Abstract/Summary

Initialised hindcasts inherit knowledge of the observed climate state, so studies of multidecadal trends in seasonal and decadal hindcast models have focused on the ensemble-mean when benchmarking against observed trends. However, this neglects the role of short-timescale variability in contributing to longer-term trends, and hence trend errors. Using a single-model coupled hindcast ensemble, we generate a distribution of 10,000 hindcast trends over 1981-2022 by randomly sampling a single ensemble member in each year. We find that the hindcast model supports a wide range of trends in various features of the large-scale climate, even when sampled at leads of just 1-3 months following initialisation. The spread in hindcast global surface temperature trends is equivalent to approximately a sixth of the total observed warming over the same period, driven by large seasonal variability of temperatures over land. The hindcasts also lend support for observed poleward jet shifts, but the magnitude of the shifts varies widely across the ensemble. Our results show that a fair comparison of hindcast trends to observations should consider the full range of model trends, not only the ensemble mean. More broadly, we argue that the hindcast trend distribution offers a largely untapped tool for studying multidecadal climate trends in a very large ensemble, through exploiting existing hindcast data.