Two kinds of headline: what the proactive/reactive label adds to the score

1. The question

A headline in this dataset carries two fields written at the moment it is ingested: a score on $[-1, 1]$ for its expected effect on the price of the coin it mentions, and a type. Reactive means the headline reports something that has already happened to the price or to the market: a move, a liquidation, a comment on a move. Proactive means it reports something that has not yet played out: a listing, a filing, a hack, a vote, a scheduled upgrade. Sentiment scores are sold by many; the type is the field that is not. The question of this note is whether it is a second piece of information or a restatement of the first.

The first note gave one answer in passing: aggregated into a 12-hour index per coin, the reactive headlines carry the whole predictive effect and the proactive ones carry none. That is a result about the two indices, each on its own sample of windows. It does not say what happens when the two are put side by side on the same days, nor what the label means at the level of the single headline, where most users of the data will meet it. This note does both. It is an ablation in the usual sense: the same tests are run with the label used and with it ignored, on the same observations, and the difference is what the label is worth.

2. Data

The sample is the one of the first note: every headline collected between 4 October 2024 and 28 September 2026 that mentions one of 37 coins with at least 500 headlines over the period and a Binance USDT spot pair. 202,222 headlines give 474,597 coin–headline pairs. Prices are Binance spot closes at 30-minute resolution; all returns are log returns; the abnormal return of a coin is its return minus the BTC return over the same window, and BTC itself enters with its raw return.

Of the headlines, 25.4% are proactive; of the coin–headline pairs, 29.3%, because a proactive headline tags more coins on average than a reactive one.¹A listing or a regulatory filing often names several assets; a comment on a move usually names one. The share differs by coin: 24.6% for BTC, 27.3% for ETH, 29.6% for XRP, 32–34% for the next seven large coins, and 17.5% and 12.5% for DOGE and SHIB, whose coverage is mostly commentary on price. It is not constant over time either, and Section 5.3 returns to that.

For the index-level tests the main construction uses 24-hour windows rather than the 12-hour windows of the first note. The reason is coverage. The proactive index needs at least two proactive headlines about a coin in a window; with 12-hour windows this holds for enough coins to form a cross-section in the first three quarters of the sample and then rarely, because the proactive share fell in mid-2025 and small coins seldom receive two proactive headlines in half a day. With 24-hour windows the proactive index has a cross-section of at least eight coins in 32 to 90 windows in every quarter. The 12-hour construction is reported as a variant in Section 6.

Table 1. Coverage of the three indices on 24-hour windows, after the history needed for a z-score. The base sample is where the "drop proactive" comparison is made; the common sample, 33 coins, is where the proactive index is compared with the others. The z-score of each index is taken against the coin's previous 30 windows in which that index exists, at least 20 of them, however far back in calendar time they reach, so the counts are below the raw number of windows meeting each rule.

3. Method

3.1 The headline level

For headline $i$ about coin $c$, published at $t_i$ with score $s_i$, the abnormal return at horizon $h$ is defined as in the first note, and extended to the hours before publication:

so that $AR_{i,-24}$ is the return over the 24 hours ending at publication. Prices sit on a 30-minute grid. For $h > 0$, $P_c(t_i)$ is the close of the first bar that closes at or after publication and $P_c(t_i + h)$ the close $h$ later, so a window after publication starts between 0 and 30 minutes after it; for $h < 0$, $P_c(t_i)$ is the close of the last bar that closes at or before publication and $P_c(t_i + h)$ the close $|h|$ earlier, so a window before publication ends between 0 and 30 minutes before it. Two sets of means are removed before comparison, the mean over all pairs of the same calendar day and the mean over all pairs of the same coin in the same month:

Both means are computed on the raw $AR_{i,h}$ and subtracted together in a single pass, as written; the adjustment is not iterated to a two-way fixed-effects fit. Headlines are then sorted into fifths by score, by rank rather than by a score threshold: the pairs of a group are ordered by score, ties broken by the order of publication, and cut into five groups of equal count. For the type curves the fifths are taken within type and month, so that a proactive headline is compared with other proactive headlines of the same month and the higher average score of proactive headlines does not enter.²Proactive headlines score higher on average than reactive ones, note 4, Fig. 5. Pooled quintiles would put more proactive headlines in the top fifth and more reactive ones in the bottom, and the type curves would partly measure the mix. The statistic is the spread between the top and bottom fifth, with $\rho^{\tau}_i$ the rank of pair $i$ within its month, and within its type for $\tau \ne \text{all}$, divided by the number of pairs in that group,

with the standard error taken from the series of daily differences. The quantity that measures the label is the difference between the two type curves,

with its standard error from the daily series of $D_h$. If the type carried nothing beyond the score, $D_h$ would be zero at every horizon.

