Central Limit Theorem

The theorem operates on the average, not on the observations. Draw n independent observations from any distribution with a finite variance, compute the mean, and repeat conceptually across all possible samples of that size. The resulting distribution of those means centres on the population mean and tightens as n rises, because the variance of the mean is the population variance divided by n. Taking the square root gives the standard error: sigma divided by the square root of n. Since the divisor grows with the square root, quadrupling the sample size only halves the standard error.

What improves with n is symmetry and thinness of tails in the sampling distribution. Skewness and excess kurtosis in the parent distribution are progressively diluted as more independent draws are averaged, which is why the standardised sample mean converges to a standard normal variable even when the parent is uniform, exponential or binomial.

Almost every interval estimate and mean-based test at Level I inherits its justification from this convergence. A confidence interval for a population mean is built as the sample mean plus and minus a reliability factor multiplied by the standard error; the reliability factor is a normal or t quantile only because the sampling distribution is approximately normal. Without that approximation, the analyst would need to know the exact parent distribution.

In practice, sigma is unknown and is replaced by the sample standard deviation, which introduces additional estimation noise and calls for the t distribution rather than the normal. When sampling without replacement from a finite population, the standard error should be scaled down by a finite population correction factor. Both adjustments modify the standard error or the quantile; neither replaces the underlying convergence argument.

Two conditions carry the weight: independence and finite variance. Financial return series routinely violate the first through volatility clustering and autocorrelation, so effective sample size is far smaller than the raw count of observations. Some series are modelled with distributions whose variance is undefined, in which case no amount of averaging produces normality. Extreme skewness slows convergence dramatically, and the tails of the sampling distribution are the last region to converge, precisely the region that risk measures such as value at risk and expected shortfall depend on.

The practical consequence is that a normal approximation can be adequate for a point estimate of a mean and simultaneously misleading for a one-in-a-hundred quantile. Analysts respond with heavy-tailed distributional assumptions, bootstrapping, or stress testing rather than relying on asymptotic normality alone.

Expect to move fluently between the population parameters and the sampling distribution. Given a population mean and standard deviation and a sample size, state the mean and standard error of the sample mean, then compute the probability that a sample mean falls within a stated range by standardising and reading a z table. Reverse the calculation to solve for the sample size needed to achieve a target standard error or margin of error.

Equally important is judgement. Identify whether the theorem is being invoked legitimately: are observations independent, is variance finite, is the sample large enough given the parent's skewness, and is the quantity being described the mean rather than an individual observation. Distinguishing the distribution of the data from the distribution of the estimator is the discriminating skill in exam items on this topic.