AI-generated research, checked but not peer-reviewed. It may contain errors.

E872　 KILL　 No signalTested · October 2026

Round-number price levels cannot be told apart from placebo levels

In 2024-2026 a stop order just beyond a fresh round price lost 252.5 bp over four hours, 26.3 bp less than at placebo prices, with an interval from -11.9 to +63.1 bp that includes zero.

round-numbers stop-orders liquidation-clusters perpetual-futures placebo-levels hourly-bars

Readable summary Full research record JSON recordFeedback on E872

Traders often say that stop-loss orders and liquidation prices pile up at round prices, so a price that breaks through a fresh round level should keep running. We took the first crossing of such a level in crypto perpetuals and entered with a stop order just beyond it. In 2024-2026 the entry lost 252.5 bp over four hours. The same entry at placebo levels, with the same spacing but not round, lost 278.8 bp. The 26.3 bp gap has an interval of -11.9 to +63.1 bp, so round levels were not shown to differ; the entry itself was a loser.

Validation crossings54,852round levels, 2024-2026; 53,615 at placebo levels

Round minus placebo, 4 h+26.3 bp95% interval -11.9 to +63.1 bp

Round-level entry, 4 h-252.5 bpBTC-adjusted, before fees; the pass line was +16 bp

Interval half-width37.5 bpthe test was sized for a 7-10 bp minimum detectable difference

01 · The idea being tested

Stop orders and liquidation prices cluster at round numbers, so the first crossing of a fresh round level triggers a cascade of forced orders that carries the price further.

Why people believe it

People place stop-loss orders and set leverage so that liquidation falls at prices like 60,000 or 0.50. When the price touches such a level, those orders turn into market orders whatever the price, and the push should continue for a few hours. A stop order resting just beyond the level would be filled by that cascade itself, so there is no race to react after the fact. The pre-registration cites a 2003 study of currency markets for the clustering of stops at round numbers; this test is on crypto perpetuals, not currencies.

Why it would persist The argument, as stated before testing

Stops belong to retail traders who do not choose levels by expected return, and a faster participant can exploit a cluster only by trading against its own customers' flow.

02 · How it was tested

For every liquid crypto perpetual we listed all round prices (such as 60,000, 65,000, 70,000 or 0.50, 0.55) and kept only levels the price had not touched for a week. The first time the price crossed such a level by 0.1%, we simulated a stop order that buys (or sells) at that point and exits at the close of the bar four hours later (a hold of four to five hours), or one hour later in the secondary version. As a control we repeated the whole procedure on a shifted grid of "placebo" levels with the same spacing, so any difference between the two grids is attributable to roundness, not to crossing a level in general.

Pass required

Validation round-minus-placebo above 0 with cluster 95% interval lower bound above 0, and round continuation gross of at least 16 bp at 4 hours.

Killed if

Validation round-minus-placebo interval includes 0, or the difference is 0 or below.

Costs

All returns in this record are before fees and slippage; no costs were deducted (a mechanism screen). The pass line carried the cost: a gross continuation of at least 16 bp at 4 hours, covering 5 bp fees on both sides plus slippage.

Controls

Placebo grid with the same density, coins and times (decides the verdict, statistic = round minus placebo). Both signs were stated in advance: continuation (primary, above 0) and fade (below 0, reported).

03 · What happened

Why it was rejected

The validation round-minus-placebo interval (-11.9 to +63.1 bp) includes zero, and the round-level entry itself averaged -252.5 bp against a pass line of +16 bp.

