Signal‐to‐noise and predictable modes of variability in winter seasonal forecasts

Abstract

AbstractRecent studies suggest seasonal forecasts for European winters are now skilful, but they also identify a “signal‐to‐noise paradox”, wherein models predict the real world more skilfully (higher correlation) than the evolution of their ensemble members. Here, we analyse seasonal hindcasts from the Met Office GloSea5 seasonal forecast system to identify sources of predictability and seek insight into the signal‐to‐noise problem. For the first time, we use an optimal detection method to identify predictable signals over the North Atlantic region within the forecast system on subseasonal time‐scales. We find two primary predictable modes: a Pacific North America (PNA)‐like mode and a North Atlantic oscillation (NAO)‐like mode. The latter is the leading predictable mode in December–January, and its spatial pattern closely resembles the NAO. The PNA‐like mode dominates in January–February. Whereas the PNA‐like mode is driven by Pacific Ocean sea‐surface temperatures, the NAO‐like mode is driven at least partly by Indian Ocean sea‐surface temperatures, not solely due to the common trend. We develop a novel method of comparing the magnitude of these modes in the forecast system and observations that complements previous approaches. This suggests that the signal‐to‐noise problem in GloSea5 is primarily a feature of the December–January NAO‐like mode, with the observed mode being three times larger than in the model. The magnitude of the PNA‐like mode is better captured by the forecasts, although there is still evidence of a weaker signal‐to‐noise problem. This suggests particular mechanisms may lead to the lower signal to noise seen in NAO hindcasts, rather than a global weakness of the forecast system in responding to initialization and external forcing. Our results, though specific to GloSea5, provide insights into the causes of the signal‐to‐noise problem in seasonal forecasts of European winters. They also imply there is significant potential for improving such forecasts and suggest how such improvements may be achieved.