AI Glossary / Statistics and probability / Intermediate
EWMA chart
Definition
An EWMA chart is a statistical process-control chart that monitors an exponentially weighted moving average of successive observations. Recent observations receive more weight, while earlier observations retain progressively smaller influence. Comparing this smoothed statistic with suitable control limits can help detect sustained shifts in a process, including changes that individual measurements may not clearly reveal.
What it means
EWMA stands for exponentially weighted moving average. Its memory helps a monitoring system accumulate evidence rather than react only to the latest observation. A sequence of modest increases may gradually lift the chart even when none of those observations is individually extreme.
The chart is a tool for detecting changes relative to a reference process. A signal is a reason to investigate. It does not identify the cause, prove that a product is defective, or show that a model's predictions have become unfair.
How it works
Start with reference data representing a reasonably stable process. Define the monitored quantity, its baseline mean μ, and its baseline standard deviation σ. If observations are subgroup averages, use the standard deviation appropriate to those averages rather than the spread of individual items.
Choose a smoothing parameter λ between zero and one, excluding zero. Update the statistic using Z_t = λX_t + (1 − λ)Z_(t−1), commonly initializing Z_0 at μ. A smaller λ gives longer memory. When λ equals one, the statistic is just the current observation.
For independent observations with constant variance and a fixed initialization at μ, the variance at time t is σ²λ[1 − (1 − λ)^(2t)]/(2 − λ). This follows by adding the variances of the weighted independent observations. Startup limits can therefore use μ plus or minus L times the square root of this variance.
As time increases, the variance approaches σ²λ/(2 − λ). Some charts use these wider, steady-state limits from the beginning. State which convention is used, because early signals can differ. L is a design parameter; selecting three does not automatically deliver a desired long-run false-alarm rate.
Choose λ and L with the size of shift and detection performance in mind. Successive EWMA statistics are correlated even when the original observations are independent. Evaluate performance using appropriate average-run-length calculations or simulation.
Investigate serial correlation, seasonality, changing subgroup sizes, and changing measurement conditions. A chart applied to poorly modeled data can signal predictable structure rather than a new process problem. Preserve a documented response plan and avoid automatically resetting the baseline whenever an inconvenient signal occurs.
Three practical examples
- A manufacturer monitors average fill volume to identify a small, sustained departure from a stable filling process.
- A service team charts standardized processing-time residuals after accounting for predictable workload changes.
- An AI operations team monitors an error measure over comparable evaluation batches, investigating shifts alongside changes in data and labels.
Think of it this way
Think of a steering indicator with memory. One bump has limited influence, but several pushes in the same direction gradually move the indicator away from its reference position.
Worked example
Use an invented baseline mean of 100, observation standard deviation 2, λ = 0.2, and Z_0 = 100. The next observations are 102, 103, and 104.
The EWMA values are Z_1 = 100.4, Z_2 = 100.92, and Z_3 = 101.536. With L = 3, the startup standard deviations are approximately 0.4000, 0.51225, and 0.57266. The corresponding upper limits are 101.2, 101.53675, and 101.71797.
All three EWMA values remain below their respective upper limits. The third raw observation, 104, must not be compared directly with the EWMA limit: the limit applies to the smoothed statistic.
The steady-state standard deviation is 2√(0.2/1.8) = 2/3, giving limits of 98 and 102. These calculations illustrate two limit conventions; three observations do not establish that the process is stable.
Illustrative case study
Fictional case: Maple Support introduces an EWMA chart for ticket-processing times. An initial alert coincides with a shift in ticket complexity, so the team examines its baseline model before changing staff procedures. After adjusting for predictable case mix, it charts residuals and documents which signals require investigation. This invented example illustrates that a sensible monitored quantity matters as much as the smoothing formula.
Common misconceptions and mistakes
Do not confuse a control limit with a customer specification or a confidence interval for an individual observation. Each answers a different question.
A smaller λ does not universally improve detection: it changes the balance between responsiveness and memory. An absence of signals does not prove stability, and a signal does not prove a particular cause.
Simple Shewhart charts can also be supplemented with run rules; the comparison is not universally “EWMA has memory and every other chart does not.”
Business use
EWMA charts can support earlier investigation of persistent operational changes while reducing reactions to isolated fluctuations. Organizations should compare the cost of false alarms with the cost of delayed detection. Record baseline versions, sampling frequency, handling of missing observations, and who investigates each signal so that the chart supports a repeatable decision process.
People also ask
Can EWMA monitor model drift?
It can monitor a chosen metric, but the metric and limits must reflect the evaluation design.
Is exponential smoothing always a control chart?
No. A control chart adds a reference process, limits, and a monitoring interpretation.
Does λ = 1 keep historical memory?
No. Only the current observation remains.
Check your understanding
Which quantity is compared with the limits?
Answer: Z_t, the EWMA statistic.
Why are startup limits narrower?
Answer: Less accumulated random variation has entered from a fixed starting value.
Does the worked example prove no change occurred?
Answer: No. It shows only that these observations did not trigger the specified limits.
Related terms
- Statistical process control
- Exponential smoothing
- Control limits
- Shewhart chart
- CUSUM
- control limit
- average run length
- serial correlation
Learn next
Learn the distinction between control limits and specifications, then study CUSUM and average run length. Practice checking baseline stability before implementing automated alerts.
More in Statistics and probability
- Eta squared ( h 2 ) - Eta squared, conventionally written eta squared or η², is an effect-size measure based on the proportion of total observed variation attributed to a...
- Event correlation - Event correlation is the process of connecting event records that may belong to the same activity, incident, or operational pattern.
- EVOP, Evolutionary Operation - EVOP, short for Evolutionary Operation, is an approach to improving an operating process through deliberately planned, usually small changes to its...
- Exact test - An exact test is a hypothesis test calibrated using a finite-sample null distribution rather than relying on a large-sample approximation.
- Expander walk sampling - Expander walk sampling is a method of generating useful samples by following a random walk on a suitably well-connected graph called an expander.
- Expected frequencies - Expected frequencies are the counts predicted for categories or table cells by a specified probability model, often the null hypothesis in a statistical test.
Sources
NIST/SEMATECH, “EWMA Control Charts”: https://www.itl.nist.gov/div898/handbook/pmc/section3/pmc324.htm
Previous: EVOP, Evolutionary Operation Next: Ex ante forecast