We study the robustness of block resampling procedures for time series. We first derive a set of formulas to characterize their quantile breakdown point. For the moving block bootstrap and the subsampling, we find a very low quantile breakdown point. A similar robustness problem arises in relation to data-driven methods for selecting the block size in applications. This renders inference based on standard resampling methods useless already in simple estimation and testing settings. To solve this problem, we introduce a robust fast resampling scheme that is applicable to a wide class of time series settings. Monte Carlo simulations and sensitivity analysis for the simple AR(1) model confirm the dramatic fragility of classical resampling procedures in presence of contaminations by outliers. They also show the better accuracy and efficiency of the robust resampling approach under di®erent types of data constellations. A real data application to testing for stock return predictability shows that our robust approach can detect predictability structures more consistently than classical methods.