Equity News Index
What the press says about a stock, read every day by a language model through the questions that have moved with its share price for five years, standardised and averaged into one number per name. Sixteen Nasdaq-100 names tracked, nine steady enough to be the index, all published next to the day's move with a plain account of how much of it the press explains. A description of the news flow, not a forecast.
The press movers
Every published name on the last complete press day: how the press reads, how far it moved in a week, and what the share price did. A new final value every night; pick a name to see the detail.
Made with spinfThe engine that reads the company press 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.
What the press is saying about nvda
Any of the sixteen tracked names. The read of the last complete press day, its parts, the move it accompanies, and the slot families behind the number.
The attribution cards for nvda
For the last market day: the actual move over 1, 5 and 20 days, the move implied by the change in the read, and the split between the slot families, the peers and the market read. Each card carries its own hit rate from the back-analysis and, once live, from the published record.
The read against the price of nvda
Weekly change or level, any year since 2022, with the company and market parts on request. The correlation shown is computed on the displayed window, from the same rows as the data files.
Change: EQ-S10:NVDA against NVDA
Loading the history…
The k-day change of the index against the k-day return. The steady cells were selected on this read at 5 days; the betas of the cards are fitted on it. The read on a market day is the press of the same UTC day.
Where it works, and where the press does not cover the business
The same read for every name we built, published or not, by year and pooled; and the coverage behind it. The twelve published names are the ones that clear the bar in every year.
How well it explains nvda
The numbers behind the claims for the selected name: fitted and held out, by horizon and by year. They describe the press; the read has no lead on the price, and the residual is large by construction.
How it works
What a reader needs to interpret the numbers ships with the data (spec.json). Four steps; the only one that looks at the price is the selection of the questions, and it is frozen.
- 01
Collect
Every English article of about three hundred selected outlets, from public feeds (GDELT) and private newsfeeds, de-duplicated, that names the company and its business. More private feeds are being added. From sixteen to five hundred articles a day per name; the six automated stock-note sites are excluded.
- 02
Read
An open-weights 12B language model reads each article once against the company’s questions: demand, profits, outlook, competition, regulation, costs, the share price, results against expectations, the market and the economy. No event extraction, no price language; the scoring itself is our own and not published.
- 03
Keep what is steady
A question stays in the index when its weekly change moved with the weekly return in every year from 2022 to 2026. The rule is the same for every name; on returns from the wrong year it keeps nothing. Frozen with the spec.
- 04
Aggregate and relate
Answers are combined by day, standardised on the question’s own trailing year, averaged within each slot family and across families; two thirds company, one third the general market read. One rolling beta on the weekly change gives the news-implied move; the parts split it exactly.
Two indexes, two hundred thousand answers a day, one machine
Both indexes run on an inference engine we built for this job. It reads an article once and answers every question asked of it from that single read, so the same open-weights model delivers about eighteen thousand tokens a second on two consumer graphics cards, fifteen times the prompt volume it would take to ask each question on its own. That is what lets sixteen company reads ride on the oil read for a minute of GPU a day, on one box, without a data centre.
Four things people ask
We built the sets for sixteen of the largest names and kept the twelve where the press writes about the business steadily enough for a company read: the index of every published name tracks the weekly return at 0.37 or better, held out, with no year below 0.2. The others are computed and listed with their reason:
- Thin coverage: Micron, Palantir and Cisco make it on sixteen to twenty-two articles a day; Costco does not. The English press we read simply does not cover it steadily.
- Well covered, wrong subject: Netflix is written about every day, but the press writes about the shows, not the business; Walmart reads mostly as the market.
- No price yet: SpaceX listed in June 2026, so there is no history to test a read against.
The rule is the same for every name and it is frozen with the spec: no name is added or dropped without a new version.
Because a third of a large stock's weekly move is the market's. The general read is the set of questions about earnings, rates, risk appetite and the economy, put to the whole news flow, that tracked the Nasdaq-100 steadily in every year since 2022; it is the same for every name and it is published on its own against QQQ. On the weeks when a stock parts ways with the market, the company read and the market read disagree, and the page shows both so a miss reads as a conflict rather than as noise. The chip names also carry a read of their peers at 15 %.
Not as a signal. The index describes what the press wrote on the day, and every slot peaks at lag zero: there is no lead on the price at any horizon we tried. It explains after the fact, and it says how much it does not explain. What it is good for:
- Post-fact attribution: which part of the story the press carried on a big week, and how much of the move was not in the news at all.
- Context: where the press has the company against its own last two months, next to where the price is.
- A baseline for surprises: a fixed read of the whole press against which a single headline or a broker note can be measured.
- A benchmark for news models on single stocks, free and fixed, that reports its own hit rates.
EQ-S10 is provided for information and research purposes only. It is not investment advice, not a recommendation to buy or sell any instrument, and not a forecast of prices. Past relationships between the index and prices do not guarantee future ones. Anyone using these data does so at their own risk.
Same reader, same engine, same daily job and the same shape: questions, standardised on a trailing year, averaged under a few drivers. OIL-F6 uses six fixed fundamentals with textbook signs and no selection; EQ-S10 keeps, inside ten slot families, the questions that moved steadily with each share price over five years, because a company's press has no textbook. The S says so.
Open data
Everything on this page is a static file: the history by year and by name, the latest objects, the spec, the summary and the back-analysis numbers. They live on the data API page, free with a key.
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.
Equity News Index is news analysis provided for research and entertainment only. It describes what the press is saying; it has no forecasting power, is not investment advice and should not be relied upon for trading decisions. See the terms of service.