Is the Cyprus Pound Real Effective Exchange Rate Misaligned? A BEER Approach

Abstract

This paper investigates whether the real effective exchange rate of the Cyprus pound is misaligned by generating measures of the equilibrium rate using the Behavioral Equilibrium Exchange Rate (BEER) approach. Several measures of the equilibrium exchange rate were derived and used to check for the existence of exchange rate misalignment. The results suggest that, during the 1990s, the actual real effective exchange rate and the various equilibrium measures generated move closely together and there is no evidence of any significant and persistent misalignment. However, the empirical evidence suggests persistent overvaluation during the 1980s.

Keywords:

• Equilibrium exchange rates
• misalignment
• Cyprus pound

Acknowledgements

Financial support from the Central Bank of Cyprus is gratefully acknowledged. The authors would like to thank three anonymous referees for extensive and constructive comments that greatly improved this paper. They also thank seminar participants at the Central Bank of Cyprus and at the Olayan School of Business of the American University of Beirut for useful comments. The usual disclaimer applies.

Notes

¹A detailed description of exchange rate policy in Cyprus can be found in Kyriacou (2002).

²As Gylfason (2002) argues, irrespective of nominal exchange rate arrangements, the real exchange rate always floats.

³For an interesting analysis that combines the FEER and BEER methodologies, see Égert & Lahreche-Revil (2003).

⁴The economic fundamentals employed in this study have been used extensively in the literature. In particular, the specification is based on MacDonald (1999) and Clark & MacDonald (1999) since the BEER approach has been popularized by them. Many other studies have used all or a subset of the variables used in this paper (see, for example, Maeso-Fernandez et al., 2002; and Paya et al., 2003). Stein & Lim (2002) concluded that the real effective exchange rate is indeed affected by a set of economic fundamentals but which fundamentals to include in this set is controversial.

⁵For a recent analysis of the effects of net foreign assets on real exchange rates see Lane & Milesi-Ferretti (2004).

⁶MacDonald & Ricci (2004) have suggested that if the proxy variable used for the Balassa–Samuelson effect is not perfect, the real interest rate differential may help capture empirically the effect of productivity differentials.

⁷All the variables are included in the empirical analysis, even though some of them influence the real exchange rate via similar channels. The rationale of such an approach is that each variable individually may be unable to capture all the complex effects that are operating; and because each variable may also affect the real exchange rate in different ways and via different channels. Looking at correlation coefficients we can see that most of them are very small, indicating that multicollinearity is not a serious concern. Out of a total of 15 correlation coefficients: 11 were less than 0.4 (in absolute value); one was 0.48; two were 0.6; and only one was high (equal to 0.81). The latter refers to the correlation coefficient between the terms of trade and the real price of oil, and the high correlation was, thus, expected. However, as will become clearer later on, both of these variables were statistically significant in our estimated models, perhaps indicating that their high correlation may not be a problem.

⁸In the literature one of two variables is usually used to proxy the Balassa–Samuelson effect: the relative price of non-tradables or a direct measure of sectoral productivity. For reasons given below, this paper focuses on the former, in line with many other studies (see for example, Clark & MacDonald, 1999; Alberola et al., 1999; Hansen & Roeger, 2000; and Paya et al., 2003). A problem of using a measure of productivity as a proxy for the Balassa–Samuelson effect is that productivity data are usually available at annual frequency. There is the additional problem of choosing one of the many alternative productivity measures that one can use (see for example, Maeso-Fernandez et al., 2004, for a brief discussion). Furthermore, given that both supply and demand factors affect the relative price of non-traded goods, the latter may be a more appropriate variable to use. In addition, the use of the relative price of non-traded goods may be more appropriate since changes in this variable may come from changes in factor endowments as suggested by Bhagwati (1984), who argued that as a country develops, it adopts more capital intensive techniques that lead to an increase in the wage-interest rate ratio and, therefore, to an increase in the relative price of (labor-intensive) non-traded goods. Thus, even though the relative price of non-traded goods may not be an ideal proxy for the Balassa–Samuelson effect, it is an important variable influencing the real exchange rate. For a critique of the use of such a variable (and, in particular, of the CPI/WPI ratio) as a proxy for the Balassa–Samuelson effect see Égert et al., 2004).

⁹All trade-weighted foreign variables used in the construction of variables LPNT and LTOT were geometric averages. The most important trading partners and competitors of Cyprus were used according to weights (w _i ) calculated by the IMF (and supplied to the authors). These weights are used by the IMF in constructing the real effective exchange rate of the Cyprus pound. The main trade partners and competitors of Cyprus were: Austria, Belgium, Finland, France, Germany, Greece, Italy, Japan, Korea, The Netherlands, Norway, Singapore, Spain, Sweden, Switzerland, Turkey, UK, USA and Taiwan. In those cases where no data were available for a particular country, that country was excluded and the weights recalculated.

¹⁰In the first draft of the paper, due to data unavailability, Cyprus's WPI reflected the combination of series codes 63 and 63A (WPI for home goods).

