Lorenzo Stella

What's interesting is that inverting the order of plotting also solves the issue: ```python forecast_entry.plot(prediction_intervals=(50, 90), color="g") ts_entry[-prediction_length * 4:].plot(ax=ax) ``` 

Hey @YuntianYeAWS there appear to be several modules in this PR, are you sure that all of them are supposed to be included? For example, the `AutoEstimator` folder seems weird,...

@YuntianYeAWS could you also expand a bit on the PR description? Especially when changes are large (many lines or modules affected) it is good to have some guidance on what...

@kashif let's first revert the change, then we see separately. Thinking about it, I'm not sure `aggregate_valid` should even be there: what is it supposed to do? Aggregate only metric...

> Locally this results in slower tests, I want to see how this impacts run-time in our actions. Any luck?

@dmitra79 I can confirm that the issue was solved in the `0.10.x` releases: all information about frequency comes now directly from the data (see #1997) and any mis-configuration of the...

This is potentially very breaking for anyone who might have custom models on the side. Maybe a smoother transition that keeps the `loss` method in the distribution classes, but deprecates...

> I find this refactoring useful but one request I would have is to also move the `sample()` method to the distribution output. This would make it easy to implement...

I'm not sure I agree with this change, because I believe it doesn't make it easier to remember what-to-import-from-where: now there would be `gluonts.dataset` and `gluonts.lab.datasets` which look similar. One...

@rsnirwan thanks for the detailed discussion of all these options! One note on the side: in general, we may not need to explicitly specify `cat` vs `real` to distinguish categorical...