Fig. 5. Predictive skill in the imperfect model world and prediction of global mean surface temperature evolution conditional on time series of annual global mean temperatures from version 5 of the UK Met Office Hadley Centre/Climatic Research Unit (HadCRUT) data set HadCRUT5.

(A) The root mean square error (RMSE) of constrained predictions (colored) and unconstrained (gray) relative to pseudo-observations simulated by ACCESS-ESM1.5 during the future 20-year period relative to the current climate (defined as the last 30 years of the constraining window), with lead times of 20 and 50 years under the SSP1-2.6 and SSP5-8.5 emission scenarios. The first year of constraining window is 1901, while the last year corresponds to the year before the release of the IPCC first to seventh assessments [IPCC First Assessment Report (FAR)1990, its Second Assessment Report (SAR) 1995, its Third Assessment Report (TAR) 2000, its 4th Assessment Report (AR5) 2006, its 5th Assessment Report (AR5) 2012, its 6th Assessment Report (AR6) 2020, and its anticipated 7th Assessment Report (AR7) 2026], plus two additional constraining windows that end in 2036 and 2046. ACCESS-ESM1.5 results are shown because it has climate sensitivity similar to the best estimate sensitivity of the observed climate (62). (B) Historical observations from 1901 to 2023, unconstrained CMIP6 projections, and future predictions. The left panel shows time series, with the predictions in color and unconstrained projections in gray, with their corresponding 5th to 95th percentile ranges shaded. The right panels display warming of the future periods with 20- and 50-year lead time, with mean changes shown as dots and 5th to 95th percentile ranges represented by error bars. The warming relative to the preindustrial period can be determined by applying an offset between 1994–2023 and 1851–1900, which is estimated to be approximately 0.95 [0.89, 1.02] °C based on the HadCRUT5 dataset.