[raashish1601]

move_std and move_var don't check that the number of values in the window is larger than ddof. When it isn't, they divide by zero or by a negative number, so you get inf, -0.0, or a negative variance:

>>> a = np.array([1.0, 2.0, 4.0, 7.0, 3.0])
>>> bn.move_var(a, window=3, min_count=1, ddof=2)
array([-0.        ,         inf,  4.66666667, 12.66666667,  8.66666667])
>>> bn.move_var(a, window=3, ddof=4)
array([         nan,          nan,  -4.66666667, -12.66666667,  -8.66666667])
>>> bn.slow.move_var(a, window=3, min_count=1, ddof=2)
array([        nan,         nan,  4.66666667, 12.66666667,  8.66666667])

nanvar/nanstd already return NaN when count <= ddof, and so does the slow reference (np.nanvar). This adds the same check to both the float and the integer code paths of the moving functions.

Added test_move_std_var_ddof_ge_count, which compares against bn.slow for float32/float64/int32/int64 and ddof from 0 up to window + 1. For integer input it compares against the float result, because np.var gives inf there while bn.nanvar gives NaN. The test fails on master and passes with this change. I ran the full test suite (216 passed, 1 xfailed) after building in a python:3.12 Docker image, plus ruff format --check and ruff check.

[raashish1601]

The pre-commit.ci failure is zizmor flagging .github/workflows/ci.yml (self-repository syntax). master fails the same way, so it's unrelated to this change.

[neutrinoceros]

Thanks for finding and fixing this.
Just running your trivially correct tests againt master shows that they correctly capture regressions. However I need to spend a couple minutes making sure I understand the problem and your solution well enough to then maintain it for ever. I'll try to make it happen before our upcoming release.