EXCHANGE RATE PASS-THROUGH: THE CASE OF BRAZILIAN EXPORTS OF MANUFACTURES

Transcription

1 EXCHANGE RATE PASS-THROUGH: THE CASE OF BRAZILIAN EXPORTS OF MANUFACTURES Afonso Ferreira Departamento de Economia - Universidade Federal de Minas Gerais (UFMG) and Centro de Pesquisa em Economia Internacional, Brazil. Andreu Sansó Departament d Econometria, Estadistica i Economia Espanyola - Universitat de Barcelona and Grup de Recerca en Anàlisi Quantitativa Regional, Spain. Abstract: The purpose of this paper is to determine the extent to which exchange rate movements affect the prices of Brazilian exports of manufactured goods. The paper presents estimates of the pass-through coefficient, derived from cointegration tests based on the Engle-Granger, Shin and Johansen procedures. When data for the whole / period were used, the point estimates of the pass-through coefficient, obtained from different model specifications and different cointegration tests, were comprised in the relatively narrow interval of 10%-27%. On the other hand, when the data base was split into two sub-samples to allow for the possibility of a structural change in the pass-through relationship, the estimates derived from the tests of cointegration varied from zero to 48%. Although the coefficient estimates reported in this paper are not robust to changes in the sample period, model specification and estimation procedure, all the results obtained indicate that the passthrough of exchange rate changes to prices is far from complete, in the Brazilian case. Additionaly, the pass-through coefficient may have changed over time, being much higher in the period / than in the period of pronounced macroeconomic instability from / Keywords: exchange rate pass-through, exports, Brazilian economy. JEL classfiication: F12, F14

2 2 1. INTRODUCTION The purpose of this paper is to determine the extent to which exchange rate movements affect the competitiveness of the Brazilian exports of manufactured goods. Changes in the exchange rate are normally split into changes in the destination currency prices of exported goods and changes in the profit margins of the exporting firms. A large body of literature has been devoted recently to the empirical evaluation of these two competing effects for several countries. The importance of this question for economic policy is quite obvious. If exchange rate changes are not fully or substantially reflected in the selling prices of exported goods, their impact on the demand for exports will be limited, even when the price elasticity of demand is quite large [Menon, 1995]. In such a case, the efficacy of the exchange rate as a policy tool in programs of export promotion and balance of payments adjustment may be reduced. The next section presents the simple mark up price model usually adopted in empirical studies of this kind. Section 3 brings a brief description of the data, while section 4 reviews the econometric methodology used in the paper. Estimates of the pass-through coefficient, derived from cointegration tests based on the Engle- Granger, Shin and Johansen procedures, are reported for the Brazilian case, in section 5. A final section offers some conclusions. 2. THE MODEL The starting point for the analysis is a mark up price model of the type: PX = (1 + λ) (CP / ER) (1) where PX = foreign currency price of manufactured exports, CP = production cost in domestic currency, ER = nominal exchange rate, CP/ER = production cost in foreign currency and (1 + λ) = mark up. Considering that manufactured goods are typically differentiated and traded in markets characterised by imperfect competition, the adoption of the traditional mark up model in (1), as a first approximation, seems to be justified.

3 3 Presuming that the mark up may vary, according to the competitive pressures in the world market, we would have: 1 + λ = [ PW / (CP / ER) ] α (2) where PW is the gap between the price of world exports and the exporter s CP ER production cost measured in foreign currency [Menon, 1995]. Combining (1) and (2), then, gives: ln PX = (1 - α) ln (CP/ER) + α ln PW (3) We can also postulate an unrestricted version of Equation (3), in which we do not require the sum of the ln CP/ER and ln PW coefficients to be equal to 1 and in which, additionally, we allow the ln CP and ln ER coefficients to differ both in sign and in magnitude: ln PX = φ 0 + φ 1 ln CP + φ 2 ln ER + φ 3 ln PW (4) where φ 0 is a constant. The parameter φ 2 in (4) is the pass-through coefficient. When φ 2 = 0, the exchange rate has no influence on the foreign currency price of manufactured exports and, therefore, exchange rate changes affect only the exporters profit margins, with no impact on the competitiveness of the country s exports. Conversely, when φ 2 = - 1, i.e., when pass-through is complete, any exchange rate variation is fully transmitted to the price of exports in foreign currency, thereby affecting the competitiveness of domestic production in the world market. Clearly, for values of φ 2 in the interval 1 < φ 2 < 0, pass-trough will be incomplete. 3. DATA In the empirical implementation of the model presented in the previous section, PX is measured by the price series for the Brazilian manufactured exports calculated by Guimaraes et al. [1997], the domestic cost variable (CP) is proxied by

