Comment 4 for bug 800011

I was just checking if I can run all models in the same or similar way

modols = sm.OLS(endog, exog)
modglsar = sm.GLSAR(endog, exog, rho=2)
modrlm = sm.RLM(endog, exog)
modlog = sm.Logit(endogbin, exog)
modglm = sm.GLM(endogbin, exog, family=sm.families.Binomial())
modar = sm.tsa.AR(endog)
modarma = sm.tsa.ARMA(endog)
modvar = sm.tsa.VAR(np.column_stack((endog, exog[:,1:])))

resols = modols.fit()
resglsar = modglsar.fit()
resrlm = modrlm.fit()
reslog = modlog.fit()
resglm = modglm.fit()
resar = modar.fit()
resarma = modarma.fit(order=(1,0))
resvar = modvar.fit()

ARMA is the only one without a default for everything in fit() but GenericLikelihood might also require something

I haven't looked at any details

instead of if np.any(np.all(np.diff(X,1,0)==1,0)), then has_trend.
maybe a check that the np.diff(X,1,0).var() < epsilon

for polynomial trends, I switched to rescaling trend to [-1, 1] or something like this, trend = np.linspace(-1,1,nobs)