rolling_mean returns small negative value on Series with non-negative values

[ahu2012]

We came upon this issue today while trying to take a square root of a non-negative series.

A quick search came up with this old issue: #2527 issue was reproduced on 0.15 by tweaking the code from the previous report, which is below:

from pandas import date_range, Series
s = Series(index=date_range('1999-02-03','1999-04-05'))
s['1999-02-03'] = 0.00012456
s['1999-02-04'] = 0.0003
s['1999-04-04'] = -0.0
s['1999-04-05'] = -0.0
In [21]: s.rolling(1).mean()['1999-04-05']
Out[21]: -5.421010862427522e-20

which returns -5.4210108624275222e-20.

[jreback]

does seem odd, but the value is indistinguishable from 0. so this is probably a round effect somewhere in cython. I'll mark it, but would need someone to deep dive into the code

In [43]: np.allclose([pd.rolling_mean(s, 1)['1999-04-05']],[0.0],atol=1e-20)
Out[43]: False

In [45]: np.allclose([pd.rolling_mean(s, 1)['1999-04-05']],[0.0],atol=1e-19)
Out[45]: True