Dynamical trends in initialised hindcasts

In initialised seasonal and decadal prediction systems, retrospective forecasts (“hindcasts”) are routinely used to assess model skill on seasonal or interannual timescales. Here, we introduce their use for studying multidecadal trends over the satellite era using the UK Met Office’s DePreSys model. An ensemble of 40 members is initialised twice annually since 1980 and run for between 13-66 months. Each hindcast simulation can be seen as a modelled snapshot, constrained by the initial conditions; by studying trends between successive snapshots, a distribution can be built of plausible alternative histories that are consistent with the observed climate state. Compared to equivalent histories in free-running model ensembles, the frequent initialisation of the hindcasts serves to pull the model back towards the observed state and minimise the development of biases, while still allowing some time for the coupled climate system to evolve. The resulting distribution offers a bridge between observations and free-running climate models.
We first outline the above methodology before presenting a survey of trends in the hindcasts. Our initial focus is on zonal-mean dynamics, including robust shifts of the zonal wind maxima in the hindcasts, and with trends in the tropics emerging as a region of interest. The method introduced here can be used to study the impact of model biases on modelled trends, and also aims to provide context for emerging trends in the climate system.