Information about Western Hemisphere Hemisferio Occidental
Chapter

Chapter 3. Estimating Current Account Norms in Latin American Emerging Markets: A Quantile Regression Approach

Author(s):
R. Gelos
Published Date:
March 2014
Share
  • ShareShare
Information about Western Hemisphere Hemisferio Occidental
Show Summary Details
Author(s)
Roberto Perrelli

A key tool to assess the equilibrium exchange rate is based on the estimation of the savings-investment balance (current account norms). These norms are projected on a set of macroeconomic fundamentals and demographic characteristics for a panel of heterogeneous countries. In the traditional estimation methods, the contribution of each fundamental for the savings-investment balance is constant across countries. In this chapter, we propose an alternative approach—the quantile regression (QR)—that captures the unobserved heterogeneity over time and across countries, delivering coefficients that vary throughout the sample. The QR results suggest that the responses of the current account norms to changes in the fiscal and oil balances are much higher at the tail end of the conditional distribution than at the average.

A central issue in the analysis of international trade imbalances and external competitiveness is the assessment of the equilibrium exchange rate of trade partners. One of the main methodologies used to assess exchange rate misalignments is the macroeconomic balance framework developed by researchers at the International Monetary Fund.1 The macroeconomic balance approach evaluates whether prevailing exchange rates and policies are consistent with medium-term internal and external equilibrium.

The theoretical foundations of the macroeconomic balance approach to exchange rate assessment were developed in the initial years of the IMF,2 and the methodology has been at the core of IMF economists’ work ever since. For instance, one of its most prominent applications in the early years was the assessment of the degree of overvaluation of the pound sterling in 1967. Over the last six decades, the methodology has been continuously refined by IMF researchers; the set of policy variables has been expanded; and the sample of countries now covers 54, including advanced, newly industrialized, and emerging market economies. Extensions of the method to oil-exporting countries and to tourism-dependent economies have also emerged in recent years.

The macroeconomic balance approach assesses the degree of exchange rate misalignment by comparing the savings-investment norm (that is, the savings-investment balance that would prevail under equilibrium) with the projected current account balance over the medium term and under existing policies. The distance between the two is known as the current account gap, and provides a good sense of the direction of misalignment (for example, overvaluation or undervaluation) as well as the degree of exchange rate adjustment that would be necessary to bring the underlying current account balance closer to the savings-investment norm.

The precise magnitude of the exchange rate misalignment, usually taken with a grain of salt by IMF economists, is computed by dividing the current account gap by the elasticities of exports and imports to marginal changes in the real effective exchange rate. Higher ratios imply that larger exchange rate adjustments would be necessary to shift the underlying current account toward the savings-investment norm.

In this chapter, we focus on the estimation of the savings-investment norms in the macroeconomic balance approach, leaving the discussion on trade elasticities for a separate work (Perrelli, 2011). In the estimation of the savings-investment norms, researchers have been considering that each policy variable included in the model has a constant impact on the current account balance of all economies in the sample. As the macroeconomic balance approach has extended its sample to a large set of heterogeneous economies, these methods tend to neglect a great deal of heterogeneity not captured by the explanatory variables (henceforth simply called unobserved heterogeneity). In addition, these methods rely on stringent distributional assumptions (normality) of the residuals.

This chapter aims to supplement the existing methods of estimation of the savings-investment norms with the quantile regression approach. Quantile regressions offer an alternative view of the impact of policy variables on exchange rate assessments of heterogeneous countries. The quantile regression approach provides a flexible framework where normality and identically independently distributed errors are not required. In the quantile regression approach, the model coefficients (slopes) can vary according to the quantiles (for example, percentiles) of the conditional distribution of the dependent variable, capturing the unobserved heterogeneity neglected by traditional models, and providing a more complete assessment of the role of the policy variables in the determination of the savings-investment norms. The approach may be quite useful for countries at the tail end of the conditional distribution of current account balances—for example, those with larger deficits or larger surpluses—where the unobserved heterogeneity is expected to be more accentuated than around the mean.

The Macroeconomic Balance Approach to Assessing the Equilibrium Exchange Rate

The macroeconomic balance approach to assessing the equilibrium exchange rate aims to evaluate whether prevailing exchange rates and policies are consistent with medium-term internal and external equilibrium. The approach is based on seminal work of Nurkse (1945), Meade (1951), and Metzler (1951). In early applications of the method, the exchange rate was assessed by comparing the underlying current account with the net capital flows (Isard and Mussa, 1998). By the mid-1990s, when the IMF established the Consultative Group on Exchange Rate Issues (CGER) to coordinate analytical work on the topic, the macroeconomic balance framework involved a sample of 22 advanced economies.3 Owing to limited access to capital markets, emerging market economies were not included in that sample. In the last decade, the approach was adapted to include countries with extended access to international capital markets and enhanced data standards, enlarging the sample from 22 countries to 54 countries. Recent advances to the methodology include adaptation of the model to oil-exporting countries (Bems and de Carvalho Filho, 2009) and to tourism-dependent small open economies (Piñeda, Cashin, and Sun, 2009).

