📊 statistics

Factor-Adjusted Location Tests for High-Dimensional Time Series

This paper proposes and validates three factor-adjusted statistical tests (max, quadratic, and Cauchy combination) for high-dimensional time series mean testing under strong common serial dependence, establishing their asymptotic properties and demonstrating reliable performance through simulations and real-world applications.

Original authors: Jiyang Wang, Xifen Huang, Long Feng

Published 2026-08-11

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiyang Wang, Xifen Huang, Long Feng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery in a city of thousands of people. You have a list of suspects (data points), and you want to know if the "average" person in this group is behaving normally or if something strange is going on. In the world of statistics, this is called testing a "mean." But here's the catch: in the modern world, we often have way more suspects than clues. We might have data on 300 different stocks, but only 50 weeks of history. This is the realm of "high-dimensional" data, where the number of things you are measuring is huge compared to the number of times you measured them.

Usually, if you want to find the average behavior, you just add everything up and divide. But in this high-dimensional world, things are messy. The suspects aren't acting alone; they are influenced by a few hidden "bosses" (called latent factors) that make them move in sync. If a storm hits the city, every stock might drop at the same time, not because of their own unique problems, but because of the storm. This creates a "serial correlation," a pattern where today's data looks suspiciously like yesterday's. If you ignore these hidden bosses and the patterns they create, your detective work will go wrong. You might think you found a criminal (a significant change in the average) when you actually just saw the weather change. This paper tackles the tricky problem of finding the true average in a noisy, high-dimensional world where everything is connected by these invisible, moving forces.

The authors, Jiyang Wang, Xifen Huang, and Long Feng, have built a new set of tools to solve this specific puzzle. They call their method "Factor-Adjusted Location Tests." Think of it as a three-step magic trick to clean up the data before you try to find the average.

First, they identify the "hidden bosses." They look at how the data moves over time to find the few underlying forces (like market trends or economic cycles) that are pulling the strings. Once they find these forces, they mathematically "project" the data onto a shadow wall that is perpendicular to them. In plain English, they subtract the influence of the hidden bosses, leaving behind only the unique, individual quirks of each data point. This step is crucial because it removes the strong, confusing connections that usually break standard statistical tests.

Second, they realize that the "crime" they are looking for might look different depending on the situation. Sometimes, only a tiny handful of suspects are acting up (a "sparse" signal), like one stock crashing while others are fine. Other times, almost everyone is acting a little weird at the same time (a "dense" signal), like a slow, creeping inflation affecting everything. To catch both types of criminals, they built three different detectors:

The Max Test: This is a sniper. It looks for the single loudest signal. It's perfect when only a few things are wrong.
The Sum Test: This is a net. It adds up all the tiny, small signals. It's perfect when many things are slightly wrong.
The Cauchy Combination: This is a smart referee. It combines the sniper and the net. If you don't know whether the problem is sparse or dense, this tool adapts and uses whichever detector is best for the job.

The paper proves that these tools work mathematically, even when the data is huge and the hidden bosses are strong. They showed that by removing the factors first, the math behind their tests stays reliable. They also developed a special "bootstrap" method—a way of simulating thousands of fake datasets to check if their tools are working correctly in real-world scenarios.

To see if their idea actually works, the team ran thousands of computer simulations. They created fake data with strong hidden connections and tested their new methods against older ones. The results were clear: the old methods often cried "wolf" when there was no wolf (they made too many false alarms), especially when the hidden bosses were strong. The new factor-adjusted methods, however, kept their cool, controlling the error rates exactly as they should. They also showed that their tools are powerful enough to catch the "wolves" when they are actually there, whether the wolves are hiding in the shadows (sparse) or running in a pack (dense).

Finally, they took their tools to the real world, applying them to weekly stock returns from the S&P 500. They found that the raw stock data was full of strong, shared patterns (the hidden bosses). After cleaning out these patterns, their tests detected a "sparse" signal: a small group of stocks that had a genuine, unusual average return, while the rest of the market was behaving normally. This suggests that the market isn't moving as a single, uniform block; rather, specific, isolated opportunities or risks exist that only a method like theirs could find.

In short, this paper doesn't just offer a new way to count averages; it offers a way to see through the noise of the modern, interconnected world. By first identifying and removing the invisible forces that tie everything together, their method allows statisticians to spot the true, unique signals hidden within the chaos, whether those signals are a single loud shout or a quiet chorus.

Technical Summary: Factor-Adjusted Location Tests for High-Dimensional Time Series

Problem Statement
The paper addresses the problem of testing whether the mean vector of a high-dimensional time series is zero ( $H_0: \boldsymbol{\mu} = \mathbf{0}$ ) when the dimension $p$ is comparable to or larger than the sample size $n$ . A critical challenge in this setting is the presence of strong common serial dependence driven by latent dynamic factors. Classical multivariate tests (e.g., Hotelling's $T^2$ ) fail when $p \geq n$ due to the singularity of the sample covariance matrix. Furthermore, existing high-dimensional tests designed for independent observations or weakly dependent residuals often suffer from severe size distortion (inflated Type I error) when applied to data with strong, low-rank dynamic factors, as these factors induce long-range correlations that standard long-run covariance estimators cannot adequately capture without explicit structural modeling.

Methodology
The authors propose a factor-adjusted framework that explicitly models and removes the latent dynamic component before testing the mean of the residuals. The procedure consists of the following steps:

Dynamic Factor Estimation: The data $\mathbf{Y}_t$ is modeled as $\mathbf{Y}_t = \boldsymbol{\mu} + \mathbf{A}\mathbf{f}_t + \boldsymbol{\varepsilon}_t$ , where $\mathbf{f}_t$ is a low-dimensional latent dynamic factor and $\boldsymbol{\varepsilon}_t$ is a vector white noise process. The loading space $\mathcal{M}(\mathbf{A})$ is estimated using the Lam–Yao method, which utilizes the eigenspace of a population matrix $\mathbf{M}$ constructed from lagged autocovariances. The number of factors $\hat{r}$ is selected via a ratio rule on the eigenvalues of the sample counterpart $\hat{\mathbf{M}}$ .
Projection: The observations are projected onto the orthogonal complement of the estimated loading space: $\hat{\mathbf{U}}_t = \hat{\mathbf{Q}}\mathbf{Y}_t$ , where $\hat{\mathbf{Q}} = \mathbf{I}_p - \hat{\mathbf{A}}\hat{\mathbf{A}}^\top$ . This step eliminates the common serial dependence.
Standardization: The projected residuals are standardized by their estimated variances to form $\hat{\mathbf{Z}}_t$ .
Test Statistics Construction: Three test statistics are constructed based on the standardized projected data:
- Max Statistic ( $T_M$ ): Designed for sparse alternatives where only a few coordinates of the mean vector are non-zero. It utilizes the maximum of the squared standardized means.
- Sum Statistic ( $T_S$ ): A quadratic form designed for dense alternatives where many small coordinate-wise effects accumulate. It aggregates the squared means with corrections for finite-sample variance and correlation estimation.
- Cauchy Combination Statistic ( $T_C$ ): A combination of the $p$ -values from $T_M$ and $T_S$ using the Cauchy combination rule. This is designed to be adaptive to unknown sparsity levels, maintaining power under both sparse and dense alternatives.
Calibration:
- Asymptotic Calibration: Theoretical limits are derived for the statistics under the null hypothesis.
- Bootstrap Calibration: A "random-loading residual bootstrap" is developed for finite-sample inference. This method resamples the estimated residuals and re-estimates the factor loadings using a randomly generated orthonormal matrix in each bootstrap iteration, thereby mimicking the uncertainty in the factor space estimation.

Key Theoretical Contributions
The paper establishes a rigorous asymptotic theory under the assumption that the idiosyncratic errors are sub-Gaussian vector white noise and the factors are weakly stationary and geometrically strongly mixing.

Null Limits and Independence: The authors prove that the max statistic $T_M$ converges to a Gumbel distribution, and the sum statistic $T_S$ converges to a standard normal distribution. Crucially, they establish the asymptotic independence between $T_M$ and $T_S$ . This independence justifies the validity of the Cauchy combination test, as the combination of two asymptotically independent $p$ -values (derived from Gumbel and Normal limits) yields a standard Cauchy limit.
Robustness to Factor Estimation: A central theoretical result is that estimating and removing the dynamic factor space does not alter the first-order null limits of the projected statistics. The "drift" caused by the estimation error of the loading space is shown to be negligible under specific growth conditions (e.g., $p = o(n^2)$ in the strong-factor case).
Local Power Analysis: For the sum statistic, the paper derives a non-central normal limit under local dense alternatives, providing a closed-form asymptotic power function. The analysis shows that the power depends on the signal strength relative to the correlation structure of the residuals.
Bootstrap Validity: The paper proves the first-order validity of the proposed random-loading residual bootstrap, showing that it correctly approximates the null distribution of the test statistics in finite samples.

Empirical Results

Simulation Studies: Extensive simulations demonstrate that the proposed factor-adjusted tests (FA-Max, FA-Sum, FA-CC) achieve reliable size control (close to the nominal 0.05 level) even in the presence of strong common serial dependence. In contrast, competing methods that do not account for dynamic factors (e.g., Srivastava–Du, Wang–Shao, Zhang–Wu) exhibit severe size distortion (oversizing) in these settings.
Power Performance: The simulations confirm the expected trade-off: the Max test is most powerful under sparse alternatives, the Sum test under dense alternatives, and the Cauchy combination test maintains competitive power across a wide range of sparsity levels. The proposed methods do not suffer from artificial power loss when no common factors are present.
Real Data Application: The method is applied to weekly excess returns of S&P 500 stocks ( $n=522, p=333$ ). The analysis reveals that raw returns exhibit strong serial dependence, which is effectively removed by the factor adjustment (leaving residuals close to white noise). The global test rejects the zero-mean hypothesis, driven primarily by a sparse signal (a small subset of stocks with significant excess returns) rather than a dense market-wide shift, a conclusion supported by the specific rejection of FA-Max and FA-CC but not FA-Sum.

Significance and Claims
The paper claims to provide a unified and adaptive framework for high-dimensional mean testing in time series with strong common dependence. Its primary significance lies in:

Explicit Handling of Dynamic Factors: Unlike methods relying on long-run covariance estimation, this approach explicitly models and removes the low-rank dynamic component, which is essential for valid inference in financial and macroeconomic panels.
Adaptivity: The Cauchy combination test offers a solution to the "unknown sparsity" problem, allowing analysts to test for mean shifts without prior knowledge of whether the signal is sparse or dense.
Theoretical Rigor: The proof of asymptotic independence between max and sum statistics in the presence of estimated factors is a novel contribution that enables simple analytic calibration for the combined test.
Practical Robustness: The random-loading residual bootstrap provides a practical tool for finite-sample inference that accounts for the variability in factor estimation, demonstrating reliable performance in both simulations and real-world financial data.

The authors conclude that their method achieves reliable size control and competitive power, making it a robust tool for high-dimensional inference in the presence of latent dynamic structures.

Technical Summary: Factor-Adjusted Location Tests for High-Dimensional Time Series

More like this