4 4 the wholesale price index and ER is given by the R$-US$ exchange rate (since most of the Brazilian exports are invoiced in dollars, the use of this exchange rate is appropriate, in the present context). Finally, PW is proxied by the IFS-IMF price series for the industrial countries exports. Quarterly data for the period / were used in all tests. However, since the variables entering several test equations were in lagged first differences, some observations were lost and the period of estimation was, for that reason, reduced to the interval / The series PX, (CP/ER) and PW are plotted, in log form, in Figures 1 and ECONOMETRIC METHODOLOGY Cointegration analysis is adopted in this paper to determine whether Equation (1), in logarithmic form, as well as Equations (3) and (4), represent long run equilibrium relationships. A preliminary step in this sort of analysis consists in examining whether the time series involved are stationary. This was done here by combining unit root (DF/ADF) and stationarity (KPSS) tests. The Dickey Fuller/Augmented Dickey-Fuller (DF/ADF) procedures test the null H 0 : y t I(1) against the alternative H a : y t I(0). In what follows, the critical values adopted for the DF/ADF tests are those derived by MacKinnon (1991) and available in MicroTSP 7.0 and EVIEWS 2.0. Differently from the DF/ADF test, the KPSS test, proposed by Kwiatkowski et al. [1992], establishes the null as H 0 : y t I(0) and the alternative as H a : y t I(1). Results of three different versions of the KPSS test for the series being considered in this paper are reported in section 5, one adopting a spectral quadratic window and the other two a Bartlett window (with automatic and manual lag selection). In performing these tests, a Gauss routine written by Andreu Sansó was used. Amano and Van Norden (1992) examine the consequences of using the KPSS stationarity test in conjunction with a standard unit root test, such as the DF/ADF test. Their results suggest that, when this procedure is adopted, the frequency of incorrect conclusions may be decreased relative to the application of only standard unit root tests. Also, such a joint testing procedure may in some cases permit researchers to be

5 5 more confident about their tests results (Amano and Van Norden, 1992). We have, therefore, chosen to report here the results derived from both tests. The tests for cointegration among the variables in (1), (3) and (4), which are also presented in section 5, are based on the Engle-Granger [1987], Shin [1994] and Johansen [1988, 1991] procedures. The Engle-Granger [1987] procedure consists in testing the OLS residuals of the cointegrating equation (in the present context, Equations (1), (3) or (4) above) for the presence of a unit root, using the DF/ADF statistic. Under the null of nonstationarity of the residuals, the series involved are not cointegrated. The procedure proposed by Shin [1994], on the other hand, makes use of a KPSS-type test to evaluate the null of cointegration. Similarly to the Engle-Granger method, the cointegrating equation is first estimated by OLS, with the residual series being then subjected to the KPSS test for stationarity. If the null of stationarity cannot be rejected, the series in question are said to be cointegrated. Again, following Amano and Van Norden (1992), if the diagnostic obtained from the Shin test (H 0 : cointegration) does not contradict the diagnostic derived from the Engle-Granger test (H 0 : no cointegration), we have strong evidence in favour of one of the hypotheses. Finally, the Johansen [1988, 1991, 1995] procedure is based on the following vector error correction model: Z t = Σ 1 p Γ i Z t-i + Π Z t-1 + ν t (5) where the vector Z t contains the n variables in the model. If the variables in Z t are I(1), the rank of matrix Π is ρ < n and there exists a representation of Π such that Π = αβ, where α and β are both n x ρ matrices. Matrix β is called the cointegrating matrix and has the property that β Z t is I(0), while Z t is I(1). The columns of matrix β have an economic interpretation as cointegrating vectors, i.e, after normalisation, they may be interpreted as giving estimates of the long run parameters of the model. In empirical applications, the main concerns consist in determining ρ, which gives the number of cointegrating vectors, and in estimating the cointegrating matrix β. This is done by means of an ML procedure described in Johansen [1988, 1991, 1995].

6 6 5. RESULTS Table 1 shows the DF/ADF test statistics for the presence of a unit root in the series under analysis. The DF/ADF tests suggest that all series are I(1), a diagnostic confirmed by the KPSS tests, reported in Table 2. 1 According to the Shin test, Equations (1), (3) and (4) all constitute cointegrating relationships, a result which departs from that derived from the Engle- Granger procedure, which rejects the null of no cointegration only for Equation (1). The estimated cointegrating equations point out to a pass through coefficient between 11% and 26%, suggesting that only a small proportion of a change in the exchange rate is transmitted to the export prices of Brazilian manufactured goods, in the long run. The measured impact of changes in the price of world exports, on the other hand, is larger, with 30% to 40% of a variation in those prices being reflected in the price of Brazilian exports, according to the test results presented in Table 3. Table 4 shows the results obtained when the Johansen method is applied to Equation (4). The Akaike and Schwarz statistics, as well as the log-likelihood ratio test, all pointed out to an optimal VAR lag length of 4 (the value of p, in Equation (5)). For this VAR lag length, the null of only one cointegrating vector was rejected, in favour of the alternative hypothesis of two cointegrating vectors. Only one of the two cointegrating vectors, however, displayed the theoretically expected signs. Following what is the established practice in such situations, this is the only result taken into account here, providing the following cointegrating equation: ln PX = ln CP ln ER ln PW Given the evidence of non-normality of the residuals, the Johansen procedure was also applied to Equation (4), with the addition of dummy variables controlling for a variety of shocks that affected the Brazilian economy during the period under analysis. 2 Again a VAR lag length of 4 was adopted. In this version of the Johansen test, also reported in Table 4, the LR statistic suggested the existence of only one cointegration vector, with the cointegrationg equation corresponding to: ln PX = ln CP ln ER ln PW

7 7 Reassuringly, the ln CP, ln ER and ln PW coefficients given by the Johansen exercise are similar to those derived from the Engle-Granger and Shin tests, suggesting a low pass-through coefficient, in the range of 10% to 27%. Stability Analysis A visual inspection of Figures 1 and 2 suggests the possibility of a structural change in the relationship between the price of exports in foreign currency, on the one hand, and domestic costs and the price of world exports, on the other, around the first quarter of In fact, it is possible to argue a priori for a structural break in our equations in this period. Starting with the Cruzado Plan, in the first quarter of 1986, and until the implementation of the Real Plan, in July 1994, the Brazilian economy experienced a succession of failed attempts at stabilisation, each of which involved some form of fixed exchange rate which was, after a short period of time, abandoned. All this led to a pronounced instability in the exchange rate. It seems reasonable to expect that, in such an extremely unstable environment, the price behaviour of exporters would change with respect to the relatively more stable 1978/1985 period. In order to examine this possibility, we reestimated our equations for the subsamples / and /1996.4, using the Engle-Granger and Shin tests for cointegration. The results are shown in Tables 5 and 6. The two tests again, in general, yield opposite results, with the Engle-Granger procedure favouring the hypothesis of no cointegration and the Shin procedure suggesting that the series cointegrate. Regarding the coefficient estimates, we have that Equation (1) suggests a value of the pass-through coefficient of 48%, in the first period, and 31%, in the second period. When the influence of the price of world exports is also taken into account, as in Equations (3) and (4), the estimated value of the pass-through coefficient falls to approximately 30%, in the years /1985.4, and close to zero, in the period / While these results confirm the previous conclusion, based on data for the whole sample period, of an incomplete pass-through of exchange rate changes to export prices, they also point out to a significant change in the behaviour of exporters

8 8 between the two sub-periods being considered. In the first sub-period (1978.3/1985.4), the exchange rate pass-through was incomplete, but relatively significant in economic terms (30%). In the second sub-period (1986.1/1996.4), it may have fallen to zero, i.e., changes in the exchange rate may have had no impact on the prices of the Brazilian exports of manufactures, with changes in those prices being related basically to movements in the price of world exports. 3 One possible explanation for these results is that, in the second period, the Brazilian exporters chose not to change their prices in foreign currency, after changes in CP/ER, i.e., the exchange rate pass-through became zero, because of the extreme exchange rate instability which characterised this period. In other words, we are suggesting the possibility that the pass-through coefficient is a function of, among other factors, the volatility in the exchange rate. To preserve their shares in external markets, exporters will choose to maintain their prices in foreign currency invariant to changes in the exchange rate that are perceived as transitory. For this reason, increases in the variability of the exchange rate may be accompanied by a reduction in the coefficient of pass-through. 6. CONCLUSIONS The pass-through relationship between exchange rate changes and prices of traded goods determines the degree of competitiveness achieved from variations in the exchange rate. The efficacy of the exchange rate as a policy tool in programs of export promotion and current account adjustment depends, therefore, on the coefficient of pass-trough. Newly industrializing countries like Brazil are generally considered to have very little control over prices at which they sell in international markets. The implication is that exchange rate changes may be of little relevance in determining the prices of their exports in those markets, i.e. that the pass-through coefficient is close to zero (Athukorala, 1991). The purpose of this paper was to estimate the pass-through coefficient for the Brazilian exports of manufactures. When data for the whole / period were used, the point estimates of that coefficient, obtained from different model specifications and different cointegration tests, were comprised in the relatively narrow interval of 10%-27%.

9 9 On the other hand, when the data base was split into two sub-samples to allow for the possibility of a structural change in the pass-through relationship, the estimates derived from the tests of cointegration varied from 48% (Equation (1) in the /1985.4) to zero (Equation (3) in the / period). These results seem to warrant the following two conclusions: (a) Although the coefficient estimates reported in this paper are not robust to changes in the sample period, model specification and estimation procedure, all the results obtained indicate that the pass-through of exchange rate changes to prices is far from complete, in the Brazilian case (again, the highest estimate derived for the pass-through coefficient was 48%, for the / period, when Equation (1) was adopted). (b) The pass-through coefficient may have changed over time, being much higher in the period / than in the period of pronounced macroeconomic instability from / One important implication of these results is that the adjustment of prices in foreign currency that tends to follow any exchange rate movement is relatively limited, with variations in the exchange rate being reflected, to a large extent, in variations in the price in domestic currency and, therefore, in the profit margins of the exporters. This means that, in periods of currency depreciation, the Brazilian exporters enjoy significant increases in their profit margins, while in periods of currency appreciation they are forced to squeeze their margins or to drop out of the international market altogether. As pointed out by Athukorala and Menon [1994], the coefficient φ 2 in Equation (4) measures the direct effect of an exchange rate change on the price of exports, for a given level of domestic costs, what they call the pricing to market (PTM) effect, related to strategic pricing behavior on the part of the exporting firms, which aims to protect market share during currency appreciation or to augment profit margins during currency depreciation. Exchange rate movements, however, also affect the cost of production measured in domestic currency, via their effect on the domestic price of imported inputs. The coefficient φ 2, therefore, overestimates the total impact of a variation in the exchange rate on the price of exports, since that impact is equal to the direct effect (given by φ 2 ) less the indirect effect related to the change in input costs. In view of the low degree of openess and the limited reliance on

10 10 imported inputs that characterized the Brazilian economy during the years 1977/1990, it is, however, reasonable to expect the direct effect captured by δ 2 to be, by far, the most important influence in determining how exchange rate changes affected the price of manufactured exports, during the period under analysis in this paper 4. A low pass-through coefficient implies that a devaluation has a limited effect on the demand for exports, irrespective of the value of the price elasticity of demand. A devaluation, however, may still have a significant impact on the volume of exports through its supply effect: with a low pass-through, the price of exports in domestic currency and, thus, the profit margins of the exporting firms increase, raising the supply of exports. This was probably the main channel through which exchange rate changes until recently affected the volume of exports, in the Brazilian case. A final point suggested in this paper is that the value of the pass-through coefficient may depend on the behaviour of the exchange rate itself, with periods of high exchange rate instability being characterised by low pass-through coefficients, since most exchange rate changes would, in this case, be viewed as transitory. If this latter conclusion proves to be correct, the relatively successful program of macroeconomic stabilisation being implemented since 1994 may contribute to enhance the efficacy of the exchange rate as a policy instrument in future efforts of balance of payments adjustment, in Brazil. NOTES 1. Similar results were also obtained from the stationarity test proposed by Leybourne and McCabe (1994). 2. Dummy variables corresponding to the following quarters were considered: , , , , , and Note that when we allow for the possibility of a structural change in Equations (3) or (4), the sum of the regression coefficients on ln CP/ER and ln PW is close to 1, as predicted by Equation (3). 4. The effect of exchange rate changes on input costs may have increased somewhat in the Brazilian case, since 1990, as a consequence of the program of trade liberalisation. See, however, in this respect, Oliveira Jr. (1999), who shows that

11 11 the share of imported inputs in intermediate consumption in Brazil rose from 3.1% to only 4.5%, between 1990 and 1995, a quite modest increase. REFERENCES Amano, R. and S. van Norden (1992). Unit root tests and the burden of proof. Athukorala, P. (1991). Exchange rate pass-through: the case of Korean exports of manufactures, Economics Letters 35, pp Athukorala, P. and J. Menon (1994). Pricing to market behaviour and exchange rate pass-through in Japanese exports, Economic Journal 104 (March), pp Engle, R. and C. Granger (1987). Co-integration and error correction: representation, estimation and testing, Econometrica 55, pp Guimarães, Eduardo et al. (1997). Indices de preço e quantum das exportações brasileiras, Fundação Centro de Comercio Exterior, mimeo. Johansen, S. (1988). Statistical analysis of cointegrating vectors, Journal of Economic Dynamics and Control 12, pp Johansen, S. (1991). Estimation and hypothesis testing of cointegration vectors in Gaussian vector autoregressive models, Econometrica 59, pp Johansen, S. (1995). Likelihood-Based Inference in Cointegrated Vector Auto- Regressive Models. Oxford University Press. Kwiatkowski, D., P. Phillips, P. Schmidt and Y. Shin (1992). Testing the null hypothesis of stationarity against the alternative of a unit root: how sure are we that economic time series have a unit root?, Journal of Econometrics 54, pp Leybourne, S.J. and B.P.M. McCabe (1994): "A Consistent Test for a Unit Root". Journal of Business and Economic Statistics, 12 (2), pp MacKinnon, J. (1991). Critical values for cointegration tests, in Engle, R. and C. Granger (eds.). Long run economic relationships. Oxford University Press. Menon, J. (1995). Exchange rate and import prices for a small open economy, Applied Economics 27, pp Oliveira, Jr., M. (1999). Desvalorização cambial e inflação: o papel da importação de insumos, Boletim IPEAd no. 55, May, p. 8.

12 12 Phillips, P. and P. Perron (1988). Testing for a unit root in time series regression, Biometrica 75, pp Sansó, A. (1998). Contrastes de estacionariedad: una sintesis, Departament d Econometria, Estadistica i Economia Espanyola - Universitat de Barcelona, mimeo. Sephton, P. (1995). Response surface estimates of the KPSS stationarity test, Economics Letters 47, pp Shin, Y. (1994). A residual-based test of the null of cointegration against the alternative of no cointegration, Econometric Theory 10, pp

13 13 TABLE 1 DF/ADF TESTS Period of estimation: / VARIABLE DF/ADF STATISTIC k without with trend trend Levels ln PX ln (CP/ER) ln CP ln ER ln PW First Differences ln PX *** 0 ln (CP / ER) *** 0 ln CP * ** 1 ln ER ** 1 ln PW *** Notes: The value of k corresponds to the lag length on the lagged dependent variable in the RHS of the ADF test equation that was required to produce approximately white noise residuals. An intercept term was included in all DF/ADF test equations. The critical values, which were taken from MacKinnon (1991), are: (a) without trend: (10%); (5%); (1%); (b) with trend: (10%); (5%); (1%). * = significant at the 10% level; ** = significant at the 5% level; *** = significant at the 1% level.

14 14 TABLE 2 KPSS TESTS - Period of estimation: / VARIABLE Spectral quadratic Bartlett window Bartlett window window (automatic (automatic lag (manual lag lag selection) selection) selection) without trend ln PX 0.700** 0.705** 1.059*** ln (CP/ER) 1.022*** 1.005*** 1.663*** ln CP 1.140*** 1.116*** 1.881*** ln ER 1.139*** 1.116*** 1.880*** ln PW 1.059*** 1.038*** 1.708*** with trend ln PX ln (CP / ER) ** ln CP 0.266*** 0.263*** 0.430*** ln ER 0.267*** 0.264*** 0.432*** ln PW ** Notes: The critical values, which were taken from Sephton (1995), are: (a) without trend: (10%); (5%); (1%); (b) with trend: (10%); (5%); (1%). * = significant at the 10% level; ** = significant at the 5% level; *** = significant at the 1% level.

15 15 TABLE 3 COINTEGRATION TESTS LHS Variable is ln PX - Period of estimation: / VARIABLE Equation (1) Equation (3) Equation (4) Cointegrating Equations ln CP / ER ln CP ln ER ln PW Cointegration Tests ENGLE-GRANGER (null hypothesis: no cointegration) k= k= * k= ** critical values 10% % % SHIN (null hypothesis: cointegration) Spectral quadratic window (automatic lag selection) Bartlett window (automatic lag selection) Bartlett window (manual lag selection) * critical values 10% % % Notes - The value of k, in the Engle-Granger test, corresponds to the lag length on the lagged dependent variable in the RHS of the DF/ADF test equation. * = significant at the 10% level ** = significant at the 5% level

16 16 TABLE 4 COINTEGRATION TESTS - JOHANSEN PROCEDURE - Equation (4) Period of estimation: / Null Alternative LR statistic Critical Values Hypothesis Hypothesis 5% 1% Without dummies ρ=0 ρ = ** ρ 1 ρ = * ρ 2 ρ = CV = [1, , 0.267, ] With dummies ρ=0 ρ = ** ρ 1 ρ = ρ 2 ρ = CV = [1, , 0.104, ] Notes: (1) An intercept term was included in the cointegrating equation and in the VAR, which corresponds to assuming a linear trend in the levels of the series. A VAR lag length of 4 was adopted, in both exercises. Dummy variables were included in the second version of the exercise, corresponding to the following quarters: , , , , , and (2) ρ is the cointegrating rank and gives the number of cointegrating vectors; LR is the Likelihood Ratio test statistic for the corresponding null hypothesis about the cointegrating rank; CV is the normalised cointegrating vector. (3) * significant at the 5% level; ** significant at the 1% level. (4) Only cointegrating vectors with economically meaningful signs are reported. (5) The variables entered the tests in the order [ln PX, ln CP, ln ER, ln PW]. The cointegrating vectors reported above were normalised in ln PX.

17 17 TABLE 5 COINTEGRATION TESTS LHS Variable is ln PX - Period of estimation: / VARIABLE Equation (1) Equation (3) Equation (4) Cointegrating Equations ln CP / ER ln CP ln ER ln PW Cointegration Tests ENGLE-GRANGER (null hypothesis: no cointegration) k= k= * SHIN (null hypothesis: cointegration) Spectral quadratic window (automatic lag selection) ** Notes - The value of k, in the Engle-Granger test, corresponds to the lag length on the lagged dependent variable in the RHS of the DF/ADF test equation. * = significant at the 10% level; ** = significant at the 5% level

18 18 TABLE 6 COINTEGRATION TESTS LHS Variable is ln PX - Period of estimation: / VARIABLE Equation (1) Equation (3) Equation (4) Cointegrating Equations ln CP / ER ln CP ln ER ln PW Cointegration Tests ENGLE-GRANGER (null hypothesis: no cointegration) k= k= SHIN (null hypothesis: cointegration) Spectral quadratic window (automatic lag selection) * Notes - The value of k, in the Engle-Granger test, corresponds to the lag length on the lagged dependent variable in the RHS of the DF/ADF test equation. * = significant at the 10% level

19 PX CPER PX PW