Bug: Faulty parameter estimates

[Tveten]

Hey,

I was very happy to find an online implementation of PELT, as it is very useful in a project I'm working on. I've encountered a bug with the output parameter estimates though. I think it should be a quick fix. Here's a MWE:

x <- c(rnorm(50), rnorm(50, 3), rnorm(50, -3))
res <- ocpt.mean.initialise(x, penalty = "BIC")

res@cpts
[1] 50 100 150
[email protected]
$mean
[1] -4.993291 69.187425 78.725460

I think I've narrowed the problem down to the param() method, lines 551-555. It seems to me like the mean is taken over the cumulatively summed observations rather than the observations themselves. (The same seems to be the case with param.var.)

I see that you haven't updated the package in some time, but I figured it was worth a shot.

Thanks.

[rkillick]

Hi @Tveten,

Thanks for the heads up. The package is currently under development as we found a bug following the GSOC submission (which necessitated making the code public). We will fix this alongside other aspects.
If you are interested in this package then you may also be interested in the package here:
https://github.com/grundy95/changepoint.forecast
This does online mean/var changepoint detection using forecast errors in a wide range of model structures.

[Tveten]

I see. Thanks a lot. The changepoint.forecast package looks interesting as well.