> ## Content Index
> Fetch the complete content index at: https://research.glassnode.com/llms.txt
> Use this file to discover other available public pages before exploring further.

# Book Churn: How Fast the Order Book Is Being Rewritten
- URL: https://research.glassnode.com/book-churn/
- Published: 2026-11-18T14:31:56.000Z
- Updated: 2026-11-18T14:31:55.000Z
- Description: Book Churn counts how often the price levels of an order book are revised. We define the metric, show how much of it depth and spread already capture, and test it as a volatility indicator on Binance BTCUSDT and ETHUSDT perpetuals.
- Author: Padraig ODowd
- Tags: Product & Research

In this article we introduce Book Churn, a metric that counts how often the price levels of an order book are revised. Depth and spread describe the order book at one moment; they do not measure how fast it is being rewritten. We publish Book Churn at ten-minute resolution, scored against the same instrument's trailing thirty days at the same hour of day, and it is a measure of order-book activity first. We also tested it as a volatility indicator on Binance BTCUSDT and ETHUSDT back to 2020, in-sample, against controls matched on year, trailing volatility and time of day. A reading two standard deviations above its norm is followed by an hour of above-median volatility about 63% of the time, against a 50% baseline. A reading from 24 hours earlier already gets 59%, so most of the edge reflects volatile periods that last for days, and a causal price-range signal does as well as churn. Churn moves with the price move and does not lead it. The metric is available in [Glassnode Studio](https://studio.glassnode.com/charts/derivatives.PerpetualOrderbookChurn?a=BTC).

## Why order-book turnover matters

Most order-book metrics describe the order book at a point in time: how much is resting near the price, how wide the spread is, and where the large price levels sit. They are snapshots, or averages of snapshots, and say nothing about how often the price levels behind them are revised.

Depth tells us how much is resting now. The update count tells us how many times the levels changed during an interval. On BTCUSDT in a typical month of 2025 and 2026, the 95th-percentile interval carried about 3.9x the updates of the 5th-percentile interval. For depth within 1% of the price, the same ratio was 1.9x. Depth alone therefore understates how much the order book changes.

![Chart 1. An illustration, not real data. Two order books hold the same size at every price level, so they have the same depth and spread. Bids are resting buy orders, asks are resting sell](https://storage.ghost.io/c/25/7e/257e6d23-09bb-4d1c-b3ed-16ff70eb3f9f/content/images/2026/11/book_churn_chart1_same_snapshot.png)

****Chart 1.** An illustration, not real data. Two order books hold the same size at every price level, so they have the same depth and spread. Bids are resting buy orders, asks are resting sell orders. Over the ten minutes shown, the levels of Book A are revised 14 times and those of Book B 56 times, four times as often. A snapshot cannot tell the two books apart, but a count of revisions can.

Below, we define the count, measure how much of it depth and spread already capture, and test it as a volatility indicator.

## What the metric measures

A market maker who is comfortable leaves quotes in place. One who is repricing risk cancels and replaces quotes again and again across many levels. Resting depth can look the same in both cases. Note that repricing risk is only one reason to revise quotes. Queue management and latency defense are others, and a count cannot tell them apart.

We can observe this activity. For each ten-minute interval, we count the rows the exchange's depth feed publishes, one row per revised price level. Trades add to the count too, since a fill changes the size at a level.

The count has three limits.

- One feed message carries many level updates, and several changes to the same price within one message collapse into a single row. We therefore count level revisions as the exchange publishes them, not individual order actions.
- The count depends on one exchange's feed format and update frequency. Absolute counts are specific to that exchange and cannot be compared across exchanges without recalibration.
- We leave out the rows of order-book snapshots. The exchange resends the full order book on connect and after a resync, and those rows restate levels instead of revising them. As a result, an interval in which the feed was reconnecting reads as low activity, because revisions made while it was disconnected were not captured. This is a feed event, not a market event.

The raw counts are large. A median ten-minute interval carries about 607,000 level updates on BTCUSDT and 548,000 on ETHUSDT.

We publish the count in two forms: a score (zscore), which compares each interval with the same instrument's recent history at the same hour of day, and a ratio to its trailing average. We describe both below.

