Why a 70% Win Rate Over 20 Trades Means Nothing

A vendor shows you a model with a 70% win rate. Impressive — until you ask the only question that matters: over how many trades?

Over 20 trades, 70% means 14 wins and 6 losses. Over 2,000 trades, 70% means 1,400 wins and 600 losses. Same headline, completely different evidence. The difference is the Wilson interval, and it is the single most useful statistical idea a non-quant trader can learn.

What the Wilson interval does

Every win rate is an estimate from a sample. The Wilson interval answers: given this sample, what range of true win rates is still plausible? Small samples give wide ranges. Large samples give narrow ones. The interval is honest in a way the headline never is.

Worked roughly, at 95% confidence

70% over 20 trades: the true win rate could plausibly be anywhere from about 48% to 86%. A coin flip (50%) sits comfortably inside that range. You have not demonstrated an edge; you have demonstrated 14 wins.

70% over 100 trades: roughly 60% to 78%. Better — 50% is now outside the range — but still a wide band.

70% over 600 trades: roughly 66% to 74%. Now the number starts to mean something.

70% over 40 trades, by contrast, still spans roughly 55% to 82% — better than 20, still loose.

The lesson is not that small samples are useless. It is that small samples cannot carry big claims. A 70% on 40 predictions and a 70% on 4,000 predictions are not the same claim, and any output that lets them look like one is lying by formatting.

How to use this tomorrow

Whenever you see a win rate, ask for n (the number of trades or predictions) in the same breath. A win rate without n is not a statistic; it is a slogan.

If n is under 100, treat the result as a hypothesis to test, not evidence to trade on.

If someone quotes “90% accuracy,” ask what the baselines scored. On a dataset where up-days are 55% of the sample, a model that always predicts “up” scores 55% without knowing anything. Accuracy only means something relative to the dumbest rule that beats it.

When you run your own tests, report the interval, not just the point estimate. It is the difference between marketing and measurement.

BuildIT Index computes Wilson intervals on every accuracy figure for exactly this reason: so a small sample cannot dress up as certainty. In its verified demo, 416 correct out of 600 recomputes to 69.33% — and the interval around that number, not just the number, is part of the verdict.

The next time a 70% win rate flashes at you, smile, and ask: “Over how many?” The silence that follows is also data.

A win rate without a sample size is a slogan. A win rate with a Wilson interval is a measurement.