BUG: rolling_max with center=True and min_periods=1 returns incorrect results

[apuapaquola]

rolling_max with center=True and min_periods=1 returns NaN's

Python 3.3.5 (default, Mar 10 2014, 03:21:31) 
[GCC 4.8.2 20140206 (prerelease)] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import pandas as pd
>>> df=pd.DataFrame([0,1,2,1,0], columns=['a']

### This is OK
>>> pd.rolling_max(df['a'], window=3, center=True)
0   NaN
1     2
2     2
3     2
4   NaN
dtype: float64


### This should return [1, 2, 2, 2, 1]
>>> pd.rolling_max(df['a'], window=3, center=True, min_periods=1)
0     1
1     2
2     2
3     2
4   NaN
dtype: float64

>>> pd.show_versions()

INSTALLED VERSIONS
------------------
commit: None
python: 3.3.5.final.0
python-bits: 64
OS: Linux
OS-release: 3.13.6-1-ARCH
machine: x86_64
processor: 
byteorder: little
LC_ALL: None
LANG: en_US.UTF-8

pandas: 0.13.1
Cython: 0.20.1
numpy: 1.8.1
scipy: 0.13.3
statsmodels: 0.5.0
IPython: 1.2.1
sphinx: 1.2.2
patsy: None
scikits.timeseries: None
dateutil: 2.2
pytz: 2014.1
bottleneck: None
tables: None
numexpr: None
matplotlib: 1.3.1
openpyxl: None
xlrd: None
xlwt: None
xlsxwriter: None
sqlalchemy: None
lxml: None
bs4: None
html5lib: None
bq: None
apiclient: None

[apuapaquola]

Looking at the function _rolling_moment and _center_window in pandas/stats/moments.py, it looks like the moment is calculated the same way as with center=False, and then shifted by half size of window. Calling the shift function introduces NaN's.

def _center_window(rs, window, axis):
    ...
    offset = int((window - 1) / 2.)
    if isinstance(rs, (Series, DataFrame, Panel)):
        rs = rs.shift(-offset, axis=axis)
    ...

[TomAugspurger]

Yeah. It looks like the Cython function are completely unaware about the center argument. I'm trying to come up with a fix that keeps the separation.

I suppose we could always calculate the function for the tail and then throw away the one's we don't need. So in this case, with min_periods=1, the result of func() would be array([ 0., 1., 2., 2., 2., 1, 1, 0, 0, 0]). The last 5 are the tail that aren't calculated now. Then we apply the labels and the shift to center. Does that sound like it would work? We may even be able to calculate fewer extra items...

[hayd]

came up here (I think) http://stackoverflow.com/q/25628587/1240268

[mroeschke]

Looks fixed on master. Could use a test.

In [65]: df['a'].rolling(3, center=True, min_periods=1).max()
Out[65]:
0    1.0
1    2.0
2    2.0
3    2.0
4    1.0
Name: a, dtype: float64