Series[Period] +/- NaT returns Series[datetime64]

[jbrockmendel]

per = pd.Period('2016Q1')
ser = pd.Series([per])

>>> ser - pd.NaT
0   NaT
dtype: datetime64[ns]

>>> ser + pd.NaT
0   NaT
dtype: datetime64[ns]

[jreback]

well , the + is right.

[makbigc]

Now, the behaviour is:

In [2]: per = pd.Period('2016Q1')

In [3]: ser = pd.Series([per])

In [4]: ser - pd.NaT
Out[4]: 
0   NaT
dtype: timedelta64[ns]

In [5]: ser + pd.NaT
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-5-b4d602b00491> in <module>()
----> 1 ser + pd.NaT

~/Code/pandas/pandas/core/ops.py in wrapper(left, right)
   1580                 (is_extension_array_dtype(right) and not is_scalar(right))):
   1581             # GH#22378 disallow scalar to exclude e.g. "category", "Int64"
-> 1582             return dispatch_to_extension_op(op, left, right)
   1583 
   1584         elif is_timedelta64_dtype(left):

~/Code/pandas/pandas/core/ops.py in dispatch_to_extension_op(op, left, right)
   1240         new_right = right
   1241 
-> 1242     res_values = op(new_left, new_right)
   1243     res_name = get_op_result_name(left, right)
   1244 

~/Code/pandas/pandas/core/arrays/datetimelike.py in __add__(self, other)
   1183         # scalar others
   1184         elif other is NaT:
-> 1185             result = self._add_nat()
   1186         elif isinstance(other, (Tick, timedelta, np.timedelta64)):
   1187             result = self._add_delta(other)

~/Code/pandas/pandas/core/arrays/datetimelike.py in _add_nat(self)
   1013             raise TypeError('Cannot add {cls} and {typ}'
   1014                             .format(cls=type(self).__name__,
-> 1015                                     typ=type(NaT).__name__))
   1016 
   1017         # GH#19124 pd.NaT is treated like a timedelta for both timedelta

TypeError: Cannot add PeriodArray and NaTType

Should the TypeError be thrown in the case of subtraction (func _sub_nat)?

[jreback]

this should be similar to how we treat a -/+ with a Timestamp (i think we raise)?

[jbrockmendel]

The "correct" way to solve this would be to implement #24983.

The issue is in how we treat pd.NaT. In most cases, we treat it is a datetime. In a few cases we treat it like a timedelta. The rule seems to be "1) if it can be treated as a datetime without raising, do that. 2) otherwise treat it as a timedelta" but I don't think we're consistent about part 2 there (and am not sure we should be)