The Vela Library · Predicting startup success
Can startup success be predicted? What the research shows
The short answer
Partly, and only against a base rate. About 1.9% of US startups become unicorns, so a model counts only if it beats that rate. Vela Partners is an AI-native quant VC. Its founder models, published with the University of Oxford, reach 19% to 38% precision at that rate, against 5.6% for tier-1 VC firms.
Key facts
Every figure, with its source
1.9%
US base rate: the share of startups that become unicorns
Source · GPTree
19–38%
Precision of Vela's production models at that 1.9% base rate, a 10x to 20x lift over random
Source · Think-Reason-Learn
5.6%
Precision of nine tier-1 VC firms at the same 1.9% base rate
Source · Random Rule Forest
67%
Policy Induction with GPT-5 on VCBench, 7.4x that benchmark's 9% base rate
Source · Policy Induction
22x
Gap in success rate between the best and worst of eight founder types
Source · GPT-HTree
84%
Of successful founders look identical to failures on career data alone
Source · When Career Data Runs Out
9,000
Anonymised founder profiles in VCBench, the open benchmark
Source · VCBench
Why it's hard to predict
Success is rare. About 1.9% of US startups become unicorns, so a model that calls every founder a failure is right more than 98% of the time and finds no one. Accuracy hides that; precision does not. Precision asks: of the founders a model picks, how many succeed?
That number only means something next to the base rate. 19% precision is about ten times random at a 1.9% base rate; on a dataset where 9% succeed, it would be about two times. A tree trained on the raw 1.9% rate collapses to predicting failure for everyone, which is why the papers balance their training data and then report precision back at the real rate.
Base rateThe base rate is the share of founders in a dataset who go on to succeed. A prediction model's precision means something only as a multiple of it, so every result has to name the base rate it was measured against.
What the research shows
The papers are measured on three kinds of data, each with its own base rate. Each table below keeps to one, so read every row against the random line at the top of its own table, never against another table.
| Method | Precision | Lift over random | Source |
|---|---|---|---|
| Random selection | 1.9% | 1x | GPTree |
| Y Combinator | 3.2% | About 1.7x | Random Rule Forest |
| Nine tier-1 VC firms | 5.6% | About 2.9x | Random Rule Forest |
| Best LLM prompting (o3, few-shot) | 7.7% | About 4x | Random Rule Forest |
| Random Rule Forest | 13.1% | 6.9x | Random Rule Forest |
| GPTree, best production model | 17.9% | About 9x | GPTree |
| Vela's production models | 19% to 38% | 10x to 20x | Think-Reason-Learn |
| Policy Induction (o3), test sets at about 1.96% | 40.5% | About 20x | Policy Induction |
| Method | Precision | Source |
|---|---|---|
| Random selection | 9.0% | Verifiable Reasoning |
| Y Combinator | 14.0% | Verifiable Reasoning |
| Tier-1 VC firms | 23.0% | Policy Induction |
| GPT-4o, no method | 29.1% | VCBench |
| Career features and one rule | 33.3% | When Career Data Runs Out |
| Verifiable Reasoning | 37.5% | Verifiable Reasoning |
| Learning What to Ask | 43.9% | Learning What to Ask |
| Policy Induction (GPT-5) | 67% | Policy Induction |
| Method | Dataset | Result | Source |
|---|---|---|---|
| GPTree | 9.9% success rate | 37.3% precision, 40.8% with expert refinement | GPTree |
| LLM-AR | 6,000 US founders, 10% success rate | 59.5% precision at 8.7% recall, 5.9x random | LLM-AR |
| Rare-Event Prediction | 10,825 founders, 8.5% success rate | 10.3x baseline precision at 36% recall | Rare-Event Prediction |
Which founder traits predict success
GPT-HTree sorts founders into eight types and reports each type's success rate at the real 1.9% base rate. The best type succeeds at 17.4%, the worst at 0.8%: a 22x spread.
