Title: Quantile autocovariances: A powerful tool for hard and soft partitional clustering of time series
Authors: Jose Vilar - Universidade da Coruna (Spain) [presenting]
Borja Lafuente-Rego - Universidade da Coruna (Spain)
Abstract: A distance based on estimated quantile autocovariances (QAD) is proposed to perform time series clustering when the target is to group series according to the underlying dependence structures. Unlike other extracted features, quantile autocovariances account for sophisticated dynamic features and are well-defined for a broad class of processes. Hence, a cluster procedure based on comparing quantile autocovariances should report satisfactory results, particularly in complex scenarios involving non-linear or heteroskedastic processes. The behavior of QAD in partitioning-based clustering is examined considering both crisp and fuzzy procedures. Contribution consists of three points. First, a broad simulation study shows the good behavior of the QAD-based clustering compared to other commonly used dissimilarities. Excellent scores are attained by classifying heteroskedastic processes and also when non-normal innovations are considered. Second contribution concerns the optimal selection of input parameters, i.e. the problem of determining the proper combination of lags and quantile levels is addressed. Third contribution consists in introducing a novel fuzzy procedure based on QAD, which presents high capability to clustering GARCH models, outperforming fuzzy clustering algorithms specifically designed to work in this framework. The usefulness of the proposed procedure is illustrated by its application to a case-study.