Who moves the crypto market? 61 news outlets ranked by what price does before and after their headlines

1. The question

The first note found that a headline sits at the end of a move, not at the start of one: between the most positive and the most negative fifth of headlines by score there is a 41 bps gap in the coin's abnormal return over the 24 hours before publication, and nothing after. That is an average over every source in the dataset. It need not hold for each. An outlet that writes when the chart has already moved will show a large gap before and nothing after; an outlet with sources, or with an audience that trades on it, could show the opposite. Every headline carries its publisher, so the question can be asked outlet by outlet: who reports the move, and does anyone precede it.

The answer to the second half is nobody, and that is the result. The sections below give the ranking on both halves, show that the predictive effect of the first note survives only when all outlets are pooled, and add the measure that does separate outlets cleanly: who writes a story first and who repeats it, and how much later.

2. Data

The sample is the one of the earlier notes, extended to 23 October 2026: every headline that mentions one of the 37 coins with a Binance USDT spot pair and at least 500 headlines over the period, 485,042 coin–headline pairs, of which 479,988 have a price on both sides of publication. An outlet is the publisher's domain with the www. removed; the four handles on X that the collector carried until May 2025 appear under their handle. The main set is the 61 outlets with at least 1,000 pairs, 468,049 pairs and 201,473 headlines together; an extended set of 78 outlets with at least 300 pairs, 477,490 pairs, is in the published table and is used for the robustness checks.

One fact about the collection shapes everything that follows. The collector was rebuilt on 29 May 2025, as described in note 4, and the set of outlets changed with it: 30 of the 61 outlets stop on that day and 31 run through October 2026, 21 of them present over the whole period and 10 starting that day. Benzinga appears twice, as the feed that stopped in May 2025 and as the site that started then, and is kept as two rows. Every figure and table marks which collection an outlet belongs to, and Section 5 repeats the main result on the post-May-2025 period alone.

Table 1. The sample. The 17 additional outlets of the extended set contribute 2% of its pairs. Prices are Binance spot closes at 30-minute resolution; the abnormal return of a coin is its log return minus the BTC log return over the same window, raw for BTC.

3. Method

3.1 Before and after publication, per outlet

For headline $i$ of outlet $o$ about coin $c$, published at $t_i$ with score $s_i$, the abnormal return before and after publication is defined as in the first note. The window before ends at the last 30-minute close before $t_i$; the window after starts at the first close after it:

Within each outlet, headlines are sorted into fifths by score within calendar month, so that an outlet whose scores drift with the market is compared with itself month by month; where an outlet has fewer than 25 pairs in a month the fifths are taken over its whole history, which happens for 0.3% of pairs. The fifths are cut by rank, not by a score threshold: every pair in the outlet-month receives its rank percentile $p_i$ by score, ties broken by publication time, and the top fifth is $p_i > 0.8$, the bottom $p_i \le 0.2$, so each holds exactly a fifth of the pairs whatever the score distribution. The tie rule matters because the score is discrete, 63 distinct values with heavy ties, and a threshold at the quantile would put about a quarter of the pairs into a "fifth". The two numbers that rank an outlet are the top-fifth minus bottom-fifth spread before publication and after it:

A large $\Delta^{-}_{o}$ means the outlet's positive headlines come after a rise and its negative ones after a fall: it reports the tape. A positive $\Delta^{+}_{o,24}$ means the coin keeps going in the direction of the score after the outlet has published: the outlet is ahead of the move, or moves it. Standard errors are clustered by calendar day, since many headlines of one day describe one market. A third number, the slope of $AR^{+}_{i,24}$ on the standardised score within outlet, uses all fifths rather than two and is reported with the table. On a discrete score scale the composition of the fifths, and with it the post-publication spread of an individual outlet, moves within its standard error with the tie rule; the pre-publication ranking does not.

3.2 Does the index come from the best outlets?

The 24-hour index of note 5, the mean score of a coin's headlines in a calendar day expressed as a z-score against the coin's previous 30 observed 24-hour index windows, at least 20 of them, counted over the days on which the coin had an index rather than over calendar days, is rebuilt three times: from all 111 outlets, from the ten outlets with the highest $\Delta^{+}_{o,24}$, and from the ten with the lowest. Each version is run through the cross-sectional sort of the first note: in every day with at least eight coins carrying the index, the return of the top third minus the bottom third over the next 24 and 48 hours, averaged with a Newey–West standard error, with a within-day permutation of the index across coins (500 draws) for a p-value. If the information in the flow were concentrated in the outlets that look best at the headline level, the ten-best index would carry it.

3.3 First mention and lag

Within each coin, headlines published within 12 hours of each other are clustered by title: titles are lowercased, stripped of punctuation, numbers, English stopwords and the coin's name, and two headlines belong to one story if their token sets overlap by at least half (Jaccard ≥ 0.5), joined in time order. The earliest headline of a cluster is the story's first mention. For each outlet, the share of its headlines that are first mentions, and the median delay in minutes behind the first mention for the ones that are not. The threshold is strict on purpose: it catches the same headline reworded, not the same event covered differently, so the reprint share it gives is a floor. Two looser thresholds are reported alongside.

