Learn
The Coin Flip Test: sports betting sample size and variance in one table
The Coin Flip Test answers the sports betting sample size question with one table: how wide a record with zero edge can swing over 50, 200 and 1,000 even-money bets, and roughly how many bets it takes to show a real edge. Over 50 bets a no-edge record can land anywhere from about 36 to 64 percent, so judge the price you got first. 21+. Nothing is guaranteed.
This is the table from the Coin Flip Test video. The video makes two points, the normal range and the real sample, and this page gives you both for three sample sizes, the arithmetic behind them, and what to look at instead while your sample is small.
Every number here is a hypothetical based on a fair coin at even money, not anyone's results. The point is how much luck alone can do.
The two points from the video
- The normal range. Over 50 even-money bets, a bettor with no edge lands between about 36 and 64 percent about 95 percent of the time. A 60 percent record sits inside that range. And about one run in ten, a fair coin wins 30 or more of 50.
- The real sample. Showing a real one-point edge, 51 percent instead of 50, takes roughly 7,000 bets at 95 percent confidence, and the table in the source puts it at 6,733.
The table
| Bets | No-edge range (about 95% of runs) | Chance of 60%+ with zero edge |
|---|---|---|
| 50 | 36% to 64% | about 10% (1 in 10) |
| 200 | 43% to 57% | about 0.3% |
| 1,000 | 47% to 53% | practically zero |
How to read it: if your record is inside the range for your number of bets, luck alone explains it. That does not mean you have no edge. It means the record cannot tell you yet.
Where the range comes from
A win rate over n bets has a standard error of the square root of p times one minus p, divided by n. That is the standard formula for a binomial proportion, and with p at one half the 95 percent band is about 1.96 standard errors either side of 50.
- 50 bets: the standard error is about 7.07 points, so the band is roughly 36 to 64 percent.
- 200 bets: about 3.54 points, roughly 43 to 57 percent.
- 1,000 bets: about 1.58 points, roughly 47 to 53 percent, using the same standard error formula.
Notice the shape. Four times as many bets only halves the band, which is the basic behaviour of variance when you average more independent results.
How many bets an edge needs
The source uses n of at least (Z divided by twice the edge) squared, with Z at 1.645 for one-tailed 95 percent confidence. Its table gives about 1,683 bets for a 52 percent win rate, 6,733 for 51 percent and 26,932 for 50.5 percent.
| True win rate at even money | Bets needed (about, 95% one-tailed) |
|---|---|
| 55% | 271 |
| 53% | 752 |
| 52% | 1,683 |
| 51% | 6,733 |
| 50.5% | 26,932 |
The 55 and 53 percent rows are our arithmetic on the same formula. The pattern is the one from the video: halve the edge and you need about four times the bets, because the sample grows with the inverse square of the edge.
Why a hot start feels like proof
A run of wins early on is vivid, and it arrives exactly when you have the least information. The table says a fair coin gets to 60 percent over 50 bets about one run in ten, so among ten bettors with no edge at all, one of them will probably feel like a genius after a month. Nothing about that run tells them, or you, which one is lucky. The test is a way to be that bettor and still keep your stakes where they were.
What to judge while the sample is small
The record is noisy for a long time, so look at something less noisy: the price you got. If you beat the closing price, you were paid more than the market's final estimate, and that tells you something on every bet, win or lose. SharpAPI's documentation describes it simply: positive CLV means the price beat the close. Researchers who built a strategy around closing-odds probabilities found the closing prices carried enough information to find mispriced bets.
A simple weekly check:
- Count your bets so far and find your row in the table.
- If your record is inside the range, do not change your stakes because of it.
- Look at your average closing line value instead. Positive over a few hundred bets is a better sign than any short record.
- Keep stakes small enough that a run at the bottom of the range does not hurt. A record at 36 percent over 50 bets is just as normal as one at 64.
What to watch out for
- Not even money. The table assumes even-money bets. At -110 you need about 52.4 percent just to break even, so the same record is worth less. The method still applies.
- Parlays and long shots. Bets at long odds swing much more, so the ranges get wider.
- Changing what you bet. A record that mixes sports, bet types and stake sizes is several small samples, not one big one.
- Stopping when you are up. Quitting after a hot run and keeping the number as proof is how a lucky record turns into a belief.
How to tell it worked
You will know the test is doing its job when a good week or a bad week stops changing your stake sizes, and when you can say "that is inside the range for my sample" without checking. To track the numbers that matter, see closing line value and the CLV calculator.
21+. For informational purposes only. Nothing is guaranteed. If you or someone you know has a gambling problem, call 1-800-GAMBLER.
Questions people ask
How many bets do you need to know if you're good at sports betting?
It depends on the edge. At even money, a 52 percent win rate needs about 1,683 bets and a 51 percent win rate about 6,733 to show at 95 percent confidence, per this sample-size table.
Is a 60 percent record over 50 bets good?
It is inside the normal range for a bettor with no edge, about 36 to 64 percent over 50 even-money bets, so it does not prove skill either way.
What should I track instead of my record?
The price you got against the closing price. Positive CLV means the price beat the close, and that tells you something on every bet.
Does a big enough sample mean I will profit?
No. A sample tells you how confident to be about the past. Nothing is guaranteed about the next bet.
Sources
- How long does it take for skill to show itself?, daytrading.com
- Binomial proportion confidence interval, Wikipedia
- Variance, Wikipedia
- Closing Line Value (CLV), SharpAPI docs
- Beating the bookies with their own numbers, arXiv
