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

Abstract

We present the evaluation of temperature and precipitation forecasts obtained with the MiKlip decadal climate prediction system. These decadal hindcast experiments are verified with respect to the accuracy of the ensemble mean and the ensemble spread as a representative for the forecast uncertainty. The skill assessment follows the verification framework already used by the decadal prediction community, but enhanced with additional evaluation techniques like the logarithmic ensemble spread score. The core of the MiKlip system is the coupled Max Planck Institute Earth System Model. An ensemble of 10 members is initialized annually with ocean and atmosphere reanalyses of the European Centre for Medium-Range Weather Forecasts. For assessing the effect of the initialization, we compare these predictions to uninitialized climate projections with the same model system. Initialization improves the accuracy of temperature and precipitation forecasts in year 1, particularly in the Pacific region. The ensemble spread well represents the forecast uncertainty in lead year 1, except in the tropics. This estimate of prediction skill creates confidence in the respective 2014 forecasts, which depict less precipitation in the tropics and a warming almost everywhere. However, large cooling patterns appear in the Northern Hemisphere, the Pacific South America and the Southern Ocean. Forecasts for 2015 to 2022 show even warmer temperatures than for 2014, especially over the continents. The evaluation of lead years 2 to 9 for temperature shows skill globally with the exception of the eastern Pacific. The ensemble spread can again be used as an estimate of the forecast uncertainty in many regions: It improves over the tropics compared to lead year 1. Due to a reduction of the conditional bias, the decadal predictions of the initialized system gain skill in the accuracy compared to the uninitialized simulations in the lead years 2 to 9. Furthermore, we show that increasing the ensemble size improves the MiKlip decadal climate prediction system for all lead years.