Specifying start/end day and length of a target/precursor period

[BSchilperoort]

In @Peter9192's review of AI4S2S/s2spy#113, he voiced his desire to be able to specify target and precursor periods by the end date and length;

e.g. if I want to issue a forecast by the end of may, I want to start appending precursors from 31 may backwards.

However, this leads to some non-trivial issues like:

• How do we deal with leap years? As the periods are unaware (yet) of which year they will belong to, we do not know if it is a leap year or not. A "simple" solution is to just ignore leap days.

• How do we determine which year the date specifies?

  • E.g., if the anchor is June 1st, and you specify an end date of June 5th, does this mean 4 days after or 361 days before the anchor. Does the (default) behavior depend on target=True/target=False?
  • And what if someone wants to define a date two years ahead of the anchor?

And lastly, do we want to support start_date, end_date, and length, and have the user use 2/3 to define the period?

[geek-yang]

e.g. if I want to issue a forecast by the end of may, I want to start appending precursors from 31 may backwards.

I think most of the use cases with this kind of need will be solved by AdventCalendar. Only some edge cases would need CustomCalendar for backward counting, I assume. I think it is better to come up with new type of calendars to address these special needs rather than change the BaseCalendar to allow for more flexibility.

[Peter9192]

• A "simple" solution is to just ignore leap days.

agree, or don't allow it.

• E.g., if the anchor is June 1st, and you specify an end date of June 5th, does this mean 4 days after or 361 days before the anchor.

Since s2s usecases are usually interested in predicting (ahead of time), I think this would mean 361 days before.

• And what if someone wants to define a date two years ahead of the anchor?

In the very beginning I had the thought that you could allow this by setting the periodicity of the calendar. Back then we had the adventcalendar doing something like n_targets = 365 / target_length, and that could conceptually very easily be changed to a multiple of 365. The idea came from a visualization that Sem showed, something like this:

you could drag the start and end but it would always have this cyclic pattern. The length of this cycle can be changed though.