BUG: Fix subtraction with timezone Series

Fixes unusual bug with timezone Series
subtraction in which single element Series
objects containing tz-aware objects would
return a timedelta of zero, even though it
visually could not be the case.

The bug was traced to the conversion of
the contained timezones to UTC, in which
the method call was somehow returning NaT,
even though attempts to replicate that
behaviour were unsuccessful. This new
method call fixes the issue and is in some
ways more intuitive given the comment above
the conversions.

Closes gh-12290.

Author: gfyoung

pandas/core/ops.py

@@ -487,9 +487,9 @@ def _offset(lvalues, rvalues):
 
                 # with tz, convert to UTC
                 if self.is_datetime64tz_lhs:
-                    lvalues = lvalues.tz_localize(None)
+                    lvalues = lvalues.tz_convert('UTC')
                 if self.is_datetime64tz_rhs:
-                    rvalues = rvalues.tz_localize(None)
+                    rvalues = rvalues.tz_convert('UTC')
 
                 lvalues = lvalues.view(np.int64)
                 rvalues = rvalues.view(np.int64)

pandas/tests/series/test_operators.py

@@ -1375,3 +1375,18 @@ def test_datetime64_with_index(self):
         df['expected'] = df['date'] - df.index.to_series()
         df['result'] = df['date'] - df.index
         assert_series_equal(df['result'], df['expected'], check_names=False)
+
+    # See gh-12290
+    def test_datetime64_single_non_native(self):
+        f = Series([pd.Timestamp('2016-02-08 13:43:14.605000',
+                                 tz='America/Sao_Paulo')])
+        s = Series([pd.Timestamp('2016-02-10 13:43:14.605000',
+                                 tz='America/Sao_Paulo')])
+
+        out = f - s
+        expected = Series(pd.Timedelta('-2days'))
+        assert_series_equal(out, expected)
+
+        out = s - f
+        expected = Series(pd.Timedelta('2days'))
+        assert_series_equal(out, expected)