- Email: [email protected]
Relative productivity increases and the appreciation of the Turkish lira
Economic Modelling 35 (2013) 614–621
Contents lists available at ScienceDirect
Economic Modelling journal homepage: www.elsevier.com/locate/ecmod
Relative productivity increases and the appreciation of the Turkish lira Kenan Lopcu a,⁎, Fikret Dülger b,1, Almıla Burgaç b,1 a b
Department of Econometrics, Çukurova University, Balcalı, Adana 01330, Turkey Department of Economics, Çukurova University, Balcalı, Adana 01330, Turkey
a r t i c l e
i n f o
Article history: Accepted 6 August 2013 JEL classification: C22 E31 F31 Keywords: Balassa–Samuelson hypothesis Real effective exchange rate Relative productivity Cointegration Multiple structural breaks
a b s t r a c t This paper studies the Balassa–Samuelson hypothesis between Turkey and 27 members of the European Union. More specifically, using recently developed cointegration techniques with multiple breaks, we test the relationship between the real effective exchange rate and inter-country differences in the relative productivity of the tradable and non-tradable sectors over the period 1990:Q1–2011:Q2. In recent years, the Central Bank of the Republic of Turkey (CBRT) has emphasized the importance of the B–S hypothesis for Turkey. Our findings, however, suggest that changes in relative productivity have played a limited role in explaining the real effective exchange rate appreciation. In particular, the relationship between the real effective exchange rate and productivity indicated by the Balassa–Samuelson hypothesis is not supported for the post 2001 era in Turkey. © 2013 Elsevier B.V. All rights reserved.
1. Introduction The Balassa–Samuelson (B–S) hypothesis (Balassa, 1964; Samuelson, 1964) identifies inter-country differences in relative productivity of tradable and non-tradable sectors as the main source of long-run changes of real exchange rates (and hence, deviations in purchasing power parity). The Central Bank of the Republic of Turkey (CBRT) in recent years has endorsed the importance of the Balassa–Samuelson hypothesis for Turkey by stating2: … [D]ifferences in productivity and relative price differences appear to have been moving together recently. In conclusion, productivity increases and the differentiation in price increases in the tradable and non-tradable goods are among the factors shaping the development of the real exchange rate in Turkey … [A] close relationship is being observed in the Turkish economy between the relative price differentiation in tradable and non-tradable goods, and the increase in productivity. … [S]ome part of the real appreciation of the New Turkish lira stemmed from the rapid increase in productivity in recent years. The strong stance of the New Turkish lira will continue
⁎ Corresponding author. Tel.: +90 322 338 72 65; fax: +90 322 338 72 83. E-mail addresses: [email protected] (K. Lopcu), [email protected] (F. Dülger), [email protected] (A. Burgaç). 1 Tel.: +90 322 338 72 66. 2 The CBRT has been applying explicit inflation targeting strategies since 2006. Inflation reports issued by the CBRT on a quarterly basis are closely followed by economic agents in Turkey. 0264-9993/$ – see front matter © 2013 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.econmod.2013.08.005
to be supported via this channel, under the assumption that the increases in productivity would prevail in the upcoming real convergence process with the EU (CBRT, 2006, p. 33–34). Briefly, according to the CBRT, part of the appreciation of the Turkish lira (TL) can be attributed to the productivity differentials between tradable and non-tradable sectors. In particular, in addition to the direct effects of structural reforms undertaken after the 2001 crisis, increased confidence, optimism, and macroeconomic stability in recent years have contributed to the strengthening of the national currency through the relative price differentials between tradable versus non-tradable goods (CBRT, 2006). The main purpose of this study, given the increased consideration of the B–S effect by the CBRT, is to test whether part of the appreciation of the TL is driven by the relatively rapid productivity increases in the tradable sectors in recent years. Hence, the study examines the validity of the B–S hypothesis between Turkey and 27 member countries of the European Union (EU-27), which include the majority of Turkey's main trading partners, for the post-financial liberalization era (i.e., 1990: Q1–2011:Q2), using recently proposed cointegration techniques with multiple structural breaks by Kejriwal (2008). Our analysis offers very limited support for the B–S hypothesis. In particular, the presence of the B–S effect is generally limited to the pre-2001 period, and in some cases even to the pre-1994 period. We do not find any support for the B–S hypothesis from the data for the post-2001 era. Our results are robust to alternative specifications and additional controls, and hence, undermine the increased consideration of the B–S effect by the CBRT.
K. Lopcu et al. / Economic Modelling 35 (2013) 614–621
A large body of literature exists that empirically tests the validity of the B–S hypothesis. Although many of the studies have found that movements in real exchange rates are consistent with the B–S hypothesis, the jury is still out on the effect of relatively rapid productivity changes in the tradable sector (with respect to the non-tradable sector) on the real exchange rate movements. For example, Alberola et al. (1999), Rahn (2003), MacDonald and Wojcik (2004), MacDonald and Ricci (2003, 2005), Lane and Milesi-Ferretti (2004), Gente (2006), Alberola and Navia (2008), Coudert and Couharde (2009), Apergis (2013) and Coulibaly and Gnimassoun (2013), among many others, have documented that the data are consistent with the B–S hypothesis. On the other hand, Méjean (2008), Bénnasy-Quéré et al. (2010) and Hall and Guo (2012) find no evidence for the B–S effect, while Clark and MacDonald (2000), Égert (2002, 2005), Lommatzsch and Tober (2004), Égert et al. (2006), Camarero (2008), and Dumrongrittikul (2012) have produced mixed evidence showing that the B–S hypothesis is supported for some countries included in the sample but not for others. Our paper contributes to this literature by using recently developed cointegration techniques, which allow for multiple breaks in the data, to test the B–S hypothesis. In addition, with the exception of Égert (2005), Dumrongrittikul (2012) and Alper and Civcir (2012), to the best of the authors' knowledge, there has been no study that tests this hypothesis using the Turkish data. Égert (2005) analyzed the domestic B–S effect for a group of EU-acceding, EU-accession (including Turkey) countries and the Common Wealth of Independent States (CIS). This study found that the relative price–relative productivity relationship was mostly insignificant for Turkey. Dumrongrittikul (2012) found a significant but negative relationship between the real exchange rate and relative productivities. Hence, her results for Turkey are inconsistent with the B–S hypothesis. Alper and Civcir (2012), on the other hand, found that the real exchange rate appreciated with positive productivity shocks while it depreciated in response to the increase in net foreign assets. However, the magnitude of the estimated elasticity in that study (around 5) is well above any reasonable a priori expectation and the estimates supported by the empirical literature. This may stem from their choice of a particular proxy for the productivity differentials, the way that the real exchange rate series was constructed, and the econometric technique employed. The rest of this paper is organized as follows. Section 2 introduces the empirical strategy used to test the B–S hypothesis along with the data and their constructions. Section 2.2 reports the empirical findings and studies the robustness of the results. Section 3 concludes the paper. 2. Empirical strategy and data 2.1. Empirical strategy The B–S hypothesis relates the real exchange rate between home and foreign countries to the corresponding relative productivity differentials between these countries. To test this hypothesis, following much of the literature (e.g., Clark and MacDonald, 2000), we begin our analysis by considering the following general form: Reert ¼ ƒðd–Prodt ; Nf at ; d–Rir t Þ where Reer denotes the real effective exchange rate, d_Prod denotes the corresponding relative productivity differentials, Nfa represents the net foreign assets at home, and d_Rir represents the real interest rate differentials between home and foreign economies. An increase in d_Prod leads to an appreciation of the Reer according to the B–S effect.3 The 3 More recently within the New Open Economy Macroeconomic (NOEM) Models framework, productivity increases can lead to a real depreciation due to tradable prices while they can cause a real appreciation due to non-tradable prices (the B–S effect). Thus, the sign of the productivity in reduced form is ambiguous. Similarly, an increase in net foreign assets may or may not lead to an appreciation. See Égert et al. (2006) for a discussion of both of these issues.
615
effect of Nfa, on the other hand, is ambiguous. According to the Balance of Payments (BOP) approach, an increase in Nfa is associated with an appreciation of the Reer in the long term. However, the sign is negative over the medium term if a decrease in Nfa (debt creation) is linked to the appreciation of the Reer. This may apply to countries with rapid growth prospects (Égert et al., 2006). Higher d_Rir, on the other hand, attracts foreign capital and causes the Reer to appreciate via a surplus in the capital account of BOP. In our empirical implementation, we consider the following specific versions of the above model: Model 1 : Reert ¼ cons þ β1 d–Prodt þ ut Model 2 : Reert ¼ cons þ β1 Prodtr t þ β2 Prodeut þ ut Model 3 : Reert ¼ cons þ β1 d–Prodtr t þ β2 Nf at þ ut Model 4 : Reert ¼ cons þ β1 Prodtr t þ β2 Prodeut þ β3 Nf at þ ut Model 5 : Reert ¼ cons þ β1 d–Prodt þ β2 d–Rirt þ ut Model 6 : Reert ¼ cons þ β1 Prodtr t þ β2 Prodeut þ β3 d–Rir t þ ut Model 7 : Reert ¼ cons þ β1 d–Prodt þ β2 Nf at þ β3 d–Rirt þ ut where cons refers to the regression intercept, and Prodtr and Prodeu stand for relative productivities in the tradable and non-tradable sectors in Turkey and EU-27, respectively. We start the empirical analysis by investigating the order of integration of the variables. Next, to assess the stability of the relationship between the real exchange rate and its determinants, we use the tests proposed by Kejriwal and Perron (2010) involving non-stationary but cointegrated variables with multiple structural changes of unknown timing in regression models. If the Kejriwal–Perron tests corroborate the existence of structural breaks, then we verify whether the variables are indeed cointegrated by cointegration tests following Kejriwal (2008), which are based on the extension of the one-break cointegration tests developed by Arai and Kurozumi (2007) (A–K henceforth) with a null of cointegration. Because our series seem to be trending (see Fig. 1), we include a deterministic trend in the unit root as well as cointegration tests. Finally, we estimate the model with breaks to investigate how the relationship between the real exchange rate and its determinants may have altered over time. 2.1.1. Structural break tests Kejriwal and Perron (2010) considered three types of statistics for testing multiple breaks. The first is the sub-Wald test, SubF, of the null hypothesis of no structural break against the alternative hypothesis of k breaks. The second test, a double maximum test called UDmax, checks the null hypothesis of no structural breaks against the alternative of an unknown number of breaks. The third test involves a sequential procedure (SEQ) that analyzes the null hypothesis of k breaks against the alternative hypothesis of k + 1 breaks. A useful strategy, then, is to use significant SubF and UDmax tests to decide if breaks exist and subsequently utilize the sequential procedure to determine the number of breaks (Kejriwal, 2008). As an alternative, the number of breaks can also be determined by using the Bayesian information criterion (BIC) suggested by Yao (1988) and the modified Schwarz criterion proposed by Liu et al. (1997) (LWZ). In this study, stability tests for the relationship between the real effective exchange rate and a number of regressors are performed, using the sequential procedure as well as the information criteria, following Kejriwal (2008), to investigate the existence of breaks in the real effective exchange rate regressions. 2.1.2. Cointegration tests with multiple structural breaks Kejriwal and Perron (2010) showed that the structural change tests they proposed have good size and power properties. In addition, as
616
K. Lopcu et al. / Economic Modelling 35 (2013) 614–621
pointed out in Kejriwal (2008) structural change tests also have power against a purely spurious regression. This means that when the cointegrating relation is unstable, the conventional cointegration tests are biased towards the non-rejection of the null of no cointegration. Hence, cointegration analysis should consider the structural changes. Structural change tests developed by Gregory and Hansen (G–H) (1996) under the null of no cointegration have power against the alternative of a single break, and therefore can have a low power if there is more than a single break. Finally, if the primary concern is cointegration with structural breaks, the null of cointegration is a more natural choice from the viewpoint of conventional hypothesis testing (Kejriwal, 2008). To avoid these problems, Kejriwal (2008) extends the cointegration test with the known or unknown one structural break tests proposed by A–K to analyze multiple structural breaks under the null of cointegration. In the current study, following the work of A–K and Kejriwal, we further augment the A–K model with a deterministic trend as our variables seem to be trending, and allow shifts in the trend as well. The regime and trend shift model used in this study is as follows: yt ¼ ci þ δi t þ z t′ βi þ ut
if T i−1 b t ≤ T i
for i ¼ 1; …k þ 1
where k is the number of breaks, zt is a vector of I (1) regressors, given by zt = zt − 1 + uz, yt is the dependent I (1) variable, and by convention, T0 = 0 and Tk + 1 = T. Augmenting the above regression model to deal with the simultaneity bias, we use the Dynamic Ordinary Least Squares (DOLS), adding the leads and lags of the first differences of the regressors. In this study the leads and lags are equal to 2 where span of data, the number of breaks and the number of variables permit, and one in other cases. yt ¼ ci þ δi t þ z t′ βi þ
lT X
Δz ′t− j Π j þ ut⁎ if T i−1 b t ≤ T i
j¼−lT
non-tradable sector as a whole, a weight is needed for each of the subsector productivity. To calculate the weights, we total the output for all non-tradable sub-sectors separately. Then, we calculate the percentage of the total output attributed to each sub-sector by dividing the total output in that sub-sector into the grand total output of the broad category of non-tradable sector. All the sectoral output and employment series for EU-27 as well as the sectoral output series for Turkey are obtained from the statistical office of the European Union (Eurostat). The employment series for Turkey is from the Turkish Statistical Institute (Turkstat) and the CBRT. The output and employment series for each sub-sector are seasonally adjusted using X-12, before the average productivity for each sub-sector is calculated. The dependent variable in our study is the consumer price index (CPI) based real effective exchange rate (Reer)6 obtained from Eurostat. An increase in the Reer of Turkey corresponds to an appreciation of the TL. The Reer, as well as all the productivity variables, Prodtr, Prodeu and d_Prod are in natural logarithm. The net foreign asset series for Turkey is computed by the difference in the total foreign assets minus the liabilities to non-residents divided by the Gross Domestic Product (GDP), and from the CBRT. The real interest rate differentials between Turkey and G7 are used as a proxy for d_Rir. The annual percentage rate on three month treasury bills (TB) and the CPI based inflation series are used to compute the d_Rir. Both the TB rates and CPI series for Turkey are from the Undersecretariat of the Treasury, while the TB rates and CPI based inflation series for G7 are from the International Financial Statistics (IFS). For Turkey, the inflation series is calculated by the authors using the Turkish CPI series. The variables used and the data sources are also summarized in Table 1. 3. Results
for i ¼ 1; …; k þ 1: The test statistic for k breaks, then, is given by: T −2 b ¼ e λ V k
XT
S t¼1 k
Ω11
2 b λ
:
Where Ω11 is a consistent estimation of the long run variance of u⁎t , t b ¼ ∑u b ¼ Tb1 =T; …; Tbk =T , Sk λ bb represents the residuals bb, and u λ iλ i¼1 iλ from the augmented model above. The break points Tb1 ; …Tbk are obtained by minimizing the sum of squared residuals. The above test statistics are compared with the critical values for multiple breaks generated by the authors, modifying the programs developed for Kejriwal (2008).4 2.2. Data The data set covers the period from 1990:Q1 to 2011:Q2. We take the EU-27 as the benchmark foreign country. Manufacturing represents the tradable sector, while the non-tradable sector includes construction, wholesale and retail trade, and community, social and personal services. Average labor productivity is used as a proxy for the productivity variable suggested by the theoretical model as commonly used in the literature.5 Hence, in order to compute productivity in the tradable sector, the total output in manufacturing is divided by the employment level in the manufacturing sector. To calculate productivity for the 4
The authors would like to thank Mohitosh Kejriwal for making available his Gauss Codes. The productivity measure based on the total hours worked rather than the total employment level would better proxy the theoretical productivity variable. Unfortunately, a measure based on the total hours worked is simply not available. For this reason, the productivity per worker, the GDP per worker or per capita GDP are widely used in the literature. For example, Camarero (2008), Dumrongrittikul (2012), Égert (2002, 2005), Égert et al. (2006), and MacDonald and Wojcik (2004), among others, used the productivity per worker, while Coudert and Couharde (2009), Alper and Civcir (2012), and Alberola and Naiva (2008), employed the GDP per capita and the GDP per worker, respectively. 5
In order to scrutinize the integrating level of variables, the Ng and Perron (2001) tests are employed. To corroborate the results of the Ng– Perron tests, we also employ more conventional unit root tests, namely the Augmented Dickey and Fuller (1979) (ADF) and Kwiatkowski et al. (1992) (KPSS) tests. Table 2 presents the Ng–Perron unit root tests. In Ng–Perron tests, the null of non-stationary is rejected if the test statistic is smaller than the critical value. The results show that for all the variables, the null of non-stationary in levels cannot be rejected at any conventional significance level by any of the Ng–Perron tests. ADF and KPSS tests also corroborate these results, while all three tests provide evidence that the first differences of the variables are stationary. Hence, we conclude that the variables used in the study are integrated order of one, I (1). As a next step, we test the null of no structural change in the longrun relationship. The results obtained are reported in Table 3. For all the models, overall the tests offer evidence in favor of the presence of break(s). In particular, at least one of the SubF, UDmax tests and the sequential procedure and both the BIC and LWZ information criteria select at least one break for all the models studied. It is important to point out that most break dates selected, the midnineties and early 2000s, coincide with the period of two crises in Turkey. The mid-nineties and early 2000s were the periods of financial and economic crises in Turkey, in which the Turkish lira lost its value sharply, interest rates sky-rocketed, and inflation and unemployment began to soar. The Turkish GDP was also reduced significantly in these crisis periods. The break fractions and the corresponding break dates are reported in Tables 4 and 5, respectively. Table 4 presents the results for G–H (1996) and A–K (2007) one structural break cointegration tests in addition to the multiple structural break cointegration tests. The results of the G–H tests show that we 6 The Reer is calculated by Eurostat as the sum of the nominal rate and the trade weighted price or cost deflator. The Reer attempts to show movements in prices or the production cost of domestically produced goods relative to prices or the production cost of goods produced by competitor countries when expressed in common currency. Competitors here for Turkey correspond to EU27.
K. Lopcu et al. / Economic Modelling 35 (2013) 614–621
617
Table 1 Variable definitions. Variables Reer Prodtr Prodeu d_Prod Nfa d_Rir
Real Effective Exchange Rate Relative productivity in Turkey Relative productivity in EU-27 Relative productivity differentials Net foreign assets Relative real interest rate differentials
Explanation
Source
Increase in the Reer of Turkey corresponds to an appreciation of the Turkish lira Average labor productivity ProdT − ProdNT Average labor productivity Prod⁎T − ProdNT⁎ Prodtr − Prodeu (Foreign assets − liabilities to non-resident) / GDP h ∗ i 1þi 1þi 1þπ −1 − 1þπ ∗ −1
Eurostat CBRT, Eurostat, TurkStat Eurostat CBRT, Eurostat, TurkStat CBRT Undersecretariat of Treasury, IFS
i: nominal interest rate, π: inflation rate, ProdT: labor productivity for tradable sector, ProdNT: labor productivity for non-tradable sector and ⁎ denotes the foreign countries. The Reer and all the productivity variables are in natural logarithm.
cannot reject the null of no cointegration for Models 1, 2, 3, and 5. For Model 4, the null is rejected only by the ADF test and for Models 6 and 7 both by the Zt and ADF tests. The A–K one break test, on the other hand, rejects the null of cointegration at least at the 5% significance level, except for Models 1 and 3. Putting it differently, A–K tests cannot reject the null of cointegration at the 1% significance level, except for Model 5. In short, G–H and A–K test results for one break are consistent at all conventional significance levels only for Model 5, and certainly contradict each other for Models 1 and 3. The results are mixed for the other models. Turning to the two-break A–K tests, the null of cointegration for Model 2 cannot be rejected at any conventional significance level, and for Model 7 it is rejected at only the 10% significance level. The tests reject the null of cointegration at 5% and 1% significance levels for Models 4, and 5 and for Models 1, and 3, respectively. Finally, noting that we have evidence of three structural breaks only for Models b cannot reject the null of cointegration for e3 λ 1 and 4, the A–K test, V Model 4, and rejects it only at the 10% level for Model 1. As the final step, we estimate the models for which there is evidence of cointegration and compare the coefficients for the subperiods. Table 5 shows estimated regressions. The estimated slope coefficients are denoted by β1, β2, …, β9. As an example, for Model 5, zt = {d_Prodt, Nfat, d_Rirt}, β1–β3, β4–β6 and β7–β9 show the estimated impact of d_Prod, Nfa and d_Rir on the Reer for regimes 1, 2, and 3, respectively. According to Table 5, for Models 1, 2, 3, and 7, productivity variables are significant and have positive signs, and thereby are consistent with the B–S hypothesis before the structural break in 1994–1995. However, after the 1994–95 structural-break through the 2000s, relative productivities, as well as the other explanatory variables, are not generally successful in explaining changes in the Reer. The only notable exception to this is the positive and significant coefficient for Prodtr after 1994 in Model 6. However, the magnitude of the Prodtr coefficient is small (0.17), and the other productivity coefficients of Model 6 are either insignificant or have the wrong sign. The coefficient for the real interest rate differentials either has the wrong sign or is not significant in all the models included. The effect of Nfa, on the other hand, is always positive and significant until 2001–2002. For Model 2, it is significant until 2006 and for Model 6 for the whole sample period. Table 2 Ng–Perron unit root tests. Deterministic component: constant and trend Variables Reer d_Prod Prodtr Prodeu d_Rir Nfa Critical valuesa
MZα
MZt
MSB
MPt
6 1 1 0 7 2
−4.43 −12.52 −10.56 −9.37 −2.90 −12.51
−1.47 −2.41 −2.21 −2.15 −1.18 −2.48
0.33 0.19 0.21 0.23 0.41 0.20
20.41 7.79 9.02 9.77 30.70 7.39
1% 5% 10%
−23.80 −17.30 −14.20
−3.42 −2.91 −2.62
0.143 0.168 0.185
4.03 5.48 6.67
Lag
a Critical values are taken from Table 1 of Ng and Perron (2001). Lags are selected according to modified Akaike information.
When we evaluate the results in terms of the standard B–S hypothesis (Model 1), we encounter four regimes and the results support the B–S hypothesis only for the first regime. The trend coefficients, on the other hand, indicate a tendency of the TL to depreciate for the first regime and to appreciate for the remaining regimes (after 1994). These results in general hold for the other models as well (e.g. Models 2, 3, 7). The first two of the three breaks indicated in Table 5 overlap with the two important economic and financial crises, while the beginning of the last regime (2006) is synchronized with turmoil in the exchange rate markets in Turkey. Following the 1994 crisis, the Turkish economy was subject to frequent and severe turmoil for a period of seven years due to the effects of the Customs Union with the EU (1996), the Asian and Russian economic crises (1997–1998), and the great Marmara earthquake (1999), in addition to persistent political instability. Interestingly, each of the turmoil was often associated with non-market interventions in the economy, on the one hand, while the country was subject to steady IMF stability programs, on the other hand. As for the period following the 2001 crisis, an overvalued exchange rate has been one of the vehicles of economic stability programs (with implicit and explicit inflation targeting at the heart) to keep the inflation target on tract and to sustain high growth rates financed by deficits of current account. While Alper and Civcir (2012) emphasized that the real exchange rate for the Turkish lira can be explained by both the B–S hypothesis and the BOP approach, our results are not in line with their conclusion. Overall, the results offer only very limited support to the B–S hypothesis. In particular, the results do not support the productivity–Reer relationship for the post 2001 era, contrary to the emphasis placed on the B–S effect by the CBRT, but are compatible with Égert (2005) and Dumrongrittikul (2012). Hence, given the frequent regime changes due to the instabilities in the Turkish economy explained above, the inability of the fairly restrictive B–S and BOP models to explain the tendency of the Turkish lira to appreciate is manifested, even if the regime changes are explicitly accounted for.
4. Conclusion This paper analyzed the movements of the real effective exchange rate for the Turkish economy for 1990:Q1–2011:Q2 using cointegration with multiple structural breaks. According to the CBRT Inflation Report II (2006), the experienced appreciation of the Turkish lira in recent years can be attributed to relative productivity differentials. Seven models related to the B–S hypothesis are constituted by adding other variables such as net foreign assets and real interest rate differentials. The estimated cointegration relationships offer very limited support for the B–S hypothesis, restricted to the pre-2001 and even to the pre1994 periods. Given the theoretical framework, the data and the econometric techniques employed, the results are not in line with the emphasis placed on the B–S hypothesis by the CBRT and Alper and Civcir (2012). However, our findings are compatible with the results of Égert (2005) and Dumrongrittikul (2012) for Turkey. Consequently, the B–S and BOP models are unable to explain the tendency of the Turkish lira to appreciate, even if the regime changes are explicitly accounted for.
618
K. Lopcu et al. / Economic Modelling 35 (2013) 614–621
Table 3 Structural break tests (regime and trend shift model).
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 7
yt = {Reert}
Sub FT
Specification
T1
zt zt zt zt zt zt zt
13.03⁎ 14.48⁎ 16.84⁎ 17.29⁎ 18.12⁎ 20.96⁎⁎ 24.24⁎⁎
= = = = = = =
{d_Prodt} q = 3, m = 5, e = 0.15, xt = 0, p = 2 {d_Prodt, Nfat} q = 4, m = 5, e = 0.15, xt = 0, p = 2 {d_Prodt, d_Rirt} q = 4, m = 5, e = 0.15, xt = 0, p = 2 {Prodtrt, Prodeut} q = 4, m = 5, e = 0.15, xt = 0, p = 1 {d_Prodt, Nfat, d_Rirt} q = 5, m = 5, e = 0.15, xt = 0, p = 1 {Prodtrt, Prodeut, Nfat} q = 5, m = 5, e = 0.15, xt = 0, p = 2 {Prodtrt, Prodeut, d_Rirt} q = 5, m = 5, e = 0.15, xt = 0, p = 1
SEQT (k + 1 / k) 2
3
8.91 9.93 11.56⁎ 15.14⁎⁎ 10.53 11.10 17.08⁎⁎
4
7.44 8.04 9.59# 8.93 7.86 9.11 10.92#
5
6.90 7.29 6.76 12.33⁎⁎ 6.71 8.22 21.07⁎⁎
UD max #
#
5.71 6.59 5.94 12.35⁎⁎ 6.38 12.54⁎⁎ 9.96⁎
13.03 14.48⁎ 16.84⁎ 17.29⁎ 18.12⁎ 20.96⁎⁎ 24.24⁎⁎
k1
2 #
14.14 14.26# 9.71 15.93⁎ 15.51 13.15 16.33
#
14.88 7.82 13.11 20.88⁎⁎ 16.44 9.07 22.64⁎
3
BIC
LWZ
9.29 13.97 8.46 15.62# 13.91 14.78 14.28
3 2 2 3 2 1 2
2 1 2 1 1 1 1
Critical values are from Tables 1 and 3 of Kejriwal and Perron (2010), ⁎⁎, ⁎, #, denote significance levels at 1%, 5% and 10%, respectively. q: number of regressors; m: number of maximum breaks allowed; e: trimming percentage; x: number of I (0) variables. p: number of leads and lags. The leads and lags are equal to 2 where span of data, the number of breaks and the number of variables permit, and one in other cases. lT
′ j Π j þ u∗t : yt ¼ ci þ δi t þ z ′t βi þ ∑ Δz t− j¼−lT
The main possible reason then, in our view, for the tendency of the Turkish lira to appreciate in the period of 2002–2010, despite the intervention of over 80 billion dollars (net purchase) in the exchange rate markets by the CBRT is not capital mobility due to real interest rate differentials but the excess liquidity in the world as a result of monetary expansion and the search for safe harbors for international capital to anchor. The increased current account deficit (CAD) of circa 10% of the GDP in 2010 is one obvious product of increased international capital inflows besides the overvalued home currency. However, such high levels of CAD are putting economic stability, a significant cause of continued capital inflows, in jeopardy. For this reason, for continued economic and political stability, structural problems beneath the high CAD need to be identified and the policies that would bring the CAD to sustainable levels must be implemented. From a historical perspective, in the post financial liberalization era, a sudden stop of capital inflows and/or a sudden rush of outflows have always triggered sharp depreciation of the
Turkish lira, large deviations in announced inflation targets, and compromised growth rates or even an initial reduction in the GDP. In short, bringing the CAD to reasonable levels and implementing structural reforms to keep it at sustainable levels need to be the primary objective of policy makers. It should be noted that the above results using a relatively short span of data need to be interpreted with caution. In particular, the liquidity surplus in the world in the 2000s combined with the reduced risk premium for Turkey as a result of structural reforms and a relatively stable macroeconomic environment might have contributed to the wrong signs we obtained for the real interest rate differentials. Furthermore, a change in the expectations about the future value of the real exchange rate might be the driving force for the tendency of the Turkish lira to appreciate in the 2000s. Future research needs to scrutinize these potential explanations as well as the lack of support for the B–S hypothesis for Turkey in the 2000s.
Table 4 Gregory–Hansen and Arai–Kurozumi cointegration tests (regime and trend shift model). G–H one breakb
yt = {Reert}
Z⁎t zt = {d_Prodt}
zt = {d_Prodt, Nfat}
zt = {d_Prodt, d_Rirt}
zt = {Prodtrt, Prodeut}
zt = {d_Prodt, Nfat, d_Rirt}
zt = {Prodtrt, Prodeut, Nfat}
zt = {Prodtrt, Prodeut, d_Rirt}
Model 1 ⁎⁎ 1% ⁎ 5% # 10% Model 2 ⁎⁎ 1% ⁎ 5% # 10% Model 3 ⁎⁎ 1% ⁎ 5% # 10% Model 4 ⁎⁎ 1% ⁎ 5% # 10% Model 5 ⁎⁎ 1% ⁎ 5% # 10% Model 6 ⁎⁎ 1% ⁎ 5% # 10% Model 7 ⁎⁎ 1% ⁎ 5% # 10%
−3.83 −6.02 −5.50 −5.24 −4.31 −6.45 −5.96 −5.72 −4.38 −6.45 −5.96 −5.72 −5.71 −6.45 −5.96 −5.72 −4.45 −6.89 −6.32 −6.16 −6.47⁎ −6.89 −6.32 −6.16 −6.45⁎ −6.89 −6.32 −6.16
Zα⁎ −26.48 −69.37 −58.58 −53.31 −31.97 −70.65 −68.43 −63.10 −32.68 −70.65 −68.43 −63.10 −49.19 −70.65 −68.43 −63.10 −33.82 −90.84 −78.87 −72.75 −60.03 −90.84 −78.87 −72.75 −57.63 −90.84 −78.87 −72.75
ADF⁎t
A–K one breaka e1 λ b b1 V λ
−4.22 −6.02 −5.50 −5.24 −4.42 −6.45 −5.96 −5.72 −4.46 −6.45 −5.96 −5.72 −6.21⁎ −6.45 −5.96 −5.72 −4.44 −6.89 −6.32 −6.16 −7.23⁎⁎ −6.89 −6.32 −6.16 −6.37⁎
0.060 0.118 0.085 0.069 0.093⁎ 0.100 0.068 0.057 0.040 0.088 0.061 0.049 0.080⁎ 0.100 0.068 0.057 0.115⁎⁎
−6.89 −6.32 −6.16
0.076 0.051 0.041
0.083 0.055 0.045 0.057⁎ 0.074 0.052 0.042 0.055⁎
0.18
0.18
0.25
A–K two breaksa e2 λ b b1 V λ 0.069** 0.052 0.037 0.032 0.029 0.055 0.037 0.031 0.046⁎⁎
0.18
0.044 0.032 0.027 0.039⁎ 0.044 0.032 0.027 0.035⁎
0.24
0.035 0.026 0.022 –
0.18
0.21
0.026# 0.035 0.027 0.023
b2 λ
0.18
0.52
0.18
A–K three breaksa e3 λ b b1 V λ #
b2 λ
b3 λ
0.18
0.52
0.76
0.76
0.023 0.031 0.024 0.021 –
–
–
–
0.18
0.52
–
–
–
–
0.18
0.52
0.19
0.52
0.70
0.18
0.55
0.018 0.027 0.020 0.018 –
–
–
–
–
–
–
–
–
–
0.18
0.52
–
–
–
–
⁎⁎,⁎, #, denote significance levels at 1%, 5% and 10%, respectively. The leads and lags are equal to 2 where span of data, the number of breaks and the number of variables permit, and one in other cases. lT yt ¼ ci þ δi t þ z′ t β i þ ∑ Δz′ t− j Π j þ u∗t : a b
j¼−lT
Critical values are obtained by simulations using 100 steps and 2500 replications. Critical values are taken from Gregory–Hansen (1996). The lags are selected according to the t ratio for the ADF tests.
Table 5 Estimated regressions with multiple structural breaks (regime and trend shift model). c2
c3
c4
δ1
δ2
δ3
δ4
β1
β2
β3
β4
β5
β6
β7
β8
β9
Tb1
Tb2
Tb3
5.26 (0.00) 4.77 (0.00)
4.25 (0.00) 4.23 (0.00)
3.73 (0.00) –
4.38 (0.00) –
−0.02 (0.00) −0.00 (0.98)
0.01 (0.00) 0.01 (0.00)
0.01 (0.00) –
0.01 (0.06) –
0.97 (0.00) 0.18 (0.59)
0.23 (0.18) 0.06 (0.42)
−0.55 (0.21) –
−0.14 (0.22) –
–
–
–
–
–
94:Q2
01:Q2
06:Q2
–
–
–
–
–
94:Q2
–
–
zt = {d_Prodt, Nfat} BIC SEQ 4.90 (0.00)
4.28 (0.00)
4.40 (0.00)
–
−0.00 (0.55)
0.01 (0.00)
0.01 (0.02)
–
0.42 (0.10)
0.24 (0.02)
−0.09 (0.32)
1.15 (0.01)
0.77 (0.00)
−0.17 (0.66)
–
–
–
94:Q2
06:Q2
–
zt = {d_Prodt, d_Rirt} 5.51 (0.00)
4.39 (0.00)
–
–
−0.02 (0.00)
0.01 (0.00)
–
–
1.40 (0.00)
−0.00 (0.99)
−0.12 (0.64)
−0.45 (0.00)
–
–
–
–
–
95:Q4
–
–
zt = {Prodtrt, Prodeut} SEQ BIC 6.38 (0.00) LWZ 6.27 (0.00)
3.92 (0.00) 3.63 (0.00)
6.44 (0.00) –
4.36 (0.00) –
0.01 (0.49) 0.01 (0.44)
0.01 (0.03) 0.01 (0.00)
0.02 (0.00) –
0.01 (0.00) –
0.62 (0.14) 0.20 (0.70)
0.10 (0.56) 0.11 (0.22)
−0.88 (0.05) −1.77 (0.19)
−0.16 (0.29) 0.50 (0.07)
−2.12 (0.04) –
0.05 (0.94) –
−1.94 (0.16) –
0.22 (0.47) –
−
94:Q2
01:Q2
05:Q1
–
95:Q4
–
–
4.44 (0.00)
–
−0.00 (0.57)
0.01 (0.00)
0.01 (0.03)
–
0.30 (0.17)
0.07 (0.60)
−0.08 (0.34)
1.21 (0.00)
0.76 (0.00)
0.19 (0.34)
−0.37 (0.16)
0.10 (0.18)
−0.65 (0.00)
94:Q2
02:Q1
–
–
–
−0.00 (0.87)
0.00 (0.10)
–
–
0.40 (0.36)
0.17 (0.02)
−0.50 (0.71)
0.56 (0.01)
1.11 (0.04)
0.48 (0.00)
–
–
–
94:Q2
–
–
–
–
–
−4.78 (0.00) 0.02 (0.88)
0.35 (0.12) −2.41 (0.03)
−0.51 (0.05) −0.13 (0.88)
−0.34 (0.00) 0.47 (0.03)
–
94:Q2
–
–
–
−0.01 (0.85) 0.02 (0.92)
–
0.01 (0.00)
1.53 (0.00) 0.74 (0.06)
–
–
0.01 (0.00) 0.02 (0.05)
–
3.77 (0.00)
0.01 (0.33) 0.01 (0.22)
−0.37 (0.20)
−0.05 (0.64)
−0.45 (0.00)
94:Q2
01:Q2
–
zt = {d_Prodt, Nfat, d_ Rirt} BIC 4.79 4.18 (0.00) (0.00) zt = {Prodtrt, Prodeut, Nfat} BIC LWZ 4.93 3.55 (0.00) (0.00) zt = {Prodtrt, Prodeut, d_Rirt} LWZ 8.44 4.02 (0.00) (0.00) BIC 6.56 4.12 (0.00) (0.00)
K. Lopcu et al. / Economic Modelling 35 (2013) 614–621
c1
yt = {Reert} zt = {d_Prodt} BIC SEQ
lT
p-Values are in parenthesis. yt ¼ ci þ δi t þ z′ t β i þ ∑ Δz′ t− j Π j þ u∗t : j¼−lT
619
620
K. Lopcu et al. / Economic Modelling 35 (2013) 614–621
Acknowledgement We thank the referees for their guiding critiques. We are grateful to Bülent Ünel, Faik Koray and Mahir Fisunoğlu for helpful comments on earlier versions of this paper. Special thanks to Pier Roberts for editing and proofreading the entire manuscript. All remaining errors are ours.
Appendix A
d_Prod
Reer 5.2
-.1 -.2
5.0 -.3 4.8
-.4 -.5
4.6
-.6 4.4 -.7 4.2
-.8 90
92
94
96
98
00
02
04
06
08
90
10
92
94
96
98
Prodtr
00
02
04
06
08
10
02
04
06
08
10
02
04
06
08
10
Prodeu
.2
.5 .4
.0
.3 -.2 .2 -.4 .1 -.6
.0 -.1
-.8 90
92
94
96
98
00
02
04
06
08
10
90
92
94
96
98
Nfa
00
d_Rir
.5
.8
.4
.6
.3
.4
.2 .2 .1 .0
.0
-.2
-.1 -.2
-.4 90
92
94
96
98
00
02
04
06
08
10
90
92
94
96
98
00
Fig. 1. The real effective exchange rate and the fundamental variables.
References Alberola, E., Naiva, D., 2008. Equilibrium exchange rates in the new EU members: external imbalances vs. real convergence. Rev. Dev. Econ. 12 (3), 605–619. Alberola, E., Cervero, S.G., López, H., Ubide, A., 1999. Global Equilibrium Exchange Rates: Euro, Dollar, “Ins”, “Outs” and Other Major Currencies in a Panel Cointegration Framework. Working Paper, 99/175. International Monetary Fund, Washington DC.
Alper, A.M., Civcir, I., 2012. Can overvaluation prelude to crisis and harm growth in Turkey. J. Polic. Model. 34 (1), 112–131. Apergis, N., 2013. The domestic Balassa–Samuelson effect of inflation for the Greek economy. Appl. Econ. 45 (23), 3288–3294. Arai, Y., Kurozumi, E., 2007. Testing for the null hypothesis of cointegration with a structural break. Econ. Rev. 26 (6), 705–739. Balassa, B., 1964. Purchasing power parity doctrine: a reappraisal. J. Polit. Econ. 72 (6), 584–596.
K. Lopcu et al. / Economic Modelling 35 (2013) 614–621 Bénassy-Quéré, A., Béreau, S., Mignon, V., 2010. On the complementarity of equilibrium exchange-rate approaches. Rev. Int. Econ. 18 (4), 618–632. Camarero, M., 2008. The real exchange rate of the dollar for a panel of OECD countries: Balassa–Samuelson or distribution sector effect. J. Comp. Econ. 36 (4), 620–632. Clark, P.B., MacDonald, R., 2000. Filtering the BEER: A Permanent and Transitory Decomposition. Working Paper, 00/144. International Monetary Fund, Washington DC. Coudert, V., Couharde, C., 2009. Currency misalignments and exchange rate regimes in emerging and developing countries. Rev. Int. Econ. 17 (1), 121–136. Coulibaly, I., Gnimassoun, B., 2013. Optimality of a monetary union: new evidence from exchange rate misalignments in West Africa. Econ. Model. 32, 463–482. Dickey, D., Fuller, W.A., 1979. Distribution of the estimators for autoregressive time series with a unit root. J. Am. Stat. Assoc. 74 (366a), 427–431. Dumrongrittikul, T., 2012. Real exchange rate movements in developed and developing economies: a reinterpretation of the Balassa–Samuelson hypothesis. Econ. Rec. 88 (283), 1–17. Égert, B., 2002. Estimating the impact of the Balassa–Samuelson effect on inflation and the real exchange rate during the transition. Econ. Syst. 26 (1), 1–16. Égert, B., 2005. Balassa–Samuelson meets South Eastern Europe, the CIS and Turkey: a close encounter of the third kind? Eur. J. Comp. Econ. 2 (2), 221–243. Égert, B., Lommatzsch, K., Revil, A.L., 2006. Real exchange rates in small open OECD and transition economies: comparing apples with oranges? J. Bank. Financ. 30 (12), 3393–3406. Gente, K., 2006. The Balassa–Samuelson effect in a developing country. Rev. Dev. Econ. 10 (4), 683–699. Gregory, A.W., Hansen, B.E., 1996. Tests for the cointegration in models with regime and trend shifts. Oxf. Bull. Econ. Stat. 58 (3), 555–560. Hall, S.H., Guo, Q., 2012. Spatial panel data analysis with feasible GLS techniques: an application to the Chinese real exchange rate. Econ. Model. 29 (1), 41–47. Kejriwal, M., 2008. Cointegration with structural breaks: an application to the Feldstein– Horioka puzzle. Nonlinear Dyn. Econom. 12 (1), 1–37.
621
Kejriwal, M., Perron, P., 2010. Testing for multiple structural changes in cointegrated regression models. Am. Stat. Assoc. J. Bus. Econ.Statics 28 (4), 503–522. Kwiatkowski, D., Phillips, P.C.B., Schmidt, P., Shin, Y., 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? J. Econ. 54 (1), 159–178. Lane, P.R., Milesi-Feretti, M.G., 2004. The transfer problem revisited: net foreign assets and real exchange rates. Rev. Econ. Stat. 86 (4), 841–857. Liu, J., Wu, S., Zidek, J.V., 1997. On segmented multivariate regression. Stat. Sin. 7 (2), 497–525. Lommatzsch, K., Tober, S., 2004. What is behind the real appreciation of the accession countries' currencies? An investigation of the PPI-based real exchange rate. Econ. Syst. 28 (4), 383–403. MacDonald, R., Ricci, L., 2003. Estimation of the Equilibrium Real Exchange Rate for South Africa. Working Paper, 3/44. International Monetary Fund, Washington DC. MacDonald, R., Ricci, L., 2005. The real exchange rate and the Balassa–Samuelson effect: the role of the distribution sector. Pac. Econ. Rev. 10 (1), 29–48. MacDonald, R., Wojcik, C., 2004. Catching up: the role of demand, supply and regulated price effects on the real exchange rates of four accession countries. Econ. Transit. 12 (1), 153–179. Méjean, I., 2008. Can firms location decisions counteract the Balassa–Samuelson effect? J. Int. Econ. 76 (2), 139–154. Ng, S., Perron, P., 2001. Lag length selection and the construction of unit root tests with good size and power. Econometrica 69 (6), 1519–1554. Rahn, J., 2003. Bilateral Equilibrium Exchange Rates of EU Accession Countries Against the Euro. Discussion Paper 11. BOFIT. Samuelson, P.A., 1964. Theoretical notes on trade problems. Rev. Econ. Stat. 46 (2), 145–154. The Central Bank of the Republic of Turkey (CBRT), 2006. Inflation Report II (Turkey). Yao, Y.-C., 1988. Estimating the number of change points via Schwartz criterion. Stat. Probab.Lett. 6 (3), 181–189.
