资源环境科技发展态势分析平台

Extracting predictors from a predictand-predictor correlation map is a common problem for climate prediction, but its skill is affected by sampling errors and the subjective selection of predictors; hence, it is difficult to ensure that the selected predictors are optimal. Additionally, cross validation tends to overestimate the actual prediction skill because of artificial skill. In view of these problems, the author proposes an empirically optimal screening (EOS) method to extract predictors from a correlation map. Based on hindcast cross validation, EOS empirically and objectively identifies an optimal correlation threshold for data screening. To mitigate artificial skill, cross validation completely separates the training and testing samples, not only for parameter fitting but also prior predictor selection. By using EOS, researchers avoid subjectively determining predictors directly from correlation maps, and EOS further refines potential predictors before the verification of physical mechanisms.

Plain Language Summary Statistical methods are commonly used in climate prediction. Researchers often calculate a correlation map between the target and a predictor field then subjectively select some highly correlated areas as predictors and verify their physical links to the target using numerical experiments. However, such a procedure requires researchers with deep physical insight and it cannot ensure that the selected predictors are optimal. Additionally, the prediction skill tends to be overestimated because of the issue of artificial skills. In view of these problems, the author proposes an empirically optimal screening (EOS) method to objectively extract predictors from a correlation map. EOS is a further process beyond the correlation map, and it provides a statistically optimal model for a forecaster's reference before the verification of physical mechanisms, thereby avoiding the subjectivity involved in selecting predictors directly from a correlation map.