statsmodels 0.15 removed predict(linear=) and other keywords, some without warning
Signature diff of statsmodels.api, formula.api, tsa.api and stats.api between 0.14.5 (with pandas<3 and scipy<1.17, because 0.14.5 fails to import on pandas 3) and 0.15.0, then calls. predict(linear=True) on Logit, Probit, MNLogit, Poisson, NegativeBinomial, GLM and GEE, at both model and results level, gave a FutureWarning on 0.14.5 and raises TypeError on 0.15.0 (use which='linear'). Also TypeError on 0.15.0: test_poisson_2indep(ratio_null=) (it warned on 0.14.5), and these three, which ran with no deprecation warning at all on 0.14.5: RLM.fit(init=), AutoReg(old_names=), acorr_lm(cov_kwargs=) (use cov_kwds=; same rename in linear_reset and compare_encompassing). sm.categorical is gone (AttributeError; it already raised NotImplementedError on 0.14.5). The Holt-Winters fit() keywords use_boxcox, initial_level and initial_trend also left the fit signature; they belong in the model constructor. That last one comes from the signature only, because my 0.14.5 calls raised ValueError for other reasons.
Follow-up on the Holt-Winters line, now tested with plain models: on 0.14.5, SimpleExpSmoothing(ts).fit(initial_level=10), Holt(ts).fit(initial_level=, initial_trend=) and ExponentialSmoothing(..., seasonal='add').fit(use_boxcox=True) already raise ValueError ('Initial values were set during model construction' / 'use_boxcox was set at model initialization'). On 0.15.0 they raise TypeError. ExponentialSmoothing(..., use_boxcox=True, initialization_method='estimated').fit() works on both. So this removal only changes the error type for code that was already failing. The predict(linear=) and cov_kwargs= removals are the ones that break working code.