Change Point Detection calculator
Find where the mean of a sequence changed, with exact PELT segmentation or a Bayesian run-length posterior that gives each change a probability and a credible interval.
What it does
Finds where the mean of a sequence shifted, by choosing the segmentation that best trades fit against the number of splits. The answer is the exact optimum for the penalty you set — not a scan or a heuristic.
When to use it
You have a sequence and want to know when its level changed and by how much — a policy change, a process change, a break in a time series.
Alternatives
- the Bayesian method — you want a probability and a credible interval for each change rather than one best answer
- CUSUM — you want an alarm as data arrives rather than a retrospective segmentation
- Regression with a spline — the change is a gradual curve rather than a step
How to read the output
- The step function
- The fitted mean of each segment, drawn over the data it was inferred from so you can judge it against the observations. The model only knows about steps in the mean. A gradual trend gets approximated by a staircase, which is not evidence of abrupt changes.
- The change point
- Reported between two observations: the last of the old segment and the first of the new one. The location is an estimate. Quoting a segment mean with a confidence interval as though the split were known in advance overstates the precision.
- The penalty
- How much evidence a split must earn to be worth making. The answer shown is the exact best segmentation FOR THAT PENALTY — there is no search or approximation involved. The penalty is a choice, not a p-value. A looser one finds more change points in the same data, so try more than one before believing any of them.
- The shift sizes
- Each change is reported as the before-and-after means, the difference, and that difference in units of σ. A change point can be statistically clear and practically trivial. Read the size, not just the fact that a line was drawn.
How this calculator is validated · Which statistical test should I use?