Extrapolating Exposure Curves Beyond Observed Limits

Why Extrapolation Needs an Explicit Tail Assumption

An empirical exposure curve is most informative where the underlying claims data provide support. At very high limits, relatively few claims may reach the relevant range, so the observed curve can become unstable or stop altogether. Extending it is not a mechanical continuation of the last observed point: it requires an assumption about the severity tail beyond the highest limit that is well supported by data.

The choice matters because a tail assumption affects both the limited expected loss at a higher limit and, when the exposure curve is normalized by the unlimited expected loss, its denominator. A curve that looks similar near the observed threshold can produce a different factor farther into the tail. The objective is to make the assumption explicit, calculate its implications, and test how a reasonable alternative changes the result.

Definition: Let \(X\) be ground-up loss, let \(m(u)=E[\min(X,u)]\) be the limited expected loss at limit \(u\), and let \(G(u)=m(u)/E[X]\) be the exposure curve when \(E[X]\) is finite and positive. If observations support the curve only through threshold \(a\), tail extrapolation estimates \(m(u)\), for \(u>a\), by combining the supported value \(m(a)\) with an assumed survival function above \(a\). The tail assumption is a model for losses beyond the threshold, not necessarily for the entire severity distribution.

Write \(S_X(x)=P(X>x)\). The value \(m(a)\) summarizes the limited losses up to \(a\); the additional limited expected loss between \(a\) and a higher limit \(u\) depends on the survival probabilities in that interval. The unlimited mean also depends on survival above \(a\). Thus, a usable extrapolation needs the supported limited mean, the survival probability at the threshold, and a specified tail shape.

The Tail Extension Formula

For a nonnegative loss, the increase in limited expected loss from \(a\) to \(u\) is the area under the survival function from \(a\) to \(u\). This remains true whether the tail comes from data or an assumed model. For extrapolation, use the assumed survival function over that interval. If the resulting unlimited mean is finite, it is found by extending the same integral to infinity.

m(u)=m(a)+\int_a^u S_X(x)\,dx,\qquad E[X]=m(a)+\int_a^\infty S_X(x)\,dx.

These equations use \(m(a)\) as the supported starting value. Do not add the expected loss below \(a\) a second time. Also, the survival probability at the threshold must be the probability of exceeding that threshold, \(S_X(a)\), on the same loss basis as \(m(a)\).

Justification: For each individual loss, \(\min(X,u)-\min(X,a)\) is the length of the part of the interval \((a,u)\) that lies below \(X\). Equivalently, it is \(\int_a^u \mathbf{1}_{\{X>x\}}\,dx\). Taking expectations and interchanging expectation and integration for this nonnegative integrand gives \(m(u)-m(a)=\int_a^u P(X>x)\,dx\). Taking the upper limit to infinity gives the formula for \(E[X]\), provided the integral is finite.

Two Common Tail Assumptions

Two useful assumptions illustrate how different tail shapes can be extrapolated from the same threshold. They are alternatives, not universal descriptions of insurance severity. In either case, the model is applied only for \(x\ge a\), and its survival probability at \(a\) must match the supported value \(S_X(a)\).

Exponential tail beyond \(a\): Assume \(S_X(a+t)=S_X(a)e^{-t/\beta}\) for \(t\ge0\), where \(\beta>0\) is the scale parameter. The expected excess above \(a\), conditional on exceeding \(a\), is \(\beta\). The tail integral through \(u\) is \(S_X(a)\beta[1-e^{-(u-a)/\beta}]\); the integral to infinity is \(S_X(a)\beta\).

Pareto tail beyond \(a\): Assume \(S_X(x)=S_X(a)(a/x)^\alpha\) for \(x\ge a>0\), where \(\alpha>0\). For \(\alpha>1\), the expected excess above \(a\), conditional on exceeding \(a\), is \(a/(\alpha-1)\), and the tail integral to infinity is \(S_X(a)a/(\alpha-1)\). For \(u>a\), the tail integral through \(u\) is \(S_X(a)a[1-(a/u)^{\alpha-1}]/(\alpha-1)\). If \(\alpha\le1\), the unlimited mean is infinite under this tail assumption, so a finite-mean exposure curve \(G(u)\) cannot be formed without an additional finite upper bound or another tail model.

The exponential tail declines at a constant rate relative to its current survival probability. The Pareto tail declines as a power of the loss amount and, when \(\alpha>1\), has a finite mean but a more persistent far tail than an exponential with the same conditional mean excess. Estimates can therefore diverge as the limit moves farther beyond the observed range.

