Journal of Banking & Finance 30 (2006) 31473169 www.elsevier.

com/locate/jbf

Deviations from purchasing power parity under dierent exchange rate regimes: Do they revert and, if so, how?
Lucio Sarno
a

a,*

, Giorgio Valente

Warwick Business School, University of Warwick and Centre for Economic Policy Research (CEPR), United Kingdom b Department of Finance, Chinese University of Hong Kong, Hong Kong Received 26 July 2005; accepted 9 December 2005 Available online 12 July 2006

Abstract We propose an empirical model for deviations from long-run purchasing power parity (PPP) that simultaneously accounts for three key features: (i) adjustment toward PPP may occur via nominal exchange rates and relative prices at dierent speeds; (ii) dierent exchange rate regimes may generate regime shifts in the structural dynamics of PPP deviations; (iii) nonlinear reversion toward PPP in response to shocks. This empirical framework encompasses and synthesizes much previous empirical research. Using over a century of data for the G5 countries, we provide evidence that long-run PPP holds, the relative importance of nominal exchange rates and prices in restoring PPP varies over time and across dierent exchange rate regimes, and reversion to PPP occurs nonlinearly, at a speed that is fairly consistent with the nominal rigidities suggested by conventional open economy models. 2006 Elsevier B.V. All rights reserved.
JEL classication: F31 Keywords: Real exchange rate; Purchasing power parity; Exchange rate regimes; Nonlinearity

Corresponding author. Tel.: +44 2476 528219; fax: +44 2476 572871. E-mail address: lucio.sarno@warwick.ac.uk (L. Sarno).

0378-4266/$ - see front matter 2006 Elsevier B.V. All rights reserved. doi:10.1016/j.jbankn.2005.12.007

3148

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169

1. Introduction 1.1. Overview The purchasing power parity (PPP) hypothesis states that national price levels should be equal when expressed in a common currency. A large literature in international nance has examined empirically the validity of PPP over the long run either by testing whether nominal exchange rates and relative prices move together in the long run or by testing whether the real exchange rate has a tendency to revert to a stable equilibrium level over time. The latter approach is motivated by the fact that the real exchange rate may be dened as the nominal exchange rate adjusted for relative national price levels, and therefore variations in the real exchange rate represent deviations from PPP, which must be stationary if long-run PPP holds (see the surveys of Froot and Rogo, 1995; Rogo, 1996; Sarno and Taylor, 2002; Taylor and Taylor, 2004; Sarno, 2005). Although long-run PPP is such a simple proposition about exchange rate behavior, it has attracted the attention of researchers for decades because it has important economic implications on several fronts. In particular, the degree of persistence in the real exchange rate can be used to infer what the principal impulses driving exchange rate movements are. For example, if the real exchange rate is highly persistent or close to a random walk, then the shocks are likely to be real-side, principally technology shocks, whereas if it is not very persistent, then the shocks must be principally to aggregate demand, such as, for example, innovations to monetary policy (Rogo, 1996). Further, from a theoretical perspective, if PPP is not a valid long-run international parity condition, this casts doubts on the predictions of much open economy macroeconomics that is based on the assumption of long-run PPP. Indeed, the implications of open economy dynamic models are very sensitive to the presence or absence of a unit root in the real exchange rate (e.g. Lane, 2001; Sarno, 2001). Finally, estimates of PPP exchange rates are often used for practical purposes such as determining the degree of misalignment of the nominal exchange rate and the appropriate policy response, the setting of exchange rate parities, and the international comparison of national income levels. These practical uses of the PPP concept would obviously be of limited use if PPP deviations contain a unit root. Regardless of the great interest in this area of research, manifested by the large number of papers on PPP published over the last few decades, and despite the increasing quality of data sets utilized and the econometric techniques employed, the validity of long-run PPP and the properties of PPP deviations remain the subject of ongoing controversies. Specifically, earlier cointegration studies generally reported the absence of signicant mean reversion of the real exchange rate for the recent oating experience (e.g. Mark, 1990), but were supportive of reversion toward PPP for the gold standard period (Diebold et al., 1991), for the interwar oat (Taylor and McMahon, 1988), for the 1950s USCanadian oat (McNown and Wallace, 1989), and for the exchange rates of high-ination countries (Choudhry et al., 1991). Some applied work on long-run PPP among the major industrialized economies has, however, been more favorable to the long-run PPP hypothesis for the recent oat (e.g. Corbae and Ouliaris, 1988; Cheung and Lai, 1993, 1994, 1998; Frankel and Rose, 1996; Coe and Serletis, 2002; Serletis and Gogas, 2004). One well-documented explanation for the inability to nd evidence of long-run PPP is the low power of conventional unit root and cointegration tests with a sample span corresponding to the length of the recent oat (Froot and Rogo, 1995; Lothian and Taylor,

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169

3149

1997). Researchers have sought to overcome the power problem either through long-span studies (e.g. Kim, 1990; Lothian and Taylor, 1996; Taylor, 2002) or through panel unit root studies (e.g. Abuaf and Jorion, 1990; Frankel and Rose, 1996; OConnell, 1998; Papell, 1998; Sarno and Taylor, 1998; Taylor and Sarno, 1998) or through time-series models that account for the possibility of nonlinear mean reversion toward PPP (e.g. Michael et al., 1997; Obstfeld and Taylor, 1997; Baum et al., 2001; Coakley and Fuertes, 2001; Taylor, 2001; Taylor et al., 2001; Nakagawa, 2002; Imbs et al., 2003). However, whether or not the long-span or panel-data studies do in fact answer the question whether PPP holds in the long run remains contentious (e.g. Engel, 1999, 2000).1 With respect to studies using nonlinear models of the real exchange rate, they unanimously support the validity of longrun PPP (see Michael et al., 1997; Obstfeld and Taylor, 1997; Baum et al., 2001; Coakley and Fuertes, 2001; OConnell and Wei, 2001; Taylor et al., 2001; Imbs et al., 2003). However, since they are based on univariate regressions for the real exchange rate and do not allow for the possibility that the dynamics of PPP deviations be aected by monetary and exchange rate regimes, to date these studies have not shed light on whether nominal exchange rates or prices drive the adjustment toward the PPP equilibrium and do not investigate the role of dierent nominal regimes on the behavior of PPP deviations. 1.2. Questions addressed and approach In light of the evidence provided by this literature, there remain at least three important issues in this area of research. First, it is still controversial whether long-run PPP is validated by the data, even among the major G5 economies. Second, it is debatable whether, when PPP is validated by the data, adjustment toward the long-run equilibrium level dened by PPP is driven primarily by the exchange rate, by relative prices, or by both of them. Third, it is puzzling why the majority of studies nd empirical estimates of the persistence of PPP deviations that are too high to be explained by conventional nominal rigidities and cannot be reconciled with the high short-term volatility of real exchange rates. Our empirical analysis is devoted to shed light on all of these three issues. We start from noting three features that we view as potentially important in designing a suitable model for the deviations from PPP. The rst feature is that the model needs to allow for the fact that adjustment toward PPP is likely to occur at dierent speeds via nominal exchange rates and prices. The vast majority of empirical studies on PPP is based on univariate representations of the real exchange rate.2 This approach is only valid if certain common factor restrictions in the unknown data generating process linking exchange rates and prices are satised. We nd that these common factor restrictions are generally rejected by the data, implying a loss of power in testing the null hypothesis that the real exchange rate
1 As far as the long-span studies are concerned, the long samples required to generate a reasonable level of statistical power with standard tests may potentially be inappropriate because of dierences in real exchange rate behavior both across dierent historical periods and nominal exchange rate regimes (e.g. Baxter and Stockman, 1989; Taylor, 2002). As for panel-data studies, these provide mixed evidence. While, for example, Abuaf and Jorion (1990), Frankel and Rose (1996) and Taylor and Sarno (1998) nd results favorable to long-run PPP, the empirical evidence reported by OConnell (1998) and Papell (1998) rejects PPP. 2 Exceptions which have considered this issue include, inter alia, Edison (1987), Edison et al. (1997), Goldfajn and Valdes (1999), Coakley and Fuertes (2000), Engel and Morley (2002) and Cheung et al. (2004). See also Coakley et al. (2005).

