I have statistical models that can predict if a headline will get > 10 votes or if a headline that gets > 10 votes will get a comments/votes ratio > 0.5. (We can’t think of a better headline than “Richard Stallman is dead”)

Both of those are treated as probability calibrated classifiers because of the extreme variations in outcomes. You might post something one day and get 0 upvotes, another day somebody might post it and get 500. Not that many people read the “new” page so it is really iffy if you get your first few votes.

In principle you could get more information out of the data treating it as a regression problem (e.g. give more points for predicting that an article gets 500 votes as opposed to 50) but one model I made never predicted that anything gets more than 16.7 votes which means the error function is going to always look awful and not really be a good tool for improving the model. The logit models I”m using, however, give realistic scores in the face of uncertainty.