How many trades does a backtest need to mean anything?

The one formula worth knowing

The standard error of a win rate is the square root of p(1 − p) / n, where p is the win rate and n is the number of trades. It tells you how much the observed win rate would move if you ran the same strategy on a different sample of the same size.

The useful consequence is the square-root rule: to halve your error, you need four times the trades. Precision improves, but slowly.

Numbers make it concrete. For a win rate near 50%, 30 trades give a standard error of about 9.1 percentage points, so the 95% range spans roughly ±17.9 points — a strategy showing 50% could genuinely be anywhere from about 32% to 68%. At 100 trades the standard error is 5.0 points (±9.8). At 400 trades it is 2.5 points (±4.9). At 1,600 trades it is 1.2 points (±2.5).

That is the arithmetic behind 'you need a few hundred trades'. It is not a rule of thumb someone invented; it falls out of how much precision you want.

Deciding how much precision you need

Precision is not free, and you do not need the same amount for every question. Ask what decision the number will drive.

If you are screening out obviously bad ideas, ±10 points is plenty and 100 trades will do. If you are deciding whether a strategy's edge is real enough to risk money on, you want the interval to exclude zero comfortably, which usually means several hundred trades. If you are sizing positions on the edge — using something like the Kelly fraction — you need considerably more, because position size is directly proportional to the estimated edge and an overestimate does not just mislead you, it overbets.

Write the decision down before you look at the number. Otherwise you will find that whatever precision you have is the precision you needed, which is how sample-size arguments become rationalisations.

The trap: correlated trades are not independent observations

The formula assumes each trade is an independent draw. Real strategies violate that constantly.

A strategy that holds five positions driven by the same condition does not have five independent observations; it has one observation counted five times. A strategy that trades the same instrument continuously has trades whose outcomes overlap in time and share the same market move. A strategy that only trades in one regime has a sample of that regime, however many trades it contains.

The practical implication is that the count you read from the report is an upper bound on the information you have, and often a loose one. Five hundred trades in one direction during one trend can carry less information than a hundred trades spread across different conditions.

The check is to look at the trades rather than the count: how many distinct positions were open at once, how many of the trades overlap in time, and how many different market conditions the sample covers. If the answers are 'several', 'many' and 'one', the effective sample is much smaller than it looks.

What we see in our own library

We store every candidate that survives our initial filters, and at the time of writing that is 4,015 strategies with a recorded trade count. The distribution is a useful reality check on this whole discussion.

The median trade count is 33. The lowest decile is 8 trades and the lowest quartile is 15. 45% of the library has fewer than 30 trades, 62% fewer than 50, 81% fewer than 100, and 94% fewer than 500. The maximum is 2,362.

In other words, the overwhelming majority of strategies that pass an initial performance filter do so on a sample far too small to distinguish them from luck. That is not a criticism of the filters; it is what a search process produces, and it is why we treat the trade count as a first-class field alongside the performance statistics rather than a detail.

The same pattern shows up at the extremes. Looking at our own candidates with the highest ratios, the values above 8 appear only on strategies with 4 to 44 trades — and every one of them is retired. High performance on a tiny sample is not a discovery; it is what taking a maximum over many small samples produces.

A procedure you can run today

Step one: read the trade count before anything else. If it is under about 30, stop interpreting the other statistics and treat the strategy as an untested idea.

Step two: compute the 95% range for the win rate. Square root of p(1 − p) / n, multiplied by 1.96 for a 95% interval. This takes one line and immediately shows you whether the result is capable of distinguishing anything.

Step three: check independence. Count how many positions were open simultaneously, and whether the trades cluster in one market condition. If they do, discount the sample.

Step four: check the out-of-sample behaviour. A strategy evaluated on data it was not fitted to is worth more than one with several times the trades evaluated on the data that produced it. This is the one adjustment that can partially compensate for a small sample, and it is the reason walk-forward analysis exists.

Step five: re-check the count after costs. A high-frequency strategy with 2,000 trades sounds like plenty of evidence until you notice that its edge is smaller than its spread, at which point the sample size is not the binding constraint.

What the trade count cannot fix

More trades do not guarantee independence. A very large trade count from a single regime is still a single regime.

More trades do not compensate for a bad test. If the strategy was fitted on the same data it is being measured on, 10,000 trades of that is still an in-sample result.

And more trades do not tell you what will happen next. Sample size reduces the uncertainty in your estimate of how the strategy behaved. It says nothing about whether the market that produced those trades will recur, which is a different and less tractable question.