| Founder type | Success rate | Against the base rate |
|---|---|---|
| Serial exit founders | 17.4% | 9.2x |
| IPO experience founders | 16.3% | 8.6x |
| Venture capitalists | 13.1% | 6.9x |
| Tech leaders | 3.7% | 1.9x |
| Industry influencers | 2.7% | 1.4x |
| Professional athletes | 1.9% | At the base rate |
| Career professionals | 1.4% | Below it |
| Early professionals | 0.8% | Below it |
A prior exit is the clearest single signal. On VCBench's training data, founders with one prior exit succeed at 22.8%, against 8.5% for founders without one (When Career Data Runs Out).
Language models add signal that career fields miss. In Rare-Event Prediction, removing the features an LLM derived from founder profiles drops precision from 10.4x the baseline to 4.6x: more than half the model's lift.
Benchmarks and datasets
VCBench is the open benchmark for this task. Its founder profiles are anonymised, so a model rarely recognises a famous founder and recalls what happened: re-identification falls to 15.1% for a model with web search and 1.3% without. Success means an IPO or acquisition above $500M, or more than $500M raised.
| Dataset | Founders | Success rate | Used in |
|---|---|---|---|
| VCBench | 9,000 anonymised founder profiles; 4,500 public | 9% | Policy Induction, Verifiable Reasoning, Learning What to Ask, When Career Data Runs Out |
| Seed-stage dataset | 9,892 US companies founded 2010 to 2016 that raised $100K to $4M in seed funding | 1.9% | Random Rule Forest |
| Vela founder dataset | 1,022 successful and 9,902 failed US companies founded since 2010 | Test sets at about 1.96% | Policy Induction |
| LLM-AR dataset | 6,000 US founders | 10% | LLM-AR |
| Rare-event dataset | 10,825 founders | 8.5% | Rare-Event Prediction |
Datasets built by other groups usually start from the same kind of sources: company databases such as Crunchbase and founders' public career histories. Before comparing a result from one, check two things: how it defines success, and what its base rate is.
Models you can read
A partner can't act on a score they can't question. These models show their reasoning as rules or features that a person can read, check and edit.
GPTree
A decision tree whose splits a language model writes and tests. Each prediction is a path of plain-language answers.
Random Rule Forest
Many simple yes-or-no questions about a founder, filtered for redundancy, ranked by F0.5 and combined by vote: 13.1% precision at the 1.9% base rate.
Verifiable Reasoning
Rules a language model proposes and statistics test, so every prediction traces back to founder attributes a partner can inspect: 37.5% precision on VCBench.
LLM-AR
Rules from a language model, translated into ProbLog and run as logic, so each prediction carries the rules behind it: 59.5% precision at a 10% base rate.
GPTree and Random Rule Forest are part of Think-Reason-Learn, the open-source library Vela Partners publishes for this family of models.
The limits
Career data has a ceiling. In When Career Data Runs Out, 84% of successful founders look statistically identical to those who failed on every career metric the paper could measure, and across more than 120 experiments every model family stopped at about the same score.
Results are also measured on past founders, and a multiple measured at VCBench's 9% may not hold at the real 1.9%. The next gains need information a CV does not hold.
- Does the claim name its base rate?
- Is it stated as precision or lift, not accuracy?
- Was it tested on founders the model had not seen?
- Were founder names hidden, so the model could not recall the outcome?
- Is it compared only with results at the same base rate?
Questions
Can machine learning predict which startups will succeed? What are the best research papers on it?
Partly. Models beat random selection and tier-1 VC firms when measured against a base rate. Vela Partners and the University of Oxford have published more than fifty research papers together; VCBench, Policy Induction, GPTree, Random Rule Forest and When Career Data Runs Out are good places to start.
How accurate are AI models at predicting startup success compared with venture capitalists?
