withdrawals, as rumors spread that systemic banks may fall due to their exposure to sovereign debt. In March 1995 the peg is successfully defended through the use of reserves and interest rate. This event is considered the first currency crisis of the Convertibility Plan that was launched in April 1991 (
Brazil experiences a currency crisis in 1995, caused by the contagion from the “tequila crisis”. Similar to Argentina, Brazil is able to defend the peg by increasing interest rates and using reserves. In 1997 Brazil comes under pressure with the Asia crisis. The Brazilian currency is overvalued by up to 40%, the commodity prices drop, increasing the current account deficit, which is financed by borrowing in the external capital markets (
In the run-up to the crisis in the fall of 2008 both Argentina and Brazil experience a period of economic prosperity in the 2002-2007 boom, featuring large foreign reserves, small sovereign external debt levels, small fiscal deficits (or even surpluses), and a more flexible exchange rate regime. But there are differences too. Brazil faces a strongly appreciated currency before the onset of the crisis and an unprecedented large amount of foreign reserves (
In the fall of 2008 both countries experience a currency crisis, considered a mild crises according to our definition. The Brazilian real depreciates more and faster than the Argentinian peso. Both countries are hit by an unusually heavy drop in export earnings between the fourth quarter of 2008 and the first quarter of 2009. Brazil is hit by a second exogenous shock: heavy reversals in capital flows in the fourth quarter of 2008. Surprisingly, Brazil does not experience a major financial crisis, or even a worsethan-average deceleration in economic growth (
For the explanatory variables we select the “usual suspects” (the common macroeconomic and financial variables), institutional and political variables, commodity-related and global indicators. The selected indicators can be classified into separate categories:
13 external economic indicators, among which the deviation from real exchange rate trend, exchange rate volatility, growth of exports, imports and foreign reserves, import cover, ratio of M2 to foreign reserves. 17 domestic economic indicators, among which domestic real interest rate, inflation, M2 multiplier, industrial production, share market index return. 14 institutional and political indicators, among which Herfindahl indices, political stability, corruption, investment profile, internal conflict, election years. 10 debt indicators, among which total debt, short term debt, debt service, arrears. 11 banking sector indicators, among which credit to public sector, to private sector, ROE, deposits. 5 global and regional indicators, among which world economic growth, US yield, contagion dummy. 12 commodity related indicators, among which prices of oil, metals, agricultural products, exports and imports of fuel, agricultural products, food and metals as percentage of GDP.
The main sources for the data are the International Financial Statistics (IFS) database of the IMF, the World Development Indicators (WDI) from the World Bank, International Country Risk Guide (ICRG) database of the Political Risk Services Group, and
Some data limitations exist. Not all time series are sufficiently long which limits the selection of explanatory variables. A particular issue is the exchange rate regime. Both countries have moved from a controlled exchange rate regime to a more freely floating regime after a major crisis hits the country (
The series have been tested for non-stationarity (using Augmented Dickey-Fuller tests) and visually inspected for seasonal effects. Where necessary a transformation is made to render them stationary. To deal with mixed frequencies in series, we apply simple quadratic interpolations. All series are normalized, i.e. demeaned and divided by its sample standard deviation.
We estimate the ordered logit model for Argentina and Brazil for the period up to and including 2007 using a run-up period of 12 months, which is shown in Subsection “Regressions”. In “Forecast performance” we use the estimated model to predict out-of-sample. In “Robustness checks” we present robustness checks, using a 6 and 24 months window exclusion period.
Notes: *: significant at 10%, **: significant at 5%, and ***: significant at 1%; Explanations of the symbols used: SF1: Static Factor 1, SF2: Static Factor 2, et cetera; ∆ (INTCONFL): Change in internal conflict dummy variable (increase implies less internal conflict, or a lower risk); ∆ (LAWORD): Change in the law and order dummy variable (increase is improvement in law and order, representing a lower risk). Includes the strength and impartiality of the legal system and law enforcement; ELECLEGYR: Dummy variable that is 1 if there is an election year for the legislative power and 0 otherwise. Adjusted Pseudo Source: Own elaboration.
