Trading Sample Size Calculator
Sample size is the number of trades needed before a measured win rate can be distinguished from chance, and it grows with the square of how small the edge is. Detecting a five-point edge takes about 380 trades; detecting a two-point edge takes about 2,397.
How many trades before you know
Defaults test an observed 55% against a 50% baseline at 95% confidence, at 200 trades a year.
This is the standard normal approximation for a proportion and it assumes trades are independent and identically distributed. Real trades are neither, so treat the answer as a floor on how much evidence you need rather than a precise threshold.
Runs entirely in your browser. Nothing you type is sent anywhere or stored.
How the number is built
The question is not whether your results are good. It is whether they are distinguishable from the results a coin would have produced.
n = z² × p(1 − p) ÷ (p − p₀)²
where p is what you measured, p₀ is what you are testing against, and z sets the confidence.
The denominator is squared, and that is the whole behaviour of this formula. Halving the edge does not double the sample requirement; it quadruples it.
A worked example
Take the defaults: 55% observed against a 50% baseline, at 95% confidence.
The edge is five percentage points, so the denominator is 0.05² = 0.0025.
The numerator is 1.96² × 0.55 × 0.45 = 0.9508.
And 0.9508 ÷ 0.0025 = 380 trades — 1.90 years at 200 trades a year.
Now halve the edge to 2.5 points. The requirement becomes 1,521 trades — four times as many, for an edge half as large. At a two-point edge it is 2,397, and at a fifteen-point edge it falls to 39.
Why this is the least popular number in trading
Nobody has 2,000 trades of a new idea. Most people have twenty, and twenty is enough to feel certain and nowhere near enough to be.
The specific failure is optional stopping. Checking results repeatedly and deciding once they look convincing guarantees a favourable-looking sample eventually, whether or not the edge exists — because you have given yourself many chances to find a good run.
Deciding the sample size before starting is the only defence, and it is why this calculator is worth running before a test rather than after it.
What the base rates look like
On this site’s shared series, 54% of 566 ten-bar windows ended higher than they began. That is a four-point edge over a coin flip and it belongs to simply holding, not to any strategy.
Over a single bar it was 52% across 571 bars. So a short-term strategy showing 52% has
demonstrated nothing beyond the drift already present in the data — and separating it from that
baseline would need thousands of trades. The figures are in
research/series-measurements.json.
What a long sample costs you
There is a trap on the other side. A sample of 2,000 trades covers years, and markets do not hold still for years — so by the time the sample is large enough to be statistically valid, part of it describes conditions that no longer exist.
Every trade in the sample also cost money. On this site’s shared series a round trip measures about 2% of the median bar range of 0.493, so gathering 380 trades is not a free experiment — the edge has to survive the cost of measuring it.
What to do when the number is unreachable
Most honest answers to this calculator are larger than a private trader’s whole record, and the useful response is to change what you measure rather than to pretend the sample is bigger.
Measure the payoff as well as the frequency. A strategy whose winners average two and a half times its losers can be judged on far fewer trades than one hanging on a two-point edge in win rate, because the quantity being estimated is larger relative to its own noise.
Judge the process where the sample is short. Whether entries followed the rule, whether the size matched the plan, whether the exits were the ones written down — those are countable after twenty trades and they are the part you control.
And treat any single result as provisional for far longer than feels natural. The calculator’s real output is not 380; it is the knowledge that a convincing run of thirty proves almost nothing.
The original data
Of the 24,971 unique videos in research/search-study-corpus.jsonl, 1 has an instruction-shaped
title about sample size, at 35,103 views — and 100% of that coverage is calculator-shaped, the only
subject in the study where the entire coverage is a tool rather than a discussion. Win-rate targets
appear in 161 at 13,711. The counts come from site/rank_tools2.py.
One video, and the people who search for this want a number rather than an argument. That is the clearest signal in the whole tool study about what the format should be.
The answer to the question above is that seven or more from ten happens by chance 17.19% of the time on a coin flip. Roughly one strategy in six with no edge at all produces that result on its first ten trades. A run that impressive is common enough to expect — which is why the sample size, not the win rate, is the number that decides whether you have learned anything.
When it fails
The formula assumes trades are independent, and trading results cluster. Conditions persist, so wins arrive near wins and losses near losses, which makes the effective sample smaller than the count suggests. Three hundred and eighty trades taken across two market conditions may carry less information than a hundred spread across six, and the calculator has no way to see the difference — it counts trades, not circumstances.
The second failure is testing after the fact. A strategy found by searching a dataset needs a much larger sample to mean anything.
A third is treating the answer as a target to reach. It is a floor, not a finish line.
A fourth is using a win rate without the payoff. A strategy winning 40% of the time can be excellent.
A fifth is measuring on backtested fills. They are consistently kinder than real ones.
And a sixth is stopping early because the result looks good. That is the failure the whole page exists to describe.
Related
Expectancy is the quantity a sample is really estimating. Backtesting is where most samples come from and where most of them go wrong. And win rate is the input this page keeps telling you not to trust alone.
The uncomfortable implication is that most of us will never gather enough trades to prove a small edge exists, and that is not a reason to stop — it is a reason to make decisions on process rather than on a recent run of results. If a strategy needs 2,000 trades to validate, judging it after 30 is not caution or impatience. It is just not measurement.
— Michael Whitman
This page is educational, not financial advice. Test every idea on your own charts before risking money.