International Evidence on Shock-Dependent Exchange Rate Pass-Through

Abstract

We analyse the economic conditions (the “shocks”) behind currency movements and show how that analysis can help address a range of questions, focussing on exchange rate pass-through to prices. We build on a methodology previously developed for the UK and adapt this framework so that it can be applied to a diverse sample of countries using widely available data. The paper provides three examples of how this enriched methodology can be used to provide insights into pass-through and other questions. First, it shows that exchange rate movements caused by monetary policy shocks consistently correspond to significantly higher pass-through than those caused by demand shocks in a cross-section of countries, confirming earlier results for the UK. Second, it shows that the underlying shocks (especially monetary policy shocks) are particularly important for understanding the time-series dimension of pass-through, while the standard structural variables highlighted in the previous literature are most important for the cross-section dimension. Finally, the paper explores how the methodology can be used to shed light on the effects of monetary policy and the debate on “currency wars”: it shows that the role of monetary policy shocks in driving the exchange rate has increased moderately since the global financial crisis in advanced economies.

Appendix to “International Evidence on Shock-Dependent Exchange Rate Pass-Through”

1.1 Appendix A: Data Sources and Sample Periods

See Tables 2, 3.

Table 2 Country-specific data sources

Table 3 Data sources common across countries

1.2 Appendix B: SVAR Methodology and Country-Specific Results

1.2.1 Methodology

We modify the methodology developed in Forbes et al. (2018a, b) to derive shock-dependent estimates of pass-through for each of the 26 countries for which we computed reduced-form estimates of pass-through in Sect. 2.²⁸ More specifically, we estimate an SVAR with seven variables for each country: changes in nominal trade-weighted exchange rates, consumer price inflation, real GDP growth, short-term interest rates, and changes in trade-weighted foreign GDP and consumer prices as well as foreign short-term interest rates. Detailed definitions are available in "Appendix A", and all variables are at a quarterly frequency. We allow for seven shocks that can affect each country’s exchange rate: three domestic shocks (supply, demand, and monetary policy) and three global shocks (global supply, demand, and monetary policy), and we leave one shock unidentified so as to capture any shocks that don’t satisfy the characteristics of the six well-identified shocks.

In order to identify the shocks, we use a set of standard and straightforward long- and short-run zero restrictions and sign restrictions, summarized in Appendix Table 4. We impose this identification using an algorithm based on Rubio-Ramirez et al. (2010) and Binning (2013) and estimate the model using Bayesian methods with Minnesota priors.²⁹ The resulting “shock-dependent” estimates of pass-through tend to be very close to the reduced-form estimates of pass-through (from Sect. 2). For a detailed comparison of the estimates obtained using these two different methods, see Fig. 6 in Forbes et al. (2017).

Table 4 SVAR identification

1.2.2 Contributions of Shocks to Exchange Rate Fluctuations for Individual Countries and Over Time

See Figs. 7 and 8.

Fig. 7

Forecast error variance decomposition of exchange rate changes, 1990–2015

Fig. 8

Historical shock decompositions of year-on-year exchange rate changes. Notes: The figures show the year-on-year changes in the quarterly trade-weighted exchange rate (effective exchange rate) of each country, decomposed into the historical contributions of the six SVAR-identified structural shocks. All data are expressed in percent. Unlike for the regression results in the rest of this paper, the exchange rate index has not been inverted for these charts, so that a positive change reflects an appreciation rather than a depreciation and vice versa

1.3 Appendix C: Sensitivity Tests

This appendix discusses a series of sensitivity tests assessing if various assumptions and choices significantly affect the main results. In almost all cases, the choices do not significantly affect key results—so we will only briefly summarize the tests.

1.3.1 Different Approaches to Estimating Long-Sample Pass-Through

First, we examine if changes in our baseline specification of reduced-form pass-through significantly affect key results. Section 2 of the paper discussed the parsimonious distributed lag regression in Eq. (1) which is standard in the literature and used as our baseline. We also estimate 26 alternative specifications (using different control variables and lag structures), which are detailed in "Appendix B" of Forbes et al. (2017a, b). The results for pass-through from these sensitivity tests are not systematically different from the baseline estimates, but it is possible that certain controls in the baseline specification could affect pass-through estimates in ways that are correlated with the shocks behind exchange rate movements. For example, controlling for GDP growth in the first stage could capture some of the impact of the shocks (such as a demand shock), and thereby make it more difficult to find a significant role for the shock variables in the second-stage estimates discussed in Sect. 4. Or, adding oil prices to the first stage could capture some of the impact of global shocks, and therefore reduce the significance of the global supply and demand shock variables in the second-stage estimates. To test these hypotheses, we estimate Eq. (1) with no control for GDP growth, and then with an additional control for oil prices. In each case, the main results from the second-stage regressions are very similar, but with small changes that follow these predictions of the correlations between GDP growth, oil prices, and the shock variables. The shock variables are somewhat more often significant when the control for GDP growth is excluded from the first stage, and the shock variables are slightly less often significant when the control for oil prices is included in the first stage. These changes are very marginal, however, and the general conclusions do not change.

Next, we note that the reduced-form estimates of exchange rate pass-through presented in Fig. 1 are sometimes imprecisely estimated and can take on implausibly high values for some countries. Our generalized least square estimates in the second stage utilize the variance of the estimated coefficients to attach lower weight on the more imprecise estimates. This approach, however, does not address issues arising from potentially biased estimated coefficients with narrow confidence bands. The high pass-through estimate of 67% for Turkey in Fig. 1 stands out as the most notable outlier with narrow confidence bands. We therefore ran all second-stage regressions in Table 1 without the estimates for Turkey. The key results were unchanged, but some estimated coefficients became slightly larger or more significant. For example, the trade openness in Table 1 was usually significant (at least at the 10% level). The coefficients on our shock variables were unchanged.

1.3.2 Different Shock Identification Assumptions and Excluding the Crisis Period

To more precisely identify domestic shocks, our baseline SVAR model imposes sign restrictions on the response of domestic prices within the quarter a shock occurs (see Appendix Table 4). We then study the response of domestic consumer prices relative to the exchange rate eight quarters after that shock when we discuss shock-dependent pass-through. To ensure that our results are not generated by assumption, we drop all sign restrictions on the domestic CPI and re-estimate the SVAR model relying only on sign and zero restrictions on other variables to identify the six shocks of interest.³⁰ Figure 9 shows that the results do not change when we modify the sign restrictions imposed on the CPI; pass-through is still significantly lower for domestic demand shocks compared to domestic monetary policy shocks. The second-stage regressions based on these estimates and explaining variations in pass-through are also broadly similar, with the coefficients on the shock variables of similar magnitudes, albeit the marginally significant coefficients in some cross-country regressions become insignificant.

Fig. 9

Cross-country mean group estimate for the response of consumer prices conditional on a 1% exchange rate depreciation, eight quarters after shock: SVAR model without sign restriction on uomestic CPI. Notes: The light blue range depicts the 2 standard error range of median consumer price responses corresponding to a 1% exchange rate depreciation caused by different shocks across the 26 countries

Second, our sample includes the financial crisis—an episode when many countries experienced particularly large exchange rate movements. To check if the estimated differences between pass-through after different shocks are driven by the crisis only, we re-estimate the SVAR without data for 2008 and 2009.³¹ The average shock-dependent pass-through across countries is plotted in Fig. 10 and is similar to our main result. Indeed, the relative responses of domestic prices and exchange rates are less dispersed across countries when the crisis is excluded from the estimation period, as indicated by the somewhat narrower confidence bands. Estimating the same second-stage regressions as in Sect. 4 is only feasible across countries, but not over the 6-year periods within countries (as these cover the crisis). The resulting coefficients on shock variables are similar to the baseline, but not significant.

Fig. 10

Cross-country mean group estimate for the response of consumer prices conditional on a 1% exchange rate depreciation, eight quarters after shock: excluding the financial crisis in 2008–2009. Notes: The light blue range depicts the 2 standard error range of median consumer price responses corresponding to a 1% exchange rate depreciation caused by different shocks across the 26 countries

1.3.3 Different Measures of “Structure” Variables

Many of the structural country characteristics that the previous literature has highlighted as affecting pass-through are highly correlated with each other. This, especially when combined with our limited sample size, makes it difficult to control for many variables simultaneously. Therefore, we estimate the main regressions using a range of different possible “structural” variables and show a sample of the results below. Table 5 presents an extended version of columns 1–7 of Table 1, where each set of structural characteristics is first included simultaneously and then kept in the more parsimonious regressions including the shocks only if significant. Table 6 presents a similar set of regressions, controlling for more structural characteristics, but using 6-year windows, as in columns 8–14 of Table 1. As discussed in the text, these highlight the consistent importance of inflation volatility for explaining pass-through estimates across countries and over time, supporting our choice of variables in the baseline regressions presented in the main text of this paper.

Table 5 Determinants of estimated long-sample exchange rate pass-through, controlling for a larger set of structural characteristics

Table 6 Determinants of estimated 6-year window exchange rate pass-through, controlling for a larger set of structural characteristics

1.3.4 Different Measures for Monetary Policy

One challenge during the period in our analysis is measuring changes in monetary policy, as several economies were at their effective lower bound (ELB) for their policy interest rate and adjusted monetary policy through other tools. More specifically, in the wake of the global financial crisis, the central banks of Canada, Israel, Japan, UK, Sweden, and Switzerland all lowered their interest rates to levels below 1% and some of them introduced additional unconventional easing measures such as quantitative easing or forward guidance. Given these policies, official interest rates are likely to miss changes in the monetary policy stance, affecting our identification of monetary policy shocks.

To address this potential issue, we estimated the SVAR model using several alternative interest rate series for these countries. First, we re-estimate the SVAR model for Japan and the UK—two countries in our sample that used quantitative easing most extensively—using shadow interest rates developed by Krippner (2016). The resulting mean group estimates of the relative responses of domestic consumer prices and the exchange rate, conditional on each shock identified in our SVAR, are almost identical to the baseline results using central bank rates shown in Fig. 2 of the paper. The second-stage regression results for the role of the shocks in explaining differences in exchange rate pass-through across countries and over time are reported in the left of Tables 7 and 8 alongside the baseline results from Table 1 of the paper for the cross-country results (columns 2, 3, and 7) and time-series results (in columns 9, 10, and 14). The results are almost identical.

As no shadow rates were available for Canada, Israel, Sweden, and Switzerland, we also performed two more robustness checks using financial market forward rates of different maturities to proxy for the stance of monetary policy. We re-estimated the SVAR model for the six countries that were near the zero lower bond using 3-month, 3-month and 6-month, 6-month forward rates from interest rate swap markets. The resulting mean group estimates of shock-dependent exchange rate pass-through are again very similar to the baseline results presented in the paper. Second-stage regression results for the role of shocks in explaining variations in reduced-form pass-through are also similar and shown to the right of Tables 7 and 8.

Table 7 Determinants of estimated long-sample exchange rate pass-through: coefficients on shocks from SVAR estimated with shadow or forward interest rates

Table 8 Determinants of estimated 6-year window exchange Rate pass-through: coefficients on shocks from SVAR estimated with shadow or forward interest rates

1.3.5 Other Tests

In a final set of sensitivity tests, we explore the role of commodity exporters/importers and emerging markets. As noted, these different groups of countries often have different characteristics of pass-though, such as emerging markets generally having higher pass-through and commodity exporters being differentially affected by permanent global shocks than commodity importers. To better explore these relationships, we included an emerging market dummy variable in the second-stage regressions wherever possible, and then a dummy variable capturing whether the country is a commodity exporter or importer.³² These additional dummy variables are usually insignificant and do not change the key results. We also considered the possibility that the impact of the shocks we focus on differ depending on whether the country considered is a net commodity exporter or importer. To test for these types of effects, we include an interaction between the commodity exporter/importer dummy and the shock contributions in the second-stage regressions. The coefficients on the interaction term were not significant and the key results are robust. A more thorough analysis of the importance of commodity price shocks for exporters and importers of those commodities, and how these may affect pass-through, requires a VAR model which identifies commodity shocks separately. While this could potentially be important for some countries, e.g. New Zealand as found in Parker and Wong (2014), we leave that for future research.

Finally, we explore whether changes in the VAT/GST could explain some of the sharp changes in pass-through that occur for specific countries over time. For our main analysis, we measure the pass-through of exchange rate movements to the consumer price index (not adjusted for changes in VAT or other government policies) as this measure is most widely available across countries. For several countries, however, there is a period when estimated pass-through changes around the same time that there were changes in the VAT (such as in Australia in 2000, New Zealand in 2010, and the UK in 2008, 2010 and 2011). Therefore, for these three countries, we re-estimate the rolling regressions of pass-through, but add dummy variables for increases or decreases in the VAT/GST. This flattens the short-term jumps in estimated pass-through experienced by New Zealand from 2011 to 2014, but it has minimal effect on the pass-through estimates for Australia and the UK.