Anatomy of the flow: two years of crypto headlines, described

1. Why a note about the data

The first and second notes measured what scored headlines know about price. The third checked a trading bot against its simulator. All three stand on one dataset, and none of them says what it is. A reader deciding whether the results transfer to their own work, or whether the data is worth having, needs the plain description first: how much, from whom, when, about what, in what form, how fast, and with which defects. That is this note. It makes no claim about returns.

Everything below is computed from the metadata of each record. Headline texts are not reproduced, and the data files that accompany the note are aggregates: counts by month, by source, by hour, by coin, by score.

2. What a record is

The collector reads every outlet that writes about crypto that I have been able to reach and keeps one record per headline, deduplicated by source and address. Each record carries:

• the time of publication, as stated by the source;
• the time the record entered the database, which is the moment the headline became usable;
• the source, by domain;
• the coins the headline is about, one or several;
• a type, reactive when the headline reports something that has happened, proactive when it announces something that will;
• a score from −1 to 1, the tone of the headline for the coins it names; most scores sit on steps of 0.1 and 7% on half-steps, 63 distinct values in all.

The score and the type are assigned when the record is created, within minutes of publication. The window of this note is the two years of whole months from October 2024 to September 2026: 230,270 scored headlines.¹The database holds 231,296 scored headlines as of 3 October 2026 and grows by a few hundred a day; the live count is on the About page.

3. How much

The flow is 317 headlines a day on average, 336 in the median, and eight days in ten fall between 146 and 460. The quietest day of the two years had 91, the loudest 642. By month the count ranges from 7,616 to 13,809; 2025 averaged 10,200 a month and the first nine months of 2026 averaged 8,400.

The step in June 2025 is the collector, not the market. Before it about 49 sources contributed at least 20 headlines in a month; after it about 35 do. The change was not only in the tail: 36 sources that had supplied 36% of the earlier volume stopped, others grew to fill the gap, and only two of the ten largest sources before June 2025 are among the ten largest after. The flow is a different mixture of outlets on the two sides of the line. Anyone comparing volumes or tone across the boundary should normalise by source and by coin, which is what the indices of the first two notes do.

Sources also change on their own. In the last week of September 2026 one outlet multiplied its output tenfold from one week to the next, and the week's count jumped by half; that is what a feed change at a source looks like in the data. A count index that is not normalised per source would read it as a burst of attention across the market.

4. Who writes

There are 109 sources after merging the variants of a domain, 83 of them with at least 100 headlines and 42 with at least a thousand. Concentration is moderate: the largest source supplies 5.2%, the ten largest 44%, the 25 largest 79%. No outlet dominates, which matters for the index, because it means no single editorial desk sets the tone of a coin's window.

Table 1. The sixteen largest sources: share of all headlines, share of their headlines that are proactive, mean score, and median minutes from publication to the database since June 2025. The full table for the 42 sources with at least a thousand headlines is in the data files.

Sources differ in character more than in size. The proactive share runs from 12% to 35% among the sixteen largest: some outlets cover the tape, some cover announcements. The mean score runs from 0.13 to 0.26. These are stable traits of a source, and they are one reason the index normalises tone within a coin rather than across the whole flow: a coin covered mainly by announcement-heavy outlets would otherwise look permanently cheerful.

5. When

The market trades around the clock; the press does not. The busiest hour is 13:00 UTC with 6.8% of headlines, the quietest 01:00 with 1.9%, a ratio of 3.6. The nine hours from 08:00 to 17:00 UTC carry 52% of the flow; the seven from midnight to 07:00 carry 17%. Saturday and Sunday together carry 15% of the week, against 17% for each weekday.

This shape has a consequence for anyone building a signal on headline counts. A 12-hour window ending at noon UTC contains the European morning; the window ending at midnight contains the American afternoon and the Asian night. The two are not comparable in volume, and a count has to be measured against the window's own norm, which is how the attention index was built.

6. About what

Each headline carries the coins it is about. There are 1,702 distinct tags over the two years, 147 with at least 100 headlines and 31 with at least a thousand. Bitcoin appears in 49% of all headlines and Ethereum in 24%; the ten most-tagged coins account for 81% of all tag mentions. Coverage is deep for a few dozen names and thin beyond them, which is the usual shape of attention and the reason the research notes restrict themselves to coins with enough headlines per window.

