Automated Investment Broker: Research Agents Debate, Correlate, and Copy Prediction Market Elite Forecasters

Video Overview: https://youtu.be/N5cHK6K50Ds

Monthly trading volume on prediction markets reached approximately $13B by late 2024 (Reuters, 2024), but existing platform rankings do not distinguish forecasting skill from chance. We describe a system that combines multi-agent LLM reasoning with prediction-market leaderboard data to rank forecasters by calibrated skill. The platform employs a five-layer architecture: (1) specialized analyst teams for fundamental, sentiment, technical, and macro analysis; (2) adversarial debate agents with bull-bear dialectics; (3) calibration engines using Brier scores and time-weighted metrics; (4) prediction market integration across Polymarket and Kalshi; and (5) a Forecaster Trust Score that combines per-trade quality with statistical confidence. We validate our approach through backtesting on Bitcoin price predictions and cross-market correlation discovery. The system achieves 60-70% accuracy in relationship discovery across prediction markets, with LLM ensemble predictions matching human crowd accuracy. In our sample, forecasters in the top Trust Score quintile showed lower Brier scores than the bottom half (0.18 vs. 0.32); the 30% figure is reported from prior superforecasting literature (Mellers et al., 2015), not measured here. We propose a metric for distinguishing skill from noise in forecaster track records and evaluate it on cryptocurrency prediction-market data.