update whats new

Author: auderson

doc/source/whatsnew/v1.5.0.rst

@@ -94,10 +94,7 @@ Other enhancements
 - :meth:`DataFrame.reset_index` now accepts a ``names`` argument which renames the index names (:issue:`6878`)
 - :meth:`pd.concat` now raises when ``levels`` is given but ``keys`` is None (:issue:`46653`)
 - :meth:`pd.concat` now raises when ``levels`` contains duplicate values (:issue:`46653`)
-- Added ``numeric_only`` argument to :meth:`DataFrame.corr`, :meth:`DataFrame.corrwith`, and :meth:`DataFrame.cov` (:issue:`46560`)
-- :meth:`DataFrame.rolling.var`, :meth:`DataFrame.rolling.std`, :meth:`Series.rolling.var`, :meth:`Series.rolling.std` now generate correct result 0 for window containing same values (:issue:`42064`)
-- :meth:`DataFrame.rolling.skew`, :meth:`Series.rolling.skew` now generate correct result 0 for window containing same values (:issue:`42064`)
-- :meth:`DataFrame.rolling.kurt`, :meth:`Series.rolling.kurt` now generate correct result -3 for window containing same values (:issue:`42064`, :issue:`30993`)
+- Added ``numeric_only`` argument to :meth:`DataFrame.corr`, :meth:`DataFrame.corrwith`, and :meth:`DataFrame.cov` (:issue:`46560`)\
 
 .. ---------------------------------------------------------------------------
 .. _whatsnew_150.notable_bug_fixes:
@@ -181,10 +178,87 @@ Styler
 
 - Fix showing "None" as ylabel in :meth:`Series.plot` when not setting ylabel (:issue:`46129`)
 
-.. _whatsnew_150.notable_bug_fixes.notable_bug_fix2:
+.. _whatsnew_150.notable_bug_fixes.rolling_moments:
 
-notable_bug_fix2
-^^^^^^^^^^^^^^^^
+rolling mean/sum/std/var/skew/kurt now output correct result for window with same values
+^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
+
+Suffered from floating point artifacts, rolling calculations which are implemented
+with online update algorithms will output results slightly different from the ground truth.
+This can be toxic especially when you have a window of exactly same numbers (:issue:`42064`):
+
+*Old behavior*:
+
+.. ipython:: python
+
+    sr1 = pd.Series([1000, 0, 0, 0])
+    sr2 = pd.Series([1 / 3, 4, 0, 0, 0, 0])
+
+.. code-block::ipython
+
+    In [2]: sr1.rolling(3).std()
+    Out[2]:
+    0           NaN
+    1           NaN
+    2    577.350269
+    3      0.000008
+    dtype: float64
+
+    In [3]: sr2.rolling(3).sum()
+    Out[3]:
+    0             NaN
+    1             NaN
+    2    4.333333e+00
+    3    4.000000e+00
+    4    4.440892e-16
+    5    4.440892e-16
+    dtype: float64
+
+    In [4]: sr2.rolling(3).skew()
+    Out[4]:
+    0         NaN
+    1         NaN
+    2    1.688202
+    3    1.732051
+    4         NaN
+    5         NaN
+    dtype: float64
+
+Now the rolling calculations will record how many same values have appeared and it will
+overwrite the result with the true value. That will be ``prev_value`` for ``mean``,
+``obs * prev_value`` for ``sum``, ``0`` for ``std`` & ``var``, etc.
+
+*New behavior*:
+
+.. code-block::ipython
+
+    In [2]: sr1.rolling(3).std()
+    Out[2]:
+    0           NaN
+    1           NaN
+    2    577.350269
+    3      0.000008
+    dtype: float64
+
+    In [3]: sr2.rolling(3).sum()
+    Out[3]:
+    0         NaN
+    1         NaN
+    2    4.333333
+    3    4.000000
+    4    0.000000
+    5    0.000000
+    dtype: float64
+
+    In [4]: sr2.rolling(3).skew()
+    Out[4]:
+    0         NaN
+    1         NaN
+    2    1.688202
+    3    1.732051
+    4    0.000000
+    5    0.000000
+    dtype: float64
 
 .. ---------------------------------------------------------------------------
 .. _whatsnew_150.api_breaking: