A1282
Title: K-INGARCH clustering for discrete-valued time series
Authors: Luis Sousa - University of Aveiro (Portugal) [presenting]
Magda Monteiro - University of Aveiro (Portugal)
Isabel Pereira - University of Aveiro (Portugal)
Dimitris Karlis - RC Athens University of Economics and Business (Greece)
Abstract: Clustering time series is a relevant problem in many areas, as it allows grouping individuals that share similar patterns. Within the context of count time series, the K-INGARCH algorithm is proposed, which is computationally efficient and ensures good allocation accuracy. The K-INGARCH method assumes a K-means approach within a model-based framework, where each center is a set of INGARCH(p,q) parameters, and the proximity of each time series to its center is given by the likelihood considering those parameters. Series are assigned to their closest model based on their conditional likelihood values, and this is repeated until there are no changes in the cluster memberships. The Poisson distribution is commonly used for the conditional modelling of INGARCH processes. This framework is generalized by allowing alternative discrete distributions within this model. Multiple distributions and cluster configurations are considered, with the final model selected using appropriate criteria. After selecting the best model, several methods to select a representative series of each final cluster are presented. The algorithm is assessed through a simulation study consisting of three scenarios with increasing levels of difficulty, determined by the proximity of the cluster parameters. Furthermore, the methodology is illustrated through an application to a real-world dataset.
