Abstract
Multi-model decadal predictions and probabilistic climate scenarios are compared and evaluated, using observations from 1961 to 2023. For averages of years 2–9 ahead, initialised predictions show higher skill for Atlantic Multidecadal Variability (AMV) and the winter North Atlantic Oscillation (NAO) but not for global mean surface temperature (GMST). The correlation skill for NAO depends on the composition of the multi-model ensemble, the length of hindcast epoch and choices of climatological baseline, months and regions used to define NAO index anomalies. For northern Europe and England & Wales, surface temperature and precipitation in winter and summer show positive trends, and both multi-model systems provide similar skill at the decadal time scale. The only exception is summer precipitation in England & Wales, for which the probabilistic scenarios predict a recent drying trend not yet seen in observations. Better simulation of teleconnections between AMV and downstream European anomalies, or use of the observed relationship in statistical methods, might improve initialised predictions of summer precipitation. The uncertainty ranges from both multi-model systems vary in reliability, with the probabilistic scenarios providing conservative ranges in most cases. For predictions of individual years, initialisation exerts little constraint on spread (except for GMST and AMV) but reliability could be improved by adjusting for biases in simulated variability. For decadal anomalies, modelling and initialisation uncertainties also influence spread and reliability for most of the variables assessed. In summary, the initialised and uninitialised distributions both show value, but diagnosis of near-term risks could be improved by optimising and combining their information in future work.