Why the denominator changes the story
Seven wins in ten trades and 700 wins in 1,000 both show 70%, but they carry very different sampling uncertainty. The Wilson interval stays within 0% to 100% and behaves better near the boundaries than a simple normal approximation.
Where this calculation breaks
The interval treats trades as independent Bernoulli observations with a stable success probability. Real strategies cluster by regime, instrument and signal family; copied or overlapping trades are not independent. Selection of only published winners or tuning after inspecting results can invalidate the interpretation entirely.
Win rate is not expectancy
A strategy that wins often can lose money if losses dwarf wins. Examine average win, average loss, tail events, costs, position sizing and drawdown alongside the interval. For a provider claim, ask for the complete dated trade list and the rule used to label a win before running statistical inference.
Worked uncertainty check
The preset 70 wins in 100 trials gives a 70% observed rate, but the 95% Wilson interval is wider than a single headline number. Enter seven wins in ten instead and notice how much less precisely the same percentage is measured. This is a statement about sampling uncertainty under a binomial model, not a statement that the next 100 trades will land inside the interval. It is especially unsuitable as a trading-performance certificate when a researcher selected the sample after seeing which signals worked.
Predeclare the population
Before computing any interval, define the strategy version, date window, instrument universe, rule for overlapping positions, treatment of cancellations and exact win definition. A single market event can create many correlated trades; counting them as independent observations understates uncertainty. Keep the full trade list and compare results by regime without tuning the strategy to the same period. Even a stable hit rate says little about net returns until average payoff, tail loss and execution cost are measured.
Avoid the denominator trap
A count of winning signals can change when cancelled calls, overlapping entries or partial exits are classified differently. Publish the labelling rule before reporting the fraction and keep it fixed for the full sample. Sensitivity-test one alternative defensible rule and show how much the interval moves. The exercise is more valuable than repeating a nominal 95% interval for a dataset whose population cannot be reconstructed.