The approach is based on the well-known national accounts identity that the savings-investment (S-I) balance equals net exports (that is, the external current account balance, CAB). It relies on an adjustment process function where movements in the real effective exchange rate (REER) of a given country would be gradually reflected in its CAB. All else being equal, an upward change in the country’s REER would be associated with a loss of international competitiveness and therefore with a deterioration of its external current account position. The opposite is also true (see Figure 3.1).

Figure 3.1REER, Savings-Investment Norm, and the Adjustment Process Function

Source: Author’s calculations.

A lynchpin of the methodology is the accurate assessment of the savings-investment norm, that is, the savings-investment balance that would prevail if the REER were at its equilibrium level (solid vertical line in Figure 3.1). Once that is obtained, one is able to identify the distance between the underlying CAB and the S-I norm. This distance, often called the current account gap, determines the direction of the exchange rate misalignment and helps to quantify the amount of exchange rate adjustment necessary to bring the underlying current account balance closer to the savings-investment norm.

The size of the exchange rate adjustment is assessed in three steps: (1) estimation of the S-I norm and the current account gap; (2) estimation of the elasticities of exports and imports to marginal changes in the REER; and (3) calculation of the ratio of the current account gap and the elasticities. The estimated results are used as an indicator of the degree of exchange rate misalignment. In this chapter, we concentrate on the first step of the macroeconomic balance approach, leaving the study of import and export elasticities for future work.

Note that the speed and path of the exchange rate adjustment toward a level consistent with its equilibrium are not yielded by the macroeconomic balance approach; neither do these aspects have a prominent role in the exercise. Instead, the approach provides a measurement of how distant the underlying current account balance is from the savings-investment norm (that is, the savings-investment balance at the equilibrium exchange rate), and by how much the real effective exchange rate would have to adjust to bring them together.

Determinants of the Savings-Investment Norm

The savings-investment norm is determined empirically by a set of macroeconomic fundamentals, demographic aspects, and country individual characteristics that are viewed as having a substantial impact on the long-term patterns of internal and external savings across countries. Although the REER is clearly associated with movements in the underlying current account, by design the S-I norm is orthogonal to changes in the REER (thus the vertical line in Figure 3.1 above), and serves as a benchmark to gauge signals of exchange rate misalignments.

The econometric estimation of the S-I norm involves the following set of variables (as described in Lee and others, 2008):

  • Fiscal balance: the ratio of the general government balance to GDP.

  • Real growth of purchasing power parity-GDP per capita: included only for emerging market economies.

  • Old-age dependency ratio: the ratio of the population above 65 years to the population between 30 and 64 years.

  • Population growth: the annual population growth rate of each economy.

  • PPP-GDP per capita: a proxy for relative income, measured as the ratio of per capita purchasing power parity (PPP) income to the U.S. level, both in constant 2000 international U.S. dollars.

  • Oil balance: the ratio of oil balance to GDP

  • Asian crisis: dummy variable for Asian emerging markets for the 1997–2004 period, namely China, Hong Kong SAR, India, Indonesia, Korea, Malaysia, the Philippines, Singapore, Taiwan Province of China, and Thailand.

  • Banking crisis: dummy variable for banking crisis episodes. Obtained from Demirgüç-Kunt and Detragiache (2005) and Gruber and Kamin (2005).

  • Financial center: dummy variable indicating if the country is a financial center.

  • Initial NFA: the ratio of net foreign assets (NFA) to GDP prevailing at the beginning of each four-year period. It uses the NFA data from Lane and Milesi-Ferretti (2007).

We note that four of these variables enter the econometric model as deviations from the averages for trading partners: (1) fiscal balance; (2) old-age dependency ratio; (3) population growth; and (4) real growth of PPP-GDP per capita. The justification is that the S-I norm of a given country would only be affected by these variables if they changed relative to its trading partners. The remaining variables enter the model as described above.

Also, the inclusion of the fiscal balance as a determinant of the savings-investment norm relies on the assumption that the hypothesis of Ricardian equivalence does not hold in the sample (Isard and Mussa, 1998). Simply put, if a country promoted a policy adjustment that reduced its fiscal deficit, the private savings in that country could decline but not enough to fully offset the improvement in public savings. Therefore, the fiscal adjustment would have a positive impact on the savings profile of the country, which would be reflected in a movement to the right of the S-I norm curve (Figure 3.2).

Figure 3.2Non-Ricardian Effect of Fiscal Adjustment on Savings-Investment Norm

Source: Author’s calculations.

The intuition behind the other variables is more straightforward. Higher growth rates of real PPP-GDP per capita (relative to trading partners), on the one hand, are associated with faster economic growth and elevated investment needs, which contribute to a deterioration of the S-I norm. On the other hand, a higher level of PPP-GDP per capita (relative to the U.S. standards) suggests a more advanced stage of economic development, normally characterized by slower economic growth and lower investment needs, thus positively associated with higher S-I norms.

A higher old-age dependency ratio and population growth (relative to trading partners) are expected to decrease national savings and thus to have a negative impact on the S-I norm. Higher oil balances, as displayed by oil-exporting countries, are positively associated with higher S-I norms. Currency and banking crises are positively associated with S-I norms owing to the forced external adjustment, sudden stop, and lack of external financing faced by the crisis countries. Being a financial center is a proxy for net creditor positions and is thus related to higher S-I norms.4 And countries with a higher initial NFA position have lower external indebtedness and tend to present higher current account balances.

