

- Role
- Solo developer
- Year
- 2026
- Status
- Prototype, dry-run by default
- Kind
- AI tool
An AI pipeline that reads world news every 15 minutes, matches it to open geopolitical markets on Polymarket and makes its own probability estimate for each. It flags markets where its estimate and the market price disagree.
The idea
Prediction markets put a price on world events. I wanted to see whether language models, given fresh news and checked against their own track record, could find prices that look off.
The bot focuses on geopolitical markets, which Polymarket lists without trading fees, and on prices in the 55 to 85 percent range.
How it works
Every 15 minutes the pipeline pulls news from GDELT, which covers sources in more than 100 languages. Free models match articles to markets, break each market question into sub-questions and summarise the evidence. One paid model, Claude Sonnet, makes the final probability estimate.
That estimate goes through Platt calibration, trained on markets that have already resolved, before it is compared with the market price. A large enough gap becomes a signal.
Guardrails
A risk engine checks daily loss, total exposure, number of open positions and a balance floor before anything is placed. There is a kill switch. Dry-run is the default: the bot records every signal and simulates trades without placing real orders.
A React dashboard shows the portfolio, open positions, the signal feed and calibration. A Discord bot posts alerts and a daily summary. 63 automated tests cover matching, forecasting, calibration, risk and execution, with every model call mocked.
Next project

FilmAxis
From match to movie