The goal of the present study is to re‐examine the exchange rate predictability with an approach that accounts for the negative effect of the finite‐sample estimation error on forecast accuracy in the in‐sample test. We consider various exchange rate models and find that despite the presence of significant population‐level predictive content in the exchange rate model, the coefficients of the predictive variables could be small enough that, with the available sample, they are estimated so imprecisely that a random walk model can be expected to forecast at least as well as the exchange rate model.