Can language models beat human venture capital experts?

Venture capital is a clean testbed for expert forecasting under uncertainty: signals are sparse, outcomes uncertain, and even top investors perform modestly in absolute terms. At inception the market index achieves only 1.9% precision; Y Combinator reaches ~3.2% (1.7× the index) and tier-1 firms ~5.6% (2.9×). VCBench standardizes 9,000 anonymized founder profiles (with adversarial tests cutting re-identification risk >90% while preserving predictive signal) and evaluates nine LLMs. Several surpass the human baselines — DeepSeek-V3 delivers over six times the index precision, GPT-4o achieves the highest F0.5 — and most models beat the human benchmarks.

The keeper is the structural point about where LLMs win: in low-base-rate, sparse-signal forecasting, modest absolute accuracy can still beat expert humans because the human bar is itself modest. This reframes "can AI match experts?" — in domains where expertise yields only a small edge over chance, the bar to exceed experts is correspondingly low, and anonymized profile features alone suffice.

This complements the vault's forecasting thread. Since Can LLMs actually forecast time series better than we think?, VCBench supplies the domain where even raw model capability clears a low human bar; and it extends Can retrieval-augmented language models forecast like human experts? — that result nears the competitive-crowd bar; VCBench shows that where the human bar is modest, LLMs clear it outright.