Free lesson · Statistics

Random outcomes, sample averages and the limits of the bell curve

Open interactive lesson Practice calculations Explore labs

Start with the idea

Imagine repeatedly running the same uncertain experiment. Individual outcomes vary. An average combines those outcomes, but the average itself would change if you collected a different sample.

Symbols, units & horizon

X_i: outcome i in dollars per trade
n: number of independent identically distributed observations
X̄: sample average in dollars
σ²: finite population variance in dollars squared
Var: variance operator
s: sample standard deviation in dollars
SE: estimated sampling standard deviation of the mean
Σ: add the terms indexed from 1 through n

When and why to use this

Use sampling uncertainty when deciding whether an average net payoff is distinguishable from noise. For dependent strategy returns, continue to the block-bootstrap and effective-sample-size lessons in Research & Validation.

For payoffs $10, $20 and $30, the average is (10+20+30)/3=$20. Deviations from it are −$10, $0 and $10. Their squared sum is 200 dollars squared; dividing by n−1=2 gives sample variance 100 dollars squared and square-rooting gives sample SD $10. The average describes location and SD describes spread.

Start with three different objects: the distribution of one trade, the distribution of the average of many trades, and uncertainty about the model itself. The central limit theorem concerns suitably normalized sums or averages under conditions; it does not turn the original trade returns into a normal distribution.

For independent observations with the same finite variance, adding observations increases the variance of the sum proportionally to the sample count. Dividing the sum by that count reduces the variance of the average. Standard error is the standard deviation of that average across hypothetical repetitions.

A normal approximation to the sampling distribution can be useful with enough suitable data. Small samples, shared exposures, structural changes and heavy tails can make it poor. If variance is infinite, the familiar square-root sample-size formula has no finite population variance to use. “Randomness has shape” should mean a testable distributional model, not certainty about the shape.

Assess skill using the whole decision process and its uncertainty, not one profitable outcome. Likewise, a loss can occur after a sound positive-expectancy decision. Rare large losses require separate tail scenarios: a precise estimate of an average is not a cap on individual losses.

Var (\overline{X}) = \frac{1}{n^{2}} \sum_{i = 1}^{n} Var (X_{i}) = \frac{σ^{2}}{n}, \hat{S E} (\overline{X}) = \frac{s}{\sqrt{n}}

Model assumptions, derivation and arithmetic

Random outcomes, sample averages and the limits of the bell curve

Write the average as (X₁+…+Xₙ)/n. Multiplying a random variable by 1/n multiplies its variance by 1/n².
Independence makes every covariance between distinct observations zero. The sum has variance nσ², so division by n² gives σ²/n.
Take the square root to obtain σ/√n. Replace unknown σ with the sample estimate s; this substitution introduces estimation uncertainty.
A normal sampling approximation may support an approximate interval mean ±1.96 SE under suitable conditions. It is an approximation, not an algebraic consequence of the variance identity.

Work it by hand

Suppose n=100 independent trades have mean $2 and sample SD $10. Estimated SE=10/√100=$1. The illustrative normal interval is 2±1.96×1, or [$0.04,$3.96]. With n=25 and the same SD, SE=$2. Correlation or selection invalidates this simple interpretation.

Apply it in a strategy

Define the return or dollar-payoff horizon and net cost convention.
Inspect dependence, tail behavior and regime changes before choosing an uncertainty estimator.
Compare the prospective decision with a benchmark, and report both tail scenarios and uncertainty around average performance.

Research deliverable

Explain the difference between one-trade risk, sampling error of the mean and model uncertainty using a small ledger.

Sources & evidence · reviewed 12 September 2026

Reviewed 12 September 2026. Continue to the existing research checkpoint on dependent samples, selection and block resampling. The numerical examples here are constructed teaching calculations. The supplied infographic motivates the distinction between decision quality and outcome; its general claims about randomness and normality are not used as mathematical premises.

Further reading: Research & Validation: dependence and selection ↗

Python implementation

Self-contained teaching example. Python 3.10+; dependencies and input conventions are shown in the code and notation. Run in your own Python environment.

from math import sqrt, isfinite

def mean_uncertainty(mean, sample_sd, count):
    if not all(isfinite(x) for x in [mean, sample_sd]) or sample_sd<0 or not isinstance(count,int) or count<2:
        raise ValueError("Finite mean, nonnegative SD and integer count >=2 required")
    se=sample_sd/sqrt(count)
    return se,(mean-1.96*se,mean+1.96*se) # normal approximation, iid assumptions

print(mean_uncertainty(2,10,100))

Continue learning

Statistics & Probability — all lessons

Quantitative finance and development glossary · Python resources and libraries · Research sources and limitations