Backtest Sample Size Calculator

Your backtest result is an estimate, not a measurement

A backtest returns two numbers everyone treats as facts: a win rate and an average R. Neither one is a fact. They're samples — a finite draw from a process that could have gone differently — and the smaller the sample, the further the draw can land from the truth.

This isn't a technicality. Run 20 trades on a strategy whose real win rate is 45% and you will land somewhere between about 25% and 65% a good chunk of the time. Both of those are the same strategy. One of them makes you delete the rules and start over; the other has you sizing up. The strategy never changed — only the sample did.

So the useful question is never “what did my backtest say?” It's how much does my backtest constrain the answer?That's what this tool measures.

The three numbers that matter

Expectancy.What the average trade returns in R. At 45% and 1:2 it's +0.35R. This is the number your strategy lives or dies by, and it's the thing the sample is trying to estimate.

Variation. How much an individual trade swings around that average. A 1:2 strategy has a standard deviation near 1.5R per trade — roughly four times its own edge. That ratio is why trading needs such large samples compared with most things people measure: the signal is small and the noise is loud.

Sample size. The only one of the three you control. Uncertainty shrinks with the square root of it, which is the detail that catches people out — to halve your margin of error you need four times the trades, not twice.

What “a no-edge strategy scores this well X% of the time” is telling you

This is the line worth reading twice. It answers a specific question: if your strategy had zero edge — the same win rate as a coin weighted to your break-even point — how often would a run of this many trades come out looking this good purely by chance?

When that figure is 30%, your result is what a strategy with no edge produces about a third of the time. It is not evidence. When it drops under 5%, luck becomes a strained explanation, and the conventional line for calling a result significant is exactly there.

One honest caveat the tool can't see: if you tested six variations of your rules and are reading the best one, that 5% isn't 5% any more. Testing enough variants guarantees one of them looks significant. If you did that, the fix is to fix the rules and test them again on data you haven't touched.

Why the break-even overlap matters more than the win rate

Traders anchor hard on win rate, but win rate on its own means nothing without the reward-to-risk beside it. A 40% win rate is excellent at 1:3 and terminal at 1:1. The line that separates them is the break-even win rate, which is just 1 ÷ (1 + R).

The calculator shows the range your true win rate could occupy, and it flags when that range still reaches below break-even. That overlap is the thing to act on. It means your sample has not yet ruled out the possibility that your strategy loses money — regardless of how good the headline number looked.

How many trades you actually need

There's no universal answer, and any article giving you one — 30, 100, “at least a year” — is guessing on your behalf. The required sample scales with the square of your noise-to-signal ratio, so it depends entirely on how big your edge is:

A strong edge is cheap to prove. Something like 50% at 1:2 — +0.5R a trade — clears the bar in well under 50 trades. A moderate edge of +0.2R needs a few hundred. A genuine but thin edge of +0.05R needs thousands, which is a polite way of saying you will never prove it, and you should widen the edge rather than grind out the sample. The tool draws that line for you rather than making you feel it after two years.

This is also the honest argument for backtesting over waiting. Five trades a week is 250 a year. The same 250 occurrences taken from historical price is a few focused sessions — which is the difference between knowing whether your edge is real this month and finding out in 2028.

What this model assumes — and where it breaks

Every trade is treated as independent, with a fixed win rate and a fixed reward-to-risk. Real results aren't independent: strategies work in some regimes and not others, so wins and losses cluster. When that happens the true uncertainty is widerthan what's shown here, not narrower.

It also assumes wins all pay the same multiple. Partial exits, breakeven stops and runners all change the shape of the return distribution. And it takes your inputs at face value — if the win rate you typed came from a backtest with hindsight in it, no amount of statistics repairs that. Garbage in, confidently intervalled garbage out. The seven mistakes that make backtest results worthless is the companion piece to this one for exactly that reason.

Getting a sample worth putting in the box

The number this tool wants is a measured win rate over a decent run of occurrences, taken on price you couldn't see ahead of. That means replaying history candle by candle rather than scrolling back and marking the setups that worked — the second one produces a beautiful number that means nothing.

How many backtests do you need before trusting a strategy covers the sampling side in more depth — variety of conditions, not just count — and how to backtest the CRT strategy shows the specific fields to log so the numbers you feed in here are ones you can defend.

Frequently asked questions