ENH: improved handling of NAs in binary ops with object Series, GH #737

Author: wesm

RELEASE.rst

@@ -89,6 +89,8 @@ pandas 0.7.0
     #657)
   - Implement array interface on Panel so that ufuncs work (re: #740)
   - Add ``sort`` option to ``DataFrame.join`` (GH #731)
+  - Improved handling of NAs (propagation) in binary operations with
+    dtype=object arrays (GH #737)
 
 **API Changes**
 

pandas/core/series.py

@@ -43,27 +43,46 @@ def _arith_method(op, name):
     Wrapper function for Series arithmetic operations, to avoid
     code duplication.
     """
+    def na_op(x, y):
+        try:
+            result = op(x, y)
+        except TypeError:
+            if isinstance(x, np.ndarray) and isinstance(y, np.ndarray):
+                mask = notnull(x) & notnull(y)
+                result = np.empty(len(x), dtype=x.dtype)
+                result[mask] = op(x[mask], y[mask])
+            elif isinstance(x, np.ndarray):
+                mask = notnull(x)
+                result = np.empty(len(x), dtype=x.dtype)
+                result[mask] = op(x[mask], y)
+            else:
+                mask = notnull(y)
+                result = np.empty(len(y), dtype=y.dtype)
+                result[mask] = op(x, y[mask])
+
+        return result
+
     def wrapper(self, other):
         from pandas.core.frame import DataFrame
 
         if isinstance(other, Series):
             if self.index.equals(other.index):
                 name = _maybe_match_name(self, other)
-                return Series(op(self.values, other.values), index=self.index,
-                              name=name)
+                return Series(na_op(self.values, other.values),
+                              index=self.index, name=name)
 
             this_reindexed, other_reindexed = self.align(other, join='outer',
                                                          copy=False)
-            arr = op(this_reindexed.values, other_reindexed.values)
+            arr = na_op(this_reindexed.values, other_reindexed.values)
 
             name = _maybe_match_name(self, other)
             return Series(arr, index=this_reindexed.index, name=name)
         elif isinstance(other, DataFrame):
             return NotImplemented
         else:
             # scalars
-            return Series(op(self.values, other), index=self.index,
-                          name=self.name)
+            return Series(na_op(self.values, other),
+                          index=self.index, name=self.name)
     return wrapper
 
 def _radd_compat(left, right):

pandas/tests/test_series.py

@@ -1135,6 +1135,24 @@ def test_operators_empty_int_corner(self):
     #     expected = (self.ts >= -0.5) & (self.ts <= 0.5)
     #     assert_series_equal(selector, expected)
 
+    def test_operators_na_handling(self):
+        from decimal import Decimal
+        from datetime import date
+        s = Series([Decimal('1.3'), Decimal('2.3')],
+                   index=[date(2012,1,1), date(2012,1,2)])
+
+        result = s + s.shift(1)
+        self.assert_(isnull(result[0]))
+
+        s = Series(['foo', 'bar', 'baz', np.nan])
+        result = 'prefix_' + s
+        expected = Series(['prefix_foo', 'prefix_bar', 'prefix_baz', np.nan])
+        assert_series_equal(result, expected)
+
+        result = s + '_suffix'
+        expected = Series(['foo_suffix', 'bar_suffix', 'baz_suffix', np.nan])
+        assert_series_equal(result, expected)
+
     def test_idxmin(self):
         # test idxmin
         # _check_stat_op approach can not be used here because of isnull check.