Consensus forecasts from the control runs of several operational numerical
models are compared to 1) the control-run forecasts of the individual model
s that compose the consensus and to 2) other consensus forecasts generated
by varying the initial conditions of the various individual models. It is f
ound that the multimodel consensus is superior to the individual control ru
ns and to the consensus forecasts constructed from ensembles of runs genera
ted by varying model initial conditions. The source of the forecast improve
ment by model consensus is not the result of a simple cancellation of error
s as a result of an overall positive bias in one model and an overall negat
ive bias in another. Rather the main improvement stems from overlapping dif
ferences in the sign of the errors associated with forecasts of individual
traveling disturbances. The results suggest that variations in model physic
s and numerics play a substantial role in generating the full spectrum of p
ossible solutions that can arise in a given numerical forecast.