The Rows Everyone Deletes: What 6,900 Accelerator Outcomes Taught Us

I lost close to a million dollars as an angel investor before I understood what I'm about to tell you, so read it as a confession rather than a lecture.

This week we finished building something I'd wanted for years: an outcome graph of roughly 6,900 accelerator companies — every Y Combinator batch since 2005, every Startmate cohort — with each company linked to what happened next. Raised. Acquired. Went public. Or died. The point of the exercise was never the directory. It was the outcomes — the column that turns a database into a teacher.

Three things fell out of it that are worth your time.

Failure is the training data everyone throws away

Of the companies with a known status, about one in six is dead. That number should stop you, because these companies were selected — vetted by the most competitive early-stage programmes in the world, funded, mentored, paraded at a demo day. The yes did not protect them. Now hold that against how startup-scoring models are usually built: on decks that raised, founders who made it, outcomes with a headline. The failures are missing, because failures don't issue press releases.

A model trained only on survivors hasn't learned what good looks like. It has learned what lucky looks like, and it will call almost everything promising. The fix is boring and hard: go and label the dead ones, and let them sit in the training set at full weight. A scoring engine that has read a thousand obituaries is simply harder to impress — and that scepticism is most of its value.

Does getting into a top accelerator predict success?

It's tempting to treat "got into a top accelerator" as a proxy for quality. The graph won't let you. One in six of the accepted died; meanwhile the archive is full of companies that got the no — or never applied — and outgrew the batch that got in. This isn't an argument that selection is noise. It's a sharper claim: a committee is pricing fit — this programme, this partner, this year's thesis, this narrow window — and fit is only loosely coupled to whether the market will eventually decide your company deserves to exist.

For a founder, the practical translation is unsentimental. An accelerator's rejection is information about a door, not a ruling on the building. For an investor, it's a warning against outsourcing conviction to someone else's admissions decision.

The outcome is the only signal that doesn't flatter you

When you track a startup to its result, you've also, quietly, tracked its investors to theirs — an acquisition is a realised return for every fund on that cap table. Roll that up across thousands of companies and a second graph appears beneath the first: which investors keep turning up next to the wins. The same handful of names recur, which surprises no one; what matters is that it's now a query rather than a reputation. And every new company resolved surfaces investors the dataset didn't have yet, so the map extends its own edges over time.

None of this is exotic. It's the unglamorous middle of the work — the part that doesn't fit in a launch tweet — and it's exactly the part that compounds.

The direction, though, is not in doubt. The teams building AI for venture are converging on the same commodity — an agent that reads a deck and returns a score. That harness will be free soon. The edge was never there. The edge is the slow, unfashionable ledger of what happened to the companies you scored, fed back in until the score stops being an opinion with a decimal point and starts being judgment.

You cannot prompt your way to that ledger. It accrues by waiting, and by doing the part nobody photographs: labelling the failures, correcting the bad matches, keeping the record honest when the honest record is embarrassing.

Think in decades and it's not a close call. The team that logs outcomes for three years beats the team with the cleverest model, every time. One of those compounds. The other churns.

Frequently asked questions

What share of accelerator-backed startups fail?

In our outcome graph of ~6,900 companies across Y Combinator and Startmate, about one in six of the companies with a known status is dead — despite having been selected and funded by highly competitive programmes. It's a reminder that acceptance was never a guarantee.

Why is survivorship bias a problem for startup scoring?

Because the clean, available data in venture is the winners — the decks that raised and the companies with a headline. Failures leave little structured trace, so models train mostly on survivors and learn to call almost everything promising. Labelling the failures and weighting them equally is what makes a score sceptical enough to be useful.

Does an accelerator rejection mean a startup is weak?

No. Selection prices fit — this programme, this partner, this year's thesis — far more than it rules on whether a company deserves to exist. The archive is full of companies that got the 'no', or never applied, and outgrew the batch that got in.