2011 Articles
Recalibrating and Combining Ensemble Predictions
The “model output statistics” (MOS) approach has long been used in forecasting to correct systematic errors of numerical models and to predict quantities not included in the model (Glahn and Lowry 1972). The MOS procedure is based on capturing the statistical relation between model outputs and observations and, in its simplest form, consists of a linear regression between these quantities. In theory, this procedure optimally calibrates the model forecast and provides reliable forecasts.
In practice, the regression parameters must be estimated from data. In seasonal forecasting, forecast histories are short, and skill is modest. Both factors lead to substantial sampling errors in the estimates. This work examines two problems where sampling error affects the reliability of regression-calibrated forecasts and provides solutions based on two “penalized” methods: ridge regression and lasso regression (Hoerl and Kennard 1988; Tibshirani 1996). The first problem comes from the observation that, even in a bivariate setting, ordinary least squares estimates lead to unreliable forecasts. The second problem arises in the context of multivariate MOS and is that common methods of predictor selection lead to negative skill and unreliable forecasts.
Subjects
Files
-
recalibrating-and-combining-ensemble-predictions.pdf application/pdf 488 KB Download File
Also Published In
- Title
- Science and Technology Infusion Climate Bulletin
- Publisher
- NOAA's National Weather Service
More About This Work
- Academic Units
- International Research Institute for Climate and Society
- Published Here
- March 13, 2024