DatetimeIndex + Series[timedelta64] returns DatetimeIndex

[jbrockmendel]

This is listed in the roundup #18824 but the behavior has changed since the older issue.

dti = pd.date_range('2017-01-01', periods=2)
ser = pd.Series(pd.TimedeltaIndex(['1 Day', '1Day']))
res1 = dti + ser
res2 = ser + dti
res3 = dti - ser
res4 = -ser + dti

>>> res1
DatetimeIndex(['2017-01-02', '2017-01-03'], dtype='datetime64[ns]', freq=None)
>>> res2
0   2017-01-02
1   2017-01-03
dtype: datetime64[ns]
>>> res3
DatetimeIndex(['2016-12-31', '2017-01-01'], dtype='datetime64[ns]', freq=None)
>>> res4
0   2016-12-31
1   2017-01-01
dtype: datetime64[ns]

[jbrockmendel]

I think there's an easy fix to this in indexes.datetimelike dropping a if isinstance(other, (ABCSeries, ABCDataFrame)): return NotImplemented right at the top of __add__ and __sub__.

[jreback]

seems a bit inconsistent.

[jbrockmendel]

Same deal for TimedeltaIndex * Series[int]. In fact this is specifically tested (it can't be intentional can it?)

        idx = TimedeltaIndex(np.arange(5, dtype='int64'))
        didx = TimedeltaIndex(np.arange(5, dtype='int64') ** 2)
        [...]
        result = idx * Series(np.arange(5, dtype='int64'))
        tm.assert_index_equal(result, didx)

[jbrockmendel]

The OP issue is closed by #18884. The TimedeltaIndex * Series[int] issue mentioned above is still outstanding. Update Just opened #19042 for the multiplication issue.