Why Diagnose an Exposure Curve?
An exposure curve translates a ground-up severity distribution into expected loss at different policy limits. Increased limits factors (ILFs) derived from that curve can influence pricing across a range of limits, so a mathematically plausible factor is not enough: the underlying data, adjustments, and model must also be coherent. A curve that decreases with the limit, or a tail factor driven by one observation, is a reason to investigate before relying on the indicated result.
The purpose of diagnosis is not to force every estimate to look smooth. Genuine sampling variation can create irregular empirical points, and a fitted curve can differ from empirical experience. Instead, the actuary checks whether the pattern is compatible with the exposure-curve definition, identifies which assumptions or data choices could produce it, and tests whether the conclusion is sensitive to those choices. This diagnostic approach is aligned with the CAS Exam 8 (Advanced Ratemaking) syllabus.
Structural Checks Before Interpreting a Pattern
For any fixed ground-up loss \(x\geq 0\), increasing the limit cannot reduce the capped amount: if \(u_2\geq u_1\), then \(\min(x,u_2)\geq\min(x,u_1)\). Taking expectations preserves this inequality. Thus \(G(u)\) is nondecreasing, and, when the denominator is positive, the ILF \(G(u)/G(u_0)\) is also nondecreasing as a function of \(u\). A decrease in estimated values is not a feature of the underlying limited expected loss; it signals that the estimates are not being constructed as comparable points from one coherent curve, or that estimation noise or processing has affected them.
This condition does not mean that successive increments must be equal. The amount added to expected loss as the limit rises depends on the probability of losses above the lower limit. In particular, sparse observations in the tail can make high-limit estimates uncertain even when all reported points are nondecreasing. Monotonicity is a necessary structural check, not a guarantee that the curve is accurate.
A useful diagnostic sequence is to verify the basic construction first, then examine sensitivity and model fit. In practice, check the loss basis and limit units, confirm that the same population and weighting logic are used at each limit, inspect the claims that contribute to each increment, and compare observed points with the fitted curve. A suspicious pattern is a prompt for that sequence, not by itself proof of a particular data error.
Confirm that the limits are ordered, the loss basis is stated, and each point estimates capped losses on a comparable basis.
Trace the capped amount, exposure weight, and aggregation for at least the base, an intermediate limit, and a high limit.
Identify which claims contribute to the tail increments and recalculate under reasonable claim-level or assumption-level sensitivities.
Review both levels and changes between limits, then determine whether disagreement is consistent with sampling variation, data adjustments, or model limitations.
Deriving a Useful Check and the Exponential Benchmark
For nonnegative \(X\), the capped loss can be written as the length of the interval from zero up to the smaller of \(X\) and \(u\). Equivalently, for each \(t\) between zero and \(u\), the loss contributes one unit of length exactly when \(X>t\). Integrating those indicators and taking expectations gives the survival-function representation below. This identity is useful for checking a parametric fit and seeing which portions of the severity tail affect the limited expected loss.
The expectation and integral can be interchanged here because the integrand is nonnegative. If \(X\) has an exponential distribution with mean \(\theta>0\), its survival function is \(S_X(t)=e^{-t/\theta}\). Substituting into the integral gives \(G(u)=\theta(1-e^{-u/\theta})\). This is a model-based result: it applies when the severity assumption is exponential with the stated mean, not merely because an empirical curve has a smooth shape.
Worked Examples: Finding and Interpreting Diagnostic Signals
Worked Example: Auditing a Nonmonotone Empirical Curve
A hypothetical portfolio has four ground-up claim amounts, in thousands of dollars: \(20, 40, 80,\) and \(120\). Each claim has equal weight. A summary reports limited expected losses of \(40\) at limit \(50\), \(38\) at limit \(100\), and \(65\) at limit \(200\), all in thousands. Identify the inconsistency and calculate the empirical curve points directly from the claims.
Method. For each limit, cap each claim at that limit, add the capped amounts, and divide by the same four claims. At limit \(50\), the capped amounts are \(20,40,50,50\). Their sum is \(160\), so \(G(50)=160/4=40\) thousand dollars.
At limit \(100\), the capped amounts are \(20,40,80,100\). Their sum is \(240\), so \(G(100)=240/4=60\) thousand dollars. At limit \(200\), no claim exceeds the limit, so the capped amounts are \(20,40,80,120\). Their sum is \(260\), and \(G(200)=260/4=65\) thousand dollars.
The recalculated points are therefore \(40,60,65\) thousand dollars, which are nondecreasing. The reported value \(38\) at limit \(100\) cannot be the equal-weight empirical capped mean from the stated claims. Its corresponding reported ILF relative to limit \(50\) would be \(38/40=0.9500\), also contradicting the structural check. The corrected empirical ILFs are \(1.0000\), \(60/40=1.5000\), and \(65/40=1.6250\).
Diagnostic conclusion. The calculation identifies a discrepancy, but not its exact cause. Check the source of the reported point, including its input claims, units, and aggregation. The common mistake is to accept a decreasing point as a genuine property of the exposure curve; the correction is to recalculate comparable capped amounts before attributing the pattern to model behavior.
Worked Example: Measuring the Influence of a Large Claim
A hypothetical dataset contains six ground-up claims, in thousands of dollars: \(10,20,30,40,100,\) and \(500\). Use equal claim weights. Calculate the empirical ILF at \(200\) relative to \(100\), then assess how much the result depends on the largest claim. Treat the deletion calculation only as a sensitivity test, not as a recommendation to remove a valid claim.
Full data calculation. At limit \(100\), the capped amounts are \(10,20,30,40,100,100\), totaling \(300\). Thus \(G(100)=300/6=50\) thousand dollars. At limit \(200\), they are \(10,20,30,40,100,200\), totaling \(400\), so \(G(200)=400/6=66.6667\) thousand dollars, rounded to four decimals. The ILF is \(66.6667/50=1.3333\), or exactly \(4/3\) before rounding.
Leave-one-out sensitivity. Removing the \(500\)-thousand-dollar claim leaves five claims. At limit \(100\), their capped amounts total \(10+20+30+40+100=200\), so the mean is \(200/5=40\) thousand dollars. At limit \(200\), the same five claims are all at or below the limit, so their total is also \(200\), and the mean is \(200/5=40\) thousand dollars. The leave-one-out ILF is \(40/40=1.0000\).
The change from \(1.3333\) to \(1.0000\) shows that the observed increment between limits \(100\) and \(200\) is entirely attributable to the largest claim in this small dataset. It does not establish that the claim is invalid or that the full-data result is wrong. It does show that the tail indication is highly sensitive and should be interpreted with the limited information in mind. The common mistake is to treat a single empirical tail factor as stable simply because its arithmetic is correct; the correction is to test and report its dependence on influential observations.
Worked Example: Comparing a Fitted Curve With Empirical Experience
A hypothetical severity model assumes \(X\) is exponential with mean \(\theta=80\) thousand dollars. An empirical estimate from a separate, consistently adjusted dataset gives \(G(80)=48\) and \(G(160)=62\) thousand dollars. Calculate the model values and ILFs, then interpret the discrepancy. Round monetary values to four decimals in thousands and factors to four decimals.
Model calculation. Using \(G(u)=80(1-e^{-u/80})\), at \(u=80\) the exponent is \(-1\). Thus \(G(80)=80(1-e^{-1})=80(1-0.3678794)=50.5696\) thousand dollars.
At \(u=160\), the exponent is \(-2\), and \(G(160)=80(1-e^{-2})=80(1-0.1353353)=69.1732\) thousand dollars. The model ILF at \(160\) relative to \(80\) is \((1-e^{-2})/(1-e^{-1})=1+e^{-1}=1.3679\).
Empirical comparison. The empirical ILF is \(62/48=1.2917\). At limit \(80\), the model is \(50.5696-48=2.5696\) thousand dollars above the empirical estimate. At limit \(160\), it is \(69.1732-62=7.1732\) thousand dollars above. The empirical increment from \(80\) to \(160\) is \(62-48=14\) thousand dollars; the model increment is \(69.1732-50.5696=18.6036\) thousand dollars.
The model therefore indicates a larger relative factor and a larger increment over this interval than the empirical curve. This pattern may reflect model-tail assumptions, sampling variability, or differences in the datasets or adjustments; the comparison alone cannot determine which explanation is correct. The common mistake is to compare only the ILFs and ignore the underlying limited expected losses. Reviewing levels and increments together shows where the disagreement arises.
