Key takeaways
- 30 trades tells you almost nothing. The 90% confidence band is 30.0 points wide — a true 50% strategy routinely measures anywhere from 33% to 63%.
- 100 trades halves the error, and is still loose. Band width 16.0 points, or roughly ±8 points.
- 200 trades is the first honest number. Band width 11.5 points, and the chance of being off by more than ten points falls to 0.4%.
- 500 trades gets you to ±3.7 points. Beyond that, precision improves with the square root of the count and the returns get thin fast.
- This is the floor, not the ceiling. These samples are drawn at random from one stationary pool. A real backtest also carries regime change, curve fitting and overlapping trades, all of which make the true uncertainty larger than the table below.
First, in plain English
Your backtest finishes and reports a win rate — say 55%. The question nobody asks is: how close to the truth is that number? Not whether the strategy is good. Whether the 55% is even real.
Here is the problem in one sentence. Flip a fair coin 30 times and you will not get 15 heads. You might get 11, or 19. The coin has not changed — you just did not flip it enough. A backtest is the same thing: a short one does not measure your strategy, it measures your strategy plus a large helping of luck, and it hands you the total without telling you how much was which.
So we measured the luck. We built a pile of 92,585 real trades where we already know the honest answer — 49.93% of them won. Then we pretended to be a trader running a short backtest: reach in, pull out 30 trades at random, work out the win rate of just those 30, and write it down. Then put them back and do it again. Twenty thousand times, for every sample size from 10 trades up to 1,000.
Because we know the real answer is 49.93%, we can see exactly how far each pretend backtest landed from it. That distance is the error, and the table below is nothing more than a record of it.
How to read the columns. "90% of backtests show" is the range that nine out of ten runs fell inside. At 30 trades that range is 33% to 63% — so a strategy that truly wins 49.93% of the time will, on a 30-trade test, quite normally show you anything from a disaster to a goldmine. "Badly wrong" is the share of runs that missed the truth by more than ten percentage points. At 30 trades that is 28.4% — call it one run in four.
This page is the evidence, not the advice. If what you want is a practical answer to how many trades to collect and how to gather them quickly, that is a different question and it has its own guide: how many trades you need to backtest a strategy. This page exists to show you the measured error bars that guide is built on.
How this was measured
The obstacle to answering this question honestly is that with real strategy results you never know the true win rate, so you have nothing to measure the error against. We solved that by building a pool where the truth is known by construction.
- The pool. 92,585 trades generated by the mechanical stop-versus-target race described in our stop and target study — a 1×ATR stop against a 1×ATR target, on 29 instruments of hourly data from 2010-08-31 to 2026-09-11.
- The truth. Because we hold every outcome, the true win rate of that pool is known exactly: 49.93%. That is the number every sample is trying to recover.
- The experiment. For each sample size from 10 to 1,000, draw that many trades at random from the pool, compute the win rate, and record how far it fell from 49.93%. Repeat 20,000 times per size.
- The output. The 5th and 95th percentile of those 20,000 estimates — the range your backtest lands in 90% of the time — plus the average miss and the share of runs that were off by more than ten points.
What each sample size actually buys you
Read the third column as the width of your uncertainty. If it says 30 points, a strategy that truly wins half its trades will show you anything from a disaster to a goldmine.
| Trades in your backtest | 90% of backtests show | How wide that is | Give or take | Typical miss | Badly wrong (>10 pts off) |
|---|---|---|---|---|---|
| 10 | 20.0% – 80.0% | 60.0 pts | ±30.0 pts | 12.33 pts | 55.3% |
| 20 | 30.0% – 70.0% | 40.0 pts | ±20.0 pts | 8.89 pts | 38.6% |
| 30 | 33.3% – 63.3% | 30.0 pts | ±15.0 pts | 7.30 pts | 28.4% |
| 50 | 38.0% – 62.0% | 24.0 pts | ±12.0 pts | 5.65 pts | 15.9% |
| 100 | 42.0% – 58.0% | 16.0 pts | ±8.0 pts | 3.96 pts | 4.5% |
| 200 | 44.0% – 55.5% | 11.5 pts | ±5.8 pts | 2.79 pts | 0.4% |
| 500 | 46.2% – 53.6% | 7.4 pts | ±3.7 pts | 1.77 pts | 0.0% |
| 1,000 | 47.3% – 52.5% | 5.2 pts | ±2.6 pts | 1.26 pts | 0.0% |
The row that should stop you is 30. A strategy whose true win rate is 49.93% will, on a thirty-trade backtest, show you 63% or better once every twenty attempts — and 33% or worse just as often. Run a few variations and keep the best one, and you have not found an edge. You have found the top of that band.
Why the number everyone repeats is wrong
The "30 trades" rule is a misremembering of a real statistical convention. Thirty is the sample size at which the sampling distribution of the mean becomes approximately normal, which is a statement about the shape of a curve, not about precision. It was never a claim that thirty observations give you a reliable estimate, and as the table shows, it does not.
The precision you want scales with the square root of the sample. Going from 30 to 120 trades halves your error. Halving it again costs 480. This is why traders who demand certainty from a backtest end up either running enormous samples or, far more often, fooling themselves with small ones.
How to use this without needing a thousand trades
- Judge a range, not a number. If a hundred-trade backtest shows 55%, the honest reading is "somewhere around 42% to 58%, probably". Write the range in your journal, not the point estimate.
- Know your breakeven before you start. At 1:1 you need 50%, at 1:2 you need 33.3%. A result is only meaningful if the bottom of its confidence band clears that line.
- Drop to a lower timeframe to gather trades, carefully. More trades per month is the cheapest way to a real sample — but our spread measurements show costs eat a larger share of every move as the timeframe falls.
- Do not test twenty variations on one small sample. Every extra variation raises the chance that the best-looking result is the top of the band rather than the best strategy. That is curve fitting with extra steps.
- Count trades, not months. Six months of a setup that fires twice a week is fifty trades, which the table above puts at ±12 points. Time on the chart is not evidence.
Limitations
Stated plainly, because a study without them is marketing.
- The pool has a true win rate near 50%. Sampling error is widest near 50% and narrows toward the extremes, so a strategy that truly wins 80% needs fewer trades than this table implies.
- Samples are drawn with replacement from one stationary pool. Real strategies face changing regimes, which adds error this method cannot see.
- Win rate only. Expectancy and drawdown converge more slowly still, because they depend on the size of outcomes as well as their direction.
- Trades in the pool are mechanical, not discretionary. Human execution adds its own variance on top.
- Independence is assumed. Overlapping or correlated trades effectively shrink your sample below its nominal count.
Download this dataset
Every figure on this page as raw CSV and JSON, regenerated whenever the archive updates. Free for any use, including commercial, with attribution to FxBacktest.
- what-each-sample-size-actually-buys.csvWhat each sample size actually buys you
- backtest-sampling-error.jsonAll tables plus metadata, one file
Cite or republish this data
This research is free to quote, screenshot, or republish — in an article, a video, a newsletter, or a course — as long as you credit FxBacktest with a link back to this page. Copy the attribution below.
Frequently asked
How many trades do you need to backtest a strategy?
Around 200 before the result means much, and 500 if you want to separate a small edge from noise. Measured by bootstrap on 92,585 real trades, a 200-trade sample gives a 90% confidence band 11.5 points wide, against 30.0 points at 30 trades and 16.0 points at 100.
Is 30 trades enough to backtest?
No. At 30 trades the measured win rate is off by more than ten percentage points 28.4% of the time, and 90% of results fall somewhere between 33% and 63% for a strategy whose true win rate is 49.93%. The "30 trades" rule comes from a statement about the shape of the sampling distribution, not about precision.
How accurate is a 100-trade backtest?
Roughly ±8 percentage points, nine times out of ten. If a 100-trade test shows a 55% win rate, the true value is most likely somewhere between 42% and 58%. That is often enough to tell a good strategy from a bad one, and rarely enough to rank two similar ones.
Why does my backtest win rate keep changing as I add trades?
Because early results are dominated by sampling noise. The average miss falls from 12.3 points at 10 trades to 4.0 points at 100 and 1.8 points at 500. The number settling down is not your strategy improving - it is the estimate finally converging on what was always true.
What data was used?
A pool of 92,585 mechanical trades from 29 instruments of Dukascopy hourly data between 2010-08-31 and 2026-09-11, each one a 1xATR stop raced against a 1xATR target. Because every outcome is held, the pool's true win rate is known exactly, which is what makes the error measurable.