How many trades does a backtest need to be reliable?

Forty trades and a 60% win rate feels like proof. It isn't — and the arithmetic that shows why takes about two minutes. Here is how to work out how many trades your backtest needs before its result means anything, and why the honest answer depends on your strategy rather than on a magic number.

Why a small sample lies so convincingly

A backtest is a sample. The trades it produced are one draw from everything the strategy could have done, and a small draw can land a long way from the truth purely by luck. The question is never "was the backtest profitable?" but "is the profit larger than what luck alone could produce with this many trades?"

Take win rate, the number everyone looks at first. The uncertainty around a measured win rate shrinks with the square root of the trade count. For a strategy that truly wins half the time, the 95% range you could see by chance looks like this:

Trades in the backtest	Win rate you could see by luck alone (true rate 50%)
40	roughly 35% to 65%
100	roughly 40% to 60%
200	roughly 43% to 57%
400	roughly 45% to 55%

At forty trades, a coin-flip strategy can show a 65% win rate without anything being wrong with the test. That is why a short backtest that looks brilliant should make you curious, not confident.

The number that actually matters: edge versus noise

Win rate on its own says little, because a 35% win rate with large winners can be excellent. What you want to know is whether the average result per trade — the expectancy — is genuinely above zero. That depends on two things: how big the edge is, and how noisy each individual trade is.

Measure both in R, where 1R is the amount risked on a trade. If the average trade makes E R and the spread of trade results (the standard deviation) is S R, a rough count of trades needed for the average to stand about two standard errors clear of zero is:

Trades needed ≈ (2 × S ÷ E)²

Two worked cases, both with trade results spread at 1.2R:

• An edge of 0.3R per trade needs about (2 × 1.2 ÷ 0.3)² = 64 trades.
• An edge of 0.1R per trade needs about (2 × 1.2 ÷ 0.1)² = 576 trades.

A third of the edge needs nine times the trades. This is the single most useful idea in this article: thin edges need enormous samples, and most real retail edges are thin. A scalping system making a fraction of an R per trade cannot be judged on a few weeks of data, however clean the curve looks.

Treat the formula as a floor, not a pass mark. It assumes trades are independent of one another and that the future looks like the past, and neither is fully true.

Three things that raise the number further

1. Every parameter you tuned

If you tried twenty settings and kept the best one, the winner's result is inflated by the selection itself. The more knobs a strategy has, and the more combinations were tried, the more trades it needs before the result survives that inflation. A practical habit: be far more demanding of a strategy with six optimised inputs than of one with two.

2. Trades that cluster in time

Two hundred trades taken across six years are worth much more than two hundred taken in one volatile quarter. Trades bunched into a single regime are not independent evidence — they are mostly one observation of how the strategy behaved in that one environment. Check how the trades are spread across years and market conditions, not just how many there are.

3. Rare-signal strategies

An expiry-day strategy on a weekly index option gets roughly fifty opportunities a year. Five years of history gives you about 250 trades — and those come from only a handful of distinct volatility regimes. For this kind of system, a longer history matters more than finer detail, and conclusions should be held loosely even when the count looks respectable.

So what is a reasonable target?

For a typical retail intraday or swing system with a modest edge, we treat fewer than 100 trades as no result yet, around 200 as the point where a backtest starts to be informative, and several hundred as the comfortable range for strategies with small average wins. Then we check that the trades spread across different years and conditions, and that the edge holds when the parameters are nudged.

If your strategy simply can't produce that many trades, that is not a reason to trust a small sample. It is a reason to test on more instruments with similar behaviour, use a longer history, or accept a wider range of possible outcomes and size positions accordingly.

A quick check you can run on your own backtest

1. Count the trades. Under 100, stop — you have an anecdote.
2. Work out E and S in R. Most backtest exports give you the trade list; a spreadsheet does the rest.
3. Plug them into (2 × S ÷ E)². If your trade count is below the answer, the profit could be luck.
4. Split by year. If one year produced most of the profit, the effective sample is far smaller than the count.
5. Reshuffle the trades. A Monte Carlo simulation shows how bad the drawdown could have been with the same trades in a different order.

Sample size is one of the six usual reasons a backtest doesn't match live results. It is also the easiest one to check before risking money.