VCAF: A Multi-Agent Framework for Venture Capital Decision-Making Using Synthetic Startup Data

Venture capital (VC) investment decisions rely heavily on evaluating early-stage startup data, which is frequently sparse, incomplete, or proprietary. To address this challenge, we introduce \textbf{VC-Data}, a synthetic dataset comprising 158 startup profiles generated using a multi-step large language model (LLM) pipeline with human validation, alongside the \textbf{Venture Caption Agents Framework} (VCAF), a multi-agent decision-making system powered by Claude-3.7-Sonnet. When evaluated on VC-Data with complete information, VCAF achieves 74.05\% accuracy and an 80.56\% F1 score, surpassing baseline human VC performance. The framework provides a systematic backtesting approach for venture capital analysis while generating interpretable investment recommendations that capture the nuanced, qualitative factors critical to early-stage investment decisions.