How it works
Two things happen behind this site: articles are grouped into stories, and stories are matched to markets. Both are done by software, and both can be wrong in ways worth understanding.
Grouping articles into stories
Every few minutes we fetch the feeds of several hundred publishers. For each new article we take the headline and the short lead the feed provides, and turn them into a list of numbers that represents their meaning — an embedding. Articles about the same event produce similar numbers.
A new article joins an existing story when it is close enough to that story's average, and was published within two days of it. Otherwise it starts a story of its own. "Close enough" is a threshold we tuned by hand against a test set of headlines, including deliberately tricky ones: two reports of the same Fed decision should merge, while an Apple product launch and an apple harvest should not.
That threshold is a trade-off, not a solved problem. Set it high and one story splits into three; set it low and two stories merge into one. Sports needs a higher bar than the rest, because sports writing is formulaic enough that unrelated previews look alike to the maths.
How AI is used here
Every headline, summary, section, list of named people and organisations, and the one-word read on how the coverage frames the outcome (settled, expected, contested or speculative) is written by a language model, not by a person. We run an open model (Qwen) on our own hardware. Once a story has two or more publishers, the model reads the headlines and short feed leads of the articles in it and writes a neutral headline and a one-sentence summary. It is told to report only what those articles say and to add no context of its own. A separate embedding model groups the articles in the first place.
What it does not do. It does not write articles, browse the web, or reproduce article text; we keep only each article's title, a short lead from its feed, and the link. The author of every page is NewsMarkets itself. There are no editors behind it and no invented bylines. Prices come straight from the exchanges, and the odds, charts and rankings are computed from data, not written by the model.
How it is checked. Only by software. The model's reply must fit a fixed format with length limits. A headline is rejected if it contains words found in none of a standard dictionary, recent headlines, or the story's own articles, and one written all in lower case is rejected too. No person reads a headline or summary before it is published. How often they are wrong is measured on a random sample, graded by independent AI auditors and not by people, and published with its uncertainty on how accurate NewsMarkets is. Treat a summary as a pointer to the reporting, and check the linked articles. The headline and summary are also what appears in search results and link previews, so the same caveat applies there.
Corrections. If a story is wrong, merged with another, or matched to the wrong market, flag it from the story page; fixes are listed in the corrections log. The markets matched to a story are also chosen by software, so read the market question itself. You can also write to newsmarkets@simbasystems.com.
Ranking
A story's position is mostly how many separate publishers are carrying it, damped by how long ago it was last updated. Major outlets count for a little more, and a story carrying a live market gets a small lift. A story from a single publisher can appear, but it will not lead the page.
Matching markets
We compare each story against every open market on Polymarket and Kalshi the same way we compare articles to each other. A match needs either a strong similarity, or a moderate one plus at least one named person, organisation or place that the story and the market question share. Markets that have already resolved are excluded, checked against each exchange's own API rather than a cached list.
Matches are imperfect, and more so than that sounds. In a random sample graded by independent AI auditors, most links were to a market about the same person, team or topic as the story but asking a different question from the one the story reports; only a minority asked about the very event the story covers, and a few were unrelated. The numbers are on how accurate NewsMarkets is. Read the market's own question rather than trusting the pairing, and read the analysis pages as showing what moved on a related market, not proof that it reacted to the story.
How current is it
Feeds are polled on a cycle of a few minutes, with quiet feeds checked less often and busy ones more. Market prices are refreshed each cycle for every market shown on the site.