FIGURE 3. Hindcast/forecast decadal means of selected time series, compared with observations and climate model projections including only radiative forcing changes but not initialized using ocean observations.

Model projections are twentieth century-RF followed by A1B scenario simulations ('20C-RF/A1B'). a, Maximum MOC strength at 30° N; b, Atlantic SST dipole index (60–10° W, 40–60° N minus 50–0° W, 40–60° S SST area averages), which is constructed to isolate MOC forced SST fluctuations from radiatively forced variations²⁵; c, European land surface temperature (5° W–10° E, 35–60° N average); d, North American surface temperature (120–70° W, 30–50° N average); and e, eastern Tropical Pacific SST (150–90° W, 20° S–20° N average). a–e, Each point represents a ten-year centred mean; vertical bars indicate ensemble spread; verification and forecast periods are indicated (dark shading begins 2008, indicating the start of the true forecast period); two additional forecasts with radiative forcing stabilized at year 2000 levels are also shown. b–e, Correlation of both hindcasts and climate model projections with observations are given in brackets. Owing to insufficient observations, in a the correlation with the MOC from SST-restored simulation is given. Hindcast correlation skill is significantly greater at the 5% level than that of the standard climate projection in all cases (significance test described in Methods). Scale on right is used for the SST-restored simulation and observations; note the different ranges for right and left axes in a and b. SST are from HADISST²⁷, land temperature from CRUTEMP3²⁸ ('CRU3').