Abstract Unpreparedness is often the main cause of the economic and social damages caused by floods. To mitigate these impacts, short-term forecasting has been the focus of several studies during the past decades; however, less effort has been paid to flood predictions at longer lead times. Here, we use forecasts by six models from the North American Multi-Model Ensemble project with a lead time from 0.5 to 9.5 months to predict the seasonal duration of floods above four National Weather Service flood categories (“action,” “flood,” “moderate” and “major”). We focus on 202 U.S. Geological Survey gage stations across the U.S. Midwest and use a statistical framework which considers precipitation, temperature, and antecedent wetness conditions as predictors. We find that the prediction skill of the duration of floods for the “action” and “flood” categories is overall low, largely because of the low accuracy of the climate forecasts rather than of the errors introduced by the statistical models. The prediction skill slightly improves when considering the shortest lead times (i.e., from 0.5 to 2.5 months) during spring in the Northern Great Plains, where antecedent wetness conditions play an important role in influencing the generation of floods. It is very difficult to draw strong conclusions with respect to the “moderate” and “major” flood categories because of the limited number of available events.