AI-generated research, checked but not peer-reviewed.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

E057　 KILL　 Not replicatedTested · October 2026

Extreme funding rates do not predict reversals in perpetuals

Trading against extreme funding rates in crypto perpetuals netted -76.2 to +78.5 bp per short trade across six cells, every 95% interval included zero, and the sign flipped between sub-periods in all six.

funding-rate crowded-leverage perpetual-futures contrarian event-study

Readable summary Full research record JSON recordFeedback on E057

Perpetual futures charge longs or shorts a funding payment every few hours, and a very high rate is read as a sign that leveraged traders are crowded on one side and due for a reversal. We took the first settlement per coin at which funding reached +0.1%, +0.2% or +0.4% per 8 hours and shorted for 8 or 24 hours. The six short cells netted -76.2 to +78.5 bp per trade after a 14 bp cost. Every 95% interval included zero, and in each cell the first, middle and last parts of the period disagreed in sign. Buying after very negative funding lost money in all three 8-hour cells.

Short after high funding-76.2 to +78.5 bpnet per trade, 6 cells (3 thresholds, 8 h and 24 h); about 400 events each

Intervals excluding zero1 of 12a loss: long after negative funding, 0.1% threshold, 8 h, -68.5 bp

Short cells with windows agreeing in sign0 of 6selection, validation and holdout, 60/20/20 by date

Worst trade-10,014 bpa short that lost the full -100% floor plus cost

01 · The idea being tested

When the funding rate on a perpetual future is extreme, leveraged traders are crowded on one side, and the price is likely to reverse against them.

Why people believe it

Funding is paid by the side that holds the more popular position. People argue that a very high rate means many traders are leveraged long, will be squeezed out by a small drop, and that the funding payment also pays whoever takes the other side. A very negative rate is read the same way for crowded shorts.

Why it would persist The argument, as stated before testing

People argue that crowded leverage is hard to trade against: the coins with extreme funding are thinly traded, short squeezes are violent, and the contrarian position has to be held through them.

02 · How it was tested

We collected the full funding history of every USD-margined perpetual on the largest venue, 655 contracts and 2.6 million settlements, and rescaled each to a rate per 8 hours because some contracts settle every 4 hours. At each contract's first settlement with funding at or above +0.1%, +0.2% or +0.4% (48 hours apart at the earliest) we shorted at the start of that hour, and after funding at or below the same negative levels we bought. Each trade was held 8 or 24 hours, received or paid the actual funding after entry, and was charged 14 bp. Because the full event list was large, we sampled up to 400 events per cell, evenly spaced by event order.

Pass required

Reconstructed after the test; no criteria were fixed in advance. Net mean per trade above zero with a 95% interval above zero and the same sign in the three windows.

Killed if

Reconstructed after the test; no criteria were fixed in advance. Intervals include zero, or the sign differs between windows.

Costs

14 bp round trip.

Controls

Price and funding components reported separately for every cell. No same-contract control hour and no placebo events were run, so the price component is read against zero, not against a matched baseline.

03 · What happened

Why it was rejected

Every short cell has a 95% interval that includes zero and a net sign that differs between sub-periods, so no cell shows a result that holds.

Short after funding at or above +0.1%, +0.2% and +0.4% per 8 hours netted -14.1, +36.4 and +71.3 bp at an 8-hour hold and +18.7, -76.2 and +78.5 bp at 24 hours (399 or 400 events per cell, in 197, 178 and 153 contracts for the three thresholds). Net is after the 14 bp cost.
All six short intervals include zero, for example -64 to +129 bp (0.2%, 8 h) and -124 to +284 bp (0.4%, 24 h). Intervals are 131 to 408 bp wide across the 12 cells.
In every short cell the three windows disagree in sign. At the 0.4% threshold with a 24-hour hold the selection, validation and holdout windows netted +65, -199 and +401 bp; at the 0.1% threshold with an 8-hour hold, -25, -35 and +40 bp.
The short collected 9.5, 19.1 and 26.1 bp of funding at 8 hours (0.1%, 0.2%, 0.4%) and 26.3, 44.5 and 58.0 bp at 24 hours. The price component was -9.6, +35.1 and +59.6 bp at 8 hours and +6.4, -102.3 and +31.7 bp at 24 hours.
The net sign follows the price component in all six short cells. Price is the short's return, so a negative value means the coin rose against it: the two cells with negative net, 0.1% at 8 hours (-14.1 bp net, price -9.6 bp) and 0.2% at 24 hours (-76.2 bp net, price -102.3 bp), are the two with negative price.
Buying after funding at or below -0.1%, -0.2% and -0.4% lost 68.5, 35.7 and 47.6 bp at 8 hours: the price fell 82.2, 62.5 and 95.1 bp while 27.6, 40.7 and 61.5 bp of funding was received. Only the 0.1% interval (-134 to -2 bp) excludes zero, and it is a loss.
The long cells at 24 hours netted -50.5, +113.5 and -65.0 bp, with intervals from -169 to +75, -56 to +293 and -239 to +131 bp, all including zero.
Four of the six short cells contain a trade that hit the -100% floor, a short whose price at least doubled; the worst trade lost 10,014 bp including cost.
Extreme negative funding is more common than extreme positive funding: 2.9% of settlements at or below -0.1% against 1.6% at or above +0.1%.

