One of the most frequently studied puzzles in international monetary economics is the failure of standard linear monetary models of exchange rates to forecast variations in the exchange rates in the short run. Ever since Meese and Rogoff’s (1983) work on out-of-sample forecast comparison of varieties of monetary models of exchange rates and naïve random walk model, a consensus view has emerged that monetary models are largely unsuccessful in forecasting exchange rates, at least in the short term. This literature casts doubts about the suitability of economic models based on fundamentals in forecasting exchange rates (see Cheung and Chinn, 2001, or Marsh, Cheung and Chinn, 2004) for evidence based on surveys).
The work of Mark (1995) revived interest in monetary models by focusing on long-term predictability of exchange rates. From this perspective, models based on fundamentals are essentially valid in the long run. That means there is a tendency in the exchange rates to adjust to their long-term values as suggested by the fundamentals. With the use of nonparametric bootstrapping methods he was able to show that monetary models with linear mean reversion are of better use in predicting exchange rates in long horizons than in short horizons. He found some out-of-sample predictability for Japanese Yen, German Mark and Swiss Frank exchange rates vis-à-vis US Dollar at 12 and 16 quarters forecast horizons.
Mark’s (1995) work has been subject to criticism on several grounds. Firstly, Berkowitz and Giorgianni (2001) argue that the distribution of the bootstrap test statistic as implemented by Mark depends on the assumption of cointegration between the fundamentals and exchange rates. Given that Mark assumes cointegration between fundamentals and exchange rates to generate bootstrap critical values, if fundamentals and exchange rates are not cointegrated in actual data, critical values and therefore inference from the test would be incorrect. Berkowitz and Giorgianni report very weak evidence of cointegration in the data which is corroborated by Kilian (1999). Kilian finds that even if there is cointegration between fundamentals and exchange rates mean reversion in forecast errors is very slow. Secondly, data generating process and assumed mean reversion has been criticized.
Since the work of Neftçi (1984) it has been increasingly popular to test for nonlinearities and structural instabilities in economic time series. Enders and Granger (1998) show that if nonlinearities are prevalent under the alternative of stationarity, linear tests for unit roots suffer from a lack of power. Not surprisingly, Kilian and Taylor (2003) show that if there is evidence of nonlinear mean reversion standard tests of long-horizon predictability of exchange rates are invalidated. Finally, Faust, Rogers and Wright (2003) argue that data on fundamentals are subject to continuous revisions. They show that Mark’s linear adjustment results are mainly the outcome of a certain window of vintages of the real time dataset and therefore not generally valid.
Download
Exchange Rates and Fundamentals: Is there a Role for Nonlinearities in Real Time?
