Evaluating Probabilities: Asymmetric Scoring Rules
提出一类非对称评分规则，用于更公平地评估概率预测的准确性，并以美国国家气象局的降水概率预报为例展示了其应用。
Proper scoring rules are over evaluation measures that reward accurate probabilities Specific rules encountered in the literature and used in practice are invariably symmetric in the sense that the expected score for a perfectly-calibrated probability assessor (or model generating probabilities) is minimized at a probability of one-half. A family of asymmetric scoring rules that provide better measures of the degree of skill inherent in the probabilities and render scores that are more comparable in different situations is developed here. One member of this family, a quadratic asymmetric rule, is applied to evaluate an extensive set of precipitation probability forecasts from the U.S. National Weather Service. Connections to previous characterizations of proper scoring rules are investigated, and some relevant issues pertaining to the design of specific asymmetric rules for particular inferential and decision-making problems are discussed briefly.