Why Most Crypto Arbitrage Spreads Are Not Executable

Two exchanges quote the same pair at different prices. Subtract one from the other and you have a gross spread. Every arbitrage scanner on the market can compute this, and the number it produces is real in the narrow sense that both quotes existed.

It is also the least useful number in the chain. Between the gross spread and a realised profit sit four separate deductions, each of which can independently take the entire spread: fees, depth, transfer time, and the latency of the data that showed you the gap in the first place.

This article is about those four. It does not tell you that arbitrage is impossible, and it does not tell you a particular spread is capturable. It describes what has to be subtracted before that question can even be asked.

A cross-exchange arbitrage is two trades, so it pays two sets of taker fees. Published spot taker fees on major centralised venues commonly sit in the range of a few basis points to around ten, before volume tiers and token-based discounts. The round trip is therefore roughly double whatever the headline rate is, and it is charged whether or not the second leg fills at the price you assumed.

For the majors on the largest venues, gross spreads are frequently thinner than that round trip. This is the ordinary case, not the exception: liquid markets are liquid precisely because participants with lower fee tiers than yours have already competed the gap down to their cost floor rather than yours.

The practical test is to compute the spread net of your own fee tier before anything else. A scanner showing a gross spread without subtracting fees is showing you a number that has not yet been asked the only question that matters.

A quote is a price for the next unit, not for your order. The relevant question is not what the top of the book says but how much you can transact before the price you receive stops resembling the price you saw.

This is where most apparently large spreads collapse. Wide gaps cluster on thin venues, and thin venues are thin at exactly the moment the gap appears. A 60-basis-point spread on a book that holds a few thousand dollars of depth at the touch is a 60-basis-point spread on a trade too small to be worth the operational overhead — and the spread compresses as you walk the book.

Sizing against displayed depth rather than persistent depth compounds this, particularly when book imbalance is elevated: displayed size can be withdrawn faster than an order can reach it, and reported volume says nothing about what a book will actually absorb.

• Compute the volume-weighted fill price across the levels your order would consume, on both legs
• Treat the spread as a function of size, not a scalar — it shrinks as size grows, and often crosses your fee floor before it crosses your target size
• Check whether depth at those levels has persisted, rather than sampling it once

The textbook version of the trade involves moving an asset from the cheap venue to the expensive one. In practice, on-chain transfer plus exchange crediting takes long enough that the spread that motivated the trade will usually have moved before the asset arrives, and withdrawal suspensions during volatile periods are common.

Practitioners therefore pre-position inventory on both venues and rebalance separately. This converts a timing problem into a capital problem: you now hold balances on multiple exchanges, and the returns must be assessed against that committed capital and the venue and custody risk it carries — not against the notional of a single trade.

This is the deduction that most retail analyses of arbitrage omit entirely. A strategy that yields a small edge per trade on capital that must sit idle across several venues is a very different proposition from the same edge on capital deployed once.

The gap you are looking at is, at best, as fresh as the slowest of the two feeds that produced it. This is measurable, so we measured it.

Across 16 centralised venues, Vultax recorded median order-book snapshot latency between 35 ms and 111 ms in a September 2026 sample. That range is respectable. The 95th percentile is the number that matters for arbitrage, and it ran from 288 ms to 590 ms — no venue in the sample held p95 under 100 ms.

A cross-venue comparison inherits the worse tail of its two inputs. When you compute a spread from two feeds each capable of a half-second tail, you are periodically comparing a fresh price on one venue against a stale one on the other. Some share of the spreads any scanner displays are artefacts of that skew rather than differences that existed simultaneously.

This is why Vultax describes its latency as a sub-100ms target rather than a guarantee, and why we publish the distribution rather than a single median.

Makarov and Schoar's study of cryptocurrency market arbitrage, published in the Journal of Financial Economics, remains the most careful published account of when these gaps persist and why.

Their central finding is that deviations are much larger across countries than within them, and that cross-country gaps can persist for days and weeks rather than seconds. During the period they studied, the average price ratio between the United States and South Korea reached roughly 40% at its extreme — the episode widely known as the kimchi premium.

The explanation is not that traders failed to notice. It is that capital controls, banking access and the practical difficulty of moving fiat between jurisdictions prevented arbitrage capital from closing the gap. The spreads persisted because they were not capturable, which is the general lesson: a spread that survives is usually one that something structural is preventing anyone from taking.

Within a single jurisdiction, and between venues with unconstrained capital flow, they found deviations far smaller and far shorter-lived. That is the regime most retail arbitrage takes place in, and it is the regime where fees and depth dominate.

The question to ask of any arbitrage tool is not how many opportunities it finds. Finding gross spreads is trivial. The question is what it subtracts, and whether it tells you.

Vultax reports raw price-gap context across connected venues, with the market-quality and liquidity-health context needed to assess whether a gap is supported by real depth. It does not represent gaps as executable profit, and it does not model your fee tier, your withdrawal limits, your inventory position or your own connectivity — all of which sit between a displayed gap and a filled trade. A tool that claims to have already accounted for all of that is claiming to know things about your account that it cannot know.

• Does it show gross or fee-adjusted spreads, and does it say which?
• Does it show depth alongside price, or price alone?
• Does it publish its own feed latency distribution, including the tail?
• Does it distinguish a gap that persisted from one that appeared in a single snapshot?
• Does it claim executability, and on what basis?

The strip at the top of this page is not a screenshot of the day this was written. Every ten minutes Vultax re-reads the best bid and ask on Coinbase, Kraken, OKX, Binance and Bybit, CoinGecko's depth within 2% of mid for six venues, and the trade tapes of the same five books and rewrites the figures; the Vi IQ beneath them scores the same six domains the terminal scores for a pair, computed for BTC on the major spot venues, with any domain that cannot be computed shown as unavailable rather than filled in. Read the numbers as a live check on the argument above, and the revision notes at the end for what has changed since publication.

The first figure is the widest gap between the five mid prices in basis points. The second is what remains of it after buying at the cheap venue's ask, selling at the dear venue's bid and paying an assumed 10 bp taker fee on each side; when it is negative, the gap is smaller than the cost of crossing it, which is the ordinary state of the market. The third is the mean best-quote spread across the five books, and the fourth is how many dollars sit within 2% of mid across six venues, the depth a real order would have to be measured against.

The cross-venue question has widened in 2026 rather than gone away. Polymarket launched perpetual futures on 3 September with up to 20x leverage and no expiry, sixty-seven markets within hours across crypto, equities, indices and commodities, three months after Kalshi's bitcoin perpetuals went live. A perpetual on a prediction-market venue quotes the same asset as a spot book on an exchange, and the gap between them is exactly the kind of number a scanner will show as an opportunity. It carries every deduction on this page plus one more, the funding payment that keeps a perpetual near spot, and it is barred to U.S. traders on the Polymarket side.

The venue-risk leg is also better documented than it was. Coinbase's own postmortem of its 7 May 2026 outage describes roughly eight hours down after an AWS thermal event and about twelve more to full recovery; in the October 2025 AWS outage Coinbase went dark while Binance, Kraken and OKX stayed up. An arbitrage that needs both legs to fill is exposed to whichever venue fails first, and the failures are not correlated the way a spreadsheet assumes.

Liquidity has thinned under the surface. Glassnode measured spot volume at its lowest since November 2023 in January 2026, and by July CoinDesk was describing a survival crisis for smaller exchanges as day traders disappeared. Thinner books make quoted gaps look larger and executable size smaller at the same time, which is why the depth figure above matters more than the spread.