Assessing the Use of Ensemble Regression and Time-Lagged Ensembles on GEFS Forecasts of Temperature and Precipitation

Forecasting temperature and precipitation at subseasonal timescales is difficult due to the gap in predictability contributed by atmospheric and surface forcings. Multimodel ensembles have been used to increase the skill of Subseasonal-to-Seasonal (S2S) forecasts by improving probability distributions and signal-to-noise ratios relative to single model forecasts. Time-lagged ensembles – in which the same model and valid time are used while the initialization time varies – have also been applied to ECMWF and CFSv2 forecasts at S2S timescales. However, the benefits of using a time-lagged ensemble for GEFS forecasts have yet to be evaluated. GEFS forecasts are currently produced daily and have 31 ensemble members. The daily initializations allow for a sizable time-lagged ensemble to be constructed, though only two initialization times (Tuesdays and Wednesdays) are used to create a 62-member time-lagged ensemble and predict the Week 3-4 temperature and precipitation. In addition, ensemble regression is another technique that may be used to improve the reliability of probabilistic forecasts through an adjustment of the model’s probability distribution function. We use twenty years of GEFS hindcasts (2000-2019) and a cross validation leave-1-year-out methodology to evaluate the additive skill of using ensemble regression to predict temperature and precipitation at Weeks 1-5. To analyze the skill of both aforementioned probabilistic (tercile) ensemble systems, verification is performed using metrics such as Brier Skill Score (BSS) and Heidke Skill Score (HSS), and results are displayed by season (DJF, MAM, JJA, SON).