central limit theorem

For large enough samples the sample mean is approximately normally distributed, even if the data are not.

The central limit theorem (CLT) is the reason statistics works on messy product data. It says: take many random samples of the same size $n$ , compute the mean of each, and those means will form an approximately normal distribution once $n$ is large enough, whatever the shape of the original data.

Think of CatChow's 90-day spend. Individual customers are wildly uneven: many spend $20–40, a few spend hundreds. But the average of 100 random customers is a blend of cheap and expensive ones, and the extremes largely cancel out. Averages of 100 customers cluster tightly and symmetrically around the true mean, even though single customers do not.

The distribution of all those possible sample means is called the sampling distribution, and the CLT describes its shape. It also tells you how tightly the means cluster: the spread of the sample means is the standard error, $\sigma / \sqrt{n}$ . So the distribution of the sample mean is roughly

$\bar{x} \sim \mathcal{N}\!\left(\mu,\ \frac{\sigma}{\sqrt{n}}\right)$

That single fact powers confidence intervals, t-tests and the z-test for conversion rates (a conversion rate is just the mean of 0s and 1s). It is why you can run an A/B test on revenue without the revenue itself being normal.

What "large enough" means depends on the data. For roughly symmetric metrics, 30 observations are often plenty. For heavily skewed ones, where a single whale can move the average, you may need hundreds or thousands, and for small samples of such data a rank-based test or a bootstrap is safer. The CLT is about means, not individual values: it never makes the raw data normal.

Example

Suppose the 90-day spend of all new CatChow customers has mean $\mu = \$50$ and standard deviation $\sigma = \$48$ , with a long right tail: most customers spend $20–40 and a few spend over $200.

Single customers. The rule "95% within $\mu \pm 2\sigma$ " would give $50 \pm 96$ , i.e. from −$46 to $146. A negative lower bound is impossible, a clear sign this skewed data is not normal.

Averages of 100 customers. Imagine the analyst draws a random 100 customers, computes their mean, and repeats this many times. By the CLT those means are roughly normal with

$SE = \frac{48}{\sqrt{100}} = \frac{48}{10} = \$4.80$

so about 95% of sample means fall within $50 \pm 1.96 \times 4.80 = 50 \pm 9.41$ , i.e. between $40.59 and $59.41. The spread shrank by a factor of 10, and the shape became a symmetric bell.

Averages of 4 customers. Here $SE = 48 / \sqrt{4} = \$24$ , and one $240 customer in a sample of four contributes $60 to its mean all alone, more than the population mean itself. These means are still visibly right-skewed: four is not "large enough" for data this lopsided.

Common mistakes

Believing the CLT makes the data normal. It is about the distribution of sample means; individual orders stay as skewed as ever.
Treating n = 30 as a magic threshold. For heavy-tailed revenue data 30 can be far too few; check how much the mean moves when you drop the largest values.
Forgetting that samples must be random and independent. The CLT says nothing about a sample of launch-day orders or one where the same customer appears many times.
Applying it to medians or maximums as if they were means. Other statistics have their own sampling behaviour, often slower to settle.
Using it to skip looking at the data. A normal-looking sampling distribution won't reveal a tracking bug or a B2B segment hiding in the tail.

Learn it in the course

Sample vs population and the normal distribution · Sample vs population: why we study samples to learn about all users, why the bell-shaped normal distribution shows up everywhere, and how the central limit theorem ties them together.