A prototype early warning system for floods is introduced. For a small headwater catchment, probabilistic streamflow predictions in 24-hourly steps are obtained from downscaling all members of the European Centre for Medium-Range Weather Forecasts (ECMWF) Ensemble Prediction System and feeding the resulting precipitation and temperature series into a hydrologic model. We apply “expanded downscaling,” a scheme that was previously used for climate scenarios and that is particularly suited to extreme events and the simulation of flood-triggering heavy rainfall. The entire model chain is thoroughly verified, using daily precipitation and streamflow observations and forecasts from the decade 1997–2006. It turns out that strong meteorologic (precipitation) events are skillfully predicted for at least 5 days lead time by the downscaling. That skill, however, is partly lost by deficiencies in the hydrological modeling as revealed in this study. We discuss ways to overcome these difficulties, along with the prospect of employing the whole system operationally, for example, for reservoir regulations. We close with an outlook for early flash flood warnings.