How it is calculated

RSS feeds from crypto domains, wires, newsrooms and project blogs, mostly in English. Each universe's page says how many domains it reads today. Each feed is polled on its own schedule; a feed that has failed for more than a day is switched off.

We keep the headline, link and date, and for some domains a snippet of at most 200 characters. Article text is never reprinted; the link goes to the domain.

Each narrative shows the headline and a link back to the domain that published it. Where a domain allows it, we also show a short snippet of at most 200 characters; where it does not, we show the headline alone. Article pages are read only to rate tone and are not republished.

Each domain has its own setting: headline, snippet and link; headline and link only; or not shown at all. Before collecting from a domain we read its robots.txt and stop when it asks us not to.

If a publisher asks us to remove its articles, we stop collecting from it and hide what we hold from every page. We can also stop all collection at once.

Coin logos are copies of marks published in open-source repositories under the CC0, MIT, ISC or Apache-2.0 licences, and are served from this site. Each logo is the mark of its coin's project, shown only to identify the coin; it implies no endorsement.

The source, licence and retrieval date of every logo are listed in /coins/SOURCES.md. A coin with no logo we may reuse shows the first letters of its ticker instead.

A narrative is a set of articles from different domains about the same news. Articles are grouped automatically by content; one domain repeating itself never makes a narrative.

Narratives are sorted into topics; a tile's size on the home page is the number of narratives in its topic.

The short “What's happening” summary on a topic page is written by a language model from the headlines, domain counts and tone of the topic's most-covered narratives alone. It is checked every two hours and rewritten only when those narratives have changed enough; it adds nothing the headlines do not say.

Attention is how loudly a narrative is being told in the news right now, as a number from 0 to 100. It measures news coverage only: not a price, not trading volume, not social media buzz. The narratives in the table are ordered by it.

Three things go into it. Reach: every article about the narrative counts, but an article from a heavier domain counts for more than one from a lighter domain, and a fresh one for more than an old one; an article's weight halves about every 5.5 hours. Spread: how many different domains carry the narrative, because a narrative many domains picked up has travelled and one nobody repeated has not. Speed: whether coverage is picking up or fading, from the articles of the last 6 hours against the 18 hours before them.

Reach leads; spread and speed adjust it and cannot replace it, so a narrative mentioned twice never reads as a strong signal. In numbers: ln(1 + reach) × (1 + 0.18 × spread + 0.18 × speed). Spread is the natural log of the number of domains carrying the narrative; speed is the log of the ratio between the hourly article rate over the last 6 hours and over the 18 hours before (each plus one), held between −0.6 and 1.2. The result is placed on the scale so that 4.5 reads as 100, and nothing is shown below 0 or above 100. The Trend column (rising, steady, fading) shows the speed part on its own.

The scale is fixed: a constant reference level counts as 100, not the busiest narrative of the day. So the same activity scores the same today and tomorrow, and the scores of different days can be compared. The constant is a calibration choice rather than a measurement, so even the busiest narratives usually score below 100.

Share of the day's attention, shown in the feeds and the API, is a different number: a narrative's reach divided by the reach of all narratives in the last 24 hours. The score says how loud a narrative is; the share says how much of everything being read it takes, which tells narratives apart when the whole news cycle is loud and everything scores high.

The score describes how much the press is writing about a narrative, not whether the news is good or bad and not what will happen to prices. It is not investment advice.

A language model labels each English article positive, neutral or negative, by how it frames the news for the market.

It does not guess: if an answer fails validation, the article stays unlabelled and does not count towards tone until it is rated again.

A net tone index from 0 to 100 over the last 24 hours: 50 + 50 × (positive − negative) / (positive + negative). Neutral articles are left out, so 50 means positive and negative balance, 100 that every rated article with a direction is positive and 0 that every one is negative.

Zones: 0–24 strongly negative, 25–44 negative, 45–55 neutral, 56–75 positive, 76–100 strongly positive. It is compared with the usual level, the mean of the daily index over the last 30 complete days. It describes how the press writes, not what to do.

For each domain we take its articles in the universe over the last 30 days, only those the model filed under one of the universe's topics, and compute the same net tone index: 50 + 50 × (positive − negative) / (positive + negative). Several feeds from one domain (rss.nytimes.com and nytimes.com, say) are added together.

A domain is ranked once it has at least 50 rated articles in the period. While collection is younger than 30 days and too few domains reach that, the floor steps down, never below 20 articles; the floor in use and the real period are always stated under the table.

The ranking describes the tone of the articles as the model reads them, not a domain's quality, reliability or bias: one that writes more about rallies and launches scores higher, one that writes more about hacks and lawsuits scores lower.

When it rates tone, the model also names the countries an article is about: the main subject of the article or the place where it happens, up to three per article, as ISO 3166-1 codes. No country is set when one is mentioned in passing, when the news is about the market as a whole, the EU or a whole region, or, for crypto, when the narrative is not about that country's regulator, a law, a court or a company under its rules.

For each country we compute the same net tone index over the universe's articles the model filed under one of its topics. An article about two countries counts for both. Countries with fewer than 10 articles in the period are hatched: there a single article weighs too much. Countries are recorded from the map's launch day; older articles have not been re-rated.

Territories without an ISO 3166 code are never coloured.

The BTC correlation in Pro is the Pearson coefficient between the tone index and BTC's price change over the selected period, with the tone shifted up to a day either way (up to a week, by the day, over 90 days). We show the shift at which they match best. With too few shared points it reads "not enough data"; when |r| is under 0.2, or no more than chance gives on a sample that size, it reads "next to no link". It describes the past, not a price forecast.

Prices come from the public candles of Binance, Binance.US, Coinbase and Kraken, read every five minutes, with the first exchange that answers serving each coin. No account or paid feed is involved.

Each price is checked against a second exchange. A price with no new tick for 15 minutes is stale. A price that has stood still for six ticks or more while the second exchange moved by more than 0.5% is frozen. Either shows as 'Price delayed' and the coin is not rated until it recovers. A price more than 1% from the second exchange is recorded for review but still shown.

Volume on public exchanges is far below what a paid market-data feed sees, so the figures here are for comparing coins with each other, not for trading.