CUSUM¶

A sequential change-detection method that accumulates signed deviations from a reference value, resets or branches according to a declared rule, and signals when the cumulative evidence crosses a decision threshold.

Version: v1 · 2026-09-28 · History
Domain-specific #: 8818
Domain group: Formal Sciences
Origin domain: Experimental Design & Statistics
Subdomains: Statistical Process Control, Sequential Analysis → Experimental Design & Statistics
Aliases: Cumulative sum control chart, Cumulative sum scheme, Page CUSUM

Core Idea¶

CUSUM turns a stream of small signed departures into persistent evidence of change. At each observation, a recursion updates an upward or downward statistic relative to an in-control target and reference allowance, retaining directional evidence that single-point charts may overlook.

A signal occurs only when the accumulated statistic crosses a decision interval. The allowance, threshold, sidedness, initialization, and baseline distribution jointly determine average run length and detection delay; a crossing indicates change evidence, not its substantive cause.

How would you explain it like I'm…

Adding-Up Alarm

Imagine tasting soup every day and keeping a score. If it tastes a little too salty you add a point, and if it tastes fine you take a point away, but the score never goes below zero. One salty spoonful doesn't matter much, but if the points keep piling up past a line you drew, you shout "the recipe changed!" The score tells you something changed, not why.

Adding Up Little Changes

CUSUM is a way to notice a slow change that no single measurement shows. Each time you measure, you see how far it is above (or below) what it should be, give a little wiggle room for normal ups and downs, and add what is left to a running total. Normal wobbles keep the total near zero, but a small push in one direction keeps adding up. When the total crosses a set line, you raise an alarm. The alarm says "something changed," but you still have to go find out what caused it.

Cumulative Sum Change Detector

CUSUM (cumulative sum) is a monitoring method that turns many small, same-direction departures into strong evidence of a shift. At each new observation it updates a running statistic by adding how far the value sits above (or below) a target, minus an allowance for ordinary noise; the statistic cannot drop below zero, so it forgets old good news but remembers a steady drift. A signal fires only when the running statistic crosses a decision threshold. Compared with a chart that flags a single point far outside the limits, CUSUM is better at catching small persistent shifts that no individual point reveals. Choosing the allowance and threshold is a trade-off between false alarms and how quickly a real change is caught. A signal is evidence that the process changed, not an explanation of why.

CUSUM is a sequential change-detection procedure from statistical process control. For an upward shift, the statistic follows a recursion of the form S_t = max(0, S_{t-1} + (x_t − μ0 − k)), where μ0 is the in-control target and k is the reference allowance; a mirror-image statistic tracks downward shifts, and a two-sided scheme runs both. Because the statistic accumulates signed departures, it retains directional evidence that a Shewhart-style chart, which judges each point alone, tends to miss when the shift is small. An alarm is raised when S_t exceeds the decision interval h. The allowance, the threshold, whether one or both sides are monitored, the starting value, and the assumed in-control distribution jointly determine the average run length (how long until a false alarm) and the detection delay after a real shift. A crossing is evidence that the process departed from its in-control state; identifying the substantive cause is a separate investigation.

Structural Signature¶

Sig role-phrases:

Ordered observations — Provide the sequential process evidence. It is ordered observations. Counterfactual: Reordered or aggregated samples change the detector.
In-control reference — Defines the baseline mean or distribution. It is in control reference. Counterfactual: Without a baseline deviation has no sign.
Reference allowance — Tunes the shift size receiving greatest sensitivity. It is reference allowance. Counterfactual: Zero or arbitrary allowance changes detection behavior.
Cumulative recursion — Carries evidence forward and may reset at a holding barrier. It is cumulative recursion. Counterfactual: Independent thresholding loses small persistent shifts.
Decision threshold — Converts accumulated evidence into a signal. It is decision threshold. Counterfactual: The statistic alone provides no alarm rule.
Performance calibration — Quantifies run length, false alarms, delay, and initialization. It is performance calibration. Counterfactual: A threshold cannot be interpreted without operating characteristics.

What It Is Not¶

An ordinary running total is not CUSUM when it lacks an in-control reference and decision rule.
A Shewhart chart judges the current sample rather than accumulating small past deviations.
EWMA uses geometrically declining weights rather than the standard CUSUM recursion.
A threshold cannot be interpreted across processes without recalibrating variance, dependence, and false-alarm performance.
Closest near-miss. An EWMA discounts older evidence; CUSUM normally adds deviations recursively with its own reset or holding rule.

