The basis trade's risk is a tail, not a path: 28,802 entries and the drawdown that never came

What this is about

A dated futures carry trade is one of the oldest structures there is. Buy spot, sell a future that expires in three months, hold to expiry. At settlement the future converges to spot by construction, so your profit is the basis you locked in on day one — whatever the price does in between.

Its crypto cousin is the perpetual funding carry. A perpetual future never expires, so instead of convergence it has funding: a cash transfer every eight hours between longs and shorts that pushes the contract back toward spot. Buy spot, short the perp, collect funding. Same economic position, different mechanism.

The dated version comes with a standard warning, and it is a good one: convergence is guaranteed at expiry, but nothing is guaranteed before it. The basis can widen after you enter. You mark to market daily because that is what margin is computed on. So you can be completely right about the destination and be stopped out on the way.

I had been repeating that warning. Then I measured it.

What was measured

venue        Binance USD-M
contracts    BTC and ETH quarterlies: 251226, 260327, 260626, 260925
window       2025-06-28 .. 2026-08-18 (417 days)
resolution   HOURLY
observations 35,138
entries      28,802 held-to-expiry, each with at least 30 days to run

The basis comes straight out of the mark-price stream, which carries both mark_price and index_price on the same rows, so no spot join is needed. The same stream confirms the mechanism: funding_rate is 0.00000000 on every dated row.

The quantity under test is maximum adverse excursion — the worst unrealised loss between entry and expiry. It is a path statistic, and it is the right one, because the terminal outcome is already known. The question is not what you earn. It is how far under water you go on the way to earning it.

The drawdown that never came

Against a median entry basis of 91 bps:

MAE median      9.2 bps
p95            50.7 bps
p99           110.7 bps
max           395.2 bps

At the typical worst moment you are about a tenth of your eventual profit under water. That is not a risk anyone needs to manage.

The 395 bps maximum looked like it might be the interesting part. It was not. All ten entries above 200 bps had an entry basis of zero, and every one of them sits at its contract’s first available hour. They are not dislocations — they are the ordinary post-listing basis build-up, measured from an entry with no carry to collect. Opening a carry trade at zero basis is not a carry trade. The “adverse excursion” is just the market pricing in the carry you declined to take.

Apply any entry rule at all and the tail evaporates:

min entry basis   entries   median   p99     max    MAE > entry basis
       0 bps    28,791      9.2   110.7   395.2            5.7%
      10 bps    28,052      8.9   110.5   134.2            3.2%
      25 bps    26,154      8.7   111.2   134.2            1.2%
      50 bps    22,299      8.9   112.6   134.2            0.0%

With a 50 bps floor — a modest bar against a 91 bps median — the worst excursion in the entire dataset is 134 bps, and not one of 22,299 entries was ever under water by more than its own eventual profit.

There was one apparent cluster: 237 of the top-1% excursions peaked on the same day, 23 July 2025. That is not an event either. 1,226 entries peak there, with a median entry basis of 296 bps and a median excursion of 78 bps, and none of them exceeded their own entry basis. The basis simply touched 382 bps in a week it was running 340–370.

So where is the risk?

Here is the honest statement, and it is more useful than the comfortable one: there is no stress episode anywhere in these 417 days. Not a small one. None.

That is not a clean bill of health. It is the opposite of a measurement. The warning is about what happens when funding markets dislocate — when everyone is on the same side, the arbitrage capital that normally closes the basis withdraws, and the spread goes somewhere it has never been. My window contains no such event, so it cannot speak to it.

What it can say is that the two risks are not the same shape, and conflating them leads to the wrong risk management:

The day-to-day mark-to-market risk is effectively bounded. Under a sane entry rule it never exceeded 1.34% of notional across 417 days. Sizing a buffer for normal conditions is close to a non-decision.

The tail risk is unbounded and unmeasured. Survival depends entirely on one number — how much buffer you hold against an event that is not in your sample, and by construction cannot be.

