Loud or positive? Coverage volume tells you how much a coin will move. Tone tells you which way.

1. The question

A news flow has two dimensions. What is being written, which the first note measured as tone, and how much is being written, which it ignored. The second is the obvious objection to the first: a coin that receives three times its usual number of headlines in twelve hours is in the news for a reason, whatever the scores say, and the tone effect might be nothing more than that reason showing through.

So this note measures volume on its own and asks three things. Does unusual attention predict where the price goes? Does it predict how much the price moves? And does the tone effect survive once attention is accounted for, or does it live only in the windows where a coin is suddenly everywhere?

The data, the universe and the windows are those of the first note: every English-language crypto headline from 4 October 2024 to 28 September 2026, 37 coins with at least 500 headlines and a Binance USDT pair, 12-hour windows aligned to 00:00 and 12:00 UTC.¹One difference in the panel. The tone index exists only where a window has at least three headlines, 14,138 coin–windows. Attention is defined everywhere, including windows with no headline at all, which are informative in their own right: 49,552 coin–windows on a grid that runs from each coin's first headline to the end of the sample.

2. Measuring attention

Let $N_{c,b}$ be the number of headlines about coin $c$ published in window $b$, zero included. Counts are skewed, so the index is built on the log count, against the coin's own previous 60 windows:

with $\mu$ and $\sigma$ the trailing mean and standard deviation over the coin's previous 60 windows of the grid, which has no gaps, so over 30 calendar days; at least 20 previous windows are required, which binds only in a coin's first ten days. A value of $+2$ means the window is two standard deviations louder than is normal for that coin; a value near $-1$ is a window quieter than usual, often one with no headline.

Coverage has a rhythm: weekends are quiet, the European morning is loud. A second version removes it before standardising, by subtracting the coin's mean log count in the same weekday and half-day slot over the previous eight weeks:

The slot is one of fourteen, weekday × half of the day; the mean is over the same slot in the previous eight weeks, at least four of which must be available; and $\tilde n$ is then standardised exactly as in (1), against its own previous 60 windows. Both are known at the end of the window. The tone index $z_{c,b}$ is the one of the first note, equation (5) there: the mean score of a window with at least three headlines, as a z-score against the coin's previous 60 such windows, at least 20 of them, gaps skipped. Attention is also computed separately for the two headline types, announcements (proactive) and commentary (reactive), from their own counts.

Table 1. Attention panel for the seven most covered coins. The 90th percentile shows how spiky coverage is: a coin that usually gets two headlines per window gets seventy on its loud days. The full table is in the data files.

3. Three targets

Direction, size and volatility are measured from the end of the window $e_b$ over $k$ windows ahead, $k = 1, 2, 4, 14$ (12 hours to 7 days). Direction is the abnormal log return against BTC of the first note,

size is its absolute value $|AR_{c,b,k}|$, and volatility is realised volatility from the 30-minute closes inside the horizon,

All three are in basis points. To read a spread as a share of the typical move: the median $|AR|$ over 12 hours is 116 bps, the median $RV$ 262 bps; over 48 hours 249 and 572.

The tests are those of the first note. In every window with at least eight coins, coins are ranked by the signal and the mean target of the top third minus the bottom third is recorded, $S_{b,k}$, the thirds taken by rank as there (of $n$ coins, the $\lceil n/3 \rceil$ highest-ranked against the $\lfloor n/3 \rfloor$ lowest-ranked); its mean over windows is reported with a Newey–West standard error with $k$ lags. The signal is shuffled across coins within each window 500 times for a permutation $p$. The signal is paired with targets seven days later for a placebo. Pooled regressions use the target demeaned by window on the signal standardised to unit variance, with errors clustered by window.

4. Results

4.1 Attention predicts size and volatility, not direction

Sorted by attention, the loudest third of coins does not go anywhere in particular: the signed spread is 0.5 bps at 12 hours, 0.0 at 24, 3 at 48, with $t$-statistics below 0.4 and a permutation $p$ of 0.91 and 1.00. It does move more. The absolute spread is 11, 13 and 16 bps at 12, 24 and 48 hours ($t$ = 4.5, 3.4, 2.5), and the volatility spread 16, 19 and 20 bps ($t$ = 7.0, 5.1, 3.2). No shuffled signal comes close ($p$ < 0.002), and the placebo paired with the following week is 0.7 and 0.8 bps. Against a median 12-hour realised volatility of 262 bps, the loudest third is about 6% more volatile than the quietest.

Table 2. Top third minus bottom third, bps per horizon, Newey–West t in parentheses. Bold: |t| ≥ 2. The tone row is the sort of the first note, same z-score and same 962 windows, scored on the abnormal return like the other rows; the first note's Table 2 scores its two legs on raw returns, which gives 9.3, 24.7 and 33.2 there, the difference being BTC's own place in the sort. The two signals are near-orthogonal in what they predict: attention carries size and volatility and no direction; tone carries direction and no size.

Removing the weekly rhythm costs about a third of the effect at 12 and 24 hours and most of it at 48, which says that part of what "loud" means is "it is Tuesday morning": the days when the press is busiest are also the days when markets are. What remains after the adjustment, 11 bps at 12 and 24 hours with $t$ above 3, is attention specific to the coin.

Within a coin, the information sits in the loud tail rather than along a ladder. The loudest fifth of a coin's windows is followed by 24, 30 and 37 bps more realised volatility than the market at 12, 24 and 48 hours; the other four fifths are indistinguishable from each other and slightly below the market. Attention works as an alarm, not as a dial.

