A0255
Title: Dynamic clustering of multivariate time series using DTGARCH model and spectral clustering
Authors: Sipan Aslan - King Abdullah University of Science and Technology (Saudi Arabia) [presenting]
Ceylan Yozgatligil - Middle East Technical University (Turkey)
Cem Iyigun - middle east technical university (Turkey)
Abstract: A novel dynamic clustering procedure is introduced for multivariate time series derived from complex systems such as brain circuitry or financial markets. The objective is to devise a clustering in which latent underlying data-generating mechanisms (DGMs) form the cluster centres. In other words, multivariate time series are treated as objects to be grouped according to their similarities of underlying DGMs. The proposed approach mainly leverages the distinguishable features extracted from a nonlinear time series model, such as the double threshold generalized auto-regressive conditional heteroskedasticity (DTGARCH) model, and groups them by spectral clustering methodology. Approximating the time series DGMs using a rich model, such as DTGARCH, is proposed to enable the comparisons of complex, nonlinear, time-dependent features of underlying data-generating processes and effectively track dynamic cluster changes. The efficiency of the approach is validated through synthetic and real-world datasets. Clustering accuracies compared with several distance measures designed for multivariate time series clustering. The presented framework offers significant implications for fields ranging from economics to neuroscience by providing a more nuanced understanding and analysis of time series data.
