Pareto Quantiles And Tail Index

Why Pareto Quantiles Matter

For a severity model, a quantile answers a practical question: what loss amount is exceeded only a specified fraction of the time? For a heavy-tailed model, that amount can grow quickly as the target exceedance probability becomes small. Pareto Type I provides a particularly direct way to calculate such thresholds and to see how tail behavior depends on the shape parameter.

This tutorial focuses on quantiles and tail probabilities for Pareto Type I losses. The earlier material on Pareto Type I severity and its moments establishes the model and its moment thresholds; here we use the survival function to quantify extreme-loss levels. All examples describe hypothetical insurance portfolios.

Definition: A loss $X$ has a Pareto Type I distribution with scale parameter $\theta>0$ and shape parameter $\alpha>0$ if $X\geq\theta$ and $$ F_X(x)= \begin{cases} 0, & x<\theta,\\ 1-\left(\dfrac{\theta}{x}\right)^\alpha, & x\geq\theta. \end{cases} $$ Its survival function is $S_X(x)=P(X>x)=(\theta/x)^\alpha$ for $x\geq\theta$, and $S_X(x)=1$ for $x<\theta$. The model is continuous, so $P(X=x)=0$ at every point. The scale parameter is the minimum possible loss; the shape parameter controls how quickly the upper tail decays.

For $0<p<1$, define the $p$-quantile as $F_X^{-1}(p)=\inf\{x:F_X(x)\geq p\}$. Because this Pareto distribution is continuous and strictly increasing for $x\geq\theta$, its $p$-quantile is the unique value $q_p$ satisfying $F_X(q_p)=p$. The equivalent exceedance statement is $P(X>q_p)=1-p$.

Inverting the Pareto Survival Function

A useful quantile formula follows directly from the survival function. At the $p$-quantile, the probability of exceeding the threshold is $1-p$. Substituting this into the Pareto survival function and solving for the threshold gives the result below.

Result: For a Pareto Type I loss with scale $\theta$ and shape $\alpha$, the $p$-quantile, for $0<p<1$, is $$ q_p=\theta(1-p)^{-1/\alpha}. $$ Equivalently, the threshold exceeded with probability $r$, for $0<r<1$, is $$ x_r=\theta r^{-1/\alpha}. $$ The second expression uses the exceedance probability $r$, not the cumulative probability $p$.

The derivation is short, but the probability being inverted matters. Starting with $F_X(q_p)=p$, use $F_X(x)=1-S_X(x)$ to obtain $S_X(q_p)=1-p$. Then solve the resulting equation:

\left(\frac{\theta}{q_p}\right)^\alpha=1-p \quad\Longrightarrow\quad \frac{\theta}{q_p}=(1-p)^{1/\alpha} \quad\Longrightarrow\quad q_p=\theta(1-p)^{-1/\alpha}.

The positive $1/\alpha$ power is valid because all quantities are positive. Since $0<1-p<1$, the factor $(1-p)^{-1/\alpha}$ is greater than one, so $q_p>\theta$, as required by the model's support. Substitution back into the survival function verifies the result:

S_X(q_p) =\left(\frac{\theta}{\theta(1-p)^{-1/\alpha}}\right)^\alpha =\left((1-p)^{1/\alpha}\right)^\alpha =1-p.

Worked Example: Finding a High Quantile

A hypothetical equipment insurer models the positive claim severity $X$ in dollars as Pareto Type I with $\theta=\$2{,}000$ and $\alpha=2$. Calculate the 95th percentile and verify the associated exceedance probability. Use the continuous Pareto model without rounding intermediate values.

Method. Use the quantile formula with $p=0.95$, so the target exceedance probability is $1-p=0.05$.

Step 1: Substitute the model parameters. $$ q_{0.95}=2{,}000(1-0.95)^{-1/2} =2{,}000(0.05)^{-1/2}. $$

Step 2: Evaluate the multiplier. Since $0.05=1/20$, $$ (0.05)^{-1/2}=\sqrt{20}=4.472135955\ldots, $$ so $$ q_{0.95}=2{,}000\sqrt{20} =8{,}944.27191\ldots. $$ The 95th percentile is $\$8{,}944.27$, rounded to the nearest cent.

Step 3: Check the answer. $$ P(X>8{,}944.27191\ldots) =\left(\frac{2{,}000}{8{,}944.27191\ldots}\right)^2 =\left(\frac{1}{\sqrt{20}}\right)^2 =0.05. $$ Thus the probability of exceeding the unrounded threshold is exactly $0.05$. The common mistake this method avoids is using $0.95$, the cumulative probability, as the survival probability.

