Effects of Real Exchange Rate on Trade Balance in Cote d’Ivoire: Evidence from Threshold Nonlinear ARDL Model

The objective of this paper is to shed new light on the nonlinear relationship between real exchange rate changes and trade balance in Cote d’Ivoire. In examining this issue, previous studies have relied on the nonlinear autoregressive distributed lag model developed by Shin et al. (2014) where real exchange rate is decomposed into partial sum of positive and negative changes. They fail to examine the response of the trade balance to extreme changes in the real exchange rate. In order to investigate possible sign and size-dependence in the response of the trade balance to exchange rate, this study uses multiple threshold nonlinear ARDL modeling. The approach is implemented using annual data covering the period 1975-2017. The results reveal that the effects of changes in real exchange rate are asymmetric in both time horizons. More specifically, real exchange rate appreciations deteriorate the trade balance while real depreciations improve it. Further, the effect of large depreciations is higher when compared with large appreciations.


Introduction
The relationship between the exchange rate and the trade balance is a subject of intense research in international economics, especially since 1973 when Magee (1973) introduced the concept of the J-curve effect. According to this concept, a real devaluation of domestic currency leads to an initial deterioration of the trade balance, subsequently followed by an improvement. This suggests that the effect of real depreciation of exchange rate on trade balance is negative in the short-run and becomes positive in the long-run. The study by Magee (1973) ar-Y. Keho DOI: 10.4236/tel.2021.113034 508 Theoretical Economics Letters gued that contracts negotiated before the devaluation of the domestic currency may be responsible for the deterioration of the trade balance in the short-run. Theoretically, devaluation improves the trade balance through two channels. Firstly, it makes the domestic goods cheaper for the foreign trading partners, thus increasing exports. Secondly, it brings about a decrease in imports since imports become relatively more expensive. Understanding the extent to which real exchange rate changes are passed through to trade balance is an important issue from balance payment perspective. Until recently, the existing empirical research assumed a linear relationship between real exchange rate and trade balance. The evidence from this literature is mixed and conflicting. Some studies found that depreciation of real exchange rate has a significant improving effect on trade balance (Baharumshah, 2001;Bahmani-Oskooee, 2001;Boyd et al., 2001;Kale, 2001;Lal & Lowinger, 2002;Musila & Newark, 2003;Yol & Baharumshah, 2007;Igue & Ogunleye, 2014;Anning et al., 2015;Caporale et al., 2015;Ogbonna, 2016;Hunegnaw & Kim, 2017;Ousseini et al., 2017;Sokeng Dongfack & Ouyang, 2019;Yazgan & Ozturk, 2019;Keho, 2021a), while others reported that depreciation worsens trade balance (Onafowora, 2003;Adeniyi et al., 2011;Genemo, 2017). Still, others failed to establish any significant relationship between the two variables (Hatemi-J & Irandoust, 2015;Duasa, 2007;Loto, 2011;Akpansung & Babalola 2013;Shawa & Shen, 2013;Meniago & Eita, 2017;Kamugisha & Assoua, 2020). Such conflicting findings point to the need for further examination of the relationship between exchange rate and trade balance.
It is possible that the failure to appropriately establish a significant relationship between real exchange rate and trade balance may be as a result of the assumption of symmetry. Recently, the economic literature developed some arguments that make the hypothesis of linearity unrealistic. According to this strand of literature, trade balance may respond differently to exchange rate appreciation and depreciation. This asymmetry in exchange rate pass-through to trade balance could be attributed to price rigidity, that is prices are more sticky downwards than upwards. There is evidence that trade goods prices respond to exchange rate changes asymmetrically (Karoro et al., 2009;Bussière, 2013). Others sources of asymmetry include adjustment costs, quantity restrictions, the presence of market power, government interventions, fixed contracting and negotiating terms of exporters (Magee, 1973;Bussière, 2013). In such a context, results from linear models may be misleading.
A number of empirical researches employed the nonlinear ARDL approach introduced by Shin et al. (2014) to re-examine the asymmetric response of trade balance to real exchange rate changes. For instance, Iyke and Ho (2017) examined the case of Ghana and found that real depreciation of the exchange rate improves the trade balance in the long-run, while real appreciation does not have any impact on the trade balance. Relying on the nonlinear ARDL model, Bahmani-Oskooee & Fariditavana (2015, 2016 and Arize et al., (2017)  Considering 97 commodities, they found significant short-run asymmetric effects for two-third of the commodities and significant long-run asymmetry for one-third of the commodities. Bahmani-Oskooee, M., and Arize, A. C. (2020) examined the case of 13 African countries and found asymmetric effects in many of the countries. Bahmani-Oskooee, M., and Fariditavana, H. (2020) focused on the bilateral trade balance of the U.S. with Canada for 161 industries. They found asymmetric short-run effects in all industries and significant long-run asymmetric effects in 62 industries. Bahmani-Oskooee, M. and Gelan, A. (2020) considered the bilateral trade balance of South Africa with the U.S. for 25 industries. They found short-run asymmetric effects in 19 industries, which last into long-run asymmetric effects in 14 industries. Keho (2021b) investigated the trade balance of Nigeria and found that positive and negative shocks in the exchange rate are associated with a negative impact on the trade balance, but only the effects of positive shocks are significant. Finally, in a study of Cote d'Ivoire, Nathaniel (2020) showed that the trade balance responds stronger to negative shocks in the real exchange rate than to positive shocks in the long-run. The nonlinear ARDL approach of (Shin et al., 2014) filters appreciations from depreciations and evaluates their separate effects on the trade balance. This implies a zero threshold effect depending on the direction of the change in exchange rate. Studies with respect to size of change in exchange rate are quite scanty. This study aims to strengthen the empirical literature by investigating the effects of exchange rate non-linearity on the trade balance in Cote d'Ivoire. In addition to investigating the magnitude and direction of the effect of exchange rate on the trade balance, it analyses the size effect of a change in the real exchange rate. This is the first paper that applies multiple threshold nonlinear ARDL model to test the short and long-run asymmetric effects of real exchange rate on the trade balance in Cote d'Ivoire. It is our belief that the results of this study will contribute invaluably to the existing literature on the effects of exchange rate on the trade balance in African countries.
The rest of the paper is organized as follows. Section 2 describes the econometric methodology and the data used in the study. The results are presented and discussed in Section 3. Finally, Section 4 provides some concluding remarks.

Empirical Model
The objective of this study is to ascertain the effects of real exchange rate on the trade balance and to test whether real depreciation and appreciation of currency have the same effect on the trade balance. Following existing studies, the empirical model is specified as follows: where: TB is the trade balance on goods and services defined as the ratio of exports to imports; Y is domestic real GDP used as a proxy for domestic income; YF is world real GDP used as a proxy for foreign income; RER is the real effective exchange rate defined in a way that a decline (increase) reflects a real depreciation (appreciation) of the domestic currency; D94 is a dummy variable that takes the value of zero for the period before 1994, the year of the CFA franc devaluation, and one otherwise; µ t is an error term assumed to be a white-noise process.
As far as the expected sign of each coefficient is concerned, the estimates of β 1 and β 2 can either be positive or negative depending upon whether income is spent on domestic goods or foreign goods. The coefficient β 1 is negative if higher levels of domestic income increase the demand for imports. Similarly, the coefficient β 2 is positive suggesting that higher levels of foreign income increase the demand for domestic goods by the rest of the world. Finally, a real depreciation of the real exchange rate is expected to improve the trade balance. Therefore, the coefficient β 3 is expected to be negative.

Data Description
The study uses annual data covering the period from 1975 to 2017.  Figure 1 portrays the trends of the trade balance and the real effective exchange rate over the period 1975-2017. We observe that both variables exhibit considerable fluctuations over the sample period. The real effective exchange rate has been appreciating after the devaluation of the CFA franc in January 1994. Between 1994 and 2009, the real effective exchange rate appreciated by about 24%, and started to depreciate from 2009. It is worth mentioning that structurally low inflation rate over the post-devaluation period helps limit the appreciation of the real effective exchange rate, which is still trading 25% below the 1993 level. Table 1 shows some descriptive statistics and correlations of the variables. It can be observed that trade balance averages 4.746 and ranges from 4.421 to 4.986. It can be seen that domestic and foreign income exhibit the highest variability over the period, as indicated by their respective standard deviations, while real effective exchange rate and trade balance are the least volatile. The Jarque-Bera statistic suggests that all the variables under study are normally distributed. The correlation matrix indicates a positive and significant relationship between the trade balance and foreign income. A negative and significant relationship exists between the trade balance and the real exchange rate whereas foreign income is positively and significantly associated with the trade balance. The correlation matrix also reveals a problem of collinearity between domestic income and foreign income, with a correlation value of 0.913. Hence, the foreign income variable was left out of the empirical analysis.

