Parametric change point detection with random occurrence of the change point

TL;DR

This paper investigates the detection of a single, randomly located change point in exponential family time series data, allowing the change point to depend on the data and its size to diminish as the sample grows.

Contribution

It extends existing change point detection theory to cases where the change point is random and possibly data-dependent, with the change size shrinking as data increases.

Findings

Statistical results from classical change point literature hold in the random change point setting.

Theoretical analysis confirms the consistency of detection methods under the new assumptions.

Simulation studies support the theoretical findings.

Abstract

We are concerned with the problem of detecting a single change point in the model parameters of time series data generated from an exponential family. In contrast to the existing literature, we allow that the true location of the change point is itself random, possibly depending on the data. Under the alternative, we study the case when the size of the change point converges to zero while the sample size goes to infinity. Moreover, we concentrate on change points in the "middle of the data", i.e., we assume that the change point fraction (the location of the change point relative to the sample size) converges weakly to a random variable $\lambda^*$ which takes its values almost surely in a closed subset of $(0,1).$ We show that the known statistical results from the literature also transfer to this setting. We substantiate our theoretical results with a simulation study.