## What depth and spread miss

Depth moves less than the update count. On BTCUSDT in a typical month of 2025 and 2026, combined bid and ask depth within 100 basis points (1%) of the mid-price was about 1.9x as large in the 95th-percentile interval as in the 5th-percentile interval. For the update count, the same ratio was about 3.9x. Month by month, the depth ratio ran from 1.4x to 2.3x and the update ratio from 3.0x to 5.5x. Part of the variation in both comes from the daily cycle.

The zscore is related to depth and spread, though less closely than it used to be. We scored depth and spread the same way as churn (previous thirty days, same hour of day). From 2020 to September 2026, the churn score correlates at −0.40 with depth within 100 basis points and +0.49 with spread on BTCUSDT, and at −0.36 and +0.43 on ETHUSDT. When the order book is revised heavily, it is on average somewhat thinner and the spread somewhat wider.

Together, depth within 25 and 100 basis points and spread explain 31% of the variation in the score on BTCUSDT and 25% on ETHUSDT. This share has fallen. In each year from 2020 to 2022 it was 45% to 73% on BTCUSDT and 40% to 70% on ETHUSDT. Since 2023 it is 21% and 15% (14% to 39% by year on BTCUSDT, 13% to 19% on ETHUSDT). We have not found out why.

The ten-minute high-low price range, scored the same way, explains about 70% on its own over the full history, and 61% to 76% in each year. Over the full history, adding depth and spread to it adds less than one point. Most of what the score shows is therefore also visible in price. About 30% is explained by none of the three; it includes noise and feed artifacts as well as any new information.

![Chart 2. BTCUSDT on 27 April 2026, 03:00 to 08:00 UTC: level updates per ten-minute interval, combined bid and ask depth within 100 basis points, the time-weighted spread, and the mid-price](https://storage.ghost.io/c/25/7e/257e6d23-09bb-4d1c-b3ed-16ff70eb3f9f/content/images/2026/11/book_churn_chart2_snapshot_vs_churn.png)

****Chart 2.** BTCUSDT on 27 April 2026, 03:00 to 08:00 UTC: level updates per ten-minute interval, combined bid and ask depth within 100 basis points, the time-weighted spread, and the mid-price high-low range. Each point is plotted at the end of its interval, so the interval starting at 05:10 appears at 05:20\. The shaded band marks the three intervals discussed below. One basis point (bps) is 0.01%.

Chart 2 shows one case, picked after the fact. In the three intervals starting at 05:10, 05:20 and 05:30 UTC on 27 April 2026, BTCUSDT carried 1.8 to 2.1 million updates each, up from 0.6 to 0.9 million in the hour before. The scores were 4.34, 3.83 and 3.41\. Depth within 100 basis points was $377 million to $412 million, against $400 million to $464 million in the hour before, and rose to $565 million in the next interval. The spread stayed at about one tick, between 0.013 and 0.016 basis points.

Price moved in the first interval, with a 95 basis-point range that a price-range signal also flagged. In the next two, the range was 28 and 39 basis points, while updates stayed near two million. Such intervals are uncommon. Of all BTCUSDT scores of 3 or more, about 7% came with depth close to its recent norm (within one standard deviation) and a price-range score below 2.

## How the score is built

For each instrument and interval, we take the log of the update count and score it against the same instrument's previous thirty days:

```
zscore = ( ln(1 + updates) − mean ) / standard deviation
```

The mean and standard deviation come from the instrument's ten-minute intervals at the same hour of day over the previous thirty days: up to 180 intervals, all before the interval being scored. We use the same hour because order-book activity follows a strong daily cycle; without this adjustment, a high reading would mostly reflect the time of day. The only current input is the interval's own update count, which is known as soon as the interval closes. For the interval starting at 14:30, nothing in the value comes from after 14:40\. Note that we have not published the delay between interval close and API availability. Anyone using the score for execution should measure it first.

The API returns two fields:

- **zscore**, the standardized score.
- **ratio**, the update count divided by the average count of the intervals in the previous thirty days, across all hours, that had any updates at all. A ratio of 5 means five times the trailing mean. Unlike zscore, ratio is deliberately not adjusted for hour of day, so it shows the raw amount of traffic.

The appendix lists when no score is returned and explains the log transform.