Reading Common Patterns Without Overclaiming
Different error patterns point toward different checks. A sharp drop as the limit rises is a structural warning: retrace the calculations and confirm that the points are comparable. A curve that is nondecreasing but unusually jagged may reflect finite-sample variation, a handful of claims near particular limits, or inconsistent smoothing choices. A very large factor at a high limit may be mathematically possible, but if only a few observations contribute to the added layer, it is likely to be sensitive to those observations and to the tail assumptions.
A fitted curve that lies above empirical points at higher limits may be assigning more expected loss to the tail than the empirical experience indicates. A fitted curve below those points may be assigning less. Neither direction automatically proves that the fit is defective: the empirical estimate can be noisy, and the fitted model may use additional information. Consider whether claims have been developed and trended consistently, whether coverage and loss bases match, whether risk mix or weighting changes the population represented, and whether the fitted severity form is suitable for the observed range. These are possible explanations to investigate, not conclusions to infer from a plot alone.
For a high-limit estimate, it is especially important to distinguish what the observed claims support from what the model extrapolates. An empirical curve based on a finite sample may stop providing new information once the limit exceeds the largest observed loss; a parametric curve can continue to increase because its assumed distribution has a tail beyond the data. That difference is not inherently an error, but the additional modeled tail should not be described as directly observed experience.
| Observed pattern | Diagnostic meaning | Useful first check |
|---|---|---|
| A decrease in limited expected loss as limits rise | Inconsistent with one coherent capped-loss curve | Recalculate capped claims and verify comparable inputs and aggregation |
| A high-limit factor that changes greatly when one claim is removed | Tail indication is claim-sensitive | Show the claim-level contribution and a leave-one-out sensitivity |
| Empirical points are jagged but nondecreasing | May reflect sampling variation or uneven tail information | Inspect increments, claim counts, and the effect of any smoothing |
| Fitted and empirical curves diverge increasingly at high limits | Possible tail-model or data-basis mismatch | Compare levels and increments, then verify adjustments and model assumptions |
Common Mistakes and Exam Traps
- Calling any irregularity an impossible curve. A decrease violates the limited expected loss ordering, but jagged increases do not. Separate the structural violation from uncertainty in the magnitude of increments.
- Removing an influential claim without justification. A sensitivity test measures dependence; it does not establish that a valid claim should be excluded. Investigate data quality and present the sensitivity transparently.
- Treating a fitted curve as observed experience. A fitted tail beyond the largest observed loss depends on the assumed distribution. Label extrapolation as model-based and assess its sensitivity.
- Assuming a monotone curve is necessarily correct. Monotonicity is a necessary condition, not a test of data quality, calibration, or predictive accuracy.
- Diagnosing a cause from a symptom alone. A discrepancy can have several plausible sources. State what the calculation establishes and what further checks are needed.
Check Your Understanding
For each problem, show the indicated check before interpreting the result. Monetary amounts are in thousands of dollars unless stated otherwise.
- A curve reports \(G(25)=18\) and \(G(50)=17.5\). Is this compatible with a single limited expected loss curve? What should be checked first? (Answer: No; limited expected loss cannot decrease with limit, so recalculate the comparable capped amounts and verify the inputs and aggregation.)
- Three equally weighted claims are \(15,35,\) and \(75\). Calculate \(G(50)\), \(G(100)\), and the ILF at \(100\) relative to \(50\). (Answer: \(G(50)=33.3333\), \(G(100)=41.6667\), and the ILF is \(1.2500\).)
- Claims \(20,60,140\) have weights \(0.5,0.3,0.2\), respectively. Calculate the weighted limited expected losses at limits \(100\) and \(200\), then the ILF at \(200\) relative to \(100\). (Answer: \(G(100)=48\), \(G(200)=56\), and the ILF is \(1.1667\).)
- A hypothetical exponential severity model has mean \(100\). Calculate its limited expected losses at limits \(100\) and \(200\), and its ILF at \(200\) relative to \(100\). (Answer: \(G(100)=63.2121\), \(G(200)=86.4665\), and the ILF is \(1.3679\).)
- Four equal-weight claims are \(10,20,40,\) and \(200\). Calculate the empirical ILF at \(200\) relative to \(50\), then calculate the same factor after omitting the largest claim as a sensitivity test. What does the comparison indicate? (Answer: With all claims, \(G(50)=30\) and \(G(200)=67.5\), so the ILF is \(2.2500\). Without the largest claim, both limited expected losses are \(23.3333\), so the ILF is \(1.0000\); the tail factor is highly sensitive to that claim.)