3150

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169

is nonstationary (or that PPP is invalid) in conventional testing procedures. Employing a model which does not impose these restrictions increases test power, whilst allowing us to shed light on the relative importance of nominal exchange rates and prices in restoring the PPP equilibrium. The second desirable feature is that the model allows explicitly for the possibility that dierent monetary and exchange rate regimes generate regime shifts in the structural dynamics of PPP deviations, especially when using long-spans of data.3 The third feature is that the model might be nonlinear, in accordance with the growing evidence that exchange rate dynamics displays statistically and economically important nonlinearities. In particular, some recent literature provides evidence that real exchange rates display nonlinear mean reversion toward PPP for a discussion of nonlinear PPP models, see the relevant sections in the surveys of Taylor and Taylor (2004) and Sarno (2005). These nonlinear models generally imply an equilibrium level of the real exchange rate in the neighborhood of which the behavior of the log-level of the real exchange rate is close to a random walk, becoming increasingly mean reverting with the absolute size of the deviation from equilibrium. This is consistent with a number of recent theoretical contributions on the nature of real exchange rate dynamics in the presence of international arbitrage costs (e.g. Dumas, 1992; Sercu et al., 1995). Also, the impulse response functions implied by these real exchange rate models generate, because of the nonlinearity, half-lives of shocks to real exchange rates that vary both with the size of the shock and with the initial conditions. By taking account of statistically signicant nonlinearities, the speed of real exchange rate adjustment implied by these models is found to be much faster than typically recorded in the relevant literature based on data for the recent oat (e.g. Taylor et al., 2001). In this paper we extend the long-span data used by Obstfeld and Taylor (2004) and apply a general modelling methodology in which regime changes in the data generating process are explicitly allowed for, focusing on the G5 countries. Taylor (2002) has examined these time series employing single-equation and panel linear econometric methods. His empirical evidence is generally supportive of long-run PPP and shows that PPP deviations have similar half-lives across the dierent monetary regimes of the last century, but much larger shocks to the real exchange rate process have characterized oating exchange rate regimes relative to xed exchange rate regimes. Our paper builds on and extends the work of Taylor (2002) and other scholars who have contributed to the literature on longspan real exchange rate behavior including Edison (1987), Kim (1990), Lothian and Taylor (1996), Michael et al. (1997) in several directions. We re-examine whether long-run PPP is valid and measure the relative importance of exchange rates and relative prices in driving the adjustment toward the long-run PPP equilibrium across dierent exchange rate regimes, including the Gold Standard, the Bretton Woods period, and the recent oat. With over a century of data and dierent exchange rate arrangements over our sample period, we tentatively hypothesize that during oating exchange rate periods, the nominal exchange rate may be relatively more important in restoring departures from long-run equilibrium, while the relative price should restore long-run equilibrium during xed

Indeed, a large literature has provided mounting evidence that the distribution of nominal and real exchange rate changes is well described by a mixture of normal distributions and that a regime-switching model may be a good characterization of exchange rate behavior (e.g. see Engel and Hamilton, 1990; LeBaron, 1992; Engel, 1994; Engel and Kim, 1999; Clarida et al., 2003; Dueker and Neely, 2005; Sarno and Valente, 2005).

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169

3151

exchange rate periods.4 Over the sample period examined, the economic history of the countries involved has seen a number of fundamental changes in monetary and exchange rate regimes, institutional structure and policy targets which, in addition to the continuous evolution of the nancial system and various nominal and real shocks, represent serious potential pitfalls to researchers attempting to nd an empirical model of the deviations from PPP that is stable over the full sample. The time-invariant, linear framework generally adopted by the literature is not suitable for the purposes of this paper. We investigate whether allowing for regime-switching in the underlying data-generating process is an adequate characterization that is capable of capturing the impact of the dierent monetary regimes of the last century on the dynamics of exchange rates and relative prices. This is done through estimating a fairly general Markov-switching vector error correction model (MS-VECM) which characterizes the dynamic relationship between exchange rates and relative prices allowing for regime shifts in the parameters as well as for nonlinear reversion toward long-run PPP in each regime. 1.3. Main results The results are supportive of long-run PPP for each of the four major exchange rates examined and of our basic conjecture: we nd that during xed exchange rate regimes, relative prices adjust to restore deviations from long-run equilibrium, while nominal exchange rates bear most of the burden of adjustment during exible exchange rate regimes. The estimated transition probabilities are consistent with the general result that the relative importance of exchange rates and relative prices in restoring the long-run equilibrium level of the exchange rate varies over time and is aected by the nominal exchange rate arrangement in operation. Further, the estimated half-lives of the regime-dependent nonlinear exchange rate models are sensibly dierent for xed and oating regimes. During xed exchange rate regimes, shocks to the PPP equilibrium relationship may be very persistent, implying half liveson average across the exchange rates considered from over ve years for large real exchange rate shocks of 20% to almost ten years for small shocks of 1%. However, the corresponding half-lives during oating exchange rate regimes are drastically shorter, since the nominal exchange rate is allowed to operate and contribute to restoring PPP. In fact, shocks will last for less than one year on average for 20% shocks, with a maximum of 1.78 years for the smallest shocks of 1%. We conclude that the PPP puzzles recorded in the literature using data for the recent oat may have been generated to some extent by the fact that none of the studies in the literature has explicitly allowed for all of the three features of PPP deviations described above, either imposing common factor restrictions to examine the stochastic properties

It is important to note that the latter statement is subject to the caveat that during xed exchange rate regimes devaluations may be used by policy makers to induce adjustments in the real exchange rate. Conversely, during a oating regime, policy makers may use heavily foreign exchange market intervention for the purpose of reducing the variability of the exchange rate (e.g. Sarno and Taylor, 2001; Calvo and Reinhart, 2002). Given these caveats, it may be that, in fact, irregular shifts in nominal exchange rates may be responsible for inducing reversion to PPP in xed exchange rate regimes or, vice versa, the role of nominal exchange rates in driving the adjustment to PPP during oating regimes is not more important than the role of relative prices. Hence, the question of whether and how PPP deviations dissipate remains an empirical issue.

3152

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169

of the real exchange rate rather than the full dynamic interaction of nominal exchange rates and relative prices, or ignoring the impact of dierent monetary and exchange rate regimes on PPP deviations, or else neglecting the possibility of signicant nonlinearities in the adjustment to PPP. Allowing for all of these features simultaneously allows us to build a generalization of existing empirical frameworks that sheds new light on the behavior of PPP deviations. The properties of PPP deviations during oating exchange regimes implied by our model appear to be fairly consistent with standard models of open economy macroeconomics and with their dynamic properties under conventional nominal rigidities.5 It is only during xed exchange rate regimes, when the burden of adjustment toward PPP relies exclusively on relative prices, that we observe remarkably long half lives of PPP shocks. 1.4. Organization The remainder of the paper is set as follows. Section 2 denes long-run PPP and describes the dynamic relationship between exchange rates and relative prices, outlining the importance of common factor restrictions in this context. In Section 3 we set out the econometrics of nonlinear Markov-switching multivariate models as applied to nonstationary processes and cointegrated systems. In Section 4 we describe our data set, while in the following section we report and discuss our empirical results. A nal section concludes. 2. Long-run PPP and the dynamic interaction between exchange rates and prices Dening st as the logarithm of the nominal exchange rate (expressed as domestic price of foreign currency) and pt as the logarithm of the ratio of domestic to foreign prices, then q t s t pt 1 may be seen as the deviation from PPP or the logarithm of the real exchange rate. Absolute long-run PPP would allow qt 5 0 in the short run, but it would require qt = 0 in the long run. A less strict version of long-run PPP postulates that qt may have a non-zero mean but it has to be a realization of a stationary process. If both st and pt have a stationary, invertible, non-deterministic ARMA representation after dierencing once (i.e. st, pt $ I(1)), this denition of long-run PPP implies that st and pt move together in the long run and exhibit a common stochastic trend, cointegrating with one cointegrating vector b 0 = [1 1]. The relationship between st and pt can be described by a cointegrating vector autoregression (VAR) of the form: gLy t et ; 2

Notably Chari et al. (2002) nd that theoretical autocorrelations for the real exchange rate at the quarterly frequency are in the range between about 0.4 and 0.8, corresponding to half lives between less than one quarter and about three quarters. Indeed, the inability of this general equilibrium model to match the persistence of real exchange rates between 3 and 5 years recorded by much empirical literature constitutes one of the main criticisms against this important theoretical paper. However, our result that the half-life in a oating rate regime can be shorter than one year reduces the relevance of this criticism.

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169

3153

where yt is a two-dimensional observed time series vector, yt = [st,pt] 0 ; g(L) is a suitable p-th order, 2 2 matrix polynomial in the lag operator L; et = [e1t,e2t] 0 is a vector of white noise processes with covariance matrix R, et $ NID(0,R).6 By the Granger Representation Theorem (Engle and Granger, 1987), st and pt must possess a vector error correction model (VECM) representation where the deviation from PPP, qt (the real exchange rate) plays the part of the equilibrium error: Dy t CLDy t1 Py t1 et ;
0

where C(L) is a 2 2 matrix polynomial; and the long-run impact matrix P = ab , where a and b are 2 1 vectors, with b denoting the cointegrating vector (assumed to be [1, 1] under PPP) and a the vector of weights on the cointegrating vector in each of the two equations of the VECM. Note that g(L) in Eq. (2) is unrestricted; hence, C(L) and a are also unrestricted. The existence of one cointegrating relationship between st and pt implies that the rank of P equals unity. To understand the relationship between tests of PPP using the VECM(3) and unit root tests based on the augmented DickeyFuller (ADF) auxiliary regression or variants of it often applied by researchers to the real exchange rate, let us start from noting that the latter focus on the roots of qt = b 0 yt rather than the properties of yt itself. This implicitly imposes common factor restrictions, as it can be seen by pre-multiplying (3) by b 0 to obtain b0 Dy t b0 CLDy t1 b0 ab0 y t1 b0 et or 1 GLLDqt qqt1 wt where the coecient q = b 0 a; G(L) is a scalar polynomial in L, and wt b0 CL GLb0 Dy t1 b0 et : 6 5 4

Eq. (5) is the conventional ADF regression used by a number of researchers for testing the null hypothesis of a unit root in the real exchange rate. It is now apparent that the disturbance term wt may contain valuable information for two reasons. First, unless b 0 C(L) = G(L)b 0 , lags of D yt enter wt. Second, if pt is not weakly exogenous and responds to PPP disequilibria, then b 0 et may be explained partly by the current value of p. Both of these reasons imply a loss of information from testing for PPP by testing for a unit root in the real exchange rate qt using an ADF regression of the form (5) rather than by analyzing the full VECM linking the nominal exchange rate st and the relative price pt, namely Eq. (3). In turn, as demonstrated for several dierent data generating processes by Kremers et al. (1992), this loss of information leads to a substantial loss of power of the ADF test in rejecting the null hypothesis of a unit root, which, in this context, would bias the outcome of the test toward concluding that the real exchange rate is nonstationary and longrun PPP is invalid. The discussion in this section suggests yet another reason, unexplored by the literature to date, why unit root tests have low power in the context of testing for PPP. Several
For ease of exposition, in this section we do not allow for a constant term in the cointegrating VAR and do not allow for regime-switching in the VAR. However, it is important to note that none of the points made below is dependent on this simplication. In fact, allowance for a constant term or generalization to a nonlinear VAR would simply increase the number of common factor restrictions on which we focus in this section.
6

3154

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169

researchers have noted that conventional unit root tests have low power in rejecting the null of real exchange rate nonstationarity for the recent oat since 1973 or so, because the sample size is too short to yield sucient test power (see Froot and Rogo, 1995). This has led several researchers to develop panel unit root tests which exploit the cross-correlation in exchange rate data in an attempt to increase test power (since at least Hakkio, 1984; Abuaf and Jorion, 1990). However, these tests also impose the same common factor restrictions as single-equation unit root tests. Indeed, the number of restrictions increases with the size of the panel. If the common factor restrictions are not satised, for any given sample size, the power of (single-equation and panel) unit root tests may be substantially lower than the power of tests based on the full VECM because unit root tests applied to the real exchange rate eectively ignore valuable information by assuming error rather than structural dynamics. We investigate the empirical relevance of these issues by using the VECM representation of st and pt to test for PPP and to shed light on the relative importance of the nominal exchange rate and relative prices in restoring the PPP equilibrium level across dierent nominal regimes since the 19th century. Further, in order to take seriously into account the possibility of regime shifts over long-spans of data, we use a generalization of the standard linear VECM (3) which is capable of allowing all of the VECM parameters to change over time and to identify the various regimes that characterize the long sample periods examined. 3. Nonlinear Markov-switching error correction Consider the following M-regime pth order Markov-switching vector autoregression (MS(M)-VAR(p)) which allows for regime shifts in the intercept term: gLy t mzt et ; 7 where yt is a K-dimensional observed time series vector, yt = [y1t, y2t, . . . , yKt] 0 ; m(zt) is a Kdimensional column vector of regime-dependent intercept terms, m(zt) = [m1(zt), m2(zt), . . . , mK(zt)] 0 ; g(L) is a suitable pth order, K K matrix polynomial in the lag operator L; et = [e1t, e2t, . . . , eKt] 0 is a K-dimensional vector of white noise processes with covariance matrix R, et $ NID(0,R). The regime-generating process is assumed to be an ergodic Markov chain with a nite number of states zt 2 {1, . . . , M} governed by the transition probPM abilities pij = Pr(zt+1 = jjzt = i), and j1 pij 1 8i; j 2 f1; . . . ; Mg. A standard case in economics and nance is that yt is nonstationary but rst-dierence stationary, i.e. yt $ I(1). Then, given yt $ I(1), there may be up to K 1 linearly independent cointegrating relationships, which represent the long-run equilibrium of the system (Granger, 1986; Engle and Granger, 1987). If indeed there is cointegration, the cointegrated MS-VAR (7) implies a Markov-switching vector error correction model or MSVECM of the form: Dy t mzt CLDy t1 Py t1 et ; 8

where C(L) is a suitable K K matrix polynomial in the lag operator L of order P 1, such p Pp p that Ci ji1 gj for i = 1, . . . ,p 1 are matrices of parameters, and P i1 Pi I is the long-run impact matrix whose rank r determines the number of cointegrating vectors (e.g. Johansen, 1988, 1991). An MS-VECM of the form (8) can be estimated using an expectation-maximization (EM) algorithm for maximum likelihood estimation (Dempster et al., 1977; Hamilton, 1993; Krolzig, 1997; Kim and Nelson, 1999).

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169

3155

Although, for expositional purposes, we have outlined the MS-VECM framework for the case of regime shifts in the intercept alone, shifts may be allowed for elsewhere. The present application focuses on a multivariate model comprising, for each country analyzed, the spot exchange rate and the relative price (hence yt = [st,pt] 0 ), for which, following the reasoning of Section 2, one unique, independent cointegrating relationship, represented by the deviation from PPP qt, should exist. As discussed in Section 5 below, in our empirical work, after considerable experimentation, we selected a specication of the MS-VECM which allows for regime shifts in the intercept, the variance-covariance matrix and the whole set of parameters (including the autoregressive component, C(L) and cointegration matrix, P). In each regime, moreover, the VECM allows for nonlinear reversion toward PPP of the same exponential form proposed by the growing literature on nonlinear real exchange rate dynamics (e.g. Michael et al., 1997; Baum et al., 2001; Taylor et al., 2001). This model, the nonlinear Markov-Switching-Intercept-Autoregressive-Heteroskedastic-VECM or MSIAH-VECM, may be written as follows: Dy t mzt CLzt Dy t1 Pzt y t1 U ut ;
0

where P(zt) = a(zt)b , ut $ NIID(0,R(zt)), and zt 2 {1, . . . , M}. The nonlinear function q2 U 1 exp r2tdt is an exponential function, where qtd = b 0 ytd and the integer z
q

d > 0 is a delay parameter. U() is bounded between zero and unity, U : R ! 0; 1, has the properties U[0] = 0 and limx!1U[x] = 1, and is symmetrically inverse-bell shaped around zero. These properties are attractive in the present modelling context because they allow symmetric adjustment toward PPP for deviations above and below the PPP equilibrium level (zero), as predicted by theories of real exchange rate determination under transactions costs (e.g. Dumas, 1992; Sercu et al., 1995). The size of the deviation from PPP qtd, standardized by dividing it by its regime-dependent variance, r2 zt , then determines q the speed of mean reversion toward PPP, with lower PPP deviations being more persistent than large deviations (of either sign), again as predicted by general equilibrium models of real exchange rate determination under transactions costs. In our empirical work, consistent with previous research in this context (e.g. Obstfeld and Taylor, 1997; Taylor, 2001; Taylor et al., 2001), we set d = 1, since economic intuition suggests a presumption in favor of smaller values of the delay parameter d rather than larger values, in that it is hard to imagine why there should be very long lags before the real exchange rate begins to adjust in response to a shock.7 4. Data The data set used in this study comprises annual observations for the nominal exchange rate (domestic price of foreign currency) and the price levels based on the consumer price index (CPI) or the gross domestic product (GDP) deator, depending on availability relative to the US for each of the G5 countries. Our data set is obtained from updating the

Although the transition probabilities are exogenous in our model, the speed of reversion to PPP is a function of the absolute size of the deviation from PPP itself. An alternative formulation would involve endogeneizing the switching probabilities, but this is not straightforward in a rich model of the type considered in this paper (e.g. see Kim et al., 2003).

3156

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169

relevant time series in the data set constructed by Obstfeld and Taylor (2004) using the International Financial Statistics of the International Monetary Fund.8 From these data we calculate the logarithm of the nominal exchange rate, st, the relative price, pt, and the real exchange rate, qt = st pt, as dened in Eq. (1). The sample spans from the late 19th or early 20th century to the late 20th century and thus covers a variety of international monetary arrangements, including the classical Gold Standard, the Bretton Woods era, the modern oat and, for some countries, the Exchange Rate Mechanism of the European Monetary System. The exact sample period for each country is as follows: 18702000 for the UK; 18801998 for France and Germany; 18852000 for Japan. The start date was in each case dictated by data availability, whereas the end date is 2000 except for France and Germany, which joined the European Monetary Union and replaced their respective national currencies with the euro in January 1999.9 5. Empirical results 5.1. Unit roots, cointegration, and common factor restrictions10 As a preliminary exercise, we test for unit root behavior of the nominal exchange rate (st) and the relative price (pt) time series, employing various unit root tests. These tests indicate that both st and pt are realizations from a stochastic process integrated of order one. We then tested the common factor restrictions required for the VAR to reduce to a univariate autoregression for the real exchange rate, using likelihood ratio (LR) tests. The results indicate that, in each case, the restrictions are strongly rejected, providing the case for using unrestricted VAR and VECM estimation for testing PPP rather than real exchange rate autoregressions. Thus, we employed the Johansen (1988, 1991) maximum likelihood procedure in a VAR for yt = [st, pt] 0 , which does not impose any common factor restrictions. However, these cointegration tests do not detect a (proportional) cointegrating relationship between exchange rates and relative prices, leading to the conclusion that long-run PPP does not hold for any of the countries examined over the full sample. There are a variety of reasons for the failure to nd a unique cointegrating relationship between economic time series where one would normally be expected on the basis of economic theory (see Siklos and Granger, 1997; and the references therein). In particular, it is

We are thankful to Alan Taylor for graciously providing the data. We focus on the G5 countries rather than on the broader set of countries analyzed, for example, in Taylor (2002), because the G5 countries are the subset of the world economy for which PPP is more likely to be a reasonable approximation to the long-run behavior of the real exchange rate. In particular, given the relatively fast technology diusion that characterizes the G5 economies compared to other economies, HarrodBalassa Samuelson eects (shocks to productivity dierentials) are unlikely to last forever. Even if some evidence exists that the productivity dierential is an important determinant of the real exchange rate, this set of countries is the one where long-run PPP is more likely to apply (see Rogo, 1996; Sarno and Taylor, 2002). The only exception might be Japan, occasionally referred to as a possible example of the HarrodBalassaSamuelson eect in operation among major economies, since it has been on average the fastest-growing economy for much of the post World War II period (see Rogo, 1996). This interpretation seems consistent with the ndings in Taylor (2002), where a deterministic trend is often needed for a variety of countries outside the G5 to capture trending departures from long-run PPP. 10 The results discussed in this sub-section are not reported to conserve space but are available upon request.
9

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169

3157

well known that the presence of structural breaks or regime shifts may perversely aect cointegration tests, which appears to be especially likely when using long-spans of data. Such shifts may signicantly alter the dynamic relationship that may exist between the variables, and tests of the long term behavior of these variables should account for them. Hence, as a check of adequacy of the models as well as an additional motivation for employing a more general modelling framework, we test for the stability of the cointegrating rank over time. In particular, we carry out two recently developed tests that are specically designed to test for stability within a cointegrating framework: the Hansen and Johansen (1999) test of the null hypothesis of stability of the eigenvalues of the stochastic matrix; and the test statistic developed by Quintos (1997) for testing the null hypothesis that the cointegrating rank is constant among breakpoints. The tests results suggest that the null hypothesis of temporal stability of the cointegrating rank is strongly rejected, further motivating the use of a regime-switching framework, to which we now turn. 5.2. Regime shifts and model selection We investigate the presence of regime shifts by applying the bottomup procedure designed to detect Markovian shifts in order to select the most adequate characterization of a nonlinear M-regime pth order MS-VECM for Dyt. Essentially, the bottomup procedure consists of starting with a simple but statistically reliable Markov-switching model by restricting the eects of regime shifts on a limited number of parameters and checking the model against alternatives. In such a procedure, most of the structure contained in the data is not attributed to regime shifts, but explained by observable variables, consistent with the general-to-specic approach to econometric modelling (Krolzig, 1997).11 The VARMA representations of the time series suggests in each case that the number of regimes is three. However, for any MS-VECM estimated the implicit assumption that the regime shifts aect alternatively only the variance-covariance matrix, the intercept term, or the autoregressive component of the VECM was found to be inappropriate. In fact, we checked the relevance of all these components by using three likelihood ratio (LR) tests of the type suggested by Krolzig (1997, pp. 135136). The results, reported in the rst three rows of Table 1 (LR1, LR2 and LR3, respectively), indicate strong rejections of the null of no regime dependence, clearly suggesting that an MS-VECM that allows for shifts in the intercept, the variance-covariance matrix and the autoregressive component, namely an

In this paper we employ the testing procedure suggested by Krolzig (1997) to determine the number of regimes, but we should point out several relevant caveats. It is well known that the determination of the number of regimes in the data is a very cumbersome problem, which does not have a robust solution to date. Specically, the presence of nuisance parameters gives the likelihood surface sucient freedom so that the nding of some statistically signicant parameters could simply be due to sampling variation. The scores associated with the parameters of interest under the alternative hypothesis may be identically equal to zero under the null. Several testing procedures have been proposed, none of which is without problems. It is important to note here that the regularity conditions under which the tests suggested by Krolzig (1997) are valid may be violated here. Therefore, the distribution of the likelihood ratio tests may dier from the adjusted v2 distribution. We rely on the asymptotic distribution in our empirical work and do not attempt to resolve the econometric problems surrounding the above issues; this would be an ambitious aim beyond the scope of this paper, given the complex nature of our cointegrated Markov model and the presence of a nonlinear transition function in the error correction term. For discussions of the problems related to hypothesis testing in this context, see Hansen (1992) and Garcia (1998).

11

3158

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169

Table 1 Bottomup identication procedure France M LR1 LR2 LR3 LR4 LR5 LR6 3 1.86 1068 9.90 105 1.68 1010 2.58 101 1.82 1020 5.38 1076 Germany 3 3.73 1046 3.28 108 1.01 1014 1.25 101 9.55 1051 1.68 1074 Japan 3 2.90 1022 7.76 107 2.14 103 2.74 101 1.33 1033 3.67 1057 UK 3 1.19 1040 9.85 104 1.01 103 3.44 101 1.41 1012 5.75 1053

Notes: LR1 is a test statistic of the null hypothesis of no regime dependent variance-covariance matrix (i.e. MSIA(M)-VECM(p) versus MSIAH(M)-VECM(p)). LR2 is a test statistic of the null hypothesis of no regime dependent intercept (i.e. MSAH(M)-VECM(p) versus MSIAH(M )-VECM(p)). LR3 is a test statistic of the null hypothesis of no regime dependent autoregressive component (i.e. MSIH(M)-VECM(p) versus MSIAH(M)VECM(p)). LR4 tests the null hypothesis that the model with one lag in the autoregressive component is equivalent to another with a higher autoregressive order (i.e. MSIAH(M)-VECM(1) versus MSIAH(M)VECM(3)). LR1,LR2,LR3,LR4 are constructed as 2(lnL* lnL), where L* and L represent the unconstrained and the constrained maximum likelihood, respectively. These tests are distributed as v2(g) where g is the number of restrictions. LR5 is the likelihood ratio test for the null hypothesis that the MSIAH-VECM(1) with 3 regimes is equivalent to the MSIAH-VECM(1) with 2 regimes. LR6 is a linearity test for the null hypothesis that the selected MSIAH-VECM(1) is equivalent to a linear VECM(1). p-values relative to the LR5 and LR6 tests are calculated as in Ang and Bekaert (1998). For all test statistics only p-values are reported.

MSIAH-VECM(p), is the most appropriate model within its class in the present application. Further, in the same spirit of the tests previously computed, we carry out a further LR test in order to select the most parsimonious MSIAH-VECM for characterizing the dynamic relationship between spot exchange rates and relative prices. In particular, we test the null of MSIAH-VECM(1) against the alternative of a higher-order MSIAH-VECM(3), since three lags is the maximum lag length suggested by conventional information criteria. The fourth row of Table 1 (LR4) indicates that we are not able to reject this null hypothesis at standard signicance levels, hence concluding that one lag is appropriate for each MSIAH-VECM estimated. In order to discriminate between models allowing for two regimes against models governed by a higher number of regimes we execute two further likelihood ratio tests, one (LR5) for the null hypothesis that the MSIAH-VECM(1) with 3 regimes is equivalent to the MSIAH-VECM(1) with 2 regimes, and another (LR6) as a linearity test of the null hypothesis that the selected MSIAH-VECM(1) is equivalent to a linear VECM(1). The results reported in Table 1 (fth row) show very large test statistics and the corresponding p-values suggest that three regimes may be appropriate in all cases to describe the dynamics of nominal exchange rates and relative prices. Finally, the linearity tests, reported in the last row of Table 1, indicate in each case the rejection of the linear VECM(1) in favor of its nonlinear, Markov-switching counterpart. Hence, the nal result of this procedure identies for all countries an MSIAH-VECM governed by three dierent regimes and one lag that can be written as follows: Dy t mzt CLzt Dy t1 Pzt y t1 U ut ; where P(zt) = a(zt)b 0 , U 1
q2 exp r2t1t , q z

10

xt $ NIID(0,R(zt)) and zt = 1,2,3. We esti-

mate the MSIAH-VECM in Eq. (10) using an EM algorithm for maximum likelihood

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169 Table 2 Tests for common factor restrictions: nonlinear MSIAH-VECM LRcf(z = 1) France Germany Japan UK 1.51 102 4.86 101 1.42 101 2.42 104 LRcf(z = 2) 3.58 102 4.91 102 3.14 102 2.26 102 LRcf(z = 3) 9.21 103 9.90 103 4.56 102 5.75 102

3159

Notes: LRcf(z = i) is the likelihood ratio test for the null hypothesis that b 0 C(Ljz = i) = G(Ljz = i)b 0 conditional on regime i = 1,2,3. Under the null hypothesis that the common factor restrictions are valid, the test statistic is distributed as a v2(d), where d is the number of restrictions. Figures reported are p-values.

(Dempster et al., 1977), for each of the countries under investigation. This algorithm has been found to be fairly accurate in estimating MS-VECMs up to 4 states when the number of observations available are larger than 100 (see Krolzig, 1997).12 Since unrestricted estimation of the MSIAH-VECM yielded parameter estimates for b 0 close to [1, 1], in the nal estimation of the model we set the long-run cointegrating vector b 0 = [1, 1], consistent with long-run PPP, to increase the accuracy of the remaining parameters estimated. The estimation yielded fairly plausible estimates of the coecients, including the regime-dependent adjustment coecients in a(zt), which are generally found to be statistically signicantly dierent from zero in at least one regime. Before turning to the empirical results from estimating the MSIAH-VECMs, however, we carry out tests of regime-dependent tests of the common factor restrictions which would make the MSIAH-VECM(10) reduce to a univariate regime-switching real exchange rate autoregression. The likelihood ratio tests results, reported in Table 2, indicate rejections of the common factor restrictions in at least two regimes for each MSIAH-VECM with fairly low p-values, strengthening the case for testing for PPP in the context of a full VECM rather than in the context of univariate models of the real exchange rate. 5.3. MSIAH-VECM estimation results Table 3 reports results relating to the classication of regimes implied by our estimated MSIAH-VECM. In particular we calculated the average smoothed probabilities over the Gold Standard (18701914), interwar (19151944), Bretton Woods (19451972) and recent oat (19732000) periods in order to see whether our preferred MSIAH-VECMs were able to correctly identify dierent nominal exchange rate regimes. These results indicate several interesting patterns. First, the estimated MSIAH-VECMs are able to disentangle quite precisely the dierent nominal regimes on the basis of the calculation of the regime classication measure (RCM) proposed by Ang and Bekaert (1998, 2002). The RCM, which is dened between 0 and 100, is easily interpretable as follows: when the RCM is 0, identication is perfectly accurate, whereas when the RCM is 100

This EM algorithm carries out iteration steps that are repeated until convergence is ensured on the basis of several criteria based on the relative change in the log-likelihood and the parameters; see Krolzig (1997), Section 6.4, pp. 109110. An alternative way to carry out estimation for this class of regime-switching models involves using Markov Chain Monte Carlo (MCMC) methods (e.g. see Chib, 2001; and the references therein). We chose to estimate the MS-VECM by maximum likelihood methods given their simplicity and faster convergence properties.

12

3160

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169

Table 3 Regime classication measure (RCM) 18701914 France 1 p 2 p p 3 RMC(3) 1 p p 2 3 p RMC(3) p 1 2 p 3 p RMC(3) 1 p 2 p p 3 RMC(3) 0.002 0.994 0.004 8.15 102 0.012 0.988 0.000 6.07 103 0.037 0.689 0.273 3.19 106 0.038 0.928 0.034 9.58 101 19151944 0.734 0.199 0.067 2.74 106 0.333 0.587 0.080 3.49 103 0.242 0.467 0.291 8.78 101 0.207 0.262 0.531 5.98 101 19451972 0.215 0.587 0.197 1.44 102 0.072 0.807 0.122 9.90 103 0.215 0.709 0.076 5.48 105 0.216 0.656 0.128 3.14 101 19732000 0.002 0.002 0.996 4.13 107 0.001 0.004 0.995 6.00 108 0.001 0.000 0.999 1.59 106

Germany

Japan

0.346 0.060 0.595 9.96 101 PT Notes: j is the average smoothed probability relative to regime j = 1,..,3 calculated as j 1 j pj;t =T j , where Tj p p is the number of observations in sub-period j, and pj,t is the smoothed (ex-post) regime probability relative to regime j = 1,..,M at time t. RCM(M) is the regime classication measure proposed by Ang and Bekaert (2002). PT j Q The statistic is calculated as RCMM 100M 2 T1j i1 M pj;t . j1

UK

no information about the regimes is revealed.13 It is comforting that our results in Table 3 indicate that in all cases the RCM is very small and in some cases close to zero. Second, the MSIAH-VECMs are able to identify regimes with xed exchange rates from regimes with exible exchange rates. In fact, in all cases we can see that the regime with the highest probability under the Gold Standard period is the same exhibiting the highest probability under the Bretton Woods period.14 However, we note that the results for the interwar period are less clear-cut in that the average smoothed probabilities display more variation across countries, unlike other periods. This is understandable since the interwar period is somewhat mixed because it includes both oating and xed (the return to the Gold Standard) regimes. It is worth noting that our regime classication is implied by the estimated MSIAHVECMs and is, therefore, data and model driven. This diers from analyses of real

13 Since the true regime is a Bernoulli random variable, the RCM is essentially a sample estimate of its variance (Ang and Bekaert, 2002). 14 While two of the regimes clearly identify xed and oating exchange rate regimes, the third regime does not lend itself to any easy economic interpretation. However, inspection of the transition probabilities suggests that this regime is largely characterized by outliers in the data, which do not t easily in any of the rst two regimes. Indeed, estimation of a more parsimonious two-regime VECM yielded similar results and economic implications as the three-regime VECM, but in a less clear-cut fashion. This is because the allowance for more than two regimes, which was suggested by the bottom-up procedure, is able to remove the outliers in the data and to identify more precisely the xed and oating exchange rate regimes, ultimately yielding a statistically superior model. We could not discriminate whether the outliers are extreme observations or inaccurate observations, but decided to estimated the model using all of the available data rather than arbitrarily take out the outliers.

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169 Table 4 Regime-dependent adjustment coecients Fixed regime aDst France Germany Japan UK 0.002 (0.004) 0.004 (0.069) 0.012 (0.009) 0.002 (0.004) aDpt 0.227 (0.091) 0.028 (0.012) 0.073 (0.028) 0.058 (0.017) Flexible regime aDst 0.302 (0.129) 0.340 (0.079) 0.137 (0.038) 0.659 (0.144) aDpt 0.041 (0.033) 0.024 (0.132) 0.161 (0.055) 0.067 (0.072)

3161

Notes: aDst denotes the estimated regime-conditional error-correction coecient relative to the spot exchange rate equation. aDpt denotes the estimated regime-conditional error-correction coecient relative to the relative price equation, obtained from estimating the nonlinear MSIAH-VECM(10). Values in parentheses are asymptotic standard errors.

exchange rate behavior across exchange rate regimes which are typically carried out using regime classications chosen a priori on the basis of the historical experience of the countries examined (e.g. Taylor, 2002). The interpretation of regimes also benets from examining jointly the results in Table 3 with the ones in Table 4, where we report the regime-dependent error correction coecients for xed and oating regimes. The results in Table 4 suggest that during xed exchange rate regimes the error correction coecients in the exchange rate equations are not signicantly dierent from zero at conventional signicance levels, while the error correction coecients in the relative price equations are statistically signicant and correctly signed. This implies that during xed exchange rate regimes the relative price responds to deviations from PPP in a way to restore the long-run exchange rate equilibrium. Vice versa during exible rate regimes we nd that the error correction coecients in the exchange rate equations are strongly statistically signicant, while the error correction coecients in the relative price equations, albeit correctly signed, are not statistically dierent from zero except for Japan.15 Even if one ignores the lack of statistical signicance of the error correction coecients associated with relative prices, moreover, these error correction coecients are quite small in size relative to the corresponding error correction coecients associated with nominal exchange rates, again except for Japan. This nding is consistent with the view that during exible rate regimes the exchange rate is primarily responsible for restoring deviations from long-run PPP equilibrium.

15 The greater role of relative prices in bearing the adjustment toward PPP in Japan does not have an obvious interpretation. However, we conjecture that one possible explanation may be that Japan is, within the G5, the country which has engaged most in foreign exchange market intervention in order to manage exchange rates during the post-Bretton Woods period (e.g. Sarno and Taylor, 2001; McKinnon and Schnabl, 2003). This means that the yen might not have been allowed to bear as much of the adjustment toward the PPP equilibrium as it would have under a pure oating regime, imposing that more of the adjustment ought to be borne by relative prices.

3162

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169

The above results are quite interesting. It is rare in the use of Markov-switching models that one nds such strong evidence of being in one particular regime or the other. The results of the MSIAH-VECM in terms of classication of nominal regimes provide clear evidence that adjustment to long-run equilibrium takes dierent forms in xed versus oating exchange rate regimes. This nding appears to be in contrast, prima facie, with the ndings of Engel and Morley (2002), who use a state-space representation of the VECM linking exchange rates and relative prices for the recent oat alone and nd that nominal exchange rates respond less than relative prices to PPP disequilibria. Note, however, that Engel and Morley focus on the speed of adjustment of exchange rate and prices toward their equilibrium levels i.e. in response to what they term the exchange rate gap and the price gap, respectively, neither of which is directly related to PPP deviations rather than toward the PPP equilibrium level. In this sense, therefore, their estimates of speeds of adjustment are not comparable to the ones we report in this paper, where we focus on the response of exchange rates and prices to PPP deviations. Our results are also in line with the evidence provided by Goldfajn and Valdes (1999) and Cheung et al. (2004), who argue that the nominal exchange rate is responsible for most of the adjustment toward PPP, rather than prices. However, our empirical framework, by generalizing this line of research to a nonlinear regime-switching VECM system, allows us to pin down the speed of adjustment toward PPP for a long-span of data and across regimes. 5.4. PPP deviations are not as puzzling as we thought We examined the dynamic adjustment in response to shocks through impulse response functions which record the expected eect of a shock at time t on the system at time t + j. Estimating the impulse response function for a nonlinear model raises special problems both of interpretation and of computation since the shape of the impulse response function is not independent with respect to either the history of the system at the time the shock occurs, the size of the shock considered, or the distribution of future exogenous innovations. Exact estimates can be produced by multiple integration of the nonlinear model with respect to the distribution function of the j future innovations, which is computationally impracticable for the long forecast horizons required in impulse response analysis. We calculate impulse response functions to system-wide shocks conditional on average initial history using the Monte Carlo integration method proposed by Koop et al. (1996). We estimated regime-dependent impulse response functions and half lives for four dierent sizes of shock k 2 {1, 5, 10, 20}, conditional on average initial history. This range of shocks seems plausible in this context since it is compatible with the observed standard deviation of the real exchange rates examined.16 The shocks considered
16 It is well documented that real exchange rates are more volatile during oating than xed exchange rate regimes (e.g. Baxter and Stockman, 1989). For example, on our data, the demeaned real exchange rates have the following standard deviations during the Bretton Woods period: 0.038 for France, 0.051 for Germany, 0.053 for Japan, and 0.039 for the UK. The corresponding standard deviations during the recent oat are: 0.065 for France, 0.072 for Germany, 0.068 for Japan, and 0.057 for the UK. Therefore, the range of shocks from 1% to 20% allows us to study deviations from PPP up 3 or 4 times the standard deviation of observed real exchange rates, depending on the regime. While other authors (e.g. Taylor et al., 2001), consider larger shocks up to 40%, we refrain from analyzing these shocks which would imply faster half lives because they are likely to be very rare occurrences.

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169 Table 5 Regime-conditional half lives Fixed regime Shock = 1% France Germany Japan UK Average Shock = 5% France Germany Japan UK Average Shock = 10% France Germany Japan UK Average Shock = 20% France Germany Japan UK Average 4.05 15.01 7.02 13.01 9.77 3.03 13.00 6.04 8.02 7.53 2.07 10.02 6.02 7.00 6.28 2.05 8.02 6.02 6.04 5.53 [1.12,6.04] [6.03,17.01] [5.03,9.01] [5.02,16.00] [4.30,12.02] [1.00,5.02] [6.03,16.00] [5.04,8.00] [5.03,13.02] [4.28,10.51] [1.01,3.04] [6.03,12.02] [5.04,7.01] [5.05,9.02] [4.28,7.77] [2.03,2.08] [7.00,9.01] [6.01,7.01] [6.01,7.02] [5.26,6.28] Flexible regime 2.00 2.01 2.09 1.02 1.78 1.08 1.11 2.03 0.70 1.23 1.06 1.04 1.12 0.64 0.97 1.06 1.02 1.10 0.61 0.95 [1.00,2.12] [0.99,3.00] [1.04,4.01] [0.29,2.13] [0.83,2.82] [1.01,2.01] [0.99,2.14] [1.05,3.04] [0.38,1.08] [0.86,2.07] [1.04,1.08] [1.00,2.04] [1.06,2.06] [0.53,0.85] [0.91,1.51] [1.04,1.08] [1.01,1.09] [1.08,1.13] [0.56,0.81] [0.92,1.03]

3163

Notes: Fixed and Flexible exchange rate regimes denote the regimes identied by the nonlinear MSIAH(3)VECM(1) corresponding to Fixed and Flexible. The gures reported are the half-lives of the real exchange rate, whereas gures in brackets are the 0.025 and 0.975 quantiles of the empirical distribution of the half lives, calculated by Monte Carlo integration methods.

are shocks to the log-level of the real exchange rate, i.e. shocks to the PPP equilibrium, essentially focusing on the analysis of the impulse response function of qt to an innovation on qt1 = b 0 yt1, by taking into account all of the dynamic interactions implied by the full VECM. This is a particularly appealing way of measuring the response to PPP shocks in our context. The estimated half lives of shocks, reported in Table 5, illustrate the nonlinear nature of our models, with larger shocks dissipating much faster than smaller shocks. The regimedependent half-lives are sensibly dierent between the xed and oating regimes and for dierent shock sizes. During the xed exchange rate regimes, when the relative price is primarily responsible for bearing the burden of adjustment toward PPP, the eects of shocks on the PPP equilibrium relationship will last for more than ve years on average across countries for a 20% shock, and for almost ten years on average for a 1% shock. These half lives are indeed longer than the half lives typically recorded by the literature using linear methods on data for the recent oat (Rogo, 1996). However, the half-lives of system-wide shocks on the equilibrium relationships during exible exchange rate regimes are drastically shorter. On average across countries, 10% and 20% shocks will last for less than one year, as one might expect given that under

3164

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169

exible exchange rate regimes the nominal exchange rate can also contribute to restoring PPP in response to shocks. Even for very small shocks of 1 percent, on average across the G5 dollar exchange rates, the half life is well below two years. In comparison to the half lives reported by the literature on nonlinear mean reversion during the recent oat using univariate nonlinear models (e.g. Taylor et al., 2001), our nonlinear MSIAH-VECM produces shorter half lives, presumably because of the lack of any common factor restrictions in our multivariate nonlinear model. The properties of PPP deviations implied by our model also appear to be fairly consistent with standard models of open economy macroeconomics and with their dynamic properties under conventional nominal rigidities. For example, Chari et al. (2002) calibrate a general equilibrium model which generates theoretical autocorrelations for the real exchange rate at quarterly frequency between about 0.4 and 0.8. These autocorrelations correspond to half lives between less than one quarter and about three quarters. However, the inability of this general equilibrium model to match the persistence of real exchange rates between 3 and 5 years recorded by much empirical literature constitutes one of the main criticisms against this important theoretical paper. Our result that the half life in a oating rate regime is below one year for shocks of ten percent or higher somewhat reduces the relevance of this criticism. Our results are related to the work of Taylor (2002), who studied twenty countries over a very similar sample and concluded that half-lives of PPP deviations are roughly the same across exchange rate regimes. In Taylor (2002), countries and regimes are analyzed either one by one or using pooled samples across both countries and time periods. In our paper we nd that half-lives are much shorter for exible exchange rate periods than for xed ones. Specically, we report longer half-lives than Taylor for the xed exchange rate regimes but much lower ones for the exible regimes. Clearly, this dierence in results is a by-product of the dierent framework used in our paper where, unlike in Taylor (2002), we do not impose any common factor restrictions, select regimes on the basis of the regime classication implied by the estimates of our Markov switching models, and allow for nonlinear reversion to PPP. Further, this dierence in results is also a by-product of the calculation of the half-life, obtained in our paper using Koop-Pesaran-Potter approach because of the intrinsic nonlinearity of our model. Taylors (2002) nding that shocks have been much larger during oating regimes than during xed regimes then provides further insight in understanding our results: clearly, if this is the case, ceteris paribus, the larger shocks of the oating regime would imply faster mean reversion during these periods in our nonlinear model. Overall, the fact that consensus estimates for the rate at which PPP deviations damps generally range between three and ve years may be due to several reasons. First, several previous studies using long-span data acknowledged the fact that the results were obtained by blending xed and oating rate data, but to the best of our knowledge this is the rst explicit attempt to allow for the presence of regime shifts as well as nonlinear adjustment in investigating PPP with long-span data across countries. Second, much previous research has employed (single-equation or panel, linear or nonlinear) autoregressions for the real exchange rate rather than focusing on the full VECM linking nominal exchange rates and relative prices, hence biasing the test outcome against long-run PPP. Our nding of very short half-lives for the recent oat is presumably the product of taking into consideration simultaneously three issues that are relevant in determining the power of tests for PPP: examining the full VECM for exchange rates and prices rather than using