Scope of Application¶

Manufacturing control. Detects sustained mean shifts.
Clinical surveillance. Monitors risk-adjusted outcome sequences.
Reliability. Signals changes in failure behavior.
Online analytics. Finds distributional changes under calibrated assumptions.

Clarity¶

Give the ordered statistic, baseline estimate, standardization, reference allowance, one- or two-sided recursion, head start or initial state, threshold, and reset policy. Report in-control run-length performance and the shift size for which the design was tuned.

Manages Complexity¶

CUSUM compresses an entire ordered history into one or two state variables, gaining sensitivity to sustained small changes. That memory also propagates baseline error, autocorrelation, and transient disturbances, so calibration and post-signal diagnosis remain separate from the recursive calculation.

Abstract Reasoning¶

Choose the process statistic and establish an in-control reference.
Select one- or two-sided change alternatives and reference allowance.
Set recursion, initialization, and decision threshold.
Update in order and record the first threshold crossing.
Evaluate average run length and diagnose rather than treating a signal as a known cause.

Knowledge Transfer¶

CUSUM transfers among applications only after rescaling observations and recalibrating reference, shift size, dependence, and threshold. Any cumulative total is not a CUSUM; the invariant is sequential evidence for a specified change with a decision rule.

Examples¶

Canonical¶

A standardized one-sided CUSUM adds each measurement minus a reference allowance, resets negative values to zero, and signals when the statistic exceeds h.

Mapped back: input → ordered standardized values; reference → allowance k; recursion → max at zero; decision → threshold h.

Applied / In Practice¶

Plotting cumulative production volume shows a running total but is not CUSUM change detection because no in-control reference or alarm threshold is defined.

Mapped back: sum → present; baseline → absent; decision → absent; verdict → not CUSUM.

Structural Tensions¶

T1 — Sensitivity versus False Alarms. A lower threshold or smaller allowance detects shifts earlier while shortening in-control run length.

Diagnostic: Which shift size and false-alarm rate are being optimized?

T2 — Memory versus Baseline Drift. Accumulation reveals small shifts but can mislead if the nominal reference itself moves.

Diagnostic: Was the baseline stable and independently estimated?

Structural–Framed Character¶

CUSUM is structural as recursive accumulation plus thresholded change evidence and framed by sequential statistics. Its identity depends on ordered updating against a reference, not on the visual appearance of a cumulative plot.

Structural Core vs. Domain Accent¶

The reusable core is stateful evidence accumulation. Statistical process control supplies distributional standardization, target shift, average run length, and alarm calibration; applying the recurrence elsewhere requires rebuilding those performance meanings rather than copying constants.

This entry is a kind of Sequential analysis.

Approved unparented root. No reviewed parent entails the CUSUM combination of referenced increments, directional recursion, and calibrated decision threshold.
Related — Shewhart chart, EWMA, and SPRT. Each performs sequential monitoring, but their memory and decision structures are not interchangeable.

Relationships to Other Abstractions¶

Current abstraction CUSUM Domain-specific

Parents (1) — more general patterns this builds on

CUSUM is a kind of Sequential analysis Domain-specific

CUSUM is Sequential Analysis that accumulates signed deviations and signals when cumulative evidence crosses a decision threshold.

Hierarchy path (1) — routes to 1 parentless root

CUSUM → Sequential analysis → Threshold

Neighborhood in Abstraction Space¶

CUSUM sits in a crowded region of the domain-specific corpus (31^st percentile for distinctiveness): several abstractions share nearly its structure, so a description that fits it tends to fit its neighbors too.

Family — Statistical Hypothesis Tests & Diagnostics (9 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-08

Not to Be Confused With¶

Shewhart chart. Tell: Signals on individual sample statistics rather than cumulative small deviations.
EWMA chart. Tell: Exponentially discounts older observations.
SPRT. Tell: Uses likelihood-ratio boundaries under explicit hypotheses and is not identical to every CUSUM.
Cumulative total. Tell: May have no reference, reset, or alarm semantics.

References¶

Frozen Wikipedia discovery revision: https://en.wikipedia.org/wiki/CUSUM (revision 1328039568).
Preserved source candidate: http://www.irisa.fr/sisthem/kniga/
Preserved source candidate: http://www.phmsociety.org/sites/phmsociety.org/files/phm_submission/2015/ijphm_15_015.pdf
Preserved source candidate: http://www.itl.nist.gov/div898/handbook/pmc/section3/pmc323.htm

The frozen Wikipedia revision is discovery provenance. The retained source set was reviewed for identity, formal or operational relation, and scope. The encyclopedia's structural synthesis is bounded to those claims; a thin authority surface is recorded as a nonblocking source-strengthening repair rather than concealed.