BUG: GH12902 fixed coercion of complex values to float when using groupby

Author: adneu

doc/source/whatsnew/v0.18.1.txt

@@ -325,3 +325,4 @@ Bug Fixes
 - ``pd.read_excel()`` now accepts column names associated with keyword argument ``names``(:issue `12870`)
 
 - Bug in ``fill_value`` is ignored if the argument to a binary operator is a constant (:issue `12723`)
+- Bug in ``groupby`` where complex types are coerced to float (:issue:`12902`)

pandas/core/groupby.py

@@ -1747,7 +1747,7 @@ def _cython_operation(self, kind, values, how, axis):
             values = _algos.ensure_float64(values)
         elif com.is_integer_dtype(values):
             values = values.astype('int64', copy=False)
-        elif is_numeric:
+        elif is_numeric and not com.is_complex_dtype(values):
             values = _algos.ensure_float64(values)
         else:
             values = values.astype(object)

pandas/tests/test_groupby.py

@@ -2475,6 +2475,13 @@ def test_groupby_level_0_nonmulti(self):
         result = a.groupby(level=0).sum()
         self.assertEqual(result.index.name, a.index.name)
 
+    def test_groupby_complex(self):
+        # GH 12902
+        a = Series(data=np.arange(4) * (1 + 2j), index=[0, 0, 1, 1])
+        result = a.groupby(level=0).sum()
+        expected = Series((1 + 2j, 5 + 10j))
+        assert_series_equal(result, expected)
+
     def test_level_preserve_order(self):
         grouped = self.mframe.groupby(level=0)
         exp_labels = np.array([0, 0, 0, 1, 1, 2, 2, 3, 3, 3])