You take your twentieth trade of the new strategy, close it out a winner, and pull up the stats: 13 wins, 7 losses, a 65% win rate. That feels like a number worth trusting — good enough to size up, good enough to tell yourself the strategy works. Statistically, it isn't. With only 20 trades behind it, that 65% is a rough guess with a wide margin of error, and the honest range of what your true win rate could be is a lot wider than the single number on the screen suggests.
A win rate is a poll, not a fact
Every win rate you compute is a sample estimate of some unknown, true long-run probability — the same way a political poll estimates support for a candidate. Nobody would trust "65% support" from a poll of 20 people the same way they'd trust "65% support" from a poll of 2,000 people, even though the headline number is identical. Small samples swing wildly from the true value just by chance; large samples settle down close to it. Your win rate works exactly the same way, and 20 trades is a very small poll.
Why the naive margin of error breaks down
Most traders who do think about this reach for the standard "margin of error" formula from a statistics class — take the win rate, multiply by its own uncertainty, and slap a plus-or-minus on it. That naive approach (a normal approximation) works fine with large samples, but it quietly falls apart in exactly the situation traders care about most: small samples and win rates near the extremes. Push it hard enough on a small sample and it can hand back nonsense — an upper bound over 100%, or a lower bound below 0%, which obviously can't be a real win rate. When the formula itself produces impossible numbers, that's a sign it isn't the right tool for the job.
The statistically sound alternative is called the Wilson score interval — a more reliable interval-estimation method built specifically for proportions estimated from small-to-moderate samples. It handles low trade counts and extreme win rates (very high or very low) without breaking down the way the naive margin-of-error calculation does, and it's the method statisticians generally recommend over the naive approach for exactly this kind of estimate. You don't need to run the formula by hand to use the idea — you just need to internalize what it's telling you: a win rate from a small sample comes with a plausible range, not a point.
The 65%-over-20-trades example, worked through
Go back to that 13-for-20 record. A proper interval around it — the kind the Wilson score method produces — is genuinely wide. The true underlying win rate could reasonably sit anywhere from roughly the low 40s% up to the mid-80s%. In other words, a strategy that actually wins 45% of the time long-run could easily have produced this exact 20-trade stretch just by chance, and so could one that actually wins 80% of the time. The 65% you're looking at isn't wrong, but it's one plausible draw from a range that spans nearly 40 percentage points. Treating it as a precise, load-bearing number — sizing up, telling other traders it works, walking away from a different setup to focus on this one — is exactly the mistake this math warns against.
Now run the same trade count up into the hundreds. As the sample grows, that plausible range narrows substantially — not because the math changes, but because a larger poll is simply a better estimate of the true population. A 65% win rate over 400 trades is a different kind of claim entirely than a 65% win rate over 20; the interval around it might only span a few percentage points, which is a number you can actually build conclusions on.
So how many trades is enough?
There's no single universal magic number — it depends on the win rate itself and how much precision you actually need before acting on it. But as a rough rule of thumb: tens of trades is not enough to draw firm conclusions, especially about a narrow slice of your data like one tagged setup. Real confidence generally starts to build somewhere in the range of 100-plus trades for an overall win rate, and you should expect to need even more than that for a specific subset — a particular strategy tag, a specific time-of-day window, a single instrument — because splitting your data into smaller buckets just recreates the small-sample problem inside each bucket.
Why raw win rate alone is the wrong headline number
This is also why win rate, by itself, is a misleading number to lead with — not because the calculation is wrong, but because a percentage with no sample size attached hides how much to trust it. The same 65% means something completely different at 20 trades versus 400, and a dashboard that shows you only the percentage, with no sense of how many trades it's standing on, invites exactly the overconfidence this post is warning about. It's the same underlying issue as why individual strategy tags can look like they don't work when they're really just thin on data — small samples, sliced too finely, producing conclusions that feel solid and aren't. And it compounds with the fact that different tools can even disagree on what your win rate is in the first place.
A win rate is a useful number, but only once you know how much weight it can bear. ExpectancyIQ tracks your trades over time rather than freezing on one flattering early stretch, so you can see when a stat is actually backed by enough history to trust it — import your trades and start building a real sample, free to start.