How big is a real arbitrage edge?
Everyone in this category quotes a number for the size of a typical arbitrage, and nobody shows their working. So here is ours, with the sample stated: every arbitrage our scanner recorded across 35+ crypto and no-KYC sportsbooks over 37 days.
1. The distribution
After removing the ones our own data-quality checks flagged (more on that below), 2,563 arbitrages remained. Their sizes:
| Percentile | Edge |
|---|---|
| 25th | 0.32% |
| 50th — median | 0.86% |
| 75th | 2.13% |
| 90th | 4.35% |
| 95th | 6.22% |
| 99th | 12.01% |
n = 2,563 · arbitrages passing data-quality checks
The median is 0.86%. Three quarters of everything found was under 2.13%. That is the honest shape of this activity: it is a low-single-digit business, and the interesting question is not how big the edges are but whether you can actually get them on.
For context on what that means in practice: on a 1,000 stake, a median arbitrage is a shade under 9 back — before anything goes wrong, and assuming every leg is accepted at the price shown. This is not a page arguing you should do this.
2. The part nobody publishes
Our scanner flags an arbitrage when it fails a data-quality check — the two books turn out to be the same platform behind different brands, or the market shape does not survive verification, or the price is an outlier against the consensus of every other book. We keep those rows rather than deleting them, which means we can ask a question the rest of the category cannot: how does the flag rate change as the claimed edge gets bigger?
share of arbitrages in each band that failed a data-quality check · n = 3,326
| Claimed edge | Found | Flagged | Flag rate |
|---|---|---|---|
| under 1% | 1,620 | 220 | 13.6% |
| 1–2% | 600 | 119 | 19.8% |
| 2–3% | 309 | 63 | 20.4% |
| 3–5% | 303 | 72 | 23.8% |
| 5–10% | 210 | 57 | 27.1% |
| 10–20% | 114 | 62 | 54.4% |
| 20%+ | 170 | 170 | 100.0% |
Every single arbitrage we recorded above 20% was bad data. All 170 of them. Not most, not nearly all — every one. Above 10% it is already worse than a coin flip.
This is the most useful thing in the dataset, and it is why we publish it. The instinct when a scanner shows a huge number is excitement. The correct instinct is suspicion. A 30% arbitrage is not a windfall; it is an alert that two prices are being compared that should not be.
There is a second reason big numbers are the wrong thing to chase even when they are real: an edge that large usually means a book has made an obvious pricing mistake, and every book's terms let it void bets taken at an obviously wrong price. The bet most likely to be cancelled is exactly the one that looked best.
3. What the flags actually were
These are not hypothetical categories. Each is a bug we found, diagnosed and fixed in our own adapters, and each produced convincing-looking arbitrages before it was caught:
- A first-quarter handicap read as a full-game money line. One book's basketball feed put a quarter handicap in the field we were reading as the match result. The prices were real; the market was not the one we thought.
- Tennis doubles matched onto the singles match. The two surnames in a singles tie are a subset of the four in a doubles pair, so fuzzy name matching links them. A coin-flip doubles line against a lopsided singles line looks like a spectacular arbitrage.
- A suspended market whose selections still looked open. The book signalled the suspension at market level while every selection under it continued to report itself as available at placeholder prices.
- Novelty markets shipped as fixtures. One feed returns "Team vs Player" and matchday-statistics specials in the same shape as real events, and they then match against real events.
- A silent sport-switch failure. A board that failed to change sport served a three-way soccer market as a two-way money line, quietly turning the draw price into the away price.
The common thread is that none of these is an arithmetic error. The arithmetic was right every time. The inputs were two different markets.
4. Not all sports are equally trustworthy
| Sport | Arbitrages | Flag rate | Median edge |
|---|---|---|---|
| Tennis | 938 | 34.5% | 1.63% |
| Soccer | 1,721 | 19.4% | 0.90% |
| Basketball | 616 | 15.9% | 0.99% |
| Esports | 51 | 13.7% | 1.00% |
sports with at least 20 recorded arbitrages
Tennis has roughly double soccer's rate of bad data, and the highest median edge — which, given everything above, should be read as a warning rather than an opportunity. Both facts have the same cause: tennis player names are short, inconsistently formatted across books, and shared between singles and doubles draws, so tennis is where fixture matching goes wrong. Its apparently fatter edges are substantially an artefact of that.
5. How long an arbitrage lasts
Size is only half the story. The other half is whether it is still there when you have finished placing the second leg. For the 2,560 arbitrages we watched from appearance to disappearance:
| Gone within | Share |
|---|---|
| 5 minutes | 38.8% |
| 15 minutes | 60.7% |
| 1 hour | 71.4% |
median lifetime 8.1 minutes · 25th percentile 3.0 minutes
The median arbitrage survives about eight minutes. Nearly two in five are gone inside five. That number, not the edge size, is what determines whether this is workable for a given person: if it takes you six minutes to log into two books and place two bets, the median opportunity is a coin flip on whether the second price is still there.
Method, and what this doesn't tell you
Every figure above comes from one query against our own database over 19 June – 26 July 2026. We have not smoothed, resampled, or excluded anything except where a table says so.
Real limitations, stated plainly:
- These are detections, not bets. Nothing here was placed. It is a measurement of what a scanner sees, not of what a bettor gets — and the gap between those is the entire difficulty of the activity.
- Our flags are not ground truth. They are our checks. Some flagged rows were probably real, and some unflagged rows are probably still bad — the 99th percentile of the "clean" set is 12%, which by this page's own argument is suspicious.
- 37 days is one sample. It covers a particular slice of the calendar and a particular set of books. Do not read it as a constant of nature.
- It is our book list, not the market. These are crypto and no-KYC sportsbooks. A set of regulated books would produce a different distribution.
If you want to check any of this, the calculators are free and the prices are live: sure bet calculator, de-vig calculator.