A causality framework to decipher prediction windows of opportunity

The evolution of the atmosphere on weekly to sub-seasonal time scales is driven by a combination of factors from different compartments of the Earth system. At these time scales, the atmospheric predictability can arise from the initial state of the atmosphere, the land surface and the ocean. The weight of these different sources and their joint contribution can change quickly depending on the season and the time scale considered. As an illustration, while subseasonal forecasts often exhibit limited skill across mid-latitudes, occasional improvements are observed in specific locations during certain periods, known as "windows of opportunity." Understanding these windows is complex due to the diverse and interdependent nature of predictors, their spatial and temporal variability, and the challenges in establishing causality relationships. A typical strategy could consist of assessing the time-lagged correlation or regression with potential predictors. The main caveat of this approach is that even a lagged relationship between two variables X and Y does not ensure causality, because of the many confounding factors involved in complex earth system interactions. Furthermore, a high correlation between X(t) and Y(t+dt) may just be the consequence of Y(t) causing X(t). In this study, we introduce a novel approach based on Liang-Kleeman information flow, allowing the assessment of causal links across various lead times. Applied to reforecast and reanalysis data, our method identifies significant predictability drivers, revealing their evolving patterns and prevalence from seasonal to subseasonal scales. Additionally, the comparison between reanalysis and reforecast results aids in assessing the capability of models to capture these causality features. We will illustrate the theoretical background by showcasing the causal factors influencing a window of opportunity identified from a multimodel subseasonal reforecast.