Probabilistic interpretation of regression-based downscaled seasonal ensemble predictions with the estimation of uncertainty

A regression‐based method of statistical downscaling from global multimodel ensemble (MME) forecasts has been developed. This method is appropriate for seasonal forecasts at stations with the use of model outputs as predictors; it refers to a technique that is known as “model output statistics” (MOS). Downscaled forecasts are formulated in terms of tercile probabilities based on the probabilistic interpretation of the forecast uncertainty. The novelty of the method is in the estimation of uncertainties originating from both regression and ensemble spread of model forecasts within the framework of the regression analysis. The method has been tested on the prediction of wintertime temperature and precipitation for 60 Korean stations by downscaling from the MME forecasts of 850 hPa temperature, sea level pressure, and 500 hPa geopotential height. Different sources of uncertainty associated with regression and ensemble spread have been evaluated and their contributions compared. It is shown that although the uncertainty associated with the deviation from the linear model is usually the largest, a comparable contribution to the uncertainty can come from the ensemble spread. Verification assessments of the method show that downscaled probabilistic MME forecasts essentially outperform the forecasts interpolated from the raw model predicted anomalies for both temperature and precipitation. Key Points Regression‐based method of statistical downscaling Estimation of uncertainties from both regression and ensemble spread Downscaled probabilistic multimodel ensemble forecasts