Is there any role for the Taylor rule in helping predict exchange rates?
Figure 1: Log nominal value of the dollar (vs. major currencies, Federal Reserve measure), and deviation of industrial production from HP filtered series (over 1967-2006 period). Source: FRED II and author’s calculations.
Figure 2: Log nominal value of the dollar (vs. major currencies, Federal Reserve measure), and deviation of CPI inflation rate (12 month difference of logged series) from HP filtered series (over 1967-2006 period). Source: FRED II and author’s calculations.
My colleagues Charles Engel and Kenneth West at the University of Wisconsin have written a paper [pdf] suggesting that the presence of Taylor rules (the tendency for central banks to adjust overnight interbank interest rates in response to deviations from targetted GDP and inflation) might explain part of the movements of the DM/dollar rate, but not more than 40% of the variation. In my view, this is an important result, especially against the backdrop of generally dismal results in the exchange rate prediction business.
More recently, Tanya Molodtsova and David Papell at the University of Houston have analyzed the out of sample forecasting performance of models based upon Taylor rule fundamentals, namely relative output gaps and inflation rates (they also assess a Taylor rule specification augmented with the real exchange rate) [pdf]. Since they are using monthly data, the output gaps are calculated using linear and quadratic detrending of industrial production, as well as HP filtered industrial production. They conclude:
‘Research on exchange rate predictability has come full circle from the “no predictability at short horizons” results of Meese and Rogoff (1983a, b) to the “predictability at long horizons but not short horizons” results of Mark (1995) and Chen and Mark (1996) to the “no predictability at any horizons” results of Cheung, Chinn, and Pascual (2005). We come to a very different conclusion, reporting strong evidence of exchange rate predictability at the one-month horizon, slight evidence of predictability at the three-month horizon, and no evidence of predictability at longer horizons.’
In addition they find that the short horizon outperformance is most pronounced for the Taylor model at the short horizon. In their inferences, it turns out to be important to use the Clark-West statistic, which yields properly sized tests in this case, as opposed to the Diebold-Mariano statistic. (For a discussion of the forecasting power of other fundamentals, also using the Clark-West test statistic, see this paper,).