Decadal predictability of seasonal temperature distriubutions

Climate predictions focus regularly on the predictability of single values, like means or extremes. While these information offer important insight into the quality of a prediction system, some stakeholders might be interested in the predictability of the full underlying distribution. These allow beside evaluating the amplitude of an extreme also to estimate their frequency. Especially on decadal time scales, where we verify multiple lead years at a time, the prediction quality of full distributions may offer in some applications important additional value.

In this study we investigate the predictability of the seasonal daily 2m-temperature on time scales of up to ten lead years within the MPI-ESM decadal prediction system. We compare yearly initialised hindcast simulations from 1960 onwards against estimates for climatology and uninitialised historical simulations. To verify the predictions we demonstrate a novel approach based on the non-parametric comparison of distributions with the integrated quadratic distance (IQD).

We show that the initialised prediction system has advantages in particular in the North Atlantic area and allow so to make reliable predictions for the whole temperature distribution for two to ten years ahead. It also demonstrates that the capability of initialised climate predictions to predict the temperature distribution depends on the season. Finally, we will also discuss potential opportunities and pitfalls of such approaches.