Model Output Statistics (MOS) have been used in conjunction with Numerical Weather Prediction (NWP) models for decades. Early NWP model's were relatively crude and thus MOS played an essential role in extracting useful forecasts from the NWP model's data. However, the quality of NWP models has improved radically since, hence the margin for improvement of the NWP forecasts by the MOS process has diminished. Therefore, it is reasonable to ask whether or not MOS continues to provide value to the forecasts and if so by how much and when.

MOS systems generally operate by applying linear regression equations derived from histories of NWP forecast data to future runs of the model. The regression serves two purposes: it eliminates biases in the original NWP data, and it optimizes the conversion of the model data into forecast information. Since the bias correction can be achieved without employing a sophisticated regression system, the important part of the MOS process lies in the latter conversion.

It is the goal of this study to ascertain the value of using linear regression to improve upon the NWP forecasts after the data is extracted from the model and bias corrected. To that end, a local MOS system was constructed whereby forecasts are computed using an equation:

Final_Forecast=NWP_Forecast - HistoricalBias + MOS_RegressionCorrection

We analyze the characteristics of each of the terms in this equation to ascertain where value in the forecast arises. Data from the GFS model from the winter and spring of 2003 were used. The analysis was done at 50 cities around the continental U.S. Individual regressions were computed for each forecast variable, each forecast day (1-8) and each city. Maximum temperatures, probability of precipitation and wind speed forecasts were assessed. Regressions were screened using various thresholds of the R-squared metric and the results stratified by this metric. The NWP forecasts were extracted from the model data using simple, generic algorithms that converted the model's fields to forecast variables (e.g. convert precipitation rates and RH to probability of precipitation).

The results indicate that it is very difficult to find useful MOS-regression corrections to the extracted and bias-corrected GFS data. Averaged over all forecast days (1-8), the MOS correction terms generally caused an increase in overall RMS error, and reduced the forecast error less than half the time. For temperature forecasts, some value to the regression process was noted, but only in the first few days of the forecast period, and only when regression was used very judiciously (i.e. only when robust regressions were identified). For the POP forecasts, we found no circumstance in which the regression correction added value to the forecasts. The direct NWP-generic algorithm performed best and even the removal of the bias term was not found to add value.

Based on these experiments, we are led to believe that the value of traditional MOS applications to NWP is questionable. We find that using simple, generic methods to extract parameters from the NWP model, and applying local bias correction alone produces forecasts that generally can not be improved upon further using MOS/regression applications. Although the process we employ to extract and correct the NWP data could be considered a statistical post-processing of the data, it is far simpler and equally if not more skillful than employing a cumbersome regression system.