How Many Trades Before a Result Means Anything? Sample Size and Standard Error

After 40 or 50 trades, many traders decide a system “works” or “doesn’t work”. But numbers from a small number of trades swing much wider than it feels. The exact same system, in the hands of two people, can leave one in profit and the other in a loss. This article works out how many trades it takes before the numbers start to say something, and why the answer depends on the shape of the system.

This is different from data snooping, which is about trying many strategies and picking the best. Here there is one strategy, but too few trades.

Why small samples fool you

Every trade’s result mixes two things: the system’s real edge, and the luck of that trade. In a single trade, luck drowns out everything. Over many trades, luck partly cancels itself out, but it cancels slowly. The swing of an average shrinks with √n, not with n.

The measure of that swing is the SE (standard error). The range where the true value probably sits, about 95% of the time, is the measured value ± 1.96 × SE.

SE shrinks with √n. To cut the error in half, you need 4 times as many trades, not 2 times.

The error in a win rate

If the true win rate is p and you count n trades:

SE of the win rate = √(p × (1 − p) ÷ n)

Example, p = 55% over 20 trades:

1. p × (1 − p) = 0.55 × 0.45 = 0.2475
2. ÷ 20 = 0.012375, square root 0.111, or 11.1 points
3. 95% range = 55% ± 1.96 × 11.1 = 33.2% to 76.8%

Trades	SE	95% range when you measure 55%	Straddles the 45.5% break-even?
20	11.1 points	33.2–76.8%	Yes
50	7.0 points	41.2–68.8%	Yes
100	5.0 points	45.2–64.8%	Yes (just)
300	2.9 points	49.4–60.6%	No, clear
1,000	1.6 points	51.9–58.1%	No, clear

The 45.5% break-even comes from RR 1.2: 1 ÷ (1 + 1.2) = 1 ÷ 2.2 = 45.45%. The reasoning is in Win rate vs break-even. These ranges use the normal approximation, which is rough when trades are few or the win rate is near 0% or 100%, but it is good enough to see the picture.

A confidence interval for expectancy in R

The win rate isn’t the final answer. What you need is expectancy, the average profit per trade, measured in R (1R = the money risked per trade, the distance to the Stop loss). Measuring in R keeps results comparable even when position size changes.

SE of expectancy = SD ÷ √n

SD (standard deviation) is how far each trade’s result swings around the average. You can compute it from your trade list with the STDEV function in Excel or Google Sheets. If the system has only two outcomes (win RR or lose 1R), the shortcut is SD = (RR + 1) × √(p × (1 − p)).

Worked example: how many trades for 55% at 1.2R

1. Expectancy = 0.55 × 1.2 − 0.45 × 1 = 0.66 − 0.45 = 0.21R per trade
2. SD = (1.2 + 1) × √0.2475 = 2.2 × 0.497 = 1.09R
3. For the lower 95% bound to clear zero you need 0.21 > 1.96 × 1.09 ÷ √n
4. √n > 1.96 × 1.09 ÷ 0.21 = 10.2, squared gives n > 104.4, so about 105 trades

Trades	SE	95% range of expectancy
50	0.155R	−0.09R to +0.51R
100	0.109R	−0.004R to +0.42R
200	0.077R	+0.06R to +0.36R
300	0.063R	+0.09R to +0.33R
1,000	0.035R	+0.14R to +0.28R

There is a hidden trap. 105 trades is the case where your measured result comes out at exactly 0.21R. But the measured result swings too. We simulated this system 20,000 times, 105 trades each, and only 51.9% of runs had a lower bound above zero. That’s a coin flip. To see it pass 4 times out of 5, you need about (1.96 + 0.84)² × (SD ÷ expectancy)² ≈ 213.2, rounded up to 214 trades (simulated at 214 trades: 80.6% pass). An easy rule: roughly twice the minimum.

Costs make it worse. Subtract a cost of 0.05R per trade and expectancy drops to 0.16R. The minimum rises to about 180 trades, and about 368 to pass 4 times out of 5. To get this range from your real trade list with a bootstrap, see Monte Carlo strategy testing.

Low win rate, big payoff: you need many more trades

The number of trades you need depends on the ratio SD ÷ expectancy, squared. A system that wins rarely but big swings hard on every trade, so you have to wait longer. Compare systems that all have the same expectancy of 0.2R.

A system winning 15% at RR 7 needs 784 trades (and about 1,600 to pass 4 times out of 5). At 5 trades a week, that is about 3 years before you know. On top of that, with systems like this long losing streaks are normal, so people tend to quit before the statistics have had time to say anything.

Watch the other side too. A high-win-rate system looks quick to prove, but the formula only knows the swings that have already happened. If the system can occasionally take a much larger loss (for example a Stop loss jumped by a price gap), and that hasn’t happened in your sample yet, the measured SD is lower than the truth and the trade count you compute is too small. Read more in Fat tails and black swans.