Worked Extrapolations

Worked Example: Extending an Exponential Tail

A hypothetical portfolio has supported limited expected loss \(m(a)=40\) thousand dollars at \(a=100\) thousand dollars, with \(S_X(a)=0.08\). Assume an exponential tail beyond \(a\) with \(\beta=50\) thousand dollars. Estimate the exposure curve at \(u=200\) thousand dollars. Report expected losses in thousands of dollars and the curve to four decimals.

Method. Use the exponential tail integral for both the limited expected loss at \(u\) and the unlimited mean. The same tail assumption must be used in both calculations.

Step 1: Calculate the tail addition through \(u\). Here \(u-a=200-100=100\) thousand dollars, so

\int_a^u S_X(x)\,dx =0.08(50)\left(1-e^{-100/50}\right) =4(1-e^{-2}) =4(0.8646647) =3.4587.

Therefore \(m(200)=40+3.4587=43.4587\) thousand dollars.

Step 2: Calculate the unlimited mean. The remaining tail integral is \(S_X(a)\beta=0.08(50)=4\) thousand dollars. Hence \(E[X]=40+4=44\) thousand dollars.

Step 3: Normalize and check.

G(200)=\frac{43.4587}{44}=0.9877.

The result is less than 1, as an exposure curve at a finite limit should be when the loss has a positive probability of exceeding that limit. A common mistake is to divide by \(m(a)=40\), giving \(43.4587/40=1.0865\). That uses the limited mean at the threshold instead of the unlimited mean, so it is not the exposure curve defined here.

Worked Example: Comparing Two Tails With the Same Mean Excess

Use the threshold information from the preceding hypothetical portfolio: \(a=100\) thousand dollars, \(m(a)=40\) thousand dollars, and \(S_X(a)=0.08\). Compare an exponential tail with \(\beta=50\) to a Pareto tail with \(\alpha=3\) at \(u=200\) and \(u=500\) thousand dollars. Report expected losses in thousands and curves to four decimals.

Method. First verify that the tails have the same threshold survival and conditional mean excess. Then calculate their limited means and normalize each by its own unlimited mean.

For the exponential tail, the conditional mean excess is \(\beta=50\). For the Pareto tail it is \(a/(\alpha-1)=100/(3-1)=50\). Both therefore imply an unlimited tail addition of \(0.08(50)=4\), giving \(E[X]=40+4=44\) thousand dollars.

At \(u=200\). The exponential tail addition is \(4(1-e^{-2})=3.4587\), so \(m(200)=43.4587\) and \(G(200)=43.4587/44=0.9877\). For the Pareto tail,

\int_{100}^{200}S_X(x)\,dx =0.08(100)\left[1-\left(\frac{100}{200}\right)^2\right] =8(0.75)=6.

Thus \(m(200)=40+6=46\) thousand dollars and \(G(200)=46/44=1.0455\). This result signals an inconsistency: the Pareto tail integral above was misapplied. For \(\alpha=3\), the correct denominator in the finite-interval formula is \(\alpha-1=2\), so the integral is \(0.08(100)[1-(100/200)^2]/2=3\). Therefore the corrected values are \(m(200)=43\) thousand dollars and \(G(200)=43/44=0.9773\).

At \(u=500\). For the exponential tail the addition is \(4(1-e^{-8})=3.9987\), so \(m(500)=43.9987\) and \(G(500)=43.9987/44=1.0000\) when rounded to four decimals. The unrounded curve is below 1. For the Pareto tail,

\int_{100}^{500}S_X(x)\,dx =\frac{0.08(100)}{2}\left[1-\left(\frac{100}{500}\right)^2\right] =4(0.96)=3.84.

Thus \(m(500)=43.84\) thousand dollars and \(G(500)=43.84/44=0.9964\). The exponential curve is higher at these finite limits, even though the Pareto tail ultimately declines more slowly. With equal total tail means, the models distribute expected loss differently across the tail; the Pareto model leaves more of its tail area beyond the higher limits.

Check. Each unrounded limited mean is no greater than the common unlimited mean of 44 thousand dollars. The example also shows why the phrase “heavier tail” alone does not determine which curve is higher at every finite limit.

Worked Example: Fitting a Pareto Exponent From Two Survival Values

A hypothetical severity study supports \(m(100)=30\) thousand dollars and \(S_X(100)=0.06\). At 200 thousand dollars, the supported survival estimate is \(S_X(200)=0.015\). Assume a Pareto tail beyond \(a=100\) thousand dollars. Estimate \(G(300)\). Report expected losses in thousands of dollars and the curve to four decimals.

