A least square support vector machine (LS-SVM) based on the independent component analysis(ICA) is proposed
to predict noisy non-stationary time series. ICA is used to estimate the independent components(IC) in the forecasting
variables. After identifying and removing the ICs containing the noise, the rest of the ICs are then used to reconstruct the
forecasting variables which contain less noise. A k-nearest neighbors(k-NN) approach is used to reduce the size of training
dataset and a new distance function is defined. By selecting similar instances in the training dataset, the complexity of
training a LS-SVM is reduced significantly. A boundary constraint component is developed to limit the predicted values to
a reasonable range. The experimental results show that the proposed approach outperforms both traditional LS-SVM and
BP-artificial neural network(BP-ANN) in the prediction performance of several time series.