Automatic history matching techniques such as the Ensemble Kalman Filter (EnKF) have been shown to provide reliable results for matching and prediction at existing wells. It is much less clear if the prediction outside of existing wells improves as a result of history matching. The amount of information in production data on the spatial distribution of properties is limited, thus the prior information derived from well logs and seismic data is likely to be important for the predictions outside of existing wells. To test this, a twin experiment was done in a fluvial system dominated by channel belts. The facies distribution of the synthetic truth was simulated with a process based model. Three different cases were tested: the position of the channel belts is perfectly known, is approximately known from seismic data and is not known except in the wells. The EnKF was used to estimate the distribution of the permeability and porosity. The resulting models were used to predict production at existing wells and at new wells not used in the history match. The results showed that the EnKF always improved the predictions at existing wells, but was often not able to improve the prediction at the new wells.