A1316
Title: Change point detection in time series using mixed integer programming
Authors: Alexander Semenov - University of Florida (United States) [presenting]
Anton Skrobotov - Russian Presidential Academy of National Economy and Public Administration and SPBU (Russia)
Peter Radchenko - University of Sydney (Australia)
Artem Prokhorov - University of Sydney (Australia)
Abstract: Recent advances in mixed-integer optimization (MIO) methods are used to develop a framework for identifying and estimating structural breaks in time series regressions. The framework requires transforming the classical structural break detection problem into a Mixed Integer Quadratic Programming problem. The problem is restated as a $l_0$ penalized regression and is compared to the infamous $l_1$ penalized regression (LASSO). MIO can find provably optimal solutions to the problem using a well-known optimization solver. The framework determines the unknown number of structural breaks and the break locations. Additionally, the accommodation of a specific number of breaks, or their minimal required number, is demonstrated. The approach's effectiveness is further presented through extensive numerical experiments, obtaining a more accurate estimation of the number of breaks compared to the popular methods.
