1060 The Econometrics of Exchange Rates
more short horizon predictability for Germany and more long-horizon predictabil-
ity for Canada and Switzerland. Moreover, tests of forecast encompassing appear
to be superior in detecting predictability of exchange rates compared to tests of
forecast accuracy.
22.6.2 Nonlinear models
Taylor and Kilian (2003) investigate the forecasting performance of long-horizon
regressions in the presence of ESTAR dynamics in the real exchange rate. They
employ the long-horizon regression of Mark (1995) and draw inferences from boot-
strap distributions of the test statistics generated under the null hypothesis that the
nominal exchange ratestfollows an RW and the real exchange rateztis an ESTAR
process:^69
st=μs+us,t, (22.91)
zt=μz+[
φ 1 (zt− 1 −μz)+( 1 −φ 1 )(zt− 2 −μz)]×exp⎛
⎝−γ
∑^5d= 1(zt−d−μz)^2⎞
⎠+uz,t. (22.92)Taylor and Kilian (2003) use quarterly data for seven OECD countries for the period
1973:1 to 1998:4. Long-horizon regressions appear to be significantly more accu-
rate than the naive RW model in several cases, especially when the Newey–West
standard errors are used with the truncation lag specified by Andrews’ procedure.
However, the out-of-sample results are not as encouraging.^70 TheDM-statistics
indicate that the long-horizon regression model is capable of beating the RW model
only for the UK and Switzerland at the three-year horizon. The authors conclude
that incorporating nonlinearities increases the predictability of models based on
macroeconomic fundamentals. However, it is difficult to detect the improvement
in the forecast accuracy due to the small time span and the rarity of large deviations
from the fundamentals.
Another group of recent studies focuses on the Markov-switching (MS)
model, which allows exchange rate dynamics to alternate between regimes.^71
Claridaet al.(2003) argue that the forward rate has predictive content regarding
the spot rate. The authors build upon the results of Clarida and Taylor (1997) and
apply an MS-VECM model, which allows for shifts in the intercept and the error
variance. Their findings indicate that the MS-VECM outperforms both the linear
VECM and the naive RW model, especially at long horizons.
Sarnoet al.(2004b) employ a long span of data for the US dollar exchange rate
and show that fundamentals are useful in explaining the behavior of numerous
exchange rates under different monetary regimes by estimating MS-VECM models.
Frömmelet al.(2005), motivated by the market microstructure literature (Lyons,
2001) and questionnaire surveys (e.g., Cheung and Chinn, 2001) showing that mar-
ket participants regard the importance of fundamentals as time-varying, establish
that there are significant regime changes in real interest differential (RID) variants
of the monetary model (Frankel, 1979) for three major US dollar exchange rates.^72