Two descriptive measures complete the table: the share of an outlet's headlines typed proactive, and, for headlines since June 2025, the median time between publication and arrival in the database.

4. Results

4.1 Who reports the move

The horizontal axis is where outlets differ. Pooled over the main set, the top fifth of headlines by score follows the bottom fifth by 93 bps over the previous day ($t = 11.6$), twice the 41 bps of the first note, because fifths are now taken within outlet and month and an outlet's own scale is what matters. Across outlets the number runs from $-41$ to $+335$ bps, median 80.

Table 2. The ten outlets that report the most of the move and the four that report the least, by the 24-hour pre-publication spread, standard error in parentheses. Bold: more than 200 bps. "First mention" is the share of the outlet's headlines that are the earliest headline of their story; "lag" the median delay behind the first mention for the others. The full table of 78 outlets with every column is table_outlets.csv.

The top of the list is market commentary: a price piece on FXEmpire or CryptoTicker is published, on average, after the coin has moved more than three percent relative to Bitcoin in the direction of the piece's tone, and a CryptoPotato, Benzinga or Coindesk headline after more than two. These are not accusations; a market report that did not follow the market would be a strange one. The bottom of the list is announcements: Coindoo, Catenaa, coinpaprika and Nftgators publish listings, launches and partnerships whose tone has nothing to do with the previous day's chart, and their pre-publication spread is zero. The proactive share lines up with this, from 11% at Protos, Benzinga and daancrypto to 43–51% at coinpaprika, Coincu and Altcoinbuzz.

The two largest outlets sit in the middle: Cryptopolitan at 56 bps (38,672 pairs) and CoinTelegraph at 62 (33,993). UToday, Crypto Economy and coinpedia, each above 130, are closer to the commentary end; Decrypt and The Block, at 105 and 99, in between. The pre-publication spread is unrelated to outlet size (rank correlation with the number of pairs 0.16) and unrelated to what happens afterwards (rank correlation with the post-publication spread $-0.01$).

4.2 Nobody moves the market

Pooled over the main set, the spread after publication is $-0.7$ bps at 4 hours, $+0.4$ at 24 and $-5.0$ at 48, none distinguishable from zero. Per outlet the numbers run from $-74$ to $+92$ bps with a median standard error of 24 bps, and the picture in Fig. 2 is the one such a distribution produces: a few outlets at either end with wide intervals, and the large ones in the middle. Of 61 outlets, under the rank cut of Section 3.1, two have an interval entirely above zero, bankless (+92, interval 22 to 162) and FXEmpire (+64, 8 to 120), and three entirely below, daily.shib.io ($-74$), tokenist ($-63$) and, at the edge, coinpedia ($-26$, interval $-52$ to $+0.4$). The count depends on the tie rule: cut the fifths at the score quantiles instead of by rank and six outlets sit above zero and none below. Between two and six above and zero to three below is never more than the five or six that chance predicts among 61 intervals at the 5% level, and the two at the top under the rank cut are among the smallest outlets of the set, 2,242 and 1,293 pairs. With all five fifths used in a regression slope instead of two, seven outlets have $|t| \ge 2$, six positive and one negative, the same count a run of 61 tests gives when nothing is there.

Table 3. The twelve largest outlets by coin–headline pairs, and Coindesk. Spreads of eq. 2 with standard errors in parentheses; the slope is the regression of the 24-hour post-publication abnormal return on the standardised score within outlet, bps per standard deviation, t in parentheses, errors clustered by day.

For the twelve outlets that supply 56% of the sample, the 24-hour spread after publication is between $-26$ and $+27$ bps and the slope between $-7$ and $+12$ bps per standard deviation of score; the only $|t| > 2$ among them are NewsBTC on the spread ($-2.2$) and UToday and Cryptonews on the slope ($2.1$). Coindesk, the outlet most often named as the one that moves prices, reports the most of the move among large outlets, 216 bps, and is followed by 9 bps ($t = 0.6$) over the next day; its regression slope, 17 bps per standard deviation ($t = 2.7$), is the largest among large outlets and worth about a sixth of a percent on a headline two standard deviations from its mean. A reader who wants to find the outlet that moves the market can find one or two candidates in Table 3 at the 5% level, and would find the same number in a table of random numbers.

4.3 The crowd moves it

Table 4. Cross-sectional sort on the 24-hour index, eq. of note 5, with the index computed from different sets of outlets. "Days" counts days with at least eight coins carrying the index; coin–days are those with a z-score. Newey–West t in parentheses; p from the within-day permutation.

