The Pareto Distribution and Power Laws

City sizes, trading volume across stocks, wealth, and the size of insurance claims all share an odd feature: a small number of instances are enormously larger than the typical one, and this isn't a fluke — it's the reliable pattern. A normal distribution would predict those giants essentially never happen. The Pareto distribution is built specifically to describe data where a few extreme values dominate the total, and to make precise exactly how likely those giants are.

The analogy: the 80/20 rule, made exact

"20% of clients generate 80% of trading volume" is the informal version of a Pareto pattern. The Pareto distribution turns that folk rule into a formula: instead of most values clustering near an average, values are spread across scales — some near the minimum, progressively fewer as you go up, but the "progressively fewer" happens slowly enough that genuinely enormous values keep showing up far more often than a bell curve would ever allow. One parameter, the tail index, controls exactly how slowly that thinning-out happens — and how heavy the consequences are.

Writing it down

The Pareto distribution with minimum value $x_m>0$ and tail index $\alpha>0$ has survival function (probability of exceeding $x$):

$$P(X > x) = \left(\frac{x_m}{x}\right)^{\alpha}, \qquad x \ge x_m.$$

In words: the chance of seeing a value bigger than $x$ shrinks as a power of $x$, not an exponential of $x$ — this is the defining feature of a "fat tail," and it's why extreme values stay far more plausible, far further out, than under a normal or exponential model. The mean exists only if $\alpha>1$: $E[X] = \frac{\alpha x_m}{\alpha - 1}$. The variance exists only if $\alpha>2$. For $\alpha \le 2$, variance is infinite; for $\alpha \le 1$, even the mean is infinite — a direct generalization of the Cauchy distribution's broken-mean behavior, with $\alpha$ as the dial controlling exactly how broken it gets.

Set a low (negative, since this is a decay) exponent and watch the curve fall off slowly, staying meaningfully above zero far to the right — that's the visual signature of a small $\alpha$: a heavy tail where big values remain plausible long after a normal distribution would have written them off entirely.

Worked example 1: insurance claim sizes

Claims above a $1,000 deductible follow $\text{Pareto}(x_m=1000, \alpha=2.5)$. What's the probability a claim exceeds $10,000? $P(X>10{,}000) = (1000/10000)^{2.5} = (0.1)^{2.5} \approx 0.00316$ — about 1 in 316 claims. Mean claim size: \frac{2.5 \times 1000}{1.5} \approx \1{,}667$. Now compare to$\alpha=1.5$(a fatter tail, same minimum):$P(X>10{,}000) = (0.1)^{1.5} \approx 0.0316$— ten times more likely, and worse, the mean is now$\frac{1.5\times1000}{0.5}=$3{,}000$, nearly double, entirely driven by how much heavier the tail got. A one-point drop in$\alpha$ changed both the typical claim size and the frequency of catastrophic ones dramatically.

Worked example 2: does trading volume across stocks fit a power law?

Suppose daily dollar volume across 500 stocks has a minimum (screened) value of x_m = \1\text{m}$and an estimated tail index$\hat\alpha = 1.2$(fit via maximum likelihood:$\hat\alpha = n / \sum \ln(x_i/x_m)$, a standard closed-form Pareto MLE). Since$\hat\alpha < 2$, variance is formally infinite — meaning as you add more stocks to a volume-weighted average, the estimate never stabilizes the way it would for a well-behaved distribution; the largest stock in the sample keeps dominating no matter how many are added. Since$\hat\alpha$ is barely above 1, even the mean daily volume is barely well-defined — a single mega-cap name's flow can distort a "total market volume" statistic in a way that never averages out, matching the well-known empirical fact that trading volume is dominated by a handful of names.

Hold this thin-tailed curve in mind as the baseline, then compare it against the power-law shape above: a normal distribution's tail is essentially zero past 4 or 5 standard deviations, while a Pareto tail with a small $\alpha$ stays meaningfully above zero for values dozens of "typical" widths away — the entire practical difference between "extreme events don't happen" and "extreme events are a routine part of the pattern."

What this means in practice

• Position sizing and concentration risk. If a book's P&L or a market's volume genuinely follows a power law with $\alpha$ near or below 2, standard deviation is a misleading risk measure — it may not even be estimating anything finite.
• Tail index estimation matters more than the mean. For fat-tailed data, the single most important number to estimate accurately is $\alpha$ — it directly answers "how much worse can the worst case realistically be," which the sample mean cannot.
• Related everywhere fat tails appear. City populations, word frequencies, insurance losses, and crash magnitudes are classic empirical Pareto-like fits — always sanity check the tail index rather than assuming a normal or lognormal shape by default.