The per-instrument window is what makes the score readable across market regimes: it is always relative to the instrument's own recent behavior, not to an absolute update rate. The score is designed to be comparable across instruments too, but we have only checked this for the two instruments analyzed here.

## How often high readings occur

On Binance BTCUSDT and ETHUSDT perpetuals since 2020, among intervals with a score (BTCUSDT / ETHUSDT):

| Threshold  | Share of scored intervals | Intervals per day | Episodes per day |
| ---------- | ------------------------- | ----------------- | ---------------- |
| zscore ≥ 2 | 4.3% / 4.6%               | 6.2 / 6.5         | 2.7 / 2.9        |
| zscore ≥ 3 | 0.94% / 0.98%             | 1.4 / 1.4         | 0.7 / 0.8        |

The score does not follow a normal distribution. High readings are more common than a normal distribution would suggest, so set thresholds from the observed frequencies in the table, not from standard deviations. The appendix has the details.

For alerts, count episodes, not intervals. The score crosses 2 in 4% to 5% of intervals, roughly six times a day per instrument, but consecutive crossings form one event. Counted as episodes, a reading above 2 occurs fewer than three times a day, and a reading above 3 less than once a day.

![Chart 3. BTCUSDT zscore over the three months to 30 September 2026, with reference lines at 2 and 3, and the ten-minute mid-price high-low range in basis points below on the same time axis.](https://storage.ghost.io/c/25/7e/257e6d23-09bb-4d1c-b3ed-16ff70eb3f9f/content/images/2026/11/book_churn_chart3_zscore_and_range.png)

****Chart 3.** BTCUSDT zscore over the three months to 30 September 2026, with reference lines at 2 and 3, and the ten-minute mid-price high-low range in basis points below on the same time axis. High scores cluster around volatile periods; the highest, on 19–21 August, reached 8.2\. The deep negative readings in July sit just above the low-activity cutoff (see appendix) and are most likely partial data gaps rather than quiet markets.

[Live chart: Book Churn for BTCUSDT in Glassnode Studio](https://studio.glassnode.com/charts/derivatives.PerpetualOrderbookChurn?a=BTC)

## Is the next hour more volatile?

An interval with a sharp price move also carries heavy feed traffic, so churn and volatility will tend to move together. We therefore test whether a churn reading adds anything to what the price action already showed.

For every reading of 2 or more on BTCUSDT and ETHUSDT from 1 January 2020 to 30 September 2026, we compare the volatility of the next hour with the median of ordinary intervals from the same year, at the same hour of day and with similar recent volatility (including the interval just closed). A run of consecutive readings counts once, as an episode. The hit rate is the share of episodes followed by an hour more volatile than that median. Figures shown with ± are one standard error. The appendix describes the method in full.

| Threshold  | BTCUSDT     | ETHUSDT     | Median volatility ratio | Episodes      |
| ---------- | ----------- | ----------- | ----------------------- | ------------- |
| zscore ≥ 2 | 63.2% ± 1.8 | 63.4% ± 1.5 | 1.18 / 1.19             | 6,699 / 7,236 |
| zscore ≥ 3 | 69.1% ± 2.5 | 68.2% ± 2.2 | 1.31 / 1.29             | 1,780 / 1,935 |

About 63% of readings are followed by an hour of above-median volatility, and that hour is typically about a fifth more volatile. Above 3, the hit rate rises to 68% to 69%.

We checked the control with two placebo tests. Fake spikes placed at random, at the same rate, score 51.2% and 50.5%, close to 50% as expected. The churn score from 24 hours earlier, however, scores 58.9% and 58.4%, well above 50%. This is the main caveat: a day-old reading retains much of the apparent edge. A large part of what the metric picks up is therefore a volatile period that lasts for days, not information about the interval just closed. Over the full sample, the current reading adds four to five points over a day-old one; from 2022 onward the gap is wider.

The hit rate is above 50% in every year, but 2020 and 2021 are close to chance, at 53% to 57%, and their 95% confidence intervals include 50%. From 2022 onward it is 67.1% ± 2.0 on BTCUSDT and 66.8% ± 1.5 on ETHUSDT, about two thirds. The appendix gives the figures by year and by volatility decile.

## Churn against a price-range signal

A message count could simply stand in for price range. To test this, we compared churn with a signal built on price range.

