From Sunday evening of this week, the weather forecasts predicted 'a widespread risk of snow and ice' in the South of England. The snow duly arrived on Tuesday night, but only in Sussex and Kent, where it normally seems to fall. At no stage did I hear or see a forecast which was willing to specify the counties in which the snowfall could be expected. This is exactly the type of information that commuters desire, and its absence is puzzling because the Met Office claims that its 3-day weather forecasts are now as accurate as their 1-day forecasts were 20 years ago. Perhaps the overall accuracy the Met Office claims for its forecasts is skewed by the accuracy of the forecasts issued under inherently predictable conditions, such as a 'blocking' high (pressure zone) during the summer or winter.
This week's New Scientist contains an article on the verification of weather forecast accuracy (27 January 2007, p32-35), which claims that forecasters at the Met Office receive an annual bonus determined by the verification scores for their forecasts. The article points out that forecasters can improve their scores by hedging, and by blurring the specifics of their forecasts. This perplexes me on two counts: firstly, that forecasters should be incentivized to make their forecasts less specific; and secondly, the very notion that the accuracy of a forecast is still dependent upon the forecaster producing it, seems to contradict the impression which the Met Office tries to convey about its prediction methodology. Atmospheric and oceanic data from weather stations and satellites are fed into meteorological models as initial conditions, and the Met Office supercomputers then predict the evolution of the meteorological parameters. In principle, making a forecast should then involve nothing more than reading off the predicted temperatures, pressures, cloudcover, and rainfall from these simulations, and translating them into a digestible package for the public. However, the very fact that forecasters receive bonuses based upon the accuracy of their forecasts suggests that there is actually still a large discretionary element to the production of a forecast. Perhaps one needs to have a good understanding of the idealisations and assumptions which have been incorporated into the atmospheric models if one is to judge the reliability of their predictions; there are a variety of atmospheric models, and perhaps the skill comes from understanding the circumstances in which each model is reliable or unreliable.