Verification Data and the Skill of Decadal Predictions

The utility of a forecast depends on its skill as demonstrated by past performance. For most forecasts errors rapidly become large compared to uncertainties in the observation-based state of the system and, for this reason, it is usually deemed adequate to assess predictions against a single verification dataset. Eleven reanalyses and station-based analyses of annual mean surface air temperature are compared as are basic skill measures obtained when using them to verify decadal prediction hindcasts from the Canadian Centre for Climate Modelling and Analysis forecasting system. There are differences between reanalysis and station-based analyses which translate also into differences in basic skill scores. In an average sense, using station-based verification data results in somewhat better correlation skill. The spread between the locally best and worst scores is obtained for individual forecast ensemble members and for ensemble mean forecasts compared to individual analyses. The comparison of ensemble mean forecasts against different analyses can result in apparent skill differences, and using a “favorable” analysis for verification can improve apparent forecast skill. These differences may be more pertinent for longer time averages and should be considered when verifying decadal predictions and when comparing the skill of decadal prediction systems as part of a model intercomparison project. Either a particular analysis could be recommended by the decadal prediction community, if such could be agreed on, or the ensemble average of a subset of recent analyses could be used, assuming that ensemble averaging will act to average out errors.