Evaluation of forecasts by accuracy and spread in the MiKlip decadal climate prediction system

Decadal climate prediction research gains progressively more attention in climate science as well as in society, industry and economy. The research aims to close the gap between short term forecasts and long term projections. Numerical weather predictions focus on an initial value problem in the beginning of a forecast. On the other hand, climate projections as a boundary condition problem examine the long-term development. In order to accommodate the demand for reliable information on near-term climate variability on the crucial timescales of a year up to a decade, different national and international initiatives like MiKlip have been launched.

Figure 1. For explanation, see text below.

In this study the MiKlip projects INTEGRATION and FLEXFORDEC present the forecasts and the skill assessment of the MiKlip decadal climate predictions. For this purpose, they employ a decadal evaluation tool called ‘MurCSS’, which was developed as part of the MiKlip Central Evaluation System. The authors point out the importance of a detailed evaluation, by combining initialized decadal climate predictions with their prediction skill using the MiKlip system. They accentuate the importance of specific statistical methods to evaluate the accuracy and the spread of the ensemble hindcast (retrospective forecasts) experiment. They present decadal forecasts and their prediction skill for near surface air temperature and precipitation for lead year 1 and lead years 2 to 9, as well as the improvement due to increased ensemble size.

As an example, shown here in Figure 1; the forecast is presented next to a specific evaluation of its appropriate hindcast set. The forecast (left) says, that we will have more (red) or less (blue) mm/day of rain or degrees Celsius than in the reference period 1981-2010. Next to the 2014 forecast of precipitation we put the lead year 1 correlation (bottom right) against the observation data set GPCP (1979-2012). And for the temperature forecast for 2015-2022 we show the lead year 2 to 9 correlation against the observation data set HadCRUT3v (1962-2012) (top right). In simple words: the higher the value in the correlation plot, the redder it is, the more we can trust that the forecast next to it appropriate.

This is just a simple illustration, of the aim of the study, the whole set up is a little bit more complex of course, the details of which can be found in the paper.


Back to List