This paper presents consistent and fast bootstrap methods for constructing nonparametric prediction intervals for autoregressive processes. These methods are compared, in a simulation study, with the Box-Jenkins approach and Thombs-Schucany's bootstrap method.