Probabilistic time series returning results above thresholds (GluonTS)

Question

I have a dataset with an upper numerical limit of 50, none of the samples go above this figure or below zero.

After running a deepAREstimator in GluonTS I get predictions way above 50 and also in the minuses. How can/could I fix this?

Answer 1

I would probably try to train the model slightly longer, and if possible on a larger dataset. To prevent getting negative values just try a different distribution: distr_output=gluonts.distribution.NegativeBinomial

But other than that its hard to tell without additional information what the problem is.