ENH/BUG: Period and PeriodIndex ops supports timedelta-like

[sinhrks]

Must be revisited after #7954 to remove _skip_if_not_numpy17_friendly

Closes #7740.

Allow Period and PeriodIndex add and sub to support timedelta-like. If period freq is offsets.Tick, offsets can be added if the result can have same freq. Otherwise, ValueError will be raised.

idx = pd.period_range('2014-07-01 09:00', periods=5, freq='H')
idx + pd.offsets.Hour(2)
# <class 'pandas.tseries.period.PeriodIndex'>
# [2014-07-01 11:00, ..., 2014-07-01 15:00]
# Length: 5, Freq: H

idx + datetime.timedelta(minutes=120)
# <class 'pandas.tseries.period.PeriodIndex'>
# [2014-07-01 11:00, ..., 2014-07-01 15:00]
# Length: 5, Freq: H

idx + np.timedelta64(7200, 's')
# <class 'pandas.tseries.period.PeriodIndex'>
# [2014-07-01 11:00, ..., 2014-07-01 15:00]
# Length: 5, Freq: H

idx + pd.offsets.Minute(5)
# ValueError: Input has different freq from PeriodIndex(freq=H)

If period freq isn't Tick, only the same offset can be added. Otherwise, ValueError will be raised.

idx = pd.period_range('2014-07', periods=5, freq='M')
idx + pd.offsets.MonthEnd(3)
# <class 'pandas.tseries.period.PeriodIndex'>
# [2014-10, ..., 2015-02]
# Length: 5, Freq: M
idx + pd.offsets.MonthBegin(3)
# ValueError: Input has different freq from PeriodIndex(freq=M)

[jreback]

hmm, does it make sense to push _add_delta (and DatetimeIndex._add_delta) to base into order to combine them (its currently marked as non-implemented in DatetimeIndexOpsMixin)?

[sinhrks]

I feel this gets more complicated, because each _add_delta handles its internal representation (UTC time in DatetimeIndex and ordinal in PeriodIndex).
https://github.com/pydata/pandas/blob/master/pandas/tseries/index.py#L615

[jreback]

ok, not a big deal (always like to have centralized stuff like this), but sometimes makes too complicated. if in the future you think can integrate then ok.

[jreback]

maybe give the example here, (and add to timeseries.rst as well)

[sinhrks]

Yeah, added docs.

[jreback]

thanks @sinhrks