Why
The closing observation in the source is the best part, and it deserves to be the headline. A perpetual future has no central price-correcting process. The protocol emits one incentive — funding — and dozens of independently-owned bots, each trying to take money off the others, collectively produce the tracking behaviour. The mechanism is not a component of the system; it is the population of participants competing. That is a genuinely elegant piece of market design and it is worth naming precisely, because the same shape recurs everywhere in this catalogue: liquidations, oracle updates, MEV, bridge relaying — all of them are jobs nobody is assigned and everybody races for.
But the worked example has a unit error worth fixing, and fixing it exposes where the risk actually lives. The comparison offered is 2% basis against 0.75% total cost, and in shape that is exactly right — enter only when expected reward clears cost plus a risk premium. The problem is that the 2% is never paid to you. A dated future converges because delivery forces it: hold to expiry and the gap is realised, so basis is the payoff. A perpetual has no expiry, so there is no moment at which the gap must close and nothing that hands it over. What is actually paid is funding, per interval, for as long as you hold.
Which means the honest inequality contains a term the example hides. Not basis > cost, but expected funding per interval × number of intervals held > total cost. And the holding period is unknown at entry. A basis of 2% is a signal that funding is currently rich, not a receipt for 2%. If it compresses in one interval you collect one interval's funding and pay the full round trip; if it persists for a week you collect a week of it. The trade is therefore a bet on persistence, not on convergence, and that reframing changes what should be modelled — the source's more advanced list (funding forecasts, volatility, execution probability) is exactly the machinery for estimating the holding period, which is why professional systems have it and the simple version does not.
The asymmetric-liquidation problem the source raises is the sharpest practical point, and it is worth stating harder: delta-neutral is not risk-neutral. The position has no directional exposure to price. It has enormous exposure to where margin sits. Spot long on one venue and perp short on another means two independent margin engines, and a fast move makes one leg profitable and the other liquidatable — the profit is unrealised on venue A while venue B closes you at the worst moment. Nothing about the hedge is wrong; the hedge is only a hedge if both legs survive to be unwound together, and survival is a funding-and-collateral logistics problem, not a pricing one. That is why the professional answer is continuous margin monitoring and capital movement rather than a cleverer entry rule.
And the third thing follows from the first: the observed basis is not disagreement about price. If the band were wider than the cheapest arbitrageur's total cost, that arbitrageur would already have traded it away. So the width of the band the basis lives inside is a readout of the cost structure of the cheapest participant — venue fees, their size, their cost of capital, their required premium. Read that way, a persistent spread stops being noise and becomes information about who is present, which is a measurement anyone can take from public data without holding a position.
How it works
The loop, in eight steps
| # | Step | The part that is not obvious |
|---|---|---|
| 1–2 | Observe spot and perp, compute basis = (perp − spot) / spot | Sign matters: positive basis pays shorts, negative pays longs |
| 3 | Estimate funding income | This, not the basis, is the payoff |
| 4 | Price fees, slippage, cost of capital | Slippage is a function of your size, not the venue's |
| 5 | Enter only if reward > cost + risk premium | The premium is what pays for step 7 going wrong |
| 6 | Buy spot, sell perp — near-simultaneously | The gap between the two fills is unhedged, and it is where execution risk lives |
| 7 | Monitor basis and margin on both venues | The margin half is the one that actually kills positions |
| 8 | Unwind both legs together on convergence | "Convergence" is not guaranteed — there is no expiry to force it |
The worked example, and the unit error in it
Spot $100, perp $102 → basis +2%. Costs: round trip 0.15%, slippage 0.10%, cost of capital 0.2%, risk premium 0.3% → 0.75%.
| As stated | Corrected | |
|---|---|---|
| Test | 2% > 0.75% → enter | Expected funding × intervals held > 0.75% |
| Why the difference | Assumes the 2% is collected | Nothing pays you the 2% — a perp never expires |
| What the 2% actually is | The payoff | A signal that funding is currently rich |
| Hidden variable | — | Holding period, unknown at entry |
| So the trade is a bet on | Convergence | Persistence |
And at basis 0.3%: 0.3% < 0.75% → do nothing. That second case is the important one — an algorithm that never declines is not an algorithm, and the width of the no-trade band is the whole product.
Delta-neutral is not risk-neutral
| Exposure | Present? |
|---|---|
| Direction of price | No — that is the point of the hedge |
| Funding rate flipping sign | Yes |
| Margin fragmentation across venues | Yes, and it is the one that liquidates you |
| Execution gap between the two fills | Yes, briefly and completely |
| Venue solvency / withdrawal halt | Yes — see third-party-blast-radius |
A sharp move makes one leg deeply profitable and the other liquidatable at the same instant, and profit on venue A cannot post margin on venue B. This is why the professional system spends its complexity budget on collateral logistics rather than on entry signals, and it is the same fragmentation aqua-shared-liquidity claims to solve — a claim this card gives a way to test.
Why this is a market-design result, not a trading one
No participant is assigned the job of tracking spot. The protocol pays funding; bots compete for it; their orders move both prices toward each other. The peg is an emergent property of adversarial participants, which makes it robust in a way a designated market maker is not — and fragile in a different way: it holds only while the funding payment is large enough to cover somebody's costs. In a thin market with high venue fees, the cheapest arbitrageur's floor is high, the no-trade band is wide, and the perp is allowed to drift exactly that far.
Read alongside two-currencies-one-ledger, which makes the same move on a different asset — a claim about shared liquidity resolved into an observable spread — and the-settlement-instant, since every step above depends on a spot price somebody chose how to compute.
Why it belongs in this catalogue
A prediction market has the same structure with the arithmetic simplified: YES + NO should cost $1, and when it does not, the same population of bots is what pulls it back. The band around $1 is the same readout — the cheapest participant's cost floor — and it is measurable on any live venue today. Whatever settlement design this repository ends up with, the no-trade band is the number that says how accurate its prices can possibly be.