Archive / Crypto & DeFi / E320
A hedged funding-rate gap between two venues does not pay
Long the perpetual with the lower funding rate and short the same coin on another venue lost 8.7 bp net per 24-hour trade over 3,099 events, because only 18.4 bp of the 185 bp the gap implied was collected.
A perpetual future pays a funding rate that differs from one exchange to the next. The belief: when the same coin's funding differs by a wide margin, buy it where funding is lower, short it where funding is higher, and collect the gap with no price risk. We tested gaps of at least 50 bp per 8 hours on 3,099 events and held each pair for 24 hours. The pair lost 8.7 bp net on average (interval -24.6 to +6.7 bp), against the +20 bp required. It collected 18.4 bp of funding where the known gap implied 185 bp, and the price gap between the two venues moved against it.
When the same coin pays a much higher funding rate on one venue than on another, a hedged pair collects the difference.
Why people believe it
People argue that funding is a periodic payment between long and short holders of a perpetual, set separately on each venue. If one venue's longs pay far more than another's, holding the opposite sides of the same coin on the two venues cancels the price move and leaves the funding difference as income.
Why it would persist The argument, as stated before testing
People argue that traders keep collateral on one venue and cannot cheaply move it, that funding rates can sit at venue-set caps, and that holding positions on two venues needs margin on both, so a gap would take time to close.
Every hour we compared the last funding payment on two large derivatives venues for the same coin and flagged cases where the two rates were at least 25, 50 or 100 bp per 8 hours apart. We then bought the coin's perpetual on the venue with the lower funding rate, shorted the same coin on the other venue with equal size, and closed both 24 hours later. The result is the funding collected, plus whatever the price gap between the venues did, minus costs. The 50 bp threshold was named as the primary test before any return was computed.
Stated as the inverse of the kill rule: primary mean net above +20 bp per 24 hours on gross notional, interval above 0, all three windows positive, worst-1% pair loss no more than 10 times the mean. Separately pre-registered: a Benjamini-Hochberg test at q = 0.10 across 9 cells (3 thresholds times pooled, larger-venue-low, other-venue-low).
Primary mean net at or below +20 bp per 24 hours on gross notional, or interval includes 0, or any of the three windows negative, or worst-1% pair loss above 10 times the mean.
7 bp per leg per side, 4 legs: 28 bp on one-leg notional, which is 14 bp on gross notional (2N). Returns are quoted on gross notional unless stated; per one-leg notional they are twice as large. Stress: 20 bp per side, 40 bp on gross notional. Funding paid by the long leg and received by the short leg is taken at the venues' historical rates for every settlement between entry and exit.
993 control pairs on the same symbols at random hours with a gap below 5 bp per 8 hours (up to one per symbol per 30 days, seeded), built the same way: this is the cost and basis noise baseline. Sampling of X = 25 was capped at 1,500 seeded events (1,486 priced); X = 50 and X = 100 were priced in full.
At the primary 50 bp threshold the mean net of -8.7 bp is below the +20 bp bar, the interval -24.6 to +6.7 bp includes 0, two of three windows are negative (-20.1 and -33.2 bp), and the 1st percentile of pair net, -373 bp, is 43 times the absolute mean: the result met all four pre-set rejection criteria.
- Primary (gap of at least 50 bp per 8 hours, net on gross notional): 3,099 events on 629 entry dates, mean -8.7 bp, median -14.4 bp, 95% interval -24.6 to +6.7 bp. Per one-leg notional the mean is -17.3 bp; under the 40 bp stress cost it is -34.7 bp.
- By time window the primary net was +3.3, -20.1 and -33.2 bp. Gross before the 14 bp cost the mean was +5.3 bp.
- Thresholds: at 25 bp per 8 hours the mean was +0.6 bp (interval -16.1 to +24.1, 1,486 events); at 100 bp it was -19.5 bp (interval -46.2 to +7.2, 1,578 events). None of the 9 cells is both positive and significant after Benjamini-Hochberg control at q = 0.10.
- When the larger venue was the lower-funding one, the pair lost more: -32.2 bp at 50 bp (interval -61.9 to -7.2, 998 events) and -66.8 bp at 100 bp (interval -134.0 to -13.2, 379 events). When the other venue was lower the means were +2.5 bp (interval -15.2 to +20.1, 2,101 events) and -4.5 bp (interval -31.5 to +22.3, 1,199 events), with medians of -19.0 and -24.4 bp.
- Controls with no gap (gap below 5 bp per 8 hours, 993 events) lost 13.6 bp net (interval -15.9 to -10.5), which is the 14 bp cost: before cost they earned +0.4 bp.
- Funding collected was about a tenth of what the known gap implied over 24 hours. At 50 bp the mean signal was 123 bp per 8 hours, implying 185 bp on gross notional, while 18.4 bp was collected. At 25 bp: 100 implied, 11.0 collected. At 100 bp: 345 implied, 31.5 collected.
- The price gap between the two venues moved against the pair by 10.8 bp at 50 bp (4.7 bp at 25 bp, 24.0 bp at 100 bp; a negative value means the long leg underperformed the short leg). Gross before cost was +14.6, +5.3 and -5.5 bp at the three thresholds, against the 14 bp cost; the gross also reflects the -100% cap on a doubled short leg.