The range signal scores the high-low range against the previous thirty days, built the same way as churn, and we set it to produce the same number of episodes so that neither signal is handicapped. A simpler version that ranks each interval against the whole calendar year does beat churn, but it uses intervals that had not happened yet, so it cannot be traded. The appendix describes both.

On that basis there is nothing between them. Churn scores 63.2% against the range signal's 63.7% on BTCUSDT, and 63.4% against 63.5% on ETHUSDT. Paired over thirty-day blocks, the difference is −0.53 points (95% interval −1.7 to +1.1) on BTCUSDT and −0.15 points (−1.8 to +1.3) on ETHUSDT. Both intervals include zero, so the difference cannot be told apart from zero; churn is slightly behind on the point estimates.

On hit rate per alert, churn therefore does not beat a price-range signal that uses only past data. The two scores correlate at 0.84, yet they often fire at different times: in about half the intervals where churn fires, the range signal does not fire at all (51.5% on BTCUSDT and 50.6% on ETHUSDT).

Note that this share depends on how the range signal is built. Matched on the number of intervals instead of episodes, the range signal fires more often and the share falls. Built on all hours instead of the same hour, it rises to about 58%. We report about half because that figure comes from the version that makes the hit-rate comparison fair.

In those intervals the order book is revised heavily while the price range stays below the range signal's threshold. These markets are not calm; their range is simply not unusual enough to trigger the other signal. Given the 0.84 correlation, this half is what remains after a large overlap.

## A worked example

The morning of 21 September 2026 is a clear example. Times below refer to the start of each ten-minute interval.

At 08:30 UTC, Bitcoin perpetuals recorded about 3.4 million level updates in ten minutes, 4.13x the thirty-day average of roughly 821,000, for a score of 4.60\. The mid-price high-low range in that interval was 226 basis points, a 2.26% swing from low to high in ten minutes.

The score stayed high for the next hour: 4.40 at 08:40, then 2.43, 2.77, 2.07, 3.62, and a peak of 4.70 at 09:30\. ETHUSDT moved with it throughout and peaked at 4.60\. The 09:30 peak came with a narrower range than at 08:30, 117 basis points.

The 08:30 interval scores 4.60 against its own hour of day, but only 3.65 against all hours. Activity between 08:00 and 09:00 UTC is normally slightly below the all-hours average, so the burst was more unusual for that hour than for the day as a whole. This is what the hour-of-day adjustment adds. The ratio field, which is deliberately not adjusted for hour, shows the raw traffic multiple instead: 4.13x the trailing average at 08:30, rising to 4.59x at 09:30.

Note that this is one episode, picked after the fact because it is clear. It illustrates the metric but is not evidence, and the range signal would also have fired here.

![Chart 4. The episode from 06:00 to 12:00 UTC on 21 September 2026: zscore for BTCUSDT and ETHUSDT (top) and mid-price change from 06:00 in percent (bottom). Each value is plotted at the end](https://storage.ghost.io/c/25/7e/257e6d23-09bb-4d1c-b3ed-16ff70eb3f9f/content/images/2026/11/book_churn_chart4_episode.png)

****Chart 4.** The episode from 06:00 to 12:00 UTC on 21 September 2026: zscore for BTCUSDT and ETHUSDT (top) and mid-price change from 06:00 in percent (bottom). Each value is plotted at the end of its interval, when it is known, so readings appear ten minutes after the start times used in the text: the 08:30 score sits at 08:40 and the 09:30 peak at 09:40.

## How to read the score

Churn peaks with the price move, not clearly before it. In the largest episodes, the interval with the sharpest move is also the one with the most updates. What the score adds is that the next hour has tended to stay more volatile than in comparable conditions. A high reading means the order book is being revised heavily now, and historically the next hour was more volatile than the median about 63% of the time. It does not warn of volatility in a calm market.

Much of the signal reflects volatile periods that last for days. A day-old reading scores 58.9% and 58.4%; the current reading adds four to five points.

The hour-of-day adjustment does not remove the daily pattern completely. The share of readings above 2 still varies by UTC hour and by ten-minute slot within the hour (see appendix).

The score has no direction. A high reading says the order book is being revised heavily, not which way price will move. It could serve as a filter alongside directional measures such as depth imbalance or order flow, but we have not tested this.

