Abstract
This paper explores how localized mixture models can be used for prediction using time series data. The estimation method presented in this study is a kernel-weighted version of an EM-algorithm, where exponential kernels with different bandwidths are used as weight functions. Nadaraya-Watson and local linear estimators are used to carry out localized estimations. Furthermore, in order to demonstrate suitability for prediction at a future time point, a methodology for bandwidth selection and adequate methods are outlined for each model, and then compared with competing forecasting routines. A simulation study is executed to assess the performance of these models for prediction. Furthermore, real data is used to investigate the performance of the localized mixture models for prediction. The data used is predominately taken from the International Energy Agency (IEA).