Significance Value for Change Point

[ErikWitt]

Hi there, I love the package!

I was wondering if there is any way to compute a significance for the the predicted change points. Certainly some change points are more obvious than others for the algorithm.

I would be great to know the certainty for each change point.

[oboulant]

Hi @ErikWitt ,

Thx for your interest in ruptures !

You can have a look at this page of the docs. Might not be exactly what you are looking for, but it gives some intuition on a possible approach if you want quantify the relative importance of change points.

Hope it helps !

[deepcharles]

Hi @ErikWitt, Unfortunately, computing a significance level requires some assumption on the data, and even in this situation, it remains complex. Within ruptures, what you can do is do change-point detection with a penalty (with Pelt or KernelCPD). The penalty will threshold out all change-points with a low amplitude shift. However, setting a relevant penalty level can only be done manually. Outside of ruptures, they are several articles that you can look at:

• Post-selection inference (mostly valid for Gaussian signals) https://www.stat.cmu.edu/~ryantibs/papers/binseginf.pdf
• computing a probabilty of being a change-point (https://doi.org/10.1007/s00180-013-0422-9). This can be quite general but the implementation is a bit difficult.

Hope this helps