Artículos Relacionados con Agent-Based Modeling

Is Polymarket's Pricing Wrong? 200 AI Agent Simulation of Crisis Yields Unexpected Answer

An experiment used MiroFish, an open-source multi-agent simulation platform, to model the geopolitical crisis in the Strait of Hormuz and compare the results with Polymarket's prediction market. The system generated 200 AI agents—including government officials, media, energy firms, financial traders, and civilians—and simulated 7 days of social media interaction (Twitter-like environment) based on a 5,800-character background brief. Key findings: - Organic, free-form discussions among agents produced an average probability of 47.9% for the strait reopening by April 2026, significantly higher than Polymarket's market-derived probability of 31%. - When agents were individually questioned in a formal "interview" setting, they converged to overly optimistic responses (60–75% across categories), reflecting a cooperation bias. - The most accurate predictions came from a minority of pessimistic agents (e.g., Iranian officials, financial analysts, academics) who organically expressed probabilities near 22%—aligning closely with market pricing. - The simulation revealed a structural divide: public/official statements tend toward optimism, while genuine risk assessments emerge from unstructured, adversarial discourse. The study suggests that natural interaction among specialized agents can generate valuable signals, but LLM bias and limited context remain constraints. Future work will expand data scope, use stronger models, and increase agent diversity.

marsbitAyer 06:16