Comment on “Comparison of Low-Frequency Internal Climate Variability in CMIP5 Models and Observations”

In a recent article, Cheung et al. applied a semiempirical methodology to isolate internal climate variability (ICV) in CMIP5 models and observations. The essence of their methodology is to subtract the scaled CMIP5 multimodel ensemble mean (MMEM) from individual model simulations and from the observed time series of several surface temperature indices. Cheung et al. detected large differences in both the magnitude and spatial patterns of the observed and simulated ICV, as well as large differences between the historical (simulated) ICV and preindustrial (PI) control CMIP5 simulations. Here it is shown that subtraction of the scaled MMEM from CMIP5 historical simulations produces a poor estimate of the modeled ICV due to the difference between the scaled MMEM and a given model’s true forced signal masquerading as ICV. The resulting phase and amplitude errors of the ICV so estimated are large, which compromises most of Cheung et al.’s conclusions pertaining to characterization of ICV in the historical CMIP5 simulations. By contrast, an alternative methodology based on forced signals computed from individual model ensembles produces a much more accurate estimate of the ICV in CMIP5 models, whose magnitude is consistent with the PI control simulations and is much smaller than any of the semiempirical estimates of the observed ICV on decadal and longer time scales.