Probabilistic Forecasts Based on "Reforecasts"

We provide an overview of the use of reforecasts for improving probabilistic weather forecasts at leads of days to weeks. A 25+ year set of retrospective forecasts have been created with a stable, reduced-resolution version of the NCEP GFS model. 15-member bred ensembles were run every day from 0000 UTC initial conditions to 15 days lead. Initial analyses were centered on the NCEP-NCAR initial condition. We show that operational 6-10 day and week 2 forecasts are dramatically improved through the application of statistical techniques, correcting the real-time forecast using the training data set. Similarly, we show that accurate and reliable probabilistic quantitative precipitation forecasts can be generated during week 1 using a 2-step analog forecast approach. Real-time and archived reforecast data is freely available to all.