Startup Success Forecasting Framework (SSFF): A Multi-Agent Framework for Startup Success Prediction

Xisen Wang, Yigit Ihlamur, Fuat Alican

Keywords: Startup Success Prediction, Venture Capital, Multi-Agent Framework, Machine Learning

TL;DR: We show that vanilla LLMs systematically over-predict startup success, and introduce SSFF—the first ML-based multi-agent framework for venture evaluation—that achieves up to 6× higher accuracy with structured assessments.

Abstract: Predicting startup success is highly uncertain and has long relied on VC intuition. While LLMs promise new capabilities, our experiments reveal a fundamental failure mode: when applied directly, LLMs systematically \textit{over-predict}, yielding poor precision under severe class imbalance. Hence, we introduce the \textbf{Startup Success Forecasting Framework (SSFF)}, the first multi-agent architecture for early-stage venture evaluation. By combining LLM-enhanced prediction (random forests, founder–idea fit), multi-agent analysis (segmentation, evaluation), and retrieval-augmented market intelligence, SSFF delivers structured, interpretable assessments. On a realistic dataset (10\% success rate), SSFF reduces optimism bias with nearly six-fold performance than GPT baselines. More broadly, SSFF serves as a generalizable template for integrating LLMs with conventional ML in high-stakes, imbalanced decision-making tasks.

Submission Number: 75