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. Analogue-based initialization selects initial states matching observations from large climate model ensemble simulations, but neglects model performance. We implement performance-based model weighting, favoring models consistent with observations in climate forced response and stationary dynamics. Focusing on sea-surface temperature decadal predictions, we demonstrate the effectiveness of a deviance statistic, not previously used in model weighting schemes. We conduct performance-weighted predictions of pseudo-observations, targeting model realizations instead of observations, which show large decadal forecast potential skill improvement compared to unweighted predictions. We also find skill gains in decadal hindcasts of 95-year real-world sea-surface temperature observations, however at considerably lower levels. We explain this apparent contradiction by limited intrinsic predictability, similarity between unweighted and weighted ensembles, and inherent skill sampling uncertainties. Our analysis therefore highlights previously unrecognized challenges in validating performance-based model weighting for climate forecasting.