Improving the Prediction of Winter Precipitation and Temperature over the Continental United States: Role of the ENSO State in Developing Multimodel Combinations
Abstract Recent research into seasonal climate prediction has focused on combining multiple atmospheric general circulation models (GCMs) to develop multimodel ensembles. A new approach to combining multiple GCMs is proposed by analyzing the skill levels of candidate models contingent on the relevant predictor(s) state. To demonstrate this approach, historical simulations of winter (December–February, DJF) precipitation and temperature from seven GCMs were combined by evaluating their skill—represented by mean square error (MSE)—over similar predictor (DJF Niño-3.4) conditions. The MSE estimates are converted into weights for each GCM for developing multimodel tercile probabilities. A total of six multimodel schemes are considered that include combinations based on pooling of ensembles as well as on the long-term skill of the models. To ensure the improved skill exhibited by the multimodel scheme is statistically significant, rigorous hypothesis tests were performed comparing the skill of multimodels with each individual model’s skill. The multimodel combination contingent on Niño-3.4 shows improved skill particularly for regions whose winter precipitation and temperature exhibit significant correlation with Niño-3.4. Analyses of these weights also show that the proposed multimodel combination methodology assigns higher weights for GCMs and lesser weights for climatology during El Niño and La Niña conditions. On the other hand, because of the limited skill of GCMs during neutral Niño-3.4 conditions, the methodology assigns higher weights for climatology resulting in improved skill from the multimodel combinations. Thus, analyzing GCMs’ skill contingent on the relevant predictor state provides an alternate approach for multimodel combinations such that years with limited skill could be replaced with climatology.