Seven headlines in ten are about a single coin. Twelve percent tag five or more: market roundups and lists, which the index treats as weak evidence about any one of them. Two thirds of all headlines tag at least one of the ten coins the production configurations trade.

7. What the scores look like

The mean score is 0.20 and the median 0.3; 68% of headlines score above zero, 31% below, 1.4% exactly zero. Half of all headlines have a score of at least 0.5 in magnitude: the scale is used, not hugged at the centre. The most common single value is 0.7.

The positive lean is a property of the types, not of the scoring. Reactive headlines, the 74% that report what happened, have a mean of 0.13 and are negative 38% of the time: markets fall as well as rise and the press reports both. Proactive headlines, the 26% that announce listings, launches, votes, upgrades and partnerships, have a mean of 0.43 and are negative 11% of the time. Nobody announces bad news in advance. This is the asymmetry behind the second note's finding that proactive headlines predict volatility a week out: they are a calendar of positive events, not a sample of opinion.

8. How fast it arrives

The time a record enters the database is the time its headline became usable, and the difference from the stated publication time is the latency of the whole pipeline. Since the June 2025 rebuild the median is 11 minutes; 86% of headlines are in the database within 15 minutes of publication, 99.2% within 30 and 99.8% within an hour, and the 95th percentile is 18 minutes. Before the rebuild the median was 36 minutes. Across the 42 largest sources the median delay runs from 8 to 17 minutes; the difference is mostly how often a source updates its own feed.

For the research notes this matters in one specific way. The index is computed on 12-hour windows and the bot runs five minutes after the window closes. A headline published in the last ten minutes of a window is, more often than not, counted in the next one. The event studies use publication time, which is the time a human reader could have acted; the live system sees the database time. The third note measured the cost of that gap and found it small.

9. Tone and the market

The level of the score is not stationary. The mean was 0.25 in 2025 and 0.09 in 2026, and month by month it moves with the market: the correlation between the monthly mean score and the monthly return of Bitcoin is 0.39. The press is more cheerful in a rising market, which is unsurprising, and it means the raw score cannot be compared across periods. A score of 0.3 was an ordinary headline in mid-2025 and a distinctly positive one in mid-2026.

This is the reason every index on this site measures tone against a coin's own recent norm: the question is never "is this headline positive" but "is this coin's news more positive than it has lately been". That construction removes the regime level by design, and the first two notes report that the resulting signal is stable across quarters.

10. What a user should know

Coverage changed once, and sources change by themselves. The June 2025 rebuild replaced about a third of the volume with different outlets, narrowed the set of regular contributors from about 49 to about 35 and cut the delay by two thirds. Individual sources change their output without notice, as one did in September 2026. Raw counts are not comparable across time; shares and measures normalised per source and per coin are.

Timestamps come from the source. Publication times are as stated by the outlet. 0.09% of records carry a publication time later than their arrival in the database, nearly half of them from one source that stamps a small fraction of its items with a future date and a further third from two others; those records are excluded from the delay figures above. No record arrived more than 24 hours after its stated publication.

Reprints are rare at the record level. A headline that repeats an earlier one word for word is 0.2% of records; one that shares its first six words with a headline of the previous 48 hours, 0.7%. The same story told in different words by different outlets is common and is not deduplicated: the index treats each outlet's telling as one observation, and the fifth note in this series will measure how much of a window is the same story.

Tags are the source's and the collector's. Which coins a headline is about is read from the headline and its source metadata. The long tail of 1,700 tags includes tickers that appear a handful of times; below a few hundred headlines a tag is a hint, not a sample.

Scores are discrete and the level drifts. The scale is −1 to 1, most values on steps of 0.1 and 7% on half-steps, the distribution leans positive because of the proactive type, and the mean follows the market. Use tone relative to a norm, not the raw level.

What is not in the data. Headline text is not republished on this site. Full text of articles was never collected. Posts from a handful of social accounts are 2% of the records and are treated like any other source. There is no price data in the dataset; the research notes join it to exchange candles separately.

The data files with this note contain every aggregate shown here. Access to the records themselves, with timestamps, sources, coins, types and scores, is described on the Data & access page.