All lessons

Module 8: Relationships and business data · Lesson 8.6 · about 20 min

Control charts: is this KPI drop real?

Tell real changes from everyday noise in a daily or weekly KPI. Build an individuals control chart from a calm baseline, read its signals (points outside the limits, runs on one side), and stop holding meetings about noise.

Key points, meanings and formulas on one page: choose "Save as PDF" in the print box.

Module 8 · about 20 minutes · uses Lesson 3.4. You’ll be able to build an individuals control chart for a KPI, read its signals, and explain to a manager why most day-to-day changes need no action.

A control chart draws the limits of ordinary variation, so you react to signals and leave noise alone.

In everyday terms: a fever thermometer. Your temperature wobbles a little all day and nobody calls a doctor for 36.9 °C. Above 38 °C is a signal. The control limits are the “normal range” for your KPI.

1. Two kinds of variation

Every process wobbles. Common-cause variation is the everyday noise of a stable process; special-cause variation is a real change: a broken machine, a new supplier, a software release. The costly mistakes are treating noise as a signal (the “why was Tuesday bad?” meeting) and missing real changes because the chart is always noisy.

A control chart (Shewhart, 1920s) separates the two.

Try it: signal or noise?
centre / limits (from days 1 to 20)
signals
A daily metric, such as the share of orders shipped late (%). The limits are set from a calm baseline (days 1 to 20) at the mean ± 3 sigma, with sigma estimated from the day-to-day changes. Points outside the limits, or 8 in a row on one side of the centre, are signals worth investigating; everything else is ordinary variation, and reacting to it (the "why was Tuesday bad?" meeting) only adds noise. Slide the shift and see how big a change must be before it is caught. Made-up numbers.

With no shift, the points wander inside the limits. Slide a shift of +1 point from day 25: a single point might not cross the limit, but a run of eight above the centre line soon gets flagged. A shift of +2.5 points is caught at once.

2. Building an individuals chart

  1. Pick a baseline of 20 or more points when the process was stable (no known changes).
  2. Centre line = the baseline mean.
  3. Sigma = average moving range ÷ 1.128 (the moving range is the absolute change from one point to the next).
  4. Limits = centre ± 3 sigma (lower limit at 0 if the metric can’t go negative).
  5. Plot new points against these fixed limits. Recompute only after a confirmed, lasting change.

3. Reading the signals

Common rules (Western Electric / Nelson, simplified):

  • One point outside the limits.
  • Eight in a row on one side of the centre line (a shift).
  • Six in a row steadily rising or falling (a drift).
  • Two of three points beyond 2 sigma on the same side.

Use a few rules, not all of them: each extra rule adds false alarms.

For rates (share of late orders out of varying daily volumes) a p-chart with limits that widen on low-volume days is better; for counts of rare events, a c-chart. The individuals chart is a good default for any KPI.

Centre line
The mean of a calm baseline period.
Moving range
Gap between each day and the day before; its average estimates day-to-day noise.
Limits
Centre ± 2.66 × average moving range (about ± 3 sigma).
Signals
A point outside the limits, or 8 in a row on one side of the centre.
Physical meaning
A process left alone produces values inside the limits about 99.7 % of the time. Points outside, or long runs on one side, mean the process itself changed: a new courier, a broken machine, a pricing error.

“Every bad day needs an explanation.” Chasing ordinary variation wastes time and leads to random fixes that add noise. Investigate signals; for everything inside the limits, improve the process as a whole.

Watch it happen, one step at a time:

Baseline. 20 calm days of late-shipping share: mean 4.0 %, average moving range 0.68.
Limits. 4.0 ± 2.66 × 0.68 → 2.2 % to 5.8 %.
Day 23: 5.4 %. Inside the limits, no run: ordinary variation. No root-cause meeting.
Days 27 to 34. Eight days in a row above the centre: a signal.
Find the cause. A new courier pickup time from day 26. Fix it, then recompute limits after 20 stable days.
Maths and principle

Principle: if a stable process is roughly normal, 99.7 % of points fall within 3 sigma of its mean, so a point outside is a strong hint that something changed.

Limits:

CL=x‾,σˆ=MR1.128,UCL, LCL=x‾±3σˆ

1.128 is the expected range of two normal values in sigma units.

False alarm rates: a point beyond 3 sigma by chance: 0.27 %, about one in 370 points. Eight in a row on one side: 2×0.58≈0.8% for any given run of eight.

Check: baseline mean 4.0 %, average moving range 0.68: σˆ=0.60, limits 2.2 % to 5.8 % (tests/stats.mjs checks controlLimits).

4. Worked example

Made-up warehouse: share of orders shipped late, daily.

  • Baseline: 20 calm days, mean 4.0 %, average moving range 0.68, limits 2.2 % to 5.8 %.
  • Day 23: 5.4 %. The ops manager asks for a root-cause review. The chart says: inside the limits, no run; ordinary variation. No review.
  • Days 27 to 34: eight days in a row between 4.3 % and 5.5 %. Signal: a shift. Investigation finds a new courier pickup time from day 26.
  • After fixing the pickup time, the series returns below the centre line. Keep the old limits until the new level has 20 stable days, then recompute.

5. Practice

  1. Baseline mean 120 calls a day, average moving range 18. Limits?
  2. Why not set limits at the target the business wants?
  3. A chart has had no signals for 3 months, but the KPI is worse than the target. What does that mean?
Answers
  1. Sigma = 18 ÷ 1.128 ≈ 16. Limits 120 ± 48: 72 to 168.
  2. Limits describe what the process actually does (the voice of the process); targets are what you want (the voice of the customer). Mixing them hides real signals or invents false ones.
  3. The process is stable but not good enough: improvement needs a change to the system (a redesign or new resources), not chasing individual bad days.

Try it on your own data: the control charts tool draws individuals, X-bar and p-charts for your own numbers and flags the signals; normal lab shows where the 3-sigma limits sit on a bell curve.

New words: control chart, common-cause variation, special-cause variation, moving range, control limits, run rule, p-chart.

  • Set limits from a calm baseline: centre ± 2.66 × average moving range.
  • Signals: a point beyond the limits, or 8 in a row on one side.
  • Do not explain every wobble inside the limits.
  • Recompute limits only after a deliberate, stable change.

Check yourself

1A KPI's control limits are 2.2 to 5.8. Today it is 5.4, the highest in two weeks. Is it a signal?
2Eight days in a row sit just above the centre line, all inside the limits. What does that suggest?
3Why estimate sigma from the average day-to-day change (moving range) rather than the overall SD?

Sources

Your ticks stay in this browser. Save them as a file on the course home.

Updated 2026-11-22. Numbers in the examples are made up to show the idea unless a source is given. Simulations run in your browser; nothing you type is sent anywhere.