Exchange Rate Pass-through to Import Prices, and Monetary Policy in South Africa

Abstract

Understanding how import prices adjust to exchange rates helps anticipate inflation effects and monetary policy responses. This article examines exchange rate pass-through to the monthly import price index in South Africa during 1980–2009. Short-horizon pass-through estimates are calculated using both single equation equilibrium correction models and systems (Johansen) models, controlling for both domestic and foreign costs. Average pass-through is incomplete at about 50 per cent within a year and 30 per cent in six months, and in the long-run, from the Johansen analysis including feedback effects, is about 55 per cent. There is evidence of slower pass-through under inflation targeting; pass-through is found to decline with recent exchange rate volatility and there is evidence for asymmetry, with greater pass-through occurring for small appreciations.

Acknowledgements

Janine Aron acknowledges support from the British Academy (British Academy Research Development Award). This research was supported in part by grants from the Open Society Institute and the Oxford Martin School. We are grateful to Bent Nielsen (Nuffield College, Oxford) and Anindya Banerjee (Banque de France) for helpful advice and to an anonymous referee for careful  comments.

Notes

^1. It is important to recognise that exchange rate changes are not always the same as exogenous exchange rate shocks. The pass-through from exchange rate changes could well be different than from exchange rate shocks and this distinction will be explored in this article.

^2. Log price levels are usually thought to be non-stationary variables; that is, showing no tendency to revert to a constant level or a precise trend. If the first difference of a log price level is stationary or ‘integrated of order 0’, I(0), the log level is said to be ‘integrated of order 1′’, I(1). A vector of different I(1) variables is said to be co-integrated if there exists a linear combination of these variables which is I(0). If one or more of the variables drives this linear combination away from its mean (or trend), there is then a dynamic adjustment process, useful for modelling or forecasting, tending to correct this deviation.

^3. One reason for including domestic costs is because of exporters’ local distribution costs. Another reason is because of tendencies to price to the domestic market.

^4. Generally simpler versions are used in the empirical literature, omitting controls for domestic costs and commodity prices, and sometimes also demand controls. Sometimes a time trend is also included to capture long-run evolutions in, for example, global trade openness or productivity levels. Then, an intercept term appears in the differenced form of Equation (1) – see below.

^5. Including several earlier lags would mean that computation of 12-month pass-through would have to be done by other methods such as dynamic simulation.

^6. The interaction terms cover an annual change to match the short-run exchange rate terms in duration: . Various other parameterisations are possible.

^7. Autometrics is an objective and easily reproducible tool, not affected by the subjective choices of the modeller. This software examines a full set of general to simple reduction paths to select a parsimonious form of the ‘general unrestricted model’ (GUM) to satisfy a set of test criteria. The test criteria include tests for normality, heteroscedasticty, including a test of ARCH residuals, residual autocorrelation, parameter stability in the form of a Chow test, and the RESET test. There is also the option of automatically dummying out large outliers. The results are reproducible using the same data, the same specification of the general unrestricted model and the same settings in Autometrics.

^8. For instance, instead of , with PLL the lags take the form , using only 6 parameters instead of 24 parameters, but covering the same two years with monthly, quarterly, six-monthly and annual changes, see Aron and Muellbauer (2013).

^9. These include strict exogeneity of foreign producer and commodity prices, long-run homogeneity and currency translation, and are sufficient to identify the impulse response function.

^10. The chosen Autometrics parameters are for a target model size of 0.01, with robust t ratios (HACSE ratios) and switching off the normality and heteroscedasticity tests as a criterion for selection. Skewed exchange rate and price data typically fail normality tests, while outliers cause heteroscedasticity.

^11. When freely estimating the error correction terms dated at t–1 in a single equation model, the weights given to the domestic and foreign prices from the Johansen analysis are supported. Dating the error correction terms at t–12, they are, as might be expected, not quite as well determined.

^12. The best-fitting demand measure is the current quarterly rate of acceleration of the log index of manufacturing output. The quarterly rate is measured as the three-month moving average of the log index. The rate of acceleration is measured as the double three-month difference. This incorporates the current month production index. One could object to this on the grounds of endogeneity bias. Lagging the measure by one month reduces the significance quite sharply, though it remains significant.

^13. Campa and Goldberg (2005) find a positive foreign demand growth effect, capturing a cyclically varying mark-up over unit labour costs. This is probably due to their use of foreign unit labour costs to measure foreign prices. Producer prices, as used in our article, are likely already to incorporate this mark-up.

^14. The pass-through coefficients after six months are around 13 per cent and 47 per cent in the two sub-samples respectively, and 34 per cent for the full sub-sample.

^15. The outliers associated with the October 2008 financial crisis and the 1985 debt crisis were included, along with outliers in late 1990 expressed as an impulse in October and a change from December to November 1990. The last is linked with the Gulf War, but entering as a change has no effect on the pass-through coefficient.

^16. We also tested interactions with a simple linear trend plus the trend itself, finding that interaction terms were not significant.

^17. One would not expect the transition from foreign currency to domestic currency invoicing to be instantaneous, but gradual. Smoothing takes the form of a double 12-month moving average of a step dummy which is zero before 2003 and 1 from January 2003. The smoothed dummy makes the transition from 0 at the end of 2002 to 1 from January 2005. We test alternatives with different reaction times to this shock: (a) starting in 2002 or 2003 or 2004; (b) slower and faster transitions, taking one or two years. The results are robust: each produces a significant result with comparable orders of magnitude, but the 24-month transition is always preferred to the 12-month transition.

^18. Kurita and Nielsen (2009) discuss the inclusion of such shifts in dynamics in co-integration analysis. They show that conventional asymptotic tables apply for statistical inference if the shifts in dynamics are for the second difference of, in this application, the log exchange rate. In our case, shifts in dynamics apply to the first difference of the log exchange rate. Then, parameter estimates should remain consistent, but the asymptotic distributions need some adjustment. We are grateful to Bent Nielsen for advising us on this point.