Improving probabilistic QPF with reforecasts

This talk will demonstrate how probabilistic QPFs can be dramatically improved through the use of a reforecast (retrospective forecast) database. This database is comprised of a 15-member ensemble of a T62 forecast model run out to two weeks lead. Forecasts have been run every day from 0000 UTC initial conditions from 1979 to present. The model is a 1998 version of the National Centers for Environmental Prediction's Global Forecast System (NCEP GFS) at T62 resolution. The initial conditions consist of a reanalysis initial condition and seven pairs of bred modes.

While the forecast model is comparatively low-resolution, the long training data set permits the application of statistical techniques that have heretofore been impossible. In this talk we will discuss a two-step analog technique: the first step is to find past local forecast analogs to the current local weather forecast, the second step is to extract observed weather on those dates and form an ensemble from them. Using this technique, we demonstrate that precipitation forecasts can be generated which are much more skillful than precipitation forecasts from the current, higher-resolution version of the NCEP GFS.

More generally, we argue that the benefits of reforecasts are so large that they should become an integral part of the numerical weather prediction process. Methods for integrating reforecast approaches without seriously compromising the pace of model development are discussed.

Users wishing to explore their own applications of reforecasts can download the real-time and retrospective forecasts through a web interface.