Have later YCombinator classes had a larger percentage of homeruns? If you have more applicants and have gotten better at selecting, this should be the case.

Is it about selecting winners, or weeding out losers? Are there companies that you know will fail? Who would be least likely to succeed as an entrepreneur?

Home runs are so rare that it's not a matter of percentage per batch. A batch will have 1 or 0, and probably 0. It will be a few years before I can tell if the rate is increasing.

Selecting winners and weeding out losers seem the same thing to me. There are companies we think will almost certainly fail, but we can never be sure. An ineffectual person would be least likely to succeed as a startup founder.

With start-ups I think it would be too expensive an experiment to fund.

On October 7, 1998 the Journal presented the results of the 100th dartboard contest. So who won the most contests and by how much? The pros won 61 of the 100 contests versus the darts. That's better than the 50% that would be expected in an efficient market.

This is wrong. If the dartboard consistently underperformed most stockpickers (i.e. say it ranked around the 30th percentile year over year) then you could make the case that (some) people add value. If, on the other hand, the dartboard is near the mean of the distribution of outcomes, you could make a case that it's all luck.

No such market exists for startups - so selection is required.

Well, you could try ranking results and calculating a score based on how each person outperformed the dartboard, then calculating a score over time.