Econometric Methodology
The methodology of the empirical analysis proceeds as follows. As a first step, we test for the order of integration of the series by using unit root tests. Second, in order to avoid spurious estimation we must establish that the variables are cointegrated. The evidence from existing empirical studies suggest that macroeconomic time series are either integrated of order zero or one. A cointegration method that allows variables to be a combination of I(0) and I(1) was proposed by Pesaran et al. (2001) as bounds testing approach to cointegration. This approach is relatively more efficient in small sample data sizes in which the order of integration is not well known or may not be necessarily the same for all variables under study. It has shown to provide consistent estimates of the long-run parameters in the presence of endogenous regressors (Inder, 1993;Gonzalo, 1994). The bounds test for cointegration involves estimating by ordinary least square the following ARDL model: After reparametrisation, Equation (2) can be rewritten in the unrestricted error-correction form: A main feature of Equation (3) is that appreciation and depreciation of real exchange rate have the same effect on the trade balance. Accordingly, it does not allow potential asymmetry or nonlinearity in the effect of exchange rate on the trade balance. As mentioned before, this needs not be the case if imports and exports respond to exchange rate changes in an asymmetric manner. Negative shocks in the real exchange rate could impact on the trade balance at different rates compared to positive shocks. variations. This approach implies one-threshold model with two sub-variables of the real exchange rate, splitting depreciations from appreciations. This decomposition provides results that may be easily interpreted. However, the use of a zero threshold may introduce a finite sample problem in one regime (Davies, 1987). In addition, it is possible that the trade balance responds differently when facing small or large changes in the exchange rate. In the absence of a clear guideline to measure large and small changes, we can set the threshold value equal to one standard deviation of the first difference of the exchange rate. Another candidate of the threshold could be the mean value of the exchange rate growth rate. We can also use the quantiles of the real exchange rate growth (∆lnRER) and define large change as any change that is greater than the 75 th quantile, while a small change is below the 25 th quantile. In this study, we borrow the concept of partial sum from econometric literature and use a more general approach determining the threshold values endogenously. Accordingly, the partial sum series are constructed as follows: This specification allows us to investigate possible sign and size-dependence in the response of the trade balance to the real exchange rate changes. The presence of a long-run relationship between the variables is tested by restricting the coefficients of lagged level variables equal to zero, that is, H 0 : δ 0 = δ 1 = δ 2 = δ 3 = δ 4 = 0. This hypothesis is tested using F-test. The (Pesaran et al., 2001). The lower critical bound assumes all the variables to be stationary, i.e. I(0), whereas the upper critical bound assumes all the variables to be stationary at first difference, i.e. I(1). If the computed F-statistic is higher than the upper critical value, then the null hypothesis of the absence of long-run relationship among the variables is rejected. If the F-statistic is below the lower critical value, then the null hypothesis cannot be rejected. The result is inconclusive if the computed F-statistic falls within the upper and lower bound values. Shin, et al. (2014) argued that due to dependency between partial sum variables, LPOS, MED, and LNEG should be treated as one variable and consequently assume that there are two (and not four) exogenous variables. The implication is that the same critical values of the F-test should be used in both the linear and nonlinear models.
When the thresholds s 1 and s 2 are unknown, they can be consistently estimated using a grid search over the values of ∆lnRER, after removing the 15% extreme observations. The optimal threshold values are selected by minimizing the sum of squared residuals of the OLS regression of the threshold asymmetric ARDL model. Once the nonlinear ARDL model is estimated and cointegration is established, we test for symmetry in the real exchange and the trade balance nexus by means of Wald tests: 1) long-run symmetry has the null hypothesis δ 2 = δ 3 = δ 4 and 2) short-run symmetry has the null hypothesis ϕ 2j = ϕ 3j = ϕ 4j , for all j.
If the null hypotheses are rejected, then the model is asymmetric. There may be short-run asymmetry, long-run asymmetry or both short and long run asymmetries in the response of the trade balance to the real exchange rate. Given that the thresholds are not identified under the null hypotheses, the Wald statistic testing for symmetry follow a nonstandard distribution (Hansen, 1996). For this reason, their appropriate p-values are obtained by means of bootstrap procedure.

