A Self-Organizing Mixture Extreme Leaning Machine for Time Series Forecasting

Abstract

A novel self-organizing Mixture Extreme Learning Machine (SOM-LEM) model and algorithm for time series forecasting is proposed in this paper. As the stock time series is non-stationary stochastic processes which switch their dynamics from time to time or have different models in different periods, and the ELM algorithm also has some drawbacks such as when the numbers of the samples and hidden nodes are very large, the calculation of the Moore-Penrose generalized inverse of matrixH will become very complicated, and the corresponding error of the elements in the matrix will become larger, and thus the generalization performance of the network will be reduced. These imply that it is not convincing and impractical for a single parametric model to capture the dynamics of the entire time series. So a SOM competitive layer is added in front of the ELM network to form the SOM-ELM model, in which, each category samples divided by SOM is then handled by a ELM model. The better generalization performance of the SOM-ELM algorithm are verified through some experiments with practical stock time series.