Time-longitude (Hovmoller) diagrams of warm season rainfall over central North America reveal distinct diurnal patterns whose maxima are temporally in phase with surface temperature and co-located with orography. In addition, a propagating axis of nocturnal rainfall dominates the signal between the cordilleras and leads to a semi-diurnal temporal behavior over the Great Plains of the U.S. Analysis of the forecasts of warm-season rainfall from two NWP models, the NCEP Eta and the Weather Research and Forecast model reveals that neither model correctly captures any hint of propagation and therefore both models miss the nocturnal rainfall maximum over the central U.S. Furthermore, zeroth-order errors in the diurnal cycle of rainfall are apparent.

Despite their dismal performance mentioned above, current NWP models are relatively adept at forecasting the approximate latitudinal belt of rainfall at lead times of about 1 day. This result suggests that a combination of statistical and dynamical forecasts should yield improved estimates of the locations of rainfall with lead times of roughly 1 day, at least in a probabilistic sense. This idea is tested by applying a probabilistic filter to forecasts such that the NWP model determines the latitudinal and temporal probability of rainfall, and the climatology derived from previous warm seasons determines the longitudinal distribution specified times during the diurnal cycle. Results are evaluated through the use of reliability diagrams and other methods appropriate for probabilistic forecasts.