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169

3165

real exchange rate autoregressions; using a long-span of data rather than only data for the recent oat, while allowing for regime shifts; modelling the dynamics of PPP deviations in a nonlinear fashion.17 6. Conclusions In this paper we have re-examined the behavior of deviations from PPP using a century of data for the four dollar exchange rates obtaining among the G5 countries. We begun by noting the empirical success in establishing the validity of long-run PPP recorded by both long-span studies and studies based on nonlinear real exchange rate models, which contrasts with the unfavorable evidence provided by studies based on standard unit root tests and most cointegration tests as well as with the mixed evidence provided by studies employing panel unit root tests. However, long-span studies have often been accused of ignoring the possible regime shifts that may characterize PPP deviations across dierent monetary and exchange rate regimes. We also illustrated how linear and nonlinear real exchange rate autoregressions impose potentially invalid restrictions on the unknown data generating process driving the relationship between nominal exchange rates and prices, and how these restrictions may prevent us from detecting reversion toward PPP. This process of interpretation of the time series evidence of the relevant literature led us to employ a nonlinear empirical framework for PPP deviations which facilitates a study of long-span data while taking into account possible regime shifts and at the same time investigating whether, if PPP is validated by the data, adjustment toward the long-run equilibrium level dened by PPP is driven primarily by the exchange rate, by relative prices, or by both. Within this framework, we are able to shed light on three hotly debated questions: whether PPP holds; whether nominal exchange rates or relative prices or both of them are responsible for responding to PPP disequilibria; and whether the half-lives of PPP deviations during oating regimes are consistent with standard dynamic general equilibrium open economy models with nominal rigidities. Our results are encouraging on a number of fronts. First, the nonlinear MSIAHVECMs clearly indicate that long-run PPP holds for each of the four exchange rates examined over the last century and under each of the monetary regimes that characterize it. Second, we nd that during xed exchange rate regimes, relative prices adjust to restore deviations from long-run equilibrium, while exchange rates bear most of the burden of adjustment during exible exchange rate regimes. The estimated transition probabilities are consistent with the view that the relative importance of exchange rates and relative prices in restoring the long-run equilibrium level of the exchange rate varies over time and is aected by the nominal exchange rate arrangement in operation. Third, the estimated half-lives implied by the models are rather dierent for xed and oating rate regimes. During xed exchange rate regimes the eects of system-wide shocks on the PPP equilibrium relationship may be very persistent indeed, with half lives ranging from two to almost ten years, depending on the shock size. However, the corresponding halflives during exible exchange rate regimes are drastically shorter, presumably because
A comparable result obtained from a dierent viewpoint is worth quoting. Imbs et al. (2005) show that failing to control for the heterogeneity of the speed of mean reversion of prices across dierent sectors may lead to overestimation of the degree of persistence of the real exchange rate. When they correct for such aggregation bias, their estimates of the half life of the real exchange rate fall drastically.
17

3166

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169

the nominal exchange rate is allowed to contribute to restoring PPP. In fact, large shocks to PPP will last for less than one year on average across countries. Overall, we leave this study with some degree of optimism, which may be summarized in the following three points: long-run PPP may hold after all; PPP deviations are reversed more quickly under oating rate regimes, when nominal exchange rates are allowed to respond to shocks; the speed at which PPP deviations dissipate is fairly consistent with the predictions of standard open economy macro theory. Acknowledgements This paper was partly written while Lucio Sarno was a Visiting Scholar at the International Monetary Fund and at Norges Bank. The research was funded by the Economic and Social Research Council (ESRC, Grant No. RES-000-22-0404). The authors are grateful to Giorgio Szego (editor), two anonymous referees, Yin-Wong Cheung, Jerry Coakley, Stefan Gerlach, Andrew Rose, Shang-Jin Wei and participants to presentations at the Australasian Macroeconomic Workshop in Hong Kong, the European Central Bank, the Bank of Italy, and the 2005 Annual Conference of the Royal Economic Society for helpful comments on previous drafts, in addition to Alan Taylor for kindly providing the data set. The authors alone are responsible for any errors that may remain. References
Abuaf, N., Jorion, P., 1990. Purchasing power parity in the long run. Journal of Finance 45, 157174. Ang, A., Bekaert, G., 1998. Regime Switches in Interest Rates, National Bureau of Economic Research Working Paper No. 6508. Ang, A., Bekaert, G., 2002. Regime switches in interest rates. Journal of Business and Economic Statistics 20, 163182. Baum, C.F., Caglayan, M., Barkoulas, J.T., 2001. Nonlinear adjustment to purchasing power parity in the PostBretton Woods era. Journal of International Money and Finance 20, 379399. Baxter, M., Stockman, A.C., 1989. Business cycles and the exchange rate system. Journal of Monetary Economics 23, 377400. Calvo, G.A., Reinhart, C.M., 2002. Fear of oating. Quarterly Journal of Economics 117, 379408. Chari, V.V., Kehoe, P.J., McGrattan, E.R., 2002. Can sticky price models generate persistent and volatile real exchange rates? Review of Economic Studies 69, 533563. Cheung, Y.-W., Lai, K.S., 1993. Long-run purchasing power parity during the recent oat. Journal of International Economics 34, 181192. Cheung, Y.-W., Lai, K.S., 1994. Mean reversion in real exchange rates. Economics Letters 46, 251256. Cheung, Y.-W., Lai, K.S., 1998. Parity reversion in real exchange rates during the Post-Bretton Woods period. Journal of International Money and Finance 17, 597614. Cheung, Y.-W., Lai, K.S., Bergman, M., 2004. Dissecting the PPP puzzle: The unconventional roles of nominal exchange rate and price adjustments. Journal of International Economics 64, 135150. Chib, S., 2001. Markov chain monte carlo methods: Computation and inference. In: Heckman, J.J., Leamer, E. (Eds.), Handbook of Econometrics, vol. 5. North Holland, Amsterdam, pp. 35693649. Choudhry, T., McNown, R., Wallace, M., 1991. Purchasing power parity and the Canadian oat in the 1950s. Review of Economics and Statistics 73, 558563. Clarida, R.H., Sarno, L., Taylor, M.P., Valente, G., 2003. The Out-of-sample success of term structure models as exchange rate predictors: A step beyond. Journal of International Economics 60, 6183. Coakley, J., Fuertes, A.M., 2000. Short run dynamics of real exchange rates. Manchester School 68, 461 475. Coakley, J., Fuertes, A.M., 2001. Border costs and real exchange rate dynamics in europe. Journal of Policy Modeling 23, 669676.

