Does insufficient oceanic resolution contribute to the signal to noise problem in seasonal forecasts?

A major limitation of present seasonal prediction systems is the well-known signal to noise problem. Ensemble climate model simulations that are initialised with real world data show a remarkable degree of prediction skill for certain variables. For example, the UK Met Office GloSea5, initialised with observations in November can predict the subsequent winter North Atlantic Oscillation index with an average skill in excess of 0.6 based on the correlation of the ensemble mean simulated winter NAOI with the corresponding observed NAOI, verified from comparing more than two decades of hindcasts with observations.

The problem arises because although the correlation of the ensemble mean prediction with observations is high, the absolute magnitude of the predicted signal is low, and the ensemble mean is poorly correlated with individual ensemble members, leading to the apparent paradox that the model is better able to predict the real world than its own ensemble members. Two deleterious consequences of the signal to noise problem are that large ensembles are required to give robust skill, making seasonal forecasts expensive, and that the underprediction of the signal lessens the societal value of the forecasts.

Despite much research, the origin of the signal to noise problem remains mysterious. Here we test the hypothesis that the signal to noise problem arises at least partly because current forecast systems do not adequately represent air-sea interaction due to insufficient oceanic resolution. We run model hindcast sets using the HadGEM3 GC3.1 climate model identical in all respects except in ocean model resolution (1/4 vs 1/12 degree), evaluate differences in how well the two configurations are able to predict their own ensemble members, and attribute these to corresponding changes in air-sea interaction, including factors such as a better resolved mesoscale eddy field and more realistic boundary currents in the higher resolution configuration.