Title: Multiple change-point detection in regression models via U-statistic type processes
Authors: William Pouliot - University of Birmingham (United Kingdom) [presenting]
Shixuan Wang - University of Reading (United Kingdom)
Abstract: Many statistical and econometric procedures have been developed that are suited to testing for multiple changes in parameters of regression models which may occur at unknown times. Techniques have been developed or extended, but said extensions lack power for detecting changes in the intercept of linear regression models. A stochastic process that easily accommodates testing for many change-points that occur at unknown times has also been developed. It is shown via simulation that this U-statistic based processes lack power in finite samples for detecting change-points, even though the consistency of said tests has been established. A slight modification of his process is suggested which corrects for this problem. This slightly altered process is then used to fashion statistics which can be used to construct tests to detect multiple changes in intercept or variance of linear regression models, and will do so with much higher power than the original process. It is also shown that this slightly altered process, when weighted by appropriately chosen functions, is sensitive to detection of multiple changes in intercept that occur both early and later on in the sample, while maintaining sensitivity to changes that occur in the middle of the sample.