ART: distribution-free and model-agnostic changepoint detection with finite-sample guarantees

Abstract We introduce ART, a distribution-free and model-agnostic framework for changepoint analysis with finite-sample guarantees. ART transforms independent observations into real-valued scores via a symmetric function; under the null hypothesis of no changepoint these scores are exchangeable. Ranking and aggregating the scores yields test statistics whose null distribution is known exactly from the permutation law of ranks, enabling exact finite-sample Type I error control without repeated refitting under permutations. ART extends naturally to a multi-scale setting: by locally ranking scores over a family of intervals and aggregating them, it supports multiple changepoint testing, localization with inference, and post-detection inference, while retaining distribution-free calibration. The approach is model-agnostic: it imposes minimal structural or distributional assumptions and accommodates diverse score constructions, including features learned by statistical or machine-learning models. Across simulations and real-data applications, ART delivers valid error control and competitive power across a range of models and distributions. These properties make ART a reliable and versatile tool for modern changepoint analysis.