Distribution-free changepoint localization after sequential change detection

2026-07-08 · Source: stat.ML updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

This paper introduces a novel distribution-free framework for constructing post-detection confidence sets for changepoints after a sequential change detection procedure stops. Unlike prior methods that require known pre- and post-change distribution classes, this framework localizes changepoints without any distributional assumptions. It establishes finite-sample coverage guarantees, conditional on correct detection, and provides non-asymptotic bounds on the conditional expected size of the confidence sets. The framework leverages conformal test martingales, extending their use from change detection to localization, and demonstrates strong empirical performance on simulated and real data. This is presented as the first general distribution-free framework for sequential changepoint localization with valid post-detection coverage.

Key takeaway

For research scientists developing robust monitoring systems, this distribution-free changepoint localization framework offers a critical advancement. You can now quantify changepoint uncertainty with valid confidence sets, even when pre- and post-change distributions are unknown. This enables more reliable decision-making in complex, real-world streaming data applications, reducing reliance on restrictive modeling assumptions. Consider integrating this wrapper around existing detection algorithms for enhanced inference.

Key insights

A distribution-free framework localizes changepoints after sequential detection, offering valid confidence sets without prior distribution knowledge.

Principles

• Changepoint localization can be decoupled from detection.
• Exchangeability within pre- and post-change segments is a key assumption.
• Lower confidence sets isolate the "post-change part" for process monitoring.

Method

The method constructs lower and upper confidence sets using sequential conformal p-values and calibrator functions, then combines them into a two-sided interval. It tests the null hypothesis $T=t$ against alternatives $T>t$ or $T<t$.

In practice

• Use lower confidence bounds in quality control to identify earliest failure onset.
• Apply two-sided intervals to localize shifts in public sentiment for issue understanding.

Topics

• Changepoint Localization
• Sequential Change Detection
• Distribution-Free Methods
• Conformal Prediction
• Statistical Inference
• Time Series Analysis

Best for: AI Scientist, Research Scientist

Related on AIssential

• The Cost of Learning under Multiple Change Points — Endogenous confounding in multiple change-point environments necessitates selective, adaptive detection to achieve optimal online learning.
• Sequential statistical inference for Large Language Models: Representation, validity, and monitoring — Sequential statistical inference enhances LLM trustworthiness by treating interactions as dependent stochastic processes.
• Exploring Statistical Change Point Detection Techniques for Performance Anomaly Detection at Mozilla — Evaluating diverse change-point detection methods and ensembles can significantly improve performance anomaly detection accuracy over traditional statistical tests.
• High-Dimensional Robust Change-Point Detection via Angular Kernel Statistics — A new angular kernel scan framework robustly detects high-dimensional change-points in low sample size, even with heavy-tailed data.
• Distill-Belief: Closed-Loop Inverse Source Localization and Characterization in Physical Fields — Distill-Belief decouples Bayesian correctness from efficiency for inverse source localization using a teacher-student framework.
• A probabilistic framework for online test-time adaptation — The framework uses state-space modeling for robust online test-time adaptation under distributional shift.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.