Use of the 'odds ratio' for diagnosing forecast skill

Abstract

This study investigates ways to quantifying the skill in forecasts of dichotomous weather events. The odds ratio, widely used in medical studies, can provide a powerful way of testing the association between categorical forecasts and observations. A skill score can be constructed from the odds ratio that is less sensitive to hedging than previously used scores. Furthermore, significance tests can easily be performed on the logarithm of the odds ratio to test whether the skill is purely due to chance sampling. Functions of the odds ratio and the Peirce skill score define a general class of skill scores that are symmetrie with respect to taking the complement of the event. The study illustrates the ideas using Finley's classic set of tornado forecasts.