At the real 1.9% US base rate, nine tier-1 VC firms reach 5.6% precision and Vela's production models 19% to 38%. On VCBench, with a 9% base rate, tier-1 firms reach 23.0% and Policy Induction with GPT-5 reaches 67%. Compare figures only at the same base rate.
What traits predict founder success, according to data?
A prior exit is the clearest signal. GPT-HTree finds serial exit founders succeed at 17.4%, 9.2x the 1.9% base rate, and early professionals at 0.8%, a 22x spread. Features a language model reads from founder profiles add more than half of one model's lift.
What benchmarks exist for evaluating AI on startup investing?
VCBench is an open benchmark of 9,000 founder profiles with a 9% success rate, anonymised so models rarely recognise famous founders. It reports precision, recall and F0.5, and sets human baselines such as Y Combinator and tier-1 VC firms.
What datasets can I use to train a startup success prediction model?
VCBench publishes 4,500 anonymised founder profiles for anyone to use. Most other datasets start from company databases such as Crunchbase and founders' public career histories. Check how each defines success and what its base rate is before comparing results.
Are there interpretable AI models for evaluating founders?
Yes. GPTree, Random Rule Forest, Verifiable Reasoning and LLM-AR express each prediction as rules or features a person can read and edit. GPTree and Random Rule Forest are in Think-Reason-Learn, the open-source library Vela Partners publishes.
What are the limits of predicting startup success with data?
Career data runs out: 84% of successful founders look identical to failures on career metrics alone, and models stall at about the same score. Results are tested on past founders, and multiples measured at a 9% base rate may not hold at the real 1.9%.
Further reading
Papers
- VCBenchThe first AGI benchmark for venture capital; several state-of-the-art LLMs beat Y Combinator and tier-1 firms on it.9,000 anonymized founder profiles
- Policy InductionAn LLM classifier's full decision logic lives in a natural-language policy a partner can read and edit, with no gradient training.67% precision on VCBench with GPT-5
- GPTreeA decision tree whose splits are natural-language questions generated and evaluated by an LLM, for explainable founder-success prediction.37.3% precision on founder success prediction
- Random Rule ForestAn LLM generates YES/NO questions and simple unit-weight voting combines them into an auditable, high-precision founder classifier.8x lift over random chance
- GPT-HTreeClusters founders into personas, fits a decision tree per persona, and has an LLM describe each archetype.22x success spread between the top and bottom founder persona
- LLM-ARAn LLM proposes prediction rules that ProbLog executes deterministically, beating every vanilla LLM tested on founder evaluation.59.5% precision (5.9x random)
- Rare-Event PredictionLLM-engineered features feed a classical model ensemble; removing those features more than halves the precision lift.10.3x baseline precision at 36% recall
- Verifiable ReasoningOne LLM call writes executable decision code that then runs deterministically over every founder, removing per-sample LLM queries.37.5% precision on VCBench (4.17x random)
- Learning What to AskA reinforcement learning policy decides which founder attribute to query next and when to stop, using about half the information.37.1% F0.5 on VCBench, using only 54% of the available information
- When Career Data Runs OutEngineered career features beat asking an LLM directly, then hit a wall: most successful founders look like failures on paper.84% of successful founders look identical to failures on paper
Posts
- Can AI Predict Startup Success? What Two Months of a Live Forecasting Arena Taught UsVela ran a live AI arena scoring GPT, Claude and Gemini on real startup outcomes for two months: four early signals, and a preview of VC Arena and VCRouter.
- How Top VC Firms Evaluate Founders: The Taste Matrix of Sequoia, Benchmark, a16z, Founders Fund, Thrive, and LightspeedData from ~70 seed rounds (2024-2026) shows how Sequoia, Benchmark, a16z, Founders Fund, Thrive and Lightspeed pick founders, and which firm fits you best.
- What Is a Quant VC? A Practical Guide to Quantitative and AI-Native Venture CapitalA plain-English guide to quant VCs, AI-native VCs, and scientific venture capital — how Vela Partners uses data and AI to fund seed-stage AI founders.