Three further headline-level checks ask whether the label means what its name says. The size of the move around a headline, $|AR_{i,h}|$ demeaned as in (2), by type. The share of headlines published within 24 hours after a move larger than two standard deviations of the coin's 24-hour abnormal return in that month, with the standard deviation taken from all 24-hour windows on the price grid, not from the headlines. And the rank correlation between the score and the return before and after publication, computed within day and averaged over days.

3.2 The index level

Time is cut into 24-hour windows aligned to 00:00 UTC. The index of coin $c$ in window $b$ is the plain mean of the scores published in it, and the same mean is taken separately over the proactive and the reactive headlines:

with $N_{c,b} \ge 3$ for the first and $N^{\tau}_{c,b} \ge 2$ for the other two, the thresholds of the first note. Each index is expressed as a z-score, $z^{\text{all}}$, $z^{\text{pro}}$, $z^{\text{rea}}$, against the coin's previous 30 windows in which that index exists, at least 20 of them; windows where the index is undefined are skipped rather than counted, so the history spans 30 days only for a coin that has the index every day. This is equation (5) of the first note with 30 windows in place of 60. A fourth signal uses the label as a feature in the simplest way, an equal-weight mix of the two type z-scores:

Returns are measured from the end of the window for holding periods of 1, 2 and 7 days, abnormal against BTC. Two tests are run on each signal. The cross-sectional sort of the first note: in every window with at least eight coins carrying the signal, the spread in return between the top third and the bottom third, 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) and the legs scored on raw returns, averaged over windows with a Newey–West standard error. And a pooled regression of the window-demeaned forward return on the signals standardised to unit variance, with standard errors clustered by window,

run with one regressor at a time and with both. The regression is the main instrument on the common sample, because it uses every coin–day with a signal and does not need eight coins in the same window; the sort on that sample has a median of seven coins per window and is reported for completeness.

The ablation is two comparisons, each on a fixed set of observations. On the base sample, the coin–days where both the all-headline and the reactive index exist, $z^{\text{all}}$ against $z^{\text{rea}}$: what is lost by dropping the proactive headlines. On the common sample, where all three exist, $z^{\text{pro}}$ on its own and next to $z^{\text{rea}}$: what the proactive headlines carry, and whether any of it is new.

4. Results

4.1 Two shapes in event time

The left panel is the picture the labels promise. Over the 24 hours before publication, coins about to receive a top-fifth reactive headline have outperformed coins about to receive a bottom-fifth one by 56.0 bps ($t = 11.3$). For proactive headlines the same number is 9.1 bps ($t = 1.7$), and at 12 and 6 hours before publication 11.8 and 9.7 bps ($t = 3.2$ and $3.5$). A reactive headline is a report on a move that is still running when it is published: of the 56 bps, 20.8 accrue in the last six hours before publication and 4.6 in the last hour ($t = 5.8$), so the hour before publication is the fastest of the twenty-four but holds less than a tenth of the move. A proactive headline is preceded by a sixth of that move, 2.3 bps of it in the last hour ($t = 2.0$), which is not nothing: some announcements leak, some are the second headline on an event, and some are labelled proactive when they should not be.

The right panel is the part the first note did not show. After a reactive headline the first half-hour carries nothing ($+0.2$ bps, $t = 0.5$); from the second half-hour the spread is negative, $-1.0$ bps at 1 hour, $-2.1$ at 2 hours ($t = -2.1$), $-4.0$ at 6 hours, $-8.2$ at 12 hours ($t = -3.1$) and $-8.6$ at 24 hours: about a sixth of the preceding move is given back within a day. After a proactive headline the spread is positive at every horizon: $+0.5$ bps at 30 minutes, $+1.3$ at 1 hour, $+3.7$ at 4 hours, $+6.4$ at 6 hours ($t = 2.3$), then $+5.1$ at 12 hours and $+2.0$ at 24, with an interval that excludes zero only at 6 hours. The drift after a proactive headline is small, it is positive at all nine points, and it is over in about half a day.