Results and Discussion
The empirical analysis begins with determination of the order of integration of the variables by means of unit root tests. This step is necessary to ascertain that none of the variables is integrated of order 2. For this purpose, two unit root tests are used: the PP test of (Phillips & Perron, 1988) and the DF-GLS test suggested by (Elliott, et al. 1996). The tests are done both in the level and first difference of the variables and the results are presented in Table 2. Both tests are conclusively in favor of unit root for all variables in levels. However, when applied to the first differences, the null hypothesis of unit root is rejected. These findings suggest that all variables are non-stationary in the level and stationary in the first difference, implying that they are I(1) processes.
After checking the stationarity of the series, the next step is devoted to cointegration analysis and estimation of the short and long run dynamics of the trade balance model.  The optimal lag structure of the DF-GLS test is chosen based on the Schwarz Information Criterion. The 5% critical values for the Phillips-Perron unit root tests are −2.932 and −3.520 for models C and C/T respectively. Those for the DF-GLS test are −1.949 and −3.190 for models C and C/T, respectively. * denotes rejection of the null hypothesis of unit root at the 5% level. Notes: The dependent variable is the trade balance (TB) defined as log(X/M), where X and M are exports and imports, respectively. Y denotes domestic real income, LPOS, MED and LNEG denote large appreciation, medium change and large depreciation in the real effective exchange rate. * and ** indicate significance at the 5% and 10% levels, respectively. The model estimated is an ARDL(1, 1, 1) model. The upper bound critical value of the F-test for cointegration when there are two exogenous variables is 4.85 at the 5% level of significance. The comparable figures when k = 4 in the nonlinear model is −4.01 at the 5% level of significance. LM is the Lagrange multiplier statistic to test for autocorrelation of order two with p-value in brackets. LR is the Likelihood ratio statistic of Breusch-Pagan-Godfrey heteroskedasticity test with p-value in brackets. W SR denotes the Wald test for the short-run symmetry, which tests the null hypothesis that ϕ 2j = ϕ 3j = ϕ 4j in Equation (7). W LR refers to the Wald test for the long-run symmetry, which tests the null hypothesis that δ 2 = δ 3 = δ 4 in Equation (7). The associated p-values are given in brackets. These results suggest that the nonlinear ARDL model combining both the short and long run asymmetry is suitable for this study.
As cointegration exists among the variables, we further proceed with the esti-

Conclusion
This paper re-examines the effect of the real exchange rate on the trade balance and short-run, suggesting that real depreciation of domestic currency improves the trade balance while real appreciation worsens it. Therefore, our results do not support the evidence of the short-run worsening of trade balance suggested by the J-curve effect. We found that large depreciations are passed to the trade balance in a larger extent than large appreciations. This confirms that the trade balance responds asymmetrically to the real exchange rate. Our findings also suggest that in monitoring the exchange rate movements, attention should not only be focused on large changes but also on relatively small changes. On the other hand, as economic growth reduces the trade balance, the country should implement economic policies that focus on the production of imported-substituted goods. Import-substitution policies will help to mitigate the deterioration in the trade balance caused by real exchange rate appreciation.
This study is not free of shortcomings. First, the study uses aggregate trade balance of Cote d'Ivoire with the rest of the world. As Cote d'Ivoire has different export and import prices with its trade partners, the results may suffer from aggregation bias. Therefore, future research should be done at disaggregated level using the bilateral trade balance of Cote d'Ivoire with its major trade partners.
Second, it may be promising to examine the exchange rate and trade balance relationship at commodity level. We intend to investigate these lines of research in future works.

Conflicts of Interest
The author declares no conflicts of interest regarding the publication of this paper.