People say “the basis trade marks to market, be careful.” The data says the marking is a non-issue and the carefulness should all be pointed at the thing you have never seen. Those are different jobs.

The thing that actually separates them

If the risk difference is mostly unmeasurable, what is the difference between a dated carry and a perpetual funding carry? It turns out to be the income, and it is not subtle.

Perpetual funding can go negative, and does. When funding is negative, a long-spot/short-perp book does not earn a reduced carry — it pays. Measured on Binance over the same 417 days:

token    % periods negative   longest run      net carry
BTC                  21.0%    14 (~4.7 d)         +3.86%/yr
ETH                  24.9%    16                  +2.99%/yr
DOGE                 30.6%    13                  +3.48%/yr
SOL                  43.3%    22 (~7.3 d)         -0.75%/yr
PENGU                38.8%    19                  -1.64%/yr
SEI                  57.8%    29 (~9.7 d)        -10.31%/yr

Three of six sampled tokens had net negative perpetual carry over the window. A hedged book holding SEI did not earn less. It paid 10.31%/yr for the privilege of being hedged.

A dated trade entered at a positive basis cannot do this. The terminal value is fixed at entry. The basis can collapse, go negative, do anything at all — and because the position is short the basis, a falling basis is a gain. At expiry it is zero by settlement and you collect what you locked in.

That is the real trade, and it is an option sale: dated futures convert an uncertain income stream into a certain number. That is insurance against negative funding and, in the same move, a cap on a funding spike.

Two markets that disagree, persistently

Across the window the two delivered almost the same median carry — 3.55%/yr dated against 3.26%/yr perpetual — and the dated side won 54.2% of 28,802 matched observations. A coin flip.

The average hides the finding entirely:

quarter    BTC dated minus perp    ETH dated minus perp
2025Q3                  +0.71                   +0.51
2025Q4                  +0.28                   -0.11
2026Q1                  +2.44                   +2.43
2026Q2                  +1.31                   +0.14
2026Q3                  -2.29                   -2.37

In 2026Q1 BTC perpetual funding was negative in 37.4% of periods and the carry collapsed to roughly 1%/yr, while the dated basis — locked at entry — held 3.50%. The insurance paid. In 2026Q3 perpetual funding ran to 6.21% and dated gave 3.92%. The cap cost. Same mechanism, opposite sign.

These are two markets pricing the same service and disagreeing by more than two percentage points for quarters at a time, in both directions. They are not substitutes, and the interesting book is probably not “pick one”.

The cost of turning over

One more structural difference, and it favours the dated side. A dated position is rolled when the calendar says so — four times a year. At a measured 30 bps round trip that is about 120 bps/yr, or 33.8% of the median basis on offer.

A perpetual carry book rotating monthly consumes 72% of a 5%/yr carry; weekly rotation consumes 312% and is loss-making outright. The calendar imposes a long holding period that discipline otherwise has to supply, and that is worth roughly half the turnover cost.

What this does not say

One venue, two tokens. Binance BTC and ETH quarterlies. Nothing here generalises to alt-coin dated contracts, which are thinner and would behave worse.

417 days, one regime, no dislocation. This is the load-bearing limit, and it is the whole reason the piece is framed around where the risk is rather than whether it exists.

No execution cost was measured. I pulled mark prices, not order books, for the dated contracts. The roll cost above imports a perpetual round trip as a proxy, and dated contracts are thinner, so the real figure is higher.

Capacity is unmeasured and almost certainly much smaller. Dated open interest is a fraction of perpetual open interest.

The one-line version

The mark-to-market risk everyone warns about in the cash-and-carry trade did not show up in 28,802 entries over 417 days — the worst excursion under any sane entry rule was 1.34% of notional, and nothing was ever under water by more than its own profit. The risk is real, it is entirely in a tail this data does not contain, and in the meantime the thing that actually separates dated futures from perpetual funding is that one of them cannot pay you a negative carry and the other one did, on half my sample.