- The short leg's 5-minute high reached at least +20% above entry in 15.4% of events at 50 bp (13.7% at 25 bp, 17.7% at 100 bp) against 2.2% in controls; at least +50% in 4.2% against 0.3%; at least +100% in 1.1% against 0.1%. The worst single pair lost 8,560 bp and the 1st percentile was -373 bp.
- In 39 events where the extreme venue's last two settled rates were identical and at least 50 bp per 8 hours, a possible venue cap, the short leg rose 20% or more in 33.3% and 100% or more in 7.7%; mean net was -213 bp (interval -664 to +11).
- Removing 47 events from 10 symbols whose two perpetuals were mismatched or redenominated (entry price ratio more than 10% apart) left the primary at -5.1 bp (interval -20.5 to +9.5, 3,074 events).
- A descriptive replication against a third venue, 58 events on 49 dates from 2026-07 to 2026-10, averaged -144.5 bp (median -21.2): funding collected +50.6 bp, price gap -41.4 bp, and one event in which the short leg rose 154% lost 6,766 bp.
No cell is significantly positive after BH; two are significantly negative
Mean 24-hour net with its 95% entry-date-clustered interval. Every interval includes 0 except the two cells where the larger venue had the lower funding rate, which are significantly negative. The pooled means are +0.6, -8.7 and -19.5 bp, and the no-gap control sits at the 14 bp cost.
Chart data (bp net per pair (gross notional))
| Mean | 95% low | 95% high | |
|---|---|---|---|
| Gap 25 bp/8h, pooled (n 1,486) | 0.6 | -16.1 | 24.1 |
| Gap 50 bp/8h, pooled, primary (n 3,099) | -8.7 | -24.6 | 6.7 |
| Gap 100 bp/8h, pooled (n 1,578) | -19.5 | -46.2 | 7.2 |
| 50 bp, larger venue lower (n 998) | -32.2 | -61.9 | -7.2 |
| 50 bp, other venue lower (n 2,101) | 2.5 | -15.2 | 20.1 |
| 100 bp, larger venue lower (n 379) | -66.8 | -134 | -13.2 |
| 100 bp, other venue lower (n 1,199) | -4.5 | -31.5 | 22.3 |
| Control, no gap (n 993) | -13.6 | -15.9 | -10.5 |
The known gap is mostly gone by the time it is collected
Funding implied by the gapFunding collectedPrice gap move (negative = against the pair)
Per gap threshold: funding the known gap implied over 24 hours, funding actually collected, and the price-gap move against the pair. Collected funding is about a tenth of the implied amount, and the price move takes back more than half of it at 50 and 100 bp.
Chart data (bp on gross notional over 24 h)
| Funding implied by the gap | Funding collected | Price gap move (negative = against the pair) | |
|---|---|---|---|
| Gap 25 bp/8h (n 1,486) | 100 | 11 | -4.7 |
| Gap 50 bp/8h (n 3,099) | 185 | 18.4 | -10.8 |
| Gap 100 bp/8h (n 1,578) | 345 | 31.5 | -24 |
The short leg squeezes far more often than in controls
Gap 25 bp/8hGap 50 bp/8h (primary)Gap 100 bp/8hControl, no gap
Share of events in which the short leg's 5-minute high rose at least 20%, 50% or 100% above entry during the 24-hour hold. At the 50 bp threshold 15.4% reached +20% against 2.2% in controls; the +100% case, 1.1% against 0.1%, is where a leg can be liquidated and the pair loses far more than the funding it collected.
Chart data (% of events)
| Gap 25 bp/8h | Gap 50 bp/8h (primary) | Gap 100 bp/8h | Control, no gap | |
|---|---|---|---|---|
| Short leg up 20% or more | 13.7 | 15.4 | 17.7 | 2.2 |
| Short leg up 50% or more | 3.3 | 4.2 | 5.1 | 0.3 |
| Short leg up 100% or more | 0.6 | 1.1 | 1.9 | 0.1 |
Why it failed Interpretation · not established
The gap is measured from the last settled rate, which reflects the premium of the previous interval. By the time the pair is open, the premium that produced the gap has largely converged: only about a tenth of the implied funding was collected, and the price gap between the venues moved by an amount that offsets much of what was. Large gaps also cluster in coins that are crowded or capped, where short-leg spikes make one-sided losses far larger than the carry. This explanation fits the numbers but the test did not isolate it.
Lesson
A hedged carry built on a rate that has already been paid is a bet that the rate persists. Collected carry as a share of implied carry (about 10% here) can be checked before price risk enters, and a hedge that cancels the coin's price move leaves the gap between the two venues' prices and the liquidation risk of each leg, neither of which is carry.
Scope of this result
Covers USDT-margined perpetuals with identical symbols on two large venues, January 2024 to October 2026, entries at the hour after the last settled rates, 24-hour holds, equal notional, 7 bp per leg per side. The third-venue replication is descriptive, 58 events over about three months. It says nothing about periods or venues outside these.
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-registeredDataUniverseSignalExecutionCostsControlsSample splitPass criterionKill criterionPower2. Result
ObservationsInstrumentsNet mean95% intervalMedianWorstFinding 1Finding 2Finding 3Finding 4Finding 5Finding 6Finding 7Finding 8Finding 9Finding 10Finding 113. Interpretation
Why it failedLessonScope4. Reproducibility
StepsGrade