Win Rate and Sample Uncertainty: More than a Percentage

Define a win and a trial

Specify whether a win means positive gross or net outcome and whether the unit is a fill, completed position or campaign. Count breakeven outcomes consistently. A report with 12 profitable completed positions out of 20 has an observed rate of 60%; changing the observation unit can change both numerator and denominator.

A basic binomial model treats outcomes as independent trials with a stable winning probability. That assumption can be questionable for overlapping trades, multiple followers copying one source, or a strategy that changed during the sample. Counting duplicated account outcomes as independent evidence can exaggerate precision.

Compare two samples with the same headline

Sample A has 12 wins in 20 trials. Sample B has 120 wins in 200. Both report 60%, but the larger sample provides more information under the same independent, stable-probability model. Using a 95% Wilson interval with z approximately 1.96 gives about 38.7%–78.1% for A and 53.1%–66.5% for B.

NIST gives the Wilson interval construction. These rounded illustrative intervals were calculated from the stated counts; they are not intervals for any TradeCopier customer or a prediction of the next trading period.

Read a confidence level correctly

A 95% confidence procedure is designed to cover the fixed underlying parameter in about 95% of repeated samples under its model. It does not mean that 95% of future trades will win or that the next observed winning rate is guaranteed to fall within this interval. After an interval is computed, its endpoints are fixed.

A run with zero observed losses also deserves an interval rather than a claim of certainty. A finite history cannot demonstrate that a stable event probability is exactly one, much less that the process will remain stable. Smaller samples and values near the boundaries particularly expose the weakness of casual normal-approximation shortcuts.

Check the assumptions before increasing the decimal places

Preserve dates, excluded records, strategy revisions and all closed outcomes. Serial dependence can make a nominal count overstate effective information, while selection of the best-performing account after seeing many histories changes the interpretation. Market regimes and execution costs can change the underlying process itself.

Win rate must also be paired with payoff magnitudes. A 60% rate with average wins of 10 and losses of 30 gives a gross two-outcome expectancy of 6 − 12 = −6 per trial. A precisely estimated winning frequency would not fix that unfavorable payoff structure. Use uncertainty analysis to describe evidence honestly, then review losses, costs and sample construction separately. It is not a method for selecting a recommended trade or certifying a signal provider.