The Model Output Statistics (MOS) approach statistically relates observed predictand data to predictor data such as forecasts from dynamical models, surface observations, and geoclimatic information. In order to develop robust forecast relationships, a large, high quality developmental sample of model forecasts and observational data is absolutely essential. While issues pertaining to the content and stability of model archives have been documented, the vital role of the observational data sample is often not fully appreciated.

A MOS forecast equation is only as good as the observational sample on which it is developed. The main source of observational data used to compute the MOS forecast predic- tands is the hourly surface observations. While these data present one picture of the observed weather, several steps may be required to transform the observations into useful predictands. First, the data must be quality-controlled to correct inevitable reporting irregularities. Next, inconsistencies in the data sample must be dealt with. The pool of reporting stations is not constant; new stations open, while existing stations close or move their instrumentation to another location. Reporting standards also change. In July 1996, the reporting standard for surface observations changed from the SAO format to METAR. As a result, some variables once used as MOS predictands, such as opaque sky cover, are no longer reported. Once these inconsistencies have been taken into account, the data must be transformed into appropriate predictands. Defining the predictand is a critical step in developing any statistical forecast equation. Often, forecast guidance is required for weather elements that are not routinely observed. One example is the daytime maximum temperature, a quantity not reported by observing sites, but very important for public forecasts. This predictand must be estimated for the required daytime period by using the hourly temperatures and the available 6-h maximum temperatures. In other cases, instrumentation can limit the ability to observe some phenomenon. For example, ASOS instruments can not report clouds above 12,000 ft. Therefore, a station can report clear skies when, in fact, the sky is totally covered with a deck of clouds at 13,000 ft. In cases like this, other data must be used to augment the information in the METAR reports. Finally, certain weather elements, like thunderstorms, are better represented by using remote-sensing data, such as lightning detection reports, rather than station-specific surface observations. How does one incorporate these varying types of data into one coherent forecast system?

In this paper, we describe the various sources of data used in our MOS development. The archival and quality control processes are outlined. We describe some of the algorithms used to derive various predictands, and discuss techniques used to augment the METAR observations. Finally, we present two predictands, snow amount and thunderstorms, that have been developed with data sources other than the hourly surface observations.