Shorting after high funding: nothing separable from zero

Mean net return per short trade after funding at or above each threshold, with day-clustered 95% intervals, after the 14 bp cost. The means range from -76.2 to +78.5 bp and all six intervals include zero; the short intervals are 131 to 408 bp wide, much wider than the cost.

Chart data (bp)

	Mean	95% low	95% high
Funding at or above +0.1%, 8 h hold	-14.1	-81	50
Funding at or above +0.1%, 24 h hold	18.7	-96	119
Funding at or above +0.2%, 8 h hold	36.4	-64	129
Funding at or above +0.2%, 24 h hold	-76.2	-245	88
Funding at or above +0.4%, 8 h hold	71.3	-68	216
Funding at or above +0.4%, 24 h hold	78.5	-124	284

The sign flips between sub-periods

Selection (first 60%)Validation (next 20%)Holdout (last 20%)

Mean net return of the short by date window (60% / 20% / 20% of events) for each threshold and hold. In every one of the six cells the windows do not agree in sign; the 0.4% threshold with a 24-hour hold runs from +65 to -199 to +401 bp.

Chart data (bp net per trade)

	Selection (first 60%)	Validation (next 20%)	Holdout (last 20%)
+0.1%, 8 h	-25	-35	40
+0.1%, 24 h	-37	74	133
+0.2%, 8 h	-39	231	68
+0.2%, 24 h	-148	126	-64
+0.4%, 8 h	76	-151	282
+0.4%, 24 h	65	-199	401

Short side: funding, price and net per cell

Funding receivedPrice componentNet after cost

Mean funding received, price component and net return (after the 14 bp cost) per short trade for each threshold and hold. Net has the same sign as the price component in all six cells; the price component runs from -102.3 to +59.6 bp.

Chart data (bp per trade)

	Funding received	Price component	Net after cost
+0.1%, 8 h	9.5	-9.6	-14.1
+0.1%, 24 h	26.3	6.4	18.7
+0.2%, 8 h	19.1	35.1	36.4
+0.2%, 24 h	44.5	-102.3	-76.2
+0.4%, 8 h	26.1	59.6	71.3
+0.4%, 24 h	58	31.7	78.5

04 · What it means

Why it failed Interpretation · not established

Two readings fit the numbers. Funding at these levels may already be reflected in prices, so there is no systematic reversal to collect: the short's price component swings from -102.3 to +59.6 bp and has no stable sign. Second, the crowded side may simply be right, as the long side suggests: after negative funding the price kept falling by 62.5 to 95.1 bp over 8 hours, more than the funding received. With a per-trade standard deviation of 666 to 1,970 bp, none of the short cells has the power to show an effect the size of the 14 bp cost.

Lesson

In this test the funding rate alone gave no stable direction: the price component of the short changed sign across thresholds and holds. When coins with fat-tailed returns are traded in a handful of hundred-event cells, individual cell means of tens of basis points are inside noise of hundreds of basis points, and only sign agreement across independent windows and an interval that excludes zero separate an effect from the best of twelve cells.

Scope of this result

Covers first settlements per contract at funding of 0.1%, 0.2% and 0.4% per 8 hours (positive and negative), on USD-margined perpetuals of one large venue, entry at the settlement hour open, holds of 8 and 24 hours, a 14 bp cost, a sample of up to 400 events per cell, 2019 to 2026.

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

Data

Funding-rate history of all USD-margined crypto perpetuals, one record per settlement — largest USD-margined crypto perpetual venue, public market-data API; per settlement (1, 4 or 8 hours, normalised to a per-8h rate); 2019-09..2026-10; public | Perpetual futures 1-hour candle opens around each event — largest USD-margined crypto perpetual venue, public market-data API; 1 hour; 2019-12..2026-10; public

Universe

All 655 USD-margined perpetuals with at least 30 funding records: 483 settle every 4 hours, 171 every 8 hours and 1 every hour. Funding is normalised to a per-8h rate by the median spacing of the contract's settlements. Normalised funding was at or above +0.1% in 1.6% of settlements, at or above +0.2% in 0.4% and at or above +0.4% in 0.06%; it was at or below -0.1% in 2.9% and at or below -0.2% in 1.6%. The universe is not restricted to small coins; the 2,395 sampled events fall in 519 contracts, including large ones.

Signal

Event = first settlement per contract at which normalised funding is at or above +theta (short trade) or at or below -theta (long trade), theta in {0.1%, 0.2%, 0.4%} per 8 hours, with a 48-hour cooldown per contract. At most 400 events per theta and side are kept, spaced evenly by event order. Six trade cells by theta and side; each is held for 8 h and 24 h.

Execution

Enter at the open of the 1-hour candle that starts at the event settlement time, so the funding paid at that settlement is not received. Hold 8 or 24 hours. Funding received equals the actual rates at all settlements after entry up to exit (a short receives positive funding, a long receives negative funding). Simple returns, loss floored at -100%.

Costs

14 bp round trip.

Controls

Sample split

Events split by date, 60% / 20% / 20%: selection, validation, holdout. Day-clustered bootstrap 95% intervals (2,000 draws, resampling days).

Pass criterion

Reconstructed after the test; no criteria were fixed in advance. Net mean per trade above zero with a 95% interval above zero and the same sign in the three windows.

Kill criterion

Reconstructed after the test; no criteria were fixed in advance. Intervals include zero, or the sign differs between windows.

Power

About 400 events per cell. Per-trade standard deviation of net return is 666 to 1,970 bp across the 12 cells, so the 95% intervals are 131 to 408 bp wide. An effect near the 14 bp cost cannot be separated from zero in any cell. Sampling up to 400 events per cell, not using every event, is part of why the intervals are wide.

2. Result

Observations

2,395

Instruments

519

Net mean

—

95% interval

Not reported

Median

—

Worst

-10,014 bp

Finding 1

Short after funding at or above +0.1%, +0.2% and +0.4% per 8 hours netted -14.1, +36.4 and +71.3 bp at an 8-hour hold and +18.7, -76.2 and +78.5 bp at 24 hours (399 or 400 events per cell, in 197, 178 and 153 contracts for the three thresholds). Net is after the 14 bp cost.

Finding 2

All six short intervals include zero, for example -64 to +129 bp (0.2%, 8 h) and -124 to +284 bp (0.4%, 24 h). Intervals are 131 to 408 bp wide across the 12 cells.

Finding 3

In every short cell the three windows disagree in sign. At the 0.4% threshold with a 24-hour hold the selection, validation and holdout windows netted +65, -199 and +401 bp; at the 0.1% threshold with an 8-hour hold, -25, -35 and +40 bp.

Finding 4

The short collected 9.5, 19.1 and 26.1 bp of funding at 8 hours (0.1%, 0.2%, 0.4%) and 26.3, 44.5 and 58.0 bp at 24 hours. The price component was -9.6, +35.1 and +59.6 bp at 8 hours and +6.4, -102.3 and +31.7 bp at 24 hours.

Finding 5

The net sign follows the price component in all six short cells. Price is the short's return, so a negative value means the coin rose against it: the two cells with negative net, 0.1% at 8 hours (-14.1 bp net, price -9.6 bp) and 0.2% at 24 hours (-76.2 bp net, price -102.3 bp), are the two with negative price.

Finding 6

Buying after funding at or below -0.1%, -0.2% and -0.4% lost 68.5, 35.7 and 47.6 bp at 8 hours: the price fell 82.2, 62.5 and 95.1 bp while 27.6, 40.7 and 61.5 bp of funding was received. Only the 0.1% interval (-134 to -2 bp) excludes zero, and it is a loss.

Finding 7

The long cells at 24 hours netted -50.5, +113.5 and -65.0 bp, with intervals from -169 to +75, -56 to +293 and -239 to +131 bp, all including zero.

Finding 8

Four of the six short cells contain a trade that hit the -100% floor, a short whose price at least doubled; the worst trade lost 10,014 bp including cost.

Finding 9

Extreme negative funding is more common than extreme positive funding: 2.9% of settlements at or below -0.1% against 1.6% at or above +0.1%.

3. Interpretation

Why it failed

Lesson

Scope

4. Reproducibility

Steps

1) Download the full funding history of every USD-margined perpetual from the venue's public API and normalise each contract's rate to a per-8h value by the median spacing of its settlements. 2) Select, per contract, the first settlement with normalised funding at or above +theta (short) or at or below -theta (long) with a 48-hour cooldown; theta in 0.1%, 0.2%, 0.4% per 8 hours; keep at most 400 events per theta and side, evenly spaced by event order. 3) Download the 1-hour candles from the event hour and enter at the open of the event hour; exit 8 or 24 hours later. 4) Net return = price return (simple, floored at -100%) plus actual funding received after entry minus 14 bp. 5) Split events by date 60/20/20 and report mean, median, worst, win rate and day-clustered bootstrap intervals per cell.

Grade

COMPLETE