A causality-based perspective on Atlantic and tropical Pacific teleconnections to Northeast Brazil in seasonal forecast systems

Abstract. Seasonal forecast systems are routinely evaluated with skill metrics, which quantify forecast performance without revealing through which physical pathways it is achieved. In this study, we use the Liang–Kleeman (LK) causality formalism to evaluate how the Copernicus Climate Change Service (C3S) seasonal multi-system represents the causal influence of large-scale drivers from the tropical Atlantic, tropical Pacific and South Atlantic on surface wind and precipitation over Northeast Brazil (NEB), a region of strategic importance for renewable energy production where climate variability is strongly shaped by these drivers. The hindcast skill assessment, performed through the anomaly correlation coefficient (ACC) and mean squared error skill score (MSSS), reveals that for both variables the highest and most consistent predictive skill is concentrated around the February–May rainy season. In ERA5, the causal influences diagnosed with the LK formalism recover the known seasonality of the most prominent drivers, namely the Atlantic Meridional Mode, El Niño–Southern Oscillation (ENSO) and the South Atlantic Subtropical High, as well as their global teleconnection patterns. The C3S systems generally capture the emergence and disappearance of the causal links, although differing substantially in their strength, with no system performing consistently across drivers and initialisation months, and none capturing the austral winter ENSO influence on precipitation. Relating each system's fidelity in capturing the ERA5 causal strengths to its average skill shows that, for surface wind, systems with higher fidelity achieve higher MSSS, whereas no robust relationship is found for the ACC. Decomposing the MSSS reveals that this link, which is not explained by driver skill or ensemble size, arises from amplitude errors, where systems with less realistic teleconnections overestimate the amplitude of their forecast signal more strongly. Overall, this study shows that causality-based evaluation can complement conventional verification by identifying which forecast systems better reproduce the teleconnections underlying regional predictability, and positions the LK formalism as a diagnostic tool for process-oriented inter-model comparison across regions and timescales.