On the Asymmetric Effects of Exchange Rate Volatility on Trade Flows: Evidence from Africa

ABSTRACT

A glance through the literature on the effects of exchange rate uncertainty on the trade flows reveals that African countries have received the least attention. We consider the response of exports and imports of 13 African nations to a Generalized Autoregressive Conditional Heteroskedasticity (GARCH)-based measure of exchange rate uncertainty. Like previous research, when we used linear models in which the effects are assumed to be symmetric, we found significant long-run effects in almost one-third of the countries in our sample. However, when we shifted to nonlinear export and import demand models, we found significant long-run effects of exchange rate uncertainty on trade flows of almost all countries. These effects were asymmetric in nature.

KEY WORDS:

• Africa
• asymmetry effects
• exchange rate volatility
• trade flows

JEL CLASSIFICATION:

• F31

Notes

1. Note that Bussiere (2013) has shown that import and export prices react to nominal exchange rate changes in an asymmetric manner. This could be another source of asymmetric response of trade flows to volatility of the real exchange rate which combines nominal exchange rate and prices.

2. See the Appendix for details.

3. Note that income elasticities in Equations (1) and (2) could be negative if economic growth is due to increased production of import substitute goods (Bahmani-Oskooee 1986).

4. An upper bound critical value is tabulated by assuming all variables in a given model to be integrated of order one or I(1). A lower bound value is tabulated if all variables are stationary or integrated of order zero, I(0). For cointegration, the calculated F statistic must be greater than upper bound critical value. Note that as they demonstrate, the upper bound critical value could also be used even if some variables are I(1) and some I(0), and this is another advantage of this approach. The third advantage of this approach is that since short-run dynamics is included in estimating the long-run effects, it allows the feedback among the variables to exert their impact (Pesaran, Shin, and Smith 2001, p. 299).

5. Indeed, Shin, Yu, and Greenwood-Nimmo (2014, p. 291) argue that due to dependency between the partial sum variables, they should be treated as one variable so that the critical values of the F test do not change when we move from linear to nonlinear models.

6. For some other application of these methods, see Gogas and Pragidis (2015), Durmaz (2015), Baghestani and Kherfi (2015), Al-Shayeb and Hatemi-Jarabad (2016), Lima et al. (2016), Aftab, Shah Syed, and Katper (2017), Arize, Malindretos, and Igwe (2017), and Gregoriou (2017).

7. Under this alternative test, we use normalized long-run coefficient estimates and long-run model (1) and generate the error term, and we denote it by ECM. We then go back to error-correction model (2) and replace the linear combination of lagged level variables by ECM_t-1 and estimate this new specification after imposing the same optimum lags from Panel A. A significant coefficient obtained for ECM_t-1 will support cointegration. Note that the t test that is used to judge significance of this estimate has new critical values that Pesaran, Shin, and Smith (2001, p. 303) tabulate. This test is also known as the t test for cointegration.

8. As for other diagnostic statistics, clearly there is evidence of autocorrelation free residuals in most models since the Lm statistic is insignificant in most cases. So are the RESET statistics, supporting correctly specified optimum models in most countries. Furthermore, estimates are stable in most models as reflected by the CS and CS² tests.

9. For determinants of real exchange rate volatility, see Calderon and Kubota (2018).