Method-of-Moments Estimation for Pareto Losses

Why Pareto Moment Estimation Requires a Range Check

The Pareto Type I model is useful for representing positive losses with a heavy upper tail. Its tail can be sufficiently heavy that some population moments are infinite. Consequently, a method-of-moments calculation is not just a matter of solving equations: the moments used must exist under the model, and the observed sample moments must produce parameters in the model's permitted range.

This tutorial uses the Pareto Type I parameterization with minimum loss $\theta$ and shape parameter $\alpha$. State the parameterization before estimating: other distributions also called Pareto use different support and moment formulas.

Definition: A loss $X$ has a Pareto Type I distribution with scale $\theta>0$ and shape $\alpha>0$ if its support is $x\geq\theta$ and $$ F_X(x)=1-\left(\frac{\theta}{x}\right)^\alpha,\qquad S_X(x)=\left(\frac{\theta}{x}\right)^\alpha,\qquad x\geq\theta. $$ Its density on that support is $f_X(x)=\alpha\theta^\alpha x^{-(\alpha+1)}$.

The scale parameter is the model's lower bound: $P(X\geq\theta)=1$. The shape parameter controls tail heaviness. Larger $\alpha$ means a thinner tail; it does not change the lower bound independently of $\theta$.

Population Moments and Their Existence

For a positive real number $r$, the $r$th raw moment is $E[X^r]$. For the Pareto Type I density, its defining integral is

E[X^r] =\int_\theta^\infty x^r\alpha\theta^\alpha x^{-(\alpha+1)}\,dx =\alpha\theta^\alpha\int_\theta^\infty x^{r-\alpha-1}\,dx.

The integral converges only if $r-\alpha-1<-1$, or equivalently $\alpha>r$. Under that condition, evaluating the integral gives

E[X^r]=\frac{\alpha\theta^r}{\alpha-r},\qquad \alpha>r.

This condition is part of the formula, not an optional technical detail. In particular, the mean requires $\alpha>1$, and the second raw moment requires $\alpha>2$. When $\alpha>2$, the variance is

\operatorname{Var}(X) =E[X^2]-E[X]^2 =\frac{\alpha\theta^2}{(\alpha-1)^2(\alpha-2)}.

If $1<\alpha\leq2$, the mean is finite but the second raw moment and variance are infinite. Matching both the first and second population moments is then not justified.

Moment conditions: The Pareto Type I moment $E[X^r]$ is finite exactly when $\alpha>r$. Thus a two-raw-moment method requires $\alpha>2$; a parameter estimate with $\alpha\leq2$ cannot support the second-moment equation.

Deriving the Two-Moment Estimator

Given positive observations $X_1,\ldots,X_n$, define the first two raw sample moments by $m_1=n^{-1}\sum_{j=1}^n X_j$ and $m_2=n^{-1}\sum_{j=1}^n X_j^2$. Method of moments sets these equal to the corresponding population moments. For this model, the equations are

m_1=\frac{\alpha\theta}{\alpha-1},\qquad m_2=\frac{\alpha\theta^2}{\alpha-2},\qquad \alpha>2.

The condition $\alpha>2$ ensures both equations are defined and finite. Divide the second equation by the square of the first to eliminate $\theta$. With $q=m_2/m_1^2$, the population ratio simplifies to

q=\frac{\alpha\theta^2}{\alpha-2} \left(\frac{\alpha-1}{\alpha\theta}\right)^2 =\frac{(\alpha-1)^2}{\alpha(\alpha-2)} =1+\frac{1}{\alpha(\alpha-2)}.

This ratio explains why the sample-moment check matters. For every finite $\alpha>2$, $q>1$. Solving $q-1=1/[\alpha(\alpha-2)]$ gives $\alpha^2-2\alpha=q/(q-1)$. The root satisfying $\alpha>2$ is then substituted into the first moment equation to obtain $\theta$.

