The Strategic Foresight of LLMs: Evidence from a Fully Prospective Venture Tournament

TL;DR

This study asks whether AI can outperform humans in strategic foresight by conducting a fully prospective, field-based benchmark using live Kickstarter campaigns. A diverse set of frontier LLMs and human forecasters compete in a double round-robin task, predicting fundraising success before outcomes are known. Frontier LLMs achieve substantially higher predictive accuracy (e.g., up to $\rho = 0.74$) than humans (max $\rho \approx 0.45$), with several models also delivering superior pairwise accuracy and value capture; aggregation provides limited or no gains over the best model. The results support AI-assisted foresight in external-market, prediction-heavy tasks and introduce a robust benchmarking template for evaluating strategic prediction, while highlighting the importance of inputs, framing, and execution capabilities as future sources of advantage.

Abstract

Can artificial intelligence outperform humans at strategic foresight -- the capacity to form accurate judgments about uncertain, high-stakes outcomes before they unfold? We address this question through a fully prospective prediction tournament using live Kickstarter crowdfunding projects. Thirty U.S.-based technology ventures, launched after the training cutoffs of all models studied, were evaluated while fundraising remained in progress and outcomes were unknown. A diverse suite of frontier and open-weight large language models (LLMs) completed 870 pairwise comparisons, producing complete rankings of predicted fundraising success. We benchmarked these forecasts against 346 experienced managers recruited via Prolific and three MBA-trained investors working under monitored conditions. The results are striking: human evaluators achieved rank correlations with actual outcomes between 0.04 and 0.45, while several frontier LLMs exceeded 0.60, with the best (Gemini 2.5 Pro) reaching 0.74 -- correctly ordering nearly four of every five venture pairs. These differences persist across multiple performance metrics and robustness checks. Neither wisdom-of-the-crowd ensembles nor human-AI hybrid teams outperformed the best standalone model.

Figures (11)

• Figure 1: Spearman's rank correlation ($\rho$) between predicted and realized project rankings, with 90% confidence intervals. The dashed line at zero represents chance performance; human evaluators are shown in blue with diamond markers.
• Figure 2: Pairwise comparison accuracy, with 90% confidence intervals. The dashed line at 50% represents chance performance; human evaluators are shown in blue with diamond markers.
• Figure 3: Scatter plots of predicted versus actual project rankings for the best LLM (Gemini 2.5 Pro), the best human (Expert 3), and the Prolific crowd ranking. Lines show linear fit.
• Figure 4: Spearman correlation matrix among all evaluators. The rightmost column shows each evaluator's correlation with actual funds raised.
• Figure 5: Percentage of top-5 project value captured by each evaluator's top-5 predictions, with 90% confidence intervals. The dashed line at 18.9% represents expected capture under random selection; human evaluators are shown in blue with diamond markers.