Polymarket Signals

The problem

A prediction market price is a crowd's probability. News moves that probability, but not instantly and not always correctly. The gap between what the news says and what the price says is where an edge would live, if there is one.

Finding it by hand means reading a lot of articles and doing arithmetic in your head, which is exactly the combination people are worst at and most confident about.

What I built

A toolkit that pulls the price history for a market from the Polymarket order book, aggregates news from eight always-available feeds (BBC, the Guardian, Al Jazeera, the New York Times, NPR and government sources) plus two optional APIs, filters by keyword and date, and has a model read each article for direction, confidence and a one-sentence reason.

Those scores become an estimate, the estimate is compared to the market price, and the difference goes through an expected-value calculation with a Kelly sizing suggestion. The chart plots every scored article against the price line, so a claimed edge can be traced back to the specific stories that produced it.

The design decisions

It does not trade. The README says so in the second line. It produces analysis and signals; the decision and the click are yours. A tool that both forms an opinion and acts on it removes the one step where a human can notice it has gone wrong.

It scores with more than one model. Gemini, Claude and GPT-4o can read the same article and disagree, and an ensemble that disagrees is more informative than a single model that sounds certain. Sources also carry their own reliability weighting, so a wire report and a blog do not count equally.

It is happy to say no. The default answer is PASS: no actionable edge, stay out. The screenshot above is a real market where the alpha came out at minus 1.8 per cent against a 10 per cent floor, and the tool's entire output is a recommendation not to bother. Most markets look like that, and a signal tool that always finds something is not finding anything.

It backtests. A separate command measures how the scoring actually performed against resolved markets, because the alternative is trusting a pipeline that has never been checked against an outcome.