Table 2. Top-fifth minus bottom-fifth spread in basis points, t-statistic in parentheses, from daily differences. The last column is eq. 4. Bold: |t| ≥ 3 after publication.

The difference $D_h$ is the measure of the label at the headline level. It is $+2.2$ bps one hour after publication ($t = 1.8$), $+5.0$ at two hours ($t = 2.7$), $+10.4$ at six hours ($t = 3.3$), $+13.3$ at twelve ($t = 3.4$), and then stops growing. Given two headlines with the same score, the one typed proactive is followed by a path about 10 bps higher over the next half day than the one typed reactive. That is information the score does not contain. It is also small: a tenth of a percent, on a single headline, before any cost.

4.2 What the label does not do

The label separates the direction of the path after publication relative to the score. It does not separate its size. The mean absolute abnormal return over the 24 hours after a headline is 233 bps after a proactive one and 229 after a reactive one; demeaned by day and coin-month, the difference is $-1.4 \pm 1.1$ bps. Over the hour after publication the two are 49 and 48 bps. The 24 hours before publication do differ, by $-11.6 \pm 1.5$ bps, proactive headlines being preceded by slightly quieter periods, which is again what the label says. A proactive headline is not a warning that the coin is about to move more than usual; it is a statement that whatever move follows has not yet started.

Table 3. Headline-level diagnostics by type. Rank correlations are Spearman coefficients computed within calendar day over all coin–headline pairs of that day and averaged over 725 days, with the t-statistic of the average. The 2σ threshold is the standard deviation of the coin's 24-hour abnormal return in that month, from the price grid. 139,205 proactive and 335,392 reactive pairs.

Table 3 restates Fig. 1 in other units. The score of a reactive headline is correlated with the return of the previous day at 0.074, seventeen standard errors from zero; the score of a proactive headline at 0.009. After publication the reactive score is correlated with the next day at $-0.017$, the proactive at zero. The share of headlines that follow a two-sigma move is close for the two types, 6.7% against 7.6%, because the large moves that attract commentary attract announcements too; the difference is in whether the score agrees with the move, which it does in 65% of such cases for reactive headlines and in 56% for proactive ones, the latter close to chance.

4.3 At the index level

Table 4. Slope of the window-demeaned forward abnormal return on the standardised signal, basis points per standard deviation, t in parentheses with standard errors clustered by window. 24-hour windows; n is the number of coin–days with a return at the horizon.

On the base sample, dropping the proactive headlines changes the slope from 9.8 to 8.8 bps per standard deviation at 24 hours and from 15.5 to 14.4 at 48, differences well inside one standard error. The reactive headlines, 71% of the flow, carry 90% of the slope; the other 29% of headlines add about a basis point, which is not distinguishable from zero.

On the common sample, the proactive index on its own has a slope of 3.0 bps at 24 hours ($t = 1.1$) and 1.6 at 48 ($t = 0.4$). Put next to the reactive index it has 0.7 and $-1.5$: whatever the proactive index knows at the daily horizon, the reactive index already knows. Put next to the all-headline index, its slope is negative, $-2.2$ and $-6.4$ ($t = -0.7$ and $-1.4$), which is the statistical way of saying that the all-headline index would be slightly better off without the proactive headlines in it, and that this too is within noise. At seven days the proactive index has a slope of 15.0 ($t = 2.3$) alone and 10.1 ($t = 1.5$) next to the reactive one; a week is the horizon at which the first note found a conventional momentum pattern, and the interval is wide.

The cross-sectional sorts agree where they have the coins to be run.

Table 5. Top-third minus bottom-third spread in basis points per holding period, Newey–West t in parentheses, 24-hour windows. "Windows" is the number of days with at least eight coins carrying the signal.

On the base sample, the same 629 days, the all-headline sort earns 21.5 bps at 24 hours and the reactive-only sort 20.0; at 48 hours 34.5 against 31.5. On the common sample the cross-section is seven coins in the median day and only 299 days reach eight, so every spread there has a standard error of 13 to 20 bps and none is individually significant; the proactive sort is the smallest of the three at every horizon, and the equal mix of eq. 6 does no better than the all-headline index at one and two days. Its 99 bps at seven days ($t = 2.3$, permutation $p = 0.03$ at 48 hours) is the one number in this table that would argue for the proactive index, and it rests on 296 windows of a thin cross-section at the horizon where the first note's placebo was also weakest.

