sktime: [BUG] bug in `NaiveForecaster` variance formulae?

From https://github.com/alan-turing-institute/sktime/pull/3435/files which I’ve merged and only now noticed potential issues in the formula for predict_var.

The main issue (which I missed on first read) was that I think we need rolling forecasts, no?

  • should the mean strategy not have rolling mean predictions used, instead of the single prediction at the cutoff? So, y.rolling(window_length).mean instead of y_pred? With an appropriate shift per fh?
  • last seems correct only for fh indices that are smaller or equal to sp.
  • the slope strategy should use window_length instead of T, and also do this on rolling basis?

FYI @topher-lo

About this issue

  • Original URL
  • State: open
  • Created 2 years ago
  • Comments: 32 (15 by maintainers)

Commits related to this issue

Most upvoted comments

Hm. Is it documented somewhere what the R code assumes?

The answer might be simple, there is a bug in both pieces of code? We could ask one of its authors.

Either way, I also think we ought to write up somehwere what the methods do, or at least, what we think it should do. I feel the above discussion has made clear that a reference to a section of Hyndman is not precise enough, since there are too many gaps to fill in.

@fkiraly I just started a long writeup on this issue. Would you be available on Discord for a quick call right now?

Not right now (have to get to work), but the forecasting meet-up and the collab session will take place (see schedule in slack general channel), I’ll try to make it. If that doesn’t work, we can try later on the day.