What percentage of registered participants eventually show up at conferences?

I am afraid that, especially for a small meeting, the variation of how many do not show up will be too large to be useful. For example, if I am organizing a 50-person event I would expect that anything from 2 to 10 would be a "normal" number of no-shows. That is, if you really can't have more than, say, 50 participants you can't really safely overbook at all. However, note that oftentimes a few participants more don't really hurt, but of course this depends on your infrastructure and plans for the meeting.

Some further considerations:

• Generally, free events have a very large number of no-shows (sometimes in the 50%+ range, from what I have heard) while paid ones do not. This has led to some meetings I know introducing a nominal attendance fee, just to make sure that people think a little before signing up. Some other meetings have also tried to levy a fee only if a registered participant does not show up, but I am not sure how well this works and if it is worth the administrative and inter-personal hassle.
• Note that no-shows are often not independent events, statistically speaking. For example, if you have a bunch of participants from the same group / team and they have an unexpected important deadline come up, they all will cancel at the same time.
• If you have a very small meeting and you know (some) of your participants well you can attempt to use this to make a stab at estimating no-shows. For instance, I once planned an event for about 30 people, and I knew that some of the senior professors who initially expressed interest where <50% to actually come. I also knew that for some fraction of the industry participants an important client meeting may come up making them cancel last-minute. Based on this information I decided that 35 registrations for a 30-person meeting was fairly safe (we ended up being 28).

Having "no shows" at a scientific conference is common (for a multitude od reasons). I could not find a general statistic of this phenomenon and I guess that the numbers vary a lot depending on the field and also on type of event.

Here is one thing I found: The Australasian Telecommunication Networks and Applications Conference, (ATNAC) published a welcome address in IEEE where is says that

The conference paper statistics were: Pending (no manuscript): 3; Withdrawn after review: 0; Rejected: 33; No Show: 4; Published: 40; Total valid: 78; Published Ratio: 51%; Acceptance Ratio: 56%.

(The numbers in the pdf are different, this is a note on the website.)

Hence, there where 78 submissions, 44 got accepted and 4 of them did not show up at the conference.

Note that the above example is about no-shows of accepted presentations. No shows among registered people without contribution is usually much higher.