One of the models also includes the lagged dependent variable (that is, the lagged current account balance in percent of GDP) as an explanatory variable in the regression. This variable is expected to have a positive sign due to the persistence of the current account balances in the sample.

Data and Econometric Models

The latest edition of the macroeconomic balance approach, as described in Lee and others (2008), uses a sample of the following 54 countries, including advanced, newly industrialized, and emerging market economies: Algeria, Argentina, Australia, Austria, Belgium, Brazil, Canada, Chile, China, Colombia, Croatia, the Czech Republic, Denmark, Egypt, Finland, France, Germany, Greece, Hong Kong SAR, Hungary, India, Indonesia, Ireland, Israel, Italy, Korea, Japan, Luxembourg, Malaysia, Mexico, Morocco, the Netherlands, New Zealand, Norway, Pakistan, Peru, the Philippines, Poland, Portugal, Russia, Singapore, the Slovak Republic, Slovenia, South Africa, Spain, Sweden, Switzerland, Taiwan Province of China, Thailand, Tunisia, Turkey, the United Kingdom, the United States, and Venezuela.

The data cover the period from 1973 to 2004, and are structured in eight blocks of four-year averages: 1973–76; 1977–80; 1981–84; 1985–88; 1989–92; 1993–96; 1997–2000; and 2001–04. The estimation methods used in the macroeconomic balance approach are the pooled ordinary least squares (OLS—where the sample is treated as a cross section of 432 observations), and the fixed-effects regression (where the sample is treated as a panel with 54 cross sections and eight time series).5

The savings-investment norms are obtained according to three econometric models:

  • Pooled OLS with lagged current account balance: in this model, besides the variables listed above, Lee and others (2008) mention the inclusion of country-specific dummy variables for six mid-sized economies and four Asian crisis countries, with their need justified by statistical tests.6

  • Pooled OLS with initial net foreign assets: the same variables as above are applied, except that the lagged current account balance is replaced by the initial net foreign assets position.

  • Fixed effects: owing to the standard cross-section dummies, other time-fixed indicators (for example, being a financial center) are naturally dropped. The model also excludes variables that present a high correlation with the cross-section dummies, such as banking crisis and PPP-GDP per capita (relative to the U.S. standards). Finally, the model excludes the initial NFA and the lagged current account balance from the list of explanatory variables.

Results of these models are discussed in detail below (see the section on Results), along with the estimates provided by the quantile regression models.

The Quantile Regression Framework

Mainstream methods used to assess the equilibrium exchange rate estimate the impact of macroeconomic fundamentals on the savings-investment balance (current account norm) of a representative country (for example, the “average” country), without appropriately assessing the role of fundamentals at the tail end of the conditional distribution of the dependent variable. By delivering coefficients that are constant throughout the sample, these models restrict the role of explanatory variables to a mere “location-shift” of the conditional distribution of the dependent variable, neglecting important features of the data. In other words, these methods assume that the explanatory variables would change the mean but not other aspects of the conditional distribution of the dependent variable. Given the diversity of countries we deal with in international economics, it is important to evaluate the effects of the covariates not only on the average response but also on the outstanding ones. The quantile regression approach was introduced by Koenker and Bassett (1978a) as a supplement to traditional econometric techniques. Rather than focusing on the conditional mean, the QR approach offers a characterization of the entire conditional distribution of the dependent variable. It is based on the principle of ranking optimization, where the dependent variable (response), given a set of explanatory variables (covariates), is ranked according to the quantile where it is located in the conditional distribution. Given its rank nature, the QR estimates of the regression coefficients are robust to outliers in the dependent variable.

The QR regression does not require that the error term follow a normal distribution (or any parametric profile) because it is more efficient than traditional least squares estimators in the presence of non-normal residuals.

Econometric Setup

Let Y be a random variable with distribution function F(y) = Pr(Yy). For τ ϵ (0,1), the τ-th quantile is defined as Q(τ) = inf{y: F(y) ≥ τ}. Simply put, an observation situated in the τ-th sample quantile is greater than τ percent of the observations. Well-known unconditional quantiles include the median (τ = 0.5) and the first and third quartiles (respectively, τ = 0.25 and τ = 0.75).

Next, consider a basic linear (in parameters) regression model:

for i = 1, …, n, where β is a K × 1 vector of coefficients, X is a n × K matrix of explanatory variables, y is the response variable, and u is the n × 1 vector of residuals. In the present application, y would represent the current account norm, and X would be the matrix of macroeconomic fundamentals, demographic factors, and country-specific dummies. In traditional estimation methods, such as the OLS, the vector of coefficient β^ is found through the minimization of the sum of the squared residuals. In contrast, the vector of coefficients in the QR setup is obtained by the solution of the following asymmetric penalty function:

The asymmetric penalty function above is equivalent to the weighted sum of the absolute values of the residuals, where the weights vary according to each quantile (τ). The QR coefficients β^(τ) are the solution to Equation (3.1) and naturally depend on each x of the conditional distribution of the dependent variable F(y|X)=Pr(Yy|X).

