Robust inference for change points using confidence sets

Change point problems can be loosely defined as situations where one observes a noisy signal which undergoes “sharp changes” at unknown points in time. Typically the interest lies in accurately estimating the so called change point locations. In this talk I will discuss the problem of making valid inference statements about change point locations when (nearly) nothing is known about the structure of the contaminating noise.

A new method will be be presented which works by performing a large number of local tests for the homogeneity of the underlying signal. The local tests are based on the structure of certain robust confidence sets for the signal. The procedure always has finite sample coverage. If time permits I will outline the proof for consistent recovery of the change point locations.