Improving Seasonal Prediction Skill for Northern Winter through the Use of Simulated and Observed Arctic Oscillation

Based on the hindcast experiments of recent 24 years using CESM2 model, we find a relatively lower predictive skills over the North American and Eurasian regions. In this study, we devised a post-processing method for the seasonal forecast model utilizing the observed and predicted relationship between the Arctic Oscillation and northern surface air temperature which is helpful for those landmass regions of less predictive skills.

Although the CESM2 predicts the AO index reasonably well, linear regression analysis applied to the hindcast data revealed that the predicted surface air temperature over Eurasian continent in the model is much less controlled by the Arctic Oscillation. Motivated by this finding, we replaced and adjusted the model's underestimated surface temperature variability associated with the Arctic Oscillation with that in the reanalysis data thereby improving the predictive skill of mid-latitude surface temperature in the seasonal prediction model. This approach lead to significant improvements in surface temperature prediction when the model exhibits high predictive skill for the Arctic Oscillation index. However, when evaluating the predictive skill over consecutive periods of more than ten years in the hindcast experiment data, the skill varied greatly depending on the evaluation period, ranging from an anomaly correlation coefficient (ACC) of 0.1 to 0.7. Here, we demonstrate that, when AO predictability itself is low, utilizing the statistically meaningful predictor for the winter AO such as Eurasian snowcover-related variable sometimes can be helpful.