Numerical Integration for Heavy-Tailed Aggregate Loss

Why Heavy-Tail Integration Needs Care

For a hypothetical aggregate loss $X$ and attachment $a$, the stop-loss payment is $(X-a)^+$, where $z^+=\max(z,0)$. Its expected value depends on the entire part of the aggregate survival curve above $a$. When that curve decreases slowly, a calculation over a seemingly wide range of losses can still omit a substantial share of the premium.

This tutorial focuses on approximating that omitted area responsibly. The numerical integral and the tail treatment are separate parts of the calculation: a fine grid can improve the integral up to a chosen endpoint, but it cannot by itself account for losses beyond that endpoint. The values below are hypothetical and are not company or industry data.

Definition: For a nonnegative aggregate loss $X$, let $S_X(x)=P(X>x)$. At attachment $a$, the stop-loss premium per unit of coverage, before expenses or other loadings, is $E[(X-a)^+]$. It equals the area under $S_X(x)$ from $a$ to infinity. This identity allows numerical integration directly from survival probabilities.

E[(X-a)^+]=\int_a^\infty S_X(x)\,dx =\int_a^T S_X(x)\,dx+\int_T^\infty S_X(x)\,dx,\qquad T>a.

The first integral is the part that can be approximated using survival values up to a finite truncation point $T$. The second is the remaining tail. The identity applies to nonnegative losses with a well-defined expectation; it does not require $X$ to have a density. In particular, an atom at the attachment does not invalidate it.

Numerical Integration and a Change of Scale

If survival probabilities are available at equally spaced points $x_0,\ldots,x_n$, with spacing $h$, the composite trapezoidal rule approximates the integral over that grid by

\int_{x_0}^{x_n}S_X(x)\,dx \approx h\left\{\frac{S_X(x_0)+S_X(x_n)}{2} +\sum_{j=1}^{n-1}S_X(x_j)\right\}.

This rule replaces the curve between adjacent points by a straight line. A finer grid can reduce this interpolation error, but it does not solve the truncation problem: the result still excludes all area above $x_n$. With heavy tails, equal spacing on the loss scale can also place many points where survival is already small while representing the broad upper range inefficiently.

A logarithmic scale is often more useful when the loss values span several multiples of the attachment. Set $t=\ln(x/a)$, so $x=ae^t$ and $dx=x\,dt$. The integral from $a$ to $T$ becomes an integral of $xS_X(x)$ over $t$. Equal steps in $t$ correspond to equal ratios, rather than equal dollar increments, between successive loss values.

\int_a^T S_X(x)\,dx =\int_0^{\ln(T/a)} ae^t S_X(ae^t)\,dt =\int_0^{\ln(T/a)} xS_X(x)\,dt.

Trapezoidal or Simpson quadrature can then be applied to the transformed integrand $xS_X(x)$. For Simpson’s rule, the transformed points must be equally spaced in $t$, and the number of subintervals must be even. Regardless of the rule, the survival values must correspond to the same aggregate-loss definition and scale as the attachment.

Tail approximation: If the survival tail is modeled as $S_X(x)=C x^{-\alpha}$ for $x\ge T$, with $C>0$ and $\alpha>1$, then $$ \int_T^\infty S_X(x)\,dx =\frac{C T^{1-\alpha}}{\alpha-1} =\frac{T S_X(T)}{\alpha-1}. $$ The final equality follows by substituting $C=T^\alpha S_X(T)$. The condition $\alpha>1$ is essential: for an exact power tail with $\alpha\le1$, the remaining integral diverges.

A fitted Pareto tail is an assumption, not a consequence of having a finite table of survival values. A slowly varying tail may not follow one power law over the entire range beyond $T$. If a tail index is estimated from data, the estimate can be particularly uncertain when only a few high-loss observations are available. Therefore, a tail-corrected premium should be accompanied by a stated tail model and a sensitivity check.

Worked Numerical Approaches

Worked Example: Simpson Integration with a Pareto Tail

Assume a hypothetical continuous aggregate loss $X$, measured in dollars, has Pareto survival $S_X(x)=(100/x)^{1.5}$ for $x\ge100$. Find the stop-loss premium at $a=\$100$ by integrating to $T=\$400$ on a logarithmic scale and applying the stated Pareto model beyond $T$. No interest adjustment or mortality assumption is involved.

Method and setup. Use the three equally spaced logarithmic points $x=100,200,400$. Their transformed integrand values $xS_X(x)$ are

100(1)=100,\qquad 200\left(\frac{1}{2}\right)^{1.5}=70.7107,\qquad 400\left(\frac{1}{4}\right)^{1.5}=50.

The transformed interval has length $\ln(400/100)=\ln4$, and its midpoint is $\ln2$. Simpson’s rule gives

\int_{100}^{400}S_X(x)\,dx \approx \frac{\ln4}{6}\left(100+4(70.7107)+50\right) =100.008.

Tail beyond $\$400$. Here $S_X(400)=0.125$, and the tail index is $\alpha=1.5$. Therefore,

\int_{400}^{\infty}S_X(x)\,dx \approx\frac{400(0.125)}{1.5-1}=100.

Estimate and check. The estimated premium is $100.008+100=\$200.008$, or \$200.01 to the nearest cent. For a check, direct integration of the specified Pareto survival gives $100/(1.5-1)=\$200$. The small difference is quadrature error on the finite interval. The common mistake is to report only the $\$100.008$ numerical integral and omit the equally large tail area.

Worked Example: Fitting a Tail to a Survival Table

A hypothetical aggregate-loss study supplies $S_X(x)=0.4000$ at $\$100$, $0.141421$ at $\$200$, $0.050000$ at $\$400$, and $0.017678$ at $\$800$. Estimate the stop-loss premium at $a=\$100$, using logarithmic trapezoidal integration through $\$800$ and a Pareto tail beyond $\$800$. Assume no interest adjustment or mortality component.

Method and tail assumption. Each doubling of $x$ multiplies the displayed survival probability by approximately $2^{-1.5}$. For this calculation, fit a Pareto tail with $\alpha=1.5$ through the endpoint. The transformed values $xS_X(x)$, rounded to six decimals, are $40.000000$, $28.284200$, $20.000000$, and $14.142400$. The logarithmic step is $h=\ln2=0.693147$.

Finite-range integral. The trapezoidal rule on the three intervals gives

\int_{100}^{800}S_X(x)\,dx \approx \ln2\left(\frac{40+14.1424}{2}+28.2842+20\right) =52.232.

Remaining tail. Using the fitted $\alpha=1.5$ at $T=800$,

\int_{800}^{\infty}S_X(x)\,dx \approx \frac{800(0.017678)}{0.5}=28.2848.

The estimated stop-loss premium is $52.232+28.2848=\$80.5168$, or \$80.52 to the nearest cent. As a check under the fitted Pareto model, integrating its survival from $\$100$ gives $100(0.4000)/0.5=\$80$; the difference reflects trapezoidal approximation and rounded table values. The common mistake is to treat the last observed survival probability as a probability of payment and stop there: the premium is an area, not the endpoint probability.

Worked Example: A Long Tail Can Dominate the Estimate

Suppose a hypothetical continuous aggregate loss $X$, in dollars, has survival $S_X(x)=0.3000(200/x)^{1.25}$ for $x\ge200$. Estimate the stop-loss premium at $a=\$200$ by logarithmic trapezoidal integration through $T=\$1{,}600$, followed by a Pareto tail correction. There is no interest adjustment or mortality assumption.

Finite-range approximation. Use the equally spaced logarithmic points $200,400,800,1{,}600$. The values of $xS_X(x)$ are $60.0000$, $50.4538$, $42.4264$, and $35.6763$. With $h=\ln2$, the trapezoidal approximation is

\int_{200}^{1600}S_X(x)\,dx \approx\ln2\left(\frac{60+35.6763}{2}+50.4538+42.4264\right) =97.539.