How the Shape Parameter Controls the Tail

For a fixed scale $\theta$, compare the probability of exceeding a threshold $x\geq\theta$ under different values of $\alpha$: $$ S_X(x)=\left(\frac{\theta}{x}\right)^\alpha. $$ The ratio $\theta/x$ is at most one. Increasing $\alpha$ therefore makes this probability smaller at every $x>\theta$. A larger shape parameter means faster tail decay and fewer very large losses; a smaller shape parameter means slower tail decay.

The power-law form also gives an exact scaling relationship. If the threshold is multiplied by a constant $c\geq1$, the exceedance probability is multiplied by $c^{-\alpha}$:

\frac{S_X(cx)}{S_X(x)} = \frac{(\theta/(cx))^\alpha}{(\theta/x)^\alpha} =c^{-\alpha}, \qquad x\geq\theta.

Thus $\alpha$ is often called the Pareto tail index: it determines the power governing the decline in tail probability as the loss threshold grows. This conclusion is a comparison with the scale held fixed. If two models have different scale parameters, their tail probabilities must be compared using both parameters.

The quantile formula gives the same comparison from another direction. For fixed $\theta$ and $p$, the factor $(1-p)^{-1/\alpha}$ decreases as $\alpha$ increases. Accordingly, the $p$-quantile falls as the tail becomes lighter. As $p$ approaches one, $1-p$ approaches zero and the quantile grows without bound for every $\alpha>0$.

Worked Example: Comparing Two Tail Shapes

Two hypothetical portfolios have Pareto Type I claim severities with the same scale, $\theta=\$1{,}000$. Portfolio A has $\alpha=1.5$; Portfolio B has $\alpha=3$. Find each 99th percentile and each probability of exceeding $\$10{,}000$, then state which 99th percentile exceeds $\$10{,}000$.

Method. Calculate quantiles using $q_p=\theta(1-p)^{-1/\alpha}$, and calculate exceedance probabilities directly using $S_X(x)=(\theta/x)^\alpha$. This separate check prevents the common error of treating a percentile and an exceedance probability as though they had the same probability level.

Portfolio A: 99th percentile. Here $1-p=0.01$, so $$ q_{0.99,A} =1{,}000(0.01)^{-1/1.5} =1{,}000(100)^{2/3}. $$ Since $100^{2/3}=21.5443469\ldots$, this gives $q_{0.99,A}=\$21{,}544.35$, rounded to the nearest cent.

Portfolio A: exceedance at $\$10{,}000$. $$ P(X_A>10{,}000) =\left(\frac{1{,}000}{10{,}000}\right)^{1.5} =(0.1)^{1.5} =0.0316228\ldots $$ To four decimal places, this probability is $0.0316$.

Portfolio B: 99th percentile. $$ q_{0.99,B} =1{,}000(0.01)^{-1/3} =1{,}000(100)^{1/3} =1{,}000(4.6415888\ldots) =\$4{,}641.59. $$

Portfolio B: exceedance at $\$10{,}000$. $$ P(X_B>10{,}000) =(0.1)^3=0.001, $$ which is $0.0010$ to four decimal places.

Check and comparison. Only Portfolio A's 99th percentile exceeds $\$10{,}000$. At that threshold, Portfolio A's exceedance probability is $0.0316>0.01$, whereas Portfolio B's is $0.001<0.01$. The values are consistent: the threshold is below A's 99th percentile, but above B's 99th percentile. The smaller shape parameter in A produces the heavier tail.

Solving for a Tail-Probability Requirement

The same survival formula can be rearranged to determine the shape parameter needed to satisfy a specified exceedance limit. Suppose $\theta$ and a threshold $x>\theta$ are fixed, and the requirement is $P(X>x)\leq r$, where $0<r<1$. Taking natural logarithms of $(\theta/x)^\alpha\leq r$ gives $$ \alpha\ln(\theta/x)\leq\ln(r). $$ Because $\ln(\theta/x)<0$, dividing by it reverses the inequality. The requirement is therefore $$ \alpha\geq\frac{\ln(r)}{\ln(\theta/x)}. $$ This direction reversal is essential: dividing by a negative number without reversing the inequality gives the wrong minimum shape.

