Why Large Losses Require Special Attention
An unpaid claim estimate can be uncertain because many claims remain unreported or immature. A different source of uncertainty arises when a small number of claims may account for a large share of the liability. A single severe injury, complex liability dispute, or event affecting many insureds can shift the estimate substantially even when the rest of the portfolio develops as expected.
The expected unpaid amount alone does not describe this risk. Two portfolios can have the same expected liability but very different possible outcomes: one may produce many moderate claims, while another may usually produce little and occasionally produce a very large amount. Catastrophe accumulation adds another issue: claims arising from one event are often dependent, so treating them as unrelated can misstate the range of possible aggregate outcomes.
These examples use stylized probability models to show the calculations. They are not estimates for any actual insurer or event. In practice, selections also depend on claim facts, policy terms, valuation maturity, coverage disputes, inflation, and the quality of the available data.
Individual Large Losses: Severity and Retention
For a claim with unpaid amount \(X\), a high possible severity can make the liability uncertain even if the claim is already reported. A retention changes how that uncertainty is allocated between the insurer and its reinsurer. Under a simple per-claim excess-of-loss arrangement with retention \(R\) and no upper recovery limit, the insurer’s net payment per claim is \(\min(X,R)\), and the ceded recovery is \(\max(X-R,0)\). This assumes the claim is covered, the reinsurance responds as stated, and the recovery is collectible.
The identity holds in both cases. If \(X\le R\), the two terms on the right are \(X\) and \(0\). If \(X>R\), they are \(R\) and \(X-R\), which sum to \(X\). Taking expectations therefore separates the expected gross amount into expected net and expected ceded amounts, provided the same coverage basis is used throughout.
Worked Example: A Reported Claim With Uncertain Severity
A hypothetical liability claim is unpaid at the valuation date. Its ultimate covered payment \(X\), in thousands of dollars, is modeled as 2,000 with probability 0.5000, 6,000 with probability 0.3000, and 14,000 with probability 0.2000. The insurer has a 5,000 per-claim retention with unlimited excess recovery. Use this discrete severity distribution; no interest discounting or mortality assumption is relevant to the claim-payment model.
Method: calculate the expected gross payment and then apply the retention to each possible outcome. The gross expected payment is
The insurer’s net payment is 2,000 when \(X=2{,}000\), and 5,000 in each of the other two outcomes. Thus
The ceded amounts are 0, 1,000, and 9,000, respectively. Their expected value is
The check is \(3{,}500+2{,}100=5{,}600\), equal to the gross expected payment. The retention also reduces variability in the insurer’s net payment. Its second moment is \(0.5000(2{,}000^2)+0.3000(5{,}000^2)+0.2000(5{,}000^2)=14{,}500{,}000\) in squared thousands. Subtracting the squared mean gives a net variance of \(14{,}500{,}000-3{,}500^2=2{,}250{,}000\), or a standard deviation of 1,500 thousand dollars.
Check and common mistake: The gross mean of 5,600 is not the insurer’s net liability under this reinsurance assumption. Applying the retention to the mean would give \(\min(5{,}600,5{,}000)=5{,}000\), which is incorrect: the retention must be applied to each possible claim amount before averaging. The correct net expected amount is 3,500 thousand dollars.
Catastrophe Accumulation and Dependence
A catastrophe can affect several insureds, locations, or policies at once. Consequently, the total unpaid amount can depend not just on individual claim probabilities but also on whether claims move together. If claims are independent, simultaneous large losses may be relatively unlikely. If one event can cause multiple losses together, the aggregate distribution can have a much heavier upper tail than an independence model suggests.
A useful starting point is to represent the total unpaid amount as the sum of individual claim amounts. For independent claims, variances add. For dependent claims, covariance terms must also be included. Positive dependence increases aggregate variance because claims tend to be high or low together.
This follows by expanding the variance of a sum: each individual variance appears once, and each pairwise covariance appears twice. Under independence, all the covariances are zero. A shared catastrophe trigger generally makes the affected claim amounts positively dependent, so omitting the covariance terms can understate aggregate uncertainty.
