A model for assessment of weather forecasts

A general paradigm for assessment of ability to discriminate between two alternatives is described in the context of weather forecast verification. The paradigm is based on the relative operating characteristic (ROC), a graph of the variation of hit rate with false alarm rate as decision criterion changes. A model for the ROC based on the mathematical theory of signal detection is shown to provide a good fit to verification data from weather forecasts for a wide variety of predictands. A basis is thus provided for the use of some indices of forecast quality derived from the mcdel. These indices are relatively independent of calibration (i.e. the correspondence between estimated probability and relative frequency) and can be evaluated for forecasts expressed in yes/no form, as ratings of risk (e.g. low, moderate, high) or explicitly as probabilities, facilitating direct comparison of these different types of forecast.