Representation of tropical SST trends in ECMWF seasonal hindcasts and implications for recent ENSO forecasts

Operational seasonal forecasts are routinely issued with their bias removed, which is estimated from hindcasts covering a sufficiently long period. An increased number of false alarms for the occurrence of El Nino by various dynamical forecasting systems in recent years challenges the view that forecast biases are stationary. Here we assess the ability of ECMWF’s operational seasonal prediction system SEAS5 to represent observed trends in tropical SSTs since 1993, with a focus on the Pacific.

SEAS5 hindcasts overestimate SST warming in the equatorial Pacific when compared to observations. Although present for all start dates, the trend error is most pronounced for May starts. As a result, SEAS5 forecasts in recent years tended to predict too warm ENSO states despite bias correction. The hindcasts also fail to reproduce the observed meridional dipole in SST trends in the eastern Pacific, with warming in the northern and cooling in the southern subtropics. We assess several numerical experiments to investigate the role of the evolving ocean observing system, the ocean data assimilation system, and the atmospheric model. Results show that the increase in Argo observations amplifies the spurious trends in the hindcasts, which points to biases in the ocean initial conditions when observational constraints are lacking prior to Argo. Furthermore, observed-SST experiments show that the atmospheric model is unable to reproduce the magnitude of increasingly northward winds that are observed in the eastern equatorial Pacific, which are associated with the meridional structure of observed SST trends and have been speculated to reduce ENSO variability. This suggests that shortcomings of the atmospheric model physics further contribute to the system’s inability to predict the recent triple La Nina period. The results call for more sophisticated calibration methods of seasonal forecasts and ultimately improved models and initialization to provide more reliable ENSO forecasts under varying background conditions.