A special case of Equation (3.1) is the least absolute deviation (LAD) estimator, where τ = .5 and the penalty function becomes symmetric, collapsing simply to:

In this special case, the QR model fits the median response on a set of covariates (the asymmetric distribution of the LAD estimator is discussed in Koenker and Bassett, 1978b).

Interpretation of the Quantile Regression Slopes

The following empirical application is used in this chapter: we estimate regression coefficients for a grid of quantiles spanning from the 10th to the 90th percentile (by increments of 1 percent) of the conditional distribution of the explanatory variable (the current account norms). Although the OLS regression yields a K × 1 vector of estimated coefficients (slopes) β^, the QR provides a matrix of coefficients β^(τ) with dimension K × (τ), where (τ) is the length of the grid of quantiles selected to be estimated by the model (in the present application, the length of this grid is (τ) = 90 − 10 + 1 = 81).

Just as the OLS coefficient β^k measures the impact of marginal changes in xk on the mean of the conditional distribution of y, the QR coefficient β^k(τ) captures the impact of marginal changes in xk on the τ-th quantile of the conditional distribution of y. As τ spans the entire sample, β^k(τ) captures the impact of xk at each sample quantile, providing a much richer description of the relationship between xk and y. Conversely, the OLS coefficient is constant throughout the sample.

In the empirical application below, we display the estimates of the QR coefficients, and the respective confidence intervals, for a long range of quantiles of the conditional distribution of y. To compare, we plot the OLS coefficients as horizontal lines (fixed values) for all quantiles in the sample.

Unobserved Heterogeneity and Interquartile Differences in Slopes

The unobserved heterogeneity neglected by the traditional regression models can be captured by the QR coefficients at each quantile of the conditional distribution. Latent factors are embedded in the quantile regression coefficients, in the sense that a given explanatory variable is allowed to interact with the unobserved heterogeneity and present different effects throughout the sample.

Some researchers find it useful to assess the statistical significance of the changes in the slopes β^k(τ) over the grid of quantiles. In this regard, a common test in the QR framework is the estimation of the degree of variability of β^k(τ) between the first and the third quartiles. Rejecting the null hypothesis—that β^k(τ=.75)β^k(τ=.25)=0—is a good indicator that the variability of the slopes is significant. The test can be extended to other quantiles.

Results

In this section, we supplement the econometric results obtained from the mainstream methods currently used in the macroeconomic balance approach (see Lee and others, 2008) with the quantile regression estimates.

The exchange rate assessment conducted by the macroeconomic balance approach is based on three models of the savings-investment norm: (1) a pooled OLS with a lagged current account balance among the explanatory variables; (2) a pooled OLS with the initial net foreign assets replacing the lagged current account balance as an explanatory variable; and (3) a fixed-effects panel regression, which includes cross-section-specific dummies but drops the lagged current account balance, the initial NFA, and time-fixed dummies such as the financial center indicator.

Overall, the three models have high goodness of fit (R-squares up to 64 percent). In addition, in the three models, the most prominent policy variables have the expected sign and are statistically significant. Nevertheless, compared with the actual data (that is, current account balances), the models deliver fitted values that have density distributions more concentrated around the sample mean. As seen in Figure 3.3, almost 20 percent of the fitted results lie around the sample mean (a deficit close to 1 percent for all models), but the actual data have half of that mass around that region.

Figure 3.3Conditional Distribution of Current Account Norms for CGER Countries

Source: Author’s calculations.

Note: CGER = Consultative Group on Exchange Rate Issues; POLS = pooled ordinary least squares. See text for discussion of models.

The compression of the conditional distributions of the fitted values toward the mean is also reflected in the standard deviations and in other relevant rank measures, as can be observed in Table 3.1. For instance, the standard deviation of the actual current account balances is 30 percent higher than those of the current account balances fitted by mainstream methods. The interquartile range, measured as the difference between the lower and upper quartiles—a robust metric for the spread in the central region of the distribution—is also about 30 percent higher in the actual data. To better understand the reasons for this compression, we take a closer look at the coefficients estimated by each model, and compare them with quantile regression estimates.

Table 3.1Summary Statistics of the Fitted Values of the Savings-Investment Norms1
Actual DataModel 1Model 2Model 3
(Percent of GDP)
Mean−0.9−0.9−0.9−0.9
Standard deviation4.73.63.33.6
Minimum−26.5−8.8−7.7−8.5
Lower quartile−3.5−3.1−2.9−3.3
Median−1.3−1.3−1.5−1.5
Upper quartile1.20.70.30.6
Maximum22.520.420.820.6
Source: Author’s calculations.

Model 1 is the pooled ordinary least squares (OLS) with lagged current account balance; Model 2 is the pooled OLS with the initial net foreign assets; and Model 3 is the fixed effects panel regression.

Source: Author’s calculations.

Model 1 is the pooled ordinary least squares (OLS) with lagged current account balance; Model 2 is the pooled OLS with the initial net foreign assets; and Model 3 is the fixed effects panel regression.

