AI Models Surpass VCs in Predicting Startup Success
- Oxford and Vela researchers introduced VCBench, a benchmark to test AI’s ability to predict startup success.
- Large language models like GPT-4o and DeepSeek-V3 outperformed Y Combinator and top venture capital firms.
- The study used a dataset of over 9,000 anonymized founder profiles, with about 810 labeled as successful.
- DeepSeek-V3 achieved more than six times the precision of the market index, while GPT-4o had the highest F0.5 score.
Researchers found that large language models could potentially become essential tools for early-stage investing by improving hit rates beyond traditional methods. The study highlights the potential for AI to identify promising startups without relying on conventional networking approaches.
With models like GPT-4o outperforming established venture firms, this research indicates a shift towards more data-driven investment strategies in identifying successful startups early on. (Source)
