Why Event Rates Matter
A catastrophe catalog may contain historical occurrences, simulated events, or a combination of modeled event scenarios. A list of events alone does not say how frequently those events are expected to occur. To use the catalog in annual risk analysis, the event counts or weights must be interpreted relative to time. The central distinction is between an event rate, which is an expected number of occurrences per year, and an annual probability, which is the chance of at least one occurrence during a year.
That distinction matters when translating modeled event losses into annual exceedance probabilities. A rate of 0.20 events per year is not automatically a 20% annual probability. Under a Poisson occurrence model, the corresponding probability of at least one event is \(1-e^{-0.20}\), or about 18.13%. The conversion depends on the assumed occurrence process; a rate by itself does not specify a probability.
This tutorial is aligned with the CAS Exam 9 (Risk Management for Actuaries) syllabus. It focuses on annualizing catalog counts and weights, and on the assumptions required to convert an event rate into an annual probability.
Counts Over Time and Annual Event Rates
The event definition and the time denominator must match. For example, a catalog of qualifying storms over 25 calendar years has 25 years of exposure only if those years are all covered under the same event definition and observation scope. If the data cover different locations or periods with different exposure, the relevant exposure time may need to be assembled from those periods rather than assumed to be a simple calendar span.
A catalog can also divide events into categories, such as moderate and severe events. If the categories are mutually exclusive and together include every event being counted, their counts and rates add. If categories overlap, adding their rates can count some occurrences more than once.
Worked Example: Annualizing Historical Occurrence Counts
A hypothetical regional catalog records 24 qualifying wind events over 40 complete years. Of these, 16 are classified as moderate and 8 as severe. Estimate the overall rate and the severe-event rate, then calculate the probability of at least one severe event in a year under a Poisson model.
Method and assumptions. Divide each count by the same 40-year observation period. For the probability calculation, assume severe events follow a Poisson process with a constant rate from year to year.
Overall rate. $$ \hat{\lambda}_{\text{all}}=\frac{24}{40}=0.600\text{ events per year}. $$ The category check gives \(16+8=24\) events, so the category counts reconcile to the total.
Severe-event rate. $$ \hat{\lambda}_{\text{severe}}=\frac{8}{40}=0.200\text{ events per year}. $$
For a Poisson rate of 0.200 per year, the chance of at least one severe event is $$ P(N\geq 1)=1-e^{-0.200}=1-0.8187=0.1813. $$ Thus the estimated annual probability is 0.1813, or 18.13% to two decimal places. The result is less than the rate of 0.200 because a rate is an expected count, while this probability only asks whether the count is zero or positive.
Check. The severe rate is one third of the overall rate, \(0.200/0.600=1/3\), consistent with 8 of the 24 catalog events being severe. The common mistake this avoids is reporting 0.200 as the annual probability without stating an occurrence model.
How Catalog Weights Represent Annual Frequency
A stochastic catalog may assign a weight to each event record. The meaning of that weight must be established before doing arithmetic. One common convention is that \(w_j\) is the annual occurrence rate represented by event record \(j\), in events per year. Under that convention, the total rate represented by a set of records is the sum of their weights.
The weights in these formulas are annual rates, not probabilities that sum to one. Another catalog may instead give normalized scenario probabilities, which sum to one and describe the mix of event types conditional on an event occurring. Those probabilities do not, by themselves, say how often events occur. If the total event rate is \(\lambda\), and scenario \(j\) has conditional probability \(p_j\), its annual rate contribution is \(w_j=\lambda p_j\).
Some simulation outputs use weights with yet another meaning, such as sampling weights or contributions scaled to a simulation horizon. A weight should only be summed as an annual rate when its definition supports that interpretation. Check the units, normalization, and time basis: do the weights sum to an annual frequency, to one, or to some other quantity?
Worked Example: Threshold Rates From Annual-Rate Weights
A hypothetical stochastic catalog has four event records. Each weight is an annual event rate, and the losses are in millions of dollars. Find the rate of events with loss greater than $1 million and the probability of at least one such event in a year under a Poisson assumption. Also calculate the catalog-weighted expected annual loss.
| Record | Annual rate \(w_j\) | Event loss \(L_j\), millions of dollars |
|---|---|---|
| A | 0.08 | 0.5 |
| B | 0.12 | 1.2 |
| C | 0.05 | 2.0 |
| D | 0.15 | 3.5 |
Method. Sum the weights of records whose losses exceed the threshold. Then convert that qualifying rate to an annual probability, assuming qualifying events follow a Poisson process. For expected annual loss, multiply each record's annual rate by its event loss and sum the contributions.
