The problems of incident detection and short term forecasting were originally closely connected. A development phase has followed where the problems have been largely considered in isolation. These stages of development are critically reviewed. It seems that short term forecasting as a method of incident detection has largely fallen out of fashion, mainly because other algorithms of superior performance have been developed. This is unfortunate because algorithms which define an incident as any event which cannot be forecast have several useful properties. For example, it is not necessary to carry out an exensive data collection procedure in order to buid up a data base of real incidents for training. A self-bootstrapping algorithm based on principal component analysis is presented wihich is not a forecasting based method, but which shares the same definition of an incident. Initial results are very promising.
展开▼
