A new attempt oil ensemble flood forecasting is introduced by use of a radar image extrapolation along with a stochastic error field simulation, and a distributed hydrologic model. In contrast to the conventional ensemble simulation that uses initial condition control to obtain a statistical outcome, stochastic prediction en-or fields were externally generated to offer probable variations of deterministic predictions. Firstly, a radar extrapolation model provided deterministic rainfall prediction, and its prediction error structure was analysed by comparing the observed rainfall fields. Secondly, on the basis of the analysed error structure, spatially correlated error fields Were Simulated using a covariance matrix decomposition method oil a real-time basis. The simulated error fields enabled not only the production of probable rainfall field variations for ensemble simulation but also an improvement in prediction accuracy by offsetting the deterministic prediction error, Finally, the simulated error fields along with the deterministic fields were tested with a distributed hydrologic model to measure the validity of the ensemble runoff prediction. Copyright (C) 2008 John Wiley & Sons, Ltd.