A1803
Title: Multiple change point detection in time series with non-stationary dynamics
Authors: Yuhan Tian - Technische Universitat Munchen (Germany) [presenting]
Abolfazl Safikhani - George Mason University (United States)
Abstract: Change point detection (CPD) is fundamental for analyzing time series data, with applications in finance, climate science, and quality control. However, most existing methods rely on piecewise stationarity, an assumption often violated in practice due to continuous non-stationary dynamics. Such dynamics can mask abrupt changes or lead to false detections. A model-based CPD framework is proposed that explicitly accounts for non-stationarity and cross-correlations. The method decomposes the time series into a random walk component capturing non-stationary behavior and a vector autoregressive component modeling dependencies. Change points are identified by evaluating error reduction within moving windows when structural changes are introduced. The proposed approach is flexible and supports a broad class of component selection procedures, while admitting theoretical guarantees on consistent estimation of both the number and locations of change points. Although derived under a model-based framework, the method demonstrates strong empirical robustness. Simulation studies show that it performs well across a wide range of non-stationary settings, including seasonality and long-memory dependence. Experiments on simulated and real data confirm the effectiveness of the method. In particular, it successfully identifies surface defects in steel rolling images despite gradual background variations, highlighting its robustness in complex non-stationary environments.