Worked Example: Meeting an Extreme-Loss Probability Limit

A hypothetical specialty insurer uses a Pareto Type I model with scale $\theta=\$5{,}000$. It wants the probability of a claim exceeding $\$50{,}000$ to be no more than $0.002$. Find the minimum shape parameter that meets the requirement, and check two candidate values, $\alpha=2.5$ and $\alpha=3$.

Method. Convert the tail-probability condition into an inequality for $\alpha$. The threshold-to-scale ratio is $50{,}000/5{,}000=10$, so the scale-to-threshold ratio is $0.1$.

Step 1: Set up the inequality. $$ P(X>50{,}000)=(0.1)^\alpha\leq0.002. $$ Take natural logarithms: $$ \alpha\ln(0.1)\leq\ln(0.002). $$ Since $\ln(0.1)=-2.302585093\ldots<0$, dividing reverses the inequality: $$ \alpha\geq\frac{\ln(0.002)}{\ln(0.1)} =\frac{-6.214608098\ldots}{-2.302585093\ldots} =2.698970\ldots. $$ Thus the minimum permissible shape is approximately $2.699$.

Step 2: Check $\alpha=2.5$. $$ (0.1)^{2.5}=0.00316228\ldots>0.002. $$ This candidate fails the requirement.

Step 3: Check $\alpha=3$. $$ (0.1)^3=0.001\leq0.002. $$ This candidate meets the requirement. The common mistake avoided here is dividing by $\ln(0.1)$ without reversing the inequality; that would incorrectly conclude that a smaller shape parameter is required.

Interpreting Quantiles and Avoiding Traps

A quantile statement and an exceedance statement describe the same threshold from opposite sides of the distribution. If $q_p$ is the $p$-quantile, then $P(X\leq q_p)=p$ and $P(X>q_p)=1-p$. For example, the 99th percentile is exceeded with probability $0.01$, not $0.99$. Keep the scale parameter and the threshold in the same monetary units before taking their ratio.

Another important distinction is between comparing models with a shared scale and comparing models with different scales. For the same $\theta$, a larger $\alpha$ means smaller exceedance probabilities above $\theta$. With different scales, that ordering need not hold at a given threshold: compute each survival probability using its own $\theta$ and $\alpha$. A fitted shape parameter is not, by itself, a complete description of the probability of exceeding a particular dollar amount.

Wrong tail level: Substituting $p$ in place of $1-p$ in the quantile formula confuses cumulative probability with exceedance probability. Use $q_p=\theta(1-p)^{-1/\alpha}$.
Wrong inequality direction: When solving a tail constraint for $\alpha$, check the sign of $\ln(\theta/x)$. For $x>\theta$, it is negative, so division reverses the inequality.
Unverified threshold: Confirm that any threshold at which the Pareto survival formula is used satisfies $x\geq\theta$. For $x<\theta$, the model gives $P(X>x)=1$.

Key takeaway: For Pareto Type I, $S_X(x)=(\theta/x)^\alpha$ and $q_p=\theta(1-p)^{-1/\alpha}$. The shape parameter is the tail index: with scale fixed, a larger value means faster tail decay, smaller extreme-loss probabilities, and lower high quantiles.

Check Your Understanding

For each problem, use the Pareto Type I model and distinguish carefully between a cumulative probability and an exceedance probability.

A hypothetical liability severity has $\theta=\$1{,}500$ and $\alpha=2$. Calculate its 96th percentile and verify the exceedance probability at that threshold. (Answer: $\$7{,}500.00$; exceedance probability $0.0400$.)
A hypothetical property severity has $\theta=\$3{,}000$ and $\alpha=1.5$. Find the probability that a loss exceeds $\$12{,}000$. (Answer: $0.1250$.)
A hypothetical equipment severity has $\theta=\$2{,}500$ and $\alpha=3$. Find the 95th percentile, rounded to the nearest cent. (Answer: $\$6{,}786.04$.)
Two hypothetical portfolios have $\theta=\$1{,}000$. Portfolio C has $\alpha=1$, and Portfolio D has $\alpha=4$. Find each 90th percentile and each probability of exceeding $\$5{,}000$. (Answer: Portfolio C: $\$10{,}000.00$ and $0.2000$; Portfolio D: $\$1{,}778.28$ and $0.0016$.)
A hypothetical model has $\theta=\$2{,}000$. What minimum value of $\alpha$ is needed to make the probability of exceeding $\$20{,}000$ no greater than $0.004$? (Answer: $\alpha\geq2.39794$, approximately $2.398$.)

What You'll Learn