Extended Data Fig. 3: Sensitivity of model-data comparison to the choice of reconstructed variable.

a, Comparing the average spectra at the location of pollen records of model simulations, instrumental data (BEST) and reconstructions for the mean summer temperature (dashed) and the mean annual temperatures (solid); logarithmically spaced axes were used. Shading indicates 90% confidence intervals around the mean. The IPSL and ECHAM5 model results exhibit a slightly lower variability in their annual temperature than in their summer temperature over all timescales, except for the longest timescale since it is dominated by leaked power from the Earth’s orbital precession, which mainly affects summer temperature in the Northern Hemisphere during the Holocene. On the other hand, TraCE-21ka generally shows a slightly higher variability in its annual compared to its summer temperature. Although the pollen-based reconstructions calibrated for annual temperature are thought to be less reliable than the summer temperature reconstructions, they give a very similar result. This shows that our conclusions are robust against uncertainties in the seasonal attribution of pollen variability. b, Same as a, but for precipitation instead of summer temperature. While most locations should reflect temperature, here we also tested the boundary case of assuming that all sites reflect precipitation. Even in this extreme case, the main results hold, namely increasing climate variability over land as a function of timescale and a corresponding deficit of variability in the climate models. The three climate models vastly disagree in terms of the amplitude of precipitation variability, but they all show temporal scaling similar to the temperature variability; this is likely caused by the temporal links between precipitation and temperature on long timescales⁸⁴.