ABM Nasira, Ashrafuddin Chowdhuryb, and Hosne Mridhac

1

Nasir,Chowdhury and Mridha, International Journal of Applied Economics, 6(2), September 2009, 51-69

Do Trade Balances Adjust to Exchange Rate Shocks? Evidence fromBilateral Trade of Bangladeshwith G-5Countries

ABM Nasira, Ashrafuddin Chowdhuryb, and Hosne Mridhac

North Carolina Central Universitya, University of Dhakab,North CarolinaCentral Universityc

Abstract This study investigates adjustment process in the bilateral trade balances of Bangladesh, a developing country,withfive of her largest sources of trade surpluses, namely the United States, the United Kingdom, France, Italyand Germany. Analysis of the quarterly data using autoregressive distributed lag (ARDL) approach to cointegration revealsonly limited adjustments in the trade balance to exchange rate shocksboth in the short- and in the long-run. The plots of generalized impulse response functions are unable to detect any persistent J-curve pattern in the trade balances. The results are robust to the application of panel cointegration techniques and fixed effect estimation.

Keywords: Trade balance, devaluation,auto-regressive distributed lag modeling, generalized impulse response functions, panel cointegration tests, fixed effect

JEL Classification: F31; F32; C32

1. Introduction

How trade balance responds to exchange rate shock has been a long debated issue in the international economics literature. Earlier attempts to resolve this debate had led to the development of the elasticity approach (Robinson, 1947 and Meltzer, 1948) to trade balance. The elasticity approachpredicts that, in the long-run,depreciationwould lead to improvement in the trade balanceif the sum of the absolute values of export and import demand elasticities is greater than unity.Much later,Junz and Rhomberg (1973)showedthat following devaluation trade balance is expected to followa J-curve like pattern during the short-run. According to Junz and Rhomberg, existence of various lags namely recognition, decision, delivery, replacement, and production lags in the trade balance adjustment process would initially lead trade balance torespond slowly and negatively. Only after the passage of time,elasticities of demand for export and import are adjusted to be sufficiently large to lead to improvement of trade balance.Review of currentliterature would indicateinconclusive evidence concerning the impact of devaluation/depreciation on trade balance, both in the short- and in the long-run. Studies analyzing link between deprecation and bilateral trade balancescan be classified into three groups. First group utilizes developed country samples. In this group, Marwah and Klein (1996), Shrivani and Wilbrate (1997), Bahmani-Oskooee and Brooks (1999), Hacker and Hatemi (2003) and Bahmani-Oskooee and Ratha (2004)identify significant short- as well as long-run responsesof trade balances to depreciation.However, Rose and Yellen (1989) investigatingU.S. bilateral trade with the member of G-7 countries are unable to estimate eithershort- or long-run adjustmentsin trade balancesto shock in the real exchange rates. Second group [Wilson (2001), Baharumshah (2001), Bahmani-Oskooee and Tatchawan(2001), Onafowora (2003)] analyzing newly industrialized country data also reports conflicting results. For example, Wilson and Baharumshah, using Malaysian data,estimate limited long-run adjustment whilefail to identifyany short-run adjustment in the trade balance following depreciation. By contrast, Bahmani-Oskooee and Tatchawanobserve limited short-run adjustment in Thai bilateral trade with 7 developed countries,but, no long-run adjustment in the trade balance. In the third group, two recent studies, respectively, by Arora, Bahmani-Oskooee and Goswami (2003) and Bahmani-Oskooee and Ratha (2003),investigatethe issue using data from developing countries. For example, Arora et al. examine bilateral trade data of India with her 7 largest trading partners while Bahamni-Oskooee et al. investigate the U.S. trade balance with 14 developing country partners. Both studies find limited support for short-run adjustment to trade balance.

Three important points emerge from the brief reviewof literature [and also from the comprehensive review on this topic by Bahmani-Oskooee and Ratha (2004)]. First,existing literature provides mostly inconclusive evidence on the issue of response of trade balance to exchange rate shock. Second, the studies investigating developing country samples are limited both in number and in the coverage of developing countries.Third, studies(for example, Bahmani-Oskooee and Brooks, 1999,Bahmani-Oskooee and Ratha, 2004, Bahmani-Oskooee and Tatchawan, 2001)utilizing a recent development, auto-regressive distributed lag approach, in the cointegration literature appears to identify some type of adjustment in the trade balance following currency depreciation.

Our goal in this paper is toinvestigatewhether a significant feedback from exchange rates to the bilateral trade balancescan be observed in the data for Bangladeshwith her five largest sources (the United States (U.S.), the United Kingdom (U.K.), France, Italy and Germany)of trade surplus. We are unaware of any previous study to investigatethe link between shocks in exchange rate and adjustments in trade balance of Bangladeshwith this particular groupof trading partners of Bangladesh. To estimate whether such link exists, we analyze quarterly data for the period of 1976Q1-2007Q2using the autoregressive distributed lag (ARDL) approach to cointegrationintroduced by Pesaran, Shin and Smith (2001). Next, to check the robustness of the findings from the estimates of individual countries and results from the bound F-testing, we, respectively, employ a residual based panel cointegration technique introduced by Pedroni (1999, 2004)and fixed effect estimation method. We also employ the generalized impulse response functions developed by Pesaran and Shin (1998) to trace out if any persistent J-curve like pattern can be observed in the time path of the bilateral trade balance to innovation in the real exchange rates.

The paper is organized as follows. Section II describes the data and defines the variables. Section III outlines the model. Section IVexplains the econometric techniques andpresents the results.And, section Vconcludes.

2.Description of Data

Data for this study are collected from three sources:the International Monetary Fund’s (IMF) direction of trade statistics (DOTS) for the bilateral import and export data; IMF’s International Financial Statistics (IFS) for the exchange rates data of the United States and United Kingdom, consumer price index (CPI), gross domestic product (GDP), GDP deflator data for the U.S.A., UK, France Italy and Germany and industrial production index data for the U.S.A. and Bangladesh; Pacific Exchange Rate services of the University of British Columbia for the exchange rate data for France, Italy and Germany; and, Bangladesh bank bulletin forthe CPI data for Bangladesh.Our sample consists of quarterly data ranging from the first quarter of 1976 to the second quarter of 2007. On the justification of employing bilateral trade data, we would like to cite Rose and Yellen(1989) and Bahmani-Oskooee and Goswami (2003). As Rose et al. indicates, the use of bilateral trade data reduces the problem of the measurement error due to aggregation bias. While Bahmani-Oskooeeet al. argues that theestimation of bilateral trade equations allows for the differences in the composition of trade across trading partners and yields estimates which are more relevant for country specific trade policy analysis than those obtained from multilateral trade equations. As for the reason forselecting thisgroupof countries,this group jointlycontributed nearly 63% of the trade surpluses for Bangladesh during 1995-2003. During the same period, the U.S., the U.K.,France, and Germanywere ranked among the top ten trading partners of Bangladesh, contributing more than 31% of the nation’s total trade. Finally, as necessary, data are adjusted to control for seasonal variation, transformed into log-form and expressed in 2000as the base year.

3. Model

To estimate the short- and long-run dynamics of the bilateral trade relationship between Bangladesh and her five major trading partners, we set up two trade balance equations-one controlsfor the change in exchange rate regime in 2003 in Bangladesh and the other captures any shift in the movement in the trade balance due to the unification of the currencies of France, Italy and Germany into Euro beginning January 1999.Rose and Yellen (1989, p. 54-57) provide detailed process through which equations 1 and 2 can be derived and the justification for utilizing them in the estimation of trade coefficients. The specifications of the reduced forms of the trade equations we estimate are as follows.

