The idea is intuitive. If a coin's funding rate stays extremely positive for weeks, longs are crowded and paying heavily to stay in. Short it: if the crowd unwinds, the price falls, and you collect their funding while you wait. Mirror it for extremely negative funding. Two sources of return from one signal.
It sounded good enough that we wanted to test it properly. We wrote the expected outcome down first ("low", based on earlier results), and it came out worse than that.
Why this is on a funding-arbitrage site. This strategy bets on price direction and treats funding as a bonus. FundingArb does the opposite: it holds a delta-neutral, hedged position so that funding is the only thing it is exposed to. This test shows what happens to funding income when the price risk is left open.
The rule, fixed before the test
- Signal: each coin's average funding rate over the past three weeks, annualised.
- Short any coin at or above +30% p.a.; long any coin at or below −40% p.a.
- At most 10 positions per side, the most extreme first, 5% of capital each. Everything else stays in cash.
- Rebalanced every Monday, no stops, volatility-scaled, trading costs and daily funding payments included.
The thresholds came from the distribution of the signal, not from returns: +30% p.a. is roughly the 96th percentile of all coin-weeks and −40% p.a. roughly the 8th. For comparison, the median coin-week sits at +4.2% p.a., and the default funding rate on most perps works out to about +11% p.a.
The universe is the one we built after our survivorship-bias test: each week, the 50 highest-volume Binance perpetuals of that week, from all 895 that ever existed, delisted ones included. January 2022 to August 2026.
Eight pass criteria, also fixed in advance: a placebo test significant after correcting for every test we had run so far; a positive Sharpe ratio; stability across both halves and the walk-forward windows; a positive result without any single coin; a positive result at double costs; acceptable losses in five historical crash windows; a positive contribution from both legs; and enough position-weeks per leg to judge it.
The result
| Universe | APR | Max drawdown | Sharpe |
|---|---|---|---|
| Point-in-time, incl. delisted (the test) | −10.7% | 61.0% | −0.33 |
| Same rule on a hindsight list of 50 surviving coins | +2.3% | 19.3% | 0.26 |
A loss of about a tenth of the capital per year, and a drawdown of more than half. It failed seven of eight pass criteria; the only one it met was the crash replay, and with an average gross exposure of just 25%, that says little.
On a list of coins that are known to have survived, the rule roughly breaks even. On the coins that were actually tradable at the time, it peaks in early 2025 and then loses more than half of its value.
Every check agrees
Placebo. On every rebalancing day we shuffled the funding values randomly among the 50 coins, so the random portfolios held the same number of extreme positions, just on the wrong coins. We did this 10,000 times. 95% of these random portfolios did at least as well as the real rule.
The real rule sits at the 5th percentile of chance (median of the random portfolios: 0.11). Picking the coins with extreme funding was worse than picking coins at random.
Stability over time. One good year cannot carry it.
2026 covers January to August only.
| Window | Sharpe | APR |
|---|---|---|
| First half (Jan 2022 – Jun 2024) | 0.28 | – |
| Second half (Jul 2024 – Aug 2026) | −0.88 | – |
| Walk-forward 1 (Jan 2022 – Mar 2023) | 0.11 | +0.1% |
| Walk-forward 2 (Mar 2023 – May 2024) | 0.45 | +8.4% |
| Walk-forward 3 (May 2024 – Jul 2025) | −0.85 | −18.6% |
| Walk-forward 4 (Jul 2025 – Aug 2026) | −0.93 | −28.9% |
Single coins. 226 coins were held at some point. Whichever one we remove, the Sharpe stays between −0.45 and −0.24. No single blow-up explains the loss.
Costs. Doubling trading costs barely matters (Sharpe −0.36), because the rule trades little. The losses come from the positions themselves, not from fees.
Crashes. Maximum drawdown inside each stress window was moderate: Luna 6.4%, FTX 7.4%, SVB 1.7%, yen carry unwind 1.4%, October 2025 flash crash 3.1%. The damage came from slow bleeding between crashes, not from the crashes.
Both sides lose
The rule mostly sat in cash or held a few longs: on average 1.1 shorts and 3.9 longs per week. Shorts were open in 68 of 241 weeks, longs in 218; 9 weeks were entirely in cash. Run on their own, both legs lose money:
| Leg on its own | APR | Max drawdown | Sharpe |
|---|---|---|---|
| Short extreme positive funding | −6.7% | 43.8% | −0.32 |
| Long extreme negative funding | −6.0% | 51.2% | −0.18 |
Why: the funding is real, the price risk is bigger
To see the mechanism, we looked at every position the rule actually opened (259 short and 944 long position-weeks) and split each week's result into the part that came from price and the part that came from funding.