We tested two instruments on one exchange. Binance BTCUSDT and ETHUSDT are the deepest perpetual markets in crypto, which is why we chose them. However, the two are highly correlated and their results move together year by year, so they amount to close to one test, not two independent ones. The measure is also specific to one exchange, so the results should not be assumed to hold elsewhere without re-running the tests there. The appendix includes a shorter re-run on five other Binance perpetuals.

The test uses information from within each year. Volatility deciles and control medians are computed per calendar year, so a trader in March 2020 would not have had them. The figures describe the sample; they are not a hit rate a trader could have achieved live.

About 37% of high readings are followed by an ordinary hour. A high reading shifts the odds; it is not a forecast.

We have not evaluated any trading or execution use. This article measures what follows a reading, not what acting on it would cost or earn.

## Common questions

**What exactly is the baseline?** The sample mean and standard deviation of ln(1 + updates) over the same instrument's ten-minute intervals in the same UTC hour (all six intervals of that hour, pooled) across the previous thirty days, excluding the interval being scored. Scores start once fourteen days of history exist; until thirty days are available, the baseline uses whatever history there is.

**Is this built from order-level (L3) data?** No. It is built from the exchange's L2 depth feed, which reports price levels rather than individual orders.

**What counts as one update?** One price-level entry, on either side of the order book, in that feed. Every entry counts, whether the change behind it came from one order or many. Order-book snapshots are not counted.

**How far back does the history go?** The metric is backfilled to 1 January 2020 for BTCUSDT and ETHUSDT, with scores from 15 January 2020, after the fourteen-day warm-up. The Point-in-Time version records each value as it was first published, so its history starts at launch in October 2026.

## Availability

Book Churn is available at ten-minute resolution for Binance BTCUSDT, ETHUSDT, BNBUSDT, DOGEUSDT, SOLUSDT, XRPUSDT, TRXUSDT and HYPEUSDT perpetual futures. History starts in January 2020 for BTCUSDT and ETHUSDT and at each contract's start date for the others. Each interval carries the standardized score and the activity ratio. Explore it in [Glassnode Studio](https://studio.glassnode.com/charts/derivatives.PerpetualOrderbookChurn?a=BTC). This article analyzes BTCUSDT and ETHUSDT in depth; the appendix covers the others in a shorter re-run.

We are considering adding more instruments and exchanges.

## Summary

Book Churn counts how often the price levels of an order book are revised, which depth and spread do not measure. Since 2023, depth and spread explain about a fifth of the variation in the score on BTCUSDT and less on ETHUSDT, while price range explains 61% to 76% in every year. As a volatility indicator, churn matches a causal price-range signal (one that uses only past data) on hit rate but does not beat it. The two scores correlate at 0.84, a day-old reading recovers most of the apparent edge, and the score moves with the price move and does not lead it. On the fairest construction, about half of churn's high readings come when the range signal is silent. We see Book Churn first as a direct measure of order-book activity, with volatility as one of its uses.

## Appendix

### When no score is returned

These rules matter if you rebuild the score from raw counts. The zscore is null:

- for the first fourteen days of an instrument's history;
- when the same-hour window holds fewer than fourteen days of usable intervals, for example after a long feed outage;
- for an interval with no updates;
- for an interval with less than a tenth of the trailing all-hours average, once that average rests on at least 84 intervals;
- when the window has no variation at all.

Intervals excluded by the third and fourth rules are also left out of later windows. The last rule applies to a window, not to an interval.

Two more details matter when reconciling early history. The warm-up ends at fourteen days but the window is thirty, so scores in between use a shorter, noisier window. An interval with partial feed coverage is scored as a quiet interval, not excluded, unless it falls below the tenth-of-average cutoff. Finally, because the window holds at most 180 values rather than 4,320, the standard deviation is estimated less precisely than a thirty-day window suggests, which makes extreme readings more common.

### Why the log transform

We take the log because update counts are heavily skewed toward high values. The 99.9th percentile interval carries 5.4x the median count on BTCUSDT and 5.6x on ETHUSDT, and the most extreme interval reaches 10x the median on BTCUSDT and 13x on ETHUSDT. Without the log, a single burst would dominate the thirty-day standard deviation. This is a design choice; we have not tested the score without it.