Worked Example: Independent Claims Versus a Shared Catastrophe
Consider four hypothetical buildings. Each has a 0.1000 probability of generating an unpaid covered loss of 1,000 thousand dollars and a 0.9000 probability of no loss. Compare two models: in the first, the four losses occur independently; in the second, one shared event causes all four losses with probability 0.1000, and none with probability 0.9000. The expected gross loss is the same in both models. A hypothetical aggregate excess cover has a 2,000 thousand dollar aggregate retention across these four losses and pays the amount above that retention without an upper limit. Frequencies are specified by these Bernoulli models; no separate frequency distribution, interest assumption, or mortality assumption applies.
Method: calculate the aggregate distribution under each dependence assumption, then apply the aggregate retention. Under independence, let \(K\) be the number of buildings with a loss. Then \(K\sim\operatorname{Binomial}(4,0.1000)\). The probabilities needed are
The gross total is \(1{,}000K\). Its expected value is \(4(0.1000)(1{,}000)=400\) thousand dollars. After the aggregate retention, the insurer’s net amounts for \(K=0\), \(K=1\), and \(K\ge2\) are 0, 1,000, and 2,000, respectively. Therefore the expected net amount is \(0.2916(1{,}000)+0.0523(2{,}000)=396.2\) thousand dollars. The expected recovery is \(400-396.2=3.8\) thousand dollars.
Under the shared-event model, the gross total is 4,000 thousand dollars with probability 0.1000 and zero otherwise. Its expected value is \(0.1000(4{,}000)=400\) thousand dollars, again. The insurer retains 2,000 thousand dollars when the event occurs and zero otherwise, so its expected net amount is \(0.1000(2{,}000)=200\) thousand dollars. The expected recovery is also 200 thousand dollars.
The gross variances reveal the difference in uncertainty. Under independence, each building has variance \(0.1(0.9)(1{,}000^2)=90{,}000\) in squared thousands, so the total variance is \(4(90{,}000)=360{,}000\). Under the shared event, the total is 0 or 4,000, giving variance \(0.1(0.9)(4{,}000^2)=1{,}440{,}000\). Thus the common-event model has four times the gross variance, despite having the same expected gross loss.
Check and common mistake: Both gross means equal 400 thousand dollars, but the expected net amounts differ because the aggregate retention is nonlinear: it is applied to the aggregate outcome, not to the expected aggregate. Treating the four claims as independent in the shared-event model would miss the accumulation and materially misstate the loss distribution.
Frequency, Severity, and Retained Catastrophe Losses
When the number of events is uncertain as well as the amount per event, a compound frequency-severity model provides a compact way to represent aggregate unpaid payments. Let \(N\) be the number of events, and let \(X_1,\ldots,X_N\) be independent and identically distributed event severities, independent of \(N\). If \(N\) has a Poisson distribution with mean \(\lambda\), the aggregate gross amount is \(S=\sum_{j=1}^{N}X_j\), with \(S=0\) if \(N=0\). Under these assumptions, \(E[S]=\lambda E[X]\) and \(\operatorname{Var}(S)=\lambda E[X^2]\).
For the variance formula, condition first on \(N\). The conditional mean is \(N E[X]\), and the conditional variance is \(N\operatorname{Var}(X)\). Applying the law of total variance gives
because a Poisson variable has both mean and variance \(\lambda\), and \(E[X^2]=\operatorname{Var}(X)+(E[X])^2\).
Worked Example: A Compound Poisson Catastrophe Layer
For a hypothetical one-year cohort of unpaid catastrophe events, let \(N\sim\operatorname{Poisson}(0.2000)\). Given an event, its covered severity \(X\), in thousands of dollars, is 1,000 with probability 0.8000 and 9,000 with probability 0.2000. Event severities are independent of one another and of \(N\). A per-event retention is 3,000 thousand dollars, with unlimited excess recovery. Use no discounting; mortality is not part of this event-payment model.
Method: use compound Poisson moments for gross losses and apply the retention to each event severity for net losses. First calculate the severity moments:
The expected gross unpaid amount is \(0.2000(2{,}600)=520\) thousand dollars. The gross aggregate variance is \(0.2000(17{,}000{,}000)=3{,}400{,}000\) in squared thousands, so its standard deviation is \(\sqrt{3{,}400{,}000}=1{,}843.9\) thousand dollars, rounded to one decimal.