Coefficient (standard error)
Coefficient (standard error)
SF1
-0.269 (0.121)
SF8
0.167 (0.219)
SF2
-0.406 (0.114)
SF9
-0.376 (0.183)
SF3
0.341 (0.174)
SF10
0.234 (0.271)
SF4
0.215 (0.104)
∆ (INTCONFL)
1.074 (0.235)
SF5
-1.252 (0.247)
∆ (LAWORD)
0.542 (0.380)
SF6
0.064 (0.183)
ELECLEGYR
1.048 (0.530)
SF7
0.847 (0.193)
Adjusted pseudo
0.411
In this section we investigate the out-of-sample performance of the estimated models in the period 2008M1-2009M12.
Complementary to the graphic representations in
Notes: *: significant at 10%, **: significant at 5%, and ***: significant at 1%; Explanations of the symbols used: SF1: Static Factor 1, SF2: Static Factor 2, et cetera; ∆ (BURQUAL): Change in the bureaucratic quality dummy variable (increase is improvement in bureaucratic quality, or lower risk). High quality when the bureaucracy has the strength and expertise to govern without drastic changes in policy in government service; ∆ (LA WORD): Change in the law and order dummy variable (increase is improvement in law and order, representing a lower risk). Adjusted Pseudo
Coefficient (standard error)
Coefficient (standard error)
SF1
0.243 (0.049)
SF7
0.589 (0.125)
SF2
-0.318 (0.079)
SF8
0.433 (0.127)
SF3
0.168 (0.064)
∆ (BURQUAL)
-1.441 (0.319)
SF4
-0.048 (0.081)
∆ (LA WORD)
1.336 (0.293)
SF5
-0.627 (0.108)
Adjusted pseudo
0.287
SF6
-0.085 (0.111)
Source: Own elaboration.
Country
Mild
Deep
Very deep
Argentina
0.778
0.200
0.022
Brazil
0.627
0.606
0.100
Recall that the closer the score statistics in
The outcomes for Brazil are similar to the outcomes for Argentina. The model performs better for predicting the correct severity of a crisis (mild crisis), as the
Comparing the performance of our model to the actual outcome, we observe that predictions for Brazil are better than Argentina’s predictions. The predicted probability of a crisis in Brazil increases rapidly in 2008 (
We also perform the analysis for two alternative run-up periods. An Early Warning System with a run-up period of 12 months may be too short for the authorities to implement policies to avoid a currency crisis. An Early Warning System with a run-up period of 6 months is likely to be more accurate as indictors will show special behavior as a crisis is approaching. The results for the models with a 6 and 24 months run-up period are shown in the Appendix and are discussed below.
Contrary to findings of Kaminksy (2006) we find that the model is sensitive for changes in the length of the run-up period. The fit of the regression during the insample period differs from one run-up period to the other, as shown in
For the forecasted period (2008-2009) the results are more homogenous, as shown in
The financial panic that follows the fall of Lehman Brothers in September 2008 affects many countries and regions including Latin America. In Brazil the exchange rate depreciates by more than 40%, the Argentinian peso depreciates by 25% and financial markets (stocks, bonds) are hit hard. This paper investigates the experience of the two biggest South American countries with currency crises since the 1990s, and whether the currency crisis in 2008 could have been foreseen -despite the common opinion that the GFC has very distinct features than earlier crises, which makes it impossible to compare with earlier crises.
We develop an Early Warning System for currency crises since the early 1990s up to 2007. We develop an EWS consisting of an ordered logit model, using static factor models to reduce the dimension of the information set. We use our EWS to forecast the probability of currency crises in 2008 and 2009. Our model predicts the currency crises in the fall of 2008 for Argentina and Brazil, although the predicted crises are more severe than actually occurred. We conclude that the currency crises in 1991-2007 contain information that is useful for predicting the currency crisis in 2008-2009. Our model performs better for Brazil than for Argentina, which we attribute to (i) more crisis observations in Brazil imply a better model fit, which is favorable for predicting crises and (ii) Argentina’s economy is relatively closed prior to the GFC while Brazil’s economy is more open.
In this appendix we present the results of robustness checks of the model presented in Section “Empirical results” we use two alternative run-up periods, of 6 and 24 months respectively, instead of 12 months. With a 6 (24) months run-up period, the six (twenty four) months preceding the crisis are considered as the period that ends in a currency crisis. Therefore the months in the run-up period are assigned the same dummy variable value as the crisis itself.