4.2 Most of it is volatility clustering

Volatility is persistent, and the press writes about moves. A loud window is therefore likely to be one where the coin has just been volatile, and the next window is likely to be volatile for that reason alone. The question is how much attention adds to what the previous day's volatility already says. The pooled regression with the previous 24 hours of realised volatility as a control answers it:

Table 3. Bucket-demeaned target on standardised attention, without and with the previous day's realised volatility. Errors clustered by window; 47,181 coin–windows in the first column, 47,165 in the other two, which also need the previous day's volatility.

Two thirds of the attention effect is the previous day's volatility wearing a byline. What attention adds on its own is 4 to 6 bps per standard deviation, significant at 12 hours ($t$ = 2.2 for volatility, 3.1 for size), marginal at 24 and gone at 48. Headlines count moves after they happen more than they foretell them, which is the first note's conclusion about tone restated for volume.

4.3 The tone effect does not depend on attention

If the 25 bps of the first note were an artefact of attention, they would be concentrated in loud windows and absent in quiet ones. They are not. Splitting the coins of every window that carry both signals into a quieter and a louder half by rank of attention, the lower $\lfloor n/2 \rfloor$ ranks forming the quieter half, and sorting by tone inside each half as in the single sorts gives a 24-hour tone spread of 9 bps in the quiet half and 3 bps in the loud half, a difference of 0.5 bps ($t$ = 0.01).²The double sort has fewer windows than the single sorts, 77 and 113, because it needs at least eight coins with both signals in the half; the louder half has one coin more when the count is odd, which is why it reaches eight more often. The regression in Table 4, which uses every coin–window with both signals, is the cleaner test. The 3 × 3 grid of attention terciles against tone terciles says the same thing in more detail:

Table 4a. Mean 24-hour abnormal return, demeaned by window, by attention tercile and tone tercile, each taken by rank within the window and independently of the other (the lowest $\lfloor n/3 \rfloor$ ranks, the highest $\lceil n/3 \rceil$, the rest in the middle), 13,642 coin–windows with both signals. The tone gradient runs left to right in every row.

The regression with both signals and their product makes it exact:

Table 4b. Signed abnormal return on standardised tone, attention and their product, bps per standard deviation, errors clustered by window; 13,642 coin–windows.

The tone coefficients are those of the first note to the decimal: 8.8 bps per standard deviation at 24 hours there, 8.8 here. Attention adds nothing to direction and the interaction is zero at every horizon up to two days. The tone index is a measure of what is being said, not of how much.

4.4 Two things found along the way

Bad news is louder. In the same regressions with size and volatility as the target, tone enters with a negative sign: −4.8 bps per standard deviation of tone on 24-hour size ($t$ = −2.8), −7.6 on 24-hour volatility ($t$ = −5.2), growing to −19 and −21 at a week. A window of unusually positive coverage is followed by smaller moves than a window of unusually negative coverage. Direction and size are both in the tone, with opposite signs.

Announcements predict volatility a week out. Counted separately, both headline types predict next-day volatility (15 and 19 bps for commentary, 11 and 14 for announcements). They part at longer horizons. Commentary fades by a week; the announcement count still shows 33 bps at 7 days ($t$ = 3.4), and its placebo, volatility in the week after that, is 8 to 9 bps ($t$ = 3.3 and 2.3) where every other placebo in this note is zero. That is not a failed placebo. Announcement headlines are about things scheduled to happen, and the things happen later: a listing, a vote, an unlock. The count of them is a calendar of future volatility, the one kind of foresight the first note found no trace of in their tone.

4.5 Over time and without the majors

The attention–volatility spread at 24 hours is positive in eight quarters of eight, between 1 and 34 bps. Without BTC and ETH it is larger, not smaller: 25 bps at 24 hours ($t$ = 7.4) and 29 at 48 ($t$ = 5.4), with the size spread at 16 and 21 bps ($t$ = 4.0, 3.2). Direction stays at zero without them too.

5. What this says

Volume and tone carry different information. How much is written about a coin tells you how much it will move over the next day or two, and nothing about the direction. What is written tells you the direction and nothing about the size. They are two instruments, not one.

Most of the volume signal is an echo. Two thirds of it is the previous day's volatility, which the press reports and which persists on its own. The part specific to attention is 4–6 bps per standard deviation and lasts half a day. Inside a coin's own history it is the loudest fifth of windows that matters, not the gradation below it.

The tone effect of the first note stands. It is the same size in quiet windows and loud ones, the interaction is zero, and the tone coefficient is unchanged by adding attention. Whatever the 25 bps a day are, they are not a coin being in the news.

Two footnotes worth their own notes. Negative tone is followed by larger moves than positive tone. And the count of announcement headlines is the first thing in this dataset that looks into the following week: it predicts volatility seven days ahead, which is what a calendar of scheduled events should do and what their scores, in the first note, did not.

Limitations. Zero-headline windows are both a measurement of silence and a consequence of which outlets are collected; a coin covered by one outlet looks quieter than it is. The weekly-rhythm adjustment uses eight weeks and a coarse slot; a finer one would remove more. Realised volatility from 30-minute bars understates the volatility of thin coins. The regression with past volatility uses one lag of one day; a richer volatility model would likely absorb more of the attention effect, not less. As in the first note, overlapping horizons are handled with Newey–West errors, and the 7-day numbers are read as suggestive.