In Table 3.2, we display the results for the pooled OLS model with the lagged dependent variable (Model 1). In addition, we provide quantile regression estimates of that specification at five selected percentiles of the conditional distribution of the current account balances: the 10th (lowest decile), the 25th (lower quartile), the 50th (median), the 75th (upper quartile), and the 90th (highest decile).

Table 3.2Current Account Norm—Estimates Including Lagged Current Account Balance
Dependent Variable: Current Account Balance
Explanatory Variables:POLSQuantile Regression
tau = 0.1tau = 0.25tau = 0.5tau = 0.75tau = 0.9
Fiscal balance0.1890.2190.1600.0680.1440.217
0.043***0.084***0.078**0.0720.061**0.084***
Real growth of PPP-GDP per capita−0.157−0.027−0.166−0.156−0.180−0.177
0.079**0.164**0.1100.0980.1490.210
Old-age dependency ratio−0.125−0.113−0.175−0.075−0.020−0.181
0.048***0.1170.062***0.0530.0630.107*
Population growth−1.034−1.093−1.159−0.600−0.468−1.100
0.405***0.616**0.559**0.4470.5100.672
PPP-GDP per capita0.0200.0380.0260.0170.0020.007
0.011*0.018**0.012**0.0120.0120.024
Oil balance0.1690.0880.1340.2000.2350.208
0.030***0.0820.081*0.058***0.054***0.067***
Asian crisis0.0340.0440.0350.0330.0420.013
0.006***0.011***0.011***0.012***0.016**0.017
Banking crisis0.0100.0140.0080.0060.0140.028
0.0070.0130.0100.0080.0120.016*
Lagged current account balance0.3660.3550.4550.4370.3990.376
0.089**0.070***0.078***0.073***0.085***0.110***
Financial center0.0310.0100.0240.0320.0380.055
0.004***0.0090.009**0.008***0.011***0.016***
Constant−0.003−0.027−0.016−0.0020.0070.021
0.0040.006***0.005***0.0050.0050.011*
R-squared0.643n.a.n.a.n.a.n.a.n.a.
Source: Author’s calculations.Note: POLS = pooled ordinary least squares; PPP = purchasing power parity. (***) significant at 1 percent; (**) significant at 5 percent; and (*) significant at 10 percent.
Source: Author’s calculations.Note: POLS = pooled ordinary least squares; PPP = purchasing power parity. (***) significant at 1 percent; (**) significant at 5 percent; and (*) significant at 10 percent.

Countries with current account balances situated below the lower quartile correspond to cases with larger current account deficits. Conversely, countries with current account balances above the upper quartile correspond to cases with the highest current account surpluses. More interestingly, countries may switch their rank over time—for example, a fast-growing economy with current account deficits below the lower quartile in the 1980s could be enjoying surpluses and be placed at the top quartile in the 2000s. Nevertheless, in light of the gradual adjustment (high persistence) of current account balances, deficit-prone countries are expected to remain associated with lower quantiles (whereas surplus economies are expected to remain at the upper quantiles) of the conditional distribution of the response.

The results indicate that, although in most cases, the coefficients present the expected sign at these quantiles, their magnitude and statistical significance change substantially throughout the quantiles. For instance, the coefficient for the fiscal balance has a higher magnitude for the countries located at the tail end of the conditional distribution rather than at the center.

A more insightful view is provided in Figure 3.4, where we plot the quantile regression coefficients for each of the policy variables on a grid of quantiles that spans from the 10th to the 90th percentile, in increments of 1 percent. The point estimates of β^(τ) are represented by the solid black line in each panel chart. Their respective confidence intervals, based on bootstrapped standard errors, are represented by the shaded gray area. For comparison, the estimated coefficients from the pooled OLS model are plotted as solid red lines (and the boundaries of their respective confidence intervals are plotted as dashed red lines). As the pooled OLS coefficients are constant for the entire conditional distribution of the dependent variable, these lines are horizontal.

Figure 3.4Quantile Regression Estimates for Pooled Ordinary Least Squares Model with Lagged Current Account Balance

Source: Author’s calculations.

Note: For simplicity, the coefficients for country and time dummies are not plotted above.

The results suggest that coefficients of the fiscal balance are almost three times higher at the tail end of the distribution (where the external imbalances reside) than in the center, with the β^(τ) function U-shaped. The lagged current account balance has a somewhat stable contribution to the dependent variable, explaining between 35 percent and 45 percent of its variation—an elevated degree of persistence for all quantiles.

In Table 3.3 and Figure 3.5, we present the quantile regression coefficients for the econometric specification of the pooled OLS model with the initial net foreign assets (Model 2).

