Eleven thousand violations, one wrong input
Deribit's mark surface breaks static arbitrage 11,593 times across 88 snapshots. Its own quotes, scanned the same way, break it never. Most of the gap turns out to be a single stale number.
Every no-arbitrage relation on an option surface is stated relative to the forward. Put-call parity is a statement about the forward. A call spread is bounded by it. A butterfly is priced around it. Get the forward wrong and nothing else on the surface has to be wrong for the whole thing to look inconsistent, because every test you run is measuring from a mark that has moved.
That is most of what I found. Deribit’s mark surface breaks static arbitrage 11,593 times across 88 snapshots of the BTC and ETH chains. The venue’s own bid and ask, scanned by the same code on the same snapshots, break it not once. And recovering the forward from the surface itself, rather than taking the one the venue publishes, improves the median violation sixteen fold.
Eleven thousand broken prices is a dramatic number. One stale input is a boring one. The evidence points at the boring one.
Why a mark is not just a quote
Deribit publishes a mark price and a mark implied volatility for every listed option. These are not decoration and they are not the same thing as the price you would trade at. They set margin requirements, and they decide liquidations. If your position is marked against a surface, the surface is the thing that can close you out, regardless of where the order book actually is.
So the question worth asking is not whether the marks are good estimates of fair value, which is unanswerable, but whether they are consistent with themselves. That is a much weaker test and a much more embarrassing one to fail.
The weakest test there is
Static arbitrage does not require a model, a volatility surface, or a view. It asks only whether a set of simultaneous prices can coexist without handing somebody a riskless profit: whether a spread costs less than nothing, whether a butterfly carries a negative price, whether more time can be worth less. No forecasting, no calibration, nothing about whether the market is right. Just internal consistency, checked in one instant.
- violations in the mark surface
- 11,593
- snapshots, BTC and ETH
- 88
- at or above one full tick
- 235
- in the venue's own bid and ask
- 0
- At or above one full tick
- 235
- Smaller than a tick
- 11,358
Two per cent of the violations are larger than the smallest increment the venue will quote. The rest are easy to wave away as rounding. The 235 are not.
That two per cent is the part I would defend in a room. Everything below a tick can be argued down to rounding, and I would rather concede it than spend the argument there.
The rate is steady rather than episodic: 155.8 violations per BTC snapshot and 107.7 per ETH snapshot. This is not a handful of bad moments. It is the ordinary state of the surface.
Wrong first
I set out to scan the quoted bid and ask for arbitrage.
There was none there at all, which should have been the end of the project. What actually happened is that my reconciliation against the marks failed first, and I spent a while assuming my own parser was broken before accepting that the clean series was the quotes and the inconsistent one was the surface the exchange margins you against.
Sixteen fold
The diagnostic that matters is not the count. It is what happens when you stop trusting the published forward.
Take the surface, discard the venue’s forward, and back out the forward implied by the marks themselves. Re-run exactly the same scan. The median violation falls by a factor of sixteen. The inconsistency does not vanish, but the bulk of it was never eleven thousand independently wrong option prices. It was one number, wrong in a way that displaced everything measured against it.
Why this matters more than the count
There is a residue. Of 1,056 slices, 1,036 fitted and 20 refused. I have not explained the 20, and I would rather say so than fold them into an average.
Why any of this is checkable
A number nobody can reproduce is a claim. So the scan exists twice: once in Rust and once in Python, written to the same specification and compared output for output. Across the capture they agree byte for byte. Two implementations agreeing does not make them right, but it does rule out the most likely way for a result like this to be an artefact, which is a bug in the one implementation that produced it.
The capture is committed to the repository. A fresh clone reproduces every figure above with no network access at all, which means the numbers cannot quietly change underneath the write-up when the venue’s API does.
What this does not show
Marks are not tradeable prices. Nothing here is an arbitrage anybody could put on, and the top of book was clean throughout, so this is a statement about the margin surface rather than about free money.
One venue. I have not checked whether other exchanges price their mark surfaces the same way, so nothing here generalises beyond Deribit.
The violation lifetimes are censored by the snapshot interval: a violation that resolves between two snapshots is recorded as lasting until the next one. Every persistence figure is therefore a lower bound, not an estimate.
The forward recovery is itself a model choice. It improves consistency by construction, so the sixteen fold figure should be read as evidence that the forward is the dominant term, not as proof that my forward is the correct one.