Time series forecasting with a hybrid clustering scheme and pattern recognition

This paper presents the development of a novel clustering algorithm and its application in time series forecasting. The common use of clustering algorithms in time series is to discover to groups sets of data with common characteristic their proximity. This property is used by several hybrid forecasting algorithms that additionally employ a function approximation technique to model interactions within each cluster. The proposed hybrid clustering algorithm (HCA) is a data analysis oriented clustering based on an iterative procedure that creates groups of data whose common property is that they are best described by the same linear relationship. A complementary pattern recognition scheme is employed to assist its implementation in time series forecasting. In this paper the HCA methodology is tested on the benchmark sunspots series, the daily closing values of the Dow Jones Index and hourly surface ozone concentrations. It exhibited a reduction of the forecasting error, in excess of 9%, when compared to other approaches met in the literature.