Table 3.3Current Account Norm—Estimates Including Initial Net Foreign Assets Position
Dependent Variable: Current Account Balance
Explanatory Variables:POLSQuantile Regression
tau = 0.1tau = 0.25tau = 0.5tau = 0.75tau = 0.9
Fiscal balance0.1980.2300.1590.0820.1450.226
0.056***0.083***0.090*0.0740.073**0.080***
Real growth of PPP-GDP per capita−0.208−0.343−0.326−0.236−0.228−0.168
0.095**0.174**0.118***0.116**0.117**0.209
Old-age dependency ratio−0.142−0.094−0.169−0.178−0.087−0.073
0.062**0.0990.065**0.059***0.051*0.086
Population growth−1.213−1.230−1.454−1.329−0.577−0.831
0.472***0.529**0.634**0.504***0.5040.552
PPP-GDP per capita0.0240.0330.0230.0170.0110.009
0.0150.019*0.012*0.0140.0130.024
Oil balance0.2310.1900.2100.1870.2960.310
0.035***0.073***0.064***0.053***0.060***0.060***
Asian crisis0.0610.0990.0670.0440.0470.038
0.007**0.041**0.031**0.026*0.023**0.023*
Banking crisis0.0080.0030.0080.0060.0190.013
0.0070.0170.0110.0090.010*0.010
Net foreign assets0.0250.0180.0300.0360.0340.014
0.010**0.0140.011***0.008***0.010***0.016
Financial center0.026−0.0050.0220.0320.0310.066
0.008***0.0230.0150.009***0.009***0.019***
Constant−0.001−0.032−0.020−0.0030.0130.023
0.0060.008***0.006***0.0050.006**0.010**
R-squared0.553n.a.n.a.n.a.n.a.n.a.
Source: Author’s calculations.Note: POLS = pooled ordinary least squares; PPP = purchasing power parity. (***) significant at 1 percent; (**) significant at 5 percent; and (*) significant at 10 percent.
Source: Author’s calculations.Note: POLS = pooled ordinary least squares; PPP = purchasing power parity. (***) significant at 1 percent; (**) significant at 5 percent; and (*) significant at 10 percent.

Figure 3.5Quantile Regression Estimates for Pooled Ordinary Least Squares Model with Initial Net Foreign Assets (NFA) Position

Source: Author’s calculations.

Note: For simplicity, the coefficients for country and time dummies are not plotted above.

The convex shape of the fiscal balance impact is observed again. Also, the contribution of the oil balances to the current account—which in Model 1 was an upward sloped curve with the quantiles—is now quite flat for most of the sample, with a peak in the upper quartile. This suggests that there is a degree of interaction between oil balance and initial net foreign balance that was not captured in Model 1.

The contribution of the initial net foreign assets position tends to be smaller at the tail end of the conditional distribution, implying that for high deficit and high surplus countries, the initial NFA position is less important for determination of external balances than for the rest of the sample. In fact, the coefficients for the initial NFA position are only statistically different from zero along the interquartile range.

Also, the contribution of the output growth (real growth of PPP-GDP per capita) is close to the pooled OLS estimates for most percentiles. However, the contribution of the relative income (PPP-GDP per capita relative to the U.S. standard) declines substantially with the quantiles in both models, even crossing the zero boundary (that is, changing sign) at the upper quantiles of the first model.

With regard to the behavior of countries’ demographic factors, both the old-age dependency ratio and the population growth tend to oscillate within the boundaries of the confidence interval delimited by the pooled OLS method. However, in Model 2, they tend to have a smaller impact on the surplus countries than on the rest of the sample.

Finally, the financial center dummy presents a higher impact in the upper quantiles, supporting the theoretical view that financial centers tend to run current account surpluses (albeit in 17 percent of the cases, a financial center presented a current account deficit).

In Table 3.4 and Figure 3.6, we present the coefficients for the econometric specification of the fixed-effects panel regression (Model 3). Coefficients are identical to those obtained from a pooled OLS regression with a dummy variable for each cross section. Hence, the quantile regression framework is applied in a similar way as in Models 1 and 2.

Table 3.4Current Account Norm—Estimates Based on Fixed-Effects Panel Regression
Dependent Variable: Current Account Balance
Explanatory Variables:POLSQuantile Regression
tau = 0.1tau = 0.25tau = 0.5tau = 0.75tau = 0.9
Fiscal balance0.3140.2940.2570.2350.1880.285
0.087***0.097***0.093***0.090***0.086**0.094***
Real growth of PPP-GDP per capita−0.268−0.199−0.183−0.183−0.112−0.139
0.1980.1430.1290.1340.1380.171
Old-age dependency ratio−0.234−0.308−0.231−0.168−0.259−0.209
0.128*0.143**0.125*0.1110.152*0.181
Population growth−0.470−1.568−1.941−1.457−0.931−1.283
1.1530.849*0.946**0.702**0.7380.965
Oil balance0.306−0.0010.2340.2500.3540.341
0.096***0.1340.123**0.110**0.101***0.123***
Asian crisis0.0670.0940.0770.0610.0520.043
0.007***0.043**0.040*0.035*0.031*0.029
Constant−0.002−0.040−0.019−0.0040.0100.013
0.0090.013***0.0130.0110.0110.013
R-squared0.633n.a.n.a.n.a.n.a.n.a.
Source: Author’s calculations.Note: POLS = pooled ordinary least squares; PPP = purchasing power parity. (***) significant at 1 percent; (**) significant at 5 percent; and (*) significant at 10 percent.
Source: Author’s calculations.Note: POLS = pooled ordinary least squares; PPP = purchasing power parity. (***) significant at 1 percent; (**) significant at 5 percent; and (*) significant at 10 percent.

