NaN is not a score. For users blindly sorting scores and taking the last value, it is a thorn in the side. We have instead chosen to raise a warning and return 0. In what situation would NaN be more practically useful?

I'm praticular, with macro-averaging, it would be very unhelpful to return NaN if a system failed to ever predict a rare class

If you feel the documentation could be clearer, please offer a PR

NaN is not a score. For users blindly sorting scores and taking the last value, it is a thorn in the side. We have instead chosen to raise a warning and return 0. In what situation would NaN be more practically useful?

A very late comment on this: I calculate statistics of metrics across various groups and classifiers. For some of those groups and classifiers, there might be no positive instances. Returning NaN would be (very) desirable for me, because otherwise the returned 0s will (falsely) distort the computed metric statistics. Replacing all returned NaNs by 0 is easy if that is what the user desires (also forcing them to think about whether that actually makes sense, which is probably a good thing?), but simply replacing all returned 0s by NaN may be wrong - the value could also actually be zero. So I currently have no easy way to get the behavior I actually need, I believe?

On a more general note, is it a good idea to support users in "blindly sorting scores" and sweeping actually occurring problems with those metrics under the rug?