Table 6. Common sample, 24-hour windows. Permutation: the real spread against 500 within-window shuffles of the signal across coins, basis points. Placebo: the same sort with the signal paired with the coin's return starting seven days later, spread with t in parentheses.

4.4 The label over time

The share of proactive headlines is 29–31% in every month from October 2024 to May 2025, 24% in June 2025, and between 19% and 25% in every month since. The step coincides with the rebuild of the collector described in note 4, when the mix of sources changed; outlets differ in their proactive share from 12% to 35%, so a change in who is collected is a change in the share. Nothing in this note depends on the level: the headline-level fifths are taken within month, and the index-level z-scores against each coin's own recent history.

Whether the label's effect is stable is a separate question, answered on the base sample by quarter.

Table 7. Mean 24-hour spread of the top-third minus bottom-third sort by calendar quarter, base sample, 24-hour windows. Standard errors are 20 to 34 bps per cell.

The two rows move together in every quarter, with a gap of 2 to 11 bps in either direction and no trend in the gap: the proactive headlines did not become more or less useful to the index after their share fell.

5. Robustness

Table 8. Sorts in bps per holding period with Newey–West t; regressions in bps per standard deviation with t from standard errors clustered by window. For 12-hour windows the z-score history is the coin's previous 60 half-day windows with the index, and "Windows" counts half-days for sorts and coin–half-days for regressions.

With 12-hour windows the base-sample comparison repeats: 16.9 against 16.4 bps in the sort, 8.8 against 7.4 per standard deviation in the regression. The common sample on 12-hour windows has a usable cross-section only until mid-2025, fewer than thirty windows a quarter after that, so only the regression is reported for it; the proactive slope is 0.0 alone and $-0.9$ next to the reactive one. Without BTC and ETH, the reactive-only sort is if anything stronger than the all-headline one, 23.3 against 19.0 bps, the mega-caps being the coins with the most proactive coverage. Requiring three proactive headlines instead of two does not help the proactive index.

6. What this says, and what it does not

The label means what it says, and the price confirms it. Reactive headlines follow a move of 56 bps between the top and bottom fifth by score; proactive headlines follow 9. The score of a reactive headline is a description of the previous day at a correlation of 0.07, seventeen standard errors from zero; the score of a proactive headline is nearly uncorrelated with it. This is a test of the label against prices, which are not what the labeller saw, and it is passed.

At the headline level the label separates two paths. After a reactive headline the coin gives back about 9 bps of the move over the next day; after a proactive headline it drifts another 6 bps in the direction of the score over six hours. The difference, 10 bps at six hours and 13 at twelve, is the one quantity in this note that the score alone cannot produce. It is small and it is fast: whoever wants to use it has six hours, a single-headline return of a tenth of a percent, and the costs of acting on single headlines.

At the daily index level the label is a filter, not a second signal. On the same coin–days, an index built from the reactive headlines alone predicts as well as the index built from all of them, 8.8 against 9.8 bps per standard deviation at 24 hours and 20.0 against 21.5 bps in the sort. The proactive index predicts 3 bps on its own, not significantly, and nothing next to the reactive one. What the label buys an index builder is the knowledge that 29% of the rows can be dropped without loss, and that any weight put on them beyond that is weight put on noise at this horizon.

It is not a measure of event risk. The absolute return after a proactive headline is the same as after a reactive one. The label says where the price is relative to the news, not how far it is about to go.

What stays open. Proactive headlines are a mixed class: listings, hacks, filings, votes, upgrades, each with its own clock. The six-hour drift is their average. Whether some kinds drift more, and whether the first proactive headline on an event behaves differently from the third, is a question about sources and redundancy that the next note takes up. The seven-day slope of the proactive index, 15 bps per standard deviation on a thin sample, is noted and not interpreted.

Limitations. One labeller, one language, one asset class, two years. The type is assigned once and never revised, so a headline about an event that had in fact already happened stays proactive; the 9 bps before proactive headlines is partly that. The common sample is the ten most covered coins most of the time, and its cross-section is thin; the regression results on it are the ones to rely on, the sorts are shown for completeness. The within-type fifths use each month's score distribution, which is known only at the end of the month, a mild look-ahead in the cut points and none in the returns. The step in the proactive share in June 2025 is a property of the collection, not of the news, and the tests are built so that it does not enter them; a user building their own index on the raw share should know it is there.