Abstract The signal‐to‐noise paradox—where climate model predictions correlate more highly with observations than with their own ensemble members—remains a challenge for interpreting predictability. We present an extended theoretical framework explaining the role of different types of model errors in the paradox. Analytical results show, despite the non‐identifiability of the latent parameters, that the paradox arises only when the shared signal is underestimated relative to observations and this underestimation is sufficiently large, rather than from differences in signal‐to‐noise ratios. The paradox therefore likely reflects intrinsic model errors that become apparent through correlation scores, without contradicting predictability theory. The paradox can also be obscured by missed and spurious signals. We further propose an improved estimate of a lower bound on real‐world predictability that exceeds conventional estimates under paradox conditions.

Read original article