Validation (2024-2026), 4-hour hold: the round-level entry averaged -252.5 bp (54,852 crossings) and the placebo-level entry -278.8 bp (53,615 crossings). Round minus placebo is +26.3 bp, interval -11.9 to +63.1 bp.
At a 1-hour hold the difference is +24.2 bp (interval -7.7 to +57.0 bp). In the selection period (2020-2023, 21,907 round and 21,851 placebo crossings) the 4-hour difference is +29.6 bp with an interval of -39.5 to +105.7 bp.
The interval half-width of 37.5 bp is about four to five times the 7-10 bp minimum detectable difference planned in the pre-registration, so a difference of a few tens of bp can be neither shown nor excluded.
The pass rule also required a gross continuation of at least +16 bp at 4 hours; the round-level entry averaged -252.5 bp, median -130.1 bp, and 35.9% of round-level entries were profitable (34.1% at placebo levels).
Up-crossings (longs) averaged -104.2 bp at round levels and -122.8 bp at placebo levels; down-crossings (shorts) averaged -384.2 and -426.6 bp. The worst single round-level trade lost 54,719 bp.

Round minus placebo: the interval includes zero

Difference in mean BTC-adjusted return between round and placebo levels, with 95% coin-day cluster intervals. Every interval crosses zero, and the validation interval is wide.

Chart data (bp)

	Mean	95% low	95% high
Validation, 4 h hold	26.3	-11.9	63.1
Validation, 1 h hold	24.2	-7.7	57
Selection, 4 h hold	29.6	-39.5	105.7

The stop entry loses at round and placebo levels alike

Round levelsPlacebo levels

Validation mean 4-hour return of the stop entry, by crossing direction. Both grids lose in both directions, far below the +16 bp gross that a pass required; the loss is largest after down-crossings.

Chart data (bp)

	Round levels	Placebo levels
Up-crossings (long)	-104.2	-122.8
Down-crossings (short)	-384.2	-426.6
All crossings	-252.5	-278.8

Same distribution at round and placebo levels

Round levelsPlacebo levels

Number of validation 4-hour trades by return in bp (54,852 round, 53,615 placebo). The two shapes are nearly identical: most trades lose, with heavy tails on both sides.

Chart data (trades)

Bin	Round levels (count)	Placebo levels (count)
< -800	7,869	8,082
-800..-400	7,263	7,357
-400..-200	8,384	8,362
-200..-100	5,777	5,851
-100..-50	2,987	2,972
-50..0	2,894	2,705
0..50	2,524	2,339
50..100	2,160	2,017
100..200	3,319	2,893
200..400	3,747	3,483
400..800	3,207	3,053
> 800	4,721	4,501

04 · What it means

Why it failed Interpretation · not established

The data show that round levels were not distinguishable from placebo levels and that a stop beyond a fresh level lost heavily on both grids. One explanation, not established by the test: on hourly bars a trigger 0.1% beyond a level is reached by price spikes that then reverse within the hour, so the entry is filled on the extreme of the bar. Roundness adds nothing visible on top of that, because the placebo grid suffers the same fills.

Lesson

Judge an entry rule against its own baseline before judging the signal. Here the average trade lost about 250 bp whether the level was round or not, so a raw result would say "stops at round numbers are bad" while the comparison says nothing about round numbers. The second lesson is about power: here the realised interval was several times wider than the one planned.

Scope of this result

Covers hourly bars of USD-margined crypto perpetuals, levels on a 0.5 x 10^k grid, a trigger 0.1% beyond the level, fresh levels only, exits at the close of bar h+1 and h+4, fills assumed at the trigger price, and returns before fees. Hourly bars do not show the order of prices within an hour, so the simulated fill is an assumption, not an observation.

A rejected experiment applies to the tested conditions and does not prove an idea can never work. bp = basis point, 0.01%.

1. Test design

Pre-registered

Yes: rule and criteria fixed before the test data was examined

Data

Perpetual futures 1-hour candles (open, high, low, close, quote volume) — largest USD-margined crypto perpetual venue, public market-data archive; 1 hour; 2020-01..2026-10; public; includes delisted contracts

Universe

Point-in-time universe of USD-margined perpetuals: trailing 30-day median daily volume of at least $1M and at least 720 hours of history. 152,225 fresh crossings in total (round and placebo), 54,852 round crossings in the validation period.

