BUG: Rolling variance is negative

[bhigy]

Pandas version checks

• I have checked that this issue has not already been reported.

• I have confirmed this bug exists on the latest version of pandas.

• I have confirmed this bug exists on the main branch of pandas.

Reproducible Example

import pandas as pd

A = [0.00000000e+00, 0.00000000e+00, 3.16188252e-18, 2.95781651e-16,
    2.23153542e-51, 0.00000000e+00, 0.00000000e+00, 5.39943432e-48,
    1.38206260e-73, 0.00000000e+00]
ts = pd.DataFrame(A)
print(ts.rolling(window=3, center=True).var(ddof=1))

Issue Description

I am trying to compute a rolling variance and for some reasons, some of the values I obtain are negative. For example, running the code above gives me:

              0
0           NaN
1  3.332500e-36
2  2.885385e-32
3  2.885385e-32
4  2.916226e-32
5 -5.473822e-48
6 -5.473822e-48
7 -5.473822e-48
8 -5.473822e-48
9           NaN

Expected Behavior

All values should be positive and match following array (with the exception of NaN values at the beginning and the end):

[0.00000000e+000 3.33250036e-036 2.88538519e-032 2.88538519e-032
 2.91622617e-032 1.65991678e-102 9.71796366e-096 9.71796366e-096
 9.71796366e-096 6.36699010e-147]

Right now, only the value from index 1 to 4 seem to be correct.

Installed Versions

INSTALLED VERSIONS

commit : 4c07e07
python : 3.10.6.final.0
python-bits : 64
OS : Linux
OS-release : 5.19.0-38-generic
Version : #39~22.04.1-Ubuntu SMP PREEMPT_DYNAMIC Fri Mar 17 21:16:15 UTC 2
machine : x86_64
processor : x86_64
byteorder : little
LC_ALL : None
LANG : en_US.UTF-8
LOCALE : en_US.UTF-8

pandas : 2.1.0.dev0+408.g4c07e0769c
numpy : 1.24.2
pytz : 2023.3
dateutil : 2.8.2
setuptools : 59.6.0
pip : 22.0.2
Cython : None
pytest : None
hypothesis : None
sphinx : None
blosc : None
feather : None
xlsxwriter : None
lxml.etree : None
html5lib : None
pymysql : None
psycopg2 : None
jinja2 : None
IPython : None
pandas_datareader: None
bs4 : None
bottleneck : None
brotli : None
fastparquet : None
fsspec : None
gcsfs : None
matplotlib : None
numba : None
numexpr : None
odfpy : None
openpyxl : None
pandas_gbq : None
pyarrow : None
pyreadstat : None
pyxlsb : None
s3fs : None
scipy : None
snappy : None
sqlalchemy : None
tables : None
tabulate : None
xarray : None
xlrd : None
zstandard : None
tzdata : 2023.3
qtpy : None
pyqt5 : None

[reddyrg1]

take

[manjalc]

The issue is probably in the roll_var() function in _libs/windows/aggregations.pyx

[topper-123]

Yes this is clearly a bug. A PR would be welcome.

[kaixiongg]

Hello, I'm encountering the same issue. Any updates on this? It seems like roll_var already uses the Welford method combined with Kahan summation for more stable precision. Is there any algorithm that offers even greater stability than that?

[kaixiongg]

take

You've already been dealing with this issue for a year without any updates. Could anyone raise this issue again?

[rhshadrach]

You've already been dealing with this issue for a year without any updates.

This is typical on the issue tracker. I think it's safe to consider any assigned issue that hasn't seen action for over a month to be abandoned. Any contributor is welcome to pick this up.

Could anyone raise this issue again?

This issue is still open, it does not need to be raised again.