How many bets it takes to prove an edge (more than you have)
A 55% bettor at standard juice needs about 2,250 settled bets before the record separates from luck. At 53% it is roughly 40,000. Where those numbers come from and what to do instead.
Break-even at minus 110 is 52.38%, so a 55% bettor is trying to demonstrate a 2.6 point difference against a process with a standard deviation near 0.5 per bet. At 95% confidence with 80% power that takes about 2,250 settled bets. Halve the edge and the requirement roughly quadruples. Almost every published betting record is far too short to distinguish skill from a good run, and saying so is not pedantry, it is the reason bankrolls disappear.
The thing being tested is smaller than it sounds
A 55% win rate sounds like a wide margin over a coin flip. It is not a coin flip you are beating. At minus 110 you need 52.38% just to break even, so the entire claim rests on 2.62 percentage points. That is the signal. The noise, per bet, is a Bernoulli outcome with a standard deviation just under 0.5.
Signal over noise squared is what sets the sample size, and it is brutal to small edges. Take a one-sided test at 95% confidence with 80% power: n comes out near 2,250. Cut the win rate to 53%, an edge of 0.62 points, and the required sample rises by roughly the square of the ratio, to about 40,000 bets. A person placing twenty bets a week reaches 2,250 in a bit over two years and 40,000 in a working lifetime.
Five things that quietly make the sample useless
Variable stake sizes: a win rate computed over bets of different sizes does not describe the money, and the profit series has a different and usually larger variance than the win-rate series. Correlated bets: five positions on the same game are not five observations. Selection: a record starting from the day you began tracking is a record chosen partly because it was going well. Stopping rules: checking significance after every bet and declaring victory the first time it crosses is a procedure that reaches significance eventually regardless of skill. And post-hoc splitting: dropping the categories where you lost turns any record into an edge.
Each of these makes your effective sample smaller than your bet count, sometimes by an order of magnitude. None of them are exotic. Most published records have at least three.
- Fix your test, your confidence level, and your sample size before you look at the result.
- Count independent events, not tickets.
- Report a Wilson interval rather than a point estimate.
- Include everything from the first bet, including the categories you abandoned.
What to use while you wait
The answer is not to stop measuring, it is to measure something with less noise. Closing line value compares your price against the market close rather than against the outcome, which removes most of the variance and produces a usable signal in dozens of bets rather than thousands. It is a weaker claim, since beating a close is evidence of a good price rather than proof of profit, but a weak claim you can actually evaluate beats a strong one you cannot.
Our position: under 500 bets you have a hypothesis. Between 500 and 2,000 you have a suggestive record and should still be sizing as though you might be wrong. Above that, run the exact binomial test and read the lower bound of the interval rather than the point estimate, because the lower bound is the number your bankroll actually lives on.
Get the Pro launch email
This guide stays free. Join the list for one email when the forecast feed, machine-readable research files, alerts, and saved research workflows open.