Detecting correlation changes in multivariate time series: A comparison of four non-parametric change point detection...

Area of Science:

• Statistics
• Data Science
• Time Series Analysis

Background:

• Change point detection in multivariate time series is challenging due to potential alterations in both mean and correlation structures.
• Existing methods like DeCon, E-divisive, Multirank, and KCP utilize diverse statistical approaches.
• A clear understanding of method differences and comparative performance, especially for correlation changes, is lacking.

Purpose of the Study:

• To elucidate the principles and algorithms of DeCon, E-divisive, Multirank, and KCP for improved accessibility.
• To conduct a direct performance comparison of these methods for detecting correlation changes in multivariate time series.
• To evaluate method robustness under varying simulation settings, including changes in mean, correlation, and the number of noise variables.

Main Methods:

• Detailed explanation of the underlying principles and algorithms for DeCon, E-divisive, Multirank, and KCP.
• Extensive simulations based on established settings (Bulteel et al., Matteson and James) to assess performance.
• Comparative analysis focusing on the detection of correlation structure changes and mean shifts.

Main Results:

• KCP demonstrated superior performance across most simulated scenarios.
• DeCon showed adequate performance in detecting correlation changes, particularly when more than two noise variables were present.
• The study highlights the strengths and weaknesses of each method under different conditions.

Conclusions:

• KCP is a highly effective method for change point detection in multivariate time series, especially for correlation changes.
• DeCon offers a robust alternative when dealing with a higher number of noise variables and correlation shifts.
• Applied researchers can leverage this comparison to select the most appropriate method for their specific multivariate time series analysis needs.