Change-Point Detection (PELT)
Also known as: Structural Break Detection, Breakpoint Analysis, Regime Change Detection, Değişim Noktası Tespiti
Change-Point Detection identifies time points at which the statistical properties of a sequence — such as mean, variance, or distribution — shift abruptly. The Pruned Exact Linear Time (PELT) algorithm, introduced by Killick, Fearnhead, and Eckley (2012), solves the penalized segmentation problem exactly while achieving linear expected computational cost, making it practical for long time series encountered in genomics, finance, climatology, and signal processing.
Read the full method
Sign in with a free account to read this section.
Method map
The neighbourhood of related methods — select a node to explore.
When to use it
Use PELT when you have a univariate or multivariate time series and need to locate abrupt shifts in mean, variance, or a general distributional parameter. The method assumes within-segment stationarity and requires specifying a cost function matched to the data's distributional family (Gaussian, Poisson, etc.) and a penalty value. It is not suited for gradual drift or smooth transitions; for those, consider state-space models or locally weighted regression. When changepoints are known to be sparse and instantaneous, PELT offers exact, scalable detection.
Strengths & limitations
- Exact optimization: PELT finds the globally optimal segmentation under the penalized criterion, unlike heuristic methods such as binary segmentation.
- Linear expected time complexity O(n): practical for time series with millions of observations when changepoints are not excessively dense.
- Flexible cost functions: supports Gaussian, Poisson, exponential, and other families, as well as nonparametric alternatives.
- Well-grounded penalty selection: AIC, BIC, and MBIC penalties provide principled, interpretable control over model complexity.
- Penalty sensitivity: results can be highly sensitive to the choice of beta; mis-specified penalties lead to over- or under-segmentation.
- Within-segment stationarity assumption: the method assumes each segment is internally homogeneous; gradual trends within segments can produce spurious changepoints.
- Distributional family must be pre-specified: incorrectly assumed cost functions reduce detection accuracy.
- Worst-case O(n^2) complexity: if changepoints are very dense, the pruning condition rarely triggers and computational savings diminish.
Frequently asked
How do I choose the penalty parameter beta?
The most common data-driven choices are BIC (beta = log(n) * p, where p is the number of model parameters per segment) and the MBIC variant recommended by Zhang and Siegmund (2007) for sparse signals. In practice, inspecting a penalty-versus-number-of-changepoints diagnostic plot — sometimes called an elbow plot — and cross-validating on held-out segments can help calibrate beta for a specific dataset and cost function.
Is PELT suitable for multivariate time series?
Yes. PELT can be applied to multivariate data by using a multivariate cost function, such as the negative log-likelihood of a multivariate Gaussian, which jointly models all dimensions. When the number of variables is large relative to the segment length, regularized covariance estimators or dimension-reduction preprocessing should be applied before running PELT to avoid rank-deficiency issues in the cost computation.
How does PELT differ from binary segmentation?
Binary segmentation greedily splits the series in two at the most significant point, then recurses on each subseries, yielding an approximate solution with O(n log n) cost. PELT instead solves the global penalized problem exactly, guaranteeing the best segmentation under the chosen criterion. This exactness matters when changepoints interact or when near-optimal greedy splits differ from the globally optimal partition.
Sources
- Killick, R., Fearnhead, P., & Eckley, I. A. (2012). Optimal detection of changepoints with a linear computational cost. Journal of the American Statistical Association, 107(500), 1590–1598. DOI: 10.1080/01621459.2012.737745 ↗
How to cite this page
ScholarGate. (2026, June 2). Change-Point Detection (PELT). ScholarGate. https://scholargate.app/en/statistics/change-point-detection
Which method?
Set this method beside its closest kin and read them side by side — the library lays the books on the table; the choice is yours.
- CUSUM ChartStatistics↔ compare
- Sequential AnalysisStatistics↔ compare