Is a break always an error?

No; it may result from a legitimate improvement in measurement.

Can the series still be useful?

Yes, with suitable interpretation or documented harmonization.

Is a statistical structural break identical to a metadata break?

No; one may describe a changed process, the other changed measurement standards.

AI Glossary / Statistics and probability / Intermediate

Time series breaks

Definition

Time series breaks are discontinuities in comparability caused by changes in how a variable is defined, classified, covered, measured, or compiled over time. In the official-statistics sense used here, a break may reflect a change in the measurement system rather than a change in the underlying phenomenon. Metadata are essential for interpreting it.

What it means

A chart can show a sudden increase because activity increased, because the measurement definition broadened, or because both happened. If the definition changes, adjacent values may no longer measure exactly the same thing.
This entry follows the OECD statistical-metadata meaning. In broader time-series modeling, a structural break can instead mean a change in the underlying process, such as a new mean or relationship. The two ideas can overlap, but a detected numerical change alone does not reveal its cause.

How it works

Review the series definition and production history. Relevant changes include new industry classifications, revised survey questions, expanded population coverage, altered reporting systems, and new estimation procedures. Record the date and scope of each change.
Distinguish a break from a routine data revision or a genuine event. A revision updates previously published values and may or may not improve comparability across a break. A business shock can change a stable measure without changing its definition.
Where possible, collect overlapping observations under both old and new methods. Parallel measurement in a bridge period helps estimate how much of the difference is methodological, though the relationship may not remain constant across all periods or subgroups.
Consider whether historical observations can be recalculated under the new definition. A carefully constructed back series may provide a more comparable history, but it requires sufficient data and explicit assumptions. Do not invent missing detail simply to make a chart continuous.
Alternative linking methods may use ratios, differences, or model-based adjustments. The appropriate choice depends on the type of measure and evidence about the relationship between methods. A proportional link is not interchangeable with an additive correction.
Retain both original and adjusted values with clear labels. Document bridge periods, formulas, uncertainty, and any periods where comparisons remain unsuitable. Users should be able to distinguish published observations from reconstructed estimates.
Mark breaks visibly in tables and charts, and explain them near reported growth rates. A small footnote detached from a prominent change claim may not provide enough context for readers or automated systems.
For AI forecasting, check whether training data mix measurement regimes. A model can learn an artificial jump as though it were a real trend or seasonality. Include justified regime information, use harmonized data, or restrict comparisons where appropriate.
Do not automatically smooth away every discontinuity. Real changes may be important signals. Investigate metadata and domain evidence alongside statistical diagnostics, and preserve unresolved uncertainty when the measurement and underlying process cannot be separated confidently.

Three practical examples

Customer analytics: flag a change from counting active accounts to counting active users.
Official statistics: explain a series affected by revised classifications or population coverage.
AI monitoring: distinguish an apparent quality shift from a change in benchmark composition or labeling policy.

Think of it this way

Think of tracking distance while changing from one ruler to another with a different scale. A jump in recorded values can come from the ruler rather than the object. To compare across the change, you need to know how the rulers relate, not merely draw a smooth line between readings.

Worked example

Consider an original illustrative series. In a bridge period, the old method gives a value of 100 and the new method gives 110 for the same underlying period. In the following period, the new method reports 121.
Comparing 121 directly with the old-method 100 gives apparent growth of 21%. Comparing like with like gives (121-110)/110 = 10% growth under the new method.
If a proportional linking assumption is justified, the bridge ratio is 110/100 = 1.10. Multiplying the old-method bridge value by 1.10 yields 110 on the new-method scale. The overall ratio 121/100 = 1.21 can be decomposed as 1.10 for the method link times 1.10 for subsequent growth.
This toy calculation does not establish that every historical value should be multiplied by 1.10. The relationship may vary over time or across components. Parallel measurement and substantive knowledge are needed before using a constant link for a reconstructed history.

Illustrative case study

Fictional case: Alder Metrics changes its definition of an AI-resolved support request to include cases closed after a suggested response. The dashboard shows an immediate improvement, which a manager attributes entirely to a new model. Analysts run the old and new definitions in parallel, identify a measurement contribution, and annotate the break. They report comparable-period changes separately from the definition change, preserving both historical versions for audit.

Common misconceptions and mistakes

Do not interpret every visible jump as a real-world change. Do not assume every break makes the entire series unusable. A linking factor estimated in one period may not apply everywhere. A model that detects a change point does not identify its cause. Smoothing a break without explanation can conceal methodological differences and create false precision.

Business use

Use break metadata to protect trend reports, forecasts, and performance claims. Assign ownership for metric definitions and record changes when systems or classifications change. Before comparing periods, confirm that the quantities are sufficiently comparable for the intended decision and communicate any adjustment transparently.

Check your understanding

What is comparable growth in the example?

Answer: Ten percent under the new method.

Why is the naive comparison 21%?

Answer: It combines a method-level change with subsequent growth.

Can the bridge ratio automatically adjust every past period?

Answer: No.

Related terms

Metadata
Structural break
Measurement error
change point
backcasting
linked series
revision policy
comparability

Learn next

Study bridge surveys and historical-series reconstruction, then compare methods for detecting genuine structural change. Practice annotating a chart so users can distinguish observed growth from methodological discontinuity.

Sources

OECD, Glossary of Statistical Terms, Time series breaks: https://www.oecd.org/content/dam/oecd/en/publications/reports/2008/09/oecd-glossary-of-statistical-terms_g1gh9ad7/9789264055087-en.pdf OECD, Data and Metadata Presentation Handbook: https://unstats.un.org/unsd/statcom/doc05/DataPresentationHandbook-OECD.pdf The bridge-period calculation is original.

Previous: Time series analysis Next: Time Series Forecasting