Tail correction. The endpoint survival is $S_X(1{,}600)=0.0222977$, and $\alpha-1=0.25$. Hence,

\int_{1600}^{\infty}S_X(x)\,dx \approx\frac{1600(0.0222977)}{0.25}=142.705.

The estimate is $97.539+142.705=\$240.244$, or \$240.24 to the nearest cent. Direct integration of the stated model gives $200(0.3000)/(1.25-1)=\$240$, confirming the scale of the answer. The correction beyond $\$1{,}600$ is larger than the integral below that point. The common mistake is to assume that a large truncation point makes the remaining tail negligible; its size depends on both endpoint survival and tail decay.

Choosing and Checking the Truncation Point

A practical calculation should show how the endpoint and tail model affect the result. A larger $T$ usually leaves less area to estimate, but it may require extrapolating from fewer reliable observations. A smaller $T$ may rely more heavily on a fitted tail. Neither choice is automatically safer.

Set the scale and attachment.
Keep the loss units consistent in the attachment, grid, survival probabilities, and final premium. Decide whether a logarithmic grid is appropriate for the range of losses.

Integrate the observed range.
Use a stated quadrature rule. Refine the grid, or compare rules when the survival curve bends appreciably between supplied points.

Specify the tail treatment.
State whether the tail uses a fitted Pareto index, another justified model, or a defensible bound. Do not silently set the tail area to zero.

Test sensitivity.
Recalculate with a nearby truncation point or plausible alternative tail index. A large change indicates that the premium is materially dependent on extrapolation.

For a Pareto tail, the estimate $TS_X(T)/(\alpha-1)$ is especially sensitive when $\alpha$ is close to one. For example, holding $T S_X(T)$ fixed, reducing $\alpha$ from $1.5$ to $1.2$ increases the estimated tail area from $2TS_X(T)$ to $5TS_X(T)$. A reported premium should not hide that sensitivity through premature rounding.

A useful numerical check is to compare the finite-range integral under a refined grid with the original calculation. Separately compare the tail correction under plausible tail assumptions. If a claimed precision is smaller than the change from either check, that precision is not supported. There is no universal truncation point or error guarantee without assumptions about the curve beyond the observed range.

Key takeaway: A heavy-tailed stop-loss estimate has two components: numerical area over the represented range and an explicit estimate of the unrepresented tail. Assess and report both; refining the first does not validate the second.

Check Your Understanding

For each problem, state or use the specified numerical method and keep any tail assumption explicit.

A hypothetical aggregate loss has $S_X(x)=(50/x)^2$ for $x\ge50$. At attachment $a=\$100$, integrate to $T=\$200$ and use the exact Pareto tail correction. What is the total premium? (Answer: $\$25.00$)
At $a=\$100$, suppose $S_X(x)=0.25(100/x)^{1.4}$ for $x\ge100$. Using $T=\$400$, what is the Pareto tail correction beyond $T$, and what is the exact premium under this model? (Answer: tail correction $\$35.90$; exact premium $\$62.50$)
A survival table gives $S_X(100)=0.4000$, $S_X(200)=0.1000$, and $S_X(400)=0.0250$. At attachment $\$100$, use logarithmic trapezoids through $\$400$, then assume a Pareto tail with $\alpha=2$. What is the approximate premium? (Answer: $\$41.19$)
At $T=\$1{,}000$, a fitted survival probability is $S_X(T)=0.0200$. Compare Pareto tail corrections for $\alpha=1.2$ and $\alpha=1.5$. What does the difference imply about sensitivity? (Answer: $\$100.00$ and $\$40.00$, respectively; the tail estimate is highly sensitive to the assumed index.)
A power survival model has $S_X(x)=0.20(250/x)^\alpha$ for $x\ge250$, with attachment $a=\$250$. For which values of $\alpha$ is the stop-loss premium finite, and what is it when $\alpha=1.25$? (Answer: finite only for $\alpha>1$; at $\alpha=1.25$, the premium is $\$200.00$.)

What You'll Learn