The whole way we measure success rates is fundamentally flawed. We need to measure the project's actual success against its own *expectation* of success. For example, a person or an organization might start a blue-sky project just to see where it goes, being careful not to pour in too many resources until they see that it has promise. They may give the project, say, a 10% chance of success at its outset, and structure their investment accordingly.
Now suppose that 10% of such projects actually do succeed. Would you say then that we have a 10% success rate? No. It would be more accurate to say we have a 100% success rate (and if 20% of them succeed, we have a 200% success rate!) as measured against originally expected success.
That's why looking at just the numbers over at SourceForge can be deceptive. Yes, the vast majority of projects there never go anywhere. But did the people who started those projects expect them to go anywhere? If so, where, and with what probability?
(Hmm, and I made the same oversimplification, at the beginning of http://producingoss.com/html-chunk/introduction.html, perhaps a footnote is in order.)