Coverage remained the same when pulling fd448b9 on rgommers:boxcox into 572aaf0 on scipy:master.

This is great, Ralf! I made a few comments inline.

I think the method="all" option in boxcox_normmax is unnecessary. It is not a lot of code, but I don't think the small increase in convenience is worth it. But that's a judgment call, and I wouldn't hold up merging because of it.

I noticed a bug in boxcox: using a negative integer lambda with an integer array gives incorrect results:

In [27]: x
Out[27]: array([1, 2, 3])

In [28]: boxcox(x, -1)
Out[28]: array([-0.,  1.,  1.])

In [29]: boxcox(x, -1.0)
Out[29]: array([-0.        ,  0.5       ,  0.66666667])

In [30]: boxcox(x, -2)
Out[30]: array([-0. ,  0.5,  0.5])

In [31]: boxcox(x, -2.0)
Out[31]: array([-0.        ,  0.375     ,  0.44444444])

I think the core Box-Cox function,

    y = np.where(lmbda == 0, log(x), (x**lmbda - 1) / lmbda)

should be implemented as a ufunc, and perhaps added to scipy.special. This would eliminate some of the spurious Runtime Warnings generated when a value is zero, and it could be reused in stats.power_divergence.

I created a pull request that implements the core Box-Cox transformation as a ufunc in scipy.special: #3150

Re method='all': there are more methods that can easily be implemented, I just didn't get around to it. There's an R function (can't find it so quickly) that has seven options. So I'd like to keep 'all'.

Thanks for the review Warren. Addressed all your comments in the last commit.

Coverage remained the same when pulling 1e1f103 on rgommers:boxcox into 881fadf on scipy:master.

Did the Travis tests run? I don't see the usual Travis status bar.

Yes they did for the second to last commit. The last commit is a one line change to a docstring, so I added [ci skip] to the commit message to not trigger Travis for it (no need to waste resources for zero code changes).

OK, I see that now (and I just noticed the little green check mark next to the penultimate commit).

Merging.