3B.7 Application of a Hybrid Statistical-Dynamical Prediction System to Seasonal Forecasts of North American Temperature

Friday, 28 July 2017: 3:00 PM
Constellation F (Hyatt Regency Baltimore)
Sarah E. Strazzo, Innovim, LLC, Greenbelt, MD; and D. C. Collins, Q. J. Wang, and A. D. Schepen

Handout (4.8 MB)

Although the dynamical models relied upon in subseasonal-to-seasonal climate forecasting generally produce more skillful forecasts than purely statistical methods, these models sometimes fail to represent observed teleconnection patterns of importance to seasonal climate variability. For example, the observed relationship between the El Niño/Southern Oscillation (ENSO) and North American wintertime 2-m temperature is not consistently reproduced by the members of the North American Multi-Model Ensemble (NMME), particularly across large swaths of the northern United States. This inability to capture an important teleconnection contributes to poor model forecast skill over these regions during the winter season (December--February, DJF). We show that application of statistical post-processing leads to improvements in forecast skill. Specifically, we apply the Calibration, Bridging, and Merging (CBaM) framework described by Schepen et al. (2014) and Schepen et al. (2016) to post-process NMME forecasts of North American temperature. Calibration models developed from dynamical model hindcasts and observations are used to statistically correct dynamical model forecast fields. Bridging models use dynamical model forecasts of relevant climate indices (e.g., ENSO) as predictors of North American temperature. Bridging and calibration models are first developed separately using Bayesian joint probability modeling and then merged using Bayesian model averaging to yield an optimal forecast. We apply the CBaM method to forecasts of North American 2-m temperature from NMME hindcast data. Bridging is implemented using model forecasts of the Niño 3.4 index. Seasonal and lead-dependent differences between calibration and bridging are examined. Forecast skill is assessed using attributes diagrams and Brier skill scores for probabilistic forecasts of above/below normal 2-m temperature. We find that although calibration generally yields more skillful 1-month lead forecasts for most of the overlapping three-month seasons, bridging improves upon calibrated model skill for forecasts of DJF 2-m temperature. The largest improvements occur over portions of the northern United States, where the observed ENSO-temperature teleconnection is not well-represented by the NMME models.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner