Time Series Prediction using LS-SVMs

This paper describes the use of LS-SVMs as an estima- tion technique in the context of the time series prediction competition of ESTSP 2008 (Finland). Given three different time series, a model is estimated for each series, and subsequent simulations of several points af- ter the last available sample are produced. For the first series, a NARX model is formulated after a careful selection of the relevant lags of inputs and outputs. The second and third series show cyclical or seasonal pat- terns. Series 2 is modelled by adding deterministic "calendar" variables into the nonlinear regression. Series 3 is first cleaned from the seasonal patterns, and a NAR model is estimated using LS-SVM on the deseason- alized series. In all cases, hyperparameters selection and input selection are made on a cross-validation basis.