asanyarray coercion in Model class breaks sparse matrices

[Skipper Seabold]

When we coerce with asanyarray in Model, it doesn't work with sparse matrices. It's a simple fix to check for sparse, but no time now. Otherwise, we can pass in block diagonal sparse matrices to GLS etc., if we want to and there's no other option for stacking (not convinced of this yet).

[joep (josef-pktd)]

We cannot replace asarray with asanyarray in any conversion!

statsmodels calculations are written to work with arrays, using matrices including sparse matrices would change multiplication and axis reduction (eg. in mean()). I expect that this would break things all over the place.

Enabling sparse matrix calculations would require a lot of changes to the internal code.

For the specific case of a sparse full sized covariance matrix, or it's cholesky (or other) decomposition, it might be worthwhile to specialcase this. As an alternative for GLS, we could use a callback function for whiten, that could then handle sparse calculations. But I'm not sure, and we have to check if changing only whiten is enough (I think it might work, since all GLS calculations go through whiten).

However, this would not be enough if we want to have exog sparse as for example for system of equations regression.

[Skipper Seabold (jsseabold)]

I meant asarray. Oops. It's a moot point I think anyway. I could have sworn I took a pseudoinverse of a sparse matrix last night and had the betas for VAR, but I can't repllicate. I am looking into rewriting it all in "companion matrix" form to see if this will be more efficient.