Attribution of Pacific trend discrepancies using the Forced Component Estimation Statistical Method Intercomparison Project (ForceSMIP)

The pattern of Pacific sea-surface temperature (SST) change since 1980 has been highlighted as a key inconsistency between climate models and observations, with widespread impacts on the hydrological cycle, hurricane activity, sea level rise, and climate sensitivity. However, it is unknown whether this trend discrepancy results from discrepancies in the forced warming pattern simulated by climate models or discrepancies in simulated internal variability. Here, we use output of ForceSMIP, where statistical and machine learning models for distinguishing between forced response and internal variability within single realizations of the climate system were evaluated with climate model large ensembles and then applied to observations, to assess the forced and unforced contributions to Pacific SST trend discrepancies. We highlight a bias-variance tradeoff amongst the statistical and machine learning methods that show skill in forced response estimation, where methods that reduce the variance in estimated trends the most exhibit biases learned from the climate-model-based training data. Low-bias high-variance methods assess the trend discrepancy to be mostly forced, whereas low-variance high-bias methods assess the trend discrepancy to be mostly due to internal variability. The latter category of methods relies on training data from climate models with documented systematic biases, and we therefore suggest that more weight be put into the former category of methods, which would lead to the conclusion that the trend discrepancy is a discrepancy in the forced response. Our work illustrates the value of statistical attribution methods that are not reliant on climate models for interpreting trend discrepancies between climate models and observations.