Simulation: 20 traders, one system, 50 trades each

We simulated in node 20 traders running the same system, 55% at 1.2R, identical in every way, 50 trades each (random draws with a fixed seed, so it can be repeated). The expected total is 50 × 0.21 = +10.5R, and the swing of a 50-trade total is 1.09 × √50 ≈ 7.7R.

• Totals range from −6R to +20.4R, and 3 of the 20 traders lose
• The win rate each trader measures ranges from 40% to 64%, while the true value is 55%
• The trader on +20.4R will believe they are skilled, and the one on −6R will drop the system. Both ran the same system. Only luck differs

Scale it up to 100,000 traders and 7.8% of them finish their first 50 trades in the red with this system. Meanwhile a system with no edge at all (50% at RR 1:1) finishes 50 trades with at least +10.5R 6.0% of the time, matching the good system’s expected result with zero edge. This is why you judge signal sellers and copy-trading providers by trade count and track length, not by a short burst of profit.

A great month is mostly noise

Say you take 20 trades a month with the same system. The expected month is 20 × 0.21 = 4.2R, but the monthly swing is 1.09 × √20 ≈ 4.9R, bigger than the expectation itself. Simulating 100,000 months:

• 25.3% of months make 8.4R or more (twice the expectation)
• 25.0% of months are negative

A genuinely good system still has about 1 losing month in 4, and about 1 double month in 4. A single month says almost nothing about whether the system got better or worse.

Regression to the mean: standout results drift back

Regression to the mean means that an unusually strong result in one period tends to be followed by something closer to average in the next, because the part that made it stand out was luck, and luck doesn’t carry over.

From the 100,000 simulated traders, we took the 10% with the best first 50 trades.

• First 50 trades: this group averaged +23.9R
• Next 50 trades, same system, same people: +10.5R, exactly the expectation
• The worst 10% did the same in reverse: from −3.0R back to +10.4R

You see this all the time in real trading. People raise their size after their best month and results go back to normal. Or they switch systems after their worst month, when the system may have done nothing wrong.

Peeking often also fools you

Another trap few people mention: checking the confidence interval every few trades and stopping the moment it “passes”. We simulated a system with no edge at all (45.45% at 1.2R) 20,000 times.

• Look once, at 300 trades: the lower 95% bound clears zero in 2.5% of runs, as theory says
• Look every 10 trades from trade 20 to 300, and count it as a pass if it ever cleared zero: 13.3%, more than 5 times as many

The fix: decide the trade count you will judge at before you start, then wait for it.

Minimum-sample rules of thumb, and their limits

Rough rules that work:

1. Under 30 trades says almost nothing, except catching a badly broken system
2. Compute the count for your own system: n ≈ (1.96 × SD ÷ expectancy)², then double it if you want a good system to reliably pass
3. Look at t = expectancy ÷ SE. Above 2 starts to separate from luck. For 55% at 1.2R over 100 trades, t = 0.21 ÷ 0.109 ≈ 1.9, not there yet. If you tried several versions before picking this one, set a higher bar, see data snooping
4. Count in R, not money or %, because changing position size adds its own swing to money figures
5. Judge at the count you set, not when results look good or bad

Limits that make this formula too optimistic:

• Trades aren’t independent. If you open 3 trades from the same signal, or several pairs that all move with the dollar, they tend to go the same way. 300 trades clustered like this carry roughly the information of 100 events, and the real SE is about √3 ≈ 1.7 times wider than computed. See Correlation and portfolio risk
• Market regimes change. The formula assumes every trade comes from one unchanging system. 300 trades collected during a strong trend may only tell you the system works in strong trends. Split your data into a first and second half and check that both point the same way. Read more in When the market regime changes and Walk-forward testing
• Changing the rules midway. Adjust the Stop loss or the entry conditions and you start counting again. The old trades belong to another system
• Demo trades and live trades differ. Fills and emotions are not the same, see From demo to a live account

None of this makes a system profitable. It only tells you how far to trust the numbers you see. While the sample is still too small, what you can control is risk per trade: keep it small enough to survive until the statistics can answer. See Risk 1% per trade.

Common mistakes

• Judging a system on 20–50 trades. The 95% range of the win rate is still about 28–44 points wide
• Looking only at the win rate. You need expectancy and SD too. A win rate above break-even is not a profit if a few losses are unusually large
• Using one minimum count for every system. At the same expectancy, a 15% system needs about 8 times as many trades as a 60% system
• Raising size after your best month. That month was mostly luck and is usually followed by an ordinary one
• Checking every day and stopping once it passes. Frequent peeking multiplies the chance of a false pass
• Counting simultaneous trades as separate trades. The real information is less than the count