Abstract
We show that the behaviour of the real exchange rates of the UK, Germany, France and Japan has been characterized by structural breaks, which changed the adjustment mechanism. In the context of a Time-Varying Smooth Transition Autoregression (TV-STAR) of the kind introduced by Lundbergh et al. (Citation2003), we show that the real exchange rate process shifted in the aftermath of Black Wednesday in the case of pound, in 1984/85 in the case of franc and, more tentatively, during the Asian crisis of 1997/98 in the case of yen.
Acknowledgements
Subject to the usual disclaimer, the authors wish to acknowledge the comments of participants in seminars at De Nederlandsche Bank, Amsterdam, and the University of Hannover for their helpful comments.
Notes
1 Major events in exchange rate history have also been revisited in the PPP framework, e.g. Taylor (Citation1992).
2 For the role of expectations in smooth transition models, see Peel and Venetis (Citation2005).
3 O’Connell (Citation1998) cites this as an example of the kind of structural change that may explain his counter-intuitive results.
4 See Akram et al. (Citation2005) for an example of a TV-STAR process fitted to the time series of nearly two centuries of annual Norwegian real exchange rates.
5 Though Sarantis (Citation1999), who finds LSTAR rather than ESTAR best, explains the behaviour of the real effective exchange rates of the franc and DM.
6 For a discussion of some of these issues, and an examination of the evidence, see Pippenger and Geppert (Citation1997) and Campa and Goldberg (Citation2005), who attribute different degrees of pass-through to differential monopoly power across both countries and industries.
7 In particular, we do not accept the argument of Taylor et al. (Citation2001) that: ‘Since … the LSTAR model implies asymmetric behaviour … we regard that model a priori as inappropriate … it is hard to think of economic reasons why the speed of adjustment should vary according to whether the dollar is overvalued or undervalued … against the currencies of other … industrialized countries’ (p. 1021). They do nonetheless test for LSTAR processes in their dataset, presumably finding ESTAR ultimately preferable.
8 Hopefully, enough observations to be immune to the upward bias in estimates of the adjustment speed in ESTAR as documented in Paya and Peel (Citation2006), though whether their conclusions apply equally to TV-STAR models and to LSTAR is unclear.
9 Though with test statistics of: −2.99, −2.92, −2.26, −3.79, against the critical value given in Kapetanios et al. (Citation2000) of −2.22, the evidence is far from overwhelming.
10 We are grateful to the authors of this survey for the use of their software.
11 One of the few exceptions is Sarantis (Citation1999).
12 We are grateful to a referee of this journal for pointing out that, given the variance shift at the start of the floating era, reliable critical values can only be derived from the wild bootstrap. In the event, this technique generated critical values, which supported the conclusions reported here, by a comfortable margin for Germany, Japan and France, less convincingly in the case of the UK.
13 In fact, as was clear from the graph of F 1(t), (not shown here, but available from the authors on request) the dataset ends well before the time transition function gets close to its upper asymptote of one. The scale of the apparent shift for Japan may also be due to other factors. For example, markets may have extrapolated the falling Japanese price level into the indefinite future. Also, the model implemented here ignores any possible Balassa–Samuelson productivity-growth effect.
14 Taking account of the fact that France was in the upper zone of the exchange rate transition function for almost the whole period (). The implausibly large size of the implied devaluation in the other two states, is almost certainly a result of the fact that we have virtually no observations in the lower regime and not many in the transition phase between the two.
15 In principle, we need not confine attention to the mean. In fact, van Dijk et al. (Citation2002) consider three different percentiles of the realized distribution of outcomes. Here, we examined the medians whenever we suspected the means may have been distorted by a small number of extreme realizations. The results, however, were not qualitatively different, so are not reported here.
16 For other values, the results (which were qualitatively similar) are available from the authors.