Funding-rate cash and carry: the honest numbers

Buy spot, short the perpetual against it, collect the funding payments. The position is market-neutral, the payments arrive every eight hours, and the pitch writes itself. It is one of the most confidently marketed "yield" strategies in crypto.

It works. It also spends much of its life doing nothing, and across the post-2022 regime it would have returned less than a savings account. Both halves of that are true at once, and the second half is the half nobody quotes.

What was tested

The strategy has two legs, and the backtest engine behind this library is single-leg, so this ran on a dedicated simulator built for it: real spot klines, real perpetual klines, and real funding history, with both legs priced and both legs charged.

Rules pinned before the run:

• Enter when the mean of the last three funding settlements exceeds 0.01% per 8h.
• Exit when it falls below 0.003%.
• Costs: roughly 0.3% per episode across both legs, entering and exiting.
• Capital: returns are computed on the total capital committed — two legs, not one. This matters. Quoting the return on the perp margin alone roughly doubles the headline.
• Basis P&L from real prices, so convergence and divergence between spot and perp are in the result rather than assumed away.

Assets: BTC, ETH, BNB. The same two independent windows and the same gates as every other sweep in the research log — positive in both windows, surviving strip-best, drawdown within the bar.

The result

BTC and ETH clear every gate in both windows — and that verdict has now been reached twice: by the August simulator, whose code was never committed, and by a simulator rebuilt from the rules above and committed to the repository in October 2026 (scripts/research/replicate-sweep8-carry.ts). The two agree on every verdict and disagree on the magnitudes, so both are shown and the difference is explained rather than averaged away.

Returns are on total capital, both legs. The worst single episode across every committed run cost −0.4% of capital (the August figure was −0.2%). For a strategy in a category where blowups are the main risk, that is the number worth staring at: the mechanism behaves.

On how few episodes that is, because it matters. The BTC and ETH windows hold between three and nine episodes — a small sample by any standard, and sweep #10 kills a different hypothesis on exactly this ground, at two to seven trades per four-year window. The distinction we are drawing, and it is worth disagreeing with: that kill was a directional strategy, where three trades means three bets on direction and a single one is a third of the evidence. This is a carry position with no directional exposure, where an episode is a count of occasions the funding rate was hot, and the return inside each is the sum of many individual settlements rather than one outcome. The strip-best check still runs and still means something, but the sample is thin and the honest framing is that this is a mechanism whose payments we can verify, not a strategy with a large track record.

The drawdown column is the one place the two simulators genuinely disagree. The August run reported mid-episode marks of 13.4% and 18.1% of capital in the first windows and called them perp-versus-spot dislocations. The committed simulator, marking both legs at every settlement from real spot and perpetual prices, never sees a mark worse than 2.4% of capital, and no reading of the pinned rules — the mean of three rates, all three rates, the latest rate — produces the larger figures; the one that reads "all three rates above the gate" does reproduce three of the six August episode counts exactly. The August code cannot be re-examined. Treat the smaller number as the one that can be checked, and keep the lesson the larger one was teaching: an under-collateralised short is still what gets margin-called in a violent move, and that risk is not priced by either simulator.

BNB fails the second window's strip-best check at the primary entry threshold, and passes at the stricter pre-registered neighbour.

That neighbour grid is the most informative part of the whole test, because it is monotone in exactly the direction the mechanism predicts — and it reproduced cell for cell:

Selectivity is the edge. The strategy is not "collect funding"; it is "collect funding only when it is paying enough to cover two legs of costs, and sit out otherwise." Loosen the gate and you convert a working strategy into a fee-donation machine.

The honest framing

Here is the part that belongs at the top of every pitch and usually isn't:

• While deployed, the position earns roughly 8–20% annualised on the committed re-run, 6–14% on the August one.
• It is idle for much of its life — 40 to 80% of the window on the committed re-run, 70 to 90% in August. Hot funding is episodic: it clusters in leveraged-long manias and disappears in quiet markets, and every 2022–26 window sits at the idle end of that range.
• Overall, across the post-2022 quiet regime, it returns little more than cash. The committed re-run's second windows make 6.7% and 6.9% on BTC and ETH over four years — under 2% a year on total capital — against a risk-free rate that spent much of the period well above zero. The August run had them lower still.

So this is an opportunistic harvester for hot-funding episodes, not a yield product. It is something to switch on when perpetual funding is visibly elevated and switch off when it is not — and if you leave it running through a quiet year, the fees and the idle capital will do the rest.

That framing is why the entry ships with its characterisation stated first rather than in a footnote. A backtest that says "6–14% APR" without saying "for a fifth to a half of the calendar, and less in a quiet year" is a true sentence assembled into a false impression.

The risks that a backtest cannot price

Market-neutral is not risk-free, and the simulator prices only what prices can express:

• Counterparty and venue risk. The position holds spot on an exchange and a short perp against it. An exchange failure, a withdrawal freeze, or an auto-deleveraging event during a violent move is a real tail that no historical funding series contains.
• Manual two-leg execution. The strategy language is single-leg, so this cannot be adopted as a Sensei plan and cannot be automated here. You would be placing and managing both legs yourself, including the rebalancing as the basis moves.
• Funding-schedule changes. The settlement mechanics and rate caps are exchange policy, not physics. They have changed before.
• Crowding. The trade is well known. Elevated funding attracts exactly this position, which is part of why elevated funding does not persist.

How this fits the rest of the research

This sweep was an angle change, and it is worth saying why.

Seven previous sweeps had established that sub-daily crypto strategies built on price shape are dead at real costs — the full story is in does any low-timeframe crypto strategy actually work?. The thing that makes those timeframes hostile is fees: the cost per trade is fixed while the edge per trade shrinks with the bar. Funding is a second charge on top, settled every eight hours a position is held through.

Worth being exact about that, because it cuts both ways: a long that enters while funding is negative is paid it rather than charged. That is precisely what the two surviving crypto entries do. So funding is not a uniform tax on holding — it is a transfer, and which side of it you are on is a choice.

So instead of trying again to predict sub-daily candles, this sweep asked the inverse question — can you be the one collecting the transfer rather than paying it? The answer is yes, conditionally, and the conditions are the whole story.

Where to read further

• Entry #6 — funding harvest, the full write-up with the episode ledger and the execution notes.
• Entry #1 and entry #5, the two directional strategies that read funding as a signal rather than as income.
• The research log for the sweep in its original chronological form.