Merge pull request #2889 from stephenwlin/nanops-int64

BUG: nanops.var produces incorrect results due to int64 overflow (fixes #2888)

Author: jreback

pandas/core/nanops.py

@@ -138,6 +138,9 @@ def get_median(x):
 
 
 def _nanvar(values, axis=None, skipna=True, ddof=1):
+    if not isinstance(values.dtype.type, np.floating):
+        values = values.astype('f8')
+
     mask = isnull(values)
 
     if axis is not None:

pandas/tests/test_series.py

@@ -1460,7 +1460,8 @@ def testit():
 
             # check the result is correct
             nona = self.series.dropna()
-            assert_almost_equal(f(nona), alternate(nona))
+            assert_almost_equal(f(nona), alternate(nona.values))
+            assert_almost_equal(f(self.series), alternate(nona.values))
 
             allna = self.series * nan
             self.assert_(np.isnan(f(allna)))
@@ -1469,6 +1470,12 @@ def testit():
             s = Series([1, 2, 3, None, 5])
             f(s)
 
+            # 2888
+            l = [0]
+            l.extend(list(range(2**40,2**40+1000)))
+            s = Series(l, dtype='int64')
+            assert_almost_equal(float(f(s)), float(alternate(s.values)))
+
             # check date range
             if check_objects:
                 s = Series(bdate_range('1/1/2000', periods=10))