Spatial Correction of Multimodel Ensemble Subseasonal Precipitation Forecasts over North America Using Local Laplacian Eigenfunctions

The extent to which submonthly forecast skill can be increased by spatial pattern correction is examined in probabilistic rainfall forecasts of weekly and week-3-4 averages, constructed with extended logistic regression (ELR) applied to three ensemble prediction systems from the Subseasonal-to-Seasonal (S2S) project database. The new spatial correction method projects the ensemble-mean rainfall neighboring each grid point onto Laplacian eigenfunctions and then uses those amplitudes as predictors in the ELR. Over North America, individual and multimodel ensemble (MME) forecasts that are based on spatially averaged rainfall (e.g., first Laplacian eigenfunction) are characterized by good reliability, better sharpness, and higher skill than those using the gridpoint ensemble mean. The skill gain is greater for week-3-4 averages than week-3 leads and is largest for MME week-3-4 outlooks that are almost 2 times as skillful as MME week-3 forecasts over land. Skill decreases when using more Laplacian eigenfunctions as predictors, likely because of the difficulty in fitting additional parameters from the relatively short common reforecast period. Higher skill when increasing reforecast length indicates potential for further improvements. However, the current design of most subseasonal forecast experiments may prove to be a limit on the complexity of correction methods. Relatively high skill for week-3-4 outlooks with winter starts during El Nino and MJO phases 2-3 and 6-7 reflects particular opportunities for skillful predictions.