¹¹It is generally recognized that it is difficult to measure NFA. This variable should reflect the international investment position of a country (see Lane & Milesi-Ferretti, 2001). No such data were available in the case of Cyprus for most of the period under investigation. Furthermore, the current account balance (whose cumulative summation can be used as a measure of the stock of NFA) was not used since it was only available with annual frequency. Thus, the paper used the IFS series 31 N as a proxy for Cyprus's stock of NFA.

¹²We are grateful to Costas Xiouros for assistance in constructing this variable.

¹³Arithmetic (rather than geometric) averages were used to calculate the foreign benchmark used in the construction of variables RID and FBAL, since the real interest rates and fiscal balances of some countries were sometimes positive and sometimes negative.

¹⁴Measuring inflation in this way is desirable since it recognizes that market participants take into account inflationary expectations. Similar specifications have been used by, amongst others, Edison & Pauls (1993) and MacDonald (1999).

¹⁵Choosing the appropriate lag length with the Akaike Information criterion using a maximum lag order of 6 for the augmentation of the ADF test suggests that NFA is I (1).

¹⁶Diagnostic testing of the individual equations in the VAR of order 2 suggests that they pass the usual battery of tests (serial correlation, functional form, normality, heteroscedasticity and autoregressive conditional heteroscedasticity—see the notes to Table 3 for more details on the actual tests used) with the following exceptions (at the 95% level): normality in the equations of LTOT, NFA, LPOIL and RID; and functional form in the equations of LTOT and RID. Increasing the lag length of the VAR did not solve most of these problems and in certain cases caused additional econometric problems.

¹⁷We have also tested for cointegration without seasonal dummies in the specification. In this case, both of Johansen's statistics clearly show that there is one cointegrating vector (at the 95% level) and the correct specification is restricted intercepts and no trends in the VAR.

¹⁸Caporale & Chui (1999) have also used a similar approach, in that they have tested for cointegration using the Johansen approach and then estimated the long-run relationship using an ARDL model. Pesaran & Shin (1999) have shown that after appropriate augmentation of the order of the ARDL model, the ARDL-based estimators of the long-run coefficients are super-consistent and valid inferences can be made using standard normal asymptotic theory. They have also shown that appropriate choice of the order of the ARDL model also tackles any endogenous regressor problems. The use of the ARDL model also avoids the problem of identifying the cointegrating vector (see Nagayasu, 1999).

¹⁹From the results reported in Table 3, the long-run coefficient (LR) for each variable was obtained as: LR=∑ _i γ _i /(1−∑ _i β _i ), where γ _i refers to the coefficients of current and lagged values of each explanatory variable and β _i to the coefficients of lagged values of the dependent variable.

²⁰ARDL models were also estimated using the ARDL option in Microfit 4.1 (see Pesaran & Pesaran. 1997). Under this option, the appropriate order of the ARDL model is chosen on the basis of model selection criteria (such as AIC and SBC). A very similar long-run relationship (not reported) was obtained using this option (and basing model selection on the AIC). Furthermore, the ARDL models estimated in this way suggest that the lagged dependent variable is an important factor influencing the real exchange rate. This is consistent with the results presented in Table 3. This is, of course, not surprising given the unit root behavior of the real effective exchange rate.

²¹We have also estimated an unrestricted error correction model (UECM) and derived the long-run equilibrium relationship based on the results. The UECM is a simple reparameterization of the ARDL model and, thus, is expected to generate the same long-run relationship. In small samples, however, the results are expected to be somewhat different since the parsimonious models that result from a ‘general-to-specific’ approach may not be identical. In the case of two variables (y and x), the UECM takes the following form:

where Δ is the first-difference operator. In our case, of course, the UECM included all variables described in the third section. In estimating the UECM we have followed a ‘general-to-specific’ approach with a maximum lag of 4, and the results of the parsimonious model are reported in Table 4. The model passes all the usual diagnostic tests. From the results presented in Table 4, the long-run coefficients (LR) are estimated as: LR=−(θ/φ), where θ is the coefficient of each lagged explanatory variable in level form (e.g. LTOT (−1)) and φ is the coefficient of LREER(−1)). The resulting long-run relationship that defines the equilibrium real exchange rate was very similar to that obtained from the ARDL model (as expected). It is given as: ERER3=4.12+1.03 LTOT−0.17 LPOIL−0.014 RID+1.9 FBAL. All coefficients are statistically significant at either the 95% or 90% levels. The plots of the actual and equilibrium real exchange rates and the conclusions regarding misalignments were also very similar and are not repeated here.

²²The finding that the coefficient of RID is very small and incorrectly signed may be due to the fact that the nominal interest rate in Cyprus was fairly constant (under government control) over the sample period and also due to the existence of restrictions on capital flows.

²³We have also estimated a static OLS regression with the current levels of all variables (including a constant and a time trend). The equilibrium exchange rates generated based on these results suggest that there is no serious misalignment as the actual and equilibrium real exchange rates move closely together throughout the sample period. However, these results are not presented since: (i) a residual-based cointegration test (in the sense of Engle & Granger, 1987) cannot be applied since no critical values are available given the number of regressors in the model; (ii) the model fails many diagnostic tests (as expected); (iii) the bias of the static OLS estimator is quite large in small samples; and (iv) statistical inference is quite difficult since the t-values follow non-standard distributions.