The Combination Forecasting Method And Its Application

So far, we have lots of theories and methods about forecast already. But each kind of sole forecast model portrays the data sequence only from one side, only reflecte partial information, so, there is its limitation. If we combine some kinds of forecasts theory to carry on combination forecast. We can use the existing informations to the greatest degree. Hopefully we can get better forecasteffect.The most major character of the gray model is that it needs only a few datas to establish model. Under the condition of information incomplete and not explicit, the gray model has its unique superiority. But, the gray model also has its inherent flaw. Firstly, its forecast graph is a smoother curve; It portrays the develop tendency of the sequence, but donnot consider the stochastic undulation; It uses a few data to modelling, completely neglects the overall situation, completely neglects the information that the comparatively early datas carried. These are not reasonable. Secondly, the gray model established on the Experience Risk Minimization principle, inevitably has the overfitting problem. So, it has poor generalization capability.In view of the preceding question, we can solve it by combination forecast through gray model and Markov model. Markov model is suitable for the predicting of undulate data. The transit probability matrix can portray the undulation information of data sequence better. Using the Markov model to dig the undulation information of the whole sequence, and use these information to revise the forecast result of the gray model. Through two kinds of model's combination, to enhance the forecast precision.In view of the gray model's second flaw, this article proposes a gray model whose parameters recognised by support vector machine. Using the SVM method based on Structure Risk Minimization principle to compute the parameters of grey model, then, utilise the gray model to forecast. Throngh this proposed combination forecast method, named SVM-GM, we may avoid the overshotting problem, and get a better predic effection.Another strategy to increase the forecast precision is throngh weighted array of several sole forecast model forecast result to get the final result. This also has many kinds of different methods. Because the nerve network has the formidable non-linear fittingability, this article uses the nerve network to study to several kind of sole model's forecast results, then, establishes non-linear combination forecast model based on the network to increase the forecast precision.