Recent experiments have been conducted to test the viability of a dynamic, ensemble MOS approach for weather forecasting. A dynamic MOS technique has been developed in which only recent NWP model output and verification data are used to form predictive regression equations. The equations are then used for subsequent weather forecasts until new regression equations are calculated. This technique is similar to that described by Mao et al. (1997) although here it has been applied to a much greater variety of weather forecast variables.

An extension of the technique has been developed in which numerous semi-independent regression equations for each weather variable are solved and stored. This entire set (or ensemble) of regressions then can be solved at forecast time to produce an ensemble of estimates of the predicted variable. A simple procedure to quality-control and median procedure is then employed to reduce the ensemble set to a single final predictor. This ensemble extension to standard MOS techniques is justified by the substantial overdetermination of regression equations using standard model data output. Experiments on this technique have shown that considerable improvement in the quality of the forecasts over the single regression dynamic MOS techniques can be acheived. In particular, the ensemble approach provides a simple way to detect and remove bad regressions that can often arise when using a short input datasets in a dynamic MOS scheme.