$\ell^2$ Inference for Change Points in High-Dimensional Time Series via a Two-Way MOSUM

J. Jenny Li, Likai Chen, Ning Wang, Wei Biao WuarXiv 2022 · DOI · arXiv

Abstract

We propose an inference method for detecting multiple change points in high-dimensional time series, targeting dense or spatially clustered signals. Our method aggregates moving sum (MOSUM) statistics cross-sectionally by an $\ell^2$-norm and maximizes them over time. We further introduce a novel Two-Way MOSUM, which utilizes spatial-temporal moving regions to search for breaks, with the added advantage of enhancing testing power when breaks occur in only a few groups. The limiting distribution of an $\ell^2$-aggregated statistic is established for testing break existence by extending a high-dimensional Gaussian approximation theorem to spatial-temporal non-stationary processes. Simulation studies exhibit promising performance of our test in detecting non-sparse weak signals. Two applications, analyzing equity returns and COVID-19 cases in the United States, showcase the real-world relevance of our proposed algorithms.

Introduction

In modern data science, high-dimensional time series—where the number of variables $p$ is large relative to the number of observations $n$—are common in fields like finance, neuroscience, and epidemiology. A fundamental challenge is identifying change points: discrete moments when the underlying statistical properties of the data, such as the mean, undergo a sudden shift.

Traditional methods for change point detection often rely on the $\ell^\infty$-norm, which identifies the single largest coordinate-wise change. While effective for "sparse" signals where only a few variables jump, these methods struggle with "dense" or "clustered" signals where many variables experience small, simultaneous shifts. In such cases, the signal-to-noise ratio for any single variable is low, making detection difficult. This paper addresses this gap by proposing an $\ell^2$-aggregation framework that "boosts" weak but widespread signals.

Methodology

The proposed framework is built upon the Moving Sum (MOSUM) technique. A standard MOSUM statistic compares the average of a data window to its left with the average of a window to its right. The authors introduce two primary innovations:

  1. $\ell^2$-Aggregated MOSUM: Instead of taking the maximum coordinate-wise change, the method aggregates differences across all $p$ dimensions using an $\ell^2$-norm. This approach leverages the fact that while noise grows at a rate of $\sqrt{p}$, a dense signal grows linearly with the number of affected components.
  2. Two-Way MOSUM: This generalizes the moving window to both time and "space" (cross-sectional groupings). By searching through spatial-temporal regions, the algorithm can pinpoint both when and where (in which group of variables) a break occurs. This is particularly useful when variables have a natural ordering or grouping, such as stocks in the same sector.

To provide a rigorous basis for statistical inference, the authors extend the High-Dimensional Gaussian Approximation (GA) theorem. They prove that the distribution of the $\ell^2$-aggregated statistic can be accurately approximated by the maximum of a Gaussian vector, even under complex temporal and spatial dependencies.

Key Results

The effectiveness of the $\ell^2$ approach was validated through extensive simulations and real-world applications.

Simulation Performance

In synthetic datasets, the $\ell^2$ MOSUM demonstrated a clear advantage over the $\ell^\infty$ method as the signal became denser.

Comparison of detection power between $\ell^2$ and $\ell^\infty$ aggregated MOSUM statistics.

As shown in Figure 2, when 10% or more of the components experience a jump, the $\ell^2$ MOSUM achieves nearly 100% detection power. In contrast, the $\ell^\infty$ method's power remains significantly lower, often struggling to reach 40%. Interestingly, the "curse of dimensionality" is partially reversed here: a larger $p$ actually improves the accuracy of the Gaussian approximation by providing more information to smooth out errors.

Real-World Applications

  • US Equity Returns: Analyzing 300 monthly stock returns (2000–2020), the algorithm identified two major structural breaks: June 2009 (post-2008 crisis rebound) and February 2020 (COVID-19 market crash).
  • COVID-19 Trends: In a dataset of daily cases across 58 US states, the Two-Way MOSUM identified major waves (Initial, Delta, Omicron) and revealed regional leads and lags, such as the Northeast's early first wave and the Midwest's early Delta surge.

Discussion and Future Work

While the $\ell^2$ MOSUM offers significant advantages for dense signals, its performance is sensitive to the choice of "bandwidth" (window size). A window that is too small increases noise, while one that is too large may merge distinct change points. Future research into multi-scale MOSUM approaches could mitigate this. Additionally, while the $O(p^2)$ complexity of the Two-Way MOSUM is manageable, further optimization may be needed for extremely high-dimensional fields like genomics. Finally, extending the theory to detect changes in variance or covariance structures remains an important open challenge.

Summary updated

Abstract

We propose an inference method for detecting multiple change points in high-dimensional time series, targeting dense or spatially clustered signals. Our method aggregates moving sum (MOSUM) statistics cross-sectionally by an $\ell^2$-norm and maximizes them over time. We further introduce a novel Two-Way MOSUM, which utilizes spatial-temporal moving regions to search for breaks, with the added advantage of enhancing testing power when breaks occur in only a few groups. The limiting distribution of an $\ell^2$-aggregated statistic is established for testing break existence by extending a high-dimensional Gaussian approximation theorem to spatial-temporal non-stationary processes. Simulation studies exhibit promising performance of our test in detecting non-sparse weak signals. Two applications, analyzing equity returns and COVID-19 cases in the United States, showcase the real-world relevance of our proposed algorithms.
BibTeX
@misc{li2022ell,
  author = {J. Jenny Li and Likai Chen and Ning Wang and Wei Biao Wu},
  title = {{$\textbackslash\{\}ell^2$ Inference for Change Points in High-Dimensional Time Series via a Two-Way MOSUM}},
  year = {2022},
  eprint = {2208.13074},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2208.13074},
}
Figure