Method. Fit the Pareto exponent from the ratio of the two survival values, then use that tail for both the limited-loss numerator and the unlimited-mean denominator.

Step 1: Solve for \(\alpha\). The Pareto assumption gives \(S_X(200)/S_X(100)=(100/200)^\alpha\). The observed ratio is \(0.015/0.06=0.25\), so \(0.25=(0.5)^\alpha\), which gives \(\alpha=2\).

Step 2: Find the unlimited mean. Since \(\alpha=2>1\), the tail mean is finite. Its contribution above 100 is

\int_{100}^\infty S_X(x)\,dx =\frac{0.06(100)}{2-1}=6.

Therefore \(E[X]=30+6=36\) thousand dollars.

Step 3: Find the limited mean at 300.

\int_{100}^{300}S_X(x)\,dx =\frac{0.06(100)}{2-1} \left[1-\left(\frac{100}{300}\right)^{1}\right] =6\left(\frac{2}{3}\right)=4.

Thus \(m(300)=30+4=34\) thousand dollars, and \(G(300)=34/36=0.9444\). As a check, the tail contribution through 300, which is 4, is less than the full tail contribution of 6; consequently the limited mean is below the unlimited mean.

Practical Checks When Choosing a Tail

A tail fitted beyond a threshold is an assumption about the part of the distribution with the least direct support. It should be treated as a modeling choice, not as an observed fact. In particular, check that the input quantities refer to the same severity basis and threshold, that the tail matches \(S_X(a)\), and that the model gives a finite mean if the exposure curve requires one.

Keep the anchor consistent. Use the survival probability at the threshold itself, not at a nearby limit, unless the tail has first been adjusted to the actual threshold.
Check the mean condition. An exponential tail has a finite tail integral for \(\beta>0\). A Pareto tail has a finite tail integral only when \(\alpha>1\).
Compare more than one plausible shape. Matching the survival and conditional mean excess at the threshold does not make two tail models equivalent at higher limits.
Interpret rounding carefully. A curve that rounds to 1.0000 may still be strictly below 1 before rounding.

The extrapolation is especially sensitive when the target limit is far above the threshold or when the assumed tail places substantial expected loss beyond the observed range. A useful calculation therefore reports the tail family and parameters alongside the resulting curve value. If the target-limit estimate changes materially under a reasonable alternative, the assumption is an important part of the result.

Key takeaway: To extend an exposure curve, add the assumed survival area above the supported threshold to \(m(a)\), and extend the same tail to infinity to obtain \(E[X]\). Verify that the tail matches the threshold data and has a finite mean before forming \(G(u)\).

Check Your Understanding

For each problem, assume the stated tail applies only above its threshold. Amounts are in thousands of dollars unless stated otherwise.

A portfolio has \(m(a)=25\), \(S_X(a)=0.10\), and an exponential tail with \(\beta=30\). Find \(E[X]\), \(m(a+30)\), and \(G(a+30)\). (Answer: \(E[X]=28\), \(m(a+30)=26.8964\), and \(G(a+30)=0.9606\).)
A portfolio has \(a=200\), \(m(a)=70\), \(S_X(a)=0.05\), and a Pareto tail with \(\alpha=2\). Find \(E[X]\), \(m(400)\), and \(G(400)\). (Answer: \(E[X]=80\), \(m(400)=75\), and \(G(400)=0.9375\).)
A Pareto tail is anchored at \(a=100\) with \(S_X(100)=0.09\), and \(S_X(300)=0.01\). If \(m(100)=45\), find \(\alpha\), \(E[X]\), \(m(600)\), and \(G(600)\). (Answer: \(\alpha=2\), \(E[X]=54\), \(m(600)=52.5\), and \(G(600)=0.9722\).)
At \(a=250\), a portfolio has \(m(a)=120\) and \(S_X(a)=0.04\). Compare exponential and Pareto tails at \(u=500\), where the exponential has \(\beta=125\) and the Pareto has \(\alpha=3\). Give both exposure-curve values to four decimals. (Answer: exponential \(G(500)=0.9946\); Pareto \(G(500)=0.9900\).)
A Pareto tail has \(a=100\), \(m(a)=60\), \(S_X(a)=0.02\), and \(\alpha=0.8\). Calculate \(m(500)\), then determine whether a finite-mean exposure curve can be calculated under this assumption. (Answer: \(m(500)=63.7973\); no finite-mean exposure curve is available because \(\alpha\le1\) makes the unlimited tail integral diverge.)

What You'll Learn