Signal

Round grid: multiples of step = 0.5 x 10^floor(log10 P), P = previous close (for example 60k/65k/70k, 0.50/0.55). A level is fresh if it lay inside no hourly high-low range in the previous 168 hours. Up-cross in hour h: close(h-1) below the level and high(h) at least 1.001 x level. Down-cross: close(h-1) above the level and low(h) at most 0.999 x level. One event per coin per level and direction per 7 days. Placebo grid: the same grid shifted by u x step, with u drawn uniformly from (0.2, 0.8) per coin-month.

Execution

Resting stop at the trigger: long at 1.001 x level after an up-cross, short at 0.999 x level after a down-cross; fill assumed at the trigger price. Exit at the close of bar h+4 (primary) and h+1 (secondary), so the hold is 4 to 5 hours after the intrabar entry (1 to 2 hours for the secondary). Return = direction x (exit / entry - 1), so a positive number means continuation in the direction of the crossing. Return is adjusted by subtracting the BTC return from the previous bar's close to the exit bar's close (beta 1); BTC itself is raw.

Costs

Controls

Sample split

Selection period 2020-2023, validation period 2024 to 2026-10 (decision on validation). Coin-day cluster bootstrap of the round-minus-placebo difference.

Pass criterion

Validation round-minus-placebo above 0 with cluster 95% interval lower bound above 0, and round continuation gross of at least 16 bp at 4 hours.

Kill criterion

Validation round-minus-placebo interval includes 0, or the difference is 0 or below.

Power

Planned: about 30,000 events (about 10,000 effective), 4-hour adjusted standard deviation about 250 bp, minimum detectable difference about 7-10 bp. Realised validation interval: -11.9 to +63.1 bp, a half-width of 37.5 bp (derived).

2. Result

Observations

54,852

Instruments

663

Net mean

-252.5 bp

95% interval

Not reported

Median

-130.1 bp

Worst

-54,719 bp

Finding 1

Validation (2024-2026), 4-hour hold: the round-level entry averaged -252.5 bp (54,852 crossings) and the placebo-level entry -278.8 bp (53,615 crossings). Round minus placebo is +26.3 bp, interval -11.9 to +63.1 bp.

Finding 2

At a 1-hour hold the difference is +24.2 bp (interval -7.7 to +57.0 bp). In the selection period (2020-2023, 21,907 round and 21,851 placebo crossings) the 4-hour difference is +29.6 bp with an interval of -39.5 to +105.7 bp.

Finding 3

The interval half-width of 37.5 bp is about four to five times the 7-10 bp minimum detectable difference planned in the pre-registration, so a difference of a few tens of bp can be neither shown nor excluded.

Finding 4

The pass rule also required a gross continuation of at least +16 bp at 4 hours; the round-level entry averaged -252.5 bp, median -130.1 bp, and 35.9% of round-level entries were profitable (34.1% at placebo levels).

Finding 5

Up-crossings (longs) averaged -104.2 bp at round levels and -122.8 bp at placebo levels; down-crossings (shorts) averaged -384.2 and -426.6 bp. The worst single round-level trade lost 54,719 bp.

3. Interpretation

Why it failed

Lesson

Scope

4. Reproducibility

Steps

1) Take 1-hour OHLC for the point-in-time perpetual universe (30-day median volume of at least $1M, at least 720 hours of history). 2) Build the round grid (step = 0.5 x 10^floor(log10 close)) and the placebo grid (shifted by u x step, u uniform in 0.2-0.8 per coin-month). 3) Mark a level fresh if no hourly range in the previous 168 hours contained it; detect crossings with the 0.1% beyond-level trigger and one event per coin, level and direction per 7 days. 4) Entry at the trigger, exit at the close of bar h+4 and h+1, subtract BTC over the same hours. 5) Coin-day cluster bootstrap of round minus placebo for 2020-2023 and 2024-2026.

Grade

COMPLETE