ENH: robust covariance matrix estimation

[esmucler]

I've been working on a Python implementation of the Orthogonalized Gnanadesikan-Kettenring (OGK) robust covariance estimator of https://www.tandfonline.com/doi/abs/10.1198/004017002188618509. I don't think statsmodels currently offers any robust alternatives for covariance estimators and thought that this might be a nice addition; would there be interest in a PR for this?

I like the OGK estimator because it is a simple idea and unlike many (most?) robust covariance estimators that can deal with "gross error" outliers, is straightforward to compute.

[josef-pkt]

I worked for a while on those #3230

It looks like I used that reference https://github.com/statsmodels/statsmodels/pull/3230/files#diff-e8bee5153f380ff69a3b9b274b3a4192a3cdd4ef211daddd801a4abaa5f8c726R285

I don't know what the status of that PR is.

I'm looking at similar things on and off for multivariate application, e.g. outlier robust multivariate analysis. I don't have any application like those yet, and so I haven't gone back to finishing up that PR.

(As usual interface design and unit test are the difficult parts at the end.)

[mdaeron]

+1. I've been looking for a Python implementation of the OGK robust covariance estimator, but so far have found nothing.