The Meteorological Development Laboratory (MDL) has just completed development of snowfall amount forecast guidance for 24-h periods. These forecasts are generated by using the Model Output Statistics (MOS) technique which statistically relates observed predictand data to predictor data such as forecasts from dynamical models and surface observations. Snowfall guidance will be available in the fall of 2003 based on the National Centers for Environmental Prediction's (NCEP) Global Forecast System (GFS, formerly AVN/MRF), with forecasts based on the Eta model to follow shortly.

In order to develop skillful forecast relationships, a high quality sample of snowfall amount observations is essential. With the advent of the Automated Surface Observing System (ASOS) in the 1990's and the conversion in the United States to METAR standards, most observing sites stopped reporting snowfall amount, with the exception being some of those with human observers. As time passed, more and more of these manual sites were replaced with ASOS equipment, resulting in snowfall amount being unavailable in the hourly observational record. The only dataset that provided significant temporal and spatial coverage for snowfall amount was the National Climatic Data Center's (NCDC) Cooperative Summary of the Day. These data consist of daily 24-h snowfall amount reports for approximately 8000 sites in the contiguous United States and Alaska.

The use of these data in a MOS development presented many challenges. First, the data had to be transformed from several NCDC formats to MDL's binary format for use in the MOS system. Secondly, a well-known limitation of the cooperative observations is that the stations report at various local times. One site might report from 7 am to 7 am, while another site 5 miles away reports midnight to midnight. This presents a challenge when producing guidance for the same time periods for all sites. Thirdly, while this dataset contains a long historical record, which is a benefit, many changes in stations, data formats, and quality control standards have occurred throughout the period of record. Although NCDC conducts extensive quality control of the data, additional work was necessary to produce the best sample of snowfall data possible.

In this paper we discuss the statistical methodology used in the development of the new MOS snowfall guidance, as well as the issues confronted in working with the NCDC data. We outline the design of our snowfall prediction system, and show how we tailored the data to fit this purpose. Finally, we present some verifications of the new forecasts and discuss future work in snowfall prediction.