📊 statistics

White noise testing for functional time series via functional quantile autocorrelation

This paper introduces a novel, robust class of nonlinear tests for serial dependence in functional time series based on functional quantile autocorrelation, which eliminates the need for moment conditions and outperforms traditional methods in detecting complex dependencies and handling outliers.

Original authors: Ángel López-Oriona, Ying Sun, Hanlin Shang

Published 2026-05-12

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ángel López-Oriona, Ying Sun, Hanlin Shang

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 watching a river. In statistics, we often treat data like a series of snapshots of that river. Sometimes, we just look at the water level at noon every day (a single number). But in functional data analysis, we look at the entire shape of the river's flow for the whole day—a continuous curve.

The paper you provided introduces a new, smarter way to check if these daily river curves are truly random (like white noise) or if they are secretly connected to each other in a pattern (serial dependence).

Here is the breakdown of their new method, using simple analogies:

1. The Problem: The "Average" is a Liar

Traditionally, statisticians check for patterns in data by looking at the average and the spread (variance).

The Old Way: Imagine trying to understand a storm by only measuring the average wind speed. If the wind is usually calm but occasionally hits a massive, dangerous gust, the "average" might look calm. You miss the danger.
The Flaw: Existing tests for functional data (like daily stock price curves) rely heavily on these "averages" and "spreads." If the data has extreme outliers (like a sudden market crash) or weird, non-linear patterns, these old tests often fail. They are like looking at a storm through a foggy window; they miss the complex, chaotic details.

2. The Solution: The "Quantile Excursion" Map

The authors propose a new tool called Functional Quantile Autocorrelation (FQA). Instead of looking at the average, they look at thresholds and percentiles.

The Analogy: Imagine you are tracking a river not by its average depth, but by asking: "On what percentage of the river's width is the water deeper than 5 feet?"
How it works:
1. They pick a specific "height" (a quantile, like the 90th percentile).
2. They draw a line across the river curve at that height.
3. They measure the length of the river section that is above that line. This is called an "excursion set."
4. They repeat this for many different heights (from very low to very high).

By tracking how these "lengths" change from one day to the next, they can see if the river's shape is repeating itself in a pattern, even if the average looks random.

3. The New Test: The "All-Seeing" Omnibus Statistic

The authors built a test statistic (a mathematical score) that aggregates all these different measurements.

The Metaphor: Think of the old tests as a security guard checking only the front door. If a thief sneaks in the back, the guard misses it.
The New Test: This new method is like a security system with cameras on every wall, floor, and ceiling. It checks the data at every possible height and every possible threshold simultaneously.
The Result: It creates a single "Omnibus" score (a summary number) that tells you: "Is there any hidden connection between today's curve and yesterday's, no matter how weird or non-linear that connection is?"

4. Why It's a Game-Changer (Robustness)

The paper highlights two major superpowers of this new test:

It ignores the "screaming" outliers: If one day the river has a massive, freak wave (an outlier), traditional tests get confused and might think the whole system is broken. This new test is "robust," meaning it can ignore that one crazy spike and still see the underlying pattern. It's like a noise-canceling headphone that filters out a sudden shout so you can still hear the music.
It catches the "whispers": Sometimes the connection between days is very subtle and hidden in the tails (the extreme ends) of the data distribution. Old tests miss these whispers; this new test hears them.

5. The Proof: Simulations and Real Money

The authors didn't just talk about it; they tested it:

The Lab Test (Simulations): They created fake data with all sorts of problems: heavy tails, crazy spikes, and rough, jagged lines. The new test consistently found the patterns that the old tests missed. In fact, in some "rough" scenarios, the old tests failed completely (like a broken compass), while the new one worked perfectly.
The Real World Test (Stock Market): They applied this to intraday stock prices (minute-by-minute prices of companies like Google and Meta).
- The Finding: The stock market data is messy, full of spikes, and heavy-tailed. The new test found strong, hidden patterns of dependence in the data that other methods missed. It successfully detected that the market isn't just random noise; there are complex, non-linear rhythms to it.

Summary

This paper introduces a super-robust detective for functional time series.

Old Detective: Looks at the average. Misses the weird stuff. Gets confused by outliers.
New Detective (FQA): Looks at the shape of the data at every possible level. Ignores the noise. Finds the hidden, complex connections that others miss.

The authors conclude that this tool is essential for anyone analyzing complex, real-world data (like weather, biology, or finance) where the data isn't perfectly smooth or normal, and where missing a hidden pattern could be costly.

Technical Summary: White Noise Testing for Functional Time Series via Functional Quantile Autocorrelation

Problem Statement
The paper addresses the challenge of detecting serial dependence in functional time series, specifically testing the null hypothesis that a sequence of random functions $\{X_t, t \in \mathbb{Z}\}$ constitutes an independent and identically distributed (i.i.d.) sequence (strong white noise) in the Hilbert space $L^2([0, 1])$ . Existing methodologies for this task predominantly rely on second-order structures, such as covariance operators, linear functional autocorrelation functions (FACF), or projections onto principal components. The authors argue that these traditional approaches have significant limitations: they are sensitive to outliers and heavy-tailed noise, and they primarily capture linear dependencies within the mean-covariance structure. Consequently, they may fail to detect complex, nonlinear serial dependence, particularly when dependence is concentrated in specific regions of the conditional distribution or in the tails of the functional observations.

Methodology
The proposed methodology introduces a novel class of nonlinear tests grounded in the Functional Quantile Autocorrelation (FQA) framework.

Functional Quantile Autocovariance and FQA:
The core of the method utilizes quantile-defined excursion sets. For a quantile level $\tau \in (0, 1)$ , the functional quantile curve $q_\tau(u)$ is defined pointwise. An excursion set $A_t^\tau$ is defined as the set of points where the observed curve $X_t(u)$ lies below the quantile curve $q_\tau(u)$ . The method measures the Lebesgue measure $L(A_t^\tau)$ of these sets.
The functional quantile autocovariance, $\gamma(\tau, \tau', l, \beta, \beta')$ , is defined as the covariance between binary indicators of whether the Lebesgue measures of excursion sets at lag $l$ fall below specific thresholds $\beta$ and $\beta'$ . The FQA, $\rho$ , is the normalized version of this covariance. Crucially, this framework does not require finite moment conditions, making it robust to heavy tails and outliers.
Omnibus Test Statistic:
To aggregate information across a range of parameters, the authors propose an omnibus test statistic, $W_T(l)$ (or its discretized version $S_T(l)$ ), which integrates the squared estimated FQA over a grid of probability levels ( $\tau, \tau'$ ) and thresholds ( $\beta, \beta'$ ). This statistic aggregates dependence signals across the entire distribution of the functional data rather than focusing solely on the mean.
Asymptotic Theory:
The paper establishes the asymptotic properties of the proposed statistics under two scenarios:
- Known Quantile Curves: Under the null hypothesis of i.i.d. functional white noise, the sample FQA converges to a normal distribution, and the omnibus statistic converges to a weighted sum of independent chi-squared variables.
- Estimated Quantile Curves: The authors prove that replacing true quantile curves with consistent empirical estimators does not alter the asymptotic distribution, provided specific regularity conditions (uniform consistency of quantile curves and stability of excursion indicators) are met.
- Consistency: Under strict stationarity and ergodicity, the test is shown to be consistent against fixed alternatives where serial dependence exists.

Key Contributions

Novel Framework: The paper formalizes FQA for hypothesis testing, moving beyond its previous use in computational clustering. It provides a rigorous inference framework that captures dependence through quantile-defined excursion sets rather than second-order moments.
Robustness: The methodology explicitly avoids moment assumptions, rendering it valid for processes with heavy-tailed noise or infinite variance, where traditional autocovariance-based tests often fail.
Omnibus Statistic Construction: The authors construct a comprehensive test statistic that aggregates dependence information over grids of quantile levels and thresholds, allowing for the detection of complex, nonlinear dependencies that might be missed by tests focusing on specific lags or linear correlations.
Theoretical Guarantees: The paper derives the asymptotic distribution of the test statistics under both known and estimated quantile curves and establishes consistency under mild stationarity and ergodicity assumptions.

Empirical Results
The authors conduct extensive simulations and a real-world application to evaluate the methodology:

Simulation Study (Size): Under the null hypothesis of strong white noise, the proposed test maintains empirical sizes close to nominal levels (0.05 and 0.01) across various scenarios, including Brownian motion, Gaussian noise, heavy-tailed Student-t noise, and Cauchy-driven processes. In contrast, competing methods (FPC, FACF, SDO, FSACF) exhibited poor size control, particularly in scenarios with discontinuous paths or heavy tails (e.g., SDO and FACF showed severe over-rejection or under-rejection).
Simulation Study (Power): In scenarios with serial dependence (linear and threshold functional autoregressive models), the proposed omnibus test demonstrated superior power compared to competitors. It successfully detected dependence in settings with heavy-tailed noise and outliers where other tests failed. Notably, the test remained robust when data was contaminated with spikes, whereas the power of FACF and SDO dropped significantly.
Computational Efficiency: The proposed test demonstrated reasonable computational times, scaling efficiently with sample size compared to some spectral and spherical methods.
Application to Financial Data: The method was applied to intraday log-returns of six S&P 500 companies. The proposed test consistently detected significant serial dependence at lags 1 through 10 (p-values $\approx 0$ ), whereas competing tests yielded heterogeneous results, often failing to detect dependence in specific series. The authors note that the test also detected dependence at longer lags, which vanished when the temporal order was permuted, suggesting genuine nonlinear and tail dependence structures in the financial data.

Significance and Claims
The paper claims that the proposed FQA-based omnibus test expands the toolkit for functional time series analysis by providing a robust framework for inference in settings where traditional methods may fail. Its primary significance lies in its ability to:

Detect complex, nonlinear serial dependence that is not captured by linear correlations or second-order summaries.
Operate effectively without requiring finite moment conditions, making it suitable for data with outliers and heavy tails.
Provide reliable diagnostics for model selection and assessment in high-frequency financial data and other applications where data may exhibit roughness or extreme values.

The authors position this work as a foundational step, offering early but strong evidence for the utility of quantile-based excursion sets in independence testing for functional data, while acknowledging that future research could explore data-driven selection of quantile grids and portmanteau extensions.