BUG: Series[td64].sum() wrong on empty series, closes pandas-dev#37151

Author: jbrockmendel

doc/source/whatsnew/v1.2.0.rst

@@ -374,6 +374,7 @@ Datetimelike
 - Bug in :class:`DatetimeIndex.shift` incorrectly raising when shifting empty indexes (:issue:`14811`)
 - :class:`Timestamp` and :class:`DatetimeIndex` comparisons between timezone-aware and timezone-naive objects now follow the standard library ``datetime`` behavior, returning ``True``/``False`` for ``!=``/``==`` and raising for inequality comparisons (:issue:`28507`)
 - Bug in :meth:`DatetimeIndex.equals` and :meth:`TimedeltaIndex.equals` incorrectly considering ``int64`` indexes as equal (:issue:`36744`)
+- Bug in :meth:`TimedeltaIndex.sum` and :meth:`Series.sum` with ``timedelta64`` dtype on an empty index or series returning ``NaT`` instead of ``Timedelta(0)`` (:issue:`31751`)
 
 Timedelta
 ^^^^^^^^^

pandas/core/arrays/timedeltas.py

@@ -381,14 +381,12 @@ def sum(
         nv.validate_sum(
             (), dict(dtype=dtype, out=out, keepdims=keepdims, initial=initial)
         )
-        if not self.size and (self.ndim == 1 or axis is None):
-            return NaT
 
         result = nanops.nansum(
-            self._data, axis=axis, skipna=skipna, min_count=min_count
+            self._ndarray, axis=axis, skipna=skipna, min_count=min_count
         )
-        if is_scalar(result):
-            return Timedelta(result)
+        if axis is None or self.ndim == 1:
+            return self._box_func(result)
         return self._from_backing_data(result)
 
     def std(
@@ -403,13 +401,11 @@ def std(
         nv.validate_stat_ddof_func(
             (), dict(dtype=dtype, out=out, keepdims=keepdims), fname="std"
         )
-        if not len(self):
-            return NaT
-        if not skipna and self._hasnans:
-            return NaT
 
-        result = nanops.nanstd(self._data, axis=axis, skipna=skipna, ddof=ddof)
-        return Timedelta(result)
+        result = nanops.nanstd(self._ndarray, axis=axis, skipna=skipna, ddof=ddof)
+        if axis is None or self.ndim == 1:
+            return self._box_func(result)
+        return self._from_backing_data(result)
 
     # ----------------------------------------------------------------
     # Rendering Methods

pandas/core/nanops.py

@@ -228,7 +228,7 @@ def _maybe_get_mask(
             # Boolean data cannot contain nulls, so signal via mask being None
             return None
 
-        if skipna:
+        if skipna or needs_i8_conversion(values.dtype):
             mask = isna(values)
 
     return mask

pandas/tests/arrays/test_timedeltas.py

@@ -2,6 +2,7 @@
 import pytest
 
 import pandas as pd
+from pandas import Timedelta
 import pandas._testing as tm
 from pandas.core.arrays import TimedeltaArray
 
@@ -175,7 +176,7 @@ def test_neg_freq(self):
 
 
 class TestReductions:
-    @pytest.mark.parametrize("name", ["sum", "std", "min", "max", "median"])
+    @pytest.mark.parametrize("name", ["std", "min", "max", "median"])
     @pytest.mark.parametrize("skipna", [True, False])
     def test_reductions_empty(self, name, skipna):
         tdi = pd.TimedeltaIndex([])
@@ -187,6 +188,19 @@ def test_reductions_empty(self, name, skipna):
         result = getattr(arr, name)(skipna=skipna)
         assert result is pd.NaT
 
+    @pytest.mark.parametrize("skipna", [True, False])
+    def test_sum_empty(self, skipna):
+        tdi = pd.TimedeltaIndex([])
+        arr = tdi.array
+
+        result = tdi.sum(skipna=skipna)
+        assert isinstance(result, Timedelta)
+        assert result == Timedelta(0)
+
+        result = arr.sum(skipna=skipna)
+        assert isinstance(result, Timedelta)
+        assert result == Timedelta(0)
+
     def test_min_max(self):
         arr = TimedeltaArray._from_sequence(["3H", "3H", "NaT", "2H", "5H", "4H"])
 

pandas/tests/series/test_reductions.py

@@ -15,6 +15,16 @@ def test_reductions_td64_with_nat():
     assert ser.max() == exp
 
 
+@pytest.mark.parametrize("skipna", [True, False])
+def test_td64_sum_empty(skipna):
+    # GH#37151
+    ser = Series([], dtype="timedelta64[ns]")
+
+    result = ser.sum(skipna=skipna)
+    assert isinstance(result, pd.Timedelta)
+    assert result == pd.Timedelta(0)
+
+
 def test_td64_summation_overflow():
     # GH#9442
     ser = Series(pd.date_range("20130101", periods=100000, freq="H"))