Law of Large Numbers & Confidence Intervals
Trading Intelligence
11 min read
Build statistical confidence in your edge by understanding sample sizes, confidence intervals, and why 10 trades prove nothing.
Loading
11 min read
Build statistical confidence in your edge by understanding sample sizes, confidence intervals, and why 10 trades prove nothing.
10 trades cannot tell you whether you have an edge or are bleeding money slowly. Most retail traders quit, tweak, or scale up long before they have the data to know. This lesson shows you the math for when to actually trust your numbers.
The Law of Large Numbers (LLN) is the statistical principle that as the number of independent trials grows, the sample mean converges to the true expected value. In trading, it means your real edge only becomes visible after hundreds of trades — not after 10.
Many traders fall into this trap:
“I just had 3 winners in a row — my system’s working!” “I hit a 6-trade losing streak — it’s broken!”
But those reactions ignore something critical — and they're so universal Kahneman & Tversky named the bias "belief in the law of small numbers" (Psychological Bulletin, 1971):
A strategy’s performance only becomes meaningful after a large enough sample size.
This is the Law of Large Numbers — and it explains:
In probability:
As the number of trials increases, the average result approaches the true expected value.
In trading:
Until then — you’re mostly seeing randomness.
Concretely: a 60% edge measured over 10 trades carries a 95% CI of roughly [26%, 88%]. That same data is consistent with both a coin-flip system and a world-class one. 10 trades genuinely prove nothing.
Note the trap: LLN says long-run averages converge to the true mean. It does not say short-run streaks "correct." Believing a 6-loss streak makes a winner "due" is the gambler's fallacy — the opposite of LLN. Each trade is independent; the past does not owe you anything.
| Concept | What it actually says | What it does NOT say |
|---|---|---|
| Law of Large Numbers (LLN) | Long-run average → true mean | Streaks correct themselves |
| Central Limit Theorem (CLT) | The sample mean's distribution becomes normal at large n | Individual outcomes are normal |
| Gambler's Fallacy | (False belief) past streaks force future reversals | This is NOT a real principle |
| Sample Size | What It Tells You |
|---|---|
| 10–20 | Noise. Not statistically meaningful |
| 50 | Early directional signal |
| 100 | Minimum for confidence in win rate/EV |
| 200–300 | Reasonable confirmation of robustness |
| 500+ | Strong evidence for long-term consistency |
Sample size needed to be 95% confident win rate > 50%:
Trades needed to confirm a win rate is above 50% (95% confidence). A 52% edge requires ~100x more trades than a 65% edge.
A 52% edge looks free on paper and takes years of trades to confirm.
The more variance and skew in your system → the larger the sample size required.
The variance intuition: with n trades, the standard error of your measured win rate is √(p(1−p)/n). At n=10 and true p=0.5, that's ~0.16 — meaning a "true 50% system" will routinely show 30%–70% over 10 trades. Variance shrinks slowly: it takes 100 trades to get standard error down to ~0.05.
Imagine this strategy:
Below 50% but profitable
When right, paid 3x risk
Fixed risk per trade
Per trade, long-run
But you only log 10 trades:
You might quit at trade #6 — never reaching the edge that would appear by trade #100.
A confidence interval shows the likely range where your true performance lies, based on your sample.
Example:
That means:
You're 95% confident your true win rate is somewhere between 35–55%.
The more trades you log, the narrower this range becomes — and the more stable your metrics become.
CI width shrinks as 1/√n. Doubling your sample only narrows the interval by ~1.41×. To halve the CI you need 4× the trades. So going from 100 → 400 trades buys you a 2× sharper estimate, not 4×.
To calculate a 95% confidence interval for your win rate:
CI = p +/- z * sqrt(p(1 - p) / n)
Wald approximation -- fast, but undercovers at small n or p near 0/1. Use only when n * p * (1 - p) >= 5. Below that threshold (small samples or extreme win rates), prefer the Wilson score interval, which stays well-calibrated even at n=20. Wald is convenient; Wilson is correct.
You’ve taken 100 trades, and 45 were winners.
p = 0.45n = 100z = 1.96Plug into formula:
CI = 0.45 +/- 1.96 * sqrt((0.45 * 0.55) / 100)
CI = 0.45 +/- 1.96 * sqrt(0.2475 / 100)
CI = 0.45 +/- 1.96 * 0.0497
CI = 0.45 +/- 0.0974
95% CI = [0.3526, 0.5474] = [35.3%, 54.7%]
This means:
You’re 95% confident your true win rate is between 35.3% and 54.7%.
The more trades you add (larger n), the narrower your confidence interval becomes — and the more precise your system measurement gets.
You can also apply confidence intervals to other metrics like:
Only consistent, full tracking can:
Instead of judging trade-by-trade, look at:
A strategy that makes +0.4R over 100 trades is likely better than one that makes +3R in 10.
Many traders:
…after just 5–20 trades.
You’re optimizing for noise — not truth.
Worse: every meaningful tweak resets n back to zero. A trader who "adjusts" every 20 trades is running a different strategy each cycle and will never accumulate a sample large enough for LLN to work on any of them.
Wait for a meaningful sample. Then evaluate with:
The Law of Large Numbers (LLN) says that as the number of independent trials grows, the sample mean converges to the true expected value. In trading, your real edge — your true win rate, average R, or EV — only becomes visible after hundreds of trades, not after 10.
It depends on the size of your edge. To be 95% confident your win rate is above 50%, you need ~270 trades for a 55% system, but ~2,700 for a 52% system. Below n=100 your CI is too wide to distinguish edge from noise.
No — they are opposites. LLN is about long-run averages converging to the true mean. The gambler's fallacy is the false belief that short-run streaks must "correct" themselves on the next trade. Each trade is independent; the past does not owe you anything.
The standard error of a proportion scales as √(p(1−p)/n), so the confidence interval narrows as 1/√n. Doubling your sample only narrows the interval by ~1.41×. To halve uncertainty you need 4× the data.
At n=10, the 95% confidence interval around an observed 60% win rate is roughly [26%, 88%]. That range is consistent with both a coin-flip system and a world-class one. With so much variance, 10 trades cannot distinguish a real edge from luck.
Trading results are deceptive — unless you give the math enough time to speak.
The Law of Large Numbers reminds you:
Great traders think in series. They trade through noise, journal consistently, and only make decisions when the math is loud enough to hear.