How many trades before you know a strategy works

You are forty trades into a new strategy and you are up. Or down. Either way the question is the same: is this the edge showing itself, or is it noise that reverses next month?

Most traders answer by feel, and the feel is wrong in both directions — they abandon something that worked after a bad twenty, and they scale up something that never worked after a good twenty. The question has an arithmetic answer. It is just not the one that gets repeated.

What "it works" means in numbers#

Judging a strategy is not asking whether you are up. It is asking whether your average result per trade is far enough above zero that randomness cannot comfortably explain it.

That means every trade has to be expressed in the same unit: R-multiples, where each outcome is a multiple of the risk you took on entry. A +£420 win and a +£90 win tell you nothing comparable if the first risked £200 and the second risked £30. In R, they are +2.1R and +3.0R, and now they belong to the same sample.

This is also why inconsistent sizing quietly destroys your ability to evaluate anything. If your risk per trade wanders between 0.4% and 3% depending on conviction, you do not have one sample of forty trades — you have forty differently-weighted observations that cannot be averaged. The position sizing tool gives you the size from balance, risk percent, and stop distance, which is the precondition for everything below.

Why "100 trades" has no arithmetic behind it#

The 100-trade rule is repeated as though it were a constant. It cannot be one.

The number of trades you need depends on two things: how large your average edge is, and how widely your individual results scatter around that average. A strategy with a big average edge and tight, repeatable outcomes reveals itself quickly. A thin edge with wild dispersion can hide inside noise for years.

One hundred happens to sit near the middle of the plausible range. For some strategies the honest number is seventy. For others it is over seven hundred. Using the midpoint as if it were the answer means you will call some strategies too late and, far more damagingly, call others far too early.

The two numbers that set your sample size#

You need your mean R — your average result per trade in R — and the standard deviation of R across your trades, which is just how much individual outcomes scatter.

The average of a sample gets more reliable as the sample grows, and it does so at a specific rate. The standard error of your mean is:

standard error = standard deviation of R / square root of n

Before treating a measured average as more than luck, you want it to sit roughly two standard errors above zero. Set mean R = 2 x (standard deviation / square root of n), rearrange for n, and you get the whole answer in one line:

trades needed ≈ ( 2 x standard deviation of R / mean R )²

Notice the square. Halve your edge and you do not need twice the sample — you need four times it.

Working it out on a real setup#

Take a swing strategy: 45% win rate, winners taken at +2R, losers at a full −1R stop.

Mean R:

(0.45 x 2R) + (0.55 x −1R) = 0.90R − 0.55R = 0.35R

For the dispersion, take the average of the squared outcomes and subtract the square of the mean:

average of squares = (0.45 x 4) + (0.55 x 1) = 1.80 + 0.55 = 2.35
variance           = 2.35 − (0.35 x 0.35) = 2.35 − 0.1225 = 2.2275
standard deviation = √2.2275 = 1.49R

Now the sample size:

( 2 x 1.49 / 0.35 )² = ( 8.53 )² ≈ 73 trades

Seventy-three. For that particular strategy, the folklore number is conservative — you would have had a defensible read before you got to 100.

Two caveats on the arithmetic. It assumes binary outcomes — full stop or full target — which is a modelling convenience, not your trading. And it assumes a fixed 1R risk. Once you have real history, do not model the dispersion at all: take the standard deviation of your own recorded R-multiples, which already contains your partial exits, your scratches, and your slippage.

The same math on three setups#

All three are profitable systems. All three would look identical in a trader's head after thirty trades. They need sample sizes that differ by a factor of ten.

The third row is the one worth sitting with. A 70% win rate feels like proof — you are right most days, and being right most days is emotionally indistinguishable from having an edge. But the losers are twice the size of the winners, so the edge per trade is tiny and the sample required to confirm it is enormous. High win rate is not evidence. It is comfort.

For my SMC strategy the mean R is +0.18R and the standard deviation is 2.15R, which implies a sample of about 565 trades. I only have 312 at the moment I am writing this.

Despite a relatively high number of journaled trades, I am still at roughly half of what I need to be confident in the edge. That surprised me.

The strategy has a ~25% win rate, but average profit-taking is between 3R and 5R. That is why the required sample is larger than the swing example above.

When the number is 500 and you take four trades a week#

Five hundred trades at four a week is roughly two and a half years. Statistical certainty is not available to you on that horizon, and pretending otherwise is how people end up p-hacking their own trading. Four things are available instead.

Set the kill line before you need it. Decide in advance what result would make you stop — a cumulative R below some level after some number of trades — and write it down while you are calm. A pre-committed exit from a strategy is the only defence against the two failure modes, quitting a good system in a drawdown and nursing a dead one out of sunk cost.

Simulate what you cannot yet observe. If you believe a win rate and a reward-to-risk ratio, you can generate the distribution of outcomes they imply rather than waiting to live through one draw from it. Feed your numbers into the Monte Carlo simulator and look at the spread of possible 50-trade results. Seeing that a genuine edge routinely produces losing 50-trade stretches is what stops a normal drawdown from reading as a verdict — the same point covered from the other side in how long a losing streak is normal.

Measure process, not just P&L. Whether you followed your entry criteria, sized correctly, and respected your stop are all things you can score on every single trade. Compliance data reaches a usable sample in weeks rather than years, and it is the input you actually control.

Resist the urge to slice. Splitting your trades by session, pair, and setup grade feels like extracting more information. It divides your sample instead, and every extra slice raises the odds that one bucket looks brilliant purely by chance.

I had a Fibonacci-based strategy I called FIB. I decided it was not worth my time after only 98 trades — enough to feel like a verdict at the time, but I stopped on feel in the moment, with no written kill line. When I later loaded the trades, the R-multiples looked profitable; the account was losing because risk in dollars was inconsistent.

What this calculation does not tell you#

Intellectual honesty about the limits matters more than the formula, because a misapplied threshold is worse than no threshold.

• It assumes your trades are independent. Three correlated FX pairs bought in the same direction on the same catalyst is closer to one observation than three, so your effective sample is smaller than your trade count. The forex use case page covers how this shows up across pairs.
• It assumes the edge is stationary. Clearing the threshold says the past sample was unlikely to be luck. It says nothing about whether the conditions that produced it still exist.
• Costs must already be inside your R-multiples. Spread, commission, and swap belong in each trade's result, not in a footnote. An edge of 0.12R is entirely capable of being a cost line.
• Statistical significance is not economic significance. A real edge can still be too small to be worth the screen time, the risk, or the drawdown it comes with.
• Testing five variations and keeping the winner voids the math. You gave yourself five chances at a fluke, so the winner needs a much higher bar than the formula above sets for a single pre-specified strategy.

My worst correlated day was five losing trades on the same day. I now try to stay off highly correlated pairs in parallel — I do not trade EUR/USD and GBP/USD together. If both look good, I pick one. If I am already in one, I stay out of the other and look elsewhere.

What to do with this#

Pull your closed trades for one strategy. Convert them to R. Calculate the mean and the standard deviation, put them through the formula, and compare the result to how many trades you have actually taken.

That comparison puts you in exactly one of three states, and each has an obvious next move:

Most traders never do this because their trade history is scattered across a spreadsheet, a broker statement, and memory. The fix is unglamorous: tag every trade with its strategy as you log it, so the R-multiples for a single system can be pulled as one clean sample instead of reconstructed from three places two years later.

You will not get certainty. You will get a number that tells you whether your opinion about your own strategy is currently worth anything — and how many more trades it takes before it is.