Contents lists available at ScienceDirect
Economic Modelling journal homepage: www.elsevier.com/locate/ecmod
Relative productivity increases and the appreciation of the Turkish lira Kenan Lopcu a,⁎, Fikret Dülger b,1, Almıla Burgaç b,1 a b
Department of Econometrics, Çukurova University, Balcalı, Adana 01330, Turkey Department of Economics, Çukurova University, Balcalı, Adana 01330, Turkey
a r t i c l e
i n f o
Article history: Accepted 6 August 2013 JEL classification: C22 E31 F31 Keywords: Balassa–Samuelson hypothesis Real effective exchange rate Relative productivity Cointegration Multiple structural breaks
a b s t r a c t This paper studies the Balassa–Samuelson hypothesis between Turkey and 27 members of the European Union. More specifically, using recently developed cointegration techniques with multiple breaks, we test the relationship between the real effective exchange rate and inter-country differences in the relative productivity of the tradable and non-tradable sectors over the period 1990:Q1–2011:Q2. In recent years, the Central Bank of the Republic of Turkey (CBRT) has emphasized the importance of the B–S hypothesis for Turkey. Our findings, however, suggest that changes in relative productivity have played a limited role in explaining the real effective exchange rate appreciation. In particular, the relationship between the real effective exchange rate and productivity indicated by the Balassa–Samuelson hypothesis is not supported for the post 2001 era in Turkey. © 2013 Elsevier B.V. All rights reserved.
1. Introduction The Balassa–Samuelson (B–S) hypothesis (Balassa, 1964; Samuelson, 1964) identifies inter-country differences in relative productivity of tradable and non-tradable sectors as the main source of long-run changes of real exchange rates (and hence, deviations in purchasing power parity). The Central Bank of the Republic of Turkey (CBRT) in recent years has endorsed the importance of the Balassa–Samuelson hypothesis for Turkey by stating2: … [D]ifferences in productivity and relative price differences appear to have been moving together recently. In conclusion, productivity increases and the differentiation in price increases in the tradable and non-tradable goods are among the factors shaping the development of the real exchange rate in Turkey … [A] close relationship is being observed in the Turkish economy between the relative price differentiation in tradable and non-tradable goods, and the increase in productivity. … [S]ome part of the real appreciation of the New Turkish lira stemmed from the rapid increase in productivity in recent years. The strong stance of the New Turkish lira will continue
⁎ Corresponding author. Tel.: +90 322 338 72 65; fax: +90 322 338 72 83. E-mail addresses: [email protected] (K. Lopcu), [email protected] (F. Dülger), [email protected] (A. Burgaç). 1 Tel.: +90 322 338 72 66. 2 The CBRT has been applying explicit inflation targeting strategies since 2006. Inflation reports issued by the CBRT on a quarterly basis are closely followed by economic agents in Turkey. 0264-9993/$ – see front matter © 2013 Elsevier B.V. All rights reserved. http://dx.doi.org/10.1016/j.econmod.2013.08.005
to be supported via this channel, under the assumption that the increases in productivity would prevail in the upcoming real convergence process with the EU (CBRT, 2006, p. 33–34). Briefly, according to the CBRT, part of the appreciation of the Turkish lira (TL) can be attributed to the productivity differentials between tradable and non-tradable sectors. In particular, in addition to the direct effects of structural reforms undertaken after the 2001 crisis, increased confidence, optimism, and macroeconomic stability in recent years have contributed to the strengthening of the national currency through the relative price differentials between tradable versus non-tradable goods (CBRT, 2006). The main purpose of this study, given the increased consideration of the B–S effect by the CBRT, is to test whether part of the appreciation of the TL is driven by the relatively rapid productivity increases in the tradable sectors in recent years. Hence, the study examines the validity of the B–S hypothesis between Turkey and 27 member countries of the European Union (EU-27), which include the majority of Turkey's main trading partners, for the post-financial liberalization era (i.e., 1990: Q1–2011:Q2), using recently proposed cointegration techniques with multiple structural breaks by Kejriwal (2008). Our analysis offers very limited support for the B–S hypothesis. In particular, the presence of the B–S effect is generally limited to the pre-2001 period, and in some cases even to the pre-1994 period. We do not find any support for the B–S hypothesis from the data for the post-2001 era. Our results are robust to alternative specifications and additional controls, and hence, undermine the increased consideration of the B–S effect by the CBRT.
K. Lopcu et al. / Economic Modelling 35 (2013) 614–621
A large body of literature exists that empirically tests the validity of the B–S hypothesis. Although many of the studies have found that movements in real exchange rates are consistent with the B–S hypothesis, the jury is still out on the effect of relatively rapid productivity changes in the tradable sector (with respect to the non-tradable sector) on the real exchange rate movements. For example, Alberola et al. (1999), Rahn (2003), MacDonald and Wojcik (2004), MacDonald and Ricci (2003, 2005), Lane and Milesi-Ferretti (2004), Gente (2006), Alberola and Navia (2008), Coudert and Couharde (2009), Apergis (2013) and Coulibaly and Gnimassoun (2013), among many others, have documented that the data are consistent with the B–S hypothesis. On the other hand, Méjean (2008), Bénnasy-Quéré et al. (2010) and Hall and Guo (2012) find no evidence for the B–S effect, while Clark and MacDonald (2000), Égert (2002, 2005), Lommatzsch and Tober (2004), Égert et al. (2006), Camarero (2008), and Dumrongrittikul (2012) have produced mixed evidence showing that the B–S hypothesis is supported for some countries included in the sample but not for others. Our paper contributes to this literature by using recently developed cointegration techniques, which allow for multiple breaks in the data, to test the B–S hypothesis. In addition, with the exception of Égert (2005), Dumrongrittikul (2012) and Alper and Civcir (2012), to the best of the authors' knowledge, there has been no study that tests this hypothesis using the Turkish data. Égert (2005) analyzed the domestic B–S effect for a group of EU-acceding, EU-accession (including Turkey) countries and the Common Wealth of Independent States (CIS). This study found that the relative price–relative productivity relationship was mostly insignificant for Turkey. Dumrongrittikul (2012) found a significant but negative relationship between the real exchange rate and relative productivities. Hence, her results for Turkey are inconsistent with the B–S hypothesis. Alper and Civcir (2012), on the other hand, found that the real exchange rate appreciated with positive productivity shocks while it depreciated in response to the increase in net foreign assets. However, the magnitude of the estimated elasticity in that study (around 5) is well above any reasonable a priori expectation and the estimates supported by the empirical literature. This may stem from their choice of a particular proxy for the productivity differentials, the way that the real exchange rate series was constructed, and the econometric technique employed. The rest of this paper is organized as follows. Section 2 introduces the empirical strategy used to test the B–S hypothesis along with the data and their constructions. Section 2.2 reports the empirical findings and studies the robustness of the results. Section 3 concludes the paper. 2. Empirical strategy and data 2.1. Empirical strategy The B–S hypothesis relates the real exchange rate between home and foreign countries to the corresponding relative productivity differentials between these countries. To test this hypothesis, following much of the literature (e.g., Clark and MacDonald, 2000), we begin our analysis by considering the following general form: Reert ¼ ƒðd–Prodt ; Nf at ; d–Rir t Þ where Reer denotes the real effective exchange rate, d_Prod denotes the corresponding relative productivity differentials, Nfa represents the net foreign assets at home, and d_Rir represents the real interest rate differentials between home and foreign economies. An increase in d_Prod leads to an appreciation of the Reer according to the B–S effect.3 The 3 More recently within the New Open Economy Macroeconomic (NOEM) Models framework, productivity increases can lead to a real depreciation due to tradable prices while they can cause a real appreciation due to non-tradable prices (the B–S effect). Thus, the sign of the productivity in reduced form is ambiguous. Similarly, an increase in net foreign assets may or may not lead to an appreciation. See Égert et al. (2006) for a discussion of both of these issues.