Record A does not qualify because \(0.5\) is not greater than \(1\). Records B, C, and D do qualify. Therefore, $$ \lambda_{1} = 0.12+0.05+0.15=0.32\text{ events per year}. $$ The annual probability of at least one event with loss greater than $1 million is $$ 1-e^{-0.32}=1-0.7261=0.2739. $$ The probability is 0.2739, or 27.39% to two decimal places.
For the expected annual loss, the four contributions are $$ 0.08(0.5)=0.040,\quad 0.12(1.2)=0.144,\quad 0.05(2.0)=0.100,\quad 0.15(3.5)=0.525 $$ in millions of dollars per year. Their sum is $$ 0.040+0.144+0.100+0.525=0.809\text{ million dollars per year}. $$ Thus the catalog-weighted expected annual loss is $809,000. The calculation is a weighted sum of event losses, not the probability of an annual aggregate loss exceeding a threshold.
Check and trap avoided. The weights total \(0.08+0.12+0.05+0.15=0.40\) events per year, while the qualifying subset totals 0.32 per year. A common error is to treat the total weight 0.40 as a probability or to include record A because its loss is close to the threshold. The weights are rates, and the stated strict condition is \(L_j>1\).
From Conditional Scenario Probabilities to Rates
A catalog may first describe which scenario occurs given that an event occurs. In that case, let \(p_j\) be the conditional probability of scenario \(j\), with \(\sum_j p_j=1\). If the total event rate is \(\lambda\), then scenario \(j\)'s annual rate is \(\lambda p_j\). For any collection of qualifying scenarios, their rates add to the total rate multiplied by their combined conditional probability.
Worked Example: Converting Scenario Probabilities Into Annual Rates
A hypothetical event model has a total occurrence rate of 0.40 events per year. Conditional on an event occurring, the modeled scenario probabilities and losses are: 0.50 for a $0.8 million loss, 0.30 for a $2.0 million loss, and 0.20 for a $5.0 million loss. Find each scenario's annual rate, the rate and annual probability of an event loss greater than $1.5 million, and the expected annual loss.
Method and assumptions. Multiply each conditional probability by the total annual event rate. Assume the resulting event occurrences follow a Poisson process, with the scenario assignment following the stated probabilities.
The scenario rates are $$ w_1=0.40(0.50)=0.20,\quad w_2=0.40(0.30)=0.12,\quad w_3=0.40(0.20)=0.08 $$ events per year. Their sum is \(0.20+0.12+0.08=0.40\), which checks against the given total rate.
Losses greater than $1.5 million occur in scenarios 2 and 3. Their combined rate is $$ \lambda_{1.5}=0.12+0.08=0.20\text{ events per year}. $$ The probability of at least one such event in a year is $$ 1-e^{-0.20}=1-0.8187=0.1813, $$ or 18.13% to two decimal places.
The expected loss conditional on an event is $$ 0.50(0.8)+0.30(2.0)+0.20(5.0) =0.40+0.60+1.00=2.00\text{ million dollars}. $$ Multiplying by the total rate gives expected annual loss: $$ 0.40(2.00)=0.80\text{ million dollars per year}. $$ This agrees with the direct weighted-rate calculation: \(0.20(0.8)+0.12(2.0)+0.08(5.0)=0.16+0.24+0.40=0.80\) million dollars per year.
Check and trap avoided. The probabilities \(0.50, 0.30, 0.20\) sum to one because they describe the scenario mix conditional on an event. They are not annual rates. Treating 0.30 and 0.20 as events per year would overstate the qualifying rate; each must first be multiplied by the total rate of 0.40.
Why the Poisson Conversion Works
The Poisson model assumes that event counts over a time interval have a Poisson distribution with mean equal to the rate times the interval length. For a constant rate \(\lambda\) per year, the count \(N\) in one year satisfies \(N\sim\operatorname{Poisson}(\lambda)\). This is a modeling assumption, often useful for random event occurrences; it is not implied merely by having a catalog.
Under the Poisson model, \(P(N=k)=e^{-\lambda}\lambda^k/k!\) for \(k=0,1,2,\ldots\).
For \(k=0\), \(P(N=0)=e^{-\lambda}\lambda^0/0!=e^{-\lambda}\).
The event occurs at least once exactly when the count is not zero, so \(P(N\geq1)=1-P(N=0)=1-e^{-\lambda}\).
For a loss threshold \(x\), apply the same reasoning to the subset of events with \(L>x\). If those qualifying occurrences form a Poisson process with rate \(\lambda_x\), then the probability of at least one threshold-exceeding event in a year is \(1-e^{-\lambda_x}\). With annual-rate catalog weights, \(\lambda_x\) is the sum of weights for records satisfying the threshold. This is an event-level exceedance probability: it is not, without additional modeling, the probability that the sum of all event losses during the year exceeds \(x\).
The conversion does not apply automatically to every catalog. A fixed historical sequence is not a