Large potential of performance-based model weighting to improve decadal climate forecast skill

Decadal climate predictions are sensitive to model initialization and simulation of climate forced response and internal variability. While analogue-based initialization selects initial states matching observations from large climate model ensemble simulations, it neglects differences in model performance. Focusing on sea-surface temperature decadal predictions, we couple analogue-based initialization with performance-based model weighting. Specifically, we favor selection of analogues from models that are statistically more consistent with observations in climate forced response and spatiotemporal variability characteristics. Through this statistical procedure, we demonstrate the effectiveness of a deviance metric that simultaneously evaluates multiple aspects of model-observation consistency and is novel to model weighting practices. We first conduct performance-weighted predictions of pseudo-observations, targeting model realizations instead of observations. Applying this exercise to more than 300 pseudo-observations to ensure robustness, we demonstrate large decadal forecast potential skill improvement compared to unweighted predictions. Second, we apply the same prediction method in decadal hindcasts of 95-year real-world sea-surface temperature observations. We find significant skill gains from performance-based weighting, however at considerably lower levels than in the pseudo-observation configuration. We explain this apparent contradiction by limited intrinsic predictability, similarity between unweighted and weighted ensembles, and inherent skill sampling uncertainties; we diagnose evidence for these three limitations in our results. Our analysis therefore highlights previously unrecognized challenges in validating performance-based model weighting, with implications for model weighting practices for climate predictions and projections across time scales.