Secondary Uncertainty in Event Losses

Why Event Losses Need More Than an Event Count

A catastrophe model may represent a possible event with an annual occurrence rate and a loss amount. The occurrence rate describes how often events of that type happen over many years. But even if an event occurs, its loss may vary: the event’s path, intensity, affected locations, and resulting damage can differ. A model that assigns a probability distribution to loss conditional on an event captures this secondary uncertainty.

This is distinct from occurrence uncertainty, which concerns the number and timing of events. The distinction matters when calculating exceedance curves. Replacing a distribution of event losses with its average may preserve expected loss, but it can remove the chance of a severe conditional outcome and materially change the tail.

Definition: Occurrence uncertainty is uncertainty in whether, when, or how many events occur. Secondary uncertainty is uncertainty in the loss conditional on an event occurring. For event type $j$, write its annual event count as $N_j$, its annual frequency parameter as $\lambda_j$, and the loss from one event as $X_j$, with conditional distribution $F_{X_j}$. Under a Poisson event model, $N_j\sim\operatorname{Poisson}(\lambda_j)$; conditional event losses are independent draws from $F_{X_j}$, independent of $N_j$.

Here, $X_j$ is an event loss in the modeled portfolio and financial basis. It may already reflect the applicable exposure and policy terms. The conditional distribution describes possible losses after an event occurs, not the probability that the event occurs. Unless stated otherwise, examples below use hypothetical losses in millions of dollars, no discounting, and no mortality component.

From Conditional Losses to Expected Annual Loss

For a Poisson event type with frequency $\lambda$ and conditional event loss $X$, the annual aggregate loss is $S=\sum_{k=1}^{N}X_k$, where $N\sim\operatorname{Poisson}(\lambda)$. The expected annual loss is

E[S]=\lambda E[X].

The frequency is measured in events per year and $E[X]$ in dollars per event, so their product is dollars per year. For multiple independent event types, expected annual losses add: $E[S]=\sum_j\lambda_j E[X_j]$. This expected value uses the entire conditional loss distribution through its mean; it does not describe the distribution’s tail.

Worked Example: Expected Loss and OEP With a Random Event Severity

A hypothetical storm type has a Poisson annual frequency of $0.20$. Conditional on one storm occurring, its loss is \$1 million with probability $0.20$, \$4 million with probability $0.50$, and \$9 million with probability $0.30$. Find expected annual loss and the probability that at least one event loss is strictly greater than \$4 million in a year. Assume independent event losses conditional on the count. No discounting or mortality applies.

Step 1: Find the conditional mean loss. The conditional distribution is discrete, so multiply each possible loss by its probability and add:

E[X]=0.20(1)+0.50(4)+0.30(9) =0.20+2.00+2.70 =4.90\text{ million dollars per event}.

Step 2: Calculate expected annual loss.

E[S]=\lambda E[X]=0.20(4.90)=0.98\text{ million dollars per year}.

Step 3: Calculate the OEP above \$4 million. A single event loss is strictly above \$4 million only when it is \$9 million, so the conditional exceedance probability is $q=0.30$. The annual rate of events whose losses exceed the threshold is $0.20(0.30)=0.060$. For a Poisson count, the probability of at least one such event is $1-e^{-0.060}=0.058235\ldots$, or 0.0582 to four decimals.

The method keeps the conditional severity probabilities in the tail calculation. A common mistake is to calculate the conditional mean as \$3.10 million; the weighted sum is \$4.90 million, as shown above. Another is to treat an event loss equal to \$4 million as exceeding a threshold defined as strictly greater than \$4 million.

Deriving the OEP Formula

For one Poisson event type, let $q=P(X>x)$, where $x$ is a loss threshold. Given $N=n$ events, the probability that none of their losses exceeds $x$ is $(1-q)^n$. Averaging over the Poisson count gives the probability of no exceedance in the year:

P(\text{no event loss}>x) =\sum_{n=0}^{\infty}P(N=n)(1-q)^n =\sum_{n=0}^{\infty}e^{-\lambda}\frac{\lambda^n}{n!}(1-q)^n =e^{-\lambda}\exp\{\lambda(1-q)\} =e^{-\lambda q}.

