jbrockmendel / sm2

A statsmodels fork focused on bugfixes and code quality

sm2

statsmodels is an excellent project and important part of the python scientific stack. But due to resource constraints, they cannot push out bugfixes often enough for my needs. sm2 is a fork focused on bugfixes and addressing technical debt.

Ideally sm2 will be a drop-in replacement for statsmodels. In places where this fails, feel free to open an issue.

With luck, fixes made here will eventually be ported upstream.

Build Status
Coverage

Changes vs Statsmodels

• sm2 contains a subset of the functionality of statsmodels. The first big difference is that statsmodels is more feature-complete.

• Test coverage statistics reported for sm2 are meaningful (:issue:4331)

• An enormous amount of code-cleanup has been done in sm2. Thousands of lines of unused, untested, or deprecated code have been removed. Many thousands of flake8 formatting issues have been cleaned up.

• MultinomialResults.params and predict will have correct column and row labels (:issue:4541)

• VARResults.cov_params will correctly return a DataFrame instead of raising ValueError.

• VARResults.acf will return correct results (:issue:4572)

• The ArmaProcess class does not have a nobs attribute.

• tsa.stattools.acf will always return (acf, confint, qstat, pvalue) here instead of a different subset of these depending on the inputs.

• stats.diagnostic.acorr_ljungbox will always return (qljungbox, pval, qboxpierce, pvalbp) here instead of a different subset of these depending on the inputs.

• summary2 methods have not been ported from upstream, will raise NotImplementedError.

• VARResults.test_whiteness has been superceeded upstream by test_whiteness_new as the older method was not an actual statistical test (:issue:4036). sm2 replaces the older version entirely and keeps only the name test_whiteness.

• ARModel.fit incorrectly sets model.df_resid upstream. That has been fixed here.

• GenericLikelihoodModelResults.__init__ incorrectly sets model.df_resid and model.df_model. That has been fixed here.

• GeneralizedLinearModel.fit incorrect sets self.mu and self.scale. This has been fixed here. (:issue:4032)

• LikelihoodModelResults._get_robustcov_results incorrectly ignores use_self argument. This has been fixed here. (:issue:4401)

Contributing

Issues and Pull Requests are welcome. If you are looking a place to start, here are some suggestions:

• Search for comments starting with # TODO: or # FIXME:

  • Some comments are copied from upstream and should have these labels but are missing them. If you find a comment that should have one of these labels (or is just unclear), add the label.

• Many tests from upstream are marked with pytest.mark.not_vetted to reflect the fact that they haven't been reviewed since being ported from statsmodels. To "vet" a test, try to determine:

  • Is this a "smoke test"? If so, it should be marked with pytest.mark.smoke.
  • Is this a test for a specific bug? Can an Issue reference (e.g. # GH#1234) be included?
  • Is there something specific being tested? If so, the test name should be made informative and often a comment should be added (e.g. # test function foo.bar in case where baz argument is near-singular)
  • Is this testing results produced by statsmodels/sm2 against results produced by another package? If so, it should be clear how those results were produced. The original authors put a lot of effort into producing these comparisons; they should be reproducible.

• There are some spots where tests are meager and could use some attention:

  • tsa.vector_ar.irf
  • regression._prediction
  • stats.sandwich_covariance

• As of 2018-03-19 there are still 390 flake8 warnings/errors. For many of these, fixing them requires figuring out what the writer's attention was upstream.

• As of 2018-03-19 about 20% of statsmodels has been ported to sm2 (though a much larger percentage of the usable, non-redundant, non-deprecated code). If there are portions of statsmodels that you want or need, don't be shy.

• If there is a change you parrticularly like, make a Pull Request upstream to get it implemented directly in statsmodels.

sm2's Issues

CI: Appveyor isn't cancelling old builds when newer commit is pushed

CI: Use Azure Pipelines?

Cov_params is method in base class, property in VARResults

codecov for master branch

TST: Try to improve precision in test assertions

CI: Update to CircleCI 2.0

Support dask/xarray?

wheels for linux, windows

CI: Appveyor running twice?

TST: Make sure test answers are reproducible

TST: Standardize on assert_allclose

skip-slow

I don't see anything wrong with skip-slow here, if this is what the question is about. When I reenable only-slow, I get (more or less) the mirror of skip-slow.

Readme rendering on PyPi page

It expects rst.