Item

Abstract

Retrospective prediction experiments have to be performed to estimate the skill of decadal prediction systems. These are necessarily restricted in the number due to the computational constraints. From weather and seasonal prediction it is known that the ensemble size is crucial to yield reliable predictions. Differences are expected for decadal predictions due to the differing time-scales of the involved processes and the longer prediction horizon. A conceptual model is applied that enables the systematic analysis of ensemble size dependencies in a framework close to that of decadal predictions. Differences are quantified in terms of the confidence intervals coverage and the power of statistical tests for prediction scores. In addition, the concepts are applied to decadal predicitions of the MiKlip Baseline1 system. It is shown that small ensemble, as well as hindcast sample sizes lead to biased test performances in a way that the detection of a present prediction skill is hampered. Experiments with ensemble sizes smaller than 10 are not recommended to evaluate decadal prediction skill or as basis for the prediction system developement. For regions with low signal-to-noise ratios much larger ensembles are required and it is shown that in this case successful decadal predictions are possible for the Central European summer temperatures.