Two-moment Pareto Type I estimates: If $m_1>0$ and $q=m_2/m_1^2>1$, the finite estimates are $$ \widehat{\alpha}=1+\sqrt{\frac{q}{q-1}},\qquad \widehat{\theta}=m_1\frac{\widehat{\alpha}-1}{\widehat{\alpha}}. $$ Equivalently, if $s_m^2=m_2-m_1^2$ is the sample variance calculated with divisor $n$, then $\widehat{\alpha}=1+\sqrt{1+m_1^2/s_m^2}$. The two expressions agree because $q=1+s_m^2/m_1^2$.

For positive observations, $m_2\geq m_1^2$, since the sample variance $m_2-m_1^2$ is nonnegative. Equality holds exactly when all observations are equal. Thus a finite two-moment estimate requires a nonconstant sample: $q>1$. If $q=1$, the displayed estimates have no finite solution; if a proposed calculation has $q<1$, the raw moments have been computed inconsistently.

Calculate raw moments.
Find $m_1$ and $m_2$ from the observations, keeping $m_2$ as the average of squared losses.

Check the moment ratio.
Compute $q=m_2/m_1^2$. A finite two-moment Pareto Type I estimate requires $q>1$.

Solve for shape, then scale.
Use $\widehat{\alpha}=1+\sqrt{q/(q-1)}$, then $\widehat{\theta}=m_1(\widehat{\alpha}-1)/\widehat{\alpha}$.

Verify the equations.
Check that the fitted first and second raw moments reproduce $m_1$ and $m_2$, and confirm $\widehat{\alpha}>2$.

Worked Example: Estimating Parameters From Two Losses

A hypothetical portfolio has two observed losses, \$120 and \$360. Fit the Pareto Type I distribution by matching the first two raw sample moments. Give the estimates to the nearest cent for the scale and to four decimal places for the shape.

Method and sample moments. There are $n=2$ observations, so $m_1=(120+360)/2=240$ dollars. The second raw moment is $m_2=(120^2+360^2)/2=(14{,}400+129{,}600)/2=72{,}000$ dollars squared.

Check and solve. The ratio is $q=72{,}000/240^2=72{,}000/57{,}600=5/4$, which is greater than 1. Therefore, $\widehat{\alpha}=1+\sqrt{(5/4)/(1/4)}=1+\sqrt{5}$, approximately $3.2361$. The scale estimate is $\widehat{\theta}=240\sqrt{5}/(1+\sqrt{5})$ dollars, approximately \$165.84.

Verification. Since $(\widehat{\alpha}-1)/\widehat{\alpha}=\sqrt{5}/(1+\sqrt{5})$, the fitted mean is $\widehat{\theta}\widehat{\alpha}/(\widehat{\alpha}-1)=240$ dollars. Also, $\widehat{\alpha}(\widehat{\alpha}-2)=(1+\sqrt{5})(\sqrt{5}-1)=4$. Hence the fitted variance is $240^2/4=14{,}400$ dollars squared, and the fitted second raw moment is $14{,}400+240^2=72{,}000$ dollars squared, as required. The estimates meet the second-moment condition because $\widehat{\alpha}>2$.

Common mistake avoided. The value $72{,}000$ is the raw second moment, not the variance. Substituting it as a variance would give the wrong moment ratio and therefore the wrong parameters.

Interpreting the Moment Ratio

The ratio $q$ is dimensionless: both $m_2$ and $m_1^2$ have units of squared dollars. It is also $1+s_m^2/m_1^2$, where $s_m^2$ uses divisor $n$. A sample whose losses have relatively little spread compared with their mean has $q$ near 1 and produces a large estimated $\alpha$. A larger $q$ produces an estimated $\alpha$ closer to 2, where the model's second moment is close to its divergence boundary.

These relationships are useful diagnostics, not evidence that the Pareto model is appropriate. Method of moments matches selected summaries; it does not guarantee that the model describes the full loss distribution. In particular, an estimated shape just above 2 permits a finite second moment but implies a heavy tail, and moments of order $r\geq\widehat{\alpha}$ are not finite under the fitted model.

Worked Example: Using Raw Moments and Checking a Higher Moment

A hypothetical sample of four losses is \$50, \$50, \$50, and \$250. Estimate the Pareto Type I parameters by two moments, then determine whether the fitted model has a finite third raw moment. Round the shape to four decimal places and the scale to the nearest cent.

