task

[m-muecke]

• Most likely it will just have an own task, which should basically be a subclass of the regression and classifcation task.

• Start with the regression task.

• The backend should inherit from the data.table backend and add a time and index column. Don't want to deal with the shit date implementation in R, i.e. I want frequency of year month, etc. various package handle this different:

• https://github.com/eddelbuettel/dtts data.table + nanotime

• https://tsibble.tidyverts.org tibble + vctrs with own vector type

I think the best implementation would be to try out the clock package since it behaves like lubridate and handles all the frequency I want. Check first how compatible it is with data.table, but I assume shouldn't be such an issue. Another challenge will be how the transformations of the target and features are handled with for example rolling mean, etc. data.table still has some open PRs for the rolling functions, check if they would work with clock.

[m-muecke]

clock doesn't work with data.table, great ...
https://clock.r-lib.org/articles/faq.html#why-doesnt-this-work-with-data-table

lets add frequency as an argument like functime and nixtla do then transformations can be inferred for tsibble (fable forecasting), check what benefits this brings