EWMA and CUSUM charts
Charts that remember the past to catch small, sustained shifts much faster than Shewhart charts.
A Shewhart chart judges each point on its own, so it is slow to notice a shift of about 1 sigma or less. EWMA and CUSUM charts accumulate information over time.
EWMA (exponentially weighted moving average)
Each plotted value is a weighted average of the newest reading and the previous EWMA value.
Smaller λ (0.05 to 0.2) gives more memory and catches smaller shifts. A common choice is λ = 0.2 with L ≈ 3 (L = 2.86 gives the same in-control ARL as a 3-sigma chart). EWMA is also robust to non-normal individual data.
CUSUM (cumulative sum)
Keeps running sums of deviations above and below the target, reset at zero.
A signal occurs when either sum exceeds H. Typical settings: K = (half the shift you want to detect) and H = or . The point where the sum started rising estimates when the shift began.
Average run length (ARL)
- ARL₀: average number of points before a false alarm in a stable process (about 370 for a 3-sigma Shewhart chart).
- ARL₁: average number of points to detect a real shift. For a 1-sigma shift, a Shewhart individuals chart needs about 44 points; a CUSUM with K = 0.5, H = 5 needs about 10.
Questions and answers
Should I replace my Xbar chart with EWMA?
Use EWMA or CUSUM where small, lasting shifts matter (chemical processes, precision machining). Keep a Shewhart chart, or Shewhart-style limits, to catch large sudden changes quickly.
