This isn't just a statistical problem, it's also a classical exploration/exploitation trade-off. You want users to notice and vote on new answers (exploration), but users only want to see the best answers (exploitation). The order you show will influence future votes (and future answers).

In addition, it's a social engineering problem. At least people with a western psychology seem to respond very strongly when a score is attributed to their person (as opposed to a group success like in a wiki). So you better make the score personal and big and visible, and do not occasionally sort by random just to discover the true score.

I think that's a great example of the "smoothing" that I was alluding to, though not in a format accessible to most programmers. However it is still just using a function of upvotes and downvotes. I think true rating can be much better when you also incorporate number of opportunities to vote. Because having the opportunity to vote (by viewing an item, or purchasing it, or whatnot) and choosing not to vote is still a really useful piece of data about the quality of an item. Especially when you are comparing old items that have had millions of opportunities against new items with only thousands.

> number of opportunities

Yep, definitely. The only challenges there are that there's less literature about doing so and that if you have both up and down votes there's no longer one right way to define a single objective for scoring.