Reliability and skill of CMIP5 multi-model decadal hindcasts

The reliability of multi-year predictions of climate is assessed using probabilistic Reliability Diagrams for near-surface air temperature, based on 63 member ensembles of initialised and non-initialised decadal hindcasts using the CMIP5 ten-year decadal climate prediction experiments. It is shown that the reliability from the ensemble system is good for most regions studied, for both initialised and non- initialised hindcasts. Over the North Atlantic the reliability and the resolution of the hindcasts show a substantial improvement for initialised predictions, confirming the importance of the initialisation for this region. Initialisition seems to play a (positive) role in other regions as well, however more experiment are necessary to assess if the improvement is significant.