Exchange Rate Predictability

This article asks whether anything forecasts exchange rates and, if so, what – and answers that it depends, systematically, on five choices the forecaster makes. Since Meese and Rogoff (1983a,b, 1988) it has been known that a simple random walk usually beats economic models out of sample, and the survey opens by insisting on what that does not mean: the finding “should not be interpreted as a validation of the efficient market hypothesis,” which “does not mean that exchange rates are unrelated to economic fundamentals, nor that exchange rates should fluctuate randomly around their past values. Hence the puzzle.” The first half reviews the predictors, models, data choices and forecast evaluation methods proposed over the preceding decade, with deliberately limited scope: a reduced-form, out-of-sample focus at monthly and quarterly frequencies, on nominal rather than real exchange rates, because structural models “are too stylized to be literally taken to the data and successfully used for forecasting.” The second half re-runs the horse race on up-to-date data for a set of 19 countries against the US dollar, using a single common data transformation (seasonal adjustment), a rolling estimation window equal to half the sample, short (one month or quarter) and long (four-year) horizons, and a battery of tests: a Granger-causality test robust to instabilities, the Diebold-Mariano-West and Clark-West out-of-sample tests against a random walk without drift, and two tests that check robustness to the forecast sample and to the estimation window size. The literature’s consensus negatives are confirmed – purchasing power parity and monetary fundamentals do not forecast at horizons under about two to three years, and non-linear models are the least successful – and the one positive consensus is that Taylor-rule and net foreign asset fundamentals have more out-of-sample content than traditional interest rate, price, output and money differentials, with the disagreement being over how much of the puzzle they resolve. In the author’s own exercise most traditional predictors show in-sample predictive ability while almost none survives out of sample; Taylor-rule fundamentals are strongly significant under Clark-West for four countries and marginally for five more at the one-month horizon, and for none at four years, but are not significant at all under Diebold-Mariano-West – a discrepancy the author reads as the difference between judging a model “in population” and judging it “at the actual estimated parameter values.” A model loading up on real interest rates, the trade balance and the current account has the strongest in-sample fit of any considered and “extremely poor” out-of-sample performance, “possibly due to the large number of parameters to estimate.” A panel monetary model beats the random walk at long horizons for four countries and a Bayesian model averaging specification beats it nowhere. The sharpest result is about instability: the Fluctuation test finds predictive ability for monetary fundamentals only occasionally and briefly – never for Canada, for France in the late 2000s, Japan in the mid-2000s, the UK in 2009 – and window-size analysis shows that only small estimation windows detect German predictability (concentrated in the late 1980s) while only large ones detect Japan’s (emerging in the late 2000s). The author’s conclusion is therefore that although some predictors work somewhere sometimes, “none of the predictors, models, or tests systematically find empirical support for superior exchange rate forecasting ability of a predictor for all models, countries and time periods,” so “Meese and Rogoff’s (1983a,b) finding does not seem to be entirely and convincingly overturned” – and that the time-varying, occasional nature of predictability is itself a new puzzle, since existing time-varying-parameter models fail to capture it.

Barbara Rossi | Macro Paper Warehouse

Exchange Rate Predictability