Large-scale fluctuations of the equatorial atmosphere and ocean over the tropical Pacific, known as El Niño–Southern Oscillation (ENSO), play an important role in the climate system. Forecasting ENSO is at the very heart of seasonal predictions because it provides the largest source of predictability on timescales of months and seasons ahead, and it is of great relevance to society. A number of global teleconnections of ENSO link the variability over the tropical Pacific with remote regions of the world and contribute substantially to seasonal forecast skill in the extratropics of both hemispheres. Retrospective seasonal forecasts at ECMWF and other centres around the world cover the well-observed recent period from the 1980s onward and are essential to estimate and calibrate the skill of operational forecasts. Here we report about a new historical retrospective research forecasting dataset, created with a version of ECMWF’s operational seasonal prediction system SEAS5, that covers all of the 20th century with forecast lead times of two years. The dataset, called SEAS5-20C, will be used to advance our understanding of the predictability of ENSO in the past and the future.
ENSO is an irregular coupled mode of variability that displays a diverse range of spatial structures, amplitudes and life cycles. The location of warm El Niño sea-surface temperature (SST) anomalies can be widely spread, from the Eastern to the central equatorial Pacific. Substantial non-monotonic variations in the amplitude of ENSO SST signals that are poorly understood have occurred during the 20th century. Nor do we understand very well the factors determining the occurrence and predictability of consecutive El Niño or La Niña years. With an average frequency of four to five years, there is only a very limited number of ENSO cases available in operational re-forecast records, and sampling the wide spectrum of ENSO flavours is not possible. In the presence of considerable variations in the coupled ocean–atmosphere system, good skill in predicting the most recent ENSO events cannot guarantee that future events will have similar predictability.
In order to further improve our confidence to predict future ENSO events and to explore potential ENSO skill beyond the seasonal timescale, we have performed a novel set of re-forecasts with a lead time of 24 months covering the extended 20th century period from 1901 to 2010. The simulations were done with a reduced-resolution version of ECMWF’s operational seasonal system SEAS5 and comprise ensembles of 10 members for forecast start dates every May and November during the 110‑year hindcast period. The initial states of the ocean, sea‑ice, waves, atmosphere and land of the reforecasts were taken from ECMWF’s first coupled reanalysis of the 20th century, CERA-20C, which assimilated temperature and salinity in the ocean and sea-level pressure and marine wind data in the atmosphere.
As an example, the figure shows retrospective forecasts of the strong El Niño event that occurred in the December–January–February (DJF) season of 1902/03. CERA-20C (panel (a)) displays a large area of warm temperature anomalies with a maximum in the central equatorial Pacific. SEAS5-20C re-forecasts of this event show the DJF SST ensemble mean anomalies from the forecasts initialised 3 months (b), 9 months (c), 15 months (d) and 21 months ahead (e). We see that on seasonal timescales the El Niño event was well predicted. Even at the longest forecast range of 21 months (e), the forecast indicated a wide-spread yet weak warming across the central equatorial Pacific.
Panel (f) compiles the individual ensemble member realisations for the NINO3.4 SST index (area-averaged SSTs over the box 5°S–5°N and 120–170°W) anomaly from the four initialisation times of panels (b)–(e) that target the DJF 1902/03 season, with the black line showing the evolution of NINO3.4 SST anomalies in the verifying CERA-20C reanalysis. The forecast from November 1902 (red) follows the CERA-20C evolution of the peak warming and the subsequent cooling extremely well, with all ensemble members indicating positive SST anomalies during DJF. The May 1902 forecast (orange) has a 90% probability for DJF SST anomalies >0, while the forecasts from November 1901 (yellow) and May 1901 (blue) give an overall indication for likely positive SST anomalies during the peak season DJF (forecast probabilities of 60%).
The re-forecasts show substantial decadal modulations of ENSO properties, and variable skill, which is the subject of ongoing investigations. Several sensitivity experiments that test the impact of sub-surface ocean observations and the role of atmospheric forcings at initial time have also been performed. SEAS5-20C is an important resource to study the complex predictability behaviour of ENSO, including multi-decadal skill variations and multi-year ENSO events. It will further be used to explore the feasibility of extending skilful ENSO predictions beyond one year. The DOI for the data discussed in this article is doi.org/10.21957/fzf9-te33.