How reliable are decadal climate predictions of near-surface air temperature?

Decadal climate predictions are being increasingly used by stakeholders interested in the evolution of climate over the coming decade. However, investigating the added value of those initialized decadal predictions over other sources of information typically used by stakeholders generally relies on forecast accuracy, while probabilistic aspects, although crucial to users, are often overlooked. In this study, the quality of the near-surface air temperature from initialized predictions has been assessed in terms of reliability, an essential characteristic of climate simulation ensembles, and compared to the reliability of noninitialized simulations performed with the same model ensembles. Here, reliability is defined as the capability to obtain a true estimate of the forecast uncertainty from the ensemble spread. We show the limited added value of initialization in terms of reliability, the initialized predictions being significantly more reliable than their noninitialized counterparts only for specific regions and the first forecast year. By analyzing reliability for different forecast system ensembles, we further highlight the fact that the combination of models seems to play a more important role than the ensemble size of each individual forecast system. This is due to sampling different model errors related to model physics, numerics, and initialization approaches involved in the multimodel, allowing for a certain level of error compensation. Finally, this study demonstrates that all forecast system ensembles are affected by systematic biases and dispersion errors that affect the reliability. This set of errors makes bias correction and calibration necessary to obtain reliable estimates of forecast probabilities that can be useful to stakeholders.