Methodological aspects of the validation of decadal predictions

Validation techniques based on past performance are well-established in seasonal forecasting. So far there is no consensus on the degree to which these are applicable to decadal predictions. We contribute to this discussion by assessing the effects of drift-correction, cross-validation and de-trending. The study employs decadal hindcasts of 2 m temperature from the EU FP6 ENSEMBLES project database and a synthetic toy model. Decadal predictions can be subject to substantial lead-time dependent model drifts. The conventional drift-correction method has a considerable sampling uncertainty, amounting to up to 40% of the potentially predictable signal. Introducing a smooth drift curve allows this uncertainty to be reduced by about 30% for annual values. For drift-corrected decadal predictions the leave-one-out cross-validation procedure may lead to biased skill estimates for decadal prediction due to the small number of hindcasts available. We identify this effect and show that 'jackknifing' represents a suitable technique for estimating potential skill without bias and to estimate sampling uncertainty. Results indicate significant correlation skill on the order of 0.7 to 0.9 for predicting global annual mean temperature on all lead-times. If linear trends are removed prior to verification, skill is still clearly above 0 in the first year for global mean temperature. On a local scale, some specific regions exhibit skill additional to the trend even at longer lead times. With the limited dataset analyzed here, the strong sampling uncertainty still prohibits drawing a final conclusion, by means of verification, on whether or not decadal predictions have skill in predicting climate variability beyond the trend.