Pandas rolling std yields inconsistent results and differs from values.std

ppmt :

Using pandas v1.0.1 and numpy 1.18.1, I want to calculate the rolling mean and std with different window sizes on a time series. In the data I am working with, the values can be constant for some subsequent points such that - depending on the window size - the rolling mean might be equal to all the values in the window and the corresponding std is expected to be 0.

However, I see a different behavior using the same df depending on the window size.

MWE:

for window in [3,5]:
    values = [1234.0, 4567.0, 6800.0, 6810.0, 6821.0, 6820.0, 6820.0, 6820.0, 6820.0, 6820.0, 6820.0]
    df = pd.DataFrame(values, columns=['values'])
    df.loc[:, 'mean'] = df.rolling(window, min_periods=1).mean()
    df.loc[:, 'std'] = df.rolling(window, min_periods=1).std(ddof=0)
    print(df.info())
    print(f'window: {window}')
    print(df)
    print('non-rolling result:', df['values'].iloc[len(df.index)-window:].values.std())
    print('')

Output:

window: 3
    values         mean          std
0   1234.0  1234.000000     0.000000
1   4567.0  2900.500000  1666.500000
2   6800.0  4200.333333  2287.053757
3   6810.0  6059.000000  1055.011216
4   6821.0  6810.333333     8.576454
5   6820.0  6817.000000     4.966555
6   6820.0  6820.333333     0.471405
7   6820.0  6820.000000     0.000000
8   6820.0  6820.000000     0.000000
9   6820.0  6820.000000     0.000000
10  6820.0  6820.000000     0.000000
non-rolling result: 0.0

window: 5
    values         mean          std
0   1234.0  1234.000000     0.000000
1   4567.0  2900.500000  1666.500000
2   6800.0  4200.333333  2287.053757
3   6810.0  4852.750000  2280.329732
4   6821.0  5246.400000  2186.267193
5   6820.0  6363.600000   898.332366
6   6820.0  6814.200000     8.158431
7   6820.0  6818.200000     4.118252
8   6820.0  6820.200000     0.400000
9   6820.0  6820.000000     0.000021
10  6820.0  6820.000000     0.000021
non-rolling result: 0.0

As expected, the std is 0 for idx 7,8,9,10 using a window size of 3. For a window size of 5, I would expect idx 9 and 10 to yield 0. However, the result is different from 0.

If I calculate the std 'manually' for the last window of each window size (using idxs 8,9,10 and 6,7,8,9,10, respectively), I get the expected result of 0 for both cases.

Does anybody have an idea what could be the issue here? Any numerical caveats?

ipj :

It seems that implementation of std() in pd.rolling prefers high performance over numerical accuracy. However You can apply np version of standard deviation:

df.loc[:, 'std'] = df.rolling(window, min_periods=1).apply(np.std)

Result:

    values          std
0   1234.0     0.000000
1   4567.0  1666.500000
2   6800.0  2287.053757
3   6810.0  2280.329732
4   6821.0  2186.267193
5   6820.0   898.332366
6   6820.0     8.158431
7   6820.0     4.118252
8   6820.0     0.400000
9   6820.0     0.000000
10  6820.0     0.000000

Now precision is better.