How Many Trades Before Your Stats Mean Anything?
A seeded simulation, with the method on the page: what a zero-edge strategy measures at 20, 100 and 500 trades, and how often a real edge looks like a loser.
3 Nov 2026 · 6 min read
On this page
There is no number of trades after which a statistic is safe, but the spread of what you can measure shrinks with the square root of the sample, and we ran it to show how slowly. We simulated a strategy whose true expectancy is exactly zero, a fair coin paying +1R or −1R, and asked what expectancy a trader would measure from 20, 50, 100, 200 and 500 trades. In 26% of the 20,000 runs of 20 trades the measured expectancy came out at +0.20R or better. Nothing was wrong with the strategy; it was a sample.
How we ran it
- Model. Every trade is an independent draw: a win of a fixed size with a fixed probability, otherwise a loss of exactly 1R. No costs, no slippage, no change in the market, no change in the trader.
- Two made-up cases. A coin flip (50% wins of +1R: true expectancy 0.00R) and a real edge (40% wins of +2R: true expectancy +0.20R).
- Sample sizes. 20, 50, 100, 200, 500 trades. For each, 20,000 separate runs, each ending in one measured expectancy (the average result per trade).
- Reproducible. The generator is mulberry32 with seed 20261102, and the tables below are rendered from the script's output file, so nothing is typed. The model above is all you need: re-run it with any generator and your figures should land close to these.
A coin flip, measured
True expectancy 0.00R. The columns show the middle 90% of what the 20,000 runs measured (5th to 95th percentile), how often the measure came out below zero or at +0.20R or better, and the longest losing run in the sample. The last column is a cross-check from a formula, 1.645 × σ ÷ √n with σ = 1.00R, which should sit near the 95th percentile.
| Trades | Measured, 5th to 95th | Below zero | +0.20R or better | Longest losing run, median (95th) | Formula |
|---|---|---|---|---|---|
| 20 | −0.40 to +0.40R | 42% | 26% | 3 (7) | ±0.37R |
| 50 | −0.24 to +0.24R | 44% | 10% | 5 (8) | ±0.23R |
| 100 | −0.16 to +0.16R | 46% | 3% | 6 (9) | ±0.16R |
| 200 | −0.11 to +0.12R | 47% | 0% | 7 (10) | ±0.12R |
| 500 | −0.07 to +0.07R | 49% | 0% | 8 (12) | ±0.07R |
At 20 trades the middle 90% runs from −0.40R to +0.40R, for a strategy that is worth nothing on average. At 100 it is −0.16R to +0.16R, and at 500, −0.07R to +0.07R. Quartering the spread takes sixteen times the trades: the width falls with the square root. The share of runs at +0.20R or better drops from 26% to 3% between 20 and 100 trades, so a +0.20R measurement from 20 trades is weak evidence and from 100 is stronger, though still not proof. The losing runs are the other lesson: with nothing wrong, the median sample of 100 trades still contains a run of 6 losses in a row, and about one in twenty contains 9 or more.
A real edge, measured
True expectancy +0.20R, with σ = 1.47R. Here the question flips: how often does a strategy that really does make money measure as a loser? (The formula column is the spread either side of the true +0.20R.)
| Trades | Measured, 5th to 95th | Below zero | Longest losing run, median (95th) | Formula |
|---|---|---|---|---|
| 20 | −0.40 to +0.80R | 25% | 4 (9) | ±0.54R |
| 50 | −0.16 to +0.56R | 16% | 6 (11) | ±0.34R |
| 100 | −0.04 to +0.44R | 10% | 7 (12) | ±0.24R |
| 200 | +0.04 to +0.36R | 3% | 9 (14) | ±0.17R |
| 500 | +0.09 to +0.31R | 0% | 11 (16) | ±0.11R |
In 25% of the runs of 20 trades, a strategy with a true +0.20R edge measured below zero, and at 200 trades it still did in 3%. A trader who abandoned it after a bad first sample would have dropped something that works; one who kept a coin flip after a good first sample would have kept something that does not. The sample cannot tell those two apart at 20 trades.
What we are not claiming
- This is not a threshold. The honest answer to "how many trades" depends on how large an effect you are trying to see against how much your results scatter. Larger scatter (a bigger win-to-loss ratio, as in the second case) needs more trades for the same precision.
- Real trading is messier than the model. Results are not independent in a trending or quiet market, sizes differ, and the trader changes. Those make real samples less reliable than these, not more.
- Both cases are made up. Neither is any trader's results, and the simulation says nothing about whether a method works. This is arithmetic, not advice.
What to do with it
- State the sample beside every figure: "+0.20R over 60 trades", never "+0.20R".
- Keep a fixed rule for what you change after how many trades, decided before the sample, not after a bad week.
- Compare your measured expectancy with the spread for your sample size in the tables, not with zero alone.
- Separate samples by method: pooling two systems into one list makes the average describe neither.
The measures this applies to are expectancy, profit factor and win rate against reward-to-risk, and a longer sample also deepens the maximum drawdown you will see.
We're working on a free journal and a backtester, with no date yet. Today Trepeat is a trade copier for MT4, MT5 and cTrader, currently on demo accounts, and you can add notes, a grade and screenshots to any copied trade once it closes, with a calendar of your results.
The journal exports to Excel (.xlsx) and calculates none of these metrics today, so the arithmetic runs in a spreadsheet on that export. See the journal documentation.
Written by the Trepeat team.