ESRL Seasonal Forecast Guidance for US Temperature and Precipitation

ESRL has developed a seasonal forecast guidance tool, which is a blend of dynamical and empirical methods. It has been derived from the dynamical climate simulation data from an ensemble of four different atmospheric general circulation models (NCEP-MRF9, GFDL-R30, NCAR-CCM3, IRI-ECHAM3) forced by known, monthly sea surface temperature fluctuations between 1950-1999.

The approach involves training a multi-variate regression model on the relationship between observed tropical SST and GCM US seasonal climate for the independent training period of 1950-94. The co-variance matrix that relates tropical SST predictors and US predictands, was computed separately for each of the 12-overlapping seasons. This method has been used to generate seasonal hindcasts for US temperature and precipitation for each season during 1995-2003. The skill scores of these hindcasts, including their comparison with CPC Official Seasonal Outlook skill scores, can be found here.

The lag-covariance matrix and the ensuing hindcasts/forecasts have been developed without any input from the observed predictands, U.S. temperature and precipitation. The only observational data entering this ESRL forecast guidance is the tropical SST predictor field. As an example, in the lag-covariance matrix, fields of observed SSTs for SON are related to GCM predictands fields over the U.S. for DJF.

In forecast mode, the ESRL method yields a "preliminary" and "final " guidance for US temperature and precipitation. The preliminary output is issued at 1-month lead, whereas the final output is issued at 0-month lead. Thus, for the DJF target season, the initial SST state used to generate the "preliminary guidance" is the observed Sept/Oct tropical SST anomalies,. The October SST anomalies are based on the first 4-weeks of the weekly Reynold's analysis, and the guidance is issued by 1 November. The "final guidance" uses as its initial SST the October/November tropical SST anomalies, and the November anomalies are based on teh first 4 weeks of the weekly Reynolds SST analysis. This guidance is issued by 1 December.

How is this tool related to the ESRL experimental seasonal prediction tool?

The approach outlined above is different from what is done on the ESRL experimental seasonal prediction site. There, the co-variance matrix has been based on simultaneous predictor/GCM predictand relations, and SST forecasts are projected onto the predictor EOF basis. This method mimicks a two-tiered system. The ESRL seasonal guidance described above is effectively a "one-tiered" approach, and uses the lag-covariance relation between predictor and predictand.

Also, the hindcasts for 1995-2003, and the real-time ESRL forecast guidance, are for a 3-category system of above, below, and normal., as opposed to the 2-category approach for the ESRL experimental prediction tool. In making this 3-category probabilistic hindcast, we make the assumption of a Gaussian probability distribution function of the GCM predictand data, and further assumes that the PDF of the GCM data is representative of the that observed.