[FEATURE] Non-continuous index in predict about hcrystalball HOT 4 CLOSED

@pavelkrizek Could you provide your opinion on the feature? The result of our discussion will determine the next steps for HCrystalBallForecaster in sktime implemented in this PR

from hcrystalball.

@MichalChromcak Is this necessary for sktime wrapper? In principle, somebody who cares just about some particular horizon could achieve the same thing by just passing custom metrics to CV, which will i.e. exponentially weigh the error terms based on time - this way gets the best model just for a particular horizon and they could filter the result themselves.

I also see bigger changes needed than just filtering out the index in the predict method. In order to make it useful, we would also need to change things in cross-validation, i.e. our Splitter is not supporting right now returning just one sample with the arbitrary horizon (horizon 10, gives you 10 samples for predict), also our CV plotting will not be probably very meaningful with these settings, so having a better picture of how many things needs to be changed to make it usable would be a better starting point (would propose to change 1 wrapper and try to use it in different contexts to see what breaks...). Some scouting of methods that are already supporting this out-of-box would be also useful to see if there are some performance benefits of passing the filtered data to predict or just post-filtering in order to decide if one uniform solution "fits all".

from hcrystalball.

@pavelkrizek At least from the test suite, there are some cases when data is passed with indices including gaps. We might also raise NotImplementedError in there, or maybe better directly in HCrystalBall if such things occur.

@mloning Would be acceptable to raise NotImplementedError in such case for sktime?

from hcrystalball.

Hi @MichalChromcak, yes that would be okay, we may have ignore any failing unit test in sktime for the HCrystalball wrapper.

from hcrystalball.