Real business data is rarely bell-shaped. Order values, salaries, call durations, claim amounts and time-on-site all bunch up at the low end with a long tail stretching to the right. Yet confidence intervals, t-tests and A/B test calculators lean on the normal distribution. The bridge between those two facts is the Central Limit Theorem, and it is worth understanding properly because it explains both why those tools work and when they stop working.
The statement in plain words
Take any population with a finite mean and finite variance, whatever its shape. Draw a random sample of size n and compute its average. Do that again and again. The averages you collect form their own distribution, called the sampling distribution of the mean. The Central Limit Theorem says that as n grows, this distribution gets closer and closer to a normal distribution, centred on the population mean, with a spread equal to the population standard deviation divided by the square root of n.
Three separate claims are packed in there. The averages centre on the true mean. Their spread shrinks with the square root of the sample size. And their shape becomes normal even if the original data is not. The first two hold for any sample size; the third is the “limit” part and needs n to be large enough.
Watching it happen
A simulation makes this concrete. Imagine an online store whose order values follow a heavily skewed pattern: a mean of about ₹915, a median of only about ₹665, and a standard deviation of about ₹867. Most orders are modest; a few are very large. A histogram of individual orders looks nothing like a bell. Its skewness, a measure of lopsidedness where zero means symmetric, is around 3.7.
Now draw 20,000 random samples of different sizes from this population and record each sample’s average.
| Sample size | Spread of sample averages | Theory: 867 divided by root n | Skewness of the averages |
|---|---|---|---|
| 5 | about ₹396 | ₹388 | about 1.8 |
| 30 | about ₹158 | ₹158 | about 0.7 |
| 100 | about ₹86 | ₹87 | about 0.4 |
Every set of averages centres close to ₹915. The spread matches the square-root formula almost exactly, even at a sample size of 5. The shape steadily loses its lopsidedness, but notice that at 30 it is still noticeably skewed. That matters, and we will come back to it.
Why analysts depend on it
Most of the questions analysts answer are about averages or proportions (a proportion is just the average of zeros and ones). What was the average order value this quarter? Did the conversion rate change? The CLT says that the uncertainty in such an estimate follows a predictable, roughly normal shape with a known width. That is what allows a statement like “₹915 plus or minus ₹31 with 95% confidence” from a sample of 3,000 orders, without needing to know the exact shape of the order distribution.
The square-root law in the spread has a very practical consequence. Precision improves slowly. Going from 100 to 400 observations halves the uncertainty; halving it again needs 1,600. When a stakeholder asks for a more precise number, this tells you how much more data it will cost.
It also explains why A/B tests for conversion work with normal-based formulas. Each visitor either converts or does not, a distribution that is about as un-bell-like as possible. But the conversion rate over thousands of visitors is an average, and its sampling distribution is very close to normal.
The “n = 30” myth and other limits
Textbooks often say 30 observations is “large enough”. That is a rule of thumb for mildly skewed data and nothing more. The simulation above shows averages of 30 still carry a skewness of about 0.7 when the underlying data is strongly skewed. For revenue per user in a freemium app, where most users pay nothing and a handful pay a lot, you may need thousands of observations before the normal approximation is trustworthy. How large is “large enough” depends on how skewed and heavy-tailed the data is.
Other conditions also matter.
- Independence. The theorem assumes observations do not influence each other. Daily sales figures that move together because of a festival, or multiple sessions from the same user counted separately, break this. The real uncertainty is larger than the formula says.
- Finite variance. Some data is so heavy-tailed that its variance is effectively unbounded; certain insurance losses, viral content views and wealth distributions come close. Averages of such data converge very slowly or erratically, and medians or trimmed means are often more stable.
- It is about averages, not individuals. The CLT does not make your data normal. A histogram of individual orders stays skewed no matter how many you collect. Using normal-based rules, such as “99.7% of orders fall within three standard deviations”, for the raw data is a misuse.
- Not every statistic. Medians, percentiles and maximums have their own sampling behaviour. The 95th percentile of response time, for example, needs larger samples to estimate stably, and bootstrapping is usually a better way to put an interval on it.
A quick self-check
If you are unsure whether the CLT applies to your data, you can test it the way the table above was built. Resample your own data with replacement, compute the statistic thousands of times, and plot the results. If the histogram of those averages looks symmetric and bell-shaped, normal-based intervals will serve you well. If it is still lopsided, use the bootstrap distribution itself to form the interval, or collect more data. A few lines of Python with NumPy are enough, and the picture is more convincing to a sceptical stakeholder than any theorem.
Written by the RP Strativa Team. The store and the simulated figures are illustrative.
Leave a Reply