Built from all outlets, the index earns 21 bps per day at 24 hours and 36 at 48, with permutation p-values of 0.008 and below 0.002, which repeats note 5 on a sample extended by a month. Built from the ten outlets that rank best at the headline level, it earns $-2$ and $+3$ bps: nothing; with the ten selected under a threshold cut of the fifths instead of the rank cut it earns $+21$ ($t = 1.1$), equally indistinguishable from zero. Built from the ten that rank worst, $-35$ at 24 hours and $-17$ at 48. The first of these carries a permutation p of 0.04 in the table, but it is not a result: the ten are selected on a ranking that Section 4.2 has shown to be noise, the membership of the ten changes with the tie rule of the fifths, and with a different ten the number changes sign; the 48-hour horizon does not confirm it. It is to be read as zero. Part of the difference is coverage: ten outlets give a cross-section of eight coins on 196 days where all outlets give 668, and the standard errors double. The rest is the point of the note. The headline-level ranking of outlets after publication is noise, so selecting on it selects noise, and the information in the flow is in the agreement of many sources about a coin in a day, not in any one of them.

4.4 Who writes first, who echoes

By the strict rule of Section 3.3, 96.2% of coin–headline pairs are the first mention of their story and 3.8% are reprints: 466,674 stories among 485,042 pairs, of which 453,000 are told once, 10,300 twice and 265 six or more times. With the overlap threshold lowered to 0.4 the reprint share is 7.0%, at 0.3 it is 13.5%. The flow is less redundant than its reputation: most headlines, at the level of their wording, are not a copy of another.

Where there is a copy, the delay is measurable and it separates outlets by an order of magnitude. daancrypto, an X account, follows the first mention by 4 minutes in the median; Crypto Briefing by 54 minutes over 688 lagged headlines; The Block by 2.1 hours, Coindesk by 2.5, crypto.news by 2.5, CoinTelegraph by 3.0. At the other end NewsBTC follows by 7.6 hours, coinpaprika by 7.3, cryptodnes by 7.9, Voice of Crypto by 7.5, and CryptoPotato, the third-largest reporter of the tape, by 6.0. The large outlets of Table 3 lag by three to five hours; a reader who wants the earliest wording of a story and reads only CoinTelegraph is, in the median, three hours behind the first outlet that had it.

Table 5. Delay behind the first mention of the story for the outlet's headlines that are not first mentions, with the interquartile range, and the share of first mentions, for a selection of outlets. All outlets: table_outlets.csv.

4.5 Two descriptive columns

The proactive share, the fraction of an outlet's headlines that announce something rather than describe a move, runs from 11% to 51% and matches the pre-publication ranking: commentary outlets are at the bottom of it, announcement outlets at the top. The time from publication to the database, measured since June 2025, is 11 minutes in the median over all outlets, 7.8 for Crypto Economy and 8.0 for Blockonomi at one end, 16.6 and 16.9 for BeInCrypto and Altcoinbuzz at the other, with the 90th percentile at or under 30 minutes for every outlet but two, at 30.0 and 30.4. The collection delay is small against the lags of Table 5 and the same for all sources to within ten minutes; it does not explain any of the ranking above.

5. Robustness

Table 6. Each variant recomputes eq. 2 and compares the sign and the rank of the 24-hour post-publication spread with the main specification. Only outlets with data in the variant are counted.

Restricting the sample to the period after the collector was rebuilt keeps the sign of the post-publication spread for 20 of the 31 outlets that have data there and ranks them with a correlation of 0.51 against the main specification: the ranking is unstable across halves of the sample, which is what a ranking of noise should be. The pre-publication ranking is stable, and the pooled post-publication spread on the later period is again indistinguishable from zero. Taking fifths within outlet over the whole period rather than within month changes little, 47 of 61 signs kept and a rank correlation of 0.82, because most outlets' score distributions drift slowly.

Three limits should be read with the tables. Half of the outlets stopped in May 2025, so for them the measures describe the first eight months of the sample only. The title clustering is lexical, and two outlets covering one event in their own words are counted as two stories, so the first-mention share overstates originality and the lags are measured on the subset of headlines that are near-verbatim repeats. And since June 2025 the stored headline includes the lead sentence of the article; only the headline proper enters the clustering, but the score was assigned on both.

6. What this says, and what it does not

Outlets differ in how much of the move they report, and the ordering is sensible: market commentary at the top, announcement feeds at the bottom, the general crypto press in the middle. They do not differ in what follows: the spread after publication is zero for the large outlets, and the two to six of sixty-one that clear a 5% threshold, depending on the tie rule, are no more than a null produces, found at the ends of the table and concentrated in the smallest outlets. No outlet in this dataset moves the market, or is reliably ahead of it, at the horizon of a day.

The predictive effect of the earlier notes is therefore not a property of any source. It appears when the index is built from all outlets and disappears when it is built from the ten that look best at the headline level, under either way of choosing them, which is exactly what one expects if the effect is an aggregate of many weak, noisy descriptions of the same day. The practical reading for a user of the data is to weight breadth, not provenance: a coin's day described by thirty outlets is a measurement, a coin's day described by one is a headline.

What outlets do differ on, by an order of magnitude, is time. Among near-verbatim repeats, some outlets follow the first mention within the hour and the large ones within three to five, and the lag is a stable property of the outlet across the sample. For the live feed that is the useful ranking: not who moves the price, but whose wording of a story arrives first.