A strategy report gives a win rate as if it were a property of the system. It is a measurement, and a measurement taken on a handful of trades comes with an error bar the report almost never prints. This note puts that error bar back.
Take a report showing 22 wins in 40 trades, a 55% win rate. A Wilson 95% interval on that count runs from 39.8% to 69.3%. That is a range of almost thirty points. Everything from a system that loses more often than it wins to one that wins seven times in ten is consistent with what the report shows.
The 55% is the middle of that range, not a fact about the system.
At a 55% win rate, the interval narrows slowly as trades are added:
Each halving of the error bar costs roughly four times the trades, because the width falls with the square root of the count. A cleaner chart does not change that arithmetic, and neither does a longer test window that adds bars but not trades. Only more independent trades do.
Precision is only useful against something. The win rate a system needs depends on its payoff: with the average win equal to the average loss, breakeven is 50%; with the average win twice the average loss, it is 33.3%.
That changes the question completely. At a 2:1 payoff, the forty-trade report above already clears breakeven: its whole interval, 39.8% to 69.3%, sits above 33.3%. At a 1:1 payoff, the same 55% needs about 385 trades before the lower end of its interval climbs above 50%. Same win rate, roughly ten times the evidence.
So the trade count a report needs is not a fixed number. It is set by how close the claimed win rate sits to the breakeven its own payoff implies. The trades-needed calculator does this search for any win rate and payoff.
The count that matters is closed trades, and only independent ones. Three years of daily data on a stock is about 756 bars, but a system that holds for a few weeks at a time might produce forty trades from it. Overlapping positions in correlated instruments behave more like one trade than several, so a system that buys five related stocks on the same morning has made one decision, not five. A report that says “tested on ten years of data” has told you the window, not the sample.
Filters improve a win rate by removing trades, and removing trades widens the interval. At 55%, 200 trades give an interval about 7 points either side; cut that to 60 with a filter and it is about 12 points either side. A filter that lifts the win rate by five points while nearly doubling the error bar has not shown anything yet.
Ask for wins and total closed trades, not the percentage alone, and compute the interval yourself (the win-rate interval calculator does it in one step). Ask for the payoff, so you know which breakeven the interval has to clear. If the report splits results by year or by market, check each piece’s count: a split that leaves fifteen trades per cell is reporting noise with a label on it. And if the report gives only a percentage and a date range, the count is the first thing to ask for, because nothing else in it can be judged without one.
Count the trades, put the interval around the win rate, then compare its lower end with breakeven. Done the other way round, starting from the headline percentage, the forty-trade report looks like a 55% system. It is a 40-to-69% system that has not yet shown which end it belongs to.
Every figure above is a Wilson 95% score interval or a breakeven, computed for this note at a 55% win rate. None of it is a result of mine.
These are the same checks I run before I believe any backtest, including my own.
Historical simulations are not live results. Nothing on this site is investment advice.