📊 statistics

Using Transcripts for Nonparametric Monitoring of Serial Dependence

This paper proposes novel nonparametric control charts based on transcripts and algebraic distances derived from ordinal patterns to effectively monitor serial dependence in process data, demonstrating their superior performance through simulations and a real-world chemical industry application.

Original authors: Christian H. Weiß, José M. Amigó

Published 2026-05-27

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Christian H. Weiß, José M. Amigó

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 quality control inspector at a factory. Your job is to watch a machine that produces widgets. Usually, the machine works perfectly, churning out widgets that are all independent of one another—like flipping a fair coin where the result of the last flip doesn't change the next one. This is called being "in control."

However, sometimes the machine gets "stuck" in a pattern. Maybe it starts making a big widget, then a small one, then a big one, then a small one, over and over. This is called "serial dependence." If your inspector tool isn't smart enough to spot this pattern, it might keep saying, "Everything is fine!" when the machine is actually broken.

The Problem with Old Tools
For over 100 years, inspectors have used "control charts" to watch these machines. But most of these old charts have a big weakness: they assume the machine is perfectly random. If the machine starts behaving in a pattern (even a weird, non-linear one), the old charts often fail. They might scream "False Alarm!" when nothing is wrong, or stay silent when the machine is actually broken.

The New Idea: Looking at the "Dance Steps"
The authors of this paper, Christian H. Weiß and José M. Amigó, wanted to build a better inspector. Instead of looking at the raw numbers (the size of the widgets), they decided to look at the order of the numbers.

Imagine you are watching a dance troupe. Instead of measuring how high each dancer jumps, you just look at the sequence of their movements: "Low, Medium, High."

Old Method (Ordinal Patterns): They looked at groups of three dancers and noted their order (e.g., "Low-Medium-High"). They then checked if these patterns happened randomly.
The New Method (Transcripts): The authors realized that looking at the difference between one dance sequence and the next is even more powerful. They invented a concept called a "Transcript."

Think of a Transcript like a translation key. If the first dance sequence was "Low-Medium-High" and the next one was "High-Low-Medium," the Transcript is the specific instruction on how to transform the first into the second. It's like asking, "What steps do I need to take to turn this dance move into that one?"

The "Distance" Meter
Once they have these "translation keys" (Transcripts), they measure the distance between them.

Imagine you have a map of a city. The "Cayley distance" is like counting how many bus stops you need to take to get from Point A to Point B.
The "Kendall distance" is like counting how many times you have to swap two people in a line to get them in the right order.

The authors found that the Kendall distance (the "swap" count) is the best ruler for spotting when the machine is acting weird.

The Three New Charts
Using these "translation keys" and "distance rulers," they built three new types of control charts:

The $\Delta\tau$ -Chart: Checks if the types of translations happening are changing.
The $\Delta K$ -Chart: Checks if the distance (number of swaps) between dance moves is changing.
The $\mu K$ -Chart: Checks the average distance. (This one turned out to be the superstar).

How They Tested It
They ran a massive simulation, like a video game where they created thousands of fake factory machines. Some machines were broken in simple ways (like a predictable up-and-down pattern), and some were broken in complex, weird ways (like sudden jumps or squiggly lines).

The Results

The Winner: The $\mu K$ -chart (the average distance checker) was the best at spotting problems, especially when the machine had a "negative" pattern (like a seesaw going up and down).
The Runner-Up: The $\Delta\tau$ -chart was also very good, especially for certain complex patterns.
Comparison: Both new charts were generally better than the old "Ordinal Pattern" charts, which often missed the subtle patterns or got confused by the weird ones.

Real-World Test
Finally, they tried their new charts on real data from a chemical factory (a batch process that makes chemicals). This factory was known to have a "seesaw" problem (if one batch is too good, the next one is usually too bad).

The old charts were slow to catch the problem.
The new charts (especially the $\mu K$ -chart) sounded the alarm much faster, catching the issue when the machine was only 23 or 24 steps into the process, whereas the old ones took much longer or missed it entirely.

The Bottom Line
The paper claims that by using these new "Transcript" and "Distance" tools, factory inspectors can spot broken patterns in their machines much faster and more reliably than before, without needing to know exactly what kind of math describes the machine's behavior. It's a smarter, more flexible way to watch the factory floor.

Technical Summary: Using Transcripts for Nonparametric Monitoring of Serial Dependence

Problem Statement
Statistical Process Monitoring (SPM) relies heavily on control charts to detect changes in a process $(X_t)$ . However, the vast majority of existing control charts assume that the monitored process (or its residuals) is serially independent and follows a specific parametric distribution (e.g., normal) under in-control (IC) conditions. When these assumptions are violated—specifically when serial dependence exists or the distribution is non-standard—the performance of standard charts degrades significantly, leading to frequent false alarms or delayed detection. While nonparametric charts based on ordinal patterns (OPs) have been proposed to address distribution-free monitoring, the authors note that recent research suggests superior power can be achieved by moving beyond ordinary OPs to utilize "transcripts" and algebraic distances derived from them.

Methodology
The paper proposes a novel framework for nonparametric control charts that monitors serial dependence by exploiting the algebraic and metric structures of the symmetric group $S_m$ .

Ordinal Patterns (OPs): The real-valued process $(X_t)$ is first discretized into a series of ordinal patterns $(\pi_t)$ of length $m=3$ . Under IC conditions (i.i.d. process), the distribution of these patterns is discrete-uniform.
Transcripts: Instead of monitoring the patterns directly, the authors compute the "transcript" series $(\tau_t)$ , defined as the permutation transforming one pattern into the next: $\tau_t = \pi_{t+1} \circ \pi_t^{-1}$ . This utilizes the group structure of permutations.
Algebraic Distances: The paper focuses on the Kendall distance ( $d_K$ ) between consecutive patterns, which is calculated directly from the transcript. The authors select $d_K$ over the Cayley distance ( $d_C$ ) based on previous findings that $d_K$ offers better discrimination power in time series analysis.
Control Chart Construction: Three new nonparametric control charts are developed using an Exponentially Weighted Moving Average (EWMA) approach to estimate the current probability mass function (PMF) of the transcripts and the Kendall distances:
- $\Delta_\tau$ -chart: Monitors the squared deviation of the estimated transcript PMF from the theoretical IC-PMF.
- $\Delta_K$ -chart: Monitors the squared deviation of the estimated Kendall distance PMF from its theoretical IC-PMF.
- $\mu_K$ -chart: Monitors the deviation of the estimated mean Kendall distance from its theoretical IC-mean ( $11/6$ ).

Key Contributions

Novel Chart Designs: The development of three specific nonparametric control charts ( $\Delta_\tau$ , $\Delta_K$ , $\mu_K$ ) that utilize transcripts and algebraic distances rather than raw ordinal patterns.
Theoretical Derivation: The paper provides the theoretical IC-distributions for transcripts and Kendall distances for $m=3$ , establishing the baseline for chart design without requiring parameter estimation.
Comprehensive Evaluation: A rigorous simulation study comparing the new charts against the state-of-the-art OP-EWMA charts (specifically the $\Delta_\pi$ , $\tau$ , and $\beta$ charts from Weiß & Testik, 2023) across various data-generating processes (DGPs).

Results
The performance of the charts was evaluated using Average Run Length (ARL) metrics under both IC and out-of-control (OOC) conditions for several types of serial dependence:

AR(1) Processes: For both positive and negative first-order autoregressive processes, the new charts generally outperformed the existing OP-EWMA charts. The $\mu_K$ -chart consistently showed the best performance, particularly for negative dependence when the smoothing parameter $\lambda = 0.05$ .
TEAR(1) and AAR(1) Processes: For the Transposed Exponential AR(1) and Absolute AR(1) processes (which exhibit asymmetry), the new charts were competitive. While the existing $\beta$ -chart remained superior for TEAR(1) due to its specific design for that asymmetry, the new $\Delta_\tau$ -chart performed best among the novel proposals.
QAR(1) and QMA(1) Processes: For Quadratic AR(1) and Quadratic Moving Average processes, the new charts performed well, often matching or exceeding the best existing charts. The $\mu_K$ -chart (for $\lambda \geq 0.10$ ) and $\Delta_\tau$ -chart (for $\lambda = 0.05$ ) were identified as the top performers among the new set.
Real-World Application: The charts were applied to a dataset of 70 batch chemical yields known to exhibit negative serial dependence (alternating high/low yields). The new charts successfully detected the violation of the i.i.d. assumption, with the $\mu_K$ and $\Delta_K$ charts (using $\lambda = 0.05$ ) triggering alarms earlier (around $t=23$ –$24$) than the competing OP-EWMA charts.

Significance and Claims
The authors claim that the proposed transcript-based control charts offer a robust, nonparametric method for monitoring serial dependence that is effective regardless of the underlying marginal distribution of the process. The study concludes that the new charts, particularly the $\mu_K$ and $\Delta_\tau$ variants, generally outperform or match the performance of existing ordinal pattern-based charts across a wide range of linear and nonlinear dependence structures. The paper emphasizes that these charts are "self-starting" in the sense that they do not require parameter estimation for the marginal distribution, making them immediately applicable in practice. The authors modestly suggest that future work could extend this transcript-based approach to spatial ordinal patterns.

More like this