Taking the complement yields the occurrence exceedance probability (OEP), the probability that at least one event loss exceeds $x$:

Result: If event counts are Poisson with frequency $\lambda$, and event losses are independent marks with conditional distribution $F_X$, then $$ \operatorname{OEP}(x)=P\left(\max_{1\leq k\leq N}X_k>x\right) =1-\exp\{-\lambda P(X>x)\}. $$ For independent Poisson event types, replace $\lambda P(X>x)$ with $\sum_j\lambda_jP(X_j>x)$.

The formula uses the conditional exceedance probability $P(X>x)$, which is $1-F_X(x)$ when $F_X(x)=P(X\leq x)$. It applies to Poisson event counts with independently assigned event losses. A deterministic annual event catalog, dependent event counts, or a non-Poisson frequency model may require a different calculation. For a discrete loss distribution, pay careful attention to whether the requested threshold is strict ($X>x$) or inclusive ($X\geq x$).

Worked Example: How Secondary Uncertainty Changes an Exceedance Curve

A hypothetical hail event type has Poisson frequency $\lambda=0.10$ per year. Conditional on an event, loss is \$2 million with probability $0.50$ and \$8 million with probability $0.50$. Find the conditional mean and variance, expected annual loss, aggregate annual variance, and OEP above \$6 million. Compare the OEP with the result of replacing every event loss by its conditional mean. The loss distribution is independent across events; no discounting or mortality applies.

Step 1: Calculate conditional moments.

E[X]=0.50(2)+0.50(8)=5\text{ million dollars}, $$ $$ E[X^2]=0.50(2^2)+0.50(8^2)=0.50(4)+0.50(64)=34\text{ million dollars}^2, $$ $$ \operatorname{Var}(X)=E[X^2]-E[X]^2=34-25=9\text{ million dollars}^2.

Step 2: Calculate expected annual loss. The frequency is $0.10$ events per year, so $E[S]=0.10(5)=0.50$ million dollars per year.

Step 3: Calculate aggregate annual variance. For a compound Poisson sum, $\operatorname{Var}(S)=\lambda E[X^2]$, so $\operatorname{Var}(S)=0.10(34)=3.40$ million dollars squared. The split is $0.10(9)=0.90$ from conditional severity variation and $0.10(5^2)=2.50$ from event-count variation; together these equal $3.40$.

Step 4: Calculate the OEP above \$6 million. Only the \$8 million outcome exceeds the threshold, so $P(X>6)=0.50$. The exceedance event rate is $0.10(0.50)=0.050$, giving

\operatorname{OEP}(6)=1-e^{-0.050}=0.0487706\ldots=0.0488.

If each event were instead assigned the mean loss of \$5 million, no single event would exceed \$6 million, producing an OEP of zero. That substitution preserves expected annual loss but erases the conditional chance of an \$8 million event. It is therefore unsuitable for this tail calculation.

Frequency and Severity Contributions to Aggregate Variance

For a compound Poisson aggregate $S=\sum_{k=1}^N X_k$, the variance formula can be justified by conditioning on the event count. Given $N$, the sum has conditional mean $NE[X]$ and variance $N\operatorname{Var}(X)$. The law of total variance then gives:

\begin{aligned} \operatorname{Var}(S) &=E[\operatorname{Var}(S\mid N)] +\operatorname{Var}(E[S\mid N])\\ &=E[N]\operatorname{Var}(X) +\operatorname{Var}(N)(E[X])^2\\ &=\lambda\operatorname{Var}(X)+\lambda(E[X])^2\\ &=\lambda E[X^2]. \end{aligned}

The first term reflects variation in loss conditional on the event count; the second reflects variation in the count. Both contribute to aggregate uncertainty under this model. This decomposition is a variance identity, not a claim that every observed year’s loss can be uniquely attributed to one source.

Worked Example: Combining Two Event Types

Two hypothetical event types occur independently, with Poisson annual frequencies of $0.05$ each. For Type A, conditional loss is \$5 million with probability $0.40$ and \$15 million with probability $0.60$. For Type B, conditional loss is \$8 million with probability $0.75$ and \$20 million with probability $0.25$. Find the OEP above \$10 million and total expected annual loss. Conditional losses are independent between events. No discounting or mortality applies.

Step 1: Find each type’s conditional mean. For Type A:

E[X_A]=0.40(5)+0.60(15)=2+9=11\text{ million dollars}.

