Information‐based probabilistic verification scores and predictability measures: Seasonal prediction examples

Abstract Assessing prediction quality through a process known as verification that compares past predictions and corresponding observations is fundamental for advancing prediction models and systems to enhance their utility. For the verification of ensemble predictions, a recently proposed approach to computing information‐based probabilistic verification scores, namely the logarithmic score and information gain, using a fixed‐distance near‐neighbour search, has the potential to improve the quality assessment of these predictions. This study demonstrates the applicability of these scores to seasonal prediction. Results illustrate that these classes of information‐based probabilistic verification scores are useful measures of prediction skill. In addition, this paper introduces information‐based potential predictability measures based on the perfect model assumption, namely entropy and relative entropy, using a fixed‐distance near‐neighbour search, consistent with the proposed logarithmic score and information gain. This study also investigates the relationship between prediction skill measured by the proposed scores and potential predictability or uncertainty in ensemble predictions measured by the proposed potential predictability measures. A seasonal prediction model analysed in this study exhibited a moderate correspondence between the logarithmic score and entropy, and a high correspondence between information gain and relative entropy, especially in the tropics. The proposed methods provide a new framework for quantifying the prediction skill and potential predictability of ensemble predictions, presenting potential avenues for advancing ensemble prediction.