PERF: rolling std/var x100 slower in current dev version

[yolking]

Pandas version checks

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

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

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

Reproducible Example

import pandas as pd
import numpy as np
import time
import platform

n = 2_000_000
window = 500
s = pd.Series(np.random.randn(n), name="x")

_ = s.rolling(window).std()

t0 = time.perf_counter()
_ = s.rolling(window).std()
t = time.perf_counter() - t0

print(f"Rolling std({window}) on {n:,} rows took {t:.4f} s")

Running on current dev (installed by pip install git+https://github.com/pandas-dev/pandas.git
or pip install --pre --extra-index https://pypi.anaconda.org/scientific-python-nightly-wheels/simple pandas)
I get

Rolling std(500) on 2,000,000 rows took 4.3563 s

Installed Versions

INSTALLED VERSIONS

commit : c4449f0
python : 3.14.0
python-bits : 64
OS : Windows
OS-release : 2022Server
Version : 10.0.22631
machine : AMD64
processor : AMD64 Family 25 Model 24 Stepping 1, AuthenticAMD
byteorder : little
LC_ALL : None
LANG : None
LOCALE : English_United States.1252

pandas : 3.0.0.dev0+2533.gc4449f0e1c
numpy : 2.3.3
dateutil : 2.9.0.post0
pip : 25.2
Cython : None
sphinx : None
IPython : None
adbc-driver-postgresql: None
adbc-driver-sqlite : None
bs4 : None
bottleneck : None
fastparquet : None
fsspec : None
html5lib : None
hypothesis : None
gcsfs : None
jinja2 : None
lxml.etree : None
matplotlib : 3.10.6
numba : None
numexpr : None
odfpy : None
openpyxl : None
psycopg2 : None
pymysql : None
pyarrow : None
pyiceberg : None
pyreadstat : None
pytest : None
python-calamine : None
pytz : 2025.2
pyxlsb : None
s3fs : None
scipy : None
sqlalchemy : None
tables : None
tabulate : None
xarray : None
xlrd : None
xlsxwriter : None
zstandard : None
qtpy : None
pyqt5 : None

Prior Performance

Running on pandas 2.3.3 I get

Rolling std(500) on 2,000,000 rows took 0.0377 s

Results on python 3.12 and 3.13 are same

[suzyahyah]

This is due to the ingesting the window size (w=500) upfront after the PR in #62514

I dont think there is a significantly faster algorithm as Welford's is already a single-pass O(n) algorithm

However, benchmarking against numpy which is O(n x w), suggests that pandas could be w times faster in theory and is probably missing some optimisations in the implementation.

import pandas as pd
import time
import numpy as np
from numpy.lib.stride_tricks import sliding_window_view

n = 2_000_000
window = 500
t0 = time.perf_counter()
s = pd.Series(np.random.randn(n), name="x")
_ = pd.Series(sliding_window_view(s, window_shape=window).var(axis=-1, ddof=1))
t = time.perf_counter() - t0
print(f"Rolling std({window}) on {n:,} rows took {t:.4f} s")

Rolling std(500) on 2,000,000 rows took 16.704 s

[suzyahyah]

Definitely a bug introduced by #62514, working on it..