An Objective Comparison of Model Output Statistics and “Perfect Prog” Systems in Producing Numerical Weather Element Forecasts

Abstract

The ?perfect prog? (PP) and model output statistics (MOS) approaches were used to develop multiple linear regression equations to forecast probabilities of more than a trace of precipitation over 6-h periods, probabilities of precipitation amounts over 12-h periods at three threshold levels, and sky cover and spot temperatures at 3-h intervals. The MOS forecasts are prepared twice a day in the Canadian Meteorological Centre production cycles and are used in the Canadian forecasting system. The experimental PP system produces forecasts twice daily in a parallel run. An independent set of forecasts was produced out to 72 h for verification purposes for the two systems, covering the 12 months of 1986. The production of maximum and minimum temperature forecasts for the climatological day based on the perfect prog approach has also been part of the operational production cycles for many years. The decomposed Brier score and signal detection theory were used to do a comparative verification of the PP and MOS forecasts. The verifications showed the attributes and advantages of each system. The PP forecasts had higher skill scores for the shorter range periods, showed greater sharpness, greater stability of the equations due to a larger dependent database, and independence from the driving model. The MOS forecasts had higher skill scores at the longer range periods and were more reliable, especially at the longer projection times since they implicitly take care of the driving model's limitations. The object of this study was to determine?if the advantages were apparent?which system to implement in the operational production cycles. If the advantages are not apparent, a consensus approach will be proposed. Some guidelines, including the use of rule-based systems, will be presented to determine how to best integrate the output of both systems.