Run / Time-Series Plot: See Whether Your Process Is Drifting Before You Blame the Data

The Run / Time-Series Plot plots data in true time order. That is the whole idea. Instead of collapsing a process into one average, you keep the sequence intact and look at how the measurement behaves from one period to the next.

In the Analyse phase of RDMAICS, this is often the first chart you should build. Before you test differences between groups, before you model anything, you need to know whether the process was stable while you were measuring it. A shifting process makes every later comparison harder to trust.

The tool applies three rules. Eight points on one side of the median. Six strictly increasing or strictly decreasing points. And a runs test when at least ten non-median points are available. Each rule is a way of asking the same question. Is this pattern more structured than random variation would produce?

Reach for it when you have a measurement collected over time, when you suspect a trend, or when someone claims an improvement and you want to see whether the sequence supports the claim.

The worked example here uses the sample data that ships with the tool. It is fictional. Treat it as a teaching case, not as evidence about any real operation.

The scenario is a service measurement taken across three stores. Each store contributes six observations. Three come from Week 1 and three come from Week 2, split by lunch, dinner and weekend trading. That gives eighteen rows in total.

The question behind the data is the kind you hear often. Store 3 looks higher than the others. Is something changing over time inside each store, or are the stores simply sitting at different levels with ordinary variation on top?

That distinction matters. If a store is drifting, you investigate what changed. If a store is stable but higher, you investigate the level itself. The run chart is the tool that separates those two conversations.

The tool asks for four columns. The observation or period label, the measurement, an optional group or stratum, and an optional series or scenario.

In the sample data the labels carry the context. Row 1 is Store 1 Lunch at 28.4. Row 8 is Store 3 Dinner at 39.8, the highest value in the set. Row 4 is Store 2 Lunch at 25.8, the lowest. The group column holds Store 1, Store 2 or Store 3. The series column holds Week 1 or Week 2.

Two settings do the heavy lifting. Separate charts by is set to Group / stratum, which splits the eighteen rows into three charts of six points each. Sort date/number labels chronologically is switched on, so the points stay in the order the process produced them.

Getting the order right is not optional. A run chart drawn out of sequence is not a run chart. It is a scatter of numbers with a line through it.

The tool returns three charts, one per store, each with n=6. Every one of them reports the same headline. No rule signal detected.

Store 1 has a median of 32.80 and 5 observed runs. Store 2 has a median of 27.90 and 4 observed runs. Store 3 has a median of 35.00 and 5 observed runs. A run is a stretch of consecutive points on the same side of the median, so the count tells you how often the series crossed back and forth.

The runs test line reads Not run for all three stores, with the note that the runs test needs 10 or more non-median points. The p value shows as a dash. That is the tool being honest rather than the tool failing. With six points per store, there is simply not enough sequence to compute the test.

So what can you say? None of the three stores shows eight points on one side of its median, because none of them has eight points at all. None shows six strictly increasing or decreasing points, for the same reason. The chart found no evidence of drift. It also did not have much opportunity to find any.

A no signal result is still useful. It tells you that the shape of the data does not justify a drift story. If someone argues that Store 3 is getting worse week on week, the chart does not support that. Store 3 sits at a median of 35.00 across six points with 5 runs, which is the pattern you would expect from a process bouncing around a level.

It also redirects you. The three medians differ. Store 2 sits at 27.90, Store 1 at 32.80 and Store 3 at 35.00. The run chart does not test whether those differences are meaningful, but it shows you that the difference lives in the level rather than in a trend. That points your next analysis at comparison between groups, not at time.

The third decision is about measurement itself. The runs test could not run. If this question matters, the action is to collect more points per store rather than to argue over six. The tool has told you exactly how many you need before the test becomes available.

You can also ask the tool for an AI review, which walks through what the charts show in plain language. Use it to check your own reading, not to replace it.

The most frequent error is reading the shape before checking the count. Six points will form a shape. All shapes look like something. The rules exist because the eye is poor at telling structure from noise, and with n=6 the rules cannot fire.

The second error is splitting too far. Separating by Group / stratum here produces three charts of six. If you also split by series, you would get charts of three. Each split halves your ability to detect anything. Split only when the groups truly represent different processes.

The third error is mixing the order up. If the period labels are sorted alphabetically rather than chronologically, a trend can appear or disappear purely from the sort. Keep the chronological setting on.

The fourth is treating the median line as a target. It is not. It is a reference line that the rules are measured against, and it moves with the data.

A run chart will not tell you why. It shows you that a pattern exists, or that one does not. The cause sits outside the chart, in the process, the people and the conditions. Every explanation you attach is a hypothesis you still have to test.

It will not set control limits. There are no upper or lower bounds here, no sigma calculation, and no capability statement. If you need those, you need a control chart, and that is a different tool with different assumptions.

It will not compare groups for you. The three medians in this case are 27.90, 32.80 and 35.00. The tool displays them. It does not test whether the gaps between them are larger than chance would produce. That requires a comparison test on the underlying data.

It will not rescue a small sample. With six points per store, the runs test does not run, and the eight point and six point rules are out of reach by construction. No setting changes that. Only more data does.

And it will not validate your measurement. If the numbers were recorded inconsistently, or if lunch and dinner are not really comparable, the chart will plot the error as faithfully as it plots the signal. The chart is downstream of your data quality, never upstream of it.