precision issue with rolling_std

[leo4183]

i tested rolling_std func with some real data (floating number with format x.xx). For a Series/DataFrame which contains let's say 5 continuous identical rows among several hundreds, the rolling_std could return a small number with 10^-7 scale instead exactly zero. this is probably a known issue which has not been officially documented. i'm just curious with the reason behind and why the std member func works better.

[porris1000]

I did the same tests and i confirm you are right. Pandas & Numpy std function return very different precisions.I think the problem doesn't only happen when there are repeated data, although it always occurs in this case. I only found your comment about this issue and have no idea about the reasons, sorry.

[kay1793]

#8424 , #8326

[mroeschke]

This may have been fixed by #18481, but we would need a reproducible example otherwise.

#27593 reproduces this. I don't think it's a bug. When you calculate the variance for each window in isolation, you can get an exact zero result in such a case. When you do a true rolling window calculation (as pandas does), you'll only get something very small in this case due to floating point issues.