The use of a new bootstrap method for small-sample inference in long-horizon regressions is illustrated by analysing the long-horizon predictability of four major exchange rates, and the findings are reconciled with those of an earlier study by Mark (1995). While there is some evidence of exchange rate predictability, contrary to earlier studies, no evidence is found of higher predictability at longer horizons. Additional evidence is presented that the linear VEC model framework underlying the empirical study is likely to be misspecified, and that the methodology for constructing bootstrap p-values for long-horizon regression tests may be fundamentally flawed.