[Python] Why does Pandas rolling std output different results for different tail sizes?

Output of pandas rolling std is not consistent when data size changes. I tried to apply rolling std for different slices of the data, while all of them are at the tail of the series. I observed that the result for s.tail(3) is different from that of s.tail(4) and s.tail(5) -

>>> s = pd.Series(np.random.default_rng(seed=123).random(size=5))
>>> s[1] = 10000000 # just a large number
>>> s
0 6.823519e-01
1 1.000000e+07
2 2.203599e-01
3 1.843718e-01
4 1.759059e-01
dtype: float64
>>> s.tail(3).rolling(window=3, min_periods=1).std()
2 NaN
3 0.025447
4 0.023604
dtype: float64
>>> s.tail(4).rolling(window=3, min_periods=1).std()
1 NaN
2 7.071068e+06
3 5.773503e+06
4 0.000000e+00
dtype: float64
>>> s.tail(5).rolling(window=3, min_periods=1).std()
0 NaN
1 7.071067e+06
2 5.773502e+06
3 5.773503e+06
4 0.000000e+00
dtype: float64


In comparison, if I apply np.std to the rolling windows, the last result would be consistent:

>>> s.tail(3).rolling(window=3, min_periods=1).apply(np.std)
2 0.000000
3 0.017994
4 0.019273
dtype: float64
>>> s.tail(4).rolling(window=3, min_periods=1).apply(np.std)
1 0.000000e+00
2 5.000000e+06
3 4.714045e+06
4 1.927280e-02
dtype: float64
>>> s.tail(5).rolling(window=3, min_periods=1).apply(np.std)
0 0.000000e+00
1 5.000000e+06
2 4.714045e+06
3 4.714045e+06
4 1.927280e-02
dtype: float64


What is causing pandas' rolling standard deviation calculation on the same windows size 3 results in different outcomes?

Pandas version: 2.2.2 (also tried 2.3.0) on Windows

Continue reading...