Impacts of RMB devaluation on China’s trade balances: a time-varying SVAR approach

ABSTRACT

In this article, I empirically examine time-varying effects of real renminbi (RMB) devaluation on China’s trade balances. To this end, I estimate a time-varying structural vector autoregression model with monthly data. Results demonstrated that effects of real RMB devaluation on China’s trade balances are time varying. In a few months after devaluation of RMB, trade balances worsen as predicted in J-curve theory in most sample periods. However, subsequent improvements of trade balances predicted in J-curve theory appear only in certain periods, particularly in 2000s. Finally, devaluation of RMB positively affects China’s GDP and OECD industrial production, although the size of effects varies across sample periods. Joining WTO, the global financial crisis and endogenous feedbacks induced by price effects seem to be important to understand these time-varying patterns.

KEYWORDS:

• RMB devaluation
• trade balances
• time-varying SVAR
• J-Curve
• China

JEL CLASSIFICATION:

• C32
• F31
• F41

Acknowledgments

I thank anonymous referees, Nak Gyoon Choi, Pil-soo Choi, Moon Hee Cho, and Yeo Joon Yoon for their helpful comments and suggestions. All remaining errors are my own.

Disclosure statement

No potential conflict of interest was reported by the author.

Notes

¹ Hereafter, devaluation and depreciation are used interchangeably.

² Bahmani-Oskooee and Ratha (2004) provide a useful literature survey on J-curve theory.

³ In fact, heteroscedasticity plays an important role in estimating the effect of RMB devaluation on trade balances. The model with homoskedastic volatility exaggerates improvements of trade balances by devaluation of real RMB. See appendix C.

⁴ Allowing time varying $A_t$ is crucial for the time varying structural VAR model. It can capture simultaneous interactions among endogenous variables in contrast to time invariant $A_t$. Time invariant $A_t$ implies that effects of an innovation of one endogenous variable on other variables are fixed. See Primiceri (2005).

⁵ Advantages and disadvantages of (geometric) random walks assumption are well described in Primiceri (2005) and Kirchner, Cimadomo, and Hauptmeier (2010).

⁶ Kirchner, Cimadomo, and Hauptmeier (2010) claim that the independence of ${\mathrm{\Omega }}_t$ and hyperparameters ($Q$, $S$ and $V$) is reasonable because innovations of VAR model ($u_t$) can capture short-term events such as business cycles events or policy shocks in contrast to innovations of parameters (${\nu }_t$, ${\zeta }_t$, ${\eta }_t$), which capture long-term events such as institutional changes. Such short-term events are not necessarily correlated with long-term institutional events.

⁷ The baseline model includes four variables. Thus, the structure of $S$ is given by

$S=\left[\begin{array}{ccc}{S}_1& 0_{1\times 2}& 0_{1\times 3}\\ 0_{2\times 1}& S_2& 0_{2\times 3}\\ 0_{3\times 1}& 0_{3\times 2}& S_3\end{array}\right]$

where $S_1=Var({\zeta }_{21,t})$, $S_2=Var([{\zeta }_{31,t},{\zeta }_{32,t}{]}^{\prime })$, $S_3=Var([{\zeta }_{41,t},{\zeta }_{42,t},{\zeta }_{43,t}{]}^{\prime })$ and $0_{i\times j}$ is the $i\times j$ matrix in which all elements are zero.

⁸ Detailed discussion in appendix A.

⁹ This normalization is usual in VAR literature. See, Kim and Roubini (2008) for example.

¹⁰ Unit root tests reveal that OECD IP, CHN GDP and real exchange rate are non-stationary while trade balance is stationary. Barsky and Sims (2011) argues that estimating the system in levels will produce consistent estimates of impulse responses and be robust to co-integrate unknown form even though some variables in VAR system are I(1). Based on their argument, I use linearly de-trended log-level data rather than the first differenced data.

¹¹ See Erceg, Guerrieri, and Gust (2006) and Laxton and Pesenti (2003).

¹² See appendix E for results. Sign restrictions are popular methods to identify shocks in VAR. Although this strategy is useful and prominent, I prefer short-run zero restrictions in TVP-VAR because one goal of this article is to identify a sign of effects of real RMB devaluation on trade balances. I do not want to impose prior restrictions on signs for trade balances even in short-run. Pre-existing evidences on effects of real exchange rate on GDP are ambiguous. Including additional variables in VAR exaggerates the over-parameterization problem. A long-run identification is also widely adopted in VAR literature. However, the model in this article only includes two lags to mitigate over-parameterization problems. Thus, estimators can be substantially biased because of the lag truncation bias. See Chari, Kehoe, and McGrattan (2008). Also, narrative approach in Kappler et al. (2013) is not straightforward to use in TVP-VAR models.

¹³ Since the system is partially identified, the order of other variables is immaterial.

¹⁴ Results are reported in section D, E and F in appendix.

¹⁵ It is worth noting that these results do not necessary mean that real exchange rate is not an important factor for China’s trade balances. There are at least two potential reasons to explain these insignificances. The first reason is that real effective exchange rate is not important to determine fluctuations of trade balances. The second reason is a shortcoming of TVP-VAR. Because TVP-VAR model has a lot of parameters to estimate, it tends to show wide confidence bands. For example, Primiceri (2005) and Kirchner, Cimadomo, and Hauptmeier (2010) show wide confidence bands with their time-varying frameworks. It is difficult to say which one is the true reason for insignificant responses. There are some evidences to support that these insignificances are partly due to a number of parameters. As shown in Figure 2, the response of trade balances to real exchange rate shocks is statistically significant in the sub-sample VAR. This evidence implies that statistically insignificant response is possibly caused by the shortcoming of TVP-VAR.

¹⁶ Because the dependent variable is statistically obtained, Bayesian method was used to fully take into account uncertainty about the dependent variable. Specifically, the method in Kirchner, Cimadomo, and Hauptmeier (2010) was used. For each horizon, standard error adjustment proceeds by using each of 1000 impulse responses in posterior distribution from the identified TVP-VAR as a dependent variable. All regressions are then estimated using a Gibbs sampling with 1100 draws and 100 burned-in draws, leaving 1,000,000 posterior draws of regression coefficients. Median and 90% credible sets are then computed using posterior draws.

¹⁷ Tang (2015) has also shown that RMB policy reform does not alter the relationship between real exchange rate and China’s growth rate using cumulative sum of residual (CUSUM) test.

¹⁸ Many studies have provided evidence on this effect by China factors. For example, Auer (2015) has shown that appreciation of RMB increases U.S. producer prices using monthly U.S. import price data covering 110 sectors. International Monetary Fund (2008) also shows that China factors contribute to low global inflation before the financial crisis.

¹⁹ These results seem to be partly caused by their sign restrictions. They impose negative sign on contemporaneous response of interest rate gap (U.S. – other countries) to appreciation of U.S. real exchange rate.

²⁰ The standard Taylor rule and Taylor principle assume that an increase in inflation by one percentage point should prompt monetary authority to increase its policy rate by more than one percentage point. This means that the real interest rate should decrease in response to deflation induced by lowering import price in foreign countries.

²¹ See Chap. 29 in Fan et al. (2014), for example.

²² In Appendix F, I show time-varying VAR results for Japan. The result clearly shows that effects of exchange rate on trade balances in Japan is time verying.