Skip to main content

LUP Student Papers

LUND UNIVERSITY LIBRARIES

Comparative Evaluation of Structural Break Detection Methods in High-dimensional Time Series: A Simulation-Based Study with Financial Applications

Yan, Liwen LU (2025) NEKP01 20251
Department of Economics
Abstract
This paper focuses on the detection of structural breaks in the second-order matrices of multivariate time series. A comprehensive simulation-based comparison is conducted to measure the performance of three widely cited nonparametric methods that allow for serial dependence in the data: the pure CUSUM procedure with binary segmentation in Aue et al.,
(2009), the CUSUM-based sparsified binary segmentation approach in Cho & Fryzlewicz,(2015), and the MuBreD method of Preuss et al., (2015). These methods are all designed to detect multiple breakpoints in covariance matrices under various data structures. Through carefully designed Monte Carlo simulations, we evaluate the methods through different break
settings, data-generating processes... (More)
This paper focuses on the detection of structural breaks in the second-order matrices of multivariate time series. A comprehensive simulation-based comparison is conducted to measure the performance of three widely cited nonparametric methods that allow for serial dependence in the data: the pure CUSUM procedure with binary segmentation in Aue et al.,
(2009), the CUSUM-based sparsified binary segmentation approach in Cho & Fryzlewicz,(2015), and the MuBreD method of Preuss et al., (2015). These methods are all designed to detect multiple breakpoints in covariance matrices under various data structures. Through carefully designed Monte Carlo simulations, we evaluate the methods through different break
settings, data-generating processes and sample sizes. Our results show that CUSUM-BS has advantages in i.i.d. and factor model data, and outperforms in large sample sizes; SBS is particularly suitable for small samples; and MuBreD has balanced performance among all sample sizes and models, but large errors, especially in large samples. To complement the simulation findings, we apply the three methods to a real-world dataset, which specifically contains the S&P 500 index, gold prices, Brent oil prices, CBOE VIX and the US dollar index (DXY). All three methods successfully detected breaks during big financial events, and the results reconfirm the simulation-based insights, with a more practical perspective. (Less)
Please use this url to cite or link to this publication:
author
Yan, Liwen LU
supervisor
organization
course
NEKP01 20251
year
type
H2 - Master's Degree (Two Years)
subject
keywords
Structural breaks detection, CUSUM, Binary segmentation, High-dimensional data analysis
language
English
id
9193967
date added to LUP
2025-09-12 11:18:27
date last changed
2025-09-12 11:18:27
@misc{9193967,
  abstract     = {{This paper focuses on the detection of structural breaks in the second-order matrices of multivariate time series. A comprehensive simulation-based comparison is conducted to measure the performance of three widely cited nonparametric methods that allow for serial dependence in the data: the pure CUSUM procedure with binary segmentation in Aue et al.,
(2009), the CUSUM-based sparsified binary segmentation approach in Cho & Fryzlewicz,(2015), and the MuBreD method of Preuss et al., (2015). These methods are all designed to detect multiple breakpoints in covariance matrices under various data structures. Through carefully designed Monte Carlo simulations, we evaluate the methods through different break
settings, data-generating processes and sample sizes. Our results show that CUSUM-BS has advantages in i.i.d. and factor model data, and outperforms in large sample sizes; SBS is particularly suitable for small samples; and MuBreD has balanced performance among all sample sizes and models, but large errors, especially in large samples. To complement the simulation findings, we apply the three methods to a real-world dataset, which specifically contains the S&P 500 index, gold prices, Brent oil prices, CBOE VIX and the US dollar index (DXY). All three methods successfully detected breaks during big financial events, and the results reconfirm the simulation-based insights, with a more practical perspective.}},
  author       = {{Yan, Liwen}},
  language     = {{eng}},
  note         = {{Student Paper}},
  title        = {{Comparative Evaluation of Structural Break Detection Methods in High-dimensional Time Series: A Simulation-Based Study with Financial Applications}},
  year         = {{2025}},
}