Figure 3.6Quantile Regression Coefficients for Fixed-Effects Panel Regression

Source: Author’s calculations.

Note: For simplicity, the coefficients for country and time dummies are not plotted above.

The results of the fixed-effects model are more streamlined, as the model forcefully drops time-fixed dummy variables (such as the financial center indicator) owing to perfect correlation with the cross-section dummies. In addition, explanatory variables that present high (but not perfect) correlation with the cross-section dummies, are not part of the original macroeconomic balance approach (fixed-effects panel regression).

The quantile regression estimates for the fixed-effects model confirm the convex shape of the fiscal balance contribution to the current account, as well as the higher impact of the oil balances on the current accounts of countries that present larger surpluses.

Interestingly, the quantile regression estimates of the impact of output growth on the current account balances are systematically lower (in absolute values) than as estimated by the fixed-effects model. The opposite happens to the contribution of population growth.

Finally, the role of the old-age dependency ratio is quantitatively similar in both approaches.

An Application to Latin American Emerging Markets

A useful application of the quantile regression framework regards the estimation of current account norms for Latin American emerging markets. To illustrate the heterogeneity among these countries, as well as in comparison with non–Latin American economies included in the CGER analysis, in Figure 3.7, we plot the percentiles of each country’s current account balance at each point in time. Large current account deficits are associated with lower percentiles; the opposite is true for sizable surpluses. These percentiles also vary over time, such that current account reversals are well captured by these statistics.

Figure 3.7Latin America: Percentiles of Current Account Balances, 1985–2004

Source: IMF staff calculations.

Note: CGER = Consultative Group on Exchange Rate Issues.

In Figure 3.7, panel A, we see that the current account balances of Brazil, Chile, Colombia, and Mexico very often lie between the 25th and the 75th percentiles, known as the interquartile range (shaded area). In contrast, the current account balances of Argentina, Peru, and Venezuela are usually located far away from the center of the distribution, as illustrated in Figure 3.7, panel B. This phenomenon is related to both cyclical and structural features of these countries during each sub-period of the sample. Such degree of differentiation sometimes is not fully captured by the CGER panel data regressions.

Econometric models that neglect heterogeneity over time and across countries may yield estimates of current account norms that tend to be much closer to the sample average than warranted by a country-year’s specific characteristics. In these cases, supplementing the CGER models with the quantile regression approach helps to uncover the neglected heterogeneity. To illustrate this point, in Table 3.5, we list five Latin American episodes of current account balances outside the interquartile range. In all of them, the current account norm based on the quantile regression model fitted for the specific percentile of each observation (“quantile norm”) was associated with narrower current account gaps (that is, smaller differences between the actual current account balance and the estimated norm) than the respective estimates of the CGER norm. All else being equal, the quantile regression gaps suggest a lower degree of exchange rate misalignment in Latin American economies than what would be expected from the examination of the CGER gaps during these episodes.

Table 3.5Latin America: Selected Episodes of Current Account Balances “Outlying” the Median, 1985–2004
PeriodEpisode PercentileCurrent Account

Balance
CGER NormQuantile NormCGER GapQuantile Gap
(Percent of GDP)
Brazil1997–20000.19−4.2−1.5−3.7−2.8−0.5
Colombia1993–960.17−4.4−0.3−3.1−4.2−1.4
Mexico1993–960.24−3.5−0.4−2.4−3.1−1.2
Peru1989–920.21−4.0−1.6−4.0−2.40.0
Venezuela1989–920.914.81.97.02.9−2.2
Source: Author’s calculations.Note: CGER = Consultative Group on Exchange Rate Issues. The CGER norm corresponds to the fitted value according to the pooled ordinary least squares model described in Table 3.2, whereas the quantile norm refers to the fitted value for the respective percentile of the episode, based on the quantile regression specification of Table 3.2. The gap refers to the difference between the actual current account balance and the fitted current account norm acording to each method.
Source: Author’s calculations.Note: CGER = Consultative Group on Exchange Rate Issues. The CGER norm corresponds to the fitted value according to the pooled ordinary least squares model described in Table 3.2, whereas the quantile norm refers to the fitted value for the respective percentile of the episode, based on the quantile regression specification of Table 3.2. The gap refers to the difference between the actual current account balance and the fitted current account norm acording to each method.

Conclusion

In this chapter, we provided an alternative framework to estimate the savings-investment norms used in the macroeconomic balance approach to assess exchange rate misalignments. The quantile regression approach provides a flexible framework to evaluate the contribution of policy variables and demographic characteristics on the determination of current account balances.

Mainstream methods used to estimate the savings-investment norms tend to disregard important features of the data, yielding fitted values that are compressed toward the mean. In these models, the contribution of each explanatory variable is fixed for all quantiles of the conditional distribution of the current account balances. These models show that the impact of any given policy variable is the same for countries carrying large deficits (that is, countries in the lower quantiles) as for countries enjoying large surpluses (those in the higher quantiles of the conditional distribution of current account balances).

