I am running the balanced accuracy in the autism kit and I get negative score.
I would think that we can future port the implementation from scikit-learn.

Do we have real test for the metrics which are not from scikit-learn?