Calculate the raw moments. The mean is $m_1=(50+50+50+250)/4=100$ dollars. The second raw moment is $m_2=(2{,}500+2{,}500+2{,}500+62{,}500)/4=70{,}000/4=17{,}500$ dollars squared. Thus $q=17{,}500/100^2=1.75=7/4$, so the sample is nonconstant and the equations are usable.

Estimate the parameters. The shape estimate is $\widehat{\alpha}=1+\sqrt{(7/4)/(3/4)}=1+\sqrt{7/3}$, approximately $2.5275$. The scale estimate is $\widehat{\theta}=100\sqrt{7/3}/(1+\sqrt{7/3})$, approximately \$60.44.

Check the moments and the third-moment condition. The fitted mean equals $100$ dollars by construction. Also, $\widehat{\alpha}(\widehat{\alpha}-2)=(1+\sqrt{7/3})(\sqrt{7/3}-1)=4/3$. The fitted variance is therefore $100^2/(4/3)=7{,}500$ dollars squared, giving a second raw moment of $7{,}500+100^2=17{,}500$ dollars squared. A third raw moment would require $\widehat{\alpha}>3$. Here $2.5275<3$, so that moment is infinite under the fitted model.

Common mistake avoided. A finite fitted variance does not imply that every higher moment is finite. Each moment has its own condition: the third requires shape greater than 3.

Common Errors and Exam Checks

Using a different Pareto parameterization. These equations apply to Type I, with support $x\geq\theta$ and survival $(\theta/x)^\alpha$. A model with support starting at zero has different formulas.
Confusing the raw second moment and variance. Use $m_2$ in $q=m_2/m_1^2$. If using variance instead, first use $m_2=m_1^2+s_m^2$.
Ignoring moment existence. The second-moment equation is usable only for $\alpha>2$. The condition must be checked before treating it as a population equation.
Taking the wrong quadratic root. The algebra also produces a root below 2. It cannot be used in a two-moment estimate because the second moment would not be finite.
Forgetting the nonconstant-sample condition. If $m_2=m_1^2$, then $q=1$, and there is no finite estimate from these two equations.

Key takeaway: For Pareto Type I, two-moment estimation is valid only when the model's second moment exists and the sample has $q=m_2/m_1^2>1$. Estimate the shape from this dimensionless ratio, recover the scale from the mean, and verify both raw moments before interpreting the fitted tail.

Check Your Understanding

For each problem, use the Pareto Type I parameterization and the two-moment method where it is valid.

A hypothetical sample has raw moments $m_1=60$ dollars and $m_2=5{,}400$ dollars squared. Find $\widehat{\alpha}$ and $\widehat{\theta}$ in exact form. (Answer: $\widehat{\alpha}=1+\sqrt{3}$, $\widehat{\theta}=60\sqrt{3}/(1+\sqrt{3})$ dollars.)
A sample has mean $200$ dollars and sample variance $8{,}000$ dollars squared, with variance calculated using divisor $n$. Find the two estimates in exact form. (Answer: $\widehat{\alpha}=1+\sqrt{6}$, $\widehat{\theta}=200\sqrt{6}/(1+\sqrt{6})$ dollars.)
A sample consists of losses \$40, \$40, \$40, and \$160. Find $m_1$, $m_2$, and $q$, then give the shape estimate in exact form and determine whether the fitted third raw moment is finite. (Answer: $m_1=70$, $m_2=7{,}600$, $q=76/49$, $\widehat{\alpha}=1+\sqrt{76/27}<3$, so the third raw moment is infinite.)
A proposed sample summary gives $m_1=100$ dollars and $m_2=9{,}000$ dollars squared. Can finite two-moment Pareto Type I estimates be obtained? Explain using the ratio condition. (Answer: No; $q=0.9<1$, which is inconsistent with raw moments of positive observations.)
A sample of positive losses has $m_1=120$ dollars and $s_m^2=0$. What does the two-moment method imply about a finite shape estimate? (Answer: No finite estimate; $q=1$, so the formula's denominator $q-1$ is zero.)

What You'll Learn