Building a prediction-market agent: when should it trust its 34.5% over the market's 20%?

Dry-Comfortable-2678 · reddit · 2026-11-10

The author is building Mantua, an AI agent researching NFL prediction markets, and shares a guardrail design problem: when its estimate diverges far enough from market price to look like an opportunity, what should it require before acting?

In Jaguars vs. Ravens, the market implied 20% for Baltimore while Mantua said 34.5% — a 14.5-point gap. The agent still favored Jacksonville and recommended waiting; Jacksonville won 33–10. But the author notes one result proves nothing, and an early forecast review revealed a tendency to pull favorites toward 50%, which could make underdogs look cheap due to model bias rather than mispricing. The proposed guardrail: show the gap, identify which inputs explain it, and recommend waiting when the gap could plausibly stem from known model bias. The post asks whether such agents need independent explaining evidence, a track record on similar forecasts, or both.

Original post →

Building a prediction-market agent: when should it trust its 34.5% over the market's 20%?

More from coding & agent