What can the last century teach us about climate models?

Climate models are an important tool in our understanding of the climate system. Among other things, we use them together with initialisation procedures to predict the climate from a few weeks to more than a decade. While the community has demonstrated prediction skill for various climate modes on these time scales in the past years, we have also encountered problems. One is the non-stationarity of prediction skill over the past century in seasonal and decadal predictions. It was shown in multiple prediction systems and for multiple variables that prediction skill varies over time. Potential reasons for this non-stationarity was found in the changing state of the North Atlantic system on multi-decadal scales and the limited representation of physical processes within the model. While on the one side this feature of climate predictions leaves uncertainties for future predictions it also highlights windows of opportunity and challenges within climate models. 

We investigate the past century for this non-stationarity with a special focus on the North Atlantic Oscillation, and how the North Atlantic sector changes during these low prediction skill periods. We will demonstrate the limited predictability of features of the North Atlantic Oscillation, like the movement of its activity centres, as well as its implication for the Signal-to-Noise paradox. We also discuss the implications of non-stationarity model prediction skill for the development on future prediction systems and which processes are most likely the reason for the current challenges the community faces.