The annual cycle of surface air temperature is examined across Northern Hemisphere land areas (north of 25°N) by comparing the results from CRUTS against four reanalysis datasets: two versions of the Twentieth Century Reanalysis (20CR and 20CRC) and two versions of the ERA-CLIM reanalyses (ERA-20C and ERA-20CM). The Modulated Annual Cycle is adaptively derived from an Ensemble Empirical Mode Decomposition (EEMD) filter, and is used to define the phase and amplitude of the annual cycle. The EEMD method does not impose a simple sinusoidal shape of the annual cycle. None of the reanalysis simulations assimilate surface temperature data, but differ in the parameters that are included: both ERA-20C and 20CR assimilate surface pressure data; ERA-20C also includes surface wind data over the oceans; ERA-20CM does not assimilate any of these synoptic data; and none of the reanalyses assimilate land-use data. It is demonstrated that synoptic variability is critical for explaining the trends and variability of the annual cycle of surface temperature across the northern hemisphere. The CMIP5 forcings alone are insufficient to explain the observed trends and decadal-scale variability, particularly with respect to the decline in the amplitude of the annual cycle throughout the twentieth century. The variability in the annual cycle during the latter half of the twentieth century was unusual in the context of the twentieth century, and was most likely related to large-scale atmospheric variability, although uncertainty in the results is greatest before ca. 1930.