A win rate is an estimate
Say you have won 60 of your last 100 games of some contest. Your observed win rate is 60 percent. That number is a fact about the past. What it says about your true underlying win rate, the long-run proportion you would reach if you kept playing under the same conditions, is less certain.
Even if the true rate were exactly 50 percent, a run of 100 games can easily give 55 or 60 wins by chance alone. A confidence interval puts a range around the observed figure to show how much room for that kind of luck there is.
The calculation
A common approximation for a win rate uses the standard error. For an observed rate p over n games it is the square root of p × (1 − p) ÷ n.
For 60 wins in 100 games:
- p = 0.6, so p × (1 − p) = 0.24.
- 0.24 ÷ 100 = 0.0024, and its square root is about 0.049.
- A 95 percent interval is roughly the observed rate plus or minus 1.96 standard errors: 1.96 × 0.049 is about 0.096.
So the interval runs from about 50.4 percent to about 69.6 percent. Rounded, “somewhere between 50 and 70 percent” is the fair summary. That is a wide range for a result that sounds clear-cut.
Now take 1,000 games with the same 60 percent rate. The standard error is the square root of 0.24 ÷ 1,000, about 0.0155. Times 1.96 gives roughly 0.030, so the interval is about 57 to 63 percent. Ten times as many games narrowed it a lot, but not tenfold.
Width shrinks slowly
The width of an interval falls with the square root of the number of games. To halve it, you need four times as many. For an observed rate of 60 percent, the approximate 95 percent margin of error looks like this:
| Games played | Margin of error (about) |
|---|---|
| 25 | 19 points |
| 100 | 10 points |
| 400 | 5 points |
| 1,600 | 2 points |
Margins are rounded from 19.2, 9.6, 4.8 and 2.4 points. The practical lesson is the one behind sample size in general: small tallies support only loose conclusions, and tightening them takes far more data than people expect.
A coin example
Suppose you flip a coin 100 times and see 55 heads. Is the coin biased? The standard error is the square root of 0.55 × 0.45 ÷ 100, about 0.0497, so the 95 percent interval is roughly 55 percent plus or minus 9.7 points, which is about 45 to 65 percent. That range comfortably includes 50 percent. The data are consistent with a fair coin, though they cannot prove it fair either. Testing dice fairness properly needs many more rolls for the same reason.
What an interval does and does not say
The wording matters. A 95 percent interval comes from a method which, used repeatedly on fresh samples from the same situation, would capture the true value about 95 times in 100. It is not a guarantee that the true value is inside your particular range, and it does not describe where individual future results will fall. A single future set of 100 games will still bounce around.
An interval also does not cover problems with the data itself. If you only counted games you remember, or if you changed your approach halfway through, the calculation is describing a sample that is not a clean set of repeated trials.
Cautions and refinements
- The simple formula is a rough tool. It behaves poorly with very small samples or with rates near 0 or 100 percent, where other methods give more reliable ranges.
- It assumes each game is an independent trial with the same chance. In skill-based games, your true rate changes with opponents, practice and conditions. A rate measured against weaker opponents does not carry over to stronger ones.
- If you look at many different win rates and report the one that looks best, the interval you quote is too optimistic. That is one of the habits behind cherry-picked statistics.
- When the true chance is known from the rules, as with a fair coin or die, a binomial distribution tells you exactly how likely each count of wins is. Intervals are mostly for cases where the chance is unknown and has to be estimated.
Using intervals sensibly
An interval is most helpful for comparisons. If one player shows 60 wins in 100 and another shows 52 in 100, the intervals overlap heavily, so the data do not clearly separate them. The same holds for claims in esports statistics, where small match counts are common and rankings can reflect noise.
None of this changes the nature of chance-based games. In casino games with a house edge, the long-run result is a loss on average, and a pleasant short-run win rate is not evidence that the game has become favourable. Outcomes stay uncertain, and anyone who finds gambling hard to control can contact a local support service.