Long-Term Prediction of Time Series using a Parsimonious Set of Inputs and LS-SVM

Time series prediction is an important problem in many areas of science and engineering. We investigate the use of a parsimonious set of autoregressive variables in the long-term prediction task using the direct prediction approach. We use a fast input selection algorithm on a large set of autoregressive variables for different direct predictors, and train nonlinear models (LS-SVM and a committee of LS-SVM) on the parsimonious set of non-contiguous set of autoregressive variables. Results will be shown for the time series competition task.