### How the score is distributed

The score has roughly unit scale: mean 0.01 and 0.03, standard deviation 1.09, ranging from −8.04 to +8.24 on BTCUSDT and from −7.61 to +9.10 on ETHUSDT. It does not follow a normal distribution, though. With a standard deviation of 1.09, a normal distribution would put about 3.3% of intervals above 2 and 0.30% above 3\. The actual shares are 4.3% and 4.6% above 2, and 0.94% and 0.98% above 3\. Readings above 2 are therefore 30% to 40% more common than a normal distribution implies, and readings above 3 more than three times as common. At the low end, the low-activity rule removes intervals below a tenth of normal activity. Intervals just above that cutoff, typically at 10% to 25% of normal activity, still score very low, down to about −8; treat them as likely data gaps rather than quiet markets. Set thresholds from the observed percentiles, not from standard deviations.

### How the volatility test was built

All figures in the volatility test, apart from the re-run on other instruments, use Binance BTCUSDT and ETHUSDT perpetuals from 1 January 2020 to 30 September 2026: 705,719 scored intervals out of 709,920 expected. Every expected interval has a row. Of the 4,201 without a score, 4,032 fall in the fourteen-day warm-up and 169 under the exclusions above.

For each interval *t*, the forward measure is the sum of the six squared ten-minute log mid-price returns from *t+1* to *t+6*, so it excludes the scored interval's own return. Ratios are reported in volatility terms, as the square root of that measure.

We compare each spike with the median forward outcome of non-spike intervals in the same group. A group is a calendar year, an hour of day, and a decile of realized volatility over the previous hour, measured up to and including the scored interval, with decile boundaries set within each year and hour. We include the current interval because a trader sees its price action at the same moment as the churn reading. We match on hour of day because the outcome itself follows a daily pattern: median forward realized variance at 16:00 UTC is roughly double that at 08:00\. The score is already adjusted for hour, so this step removes the daily pattern in what is being predicted, not in the signal. Runs of consecutive spikes count once, at their start. Standard errors come from a block bootstrap over blocks of thirty consecutive days that contain an episode. This gives wider errors than clustering by day, which suits volatile periods that last for weeks. Figures shown with ± are one standard error.

### By year

Hit rate at zscore ≥ 2 by year:

| Year | BTCUSDT | ETHUSDT | Episodes, BTC / ETH |
| ---- | ------- | ------- | ------------------- |
| 2020 | 53.3%   | 56.0%   | 1,090 / 1,210       |
| 2021 | 56.5%   | 55.0%   | 1,045 / 968         |
| 2022 | 64.6%   | 63.3%   | 900 / 980           |
| 2023 | 68.3%   | 67.2%   | 1,279 / 1,362       |
| 2024 | 70.6%   | 68.7%   | 875 / 1,019         |
| 2025 | 66.5%   | 70.0%   | 845 / 962           |
| 2026 | 64.4%   | 63.7%   | 665 / 735           |

We keep the weak early years in the sample; their 95% intervals, shown in Chart 5, include 50%. We cannot tell whether the weaker results for 2020 and 2021 reflect a change in market behavior or differences in the data. Those two years hold about a third of the episodes; without them the hit rate is 67.1% ± 2.0 on BTCUSDT and 66.8% ± 1.5 on ETHUSDT.

![Chart 5. Hit rate by year at zscore ≥ 2 for both instruments, with 95% block-bootstrap intervals and the 50% baseline, 2020 to 2026 (2026 through 30](https://storage.ghost.io/c/25/7e/257e6d23-09bb-4d1c-b3ed-16ff70eb3f9f/content/images/2026/11/book_churn_chart5_hit_rate_by_year.png)

****Chart 5.** Hit rate by year at zscore ≥ 2 for both instruments, with 95% block-bootstrap intervals and the 50% baseline, 2020 to 2026 (2026 through 30 September).