(1)

(2)

where, tbnt(with the subscripts 1 and 2, respectively, identifying equations 1 and 2) denotes the measureof the bilateral trade balance constructed as the ratio of the value of exports to value of imports, ynt *and ytare, respectively, the measures of foreign and domestic real economic activities. The income measures are constructed as the nominal GDP adjusted by the GDP deflators for all five partner countries.Since quarterly GDP data are not available for Bangladesh, we utilize industrial production index adjusted for changes in CPI as income proxy. Our variable of interest is the bilateral exchange rate denoted by excnt.Since the use of the real exchange rate, as indicated in Goldstein and Khan (1985, p. 1049), may yield insignificant estimates due to feed back effect generated from depreciation to domestic price, we employ both nominal and real measures of exchange rates alternately in the empirical investigation.The nominal bilateral exchange rates are the cross rates calculated as the ratio of bilateral rates between taka and the U.S. dollar and between currencies of partner countries and U.S. dollar. The real exchange ratesareconstructed as the product of bilateral nominal cross rates and the partner country CPIs divided by the domestic CPI. Theexchange rate measuresare constructed in such a manner that an increase would indicate nominal and/or real depreciation of taka and a decrease would indicate appreciation. Thus, in the short-run, if the trade balance is to exhibit a J-curve like patternfollowingdepreciation, estimates of the exchange coefficients,b, must be initially negative followed by positive values in the long-run.In the long-run, the signs of the coefficients of the incomes measures are expected to bepositive (c > 0) for the foreign real economic activitiesandnegative (d < 0) for domestic income measure.Finally, the dummy variable, D1, which takes the value equal to 0 for the period 1976Q1-1998Q4 and 1 for the period 1999Q1-2007Q2, in equation 1 controls for the impact of unification of currencies of France, Italy, Germany into euro on bilateral trade balance of Bangladesh with the respective countries. While the dummy variable, D2, taking 0 for the period 1976Q1-2003Q1 and 1 for the period 2003Q2-2007Q2, in equation 2, controls for the impact of the change in exchange rate regime in Bangladesh on the bilateral trade balances. Equation 1 is estimated using data for France, Italy and Germany while equation 2 is estimated using data for the U.S.A. and the U.K..

4. Econometric Technique and Empirical Results

The analysis of the data is performed in three steps. The first steptests for the existence of long-run relationship among the variables in equation 1 and 2 using the bound F-test procedure proposed by Pesaran, Shin and Smith (2001).The second step estimates the short-run adjustments in the trade balances to exchange rate shocks using theauto- regressive distributed lag (ARDL) approachto cointegration also introduced by Pesaran, Shin and Smith (2001), plots the generalized impulse responsefunctions introduced by Pesaran and Shin (1998) to visually inspect if any J-curve pattern can be observed in the movement of trade balance in response to shock in the bilateral real exchange rates and captures the long-run estimates from the short-run estimates obtained using ARDL approach. The final step checks the robustness of the findings inthe second stepby employing the panel cointegration tests developed by Pedroni (1999 & 2004) and fixed effect estimation method.

4.1 ARDL Approach to Cointegration

The (ARDL) approach to cointegration is considered an ideal technique in estimating small time series model for its ability to correct biases in the estimates due to contemporaneous and serial correlation, to estimate models containing fractionally integrated variables, and to capture dynamics in the adjustment process through a series of lagged explanatory variables.1 The estimation with ARDL approach is performed in two steps: (i) determining the long-run relationship among the four variables in equations 1and 2 using the bound F-testing procedure also proposed by Pesaran, et al.;2and, (ii) estimating an equilibrium conditional model (ECM) represented by equation 2 using the ordinary least square (OLS) technique.

(3)

Where D= (D1 or D2), zt = (tbnt, excnt, ynt, yt)’ = () and xit= (exnt, ynt, yt). Equation 3 is estimated for all possible lags of p(j=1,…,p). The subscriptsn=1, 2, …,5 and t=1,…,Tdenote, respectively, the number of partner countries and number of observations. The specification of equation 3 is based on the assumptions that the error terms inuntare serially uncorrelated and the ‘contemporaneous effects, , are, by construction, uncorrelated with the error term.’

