Multi-model decadal predictions of probabilities for seasonal mean temperature and precipitation extremes

There is great interest from a wide range of stakeholders in near-term climate prediction ranging from seasonal to decadal timescales. While seasonal forecasting is done operationally since more than 20 years now, decadal climate prediction still has to be considered mainly a research subject. The vast majority of existing decadal prediction studies focusses on skill of temporally (typically multi-annual) averaged parameters. This is in line with the general understanding of climate prediction skill to be expectable only for low-frequency climate variability on larger spatial scales. However, while such predictions of multi-annual means might be skilful, they contain little information on shorter timescale extremes.

We present a different approach, that is the temporal pooling of seasonal means to form probabilistic forecasts. Thus, rather than for example analyzing the anomaly of the summer temperature averaged over a decade, we examine the probabilities of extreme seasonal summer temperatures within this decade (exceedance of some quantile of the climatological summer temperature probability distribution). This approach complements the common multi-annual means and hence extends the usability of decadal predictions.

For this study we use decadal climate predictions produced by the CMIP5 multi-model ensemble as well as available CMIP6-DCPP contributions. We analyze these large ensembles’ skill regarding forecasting the probability of extremely warm and extremely dry seasons. A season is considered to be “extreme” if the seasonal mean temperature (precipitation) is above (below) the 5^th (1^st) sextile of the climatological probability distribution.

We will show that the forecast skill in this respect is comparable to that obtained for the common approach, based on multi-annual year averages. This means the existence of significant skill for many regions globally when considering the probability of extremely warm temperatures. Skill regarding predicting extremely dry seasons (i.e. low precipitation) is rather limited, though.

These results generally agree with studies applying the common multi-annual averaging approach for assessing skill of temperature and precipitation climate predictions but extend the existing knowledge by covering probabilities of seasonal mean extremes. Hence, this approach states an important contribution towards the extended utility of decadal climate predictions. An additional benefit of the framework proposed here is the larger sample size when pooling instead of averaging. This allows to consider extreme events of higher magnitude before reaching the limitations of statistical uncertainty hampering the derivation of robust results.