Random Rule Forest: Predicting Startup Success with LLM-Written Yes-or-No Questions

Random Rule Forest (RRF) asks a language model to write simple yes-or-no questions about founders, keeps the ones that best pick out successful founders, and combines them into a scorecard an investor can read and edit. In our paper, written by researchers at the University of Oxford and Vela Partners, half of the founders it flagged went on to succeed on a test set where 1 in 10 founders succeeded, 5× better than chance.

Why use questions to predict startup success?

Only around 2% of early-stage founders succeed in reality, so a model that helps find them has to be accurate and also explainable to the investors who act on it. Complex models such as deep neural networks are hard for an investor to interpret, whereas a list of readable questions can be checked, challenged and edited. Random Rule Forest therefore uses a language model to write readable features and keeps the final decision rule deliberately simple. Each question is a weak signal on its own, while many together become a strong one.

How does Random Rule Forest work?

1. Describe founders in words. Profile data from Crunchbase and LinkedIn is turned into short descriptions a language model can reason about, such as “MBA from Stanford University. Product Manager at Google, then VP at a health tech startup. Founded a telemedicine company acquired for $100M.”
2. Write and improve questions. OpenAI's GPT-4o mini writes 10 yes-or-no questions, and a second GPT-4o mini call answers them for founders whose outcomes are known. The first model is then told how accurate each question was, so its next 10 questions can build on the ones that picked out successful founders. Repeating this gives a pool of 100 questions.
3. Filter and rank. Questions whose precision (the share of flagged founders who succeed) is worse than chance are dropped, and the rest are ranked by precision.
4. Vote. A founder is predicted to succeed when enough of the top questions are answered yes. The best setting used the top 43 questions and needed 36 yes answers.

No single attribute is required, because a founder is flagged by gathering enough positive signals. Three of the highest-ranked questions in the paper are shown below.

• “Has the founder been involved in any venture capital or private equity firms?”
• “Is the founder's university ranked among the top 50 globally?”
• “Is the founder's educational background relevant to their startup's industry?”

What did the paper find?

The paper uses founder profiles from Crunchbase and LinkedIn, for companies founded between 2010 and 2016 that raised between $100,000 and $4 million. A founder counts as successful if the company later went public or was acquired at a valuation above $500 million, or raised more than $500 million. Random Rule Forest wrote and chose its questions using one group of founders and was then tested on 500 founders it had not seen, of whom 10% succeeded.

• On the test set, 50% of the founders Random Rule Forest flagged went on to succeed, 5× the 10% chance rate.
• Adding 20 questions guided by traits that domain experts suggested, such as prior exits or experience scaling teams, raised precision to 54% (5.4× chance).

The test set is small, and successful founders make up 10% of it against around 2% in reality, so the results still need testing at the real-world success rate. They also depend on complete and accurate founder profiles.

Where can I read more?

• Griffin, Ternasky, Alican and Ihlamur. Random Rule Forest (RRF): Interpretable Ensembles of LLM-Generated Questions for Predicting Startup Success. arXiv:2505.24622, May 2025.
• More about Random Rule Forest on the Vela Partners research page.

The founder data comes from Crunchbase and LinkedIn and includes paid sources, which may limit how much of it can be shared.