Free data service · hourly

Polymarket News Index

What the news says about every outcome of the 200 most traded Polymarket events, next to the price, every hour, with the articles behind each move.

Price against the news

What the news says, next to what the market prices

For each market, a language model reads the day's articles against every outcome. Orange is what the news implies, blue is the price; the chart shows the two over time. Pick a list, or shuffle.

Among the 40 most traded markets, the ones where the news has followed the price most closely.

spinfspinfMade with spinf

The engine that reads the news on every market is open as an API

spinf reads a document once and scores every question you ask of it, at a fraction of the usual cost. In beta: try it on your own documents.

Try spinf
Any market
Today, so far

What moved

Ready-made lists from the hourly release: where the news moved most, where it disagrees most with the price, what the press covered, what is about to resolve. Provisional until the day's final release, after 01:20 UTC.

The paper

Nine months, 257 markets, 620k articles

We replayed 2026 from January to October before going live. The paper has the method, the sample questions, the full distribution, the top ten and the rankings, what did not work, and the threats to validity.

38%

1% by chance

of markets with 12+ weeks of data where the news alone correlates with the price above 0.5 (20% above 0.7)

73%

41% by chance

of market-days where the model, reading the articles, puts the market’s favourite first (70% with no article: its bias)

78%

50% by chance

of outcome pairs ranked by the news in the market’s order

+0.074

+0.026 null

partial correlation of the news with the next day’s price change, beyond the market’s own reversion (t = 10.7)

The spread between markets is the most interesting result: the best track at 0.90–0.94, and 8% correlate negatively. It is not explained by the smoothing window, the article volume or the category alone. The reader model matters too: on the same articles, Qwen3.8 27B follows the price twice as closely as Gemma 4 12B, and neither predicts it better. The questions were written, and the reader chosen, after the period they are tested on, so the live feed, running since 3 October 2026, is the out-of-sample test.

Read the paper (PDF)

Found an error, or using the data? Write to info@thinkingtext.com.

Two markets over 2026: the price of a US–Iran deal clause (diluting Iran’s enriched uranium) and of the Georgia Senate race, each with the news read alone, which follows it closely.

The two closest markets. Blue: the price. Orange: the news alone, from the day’s articles with no market data, rescaled to the price’s range. From the paper, Figure 1.

Methodology

How it works

Four steps. The price enters only at the end, where the news is compared with it; the news line itself never sees a price.

  1. 01

    Pick the markets

    Every day, the 200 open events of both Polymarket exchanges with the most traded dollars over 30 days (from the trade tape), sports and automated markets aside. Each becomes one question with mutually exclusive outcomes, priced from its order book.

  2. 02

    Write the query, once

    Claude drafts each new market’s question, outcomes and keyword filter in about a minute; automatic checks test the prices, the filter’s volume and a sample of its articles. The queries are not published.

  3. 03

    Read the news

    English articles from 2,654 outlets (GDELT), first 260 words, de-duplicated. An open-weights 27B model (Qwen3.8 27B, via spinf) reads each one three ways against every outcome, each against the same prompt with an empty article.

  4. 04

    Compare, every hour

    Per outcome: the news read and its smoothed line, a news-implied probability next to the price, the order of the outcomes, and rule-based alerts with links to the articles behind each move. Final values once a day.

Questions

Four things people ask

Every day, the articles that a market’s keyword filter lets through (about ten a day for the median market, from GDELT’s English outlets) are read by an open-weights 27B language model (Qwen3.8 27B, via spinf) against the market’s outcomes: which outcome the article supports, whether it is this outcome, whether the chance goes up or down. Each read is taken relative to the same prompt with an empty article, so the model’s own prior cancels. The reads are averaged per day and smoothed; no price goes into them.


We do not know yet, and it is the question we find most interesting. The match is not explained by article volume, category or the smoothing window alone. Some markets are priced on information the press does not carry (polls, order flow, insiders); in others the filter lets through articles about the wrong thing. Every market carries its own track record (correlation, the same against unrelated markets, how often the favourites agree), so you can judge each one rather than the average.


Pooled over every market, the news carries a small forward signal on the next day’s price change beyond the market’s own reversion, and it grew through the year. We did not test a trading strategy, and the questions were written after the period they are tested on: the live feed is the out-of-sample test.

The index is provided for information and research purposes only. It is not investment advice, not a recommendation to trade any contract, and not a forecast. Anyone using these data does so at their own risk.


Each market’s question, outcomes and keyword filter are drafted once by Claude, checked automatically (prices, filter volume, a relevance sample) and revised when a market changes; a market that is revised gets a new version and a full new history. A drafted market matched or beat our hand-built ones. The per-market queries are not published; the data computed from them is, every hour.

Access

Open data, every hour

Everything on this page is a static file: the market list, each market's state and daily history, the alerts, daily, weekly and monthly files and top lists. They live on the data API page, free with a key, with an agent prompt, curl and Python examples.

Free with an API key

Register once with a name, a company and an email; the key comes by email and opens every file, from the browser, a script, a spreadsheet or an agent.

latest.json · manifest.jsonmarket list · state · historyalerts · jsonldaily · weekly · monthly · csv, json, parquetSCHEMA.md
Get the data

Polymarket News Index is news analysis provided for research and entertainment only. It describes what the press is saying next to what the market prices; it is not a forecast, not investment advice and should not be relied upon for trading decisions. Not affiliated with Polymarket. See the terms of service.