Change Point Detection (Binary Segmentation)

Change point detection by binary segmentation repeatedly splits a series at the point that most reduces the within-segment cost, stopping when no split's gain exceeds a penalty such as BIC.

이진 분할 변화점 탐지는 구간 내 비용을 가장 크게 줄이는 지점에서 시계열을 반복해 나누고, 어떤 분할의 이득도 BIC 같은 벌점을 넘지 못하면 멈추는 방법이다.

When did the average level of my process shift, and how many shifts were there?

공정의 평균 수준이 언제, 몇 번 바뀌었는지 찾고 싶어요.

Binary segmentation searches the whole series for the single best split in mean, then searches each resulting segment, accepting splits only while the cost reduction exceeds the penalty. The result is a set of change points and the mean of each segment.

When to use Change Point Detection (Binary Segmentation)

• You do not know in advance when or how many level shifts occurred.
• You want an automated, fast segmentation of a long series.
• Shifts in the mean are the main feature of interest.

When to use something else

• You have one hypothesised break date; use the Chow structural break test.
• You need the exact optimum for closely spaced or short segments; use optimal partitioning or PELT.

Assumptions and requirements

• Independent Gaussian noise around piecewise-constant means.
• Constant noise variance across segments.
• A minimum segment length (default 5) and at most 10 change points.

What the analysis reports

• Number and timing of change points
• Mean and SD of each segment
• Penalty value and noise scale used

How to report it

Binary segmentation with a BIC penalty detected [value] change points, at [value] and [value]; segment means were [value], [value] and [value].

Related methods

• Structural Break Test (Chow Test)
• Mann-Kendall Trend Test with Sen's Slope

Run Change Point Detection (Binary Segmentation) on Quorant