The quantile regression estimates provided in this chapter supplement those from traditional models by providing slopes that vary according to the quantile of the conditional distribution of the current account balances. Through the use of the quantile-varying slopes, the econometrician provides a much richer description of the relationships between current account balances and macroeconomic fundamentals in a large panel of heterogeneous countries.

Our results indicate that the contribution of fiscal balances to the current account tends to be higher at the tail end of the conditional distribution of current account balances (that is, where countries with the largest external imbalances are located). In addition, we found that the contribution of oil balances tends to be more accentuated in the upper quantiles. Important nonlinearities in the role of output growth and demographic characteristics were also discussed.

In an empirical application to Latin American emerging markets, we find that during periods of sizable deficits or surpluses, the current account norms based on the quantile regression approach are broadly associated with narrower current account gaps (less exchange rate misalignment) than suggested by the CGER macroeconomic balance approach. This finding may reflect the ability of the quantile regression framework to capture unobserved heterogeneity over time and across countries.

References

    BemsRudolfs and Irineude Carvalho Filho2009Exchange Rate Assessments: Methodologies for Oil Exporting Countries,IMF Working Paper 09/281 (Washington: International Monetary Fund).

    ChinnMenzie and EswarPrasad2003Medium-Term Determinants of Current Accounts in Industrial and Developing Countries: An Empirical Exploration,Journal of International Economics Vol. 59 No. 1 pp. 4776.

    Demirgüç-KuntAsli and EnricaDetragiache2005Cross-Country Empirical Studies of Systemic Bank Distress: A Survey,IMF Working Paper 05/96 (Washington: International Monetary Fund).

    GruberJoseph and SteveKamin2005Explaining the Global Pattern of Current Account Imbalances,International Finance Discussion Paper No. 846 (Washington: Board of Governors of the Federal Reserve System).

    IsardPeter and HamidFaruqeeeds. 1998Exchange Rate Assessment: Extensions of the Macroeconomic Balance ApproachOccasional Paper 167 (Washington: International Monetary Fund).

    IsardPeter and MichaelMussa1998A Methodology for Exchange Rate Assessment,” in Exchange Rate Assessment: Extensions of the Macroeconomic Balance Approach edited by PeterIsard and HamidFaruqeeOccasional Paper 167 (Washington: International Monetary Fund) chapter 2.

    KoenkerRoger2005Quantile Regression (Econometric Society Monographs) (CambridgeMassachusetts: Cambridge University Press).

    KoenkerRoger and GilbertBassett1978aRegression Quantiles,Econometrica Vol. 46 pp. 3350.

    KoenkerRoger and GilbertBassett1978bThe Asymptotic Distribution of the Least Absolute Error Estimator,Journal of the American Statistical Association Vol. 73 pp. 61822.

    KoenkerRoger and KevinHallock2001Quantile Regression,Journal of Economic Perspectives Vol. 15 pp. 14356.

    LanePhilip and Gian MariaMilesi-Ferretti2007The External Wealth of Nations Mark II: Revised and Extended Estimates of Foreign Assets and Liabilities, 1970–2004,Journal of International Economics Vol. 73 No. 2 pp. 22350.

    LeeJaewooGian MariaMilesi-FerrettiJonathanOstryAlessandroPrati and LucaAntonio Ricci2008Exchange Rate Assessments: CGER Methodologies Occasional Paper 261 (Washington: International Monetary Fund).

    MeadeJames E.1951The Theory of International Economic Policy Volume One: The Balance of Payments (London: Oxford University Press).

    MetzlerLoyd A.1951Wealth, Saving, and the Rate of Interest,Journal of Political Economy Vol. 59 pp. 93116.

    NurkseRagnar1945Conditions of International Monetary Equilibrium,Princeton Essays in International Finance No. 4 (Princeton, New Jersey: Princeton University).

    PerrelliRoberto A.2011Extending the Macroeconomic Balance Assessment of the Equilibrium Exchange Rate—A Quantile Regression Approach,” unpublished (Washington: International Monetary Fund).

    PiñedaEmilioPaulCashin and YanSun2009Assessing Exchange Rate Competitiveness in the Eastern Caribbean Currency Union,IMF Working Paper 09/78 (Washington: International Monetary Fund).

    PolakJacques1995Fifty Years of Exchange Rate Research and Policy at the International Monetary Fund,IMF Staff Papers Vol. 42 pp. 73461.

    TaylorAlan2001A Century of Current Account Dynamics,Journal of International Money and Finance Vol. 21 pp. 72548.

For a broader discussion, see Polak (1995), Isard and Faruqee (1998), and Lee and others (2008).

Namely Australia, Austria, Belgium, Canada, Denmark, Finland, France, Germany, Greece, Ireland, Italy, Japan, Luxembourg, Netherlands, New Zealand, Norway, Portugal, Spain, Sweden, Switzerland, the United Kingdom, and the United States.

In the sample, six economies are considered financial centers: Belgium, Hong Kong SAR, Luxembourg, the Netherlands, Singapore, and Switzerland.

The data are available on request.

The dummies for mid-sized economies are for Australia, Chile, Israel, New Zealand, Sweden, and Thailand. The dummies for the Asian crisis economies are for Korea, Malaysia, Singapore, and Thailand.

    Other Resources Citing This Publication