On a perpetual future, longs pay shorts every 8 hours whenever funding is positive. When funding is extreme, a crowded, leveraged long book is paying a lot just to stay open. The bet is that this book eventually unwinds, and price lags in the following weeks. There is a mirror image for extreme negative funding.
crypto derivatives flow / funding rate
Crypto extreme-funding reversal
The idea
How it was tested
While preparing the test, an earlier backtest of the long side, extreme negative funding, turned up. It had been run on 2026-09-11, never registered, and its script was later lost. It had failed. The long side was declared burned and not retested; only the short side was pre-registered, on 2026-09-13.
The signal: funding in the top decile of its own trailing 365-day distribution, computed causally, per coin. BTC and ETH only, the two with daily funding history. Short for 20 calendar days, non-overlapping episodes. The benchmark is an unconditional 20-day short of the same coin over the same window, because being short crypto loses money on its own as prices rose on average. Fewer than 30 episodes would void the test. Pass: t of at least 1.65.
Where it died
One run, the same day. The signal did worse than the plain short, with a t of −1.25: not just insignificant, but the wrong sign. The lost long-side run had pointed the same way, with t −0.48 on 80 episodes. Both sides of the mechanism were measured, and neither showed that extreme funding predicts a reversal.
The record
Thesis
When the perpetual-futures funding rate is extremely positive, leveraged longs pay shorts every 8 hours just to keep their positions open. That crowded, expensive long book should eventually unwind, so price should underperform in the following weeks. Mirror image for extreme negative funding.
Specification
- signal
- Funding rate in the top decile of its own trailing 365-day distribution (causal percentile, no look-ahead), per symbol
- universe
- BTC-USD, ETH-USD (the only two with daily funding history)
- entry
- short, non-overlapping episodes, one per symbol at a time
- exit
- 20 calendar days
- benchmark
- An unconditional 20-day short of the same coin over the same funding-data window, so the test asks whether the signal adds anything to "being short crypto", which itself loses because crypto rose on average
- void rule
- fewer than 30 episodes
Gates
Pre-registrationpass
Was the hypothesis, its parameters and its pass/fail rule written down before any data was looked at?
Long side declared burned before writing; short-side methodology written from scratch, one run.
Market-adjusted alpha (legacy)fail
Does the strategy beat SPY over the same holding windows, assuming beta = 1?
t_HAC ≥ 1.65 (single pre-registered side)
t (HAC)-1.25alpha / trade %-2.65trades69
Alpha = strategy return minus the unconditional short benchmark (beta fixed at 1). Wrong sign, not just insignificant.
Why it died
Shorting the top funding decile did worse than simply being short crypto over the same period: alpha −2.651% per 20-day episode, t −1.25, on 69 episodes. The burned long side had failed the same way (alpha −1.075%, t −0.48, 80 episodes). Neither side of the mechanism shows that extreme funding predicts a reversal. Only 4 of 7 years had the right sign.
What it taught
An obligation to pay is not an obligation to trade badly. Leveraged longs do pay funding automatically, but they choose to keep paying, and the payment itself does not force the unwind the thesis needed. The test also shows why a forgotten earlier run has to be declared: had the long-side result stayed hidden, a second threshold on the same mechanism would have looked like a first attempt.
Verdict history
- 2026-09-11Long side backtested (not registered, script later lost): alpha −1.075%, t −0.48, n 80 No alpha
- 2026-09-13Long side declared burned; short side pre-registered and locked
- 2026-09-13Single run of the short side: t −1.25, wrong sign No alpha