A new SVM-based ensemble approach for time series forecasting

Time series analysis has remained as an extremely active research area for decades, receiving a great deal of attention from very different domains like econometrics, statistics, engineering, mathematics, medicine and social sciences. To say nothing about its importance in real-world applications in a wide variety of industrial and business scenarios. However, as hardware becomes ubiquitous, the amounts of data collected is more and more overwhelming, bringing us all the so-called big data era. It is in this context where automatic time series analysis deserves especial attention as a mean of making sense of such enormous databases. Nevertheless, the automatic identification of the appropriate data modelling techniques stands in the middle as a compulsory stage of any big data implementation. Research on model selection and combination points out the benefits of such techniques in terms of forecast accuracy and reliability. This study proposes a novel ensemble approach for automatic time series forecasting as a part of a big data implementation. Given a set of alternative models, a Support Vector Machine (SVM) is trained at each forecasting origin to select the best model, according to the computed features and the past performance. The feature space embeds information of the time series itself as well as responses and parameters of the alternative models. This approach will help to reduce the risk of misusing modelling techniques when dealing with big datasets, and at the same time will provide a mechanism to assert the appropriateness of the underlying models used to analyse such data. The effects of the proposed approach are explored empirically using a set of representative forecasting methods and a dataset of 229 weekly demand series from a leading household and personal care UK manufacturer. Findings suggest that the proposed approach results in more robust predictions with lower mean forecasting error and biases than base forecasts.