Nonlinear/Non-Gaussian Time Series Prediction Based on RBF-HMM-GMM Model

Abstract

In order to cope with the nonlinear and non-Gaussian time series, a RBF-HMM-GMM model, which is based on radial basis function (RBF) neural networks with the assumption of measurement noise being hidden Markov model (HMM) and the distribution of each hidden states being approximated by Gaussian mixture models (GMM), is proposed in this paper. In the proposed model, both the orders (numbers of nodes and inputs of RBF network, numbers of hidden states of HMM, numbers of Gaussian mixture component of GMM) and the parameters change over time. Firstly, a scheme for time series forecasts based on RBF-HMM-GMM model is proposed. Then an on-line prediction algorithm based on RBF-HMM-GMM model using sequential Monte Carlo (SMC) methods is developed. At last, the monthly West Texas Intermediate crude oil future price series are analyzed, and experimental results indicate that the RBF-HMM-GMM model is able to predict the time series accurately. DOI: http://dx.doi.org/10.11591/telkomnika.v10i6.1545 Full Text: PDF