A Probabilistic Derivation of Heidke Skill Score

AbstractAn alternative derivation of Heidke skill score for 2 × 2 tables is presented, starting from the assumption that a categorical forecast is useful, if the probability of an occurrence of an event, given the forecast, is greater than the base rate of the event. A tentative measure of skill would then be the difference of these probabilities, normalized by the maximum value based on the base rate. For binary events, the Heidke skill score is then the harmonic mean of these differences for both the occurrence and the nonoccurrence of the event. This derivation differs from the usual derivation in that the concept of chance agreement is not used. It is Bayesian in nature with implied updating of prior probabilities to posterior probabilities.