After the 3,000 retention, the net severity \(Y=\min(X,3{,}000)\) is 1,000 with probability 0.8000 and 3,000 with probability 0.2000. Thus \(E[Y]=0.8000(1{,}000)+0.2000(3{,}000)=1{,}400\), and \(E[Y^2]=0.8000(1{,}000^2)+0.2000(3{,}000^2)=2{,}600{,}000\). The expected net aggregate is \(0.2000(1{,}400)=280\) thousand dollars. Its variance is \(0.2000(2{,}600{,}000)=520{,}000\) in squared thousands, giving a standard deviation of \(\sqrt{520{,}000}=721.1\) thousand dollars, rounded to one decimal. The expected ceded amount is \(520-280=240\) thousand dollars.
Check and common mistake: The gross expected amount reconciles to expected net plus expected ceded: \(280+240=520\). Applying the retention to the expected severity would instead give \(\min(2{,}600,3{,}000)=2{,}600\), which incorrectly implies no expected recovery. The retention must be applied to each event severity before taking its expectation.
Interpreting the Estimate and Avoiding Traps
Large-loss and catastrophe uncertainty matters because a central estimate can be sensitive to a small number of claim outcomes, event assumptions, and coverage interpretations. A low expected frequency does not imply a negligible liability: a rare event can still have a material expected cost when its severity is high. Similarly, a retention may reduce the insurer’s net average and variability without eliminating uncertainty, especially when several claims can arise from one event or when reinsurance terms are disputed.
- Keep gross and net bases distinct. Estimate covered gross payments and recoveries on consistent terms. A recovery should not be subtracted twice, and a stated retention should not be confused with a policy deductible or an occurrence limit.
- Respect the order of operations. For a nonlinear layer, apply the policy or reinsurance function to each modeled outcome, then calculate the mean or other summary. In general, \(E[\min(X,R)]\) is not \(\min(E[X],R)\).
- Model dependence where accumulation is plausible. Separate location-level claim probabilities do not justify an independence assumption if a common event can produce multiple claims.
- Separate modeled process variation from model and parameter uncertainty. A precise calculation under an assumed distribution does not establish that the distribution, frequency, trend, or coverage interpretation is correct.
- Use claim information carefully. A known large claim may have unique facts, limits, and settlement uncertainty. Treating it as an ordinary claim can obscure its influence on the unpaid estimate.
These calculations are aids to reserve reasoning, not substitutes for selecting a supportable liability estimate. The estimate should reflect the valuation date, known claim information, applicable coverage and reinsurance, and the uncertainty that remains in future payments. Recognizing tail severity and accumulation is aligned with the CAS Exam 5 (Basic Ratemaking and Estimated Claim Liabilities) syllabus.
Check Your Understanding
For each problem, calculate on the stated basis and distinguish expected amount from variability.
- A hypothetical unpaid claim has severity 1,000 thousand dollars with probability 0.7000 and 7,000 thousand dollars with probability 0.3000. The insurer has a 4,000 per-claim retention and unlimited excess recovery. What are the expected gross, net, and ceded amounts? (Answer: gross 2,800 thousand dollars; net 1,900 thousand dollars; ceded 900 thousand dollars.)
- A single unpaid claim is 10,000 thousand dollars with probability 0.0800 and zero otherwise. The per-claim retention is 3,000 thousand dollars. What is the expected net liability and its standard deviation? (Answer: expected net 240 thousand dollars; standard deviation 813.9 thousand dollars.)
- Three independent locations each have probability 0.2000 of an unpaid 2,000-thousand-dollar loss and probability 0.8000 of no loss. An aggregate retention of 3,000 thousand dollars applies across the three losses. What are the expected gross and net amounts? (Answer: gross 1,200 thousand dollars; net 1,080 thousand dollars.)
- Five buildings each have a 0.1000 probability of an unpaid 1,000-thousand-dollar loss. Compare independent losses with a shared-event model in which all five losses occur together with probability 0.1000. What is the gross expected amount and gross variance under each model? (Answer: expected amount is 500 thousand dollars under both models; variance is 450,000 under independence and 2,250,000 under the shared-event model, in squared thousands.)