Abstract
This paper addresses a fundamental but largely unexplored challenge insequential changepoint analysis: conducting inference following a detectedchange. We study the problem of localizing the changepoint using only the dataobserved up to a data-dependent stopping time at which a sequential detectionalgorithm $\mathcal A$ declares a change. We first construct confidence setsfor the unknown changepoint when pre- and post-change distributions are assumedto be known. We then extend our framework to composite pre- and post-changescenarios. We impose no conditions on the observation space or on $\mathcal A$-- we only need to be able to run $\mathcal A$ on simulated data sequences. Insummary, this work offers both theoretically sound and practically effectivetools for sequential changepoint localization.