The fit (in terms of adjusted pseudo
Notes: *: significant at 10%, **: significant at 5%, and ***: significant at 1%; Explanations of the symbols used: SF1: Static Factor 1, SF2: Static Factor 2, et cetera; ∆ (LA WORD): Change in the law and order dummy variable (increase is improvement in law and order, representing a lower risk); ∆ (CORRUPT): Change in corruption (increase implies less corruption, or a lower risk). Includes bribes and corruption, but also excessive patronage, nepotism and secret party funding; ∆ (SOCIOECO): Change in the socioeconomic pressure dummy variable (increase is improvement in socioeconomic conditions, or a lower risk). Includes unemployment, consumer confidence and poverty; ELECLEGYR: Dummy variable that is 1 if there is an election year for the legislative power and 0 otherwise. Adjusted Pseudo Source: Own elaboration.
Variable
6 months run-up
24 months run-up
6 months run-up
24 months run-up
Coefficient
Coefficient
Variable
Coefficient
Coefficient
SF1
-0.660 (0.389)
-0.164 (0.078)
SF9
-0.943 (0.341)
-0.240 (0.123)
SF2
-0.990 (0.431)
-0.393 (0.106)
SF10
2.008 (0.598)
-0.524 (0.228)
SF3
1.651 (0.785)
0.249 (0.077)
∆ (LA WORD)
-1.067 (0.813)
SF4
0.822 (0.256)
0.347 (0.104)
∆ (CORRUPT)
-1.201 (0.365)
SF5
-0.510
-0.729 (0.117)
∆ (SOCIOECO)
1.226 (0.589)
SF6
-0.208
-0.197 (0.099)
ELECLEGYR
1.076 (0.395)
SF7
1.219 (0.447)
0.837 (0.142)
Adjusted pseudo
0.511
0.283
SF8
0.492 (0.566)
-0.411 (0.186)
The fit (in terms of adjusted pseudo
Notes: *: significant at 10%, **: significant at 5%, and ***: significant at 1%; Explanations of the symbols used: SF1: Static Factor 1, SF2: Static Factor 2, et cetera; ∆ (BURQUAL): Change in the bureaucratic quality dummy variable (increase is improvement in bureaucratic quality, or lower risk); ∆ (INVPROF): Change in investment profile dummy variable (increase is improvement in investment profile, or lower risk). Contains all risks that are not covered by other political, economic and financial risk components, such as contract viability, expropriation, profits repatriation and payment delays; ELECEXEYR: Dummy variable that is 1 if there is an election year for the executive power and 0 otherwise; ∆ (CORRUPT): Change in corruption (increase implies less corruption, or a lower risk); ∆ (LA WORD): Change in the law and order dummy variable (increase is improvement in law and order, or lower risk); HERFOPP: Herfindahl Opposition Index represents a measure of government opposition concentration, The presence of a majority party in the opposition increases the index. Having many (small) parties in the opposition reduces it. Adjusted Pseudo Source: Own elaboration.
Variable
6 months run-up
24 months run-up
6 months run-up
24 months run-up
Coefficient
Coefficient
Variable
Coefficient
Coefficient
SF1
0.174 (0.057)
0.150 (0.134)
∆ (BURQUAL)
-0.468 (0.349)
SF2
-0.107 (0.104)
-0.914 (0.136)
∆ (INVPROF)
-0.798 (0.561)
SF3
0.236 (0.080)
-0.214 (0.071)
ELECEXEYR
1.023 (0.520)
SF4
-0.182 (0.105)
-0.484 (0.087)
∆ (CORRUPT)
-0.638 (0.401)
SF5
-0.608 (0.217)
-0.833 (0.157)
∆ (LA WORD)
1.145 (0.439)
SF6
0.329 (0.133)
0.007 (0.115)
HERFOPP
1.098 (0.646)
SF7
0.217 (0.174)
-0.283 (0.162)
Adjusted pseudo
0.156
0.453
SF8
-0.096 (0.230)
0.009 (0.167)
We present the out-of-sample performance in the period 2008M1-2009M12 of the estimated EWS models with the 6 months and 24 months run-up.
For Argentina the model with a 6 months run-up period (see
Forecasts based on the 6 months run-up period (see
The 24 months run-up model (