For Type B:

E[X_B]=0.75(8)+0.25(20)=6+5=11\text{ million dollars}.

Step 2: Calculate total expected annual loss.

0.05(11)+0.05(11)=0.55+0.55 =1.10\text{ million dollars per year}.

Step 3: Calculate the event exceedance rate above \$10 million. Type A exceeds \$10 million only in its \$15 million outcome, with conditional probability $0.60$. Type B exceeds \$10 million only in its \$20 million outcome, with conditional probability $0.25$. Thus the combined exceedance rate is

$$ 0.05(0.60)+0.05(0.25)=0.0300+0.0125=0.0425. $$

The OEP is $1-e^{-0.0425}=0.041610\ldots$, or 0.0416 to four decimals. The event types’ means happen to be equal, but their tail probabilities differ. A common arithmetic error is to understate the total expected annual loss: each type contributes $0.55$ million per year, so the total is $1.10$, not $0.82$ million.

How Secondary Uncertainty Changes Exceedance Curves

An event loss table may contain a single loss for each event, while a conditional model can represent a range of losses for that event. If the table’s single value is the conditional mean, it can support an expected-loss calculation under suitable assumptions, but it does not retain the full conditional tail. Consequently, an OEP curve built from mean event losses can misstate the probability of high event losses.

The effect depends on the threshold. In the hail example, replacing the conditional distribution with its mean removed the chance of exceeding \$6 million. At a different threshold, the replacement might create an exceedance that the original distribution did not have. There is no general rule that secondary uncertainty raises every point on an exceedance curve; the relevant comparison is between the actual conditional tail probability and the tail implied by the simplified loss representation.

OEP concerns the largest individual event loss in a year. It is not the same as the annual aggregate exceedance probability (AEP), which concerns total annual loss. Multiple individually modest events can produce a large annual aggregate, so an OEP calculation should not be used as an AEP calculation. When annual aggregate loss is needed, the event counts and conditional severity distributions must be combined to obtain the distribution of $S$.

Common exam trap: Do not multiply the event frequency by a conditional loss amount when calculating OEP. For a threshold $x$, first find $P(X>x)$; then calculate $1-\exp\{-\lambda P(X>x)\}$. Multiplying $\lambda$ by $E[X]$ gives expected annual loss, not an exceedance probability.

Check Your Understanding

For each hypothetical model, use the stated Poisson assumptions. Report probabilities to four decimals and losses in millions of dollars.

A storm type has annual frequency $0.12$; conditional loss is \$3 million with probability $0.75$ and \$11 million with probability $0.25$. Find expected annual loss and OEP strictly above \$5 million. (Answer: expected annual loss = \$0.60 million per year; OEP = 0.0296.)
An event type has annual frequency $0.08$, with conditional loss \$1 million with probability $0.60$ and \$6 million with probability $0.40$. Find expected annual loss and OEP above \$4 million. What OEP would result from substituting the conditional mean loss for every event? (Answer: expected annual loss = \$0.24 million per year; distribution-based OEP = 0.0315; mean-substitution OEP = 0.)
Two independent event types have annual frequencies $0.03$ and $0.02$. Type A has losses \$4 million and \$12 million, each with probability $0.50$; Type B has losses \$7 million and \$15 million, with probabilities $0.25$ and $0.75$. Find total expected annual loss and OEP strictly above \$10 million. (Answer: expected annual loss = \$0.50 million per year; OEP = 0.0296.)
A single event type has annual frequency $0.20$, with conditional loss \$2 million or \$10 million, each with probability $0.50$. Find the expected annual aggregate loss and its variance, separating the count and conditional-severity contributions to variance. (Answer: expected annual loss = \$1.20 million per year; aggregate variance = 10.40 million dollars squared, comprising 7.20 from count variation and 3.20 from conditional severity variation.)
Two independent event types have annual frequencies $0.04$ and $0.06$. Type A has loss \$5 million with probability $0.80$ and \$20 million with probability $0.20$; Type B has loss \$9 million or \$14 million, each with probability $0.50$. Find total expected annual loss and OEP strictly above \$10 million. (Answer: expected annual loss = \$1.01 million per year; OEP = 0.0373.)

What You'll Learn