The value of the Model Output Statistics (MOS) approach to improving 6-10 day and week 2 probabilistic forecasts of surface temperature and precipitation is demonstrated. Retrospective two-week ensemble ``re-forecasts'' were computed using a reduced-resolution T62L28 version of the NCEP medium-range forecast model with physics operational during 1998. NCEP-NCAR reanalysis initial conditions and bred modes were used to initialize the 15-member ensemble. Probabilistic forecasts of precipitation and temperature were generated using a logistic regression technique with the ensemble mean (precipitation) or ensemble mean anomaly (temperature) as the only predictor. Forecasts were computed and evaluated during 23 winter seasons from 1979 to 2001.

Evaluated over the 23 winters, these MOS-based probabilistic forecasts were skillful and highly reliable. When compared against operational NCEP forecasts for a subset of 100 days from the 2001-2002 winters, the MOS-based forecasts were comparatively much more skillful and reliable. For example, the MOS-based week 2 forecasts were more skillful than operational 6-10 day forecasts. Most of the benefit of the MOS approach could be achieved with 10 years of training data, and since sequential sample days provided correlated training data, the costs of re-forecasts could also be reduced by skipping days between forecast samples.

MOS approaches will still require a large data set of retrospective forecasts in order to achieve their full benefit. This forecast model must remain unchanged until re-forecasts have been computed for the next model version, a penalty which will slow down the implementation of model updates. Given the substantial improvements noted here, it is argued that re-forecast based MOS techniques should become an integral part of the medium-range forecast process despite this cost. Techniques for computing re-forecasts while minimizing the impact to operational weather prediction facilities and model development will be discussed.