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169

3167

Coakley, J., Flood, R., Fuertes, A.M., Taylor, M.P., 2005. Purchasing power parity and the theory of general relativity: The rst tests. Journal of International Money and Finance 24, 293316. Coe, P.J., Serletis, A., 2002. Bounds tests of the theory of purchasing power parity. Journal of Banking and Finance 26, 179199. Corbae, D., Ouliaris, S., 1988. Cointegration and tests of purchasing power parity. Review of Economics and Statistics 70, 508511. Dempster, A.P., Laird, N.M., Rubin, D.B., 1977. Maximum likelihood estimation from incomplete data via the EM algorithm. Journal of the Royal Statistical Society 39 (Series B), 138. Diebold, F.X., Husted, S., Rush, M., 1991. Real exchange rates under the gold standard. Journal of Political Economy 99, 12521271. Dueker, M., Neely, C.J., 2005. Can Markov Switching Models Predict Excess Foreign Exchange Returns? Federal Reserve Bank of St. Louis Working Paper No. 21B. Dumas, B., 1992. Dynamic equilibrium and the real exchange rate in spatially separated world. Review of Financial Studies 5, 53180. Edison, H.J., 1987. Purchasing Power Parity in the Long Run. Journal of Money, Credit and Banking 19, 376 387. Edison, H.J., Gagnon, J.E., Melick, W.R., 1997. Understanding the empirical literature on purchasing power parity: The Post-Bretton Woods era. Journal of International Money and Finance 16, 117. Engel, C., 1994. Can the markov switching model forecast exchange rates? Journal of International Economics 36, 151165. Engel, C., 1999. Accounting for US real exchange rate changes. Journal of Political Economy 107, 507538. Engel, C., 2000. Long-run PPP may not hold after all. Journal of International Economics 51, 243273. Engel, C., Hamilton, J.D., 1990. Long swings in the dollar: Are they in the data and do markets know it? American Economic Review 80, 689713. Engel, C., Kim, C.J., 1999. The Long-run US/UK real exchange rate. Journal of Money Credit and Banking 31, 335356. Engel, C., Morley, J.C., 2002. The Adjustment of Prices and the Adjustment of the Exchange Rate. University of Wisconsin and Washington University, mimeo. Engle, R.F., Granger, C.W.J., 1987. Co-integration and error correction: Representation, estimation, and testing. Econometrica 55, 251276. Frankel, J.A., Rose, A.K., 1996. A panel project on purchasing power parity: Mean reversion within and between countries. Journal of International Economics 40, 209224. Froot, K.A., Rogo, K., 1995. Perspectives on PPP and long-run real exchange rates. In: Rogo, K., Grossman, G. (Eds.), Handbook of International Economics. North Holland, Amsterdam. Garcia, R., 1998. Asymptotic null distribution of the likelihood ratio test in markov switching models. International Economic Review 39, 763788. Goldfajn, I., Valdes, R., 1999. The aftermath of appreciations. Quarterly Journal of Economics 114, 229 262. Granger, C.W.J., 1986. Developments in the study of cointegrated variables. Oxford Bulletin of Economics and Statistics 48, 213228. Hakkio, C.S., 1984. A re-examination of purchasing power parity: A multi-country and multi-period study. Journal of International Economics 17, 177265. Hamilton, J.D., 1993. Estimation, inference and forecasting of time series subject to changes in regime. In: Maddala, G.S., Rao, C.R., Vinod, H.D. (Eds.), Handbook of Econometrics, vol. 4. Elsevier, Amsterdam. Hansen, B.E., 1992. The likelihood ratio test under nonstandard conditions: Testing the Markov switching model of GNP. Journal of Applied Econometrics 7, S61S82. Hansen, H., Johansen, S., 1999. Some tests for parameter constancy in cointegrated VAR-models. Econometrics Journal 2, 306333. Imbs, J., Mumtaz, H., Ravn, M.O., Rey, H., 2003. Non linearities and real exchange rate dynamics. Journal of the European Economic Association 1, 639649. Imbs, J., Mumtaz, H., Ravn, M.O., Rey, H., 2005. PPP strikes back: Aggregation and the real exchange rate. Quarterly Journal of Economics 120, 144. Johansen, S., 1988. Statistical analysis of cointegrating vectors. Journal of Economic Dynamics and Control 12, 231254. Johansen, S., 1991. Estimation and hypothesis testing of cointegrating vectors in Gaussian vector autoregressive models. Econometrica 59, 15511580.

3168

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169

Kim, Y., 1990. Purchasing power parity in the long run: A cointegration approach. Journal of Money, Credit, and Banking 22, 491503. Kim, C.-J., Nelson, C.R., 1999. State-space models with regime switching. MIT Press, Cambridge, Mass. and London, UK. Kim, C.-J., Piger, J.M., Startz, R., 2003. Estimation of Markov Regime-Switching Regression Models with Endogenous Switching. Federal Reserve Bank of St. Louis Working Paper No. 2003-015A. Koop, G., Pesaran, M.H., Potter, S., 1996. Impulse response analysis in nonlinear multivariate models. Journal of Econometrics 74, 119147. Kremers, J.J.M., Ericsson, N.R., Dolado, J.J., 1992. The power of cointegration tests. Oxford Bulletin of Economics and Statistics 54, 325348. Krolzig, H.-M., 1997. Markov-Switching Vector Autoregressions. Springer, New York, NY. Lane, P.R., 2001. The new open economy macroeconomics: A survey. Journal of International Economics 54, 235266. LeBaron, B., 1992. Forecast improvements using a volatility index. Journal of Applied Econometrics 7, S137 S149. Lothian, J.R., Taylor, M.P., 1996. Real exchange rate behaviour: The recent oat from the perspective of the past two centuries. Journal of Political Economy 104, 488510. Lothian, J.R., Taylor, M.P., 1997. Real exchange rate behaviour: The problem of power and sample size. Journal of International Money and Finance 16, 945954. Mark, N.C., 1990. Real and nominal exchange rates in the long run: An empirical investigation. Journal of International Economics 28, 115136. McKinnon, R., Schnabl, G., 2003. Synchronized business cycles in East Asia: Fluctuations in the yen/dollar exchange rate and Chinas stabilizing role. World Economy (August), 10671088. McNown, R., Wallace, M., 1989. National price levels, purchasing power parity, and cointegration: A test of four high ination economies. Journal of International Money and Finance 8, 533545. Michael, P., Nobay, A.R., Peel, D.A., 1997. Transactions costs and nonlinear adjustment in real exchange rates: An empirical investigation. Journal of Political Economy 105, 862879. Nakagawa, H., 2002. Real exchange rates and real interest dierentials: Implications of nonlinear adjustment in real exchange rates. Journal of Monetary Economics 49, 629649. Obstfeld, M., Taylor, A.M., 1997. Nonlinear aspects of goods-market arbitrage and adjustment: Heckschers commodity points revisited. Journal of the Japanese and International Economies 11, 441479. Obstfeld, M., Taylor, A.M., 2004. Global Capital Markets: Integration, Crisis, and Growth. Cambridge University Press, Cambridge, UK and New York, USA. OConnell, P.G.J., 1998. The overvaluation of purchasing power parity. Journal of International Economics 44, 119. OConnell, P.G.J., Wei, S.-J., 2001. The bigger they are, the harder they fall: Retail Price dierences across US cities. Journal of International Economics 56, 2153. Papell, D.H., 1998. Searching for stationarity: Purchasing power parity under the current oat. Journal of International Economics 43, 313332. Quintos, C.E., 1997. Stability tests in error correction models. Journal of Econometrics 82, 289315. Rogo, K., 1996. The purchasing power parity puzzle. Journal of Economic Literature 34, 647668. Sarno, L., 2001. Towards a new paradigm in open economy modelling: Where do we stand? Federal Reserve Bank of St. Louis Review 83, 2136. Sarno, L., 2005. Towards a solution to the puzzles in exchange rate economics: Where do we stand? Canadian Journal of Economics 38, 673708. Sarno, L., Taylor, M.P., 1998. Real exchange rates under the recent oat: Unequivocal evidence of mean reversion. Economics Letters 60, 131137. Sarno, L., Taylor, M.P., 2001. Ocial intervention in the foreign exchange market: Is it eective and, if so, how does it work? Journal of Economic Literature 39, 839868. Sarno, L., Taylor, M.P., 2002. Purchasing power parity and the real exchange rate. International Monetary Fund Sta Papers 49, 65105. Sarno, L., Valente, M.P., 2005. Empirical exchange rate models and currency risk: some evidence from density forecasts. Journal of International Money and Finance 24, 363385. Sercu, P., Uppal, R., Van Hulle, C., 1995. The exchange rate in the presence of transactions costs: implications for tests of purchasing power parity. Journal of Finance 50, 13091319.

L. Sarno, G. Valente / Journal of Banking & Finance 30 (2006) 31473169

3169

Serletis, A., Gogas, P., 2004. Long-horizon regression tests of the theory of purchasing power parity. Journal of Banking and Finance 28, 19611985. Siklos, P., Granger, C.W.J., 1997. Regime-sensitive cointegration with an illustration from interest rate parity. Macroeconomic Dynamics 1, 640657. Taylor, A.M., 2001. Potential pitfalls for the purchasing power parity puzzle? Sampling and specication biases in mean reversion tests of the law of one price. Econometrica 69 (2), 473498. Taylor, A.M., 2002. A century of purchasing-power parity. Review of Economics and Statistics 84, 139150. Taylor, M.P., McMahon, P.C., 1988. Long-run purchasing power parity in the 1920s. European Economic Review 32, 179197. Taylor, M.P., Sarno, L., 1998. The behaviour of real exchange rates during the Post-Bretton Woods period. Journal of International Economics 46, 281312. Taylor, A.M., Taylor, M.P., 2004. The purchasing power parity debate. Journal of Economic Perspectives 18, 135158. Taylor, M.P., Peel, D.A., Sarno, L., 2001. Nonlinear mean-reversion in real exchange rates: Toward a solution to the purchasing power parity puzzles. International Economic Review 42, 10151042.