Impliment BusinessWindowIndexer for non-fixed offsets

[mroeschke]

• closes Add test/example for rolling mean for BusinessDay #19
• tests added / passed
• passes black pandas
• passes git diff upstream/master -u -- "*.py" | flake8 --diff

[jreback]

so we ignore the arg when using a custom indexer? or should it still be passed (and can be ignored if needed inside the custom indexer?

imagine writing VariableIWindowIndexer and FixedWindowIndwxer as a BaseIndexer (which i think we should do at some point)

[mroeschke]

All custom indexers should subclass BaseIndexer so self.window should be the custom indexer which will accept arguments from mean (_apply).

VariableWindowIndexer and FixedWindowIndexer already subclass BaseIndexer

[mroeschke]

Question, @DiegoAlbertoTorres and @jreback, on the definition of a rolling business day.

Given the example below:

In [3]: from pandas import Series, date_range

In [4]: from pandas.core.window.indexers import BusinessWindowIndexer

In [5]: index = date_range("2019-09-06", "2019-09-13", freq="D")

In [12]: df = pd.DataFrame(range(len(index)), index=index)

In [14]: df['day_names'] = df.index.day_name()

In [15]: df
Out[15]:
            0  day_names
2019-09-06  0     Friday
2019-09-07  1   Saturday
2019-09-08  2     Sunday
2019-09-09  3     Monday
2019-09-10  4    Tuesday
2019-09-11  5  Wednesday
2019-09-12  6   Thursday
2019-09-13  7     Friday

Are values that fall on non-business days aggregated in the window? (Currently they are)

In [16]: df[0].rolling('D').mean()
Out[16]:
2019-09-06    0.0
2019-09-07    1.0
2019-09-08    2.0
2019-09-09    3.0
2019-09-10    4.0
2019-09-11    5.0
2019-09-12    6.0
2019-09-13    7.0
Freq: D, Name: 0, dtype: float64


# 2019-09-09 window = (Friday -> Monday] = Saturday + Sunday + Monday = (3 + 2 + 1) / 3 = 2
# 2019-09-08 window = (Friday -> Sunday] = Saturday + Sunday = (2 + 1) / 2 = 1.5
In [17]: df[0].rolling(BusinessWindowIndexer(index=s.index, offset="B")).mean()
Out[17]:
2019-09-06    0.0
2019-09-07    1.0
2019-09-08    1.5
2019-09-09    2.0
2019-09-10    4.0
2019-09-11    5.0
2019-09-12    6.0
2019-09-13    7.0
Freq: D, Name: 0, dtype: float64

[mroeschke]

Will document this in the POC doc after this PR.

[mroeschke]

Closing this example. This example requires a non trivial reworking of the internals for a feature that can be done today.