Funding, not fees: what actually kills a crypto perp cross-section

A while back we ran the cost-floor experiment: take a stack of signals that all die net of fees, move them to a venue that charges zero maker and zero taker, and see which ones cross into the black. Nothing crossed. The conclusion was negative and slightly unsatisfying — the fee was never what killed them — because it named what wasn't responsible without naming what was.

This is the follow-up. It names it.

The setup

Cross-sectional statistical arbitrage on USDT perpetuals, single exchange, across Bybit, OKX and Binance. A $250k dollar-neutral book, equal-weight top-25 long against bottom-25 short of a point-in-time top-50 universe — reconstructed as of each date, no constituent list applied backwards — at roughly $5k a name.

The discipline was to measure the costs before looking at a single signal. Taker fees read from private account endpoints rather than published schedules, which matters more than it sounds: Bybit prices fees per symbol, with six distinct groups inside the top 50, and OKX's aggregate fee endpoint reports numbers that disagree with the per-instrument-family values actually charged. Slippage measured by walking live order books to fill the real clip, sampled across a full 24-hour cycle. Funding measured per name over five years.

Then twenty-two hypotheses across four pre-registrations, every kill criterion written down before its test.

The boring half of the answer

Trading cost behaves exactly as the zero-fee experiment implied. At weekly and daily rebalancing it is a surcharge, not a wall. At four hours and one hour it is a wall, and no fee tier closes it — repricing everything at published retail schedules moves the bar by 8–36% and flips no verdict. Venue choice matters more than tier, and the gap between venues is slippage rather than fees.

Twenty-two hypotheses, none clears the bar. The one that came closest headlines at net Sharpe 1.8–2.0 over three to five years and passes four separate kill criteria individually — hold-out, volatility neutralisation, size neutralisation, survivorship. Each one shaves it. Stacked on an honest sample it is Sharpe 0.49–0.82 at t of 0.90 to 1.20: a real effect that cannot be told apart from noise, and roughly a third of it is drift-harvesting a falling altcoin market rather than cross-sectional mispricing.

That is a null result, and it is the least interesting thing here.

The term nobody puts in the P&L

A perpetual future has no expiry. Exchanges anchor it to spot with a periodic payment between longs and shorts: a long pays the funding rate, a short receives it. It is a cash flow, and it lands every eight hours whether you modelled it or not.

A dollar-neutral book cancels the market-wide level of funding almost exactly — we measured the neutral-book contribution at approximately zero over 20,000 draws. That is presumably why the term gets dropped: if it averages out across names, it looks like a wash.

It is not a wash, because the weights are not random with respect to funding. Any signal correlated with funding — and cross-sectional crypto signals very often are, since funding is itself a crowding proxy — produces a book systematically tilted to one side of the funding distribution. Per unit gross, in basis points per day:

Funding does not dominate everywhere — across all our book constructions it exceeds trading cost in about a fifth of them, and the fast reversal signals die to turnover long before funding gets a vote. But for the candidates that got close enough for the answer to matter, funding is what decided it. Two order-flow signals had gross returns that cleared their trading cost and were pushed negative by funding alone.

The part that surprised us

Look again at the funding column. It is positive for the funding-sorted book and negative for the low-vol book. That is not a rounding artefact — it is the whole point.

Everyone's intuition is that leaving a cost out of a backtest flatters it. Funding is not a cost. It is a cash flow with a sign, and which sign you get depends on which side of the crowd your signal puts you. A book ranked on funding is short the names that pay, so it collects — leave funding out and you understate that strategy by 5.3 bp/day. A low-vol book happens to sit on the paying side, so leaving funding out flatters it by 3.4.

For the funding-sorted book the omitted term is not merely large, it is the entire economics. Gross price return is negative on both venues with long history: ranking on funding tells you essentially nothing about which coin goes up. The book earns because it is short the side that pays. Compute the P&L without funding and its Sharpe goes from +0.74 to −0.49 — you conclude the exact opposite of the truth, and you blame transaction costs on the way out.

So omitting funding is not a conservative simplification. It is a bias whose direction you cannot sign without computing the thing you skipped.

This happens in print

A 2026 SSRN working paper evaluates three cross-sectional screeners on ten Binance perpetuals and concludes, reasonably, that they fail. Its signal set includes a funding-rate rank with the hypothesis that high funding implies short — the same construction as our funding-sorted book. Its decision frequency is eight hours, matching Binance's funding grid exactly, so every position it holds crosses a funding timestamp. Its net-return equation subtracts turnover times a five basis point round trip, and nothing else. Funding is a feature in that paper and never a cash flow.

We want to be fair about this, because the paper is careful in ways many are not — it embeds a parameter-free control model, and it discloses a look-ahead leak it found in its own pipeline. Its headline conclusion also survives: we tested fifty names with point-in-time construction rather than ten fixed large caps, which is the alternative its own discussion wonders about, and we also found nothing tradable. The correction is to the attribution, not the verdict.

And before drawing a moral about the literature, we went and checked whether this is common. It is not — and for an interesting reason. Almost all crypto cross-sectional factor research runs on spot, where there is no funding term to forget. The population of published cross-sectional perpetual studies is tiny, and the peer-review-track one in it defines its headline metric as “price movement plus funding fee yields.” One preprint made a specific mistake. That is a smaller claim than we wanted to make, so it is the one we are making.

What we threw away

The honest part of this write-up is what did not survive. Three drafts came out of this project and all three were withdrawn.

The first found that the standard Grinold-style rule for sizing a cross-sectional signal misprices these books, and traced the error to two measurable components that lined up monotonically with cross-sectional kurtosis across three asset classes. It looked publishable. Then we derived the closed form instead of asserting the mechanism, and it failed on four counts at once — the tail indices implied by the two statistics contradicted each other on every panel.

The second reframed the work around a cost-penalty multiple, chosen specifically because the badly-identified constant cancels out of a ratio. An adversarial audit returned ten blocking findings. The constant cancels only conditional on the mapping we had already falsified, and the multiple priced a full round trip every period, which no real strategy pays. For our own best signal the true daily penalty is about 1.15×, not the 5–6× we advertised.

The third was the funding result written up as a methodological finding about the literature. The survey above killed it.

What is left is a measurement that has now survived a derivation, an adversarial audit and a literature check: funding is large, its sign is not signable in advance, and it decided every candidate in this study that got far enough for the answer to matter. Price it before you do signal research on perpetuals. It is free from every venue's public API and it is bigger than the friction you are already modelling.

The full study — tools, all four pre-registrations, the derived results every number is computed from, and a research log recording seventeen of our own claims that were withdrawn along the way — is public at github.com/ssanin82/dtc-stat-arb.