Let the Machine Say No: Earlybird's Retterath Argues Venture Alpha Now Lives in a Firm's Own Decision Data

51 minutes ago

4 min read

What's New

Public startup data is commoditized, and venture alpha now comes from the proprietary record of a firm's own decisions, codified so software can reject at the top of the funnel and humans can go deeper at the bottom. Andre Retterath, General Partner at Earlybird, argues this in a podcast on How I Invest. Since 2018 Earlybird has surveyed every investment committee participant on more than 200 decisions and tracked those votes against outcomes, alongside 1,000-plus memos and meeting transcripts. A 2020 paper he published found a machine learning model matched the best of 120 human investors at predicting a confident no. The payoff is fewer, better-prepared founder meetings on the 10 to 12 companies a year the firm will actually back.

Why It Matters

The consensus among data-driven VCs is that Harmonic, PitchBook, and their peers are the edge. Retterath's own 2026 landscape report shows two-thirds of firms now use technology for alpha rather than efficiency, which means vendor data confers none. On the other side are partnerships that never institutionalize judgment beyond a few individuals, and investment professionals who resist new workflows. Retterath built Earlybird's Eagle Eye platform and sells this thesis through his Data Driven VC newsletter, so the argument serves his brand. He is candid that the model rejected Lovable at pre-seed over an 8% ownership threshold and that the miss likely cost more than 50x.

Big Picture Drivers

The cost matrix is asymmetric: A false positive loses the money once. A false negative loses 50 times the upside. With 5% of a portfolio driving 95% of outcomes, the expensive error is the wrong no, which makes reliable automated rejection valuable.
Groupthink is structural: Tight networks, awareness of the power law, and the fear of missing the outlier push investors toward whatever peers are looking at. At Earlybird, "fund ABC is looking at it" counts as a negative signal.
Coverage was a measured illusion: Scraping public registers for 200 competitor funds showed Earlybird saw 72% of the deals its competitors closed in 2019 to 2020 while believing it saw everything. Eagle Eye now surfaces 19 of 20 at least 6 weeks before close.
Founder traits shift with the era, and some never move: Average unicorn founder age has held at 34 for 15 years. Founder branding barely mattered five years ago and now decides talent and capital access at the application layer. Defense founders in Europe went from unfundable to oversubscribed within the same span.
Decision data cannot be bought: IC votes, memos, transcripts, and labeled rejections train a recommendation engine that encodes Earlybird's taste. The target is a calendar auto-filled with founders at the intersection of predicted success and firm fit.
Change management beat data quality as the hard problem: Entity matching and quantifying qualitative signals proved solvable. Getting five-year veterans to abandon their CRM habits required tying geographic hit rate to performance reviews.

By The Numbers

72% to 96%: Earlybird's deal coverage across Europe before and after Eagle Eye, measured against 200 competitor funds' actual investments.
50x+: Retterath's estimate of the return forgone by rejecting Lovable because the round would have left Earlybird below its 15% ownership target.
200+: Investment committees captured with per-participant votes since 2018.
120: Investors benchmarked against the model in the 2020 paper; the model equaled the best of them at predicting a no.
8 to 3: Full-time engineers at peak, roughly 20% of the team, now down to three senior engineers on top of the data infrastructure.
10,000: Inbound opportunities a year, historically all answered by hand, now screened to roughly 200 that meet hard criteria.

Key Trends to Watch

Taste engines replacing dashboards: Retterath abandoned the dashboard model 18 months ago in favor of chat interfaces over MCPs. Watch whether other firms follow into recommendation systems trained on their own rejections.
Ownership thresholds loosening: After the Lovable miss, Earlybird became selectively flexible on shareholding. Expect more disciplined firms to relax initial ownership rules for outlier candidates.
Hit rate becoming a standard KPI: A firm that measures coverage against competitors' closed deals can hold junior investors to it. Watch for this metric spreading into performance reviews.
Source attribution reshaping calendars: Cohort analysis showed deals sourced at large investor conferences underperformed, so Earlybird shifted to research conferences like ICML. Expect more firms to cut events on cohort data.

Memorable Quotes

"Actually the machine learning model was as good as the best investor in our sample." The finding that earned the algorithm the right to say no at the top of the funnel.
"No firm will ever be able to sell this. This is unique to Earlybird." Why decision data, and not vendor data, is the moat.
"It's rather negative signal because we want independent thinkers." How Earlybird treats a peer fund's interest in a deal.
"We are not paid to make 10,000 decisions a year. We just need to make this one right decision." The reason to automate everything upstream of judgment.

The Wrap

The thesis holds if Earlybird keeps landing an outlier in every fund, as it has across 17 funds over 30 years, while spending human time on fewer companies, and if firms with codified decision records show better hit rates and fewer costly rejections than peers relying on purchased data. It also holds if the taste engine avoids the next Lovable rather than encoding the ownership bias that caused it. It fails if firm-specific data proves too small to generalize, if models trained on past decisions entrench past blind spots, or if the era-dependent traits Retterath describes shift faster than any engine can retrain. The 2024 to 2027 vintages, the first selected with the machine saying no, will show whether the record was worth building.