Model selection for time series forecasting using similarity measure

Several methods have been proposed to combine the forecasting results into single forecast namely the simple averaging, weighted average on validation performance, or non-parametric combination schemas. These methods use fixed combination of individual forecast to get the final forecast result. In this paper, quite different approach is employed to select the forecasting methods, in which every point to forecast is calculated by using the best methods used by similar training dataset. Thus, the selected methods may differ at each point to forecast. The similarity measures used in this paper are Euclidean and Dynamic Time Warping (DTW). The dataset used in the experiment is the time series data designated for NN3 Competition. The experimental result shows that the combination of methods selected based on the similarity between training and testing data may perform better compared to either the best of individual predictor or the combination of all methods.