Is CUSUM simply a cumulative total?

No. A monitoring chart combines accumulated deviations with a reference and a decision rule; tabular versions also use an allowance and resetting at zero.

Is it always better than a Shewhart chart?

No. Relative performance depends on the change pattern and design. CUSUM is especially associated with sensitivity to small sustained shifts.

Can it detect AI model drift?

It can monitor a suitable metric, but a signal does not uniquely diagnose data drift, concept drift, or a software fault.

AI Glossary / Statistics and probability / Intermediate

Cusum chart

Definition

A CUSUM chart, short for cumulative sum control chart, monitors a process by accumulating evidence of departures from a reference level. It is particularly useful for detecting small sustained shifts that may be difficult to recognize from individual observations. A signal indicates that the process deserves investigation; it does not identify the cause of the change.

What it means

A process can drift slightly upward for several observations without producing one spectacular outlier. CUSUM retains information across observations, allowing those modest deviations to build into a signal.

There are several forms, including plots of raw cumulative deviations with decision boundaries and tabular recursions that track upward and downward shifts. This entry explains a common standardized tabular version for monitoring a mean. Other CUSUM designs exist for counts, proportions, likelihood ratios, and other statistics; their formulas and calibration should not be assumed identical.

How it works

Establish an in-control reference mean and an appropriate standard deviation for the statistic being monitored. Standardize each new observation or subgroup statistic to obtain z_t. For an upper-sided chart initialized at zero, calculate C_t+ = max(0, C_(t-1)+ + z_t - k), where k is a reference allowance.

Signal when C_t+ exceeds a chosen decision limit h. A lower-sided counterpart accumulates -z_t - k instead. The zero floor discards accumulated evidence in the opposite direction, allowing a new run of evidence to begin.

Choose k and h for a meaningful shift size and acceptable false-alarm and detection-delay behavior. Specify whether equality triggers a signal, how charts reset, and what action follows. Monitoring subgroup means requires their standard error, not automatically the standard deviation of individual observations.

Three practical examples

A manufacturing team watches for a persistent upward shift in average component dimensions before many items cross specification limits.
An AI operations team monitors appropriately standardized prediction residuals to detect sustained bias, while accounting for changing case mix and dependence.
A service team tracks a stable process metric after adjusting for known workload patterns, using an agreed investigation procedure when accumulated deviations become unusual.

Think of it this way

Think of a bucket receiving small amounts of evidence. Repeated upward deviations add evidence, while the allowance prevents every tiny fluctuation from accumulating indefinitely. Opposing evidence can empty the bucket, but its level cannot fall below zero. Crossing the marked line prompts inspection. The analogy describes the tabular upper-sided chart, not every cumulative-sum formulation.

Worked example

Consider an original standardized upper-sided example. Set C_0+ = 0, k = 0.5, and h = 2. Use the explicit rule “signal only when C_t+ > 2.” These settings demonstrate arithmetic and are not a calibrated monitoring recommendation.

Suppose successive standardized observations are 0.2,0.8,1.1,0.9,1.2,1.0. The running values are max(0,0+0.2-0.5) = 0; then 0+0.8-0.5 = 0.3; then 0.3+1.1-0.5 = 0.9.

The next values are 0.9+0.9-0.5 = 1.3; 1.3+1.2-0.5 = 2.0; and 2.0+1.0-0.5 = 2.5. The first signal occurs at observation six under the stated strict inequality. Under an equality-inclusive rule it would occur at five.

No individual observation exceeds three standard deviations, yet the repeated positive deviations accumulate. This does not prove a true mean shift occurred: even an unchanged process can eventually produce a signal. Its frequency must be assessed under the intended monitoring assumptions.

Illustrative case study

In this fictional case, Cedar Forecasting monitors signed errors from a demand model. The team wants to detect a persistent tendency to underpredict, so it defines the error direction and an upper-sided CUSUM before deployment.

A pilot chart signals during a seasonal promotion. Investigation shows that the monitoring baseline did not account for that workload change. The team revises the reference model, validates the monitoring procedure on representative historical periods, and keeps a documented escalation process.

The signal helped identify an issue in the monitoring assumptions; it did not establish that the forecasting model had suddenly deteriorated. This invented case illustrates why chart design and operational interpretation belong together.

Common misconceptions and mistakes

Control limits are not specification limits. Specifications describe acceptable products or performance; chart limits describe evidence against a reference process under a monitoring design.

Do not interpret h as a universal number of standard deviations for each observation. It is a threshold for the accumulated statistic. Likewise, a CUSUM signal is not automatically an ordinary single-test p-value.

Autocorrelation, seasonality, baseline estimation error, changing variance, and changing sample sizes can alter false-alarm behavior. Repeatedly resetting or retuning a chart after seeing inconvenient signals can conceal genuine changes. Define the reset and investigation policy in advance and record deviations from it.

Business use

CUSUM charts can provide earlier attention to sustained quality or performance changes, especially where a small shift has material cumulative cost. The business value depends on what happens after an alert: confirming the data, investigating plausible causes, and deciding whether intervention is warranted.

Evaluate both false alarms and delay in detecting meaningful changes. Average run length summarizes the expected number of observations until a signal under a specified process; it is not a guaranteed interval between alarms. Account for monitoring frequency and the burden of investigating many simultaneous charts when translating statistical settings into an operational policy.

Check your understanding

Check your understanding: Why is the first accumulated value zero?

Answer: The deviation minus the allowance is negative, and the recursion floors it at zero.

Why does a value of two not signal in the example?

Answer: The stated rule requires exceeding two.

Does a signal establish causation?

Answer: No; it initiates investigation under a specified monitoring model.

Related terms

Statistical process control
Control chart
Shewhart chart
EWMA
average run length
change detection
prediction residual
process stability

Learn next

Compare CUSUM and Shewhart responses to simulated abrupt and gradual mean changes. Then assess false-alarm behavior under autocorrelated observations before choosing limits for a real monitoring process.

Sources

Sources: NIST/SEMATECH e-Handbook, CUSUM Control Charts, https://www.itl.nist.gov/div898/handbook/pmc/section3/pmc323.htm, and CUSUM Average Run Length, https://www.itl.nist.gov/div898/handbook/pmc/section3/pmc3231.htm. The standardized recursion and inequality convention are stated explicitly; the sequence and fictional case are original teaching examples.

Previous: Curvilinear relationship Next: Cut detection