About the Forecasting Agent
An AI forecasting system that analyzes online prediction markets, identifies mispriced opportunities, and paper trades them autonomously.
Multi-Step Reasoning Pipeline
Each scan is one pass through the flow below: a planner picks the research tools, and while they run, four web research loops each search and read pages on one side of the question (current state, base rate, the case for yes, the case for no). Five independent forecast draws then read all of that research unsummarised, and the forecast is their plain mean. When the draws disagree, or land far from the market's own price, one more loop researches the crux and the draws are redone. Every stage's inputs and outputs are kept as a trace, which is what the forecast pages show.
The Models
The pipeline mixes two Claude tiers and one GPT model. The frontier Claude slot handles the steps where judgment matters most: planning, three of the five forecast draws, and the trade decision. The fast Claude slot runs the search-and-read research loops. A GPT model writes the other two draws, so the forecast averages two model families rather than samples of one, and no model edits the number afterwards.
Paper Trading
All trades are simulated, no real money is at risk. The system tracks every trade as if it bought $1,000 worth of contracts at the market price when the edge was identified. P&L is calculated based on how the market moves after entry: if the agent bought YES at $0.40 and the market moves to $0.60, that's a $500 unrealized gain on $1,000 deployed. When the market resolves, winning YES contracts pay $1.00 each and losing ones pay $0.00.
Team
The forecasting agent is a research project from the Agentic Learning AI Lab at New York University.
