Change-point problem: Direct estimation using a geometry inspired identifiable reparameterization

TL;DR

This paper introduces a geometry-inspired reparameterization for change-point detection, providing a new estimation method that outperforms the traditional MLE, demonstrated on Bitcoin data.

Contribution

A novel reparameterization approach using a modified horn torus and Riemannian metric for improved change-point estimation.

Findings

The new estimator outperforms MLE in simulations.

Application to Bitcoin data shows better detection accuracy.

MLE tends to produce false alarms.

Abstract

Estimation of mean shift in a temporally ordered sequence of random variables with a possible existence of change-point is an important problem in many disciplines. In the available literature of more than fifty years the estimation methods of the mean shift is usually dealt as a two-step problem. A test for the existence of a change-point is followed by an estimation process of the mean shift, which is known as testimator. The problem suffers from over parametrization. When viewed as an estimation problem, we establish that the maximum likelihood estimator (MLE) always gives a false alarm indicting an existence of a change-point in the given sequence even though there is no change-point at all. After modelling the parameter space as a modified horn torus. We introduce a new method of estimation of the parameters. The newly introduced estimation method of the mean shift is assessed with a proper Riemannian metric on that conic manifold. It is seen that its performance is superior compared to that of the MLE. The proposed method is implemented on Bitcoin data and compared its performance with the performance of the MLE.