Bound F-TestingProcedure

The bound F-test determines whether the variables in equation 1 are cointegrated. However, unlike the widely used ADF type tests, bound F-test does not require pre-testing of each individual variableto identify the order of integration. The bound testing process involvesthe following three steps: (i) selectingthe lag-length for the first-differenced variables in equation 2; (ii) applyingOLS to the selected ECMs; and, (iii) evaluating null hypothesis of no cointegration against the alternative of long-run relationship. The evaluation step comparesthe calculated F-values from equation 2 with those of the upper and lower band critical values, computed by Pesaran et al.. If the calculated F-statistics are greater than the upper levels of the band, the null is not accepted, indicating long-run relationshipamong the variables in equation 1. Alternatively, if the calculated F-values are smaller than those of the lower critical band, the null hypothesis is not rejected. In the case, when the calculated values fall within the lower and upper bounds, the test result is inconclusive. In such situation, additional testingis required to identify long-run relationship. One less frequently used but more efficient method (compared to widely used two-step ADF-Johansen procedure) proposed by Kremers, Ericsson, and Dolado (1992) is the use of the error correction estimates (ECMt-1)from the selected models to detect long-run relationship.

4.2 Empirical Results from Single Equation Estimation

Results from the Bound F-Test

The results displayed in table 1 suggest thatthe calculated F- values from equations1 and 2 are larger than the upper band critical values,indicating that the variables in both the equations maintain long-run relationships for all five cases.

Table 1 goes about here

The relationship holds either at 95 percent or at 90 percent confidence levels. The last column of table 1 presents the lag lengths at which the long-run relationship can be observed among the four variables.

As shown in table 2, the estimates of all five error-correction parameters (ECMt-1)are found to be highly significant at 1% level and have expected negative signs, reconfirming the conclusion of the bound F-test in table 1.Then again, quite small magnitudes of the estimates indicatea very slow adjustment of the short-run deviation to the long-run equilibrium.

Short-Run Estimates

Having determined the long-run properties of the variables, our next task is to estimate theselected equilibrium conditional models (ECM) using the OLS method. In selecting the preferred models, we utilize the Akaike Information Criterion (AIC) as the lag selection criterion. The selected ECMs are estimated while alternately employingboth measures of the exchange rates along with the income measures. Table 2displays the short-run estimates of our data analysis.

Table 2 goes about here

The estimates revealthe following information.3First, the ARDL estimates yield no clearJ-curvepattern in the movement in bilateral trade balancesexcept for the U.S.Asamples whenreal exchange ratesare used as measure of exchange rates and for France when both measures of exchange rates are used. For Germany, the use of nominal exchange rate indicates immediate deterioration of the trade balance followed by improvement. However, the short-run estimates are largely insignificant indicating merely observing J-curve pattern in trade balance movement may not have any implication for trade policy.

Generalized Impulse Response Functions4

Generalized impulse response function is a new development in the time series literature. This technique is adopted to trace out the response of a variable to one standard deviation innovation to the error term. Identifying the generalized impulse response functions, unlike traditional impulse response analysis, does not require the orthogonalization of shocks and “is invariant to the ordering of the variables in the VAR” (Pesara and Shin, 1998).Plotting generalized impulse response functions based on estimating vector error correction (VEC) model require three steps: identifying the existence of long-run relation among the variables in the model; selecting optimal lag for each specification; and, estimating the VEC model to generate the generalized impulse response function (Onfowora, 2003, p. 3). Since the bound F-testing procedure in the first step has already identified (and also corroborated by the estimates of ECMs in table 2) long-run relationship among the four variables in the model, our next step is to determine the optimal lag length, which we accomplish using the AIC criterion. In the final step, the generalized impulse response functions of the bilateral trade equations to one standard deviation innovation (increase) in the real exchange rates are generated from the estimates of VEC model. The plots of the generalized impulse response functions are shown in Figure 1. As illustrated in figure 1, no persistent J-curve like pattern can be observed in the time path of the bilateral trade balances to one standard innovation (increase) in the real exchange rates. For example, after an initial overshooting through the first quarter, the U.S. bilateral trade balanceappears to deteriorate throughout the second, third and fourth quarters to shock in the real exchange rate. Similarly, trade balances of the U.K. and France appear to deteriorate only through the first two quarters with no discernible persistent upward movement afterward. By contrast, the trade balances of Italy and Germany appear to improve throughout the eighth quarter for Italy and second quarter for Germany followed by deterioration thereafter. One common pattern that can be observed in all cases is that the time paths of trade balances seem to peter out over the twenty four quarters.