Tuesday, 30 January 2024
Hall E (The Baltimore Convention Center)
ENSO forecasts from coupled GCMs frequently suffer from transient climate drifts and initialization shocks, arising from inconsistencies between the model’s preferred climate and the assimilated observations. To address this, we build on previous studies and explore a model-analog forecast (MAF) technique to complement GFDL’s traditional assimilation-initialized approach to seasonal forecasting, using the GFDL SPEAR coupled GCM. MAFs mine existing simulations for climate states that resemble the present observations, and then follow those model trajectories through time to produce an ensemble forecast. MAFs are inexpensive to produce, and avoid drifts and shocks by initializing directly on the model attractor. To create the MAF, we draw 15 states from a 900-year library of existing SPEAR historical ensemble simulations, seeking simulated tropical SST and sea surface height maps that closely resemble the observations. This is repeated for each month during 1991–2023, to produce a retrospective MAF suite. For forecast leads beyond a month or two, the SST anomaly skill for the MAF is found to match that of the state-of-the-art assimilation-initialized method in the NINO3.4 region, and actually exceeds it in the equatorial western Pacific and eastern Indian basins that drive many of ENSO’s teleconnections. Both forecast methods derive much of their skill from post-El Niño evolution, with some additional skill post-La Niña. Forecasts from neutral conditions yield the least skill for both methods — though with increasing lead, the MAF shows much less reduction in skill than the traditional method, with the MAF better able to penetrate the spring predictability barrier to provide skill at longer leads. The MAF is more skillful than the traditional method in boreal winter, but less skillful in boreal spring, suggesting that a hybrid of the two methods may be beneficial. Ranking the observations within the retrospective forecast ensembles reveals that the MAF ensemble PDFs are more consistent with observations than ensemble PDFs from the traditional forecast method — especially at short leads, where the traditional method yields overconfident forecasts. We explore ways to further improve the MAFs, by modifying the spatial and temporal weights, target variables, detrending method, ensemble size, library size, and bias corrections. We further discuss the promise of using MAFs to evaluate and learn from predictions when data assimilation is prohibitive — such as with prototype or high-resolution models.