The funding diamonds sit close to zero; the price ranges span tens of percent. Whatever the funding pays, one ordinary week of price movement can take it back many times over.
| Per position-week | Short leg | Long leg |
|---|---|---|
| Funding signal at entry (median, p.a.) | +46% | −116% |
| Price change over the previous 30 days (median) | +44% | +15% |
| Price change in the holding week (median) | −5.3% | −3.9% |
| Price change in the holding week (average) | +1.75% | −2.72% |
| Funding implied by the trailing signal | +1.11% | +4.28% |
| Funding actually collected | +0.43% | +2.29% |
| Net per week (price + funding) | −1.32% | −0.43% |
Price changes are those of the coin; for the short leg, a falling price is a gain. Equal-weighted averages over all position-weeks, before costs and volatility scaling.
- Extreme positive funding marks a pump in progress. The median short candidate had already risen 44% over the previous 30 days. In most holding weeks it drifted lower (median −5.3%), which looks like the short is working. But in one week out of ten it rose another 36% or more, and those squeezes outweighed all the small wins: on average, the coin rose 1.75% a week against the short.
- Extreme negative funding marks a coin the market is betting against, often right after a rally (median rise of 15% over the previous 30 days). These coins kept falling, by 2.72% a week on average. The funding paid to longs was substantial (2.29% a week) and still did not cover the slide.
- Trailing funding overstates what you collect. Funding fades after an extreme reading: the short leg actually collected about 39% of what the trailing three-week average implied, the long leg about 54%. The same point applies to annualised funding figures in general, as our fee-coverage article explains.
Put together: the rule buys and sells exactly the tails that keep running, and the funding it collects for doing so is small next to the moves it is exposed to.
A curated coin list would have hidden this
On a hindsight list of 50 coins that survived to 2026, the same rule roughly breaks even (APR +2.3%, Sharpe 0.26) instead of losing 10.7% a year. Threshold rules react to the universe even more strongly than ranking rules, because a curated list removes exactly the extreme cases they trade. It is the same lesson as in our survivorship-bias test, only sharper. Even on the hindsight list, the result is far from anything worth trading.
Other settings don't rescue it
These variants were fixed in advance as descriptive checks, not as candidates to pick the best one from:
| Variant | APR | Max drawdown | Sharpe |
|---|---|---|---|
| Test setting: 3-week signal, +30% / −40% | −10.7% | 61.0% | −0.33 |
| 1-week signal | −14.5% | 66.9% | −0.55 |
| 2-month signal | −6.5% | 61.1% | −0.18 |
| Looser thresholds: +20% / −20% | −16.7% | 69.2% | −0.50 |
| Stricter thresholds: +50% / −80% | −7.8% | 52.4% | −0.31 |
| Double trading costs | −11.3% | 61.5% | −0.36 |
What we take away
- Extreme funding is information about price momentum, not a mispricing to harvest. It tells you a move is underway, and in this sample the move tended to continue.
- Funding income is small next to price risk. Collected funding was a fraction of a percent to a few percent per week, against typical weekly moves of 10% to 14%. Funding only becomes the dominant source of return when the price exposure is hedged away.
- Trailing funding is not future funding. Extreme readings fade; plan with what is actually collected, not with what the trailing average annualises to.
- Tail rules are fragile to the universe. Test them on the coins that were tradable at the time, delisted ones included.
- Set thresholds from the signal, not from the returns. Otherwise the backtest finds the thresholds that happened to work.
What this means for funding traders
Funding is real money: in these extreme weeks, the long positions collected 2.29% a week. These are the tails of the distribution, not typical rates, and they faded quickly. But as long as the position is also exposed to the price, the price decides the outcome. That is the whole reason delta-neutral funding arbitrage exists: holding spot against the perp removes the price exposure, so the funding is what remains. That is a statement about mechanics, not about returns: fees, borrowing, execution and the usual, much lower rates decide what a hedged position actually earns. How that structure works, and which risks it does not remove, is covered in What Is Delta-Neutral Funding Rate Arbitrage?
Methodology
- Data
- Binance USDⓈ-M perpetuals: daily candles, quote volume and funding; 895 symbols including delisted ones.
- Period
- 2022-01-15 to 2026-08-31.
- Universe
- Weekly top 50 by trailing 30-day quote volume; minimum age 220 days; sufficient recent funding history.
- Signal
- Average daily funding over the past three weeks, annualised. Short at ≥ +30% p.a., long at ≤ −40% p.a.; thresholds set from the signal distribution before the test.
- Portfolio
- Weekly rebalancing on Mondays; at most 10 positions per side at 5% each, otherwise cash; no stops; volatility-targeted.
- Costs
- Per side and per change in position, plus daily funding payments.
- Placebo
- 10,000 permutations of the signal values among the week's universe on each rebalancing day.
- Multiple testing
- Threshold 0.05 divided by the number of logged tests run so far (60 at the time). About 120 earlier exploratory hypotheses are not in this count; including them makes the threshold stricter and changes no verdict here.
- Known limits
- Binance data only; execution on other exchanges is not modelled. Few short positions (on average about one per week), so the short leg on its own is thin evidence. Price artefacts on delisting days are deliberately included.