Signs of Surprising Skill in Dynamic Predictions Beyond the 2-Week Limit of Predictability—Evaluation of the Operational CFSv2 for Week 3 and 4

NWS is committed to issuing week3&4 (time mean for days 15-28) forecasts, to bridge the gap between weather and short-term climate prediction. Many in our field may dismiss forecasts for this range, since it is beyond the Lorenzian 2-week limit of predictability. One of the few US dynamical tools that can be used immediately, for this most challenging range of forecasting, is the NCEP Climate Forecast System version 2 (CFSv2, Saha et al, 2014). This T126L64 coupled atmosphere-ocean-seaice model produces 16 ensemble members each day, that go out to 45 days, since its operational implementation in March 2011. The initial states are from the Climate Forecast System Reanalysis (CFSR, Saha et al, 2010). In addition, retrospective forecasts are available four times daily (from each 0, 6, 12 and 18Z cycles) out to 45 days, over the period 1999-2010. This extensive dataset has been archived, every 6 hours of the forecast length, for a large selection of variables. It has also been exploited for calibration of CFSv2 in real time, as well as for computing skill estimates for forecasters to use when making the operational weekly and monthly outlooks at CPC. The verification data (for most variables) is from the CFSR on a 1x1 degree grid.

Many technical questions arise:

• How does one correct for mean error? The answer is by using harmonic climatologies for both forecasts (lead dependent) and verification over the period 1999-2010.

• Does this correction method have beneficial impacts on the independent data for 2011-15 forecasts generated in real time? The answer is yes. The calibration is applicable in spite of some unavoidable changes in the ‘system' since 2011.

CFSv2 has some modicum of skill for November – May, but almost none in NH summer. Among all the tools one can apply, the ensemble mean may be the most efficient in increasing the signal to noise ratio. Therefore 32 ensemble members (2-day lagged) may be best in the current operational set-up. For some variables, taking a 14 day mean is also helpful to reduce noise.

For 500hPa geopotential, the skill from November thru May appears related to 1 or 2 EOF modes, most notably the (N)AO. In many months, these modes can be predicted for week3&4 at 0.6 anomaly correlation, or better. The skill over the operational period 2011-2015 appears higher than for the retrospective period 1999-2010, perhaps because the initial states are getting better, or because nature projects more (by accident) onto these most predictable modes in recent years. Verification of observed/analysed surface fields (such as CPC ‘daily' temperature) is more complicated, but has been attempted. Plans for some extension of CFSv2 forecast menu (like more members from 0, 6, 12Z cycles) will be presented.

Saha et al (2010) The NCEP climate forecast system reanalysis. Bull Am Meteor Soc 91:1015–1057. doi:10.​1175/​2010BAMS3001.​1

Saha et al (2014) The NCEP climate forecast system version 2. J Clim 27:2185–2208. doi:10.​1175/​JCLI-D-12-00823.​1