Learn · updated 2026-10-28
CUSUM Chart: Catch a Small, Lasting Shift
A CUSUM chart (cumulative sum chart) plots the running total of deviations from a target, so a small shift that lasts (0.5σ to 1σ) builds up into a signal much sooner than a single point would cross a Shewhart limit.
A Shewhart chart looks at each point on its own, so a lasting shift of half a standard deviation runs for about 155 points, on average, before one lands outside the limits. A CUSUM keeps a running total of how far the process sits above (or below) target, and that total grows quickly once the mean really moves.
Upper sumS_hi = max(0, previous S_hi + x − target − k)Lower sumS_lo = max(0, previous S_lo + target − x − k)Signaleither sum above the decision limit hUsual settingsk = 0.5σ, h = 4σ to 5σ (NIST's example uses h ≈ 6.6σ, from its V-mask design)
Worked Example (NIST)
Twenty group means, target 325, σ of the means 0.635. NIST uses k = 0.3175 (half a σ) and h = 4.1959. The upper sum stays near zero for 12 groups, then the mean moves up by about 3 and the sum climbs.
| Group | Mean | x − 325 − 0.3175 | Upper sum |
|---|---|---|---|
| 12 | 325.150 | -0.168 | 0.00 |
| 13 | 328.325 | 3.007 | 3.01 |
| 14 | 327.250 | 1.933 | 4.94 |
| 15 | 327.825 | 2.507 | 7.45 |
LimitLine's sums match the ones NIST prints for all 20 groups, and both signal first at group 14. The lower sum never gets near h: the process did not shift down.
Note that this shift is large (about 4σ of the means), so a plain Shewhart chart with limits 325 ± 3 × 0.635 would already flag group 13 (328.3, above 326.9). The NIST example shows how the sums work; CUSUM's real advantage is for small shifts of 0.5σ to 1σ, where a Shewhart chart takes dozens to hundreds of points.
Choosing K and H
- k is half the shift you want to catch, in σ: k = 0.5 targets a 1σ shift.
- h sets the false-alarm rate. With k = 0.5, h = 4 gives an in-control average run length of about 170 points; h = 5 gives about 465 (Montgomery). Bigger h, fewer false alarms, slower detection.
- Use the σ of what you plot: for subgroup means, σ ÷ √n.
CUSUM, EWMA or a Shewhart Chart?
For shifts of about 1.5σ or more, a Shewhart chart with the Nelson rules is enough and easier to explain. For small, lasting shifts (0.5σ to 1σ), a CUSUM or an EWMA chart catches them much sooner; the CUSUM also tells you when the shift started (where the sum left zero).
Sources
- NIST/SEMATECH e-Handbook, §6.3.2.3 CUSUM Control Charts
- D. C. Montgomery, Introduction to Statistical Quality Control, Wiley
The LimitLine Newsletter
About one email a month: new guides, tool updates and one practical SPC tip. Unsubscribe in one click.