Abstract
In this paper, we consider detecting and estimating breaks in heterogeneous mean functions of high-dimensional functional time series which are allowed to be cross-sectionally correlated and temporally dependent. A new test statistic combining the functional CUSUM statistic and power enhancement component is proposed with asymptotic null distribution theory comparable to the conventional CUSUM theory derived for a single functional time series. In particular, the extra power enhancement component enlarges the region where the proposed test has power, and results in stable power performance when breaks are sparse in the alternative hypothesis. Furthermore, we impose a latent group structure on the subjects with heterogeneous break points and introduce an easy-to-implement clustering algorithm with an information criterion to consistently estimate the unknown group number and membership. The estimated group structure can subsequently improve the convergence property of the post-clustering break point estimate. Monte-Carlo simulation studies and empirical applications show that the proposed estimation and testing techniques have satisfactory performance in finite samples.
Introduction
Modern statistical analysis increasingly encounters functional data—observations that are represented as entire curves or functions rather than discrete scalar points. Examples include intraday stock price movements or age-specific mortality profiles. When hundreds of these curves evolve simultaneously over time, they form a "high-dimensional functional time series." A primary challenge in analyzing such data is the identification of structural breaks: sudden, significant changes in the underlying mean of the process, such as those caused by market crashes or major policy shifts.
Existing methods for change-point detection often falter in high-dimensional settings due to two main factors: sparsity and heterogeneity. Sparse breaks occur when only a small fraction of subjects experience a change, causing the signal to be "washed out" during aggregation. Heterogeneity refers to the fact that different subjects may experience breaks at different times. A new framework by Li, Li, and Shang provides a robust toolkit to overcome these hurdles, even when data is cross-sectionally correlated and temporally dependent.
Methodology
1. Detection: The PE-CUSUM Test
The researchers propose a two-stage framework, beginning with a "Power Enhanced" Cumulative Sum (PE-CUSUM) test. Traditional CUSUM statistics measure the aggregate shift across all subjects. However, if the break is sparse, the average shift may appear negligible.
The PE-CUSUM test incorporates a Power Enhancement (PE) component that acts as a high-criticism filter. It scans individual subjects and adds a boost to the test statistic if specific subjects show strong evidence of a break. This ensures that even if the average shift is small, a few strong signals can trigger a detection, maintaining high power without increasing the rate of false positives (size distortion).
2. Estimation: Latent Group Clustering
Once a break is detected, the next challenge is accurately dating it. The authors assume a latent group structure where subjects are clustered based on their break points. Their proposed algorithm involves:
- Individual Estimation: Estimating potential break points for every subject independently.
- Clustering and Information Criterion (IC): Using a clustering algorithm and an IC to determine the optimal number of distinct "break events" or groups.
- Refinement: Pooling data within each identified cluster to "borrow strength," which significantly improves the precision and convergence rate of the final break date estimates.
Empirical Results
Simulation Studies
The authors validated their approach using Monte-Carlo simulations across various signal-to-noise ratios. In scenarios where only 10% of subjects experienced a break (a Sparse Detection Ratio of 0.1) and the signal was weak, the PE-CUSUM test maintained a 96% detection power. In contrast, the standard CUSUM test's power dropped to a mere 1.8%, demonstrating the necessity of the power enhancement component.
Real-World Applications
- Financial Markets: Analyzing intraday 5-minute returns for 28 Dow Jones Industrial Average (DJIA) stocks from 2018 to 2021, the model identified a common structural break on March 16, 2020. This date corresponds to the peak of COVID-19-induced market volatility, successfully unifying heterogeneous signals into a single, interpretable event.
- Demography: Applying the method to age-specific mortality data for 32 countries (1960–2013), the model identified a common structural break around 1988 for male mortality. This likely reflects shared global advancements in healthcare and lifestyle changes across developed nations.
Conclusion and Future Directions
The PE-CUSUM framework represents a significant advancement in high-dimensional statistics by explicitly allowing for cross-sectional correlation and temporal dependence—features often ignored in simpler models. While the current methodology is optimized for detecting at least one break per subject, future research may extend this to handle multiple breaks within a single series. Additionally, as dimensions grow into the thousands, further optimization of the Information Criterion's search space will be essential for maintaining computational efficiency.
Reference: Li, D., Li, R., & Shang, H. L. (2023). Detection and Estimation of Structural Breaks in High-Dimensional Functional Time Series. arXiv:2304.07003v1.
Summary updated