Learn · updated 2026-10-28
EWMA Chart: Catch a Small, Slow Drift
An I-MR chart reacts to each point on its own, so it catches big jumps fast but can miss a shift of half to one standard deviation for a long time. An EWMA (exponentially weighted moving average) chart plots a running average that remembers past points, so a small, steady drift builds up until it crosses a limit.
EWMAz(t) = λ·x(t) + (1 − λ)·z(t−1), starting at z(0) = targetSpread of zσ(z) = σ·√(λ ÷ (2 − λ))Limitstarget ± L·σ(z), usually L = 3
Worked Example: NIST's 20 Points
NIST's handbook uses a target of 50, a historical standard deviation of 2.0539 and λ = 0.3. Then σ(z) = 2.0539 × √(0.3 ÷ 1.7) = 0.8628, so the limits are 47.41 to 52.59.
| Point | x | EWMA z |
|---|---|---|
| 1 | 52.0 | 50.60 |
| 2 | 47.0 | 49.52 |
| 3 | 53.0 | 50.56 |
| 4 | 49.3 | 50.18 |
| 5 | 50.1 | 50.16 |
| 6 | 47.0 | 49.21 |
| 7 | 51.0 | 49.75 |
| 8 | 50.1 | 49.85 |
| 9 | 51.2 | 50.26 |
| 10 | 50.5 | 50.33 |
| 11 | 49.6 | 50.11 |
| 12 | 47.6 | 49.36 |
| 13 | 49.9 | 49.52 |
| 14 | 51.3 | 50.05 |
| 15 | 47.8 | 49.38 |
| 16 | 51.2 | 49.92 |
| 17 | 52.6 | 50.73 |
| 18 | 52.4 | 51.23 |
| 19 | 53.6 | 51.94 |
| 20 | 52.1 | 51.99 |
On an I-MR chart the same 20 points stay inside limits of 44.43 to 56.50, and no single point stands out: the drift shows up only as a pattern, which is what the EWMA is built to see.
Choosing λ and L
- Small λ (0.05 to 0.1) remembers more history and is best for very small shifts; it reacts slowly to a big jump.
- λ = 0.2 to 0.3 is a common middle ground. LimitLine's live line data uses λ = 0.2 and L = 3 to flag a drift.
- L = 3 is the usual limit width. Designs with a small λ often use a slightly narrower L, such as about 2.7 for λ = 0.1 in Lucas and Saccucci's tables.
Three Common Mistakes
- Using the wide, long-run limits from point 1. The spread of z grows from λσ toward its limit over the first points; the exact limits use σ·√(λ ÷ (2 − λ) · (1 − (1 − λ)^(2t))). The plain formula is fine once t is past about 10.
- Estimating σ from the data being judged. Take σ from a stable history (or the moving range of a stable period), or a drift inflates σ and hides itself.
- Replacing the I-MR chart. EWMA is slower on a large, sudden jump. Many plants run both: I-MR for big jumps, EWMA for slow drift.
Sources
- NIST/SEMATECH e-Handbook of Statistical Methods, 6.3.2.4 “EWMA Control Charts”
- S. W. Roberts (1959), “Control Chart Tests Based on Geometric Moving Averages”, Technometrics 1(3)
- J. M. Lucas and M. S. Saccucci (1990), “Exponentially Weighted Moving Average Control Schemes”, Technometrics 32(1)
The LimitLine Newsletter
About one email a month: new guides, tool updates and one practical SPC tip. Unsubscribe in one click.