The task of finding variance change points has been the focus of considerable research in sequential data analysis. In spite of empirical success of many change point algorithms, there are several unresolved issues: (a) use various probabilistic modeling assumptions in one form and another, (b) fail when there are multiple change points, especially when a dominant change point masks other change points, (c) check each point is a change point or not, thus increase computation extensively. We present a novel offline algorithm which uses a dynamic mode decomposition based data-driven dynamical system and local adaptive window to iteratively detect variance change points. We propose a variance descriptor function which is used for guiding the focus-of-attention of change points. For detecting change points, it is used for generating regions of interest and providing coarse information, which automatically governs the location and the size of window to detect change points at different scales. The proposed algorithm is completely data driven, doesn't require a probabilistic model, and detects multiple variance change points accurately and efficiently on many time series.