Insuring Temporal and Spatial Consistency in Short Range Statistical Weather Forecasts

Statistical forecasts have been produced for many years by a number of organizations and with a variety of techniques. Purely Markov models have been used for very short range forecasts. Post processing of Numerical Weather Prediction (NWP) model data has been done in many different ways to produce interpretative guidance out to several days in advance.

Many times stepwise regression is employed where a plethora of potential predictors are "screened" to produce regression equations. Usually, the selection process is based on minimizing the Root Mean Square Error (RMSE) of a predictand, and a suitable stopping procedure is used to decide on the number of predictors to include in the equations. With suitable data samples, equations can be produced for one or more variables, for several locations, and for several projection in time. Consistency of the forecasts produced by these equations is of concern--consistency (1) among variables, (2) among spatial locations, (3) among projections made from one start time, and (4) among forecasts valid at the same time, but made at different times. Certainly, meteorological conditions may warrant significant and even rapid changes in all of these aspects, but many times the conditions are such to suggest the guidance should exhibit consistency.

This paper will describe techniques to help assure consistency spatially and temporally, and show results of applying these techniques.