### By volatility decile

An overall hit rate can hide a result that comes from only one part of the sample, so the table below splits it by volatility decile (BTCUSDT, zscore ≥ 2). The last two columns show the median realized volatility over the next hour for control intervals and for spike intervals in each decile. The ratio column is the median of each spike's ratio to its own control group (year, decile and hour), while the two level columns are medians across the whole decile. Dividing one level by the other therefore does not give the ratio.

| Trailing vol decile | Episodes | Hit rate    | Volatility ratio | Control next hour | After a spike |
| ------------------- | -------- | ----------- | ---------------- | ----------------- | ------------- |
| 1 (quietest)        | 13       | 92.3%       | 1.43             | 18 bps            | 45 bps        |
| 2                   | 31       | 83.9%       | 1.53             | 25 bps            | 39 bps        |
| 3                   | 42       | 78.6%       | 1.28             | 28 bps            | 42 bps        |
| 4                   | 91       | 71.4%       | 1.30             | 32 bps            | 43 bps        |
| 5                   | 186      | 58.6% ± 5.3 | 1.14             | 35 bps            | 40 bps        |
| 6                   | 321      | 64.8% ± 3.3 | 1.19             | 38 bps            | 46 bps        |
| 7                   | 485      | 64.5% ± 2.5 | 1.15             | 40 bps            | 48 bps        |
| 8                   | 838      | 61.7% ± 2.2 | 1.16             | 45 bps            | 52 bps        |
| 9                   | 1,570    | 61.2% ± 2.1 | 1.14             | 52 bps            | 60 bps        |
| 10 (busiest)        | 3,122    | 63.7% ± 2.3 | 1.22             | 68 bps            | 78 bps        |

The hit rate is above 50% in every decile, but not all deciles are clearly above it. Decile 5, at 58.6% ± 5.3, cannot be told apart from 50% on its own, and deciles 1 to 3 hold only 86 episodes between them and should not be leaned on either. Below decile 5 there are too few episodes for a bootstrap error to mean much, so we show none. The result rests on deciles 6 to 10, which hold 95% of the episodes and score 61% to 65%, with errors of two to three points.

The volatility ratio is flat from decile 5 to decile 10, between 1.14 and 1.22, so the size of the effect is no more concentrated in the busiest deciles than the hit rate is. In absolute terms the effect is modest. In the deciles that hold most of the sample, a spike is followed by an hour of 40 to 78 basis points of volatility, against 35 to 68 for the control, a difference of five to ten basis points. We have not tested what, if anything, this means for execution or position sizing.

On ETHUSDT the hit rate is also above 50% in every decile, but the pattern is noisier and does not rise or fall steadily. Decile 3 scores 51.2% on 43 episodes, no demonstrable edge. Absolute volatility levels are higher throughout, 29 to 87 basis points for the control.

### How the price-range signal was built

We built the range signal the same way as churn: the log of the high-low range in basis points, ln(1 + range), scored against the same hour of day over the previous thirty days, with the same warm-up. Comparing an hour-adjusted signal with an unadjusted one would handicap one of them. We match the two signals on episode count rather than interval count. The test is scored on episodes and churn readings last longer, so matching on intervals would make churn artificially more selective per episode.

The simpler version flags intervals whose high-low range falls in the top few percent of the calendar year. It beats churn, but it ranks each interval against the whole year, including intervals that had not happened yet. It cannot be traded, so we do not use it as the benchmark.

### Hour of day and ten-minute slot

By UTC hour, the share of readings above 2 still ranges from 2.3% to 5.1% on BTCUSDT and from 2.8% to 5.3% on ETHUSDT. It is lowest at 14:00, the hour with the most order-book updates. The ten-minute slot at the top of the hour has a reading above 2 in 5.6% of cases, against 3.3% for the slot at :50 on BTCUSDT (5.9% against 3.5% on ETHUSDT), because scheduled activity clusters on minute boundaries. Comparing the :00 and :30 slots with the other four, this excess has grown from about 1.2x before 2024 to roughly 1.4x since on BTCUSDT and 1.5x on ETHUSDT. We adjust for the hour rather than the ten-minute slot because a slot-level baseline would hold only thirty values over thirty days, far too few to estimate a standard deviation.

### Re-run on other instruments

A re-run on Binance BNBUSDT, DOGEUSDT, SOLUSDT, XRPUSDT and TRXUSDT, with data to 19 October 2026, gives hit rates of about 60% to 65%, against 57% to 61% for the day-old placebo. On none of them is churn measurably worse than the price-range signal. Note that thresholds for these instruments are not validated. HYPEUSDT has traded only since May 2025, too short a history for these tests to be conclusive.