Forecasting upper and lower uncertainty bands of river flood discharges with high predictive skill

River discharges flood forecasting is a complex task with multiple sources of uncertainty. Bayesian methods can incorporate multiple types of uncertainties by inferring the probability density function of ensemble forecasts based on past events. However, such methods lead to forecasts with large uncertainty bands. In order to reduce the uncertainty in the forecasts, we focus solely on the prediction of the upper and lower range of the uncertainty bands. Therefore, we develop three forecast methods in which we search for the indexes of the upper and lower forecast members of an ensemble (termed best-pairs), which provide the highest predictive skill. The results show for four distinct hindcasts of historical events in a case study in Bavaria (Germany) that the new methods have a higher predictive skill of the observations than probabilistic methods, at least for the first 4 out of 12 h’ forecasts. Moreover, the new methods are computational efficient because they considerably reduce the number of members of the ensembles required to produce a flood discharge forecast with high predictive skill.