A1375
Title: A Bayesian approach for change point detection in multivariate time series
Authors: Chi-Hsiang Chu - National University of Kaohsiung (Taiwan) [presenting]
Sangyeol Lee - Seoul National University (Korea, South)
Shih-Feng Huang - National Central University (Taiwan)
Abstract: A Bayesian framework is proposed for detecting structural change points in multivariate time series through a structural break vector autoregressive (SBVAR) model. The proposed approach extends the univariate structural break autoregressive (SBAR) model to a higher-dimensional setting, allowing for dynamic interactions among multiple time series while accommodating time-varying parameter regimes. To identify structural breaks and estimate segment-wise coefficients, a three-stage Bayesian procedure is developed. In the first stage, change points are represented by binary indicators associated with groups of regression coefficients, and a stochastic matching pursuit (SMP) algorithm is employed to explore the model space via add-delete operations. The marginal likelihood for each group is derived in closed form by integrating out the corresponding parameters under a Gaussian prior, thereby enabling efficient computation of posterior inclusion probabilities. In the second stage, candidate change points are constructed based on posterior inclusion probabilities, followed by an information criterion-based refinement to determine the final set of structural breaks. In the final stage, segment-specific VAR coefficients are estimated accordingly. Simulation studies are conducted to demonstrate the feasibility and effectiveness of the proposed approach in identifying structural changes in multivariate time series.
