+0.70Morning Minute: The CFTC Officially Greenlights Perps in the US, Starting with Backpack · Decrypt
+0.60CFTC Makes Historic Move With Blockchain Derivatives Approval · U.Today
+0.60US Approves Blockchain-Based Derivatives Exchange In Major Win For Crypto · Bitcoinist
news index, 3-day mean BTC/USDT, daily close
Each point is one UTC day, 00:00–24:00: the weighted mean of LLM scores (−1…1) of every BTC headline published that day, smoothed with a 3-day mean. 27,990 headlines over the last six months. Hover a point to see the day's strongest headlines. Updated 2026-11-17. What this index predicts is measured in note 1; why it is the tone and not the volume of news, in note 2.
Five things a trader would do with the news index, tested on the open data. The strength threshold that looked best in note 8 is a spike, not a principle. A long-only rule picks the right coins, 17 bps a day over the basket, but its level is one quarter of 2024. Waiting for a dip before buying destroys the rule, because the dips select the losers. Sizing by liquidity doubles capacity. As an overlay on a large book, the only thing that survives fees is not buying into bad news. The index is a measurement; a rule built on it needs its own construction and a live record, which is what note 3 is for.
The 24-hour sort of the earlier notes earns 21 bps a day gross and turns over 62% of its book daily: it breaks even at 8.5 bps per side, below the base spot fee. Holding until a coin leaves its group and trading only strong signals raises the break-even to 22–30 bps and leaves 13–21 bps a day at a 10 bps fee, but that is the best of 84 constructions, and with spread and impact it survives only at about ten thousand dollars per position. The information is real; as a strategy on its own it is small, and the capacity bound belongs to the naive sort, not to the index as an input to other models.
Nobody. The outlets that report the tape, Coindesk among them, publish two to three percent after the move; the largest ones about half a percent after it; and none is followed by anything: of 61 outlets, between two and six show a positive effect after publication and at most three a negative one, depending on how ties in the score are cut, never more than chance predicts. The 25 bps a day of the index comes from all outlets together, not from the best ten. The ranking that is real is the other one: who reports the tape, who writes first, and who echoes six hours later.
Every live trade of the last five weeks set against the simulated deal of the same configuration: 32 of 34 matched, entries 25 minutes apart in the median, exits identical, per-trade returns within half a point for most pairs. Live and simulated earned the same: +1.7% and +1.4% per trade, +59% and +56% summed.
Reactive headlines arrive after a 56 bps move and give back 9 bps of it; proactive ones arrive after 9 bps and are followed by a 6 bps drift that is over in half a day. At the daily index level the label's whole contribution is which quarter of the headlines to drop: the reactive-only index equals the all-headline index, and the proactive index adds nothing to it.
What the dataset behind this site actually is: 230,000 scored headlines from 109 sources over two years, how much arrives per day and from whom, when it is written, which coins it covers, what the scores look like, how fast a headline reaches the database, and what a user should know before relying on it.
The volume of a coin's news predicts the size of its next moves and nothing about their direction; the tone predicts direction and nothing about size. The two carry different information, and the tone effect of the first note is not a proxy for being in the news.
Two years, 37 coins, 202,222 scored headlines: one headline predicts nothing after publication; a coin's news tone, measured against its own norm, is followed by about ¼% a day of abnormal return for two days.
Every number on this site comes from one dataset: two years of scored and typed crypto headlines and the index built on them. It is available as a licence, as an API, or for joint research. The dataset is described in note 4; what it predicts, in notes 1 and 2.