Statistical Data Analysis and Inference › Statistical Inference: From Samples to Populations

Descriptive statistics summarize historical data, but we care about the future. Statistical inference bridges this gap by using sample data to make probabilistic statements about population parameters. This transition from description to inference is perhaps the most important conceptual leap in quantitative analysis.

The distinction between sample and population is fundamental to understanding everything that follows in this section. When we compute a mean return from historical data, we're not calculating the "true" expected return; we're estimating it from limited information.

We observe sample statistics (denoted with bars or hats: $\bar{r}$, $\hat{\sigma}$) and use them to estimate population parameters (denoted with Greek letters: $\mu$, $\sigma$). This notational convention is a constant reminder that what we calculate from data (sample statistics) and what we're trying to learn about (population parameters) are different quantities.

The key challenge: sample statistics vary from sample to sample. Under a common population model, different samples can produce different mean estimates because of sampling error. Means from 2010-2020 and 2015-2025 may also differ because the population mean or regime changed; they need not differ. With independent underlying observations of a common finite, positive variance, their shared observations give the two mean estimates positive covariance. Sampling variation is an inherent feature of estimation, not a calculation mistake. Distinguishing it from changes in the generating process matters when interpreting historical windows.

The standard error quantifies how much a sample statistic varies across different samples. It answers: if I collected a new sample of the same size, how different might my estimate be? For independent observations with a common finite, positive variance $\sigma^2$, the sample-mean variance is $\sigma^2/n$, so:

$$SE(\bar{r}) = \frac{\sigma}{\sqrt{n}}$$

where:

• $SE(\bar{r})$: standard error of the sample mean, measuring how much $\bar{r}$ would vary across repeated samples
• $\sigma$: population standard deviation (true variability of individual returns)
• $n$: sample size (number of observations)
• The $\sqrt{n}$ in the denominator reflects that averaging reduces variability, but with diminishing returns. At fixed positive population volatility, halving the standard error requires four times as much data

This formula captures one of the most important relationships in statistics under that sampling model. In general,

$$\operatorname{Var}(\bar r)=\frac{1}{n^2}\sum_{i=1}^{n}\sum_{j=1}^{n}\operatorname{Cov}(r_i,r_j),$$

Here $\operatorname{Var}$ measures variability and $\operatorname{Cov}$ measures how two observations vary together. The double sum includes $n$ individual variances and all ordered off-diagonal covariance terms. Independence makes the off-diagonal terms zero; a common variance then leaves $n\sigma^2/n^2=\sigma^2/n$. Dependence can increase or decrease mean uncertainty, depending on the covariance terms.

When the independence and positive common-variance conditions apply, more variable data produces less precise estimates and larger samples produce more precise estimates. The square root relationship means precision improves slowly with additional data.

In practice, we estimate $\sigma$ with the sample standard deviation $s$:

$$\widehat{SE}(\bar{r}) = \frac{s}{\sqrt{n}}$$

For i.i.d. normal observations, replacing $\sigma$ with $s$ leads to exact Student-t inference for the mean. For nonnormal i.i.d. observations, a t-based procedure may be a useful large-sample approximation under suitable regularity conditions. Large $n$ alone does not repair an invalid independence assumption or guarantee a good approximation for very heavy-tailed data.

Under the independent common-variance model with fixed positive population volatility, the standard error decreases with the square root of the sample size. To halve that standard error, you need four times as much data. In finance, this slow improvement can leave mean return estimates imprecise even with long samples. In this seeded example, the printed ratio shows that the standard error is large relative to the sample mean, so the mean estimate is noisy.

The Central Limit Theorem (CLT) is one of the most important results in statistics, and its implications extend beyond textbooks. It provides the theoretical justification for using normal distribution-based methods even when the data themselves are not normally distributed.

This theorem is powerful. Under its i.i.d. and finite-variance conditions, it tells us that the standardized sample mean converges to a normal distribution regardless of the population's shape. A distribution can be skewed, bimodal, or heavy-tailed and still satisfy the theorem, but infinite variance or material dependence requires different conditions and may produce a different limit.

Averaging and the CLT describe two related but distinct effects. Averaging reduces the variance of the unscaled sample mean, which concentrates it around $\mu$. Under the stated conditions, repeated convolution together with centering by $\mu$ and scaling by $\sigma/\sqrt{n}$ produces the normal limiting shape. Concentration by itself does not imply a bell curve.

Even if individual observations are not normally distributed, averages may be approximately normal when the sampling conditions and sample size make the CLT approximation adequate. Financial dependence, changing volatility, and very heavy tails can slow or invalidate that approximation, so normal-based inference for returns needs diagnostics or robust alternatives.

Let's demonstrate the CLT by repeatedly sampling and computing means: