Time series prediction by mixture of linear local models
Sang-Keon Oh, Kap-Ho Seo, Ju-Jang Lee · 2004
Local modeling approaches have emerged as one of the promising methods of time series prediction. By divide-and-conquer method, state-dependent local model can approximate a subset of training data accurately. However, the construction of local models need appropriate selection of much larger number of parameters. This paper presents a method to construct a mixture of linear prediction models for the prediction of nonlinear time series. The use of locally linear model reduces the burden on the user to specify parameters using linear optimization method. This method is applied to the modelling of the Mackey-Glass time series.