Handling a CustomBusinessDay in time-based .rolling()

[chrisaycock]

Starting with 10 business days following Christmas Eve:

from pandas.tseries.offsets import CustomBusinessDay
from pandas.tseries.holiday import USFederalHolidayCalendar

days = CustomBusinessDay(calendar=USFederalHolidayCalendar())

df = pd.DataFrame({'value': np.arange(n)},
                  index=pd.date_range('2015-12-24', periods=n, freq=days))

I can compute the three-day sum of the values with just:

In [21]: df.rolling('3d').sum()
Out[21]:
            value
2015-12-24    0.0
2015-12-28    1.0
2015-12-29    3.0
2015-12-30    6.0
2015-12-31    9.0
2016-01-04    5.0
2016-01-05   11.0
2016-01-06   18.0
2016-01-07   21.0
2016-01-08   24.0

But this is purely in terms of Gregorian calendar days, not the business calendar days that I had created the DataFrame with.

I can easily compute

df.index - 3*days

though I get a performance warning:

Non-vectorized DateOffset being applied to Series or DatetimeIndex

But I can't just pass this offset directly:

In [23]: df.rolling(3*days).sum()
...
ValueError: <3 * CustomBusinessDays> is a non-fixed frequency

I would like to be able to handle a CustomBusinessDay in .rolling(). (Because the DataFrame may come from any source, it would be easier to just pass the DateOffset object to .rolling() instead of using anything specific to the DataFrame's index.)

I know that the freq parameter was deprecated in 0.18, though 0.19 kinda brings this back in window. Is there an intrinsic reason window can't handle a CustomBusinessDay?

[chris-b1]

You could probably build a second version of this class that handled an arbitrary offset object rather than a fixed window, although it wouldn't be performant, as there would have to be a ton of calls back into python space. Unless there's some trick I'm not thinking of.
https://github.com/pydata/pandas/blob/master/pandas/window.pyx#L252

There was an issue about cythonized offsets, #11214, and I started a proof of concept, but never went any further with it.
https://github.com/chris-b1/pandas/tree/cythonize-offset

[agraboso]

Is there an intrinsic reason window can't handle a CustomBusinessDay?

As of now, because offset-based rolling windows are based on converting to nanoseconds (as we are discussing in #13968), which non-fixed frequencies do not support — the ValueError is coming from here.

This reaffirms my impression that the whole architecture around DateOffset is somewhat broken: we are calling by the same name things that are fundamentally different — regularity and anchoring being the critical concepts. I think this is something to think about long and hard for pandas 2.0

[chris-b1]

Actually, this isn't as bad as I was thinking. It's be almost like resampling, the one expensive op you'd have to do is generate the range of anchored dates.
rng = pd.date_range(df.index.min(), df.index.max(), freq=day)

Then you'd just need a cython function that generates window edges from that range and the frame's index.

[sinhrks]

xref #12351, #12377.

[jreback]

this is actually pretty easy. the point of build as @chris-b1 indicated. IS to create the indexers to the window, which by-definition are generally variable (they don't have to be though as that can be done more efficiently by a fixed offset window).

Once you have this window indexer (and this is pretty cheap to create), then rolling is the same. The current logic assumes a fixed differential indexer (e.g. it has a constant period regardless of the point in time).