Clustering of Time Series Based on Forecasting Performance of Global Models

This article proposes a new procedure to perform clustering of time series. The approach relies on the classical K-means clustering method and is based on two iterative steps: (i) K global forecasting models are fitted via pooling by using the series belonging to each group and (ii) each series is assigned to the cluster associated with the model yielding the best forecasts in accordance with a specific criterion. The resulting clustering solution includes groups which are optimal in terms of overall prediction error, and thus the procedure is able to detect the different forecasting patterns existing in a given dataset. Some simulation experiments show that our method outperforms several alternative techniques in terms of both clustering accuracy and forecasting error. The procedure is also applied to carry out clustering in three real time series databases.