A perpetual future has no expiry. To keep its price close to spot, exchanges make one side pay the other every 8 hours: when the perpetual trades above spot, leveraged longs pay the shorts. Buy the coin spot, short the same amount of perpetual, and the price risk cancels. What is left is the funding the longs pay, plus any convergence between the two prices.
carry / basis arbitrage
Crypto spot-perpetual basis arbitrage
The idea
How it was tested
Pre-registered on 2026-07-26, with a projected 10-30% a year. Entry when the basis exceeds 0.5% and 7-day average funding exceeds 0.01% per 8-hour period, on BTC and ETH. Exit on convergence or after 30-60 days. One position, both legs of equal notional. Assumed costs: 15 bps on the spot leg and 0.04 bps on the perpetual. The pass rule was written in advance: at least 5% annualized out of sample, archive below 2%.
Where it died
The day after registration, 60 days of real exchange data contradicted the model on both legs. The basis had the opposite sign to the one modelled, and funding averaged 0.0037% per 8-hour period, about an eighth of the 0.030% per period the plan assumed. Expected P&L was negative.
The two-year backtest, on 2026-08-20, confirmed it. Mean basis was 0.0125% for BTC and 0.0136% for ETH, a small fraction of the 15 bps it cost to open the spot leg. The entry threshold fired once in 730 days.
On 2026-09-09 the premise itself fell. The broker sells crypto only as spot, so the cheap perpetual leg never existed there. Its 0.04 bps cost was a decimal slip for 4 bps.
The record
Thesis
Buy BTC/ETH spot and short the same amount of perpetual futures: the position is delta-neutral and collects the funding rate that longs pay shorts, plus any convergence of the spot-perpetual basis.
Specification
- entry
- basis > 0.5% and 7-day average funding > 0.01% per 8-hour period
- exit
- basis convergence or a 30-60 day time exit
- sizing
- one position, spot and perpetual legs of equal notional
- costs assumed
- 15 bps on the spot leg, 0.04 bps on the perpetual leg
Gates
Pre-registrationpass
Was the hypothesis, its parameters and its pass/fail rule written down before any data was looked at?
Entry rule, exits and pass/fail criteria (OOS ≥ 5% annualized, archive below 2%) written before the backtest.
Gate F — Frictionfail
Is the alpha per trade large enough to pay for its own execution?
mean basis larger than the round-trip cost
mean basis btc pct+0.01mean basis eth pct+0.01cost bps15trades in 730 days1
Two-year backtest. The entry threshold (basis > 0.5%) fired once in 730 days, so there is no sample to speak of.
Why it died
The spread it was meant to capture averaged 0.0125% (BTC) and 0.0136% (ETH), about a hundredth of the 15 bps it cost to put on. Live exchange data from the same summer was worse than the model in both parts of the trade: the basis had the opposite sign to the one modelled, and funding averaged 0.0037% per 8-hour period — about 1.4% a year against the 32.85% the plan assumed. The premise underneath was also wrong: the broker sells crypto only as spot, so the perpetual leg it priced at 0.04 bps did not exist there, and that figure was itself a decimal slip for 4 bps.
What it taught
A perpetual future has no expiry date, so "the basis converges mathematically at expiry" — the pre-registration's central claim — was never true. The only anchor is funding, which is a variable carry that can turn negative, not an arbitrage. Checking the instrument's contract terms would have killed this before any backtest.
Verdict history
- 2026-07-26Pre-registered with a projected 10-30% a year
- 2026-07-2760 days of real exchange data: basis sign inverted, funding about 1.4% a year — expected P&L negative
- 2026-08-20Two-year backtest: mean basis far below commission, 1 trade; archived Eaten by friction
- 2026-09-09Found that the broker has no crypto perpetual contract; crypto there is spot only