How many trades before you trust a win rate?
There is no universal trade-count threshold. A win rate from a small sample can be uncertain even when it looks precise; a confidence interval makes that uncertainty visible.
In a hypothetical sample with a 55% observed win rate, the 95% Wilson interval is about 45.2%–64.4% after 100 trades. It narrows to about 50.1%–59.8% after 400 trades, assuming independent, comparable binary outcomes. Neither interval says what your next trades will do or proves a strategy is profitable.
A reproducible win-rate example
Suppose a trade log contains 55 wins and 45 losses in 100 trades. The observed win rate is 55/100 = 55%. A point estimate alone hides how much it could move if another sample were drawn from the same process.
This example uses the two-sided 95% Wilson score interval for a binomial proportion. With observed proportion p̂, sample size n and z = 1.96, its center is (p̂ + z²/(2n)) ÷ (1 + z²/n); its half-width is z × √[p̂(1−p̂)/n + z²/(4n²)] ÷ (1 + z²/n). The lower and upper limits are center minus and plus the half-width.
To make the comparison reproducible, hold the observed rate at 55% in each row and choose sample sizes for which the win count is an integer. These are hypothetical samples, not RiskKit user data.
| Trades | Wins / losses | Interval |
|---|---|---|
| 20 | 11 / 9 | 34.2%–74.2% |
| 40 | 22 / 18 | 39.8%–69.3% |
| 100 | 55 / 45 | 45.2%–64.4% |
| 200 | 110 / 90 | 48.1%–61.7% |
| 400 | 220 / 180 | 50.1%–59.8% |
| 1,000 | 550 / 450 | 51.9%–58.1% |
The interval gets narrower as the hypothetical sample grows, but more observations do not fix a biased or unrepresentative test. The NIST/SEMATECH handbook describes the Wilson method and its formula for confidence intervals for proportions.
A win rate is not the same as an edge
Win rate must be read with the size of wins and losses. In R units, a simplified expectancy is:
Expectancy = (win probability × average win in R) − (loss probability × average loss in R)
At 55% wins, +1R average wins and −1R average losses, the point estimate is +0.10R per trade. But the 100-trade interval above includes win rates below 50%, and the 400-trade lower bound of 50.1% would imply only about +0.002R per trade at a 1:1 payoff—before commissions, spread and slippage.
That calculation is only a lens, not a validation rule. The Wilson interval covers uncertainty in a binary win probability; it does not estimate uncertainty in the average size of wins, average size of losses, trade costs, or the full return distribution. A strategy with a lower win rate can have positive expectancy when winners are larger, and a high win rate can still lose money when losses are much larger.
Use the Trading Expectancy Calculator with realised reward-to-risk figures, not just the win rate. Treat the result as a sample estimate and include real execution costs where possible.
What makes the interval misleading?
- Dependent trades: Wilson assumes binomial observations. Trades clustered in one market regime may not be independent, so the effective information can be smaller than the raw trade count suggests.
- Changing rules or conditions: combining different setups, instruments, timeframes or market regimes can make one pooled percentage difficult to interpret.
- Strategy selection: repeatedly testing many rules and reporting only the best result creates selection bias. A simple interval does not correct for that search.
- Costs and payoff size: the win-rate interval says nothing by itself about commissions, slippage, average R, tail losses or whether a stop fills as planned.
- Sample representativeness: a large backtest from one narrow period is not automatically more informative about a different market environment.
A useful log records the rule version, instrument, timeframe, date, net result in R and costs. Keep a later period out of the rule-building process when possible, then compare the out-of-sample results with the original assumptions.
A practical workflow before changing risk
- Define one testable setup and keep its entry, exit and risk rules consistent.
- Record every eligible trade, including losses, fees and skipped signals—not only examples that look representative.
- Report the win-rate estimate alongside sample size and an interval; do not turn a round-number trade count into a pass/fail guarantee.
- Review average win, average loss and net expectancy separately. Then stress-test different sequences and drawdowns instead of assuming the average arrives smoothly.
Use the Private Trade Journal to record outcomes and R-multiples, then use the Fixed Risk Monte Carlo Simulator or Kelly Monte Carlo Simulator to examine how your stated assumptions behave across alternative trade sequences. Those simulations rearrange assumed outcomes; they do not make the inputs more reliable.