615
effect of Nfa, on the other hand, is ambiguous. According to the Balance of Payments (BOP) approach, an increase in Nfa is associated with an appreciation of the Reer in the long term. However, the sign is negative over the medium term if a decrease in Nfa (debt creation) is linked to the appreciation of the Reer. This may apply to countries with rapid growth prospects (Égert et al., 2006). Higher d_Rir, on the other hand, attracts foreign capital and causes the Reer to appreciate via a surplus in the capital account of BOP. In our empirical implementation, we consider the following specific versions of the above model: Model 1 : Reert ¼ cons þ β1 d–Prodt þ ut Model 2 : Reert ¼ cons þ β1 Prodtr t þ β2 Prodeut þ ut Model 3 : Reert ¼ cons þ β1 d–Prodtr t þ β2 Nf at þ ut Model 4 : Reert ¼ cons þ β1 Prodtr t þ β2 Prodeut þ β3 Nf at þ ut Model 5 : Reert ¼ cons þ β1 d–Prodt þ β2 d–Rirt þ ut Model 6 : Reert ¼ cons þ β1 Prodtr t þ β2 Prodeut þ β3 d–Rir t þ ut Model 7 : Reert ¼ cons þ β1 d–Prodt þ β2 Nf at þ β3 d–Rirt þ ut where cons refers to the regression intercept, and Prodtr and Prodeu stand for relative productivities in the tradable and non-tradable sectors in Turkey and EU-27, respectively. We start the empirical analysis by investigating the order of integration of the variables. Next, to assess the stability of the relationship between the real exchange rate and its determinants, we use the tests proposed by Kejriwal and Perron (2010) involving non-stationary but cointegrated variables with multiple structural changes of unknown timing in regression models. If the Kejriwal–Perron tests corroborate the existence of structural breaks, then we verify whether the variables are indeed cointegrated by cointegration tests following Kejriwal (2008), which are based on the extension of the one-break cointegration tests developed by Arai and Kurozumi (2007) (A–K henceforth) with a null of cointegration. Because our series seem to be trending (see Fig. 1), we include a deterministic trend in the unit root as well as cointegration tests. Finally, we estimate the model with breaks to investigate how the relationship between the real exchange rate and its determinants may have altered over time. 2.1.1. Structural break tests Kejriwal and Perron (2010) considered three types of statistics for testing multiple breaks. The first is the sub-Wald test, SubF, of the null hypothesis of no structural break against the alternative hypothesis of k breaks. The second test, a double maximum test called UDmax, checks the null hypothesis of no structural breaks against the alternative of an unknown number of breaks. The third test involves a sequential procedure (SEQ) that analyzes the null hypothesis of k breaks against the alternative hypothesis of k + 1 breaks. A useful strategy, then, is to use significant SubF and UDmax tests to decide if breaks exist and subsequently utilize the sequential procedure to determine the number of breaks (Kejriwal, 2008). As an alternative, the number of breaks can also be determined by using the Bayesian information criterion (BIC) suggested by Yao (1988) and the modified Schwarz criterion proposed by Liu et al. (1997) (LWZ). In this study, stability tests for the relationship between the real effective exchange rate and a number of regressors are performed, using the sequential procedure as well as the information criteria, following Kejriwal (2008), to investigate the existence of breaks in the real effective exchange rate regressions. 2.1.2. Cointegration tests with multiple structural breaks Kejriwal and Perron (2010) showed that the structural change tests they proposed have good size and power properties. In addition, as
616
K. Lopcu et al. / Economic Modelling 35 (2013) 614–621
pointed out in Kejriwal (2008) structural change tests also have power against a purely spurious regression. This means that when the cointegrating relation is unstable, the conventional cointegration tests are biased towards the non-rejection of the null of no cointegration. Hence, cointegration analysis should consider the structural changes. Structural change tests developed by Gregory and Hansen (G–H) (1996) under the null of no cointegration have power against the alternative of a single break, and therefore can have a low power if there is more than a single break. Finally, if the primary concern is cointegration with structural breaks, the null of cointegration is a more natural choice from the viewpoint of conventional hypothesis testing (Kejriwal, 2008). To avoid these problems, Kejriwal (2008) extends the cointegration test with the known or unknown one structural break tests proposed by A–K to analyze multiple structural breaks under the null of cointegration. In the current study, following the work of A–K and Kejriwal, we further augment the A–K model with a deterministic trend as our variables seem to be trending, and allow shifts in the trend as well. The regime and trend shift model used in this study is as follows: yt ¼ ci þ δi t þ z t′ βi þ ut
if T i−1 b t ≤ T i
for i ¼ 1; …k þ 1
where k is the number of breaks, zt is a vector of I (1) regressors, given by zt = zt − 1 + uz, yt is the dependent I (1) variable, and by convention, T0 = 0 and Tk + 1 = T. Augmenting the above regression model to deal with the simultaneity bias, we use the Dynamic Ordinary Least Squares (DOLS), adding the leads and lags of the first differences of the regressors. In this study the leads and lags are equal to 2 where span of data, the number of breaks and the number of variables permit, and one in other cases. yt ¼ ci þ δi t þ z t′ βi þ
lT X
Δz ′t− j Π j þ ut⁎ if T i−1 b t ≤ T i
j¼−lT
non-tradable sector as a whole, a weight is needed for each of the subsector productivity. To calculate the weights, we total the output for all non-tradable sub-sectors separately. Then, we calculate the percentage of the total output attributed to each sub-sector by dividing the total output in that sub-sector into the grand total output of the broad category of non-tradable sector. All the sectoral output and employment series for EU-27 as well as the sectoral output series for Turkey are obtained from the statistical office of the European Union (Eurostat). The employment series for Turkey is from the Turkish Statistical Institute (Turkstat) and the CBRT. The output and employment series for each sub-sector are seasonally adjusted using X-12, before the average productivity for each sub-sector is calculated. The dependent variable in our study is the consumer price index (CPI) based real effective exchange rate (Reer)6 obtained from Eurostat. An increase in the Reer of Turkey corresponds to an appreciation of the TL. The Reer, as well as all the productivity variables, Prodtr, Prodeu and d_Prod are in natural logarithm. The net foreign asset series for Turkey is computed by the difference in the total foreign assets minus the liabilities to non-residents divided by the Gross Domestic Product (GDP), and from the CBRT. The real interest rate differentials between Turkey and G7 are used as a proxy for d_Rir. The annual percentage rate on three month treasury bills (TB) and the CPI based inflation series are used to compute the d_Rir. Both the TB rates and CPI series for Turkey are from the Undersecretariat of the Treasury, while the TB rates and CPI based inflation series for G7 are from the International Financial Statistics (IFS). For Turkey, the inflation series is calculated by the authors using the Turkish CPI series. The variables used and the data sources are also summarized in Table 1. 3. Results
for i ¼ 1; …; k þ 1: The test statistic for k breaks, then, is given by: T −2 b ¼ e λ V k
XT
S t¼1 k
Ω11
2 b λ
:
Where Ω11 is a consistent estimation of the long run variance of u⁎t , t b ¼ ∑u b ¼ Tb1 =T; …; Tbk =T , Sk λ bb represents the residuals bb, and u λ iλ i¼1 iλ from the augmented model above. The break points Tb1 ; …Tbk are obtained by minimizing the sum of squared residuals. The above test statistics are compared with the critical values for multiple breaks generated by the authors, modifying the programs developed for Kejriwal (2008).4 2.2. Data The data set covers the period from 1990:Q1 to 2011:Q2. We take the EU-27 as the benchmark foreign country. Manufacturing represents the tradable sector, while the non-tradable sector includes construction, wholesale and retail trade, and community, social and personal services. Average labor productivity is used as a proxy for the productivity variable suggested by the theoretical model as commonly used in the literature.5 Hence, in order to compute productivity in the tradable sector, the total output in manufacturing is divided by the employment level in the manufacturing sector. To calculate productivity for the 4
The authors would like to thank Mohitosh Kejriwal for making available his Gauss Codes. The productivity measure based on the total hours worked rather than the total employment level would better proxy the theoretical productivity variable. Unfortunately, a measure based on the total hours worked is simply not available. For this reason, the productivity per worker, the GDP per worker or per capita GDP are widely used in the literature. For example, Camarero (2008), Dumrongrittikul (2012), Égert (2002, 2005), Égert et al. (2006), and MacDonald and Wojcik (2004), among others, used the productivity per worker, while Coudert and Couharde (2009), Alper and Civcir (2012), and Alberola and Naiva (2008), employed the GDP per capita and the GDP per worker, respectively. 5
In order to scrutinize the integrating level of variables, the Ng and Perron (2001) tests are employed. To corroborate the results of the Ng– Perron tests, we also employ more conventional unit root tests, namely the Augmented Dickey and Fuller (1979) (ADF) and Kwiatkowski et al. (1992) (KPSS) tests. Table 2 presents the Ng–Perron unit root tests. In Ng–Perron tests, the null of non-stationary is rejected if the test statistic is smaller than the critical value. The results show that for all the variables, the null of non-stationary in levels cannot be rejected at any conventional significance level by any of the Ng–Perron tests. ADF and KPSS tests also corroborate these results, while all three tests provide evidence that the first differences of the variables are stationary. Hence, we conclude that the variables used in the study are integrated order of one, I (1). As a next step, we test the null of no structural change in the longrun relationship. The results obtained are reported in Table 3. For all the models, overall the tests offer evidence in favor of the presence of break(s). In particular, at least one of the SubF, UDmax tests and the sequential procedure and both the BIC and LWZ information criteria select at least one break for all the models studied. It is important to point out that most break dates selected, the midnineties and early 2000s, coincide with the period of two crises in Turkey. The mid-nineties and early 2000s were the periods of financial and economic crises in Turkey, in which the Turkish lira lost its value sharply, interest rates sky-rocketed, and inflation and unemployment began to soar. The Turkish GDP was also reduced significantly in these crisis periods. The break fractions and the corresponding break dates are reported in Tables 4 and 5, respectively. Table 4 presents the results for G–H (1996) and A–K (2007) one structural break cointegration tests in addition to the multiple structural break cointegration tests. The results of the G–H tests show that we 6 The Reer is calculated by Eurostat as the sum of the nominal rate and the trade weighted price or cost deflator. The Reer attempts to show movements in prices or the production cost of domestically produced goods relative to prices or the production cost of goods produced by competitor countries when expressed in common currency. Competitors here for Turkey correspond to EU27.
K. Lopcu et al. / Economic Modelling 35 (2013) 614–621
617
Table 1 Variable definitions. Variables Reer Prodtr Prodeu d_Prod Nfa d_Rir
Real Effective Exchange Rate Relative productivity in Turkey Relative productivity in EU-27 Relative productivity differentials Net foreign assets Relative real interest rate differentials
Explanation
Source
Increase in the Reer of Turkey corresponds to an appreciation of the Turkish lira Average labor productivity ProdT − ProdNT Average labor productivity Prod⁎T − ProdNT⁎ Prodtr − Prodeu (Foreign assets − liabilities to non-resident) / GDP h ∗ i 1þi 1þi 1þπ −1 − 1þπ ∗ −1
Eurostat CBRT, Eurostat, TurkStat Eurostat CBRT, Eurostat, TurkStat CBRT Undersecretariat of Treasury, IFS
i: nominal interest rate, π: inflation rate, ProdT: labor productivity for tradable sector, ProdNT: labor productivity for non-tradable sector and ⁎ denotes the foreign countries. The Reer and all the productivity variables are in natural logarithm.
cannot reject the null of no cointegration for Models 1, 2, 3, and 5. For Model 4, the null is rejected only by the ADF test and for Models 6 and 7 both by the Zt and ADF tests. The A–K one break test, on the other hand, rejects the null of cointegration at least at the 5% significance level, except for Models 1 and 3. Putting it differently, A–K tests cannot reject the null of cointegration at the 1% significance level, except for Model 5. In short, G–H and A–K test results for one break are consistent at all conventional significance levels only for Model 5, and certainly contradict each other for Models 1 and 3. The results are mixed for the other models. Turning to the two-break A–K tests, the null of cointegration for Model 2 cannot be rejected at any conventional significance level, and for Model 7 it is rejected at only the 10% significance level. The tests reject the null of cointegration at 5% and 1% significance levels for Models 4, and 5 and for Models 1, and 3, respectively. Finally, noting that we have evidence of three structural breaks only for Models b cannot reject the null of cointegration for e3 λ 1 and 4, the A–K test, V Model 4, and rejects it only at the 10% level for Model 1. As the final step, we estimate the models for which there is evidence of cointegration and compare the coefficients for the subperiods. Table 5 shows estimated regressions. The estimated slope coefficients are denoted by β1, β2, …, β9. As an example, for Model 5, zt = {d_Prodt, Nfat, d_Rirt}, β1–β3, β4–β6 and β7–β9 show the estimated impact of d_Prod, Nfa and d_Rir on the Reer for regimes 1, 2, and 3, respectively. According to Table 5, for Models 1, 2, 3, and 7, productivity variables are significant and have positive signs, and thereby are consistent with the B–S hypothesis before the structural break in 1994–1995. However, after the 1994–95 structural-break through the 2000s, relative productivities, as well as the other explanatory variables, are not generally successful in explaining changes in the Reer. The only notable exception to this is the positive and significant coefficient for Prodtr after 1994 in Model 6. However, the magnitude of the Prodtr coefficient is small (0.17), and the other productivity coefficients of Model 6 are either insignificant or have the wrong sign. The coefficient for the real interest rate differentials either has the wrong sign or is not significant in all the models included. The effect of Nfa, on the other hand, is always positive and significant until 2001–2002. For Model 2, it is significant until 2006 and for Model 6 for the whole sample period. Table 2 Ng–Perron unit root tests. Deterministic component: constant and trend Variables Reer d_Prod Prodtr Prodeu d_Rir Nfa Critical valuesa
MZα
MZt
MSB
MPt
6 1 1 0 7 2
−4.43 −12.52 −10.56 −9.37 −2.90 −12.51
−1.47 −2.41 −2.21 −2.15 −1.18 −2.48
0.33 0.19 0.21 0.23 0.41 0.20
20.41 7.79 9.02 9.77 30.70 7.39
1% 5% 10%
−23.80 −17.30 −14.20
−3.42 −2.91 −2.62
0.143 0.168 0.185
4.03 5.48 6.67
Lag
a Critical values are taken from Table 1 of Ng and Perron (2001). Lags are selected according to modified Akaike information.
When we evaluate the results in terms of the standard B–S hypothesis (Model 1), we encounter four regimes and the results support the B–S hypothesis only for the first regime. The trend coefficients, on the other hand, indicate a tendency of the TL to depreciate for the first regime and to appreciate for the remaining regimes (after 1994). These results in general hold for the other models as well (e.g. Models 2, 3, 7). The first two of the three breaks indicated in Table 5 overlap with the two important economic and financial crises, while the beginning of the last regime (2006) is synchronized with turmoil in the exchange rate markets in Turkey. Following the 1994 crisis, the Turkish economy was subject to frequent and severe turmoil for a period of seven years due to the effects of the Customs Union with the EU (1996), the Asian and Russian economic crises (1997–1998), and the great Marmara earthquake (1999), in addition to persistent political instability. Interestingly, each of the turmoil was often associated with non-market interventions in the economy, on the one hand, while the country was subject to steady IMF stability programs, on the other hand. As for the period following the 2001 crisis, an overvalued exchange rate has been one of the vehicles of economic stability programs (with implicit and explicit inflation targeting at the heart) to keep the inflation target on tract and to sustain high growth rates financed by deficits of current account. While Alper and Civcir (2012) emphasized that the real exchange rate for the Turkish lira can be explained by both the B–S hypothesis and the BOP approach, our results are not in line with their conclusion. Overall, the results offer only very limited support to the B–S hypothesis. In particular, the results do not support the productivity–Reer relationship for the post 2001 era, contrary to the emphasis placed on the B–S effect by the CBRT, but are compatible with Égert (2005) and Dumrongrittikul (2012). Hence, given the frequent regime changes due to the instabilities in the Turkish economy explained above, the inability of the fairly restrictive B–S and BOP models to explain the tendency of the Turkish lira to appreciate is manifested, even if the regime changes are explicitly accounted for.
4. Conclusion This paper analyzed the movements of the real effective exchange rate for the Turkish economy for 1990:Q1–2011:Q2 using cointegration with multiple structural breaks. According to the CBRT Inflation Report II (2006), the experienced appreciation of the Turkish lira in recent years can be attributed to relative productivity differentials. Seven models related to the B–S hypothesis are constituted by adding other variables such as net foreign assets and real interest rate differentials. The estimated cointegration relationships offer very limited support for the B–S hypothesis, restricted to the pre-2001 and even to the pre1994 periods. Given the theoretical framework, the data and the econometric techniques employed, the results are not in line with the emphasis placed on the B–S hypothesis by the CBRT and Alper and Civcir (2012). However, our findings are compatible with the results of Égert (2005) and Dumrongrittikul (2012) for Turkey. Consequently, the B–S and BOP models are unable to explain the tendency of the Turkish lira to appreciate, even if the regime changes are explicitly accounted for.
618
K. Lopcu et al. / Economic Modelling 35 (2013) 614–621
Table 3 Structural break tests (regime and trend shift model).
Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 7
yt = {Reert}
Sub FT
Specification
T1
zt zt zt zt zt zt zt
13.03⁎ 14.48⁎ 16.84⁎ 17.29⁎ 18.12⁎ 20.96⁎⁎ 24.24⁎⁎
= = = = = = =
{d_Prodt} q = 3, m = 5, e = 0.15, xt = 0, p = 2 {d_Prodt, Nfat} q = 4, m = 5, e = 0.15, xt = 0, p = 2 {d_Prodt, d_Rirt} q = 4, m = 5, e = 0.15, xt = 0, p = 2 {Prodtrt, Prodeut} q = 4, m = 5, e = 0.15, xt = 0, p = 1 {d_Prodt, Nfat, d_Rirt} q = 5, m = 5, e = 0.15, xt = 0, p = 1 {Prodtrt, Prodeut, Nfat} q = 5, m = 5, e = 0.15, xt = 0, p = 2 {Prodtrt, Prodeut, d_Rirt} q = 5, m = 5, e = 0.15, xt = 0, p = 1
SEQT (k + 1 / k) 2
3
8.91 9.93 11.56⁎ 15.14⁎⁎ 10.53 11.10 17.08⁎⁎
4
7.44 8.04 9.59# 8.93 7.86 9.11 10.92#
5
6.90 7.29 6.76 12.33⁎⁎ 6.71 8.22 21.07⁎⁎
UD max #
#
5.71 6.59 5.94 12.35⁎⁎ 6.38 12.54⁎⁎ 9.96⁎
13.03 14.48⁎ 16.84⁎ 17.29⁎ 18.12⁎ 20.96⁎⁎ 24.24⁎⁎
k1
2 #
14.14 14.26# 9.71 15.93⁎ 15.51 13.15 16.33
#
14.88 7.82 13.11 20.88⁎⁎ 16.44 9.07 22.64⁎
3
BIC
LWZ
9.29 13.97 8.46 15.62# 13.91 14.78 14.28
3 2 2 3 2 1 2
2 1 2 1 1 1 1
Critical values are from Tables 1 and 3 of Kejriwal and Perron (2010), ⁎⁎, ⁎, #, denote significance levels at 1%, 5% and 10%, respectively. q: number of regressors; m: number of maximum breaks allowed; e: trimming percentage; x: number of I (0) variables. p: number of leads and lags. The leads and lags are equal to 2 where span of data, the number of breaks and the number of variables permit, and one in other cases. lT
′ j Π j þ u∗t : yt ¼ ci þ δi t þ z ′t βi þ ∑ Δz t− j¼−lT
The main possible reason then, in our view, for the tendency of the Turkish lira to appreciate in the period of 2002–2010, despite the intervention of over 80 billion dollars (net purchase) in the exchange rate markets by the CBRT is not capital mobility due to real interest rate differentials but the excess liquidity in the world as a result of monetary expansion and the search for safe harbors for international capital to anchor. The increased current account deficit (CAD) of circa 10% of the GDP in 2010 is one obvious product of increased international capital inflows besides the overvalued home currency. However, such high levels of CAD are putting economic stability, a significant cause of continued capital inflows, in jeopardy. For this reason, for continued economic and political stability, structural problems beneath the high CAD need to be identified and the policies that would bring the CAD to sustainable levels must be implemented. From a historical perspective, in the post financial liberalization era, a sudden stop of capital inflows and/or a sudden rush of outflows have always triggered sharp depreciation of the
Turkish lira, large deviations in announced inflation targets, and compromised growth rates or even an initial reduction in the GDP. In short, bringing the CAD to reasonable levels and implementing structural reforms to keep it at sustainable levels need to be the primary objective of policy makers. It should be noted that the above results using a relatively short span of data need to be interpreted with caution. In particular, the liquidity surplus in the world in the 2000s combined with the reduced risk premium for Turkey as a result of structural reforms and a relatively stable macroeconomic environment might have contributed to the wrong signs we obtained for the real interest rate differentials. Furthermore, a change in the expectations about the future value of the real exchange rate might be the driving force for the tendency of the Turkish lira to appreciate in the 2000s. Future research needs to scrutinize these potential explanations as well as the lack of support for the B–S hypothesis for Turkey in the 2000s.
Table 4 Gregory–Hansen and Arai–Kurozumi cointegration tests (regime and trend shift model). G–H one breakb
yt = {Reert}
Z⁎t zt = {d_Prodt}
zt = {d_Prodt, Nfat}
zt = {d_Prodt, d_Rirt}
zt = {Prodtrt, Prodeut}
zt = {d_Prodt, Nfat, d_Rirt}
zt = {Prodtrt, Prodeut, Nfat}
zt = {Prodtrt, Prodeut, d_Rirt}
Model 1 ⁎⁎ 1% ⁎ 5% # 10% Model 2 ⁎⁎ 1% ⁎ 5% # 10% Model 3 ⁎⁎ 1% ⁎ 5% # 10% Model 4 ⁎⁎ 1% ⁎ 5% # 10% Model 5 ⁎⁎ 1% ⁎ 5% # 10% Model 6 ⁎⁎ 1% ⁎ 5% # 10% Model 7 ⁎⁎ 1% ⁎ 5% # 10%
−3.83 −6.02 −5.50 −5.24 −4.31 −6.45 −5.96 −5.72 −4.38 −6.45 −5.96 −5.72 −5.71 −6.45 −5.96 −5.72 −4.45 −6.89 −6.32 −6.16 −6.47⁎ −6.89 −6.32 −6.16 −6.45⁎ −6.89 −6.32 −6.16
Zα⁎ −26.48 −69.37 −58.58 −53.31 −31.97 −70.65 −68.43 −63.10 −32.68 −70.65 −68.43 −63.10 −49.19 −70.65 −68.43 −63.10 −33.82 −90.84 −78.87 −72.75 −60.03 −90.84 −78.87 −72.75 −57.63 −90.84 −78.87 −72.75
ADF⁎t
A–K one breaka e1 λ b b1 V λ
−4.22 −6.02 −5.50 −5.24 −4.42 −6.45 −5.96 −5.72 −4.46 −6.45 −5.96 −5.72 −6.21⁎ −6.45 −5.96 −5.72 −4.44 −6.89 −6.32 −6.16 −7.23⁎⁎ −6.89 −6.32 −6.16 −6.37⁎
0.060 0.118 0.085 0.069 0.093⁎ 0.100 0.068 0.057 0.040 0.088 0.061 0.049 0.080⁎ 0.100 0.068 0.057 0.115⁎⁎
−6.89 −6.32 −6.16
0.076 0.051 0.041
0.083 0.055 0.045 0.057⁎ 0.074 0.052 0.042 0.055⁎
0.18
0.18
0.25
A–K two breaksa e2 λ b b1 V λ 0.069** 0.052 0.037 0.032 0.029 0.055 0.037 0.031 0.046⁎⁎
0.18
0.044 0.032 0.027 0.039⁎ 0.044 0.032 0.027 0.035⁎
0.24
0.035 0.026 0.022 –
0.18
0.21
0.026# 0.035 0.027 0.023
b2 λ
0.18
0.52
0.18
A–K three breaksa e3 λ b b1 V λ #
b2 λ
b3 λ
0.18
0.52
0.76
0.76
0.023 0.031 0.024 0.021 –
–
–
–
0.18
0.52
–
–
–
–
0.18
0.52
0.19
0.52
0.70
0.18
0.55
0.018 0.027 0.020 0.018 –
–
–
–
–
–
–
–
–
–
0.18
0.52
–
–
–
–
⁎⁎,⁎, #, denote significance levels at 1%, 5% and 10%, respectively. The leads and lags are equal to 2 where span of data, the number of breaks and the number of variables permit, and one in other cases. lT yt ¼ ci þ δi t þ z′ t β i þ ∑ Δz′ t− j Π j þ u∗t : a b
j¼−lT
Critical values are obtained by simulations using 100 steps and 2500 replications. Critical values are taken from Gregory–Hansen (1996). The lags are selected according to the t ratio for the ADF tests.
Table 5 Estimated regressions with multiple structural breaks (regime and trend shift model). c2
c3
c4
δ1
δ2
δ3
δ4
β1
β2
β3
β4
β5
β6
β7
β8
β9
Tb1
Tb2
Tb3
5.26 (0.00) 4.77 (0.00)
4.25 (0.00) 4.23 (0.00)
3.73 (0.00) –
4.38 (0.00) –
−0.02 (0.00) −0.00 (0.98)
0.01 (0.00) 0.01 (0.00)
0.01 (0.00) –
0.01 (0.06) –
0.97 (0.00) 0.18 (0.59)
0.23 (0.18) 0.06 (0.42)
−0.55 (0.21) –
−0.14 (0.22) –
–
–
–
–
–
94:Q2
01:Q2
06:Q2
–
–
–
–
–
94:Q2
–
–
zt = {d_Prodt, Nfat} BIC SEQ 4.90 (0.00)
4.28 (0.00)
4.40 (0.00)
–
−0.00 (0.55)
0.01 (0.00)
0.01 (0.02)
–
0.42 (0.10)
0.24 (0.02)
−0.09 (0.32)
1.15 (0.01)
0.77 (0.00)
−0.17 (0.66)
–
–
–
94:Q2
06:Q2
–
zt = {d_Prodt, d_Rirt} 5.51 (0.00)
4.39 (0.00)
–
–
−0.02 (0.00)
0.01 (0.00)
–
–
1.40 (0.00)
−0.00 (0.99)
−0.12 (0.64)
−0.45 (0.00)
–
–
–
–
–
95:Q4
–
–
zt = {Prodtrt, Prodeut} SEQ BIC 6.38 (0.00) LWZ 6.27 (0.00)
3.92 (0.00) 3.63 (0.00)
6.44 (0.00) –
4.36 (0.00) –
0.01 (0.49) 0.01 (0.44)
0.01 (0.03) 0.01 (0.00)
0.02 (0.00) –
0.01 (0.00) –
0.62 (0.14) 0.20 (0.70)
0.10 (0.56) 0.11 (0.22)
−0.88 (0.05) −1.77 (0.19)
−0.16 (0.29) 0.50 (0.07)
−2.12 (0.04) –
0.05 (0.94) –
−1.94 (0.16) –
0.22 (0.47) –
−
94:Q2
01:Q2
05:Q1
–
95:Q4
–
–
4.44 (0.00)
–
−0.00 (0.57)
0.01 (0.00)
0.01 (0.03)
–
0.30 (0.17)
0.07 (0.60)
−0.08 (0.34)
1.21 (0.00)
0.76 (0.00)
0.19 (0.34)
−0.37 (0.16)
0.10 (0.18)
−0.65 (0.00)
94:Q2
02:Q1
–
–
–
−0.00 (0.87)
0.00 (0.10)
–
–
0.40 (0.36)
0.17 (0.02)
−0.50 (0.71)
0.56 (0.01)
1.11 (0.04)
0.48 (0.00)
–
–
–
94:Q2
–
–
–
–
–
−4.78 (0.00) 0.02 (0.88)
0.35 (0.12) −2.41 (0.03)
−0.51 (0.05) −0.13 (0.88)
−0.34 (0.00) 0.47 (0.03)
–
94:Q2
–
–
–
−0.01 (0.85) 0.02 (0.92)
–
0.01 (0.00)
1.53 (0.00) 0.74 (0.06)
–
–
0.01 (0.00) 0.02 (0.05)
–
3.77 (0.00)
0.01 (0.33) 0.01 (0.22)
−0.37 (0.20)
−0.05 (0.64)
−0.45 (0.00)
94:Q2
01:Q2
–
zt = {d_Prodt, Nfat, d_ Rirt} BIC 4.79 4.18 (0.00) (0.00) zt = {Prodtrt, Prodeut, Nfat} BIC LWZ 4.93 3.55 (0.00) (0.00) zt = {Prodtrt, Prodeut, d_Rirt} LWZ 8.44 4.02 (0.00) (0.00) BIC 6.56 4.12 (0.00) (0.00)
K. Lopcu et al. / Economic Modelling 35 (2013) 614–621
c1
yt = {Reert} zt = {d_Prodt} BIC SEQ
lT
p-Values are in parenthesis. yt ¼ ci þ δi t þ z′ t β i þ ∑ Δz′ t− j Π j þ u∗t : j¼−lT
619
620
K. Lopcu et al. / Economic Modelling 35 (2013) 614–621
Acknowledgement We thank the referees for their guiding critiques. We are grateful to Bülent Ünel, Faik Koray and Mahir Fisunoğlu for helpful comments on earlier versions of this paper. Special thanks to Pier Roberts for editing and proofreading the entire manuscript. All remaining errors are ours.
Appendix A
d_Prod
Reer 5.2
-.1 -.2
5.0 -.3 4.8
-.4 -.5
4.6
-.6 4.4 -.7 4.2
-.8 90
92
94
96
98
00
02
04
06
08
90
10
92
94
96
98
Prodtr
00
02
04
06
08
10
02
04
06
08
10
02
04
06
08
10
Prodeu
.2
.5 .4
.0
.3 -.2 .2 -.4 .1 -.6
.0 -.1
-.8 90
92
94
96
98
00
02
04
06
08
10
90
92
94
96
98
Nfa
00
d_Rir
.5
.8
.4
.6
.3
.4
.2 .2 .1 .0
.0
-.2
-.1 -.2
-.4 90
92
94
96
98
00
02
04
06
08
10
90
92
94
96
98
00
Fig. 1. The real effective exchange rate and the fundamental variables.
References Alberola, E., Naiva, D., 2008. Equilibrium exchange rates in the new EU members: external imbalances vs. real convergence. Rev. Dev. Econ. 12 (3), 605–619. Alberola, E., Cervero, S.G., López, H., Ubide, A., 1999. Global Equilibrium Exchange Rates: Euro, Dollar, “Ins”, “Outs” and Other Major Currencies in a Panel Cointegration Framework. Working Paper, 99/175. International Monetary Fund, Washington DC.
Alper, A.M., Civcir, I., 2012. Can overvaluation prelude to crisis and harm growth in Turkey. J. Polic. Model. 34 (1), 112–131. Apergis, N., 2013. The domestic Balassa–Samuelson effect of inflation for the Greek economy. Appl. Econ. 45 (23), 3288–3294. Arai, Y., Kurozumi, E., 2007. Testing for the null hypothesis of cointegration with a structural break. Econ. Rev. 26 (6), 705–739. Balassa, B., 1964. Purchasing power parity doctrine: a reappraisal. J. Polit. Econ. 72 (6), 584–596.
K. Lopcu et al. / Economic Modelling 35 (2013) 614–621 Bénassy-Quéré, A., Béreau, S., Mignon, V., 2010. On the complementarity of equilibrium exchange-rate approaches. Rev. Int. Econ. 18 (4), 618–632. Camarero, M., 2008. The real exchange rate of the dollar for a panel of OECD countries: Balassa–Samuelson or distribution sector effect. J. Comp. Econ. 36 (4), 620–632. Clark, P.B., MacDonald, R., 2000. Filtering the BEER: A Permanent and Transitory Decomposition. Working Paper, 00/144. International Monetary Fund, Washington DC. Coudert, V., Couharde, C., 2009. Currency misalignments and exchange rate regimes in emerging and developing countries. Rev. Int. Econ. 17 (1), 121–136. Coulibaly, I., Gnimassoun, B., 2013. Optimality of a monetary union: new evidence from exchange rate misalignments in West Africa. Econ. Model. 32, 463–482. Dickey, D., Fuller, W.A., 1979. Distribution of the estimators for autoregressive time series with a unit root. J. Am. Stat. Assoc. 74 (366a), 427–431. Dumrongrittikul, T., 2012. Real exchange rate movements in developed and developing economies: a reinterpretation of the Balassa–Samuelson hypothesis. Econ. Rec. 88 (283), 1–17. Égert, B., 2002. Estimating the impact of the Balassa–Samuelson effect on inflation and the real exchange rate during the transition. Econ. Syst. 26 (1), 1–16. Égert, B., 2005. Balassa–Samuelson meets South Eastern Europe, the CIS and Turkey: a close encounter of the third kind? Eur. J. Comp. Econ. 2 (2), 221–243. Égert, B., Lommatzsch, K., Revil, A.L., 2006. Real exchange rates in small open OECD and transition economies: comparing apples with oranges? J. Bank. Financ. 30 (12), 3393–3406. Gente, K., 2006. The Balassa–Samuelson effect in a developing country. Rev. Dev. Econ. 10 (4), 683–699. Gregory, A.W., Hansen, B.E., 1996. Tests for the cointegration in models with regime and trend shifts. Oxf. Bull. Econ. Stat. 58 (3), 555–560. Hall, S.H., Guo, Q., 2012. Spatial panel data analysis with feasible GLS techniques: an application to the Chinese real exchange rate. Econ. Model. 29 (1), 41–47. Kejriwal, M., 2008. Cointegration with structural breaks: an application to the Feldstein– Horioka puzzle. Nonlinear Dyn. Econom. 12 (1), 1–37.
621
Kejriwal, M., Perron, P., 2010. Testing for multiple structural changes in cointegrated regression models. Am. Stat. Assoc. J. Bus. Econ.Statics 28 (4), 503–522. Kwiatkowski, D., Phillips, P.C.B., Schmidt, P., Shin, Y., 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? J. Econ. 54 (1), 159–178. Lane, P.R., Milesi-Feretti, M.G., 2004. The transfer problem revisited: net foreign assets and real exchange rates. Rev. Econ. Stat. 86 (4), 841–857. Liu, J., Wu, S., Zidek, J.V., 1997. On segmented multivariate regression. Stat. Sin. 7 (2), 497–525. Lommatzsch, K., Tober, S., 2004. What is behind the real appreciation of the accession countries' currencies? An investigation of the PPI-based real exchange rate. Econ. Syst. 28 (4), 383–403. MacDonald, R., Ricci, L., 2003. Estimation of the Equilibrium Real Exchange Rate for South Africa. Working Paper, 3/44. International Monetary Fund, Washington DC. MacDonald, R., Ricci, L., 2005. The real exchange rate and the Balassa–Samuelson effect: the role of the distribution sector. Pac. Econ. Rev. 10 (1), 29–48. MacDonald, R., Wojcik, C., 2004. Catching up: the role of demand, supply and regulated price effects on the real exchange rates of four accession countries. Econ. Transit. 12 (1), 153–179. Méjean, I., 2008. Can firms location decisions counteract the Balassa–Samuelson effect? J. Int. Econ. 76 (2), 139–154. Ng, S., Perron, P., 2001. Lag length selection and the construction of unit root tests with good size and power. Econometrica 69 (6), 1519–1554. Rahn, J., 2003. Bilateral Equilibrium Exchange Rates of EU Accession Countries Against the Euro. Discussion Paper 11. BOFIT. Samuelson, P.A., 1964. Theoretical notes on trade problems. Rev. Econ. Stat. 46 (2), 145–154. The Central Bank of the Republic of Turkey (CBRT), 2006. Inflation Report II (Turkey). Yao, Y.-C., 1988. Estimating the number of change points via Schwartz criterion. Stat. Probab.Lett. 6 (3), 181–189.