Persistent discrepancies exist between trends in coupled models and observations over the satellite era. Comparison is complicated by model biases, as well as differences in the phasing of coupled variability between the models and the observed system. To bridge this divide, we investigate the range of trends that are consistent with the retrospective forecasts (hindcasts) of an initialised near-term prediction system, and compare these to trends in the equivalent free-running, uninitialised model. We find that the initialisation has a large impact at low latitudes, leading to cooler surface temperature trends and less warming of the tropical upper troposphere than in the free-running model. By contrast, the initialisation appears to be a weaker influence on warming at high latitudes, despite tightly constraining sea ice trends. After only 4 months of integration, the hindcasts show a remarkably wide spread of Arctic amplification trends, pointing to the role of internal atmospheric variability in modulating the forced amplification. The hindcasts also lend support for emerging poleward shifts of the midlatitude circulation, offering a useful data point for setting uncertain observed trends in context. Regionally, the hindcasts better capture the Pacific SST trend pattern than the free-running model, and the observed winter Eurasian cooling.  Our results highlight the importance of intrinsic atmospheric variability in contributing to trends over the satellite era, and suggest plausible alternative trends that remained unseen.