A Priori Identification of Skillful Extratropical Subseasonal Forecasts

The current generation of extratropical subseasonal operational model forecasts has, on average, low skill for leads beyond 3 weeks. This is likely a fundamental property of the climate system, due to the relative high amplitude of unpredictable synoptic variability compared to potentially predictable, but generally weaker, climate signals. Thus, for subseasonal forecasts to be useful, their high versus low skill events should be identified at time of forecast. Unfortunately, spread-skill type relationships typically provide relatively limited guidance beyond forecast Week 3. We show that a linear inverse model (LIM), an empirical-dynamical model constructed from covariability statistics of wintertime (December-March) weekly-averaged observational analyses, can be used to identify, a priori, the expected subseasonal surface and mid-tropospheric forecast skill. Using the LIM’s predicted signal-to-noise ratio, we identify the small subset (10-30%) of Weeks 3-6 forecasts - of the LIM and two operational models from NCEP and ECMWF - with relatively higher skill versus the much larger remainder of forecasts whose skill cannot be distinguished from random chance.

To evaluate the skill of the LIM and to contrast our LIM-based signal-to-noise skill forecasting technique with a traditional spread-skill relationship, cross-validated hindcasts generated using the LIM are compared to bias-corrected WCRP Subseasonal-to-Seasonal Prediction Project ECWMF (1997-2016) and NCEP CFSv2 (1999-2010) winter season hindcasts. The LIM is constructed from Japanese Reanalysis (JRA-55, 1979-2016) and accounts for interactions between tropical heating, mean-sea level pressure (MSLP), and extratropical tropospheric and stratospheric circulation variability. Skill is evaluated for anomalous MSLP and 500 hPa geopotential height.

For all three models, we find that while median forecast skill at leads beyond Week 3 is quite low, a small subset of forecasts has notably higher skill occurring more frequently than expected from random chance. For example, forecasts initialized for the upper 10% of LIM signal-to-noise ratio yield actual LIM and ECMWF Week 5 Pacific MSLP forecast skills in the 0.5-0.6 range, while the remaining 90% of forecasts yield skills of only 0.1-0.15. In general, we find that the LIM-based skill forecasts are superior to those based on a standard spread-skill relationship. Similar results are achieved for the North Atlantic basin.