Source: statsmodels
Version: 0.14.2+dfsg-1
Severity: normal
Control: forwarded -1 https://github.com/statsmodels/statsmodels/issues/7866
statsmodels is failing tests on arm64 with scipy 1.14 (from
experimental)
1504s ______________ TestZeroInflatedModel_probit.test_fit_regularized
_______________
1504s
1504s self =
<statsmodels.discrete.tests.test_count_model.TestZeroInflatedModel_probit
object at 0xffff91ad24e0>
1504s
1504s @pytest.mark.skipif(PLATFORM_LINUX32, reason="Fails on 32-bit Linux")
1504s def test_fit_regularized(self):
1504s > super().test_fit_regularized()
1504s
1504s
/usr/lib/python3/dist-packages/statsmodels/discrete/tests/test_count_model.py:114:
1504s _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
_ _ _
1504s
/usr/lib/python3/dist-packages/statsmodels/discrete/tests/test_count_model.py:50:
in test_fit_regularized
1504s assert_allclose(res_reg.params[2:], self.res1.params[2:],
1504s _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
_ _ _
1504s
1504s args = (<function assert_allclose.<locals>.compare at 0xffff8c1b8540>,
array([ 0. , 0. , -0.02679517, 1.48236838]),
array([-0.08217884, 0.00856726, -0.02679518, 1.48236911]))
1504s kwds = {'equal_nan': True, 'err_msg': '', 'header': 'Not equal to
tolerance rtol=0.05, atol=0.05', 'verbose': True}
1504s
1504s @wraps(func)
1504s def inner(*args, **kwds):
1504s with self._recreate_cm():
1504s > return func(*args, **kwds)
1504s E AssertionError:
1504s E Not equal to tolerance rtol=0.05, atol=0.05
1504s E
1504s E Mismatched elements: 1 / 4 (25%)
1504s E Max absolute difference: 0.08217884
1504s E Max relative difference: 1.
1504s E x: array([ 0. , 0. , -0.026795, 1.482368])
1504s E y: array([-0.082179, 0.008567, -0.026795, 1.482369])
1504s
1504s /usr/lib/python3.12/contextlib.py:81: AssertionError
Full test log at
https://ci.debian.net/packages/s/statsmodels/unstable/arm64/51043990/
The problem was previously acknowledged upstream at
https://github.com/statsmodels/statsmodels/issues/7866
where the recommendation was
"The zeroinflated test can be skipped, fit_regularized for ZI model is
inherited, and I never checked how robust optimization is in those
cases. Starting values for optimization are not very good yet."
