fix wrong time-series alignment and add WA for StatsForcastAutoARIMA

[bachng2017]

a proposed fix for time-series alignment error and issue of StatsForcastAutoARIMA

1. see [Bug]: inconsistent result for timeseries mindsdb#3234 for detail about the alignment issue
2. see https://mindsdbcommunity.slack.com/archives/C037482KJ22/p1665911392587569 for more detail about the issue of StatsForcastAutoARIMA

the proposed fix is tested with house-sale tutorial only. It might need more review to consider the impact of the change

[paxcema]

Thanks for the PR :)

The workaround is okay. There are two other modifications that I'm not 100% convinced on yet. I will manually test a bit more, but if you have any proof it would be awesome to translate that into a test that captures the failing behavior, so that we can catch this next time that the alignment fails. If you can't, then give me some time and I will 1) merge this and 2) add tests on top.

[paxcema]

Update: the test in #1023 leads me to believe the general alignment behavior is correct (StatsForecast bug notwithstanding), and unless we have a test to showcase an edge case that justifies the above changes, I'm inclined to not merge those. I will try to find such a test, but in the meantime @bachng2017, it does seem like the issue is in mindsdb rather than lightwood, so I will check that now.

[bachng2017]

thanks @paxcema
did you check the conversation at https://mindsdbcommunity.slack.com/archives/C037482KJ22/p1665911392587569
from that it looks like an issue of StatsForcastAutoARIMA for us

[paxcema]

For now, can you please remove the changes in offset indexes?

Just leave the StatsForecast workaround, and I will merge immediately. We can debug the second problem in a separate PR. Thanks!

[bachng2017]

yes we've just left the StatsForcastAutoARIMA wa there and remove other changes

[paxcema]

Can you try to merge again please? Somehow GitHub thinks the newly added RandomForest mixer will be added again, which should not be the case.

[bachng2017]

tried to merge again, not sure is that OK or not

[paxcema]

Yes, merge is now fine 👍

There is a failing test, however. I will have to take a closer look on why before merging.

[paxcema]

Ah, it fails when min_offset = -np.inf, so this is not enough. I'm looking at alternative solutions.

EDIT: Found a way. I will merge this anyway because it has certainly been very helpful, and I will overwrite on top with another PR. Thanks @bachng2017!

[bachng2017]

thanks, the merge also very helpful for us.
We will do our own test (manual) with the latest and try to submit other issue (because we think the index issue which we've removed from this PR still be there)