Skip to content

Detection and Mode-Identification of Multiple Change Points in Tensor Factor Models

Home / Papers / Detection and Mode-Identification of Multiple Change Points in Tensor Factor Models

Detection and Mode-Identification of Multiple Change Points in Tensor Factor Models

Authors: Yuqi Zhang, Zetai Cen, Haeran Cho Date: 2026-04-13 Paper ID: openalex:2604.11300

Summary

This paper addresses structural change point detection in high-dimensional tensor-valued time series using Tucker decomposition-based factor models. The authors propose a two-stage approach: first, detecting change points by exploiting the underlying low-rank tensor structure, and second, identifying the specific modes associated with each change point. This mode-identification step is shown to improve the estimation accuracy of mode-wise loading spaces following segmentation. The methodology is validated through simulations and applied to New York City taxi usage and Fama-French portfolio returns.

Key Contributions

  • Proposes a change point detection algorithm for tensor-valued time series that leverages Tucker decomposition for computational efficiency.
  • Introduces a framework for mode-identification, enabling the pinpointing of specific tensor modes where structural shifts occur.
  • Establishes theoretical consistency for both detection and mode-identification under weak moment conditions.
  • Demonstrates superior post-segmentation estimation of loading spaces by incorporating mode-identification.

Key Concepts

Archivist Review

I have approved the core methodological contributions—the tensor-specific change point detection and the mode-identification framework—as these represent significant advancements in analyzing multi-dimensional time series. I rejected the datasets as they are common benchmarks, not unique research contributions. The approved concepts capture the structural inductive biases necessary for tensor time-series analysis.

Approved Concepts

  • Tensor Factor Model Change Point Detection: The framework addresses the non-trivial extension of change point detection from vector time series to high-dimensional tensors, which is a growing area in time-series analysis.
  • Mode-Identifiability in Tensor Time Series: Provides a novel approach to interpreting which dimensions of a tensor contribute to structural shifts, enhancing the explainability of tensor models.

Rejected Candidates

  • [dataset] New York City taxi usage (new-york-city-taxi-usage) - not_novel: This is a standard real-world benchmark used for demonstration rather than a seminal dataset worth a standalone entry.
  • [dataset] Fama—French portfolio returns (fama-french-portfolio-returns) - not_novel: This is a well-known, widely available financial dataset rather than a novel research-contributed dataset.

Metadata & Links

url
https://arxiv.org/abs/2604.11300
paper_id
2604.11300
paper_source
openalex
domain
time-series
tags
architectures
datasets
concept_slugs
tensor-factor-model-change-point-detectionmode-identifiability-in-tensor-time-series
dataset_slugs
skill
TimeSeriesSkill
created_at
2026-04-16T05:47:23Z