Multicriterion Segmentation of Demand Markets to Increase Forecasting Accuracy of Analogous Time Series: A First Investigation

Abstract

Forecasting techniques for analogous time series require the identification of related time series as a fundamental first step of the analysis. The selection of meaningful groupings at this stage is known to be a major factor for final forecasting performance. In this context, a segmentation is typically obtained using clustering of the actual time series data, clustering of suspected causal factors (associated with each time series), or, alternatively, by taking into account expert opinion. Here, we consider the potential of multicriterion segmentation techniques to allow for the simultaneous consideration of multiple types of information sources at this segmentation stage. We use experiments on synthetic data to illustrate the potential this has in feeding forward into the analytics pipeline and, thus, improving the final robustness and accuracy of forecasting results. We discuss the potential of this approach in a real-world context and highlight the role that evolutionary computation can play in contributing to further developments in this area.