Must segments have constant means?

No; that is one possible model.

Does segmentation require knowing the number of changes?

No; penalized methods can select it.

Can a smooth trend create apparent segments?

Yes, particularly if a piecewise-constant model approximates it.

AI Glossary / Statistics and probability / Intermediate

Time-series segmentation

Definition

Time-series segmentation divides an ordered sequence into contiguous intervals that are relatively coherent under a chosen criterion. The boundaries may reflect changes in mean, variance, trend, dependence, or another modeled feature. It is often used to summarize behavior, detect regimes, or prepare data for further analysis while preserving temporal order.

What it means

A long sequence may contain several distinct phases rather than one stable process. Segmentation seeks boundaries between those phases so that each part can be described more simply or consistently.
What counts as a coherent segment depends on the task. A method targeting changes in average level may miss changes in variability or frequency content. The returned boundaries are therefore results of a model and objective, not automatically objectively correct event times or causal explanations.

How it works

Define the purpose of segmentation. It may support retrospective description, online monitoring, compression, classification, or forecasting. A method that can inspect the entire recorded series has different information from one that must make decisions as observations arrive.
Choose the features and within-segment model. A piecewise-constant mean model summarizes each segment by its average. Other models can represent linear trends, changing covariance, or distributional differences. Match the model to the type of change that matters.
Specify a cost for fitting each segment. For a mean-change model, a common cost is the sum of squared deviations from the segment mean. Lower total fitting cost indicates a better fit to the observed values under that criterion.
Control complexity. Allowing every observation to form its own segment can make fitting error vanish without producing a useful explanation. A fixed number of boundaries, a penalty per boundary, or a minimum segment length constrains the solution.
Select a search algorithm. Dynamic programming can optimize certain objectives exactly, while methods such as binary segmentation or sliding-window searches make different computational trade-offs. PELT uses pruning to accelerate penalized segmentation under its conditions.
Preserve the time structure and relevant preprocessing information. Irregular spacing, missing observations, seasonal patterns, and correlated errors can affect the interpretation of costs and boundaries. Removing seasonality may be justified for one question and erase useful structure for another.
Validate the chosen settings. Excessively weak penalties produce many small segments; excessively strong penalties can hide meaningful changes. Use domain evidence, simulation, appropriate validation data, and sensitivity analysis rather than choosing parameters solely for an attractive chart.
Distinguish boundary detection from explanation. A change in the measured series could reflect a real regime shift, a sensor recalibration, a revised definition, or a data-processing issue. Examine metadata and context before attaching an operational label.
For predictive use, avoid using future observations to construct features that would not have been available at prediction time. Offline segmentation can be valuable for historical analysis while being unsuitable as a direct simulation of live detection performance.

Three practical examples

Operations monitoring: identify periods with different average processing times.
Sensor analysis: summarize a recording into regimes with distinct signal behavior.
AI workflow analysis: locate changes in model performance while checking whether evaluation data or labeling rules changed.

Think of it this way

Think of dividing a recorded journey into stretches of similar travel behavior, such as steady movement and stop-start traffic. The boundaries depend on whether you care about speed, variability, or route changes. A useful division for one purpose may not be the best division for another.

Worked example

Consider an original six-point series: 1, 1, 2, 8, 9, 9. Use a piecewise-constant mean model with squared-error cost. With no boundary, the mean is 5 and the cost is 16 + 16 + 9 + 9 + 16 + 16 = 82.
Place a boundary after the third observation. The first segment has mean 4/3 and squared-error cost 2/3. The second has mean 26/3 and the same cost, 2/3. Total fitting cost is therefore 4/3, approximately 1.333. Checking all five possible single boundaries confirms that this is the lowest single-boundary cost.
If the objective adds a penalty of 10 for each boundary, the two-segment objective is 4/3 + 10, approximately 11.333, compared with 82 for no boundary. Under a rule allowing at most one boundary, the split is preferred.
With a penalty of 100, its objective becomes approximately 101.333, so no boundary is preferred. The example shows how the penalty changes the selected complexity; it does not prove that a real-world event occurred between observations three and four.

Illustrative case study

Fictional case: Elm Monitoring segments a history of AI service latency and finds a sharp change. Engineers initially suspect a new model version, but the timestamp aligns with a logging change that began including queue time. They label the measurement break and reassess the series on a comparable definition. They also compare several reasonable penalties to see whether other boundaries persist. Segmentation helps focus investigation without determining the cause by itself.

Common misconceptions and mistakes

Do not confuse contiguous segmentation with clustering observations regardless of order. A lower fitting error alone does not justify more segments. A detected boundary is not necessarily a causal event. Offline results can use future information unavailable in live operation. Different cost functions detect different kinds of change, and no single penalty is universally appropriate.

Business use

Use segmentation to make long sequences interpretable and to identify periods needing investigation. Report the feature representation, cost, search method, penalty, and minimum segment length. Keep detected boundaries separate from confirmed operational events, and evaluate live detection separately when decisions must be made in real time.

Check your understanding

What is the no-boundary cost in the example?

Answer: Eighty-two.

Where is the best single split?

Answer: After the third observation.

Why does a larger penalty change the decision?

Answer: It charges more for additional model complexity.

Related terms

Structural break
Dynamic programming
Regularization
Time series
Change-point detection
piecewise model
PELT
online detection

Learn next

Compare mean-change and variance-change costs, then study offline and online change detection. Test sensitivity to penalties and minimum segment lengths before interpreting boundaries as meaningful regimes.

Sources

Truong, Oudre and Vayatis, ruptures: change point detection in Python: https://arxiv.org/abs/1801.00826 ruptures, PELT documentation: https://centre-borelli.github.io/ruptures-docs/code-reference/detection/pelt-reference/ ruptures, Window sliding segmentation: https://centre-borelli.github.io/ruptures-docs/user-guide/detection/window/ The six-point segmentation example is original.

Previous: Time series model Next: Time to answer