Clustering preprocessing to improve time series forecasting

Abstract

This work proposes a novel general-purpose forecasting algorithm. It first extracts patterns from time series using the information provided by certain clustering techniques, which are applied as a first step of the approach. Moreover, the occurrence of data with especially unexpected values (outliers) is also addressed in this work. To deal with these outliers, a new hybrid methodology has been proposed, by inserting and adapting an existing approach based on the discovery of frequent episodes in sequences in the general scheme of prediction.