| description abstract | It is being increasingly recognized that the uncertainty in weather forecasts should be quantified and furnished to users along with the single-value forecasts usually provided. Probabilistic forecasts of ?events? have been made in special cases; for instance, probabilistic forecasts of the event defined as 0.01 in. or more of precipitation at a point over a specified time period [i.e., the probability of precipitation (PoP)] have been disseminated to the public by the Weather Bureau/National Weather Service since 1966. Within the past decade, ensembles of operational numerical weather prediction models have been produced and used to some degree to provide probabilistic estimates of events easily dealt with, such as the occurrence of specific amounts of precipitation. In most such applications, the number of ensembles restricts this ?enumeration? method, and the ensembles are characteristically underdispersive. However, fewer attempts have been made to provide a probability density function (PDF) or cumulative distribution function (CDF) for a continuous variable. The Meteorological Development Laboratory (MDL) has used the error estimation capabilities of the linear regression framework and kernel density fitting applied to individual and aggregate ensemble members of the Global Ensemble Forecast System of the National Centers for Environmental Prediction to develop PDFs and CDFs. This paper describes the method and results for temperature, dewpoint, daytime maximum temperature, and nighttime minimum temperature. The method produces reliable forecasts with accuracy exceeding the raw ensembles. Points on the CDF for 1650 stations have been mapped to the National Digital Forecast Database 5-km grid and an example is provided. | |
