Equitable Skill Scores for Categorical Forecasts

Abstract

Many skill scores used to evaluate categorical forecasts of discrete variables are inequitable, in the sense that constant forecasts of some events lead to better scores than constant forecasts of other events. Inequitable skill scores may encourage forecasters to favor some events at the expense of other events, thereby producing forecasts that exhibit systematic biases or other undesirable characteristics. This Paper describes a method of formulating equitable skill scores for categorical forecasts of nominal and ordinal variables. Equitable skill scores are based on scoring matrices, which assign scores to the various combinations of forecast and observed events. The basic tenets of equitability require that (i) all constant forecasts?and random forecosts?receive the same expected score, and (ii) the elements of scoring matrices do not depend on the elements of performance matrices. Scoring matrices are assumed here to be symmetric and to possess other reasonable properties related to the nature of the underlying variable. To scale the elements of scoring matrices, the expected scores for constant and random forecasts are set equal to zero and the expected score for perfect forecasts is set equal to one. Taken together, these conditions are necessary but generally not sufficient to determine uniquely the elements of a scoring matrix. To obtain a unique scoring matrix, additional conditions must be imposed or some scores must be specified a priori. Equitable skill scores are illustrated here by considering specific situations as well as numerical examples. These skill scores possess several desirable properties: (i) The score assigned to a correct forecast of an event increases as the climatological probability of the event decreases and (ii) scoring, matrices in n+1?event and n-event situations may be made consistent, in the sense that the former approaches the latter as the climatological probability of one of the events approaches zero. Several possible extensions and applications of this method are discussed.