A hybrid ensemble approximation method for chaotic time series forecast

To reduce the prediction error, one method is to integrate the machine learning approximation system with multi-model ensemble of predictors. This article proposes a method of constructing a robust ensemble of predictors to forecast the chaotic time-series data. To construct the multi-ensemble hybrid system, the proposed method involves using the Bagging method to randomize the training data, training several prediction models based on SVM and ANN with different initial parameters, using improved fuzzy RBF neural networks to build error models, as well as using aggregation methods including weight/bias voting system within the ensemble and a D-S reasoning system between ensembles. By an empirical experiment, the results shows for a chaotic discrete time series prediction the proposed method can be very accurate in comparison with single-model SVM or ANN method as well as with many existing ensemble method. Copyright © 2012 Binary Information Press.