Automatic change-point detection in time series via deep learning

• DOI: 10.1093/jrsssb/qkae004
• Publication Date: 2024-01-10T21:20:01Z
• Authors: Jie Li; Paul Fearnhead; Piotr Fryzlewicz; Tengyao Wang
• Keywords: CUSUM; Gaussian Noise

Abstract Detecting change points in data is challenging because of the range possible types and behaviour when there no change. Statistically efficient methods for detecting a will depend on both these features, it can be difficult practitioner to develop an appropriate detection method their application interest. We show how automatically generate new offline based training neural network. Our approach motivated by many existing tests presence point being representable simple network, thus network trained with sufficient should have performance at least as good methods. present theory that quantifies error rate such approach, depends amount data. Empirical results that, even limited data, its competitive standard cumulative sum (CUSUM) classifier mean noise independent Gaussian, substantially outperform auto-correlated or heavy-tailed noise. also shows strong localizing changes activity accelerometer