EWMA Chart: Catch the Slow Drift Before It Undoes Your Gain

The EWMA Chart monitors the exponentially weighted moving average to detect small shifts in the process mean, and it is robust to non-normality. That single sentence carries a lot of weight in the Control phase of RDMAICS.

Most improvement projects fail quietly. Nobody reverses the change. Nobody announces a return to the old method. The process just eases back, half a unit at a time, until the gain has gone. A standard individuals chart often misses that, because no single reading looks strange enough to trip a rule.

The EWMA chart works differently. Each plotted point carries a share of every earlier point, with recent data weighted most heavily. A run of readings that are each slightly high adds up inside the statistic. The chart reacts to the pattern, not to one dramatic value.

Use it when you have handed a process back to its owner, when the thing you care about is the mean rather than the spread, and when the shift you fear is small and gradual rather than sudden.

The tool opens with sample data already loaded. This example is fictional, created to show the method rather than to report a real site.

The series is named Delivery-time drift. It records one observation each day for twelve days, with each day tagged as either a Lunch shift or a Dinner shift. The values run from Day 1 at 29 through to Day 12 at 30.9.

Read the column in order and the story is plain enough once you look for it. Day 1 is 29. Day 2 is 28.8. Then 29.2, 29.1, 29.4, 29.6, 29.8, 30.1, 30.3, 30.5, 30.6, 30.9. There is no jump. There is no bad day. There is a steady climb of roughly two units across twelve readings.

This is exactly the shape that defeats informal monitoring. A supervisor glancing at yesterday's number sees 30.9 and compares it to 30.6. The difference looks like noise. Compared to Day 1, it is not noise at all.

The input table has four columns, and only one of them is strictly required for the chart to run.

Observation / period holds the label for each row. In the sample that is Day 1 through Day 12. These labels do not enter the calculation. They exist so that when a signal appears you can name the day it belongs to and go find out what happened.

Observations in time order is the measured value. The asterisk on that column heading is a reminder that order matters. An EWMA chart assumes the rows are sequential. If you sort the table by value, or paste in data grouped by shift rather than by date, the chart becomes meaningless.

Group / stratum and Series / scenario are optional. In the sample, Group alternates between Lunch shift and Dinner shift, and Series carries the single label Delivery-time drift. Stratum labels let you ask later whether a drift sits in one part of the operation or across all of it.

The one setting you control is Lambda, the smoothing constant. The sample uses 0.2. A smaller lambda gives earlier readings more influence and makes the chart slower but more sensitive to tiny shifts. A larger lambda makes the chart behave more like a plot of the raw values.

The result panel reports four headline numbers. Points is 12. Centre is 29.775. Sigma is 0.2015. Out of control is 9.

The centre line of 29.775 is the average of all twelve readings. Sigma of 0.2015 is the control limit width used for the weighted statistic, and it is much narrower than the spread of the raw data. That narrowness is the point. The EWMA statistic is an average, and averages vary less than single values, so the limits close in and small shifts break through them.

Nine of the twelve points fall outside those limits. The signals table lists Points 1, 2, 3, 4, 5, 6, 7, 11 and 12. Every one of them carries Rule 1, one point beyond 3 sigma, with the explanation that a single reading this far out is almost never chance.

Look at which points signal and which do not. The early points, 1 through 7, sit below the centre. The late points, 11 and 12, sit above it. Points 8, 9 and 10 pass through the middle without comment. That is not a scatter of random alarms. That is the signature of a process whose mean has moved from one level to another, with the crossing happening around Day 8 to Day 10.

The tool also offers an AI review that explains the result in plain language if you want a second reading before you take it to a stakeholder.

A chart earns its place by changing what someone does. In the Control phase the decisions are usually about escalation and ownership.

The first decision is whether to act at all. Nine signals out of twelve is not a borderline call. The process mean has moved, and the chart gives you a date range to investigate rather than a vague sense that things feel slower.

The second decision is where to look. The sample alternates Lunch shift and Dinner shift, so you can ask whether the climb is confined to one shift or runs through both. The pattern here touches odd and even rows alike, which points away from a single shift as the cause.

The third decision is about the baseline. The note under the chart says the limits were calculated from all twelve points, and invites you to set a baseline above so that later data is judged against an earlier period. That is the proper way to run a Control phase chart. Freeze the limits on the period when the process performed as the project intended. Every later reading is then measured against the gain you were handed, not against the drifting average of recent weeks.

The most frequent error is recalculating limits every time new data arrives. If the limits follow the data, the chart will always look acceptable. A slow drift becomes the new centre line. Set the baseline on the stable period and leave it.

The second error is changing lambda until the chart says something comfortable. Lambda is a sensitivity choice made in advance. The sample uses 0.2. Pick a value, record why, and apply it consistently.

The third error is treating the EWMA chart as a substitute for a dispersion chart. The EWMA statistic tracks the mean. It says nothing about whether variation has widened. Run it alongside a range or dispersion chart.

A fourth error is responding to each signal as a separate incident. Here there are nine signals, but there is one problem. Investigating Day 1 and Day 2 as individual events would waste effort and miss the shift entirely.

The fifth error is sorting or filtering the data. Any reordering destroys the time structure that the whole method rests on.

The EWMA chart detects a change in the mean. It does not explain the change. Nothing in this output says why delivery times climbed from 29 to 30.9. Cause still comes from going to the process and asking.

It will not tell you whether the result matters to a customer. The chart has no specification limit in it. A two unit drift might be invisible to the people you serve, or it might be the difference between acceptable and unacceptable. That judgement sits outside the statistics.

It will not protect you from poor measurement. If the way the time is recorded changed partway through the twelve days, the chart will faithfully report a shift that exists only in the measurement system.

Twelve points is also a short series. The result reports exactly 12 points, and limits built on a short baseline are less stable than limits built on a longer one. Treat early charts as provisional and revisit them as the record grows.

Finally, the chart will not sustain anything by itself. It makes drift visible. Somebody still has to look at it on a set rhythm, and somebody still has to be accountable for responding. That is the Sustain work that follows Control, and no tool does it for you.