Multi-annual predictions of daily temperature and precipitation extremes: forecast quality and impact of model initialisation

Both global warming and internal climate variability modulate changes in the intensity and frequency of extreme climate events. Anticipating such variations years in advance may help minimise the impact on climate-vulnerable sectors and society, as well as enable short-term adaptation strategies and early-warning systems in a changing climate. Decadal climate predictions are a source of climate information for multi-annual timescales. They are provided by climate forecast systems similar to the models used for long-term climate projections but that have been initialised with the best estimate of the contemporaneous conditions of the climate system. However, before using the predictions, the forecast quality should be assessed. This is an essential step to evaluate the accuracy of the predictions and find windows of opportunity (variables/indices, regions and forecast times) to provide climate services with data of sufficient quality to satisfy the user requirements.

We evaluate the deterministic and probabilistic forecast quality of the multi-model ensemble built with all the available decadal hindcasts (i.e., retrospective decadal predictions) contributing to CMIP6, which consists of a total of 133 ensemble members from 13 forecast systems. The forecast quality assessment has been performed for predictions of seasonal and annual indices of daily temperature and precipitation extremes for the forecast years 1-5. These indices measure the intensity and frequency of hot and cold temperature extremes, and the intensity and rainfall accumulation related to heavy precipitation extremes. The prediction skill for the temperature and precipitation extreme indices is further compared to the skill for mean temperature and precipitation, respectively. In order to assess the impact of the model initialisation, the predictions are compared against historical forcing simulations (i.e., retrospective climate projections) created with the same models, consisting of a total of 134 ensemble members from the same forecast systems as the decadal hindcasts.

We find that the decadal hindcasts skillfully predict both mean and extreme temperature indices over most of the globe for multi-annual periods. The forecast quality for mean precipitation and extreme precipitation indices is generally low, and significant skill is found only over some limited regions. The reduced quality of the precipitation predictions with respect to temperature is due to the relatively smaller effect of human-induced warming for this variable. The comparison between the skill for mean variables and extreme indices shows that the extreme indices are generally predicted with lower skill, especially those related to the intensity of extreme events. We find generally small and region-dependent improvements from model initialisation compared to historical forcing simulations. The added value due to initialisation is higher for the mean variables than for the extreme indices. Besides, such skill differences differ between indices, especially those representing extreme temperature. This systematic evaluation of decadal hindcasts is essential when providing a climate service based on decadal predictions so that the user is informed about the trustworthiness of the forecasts for each specific region and extreme event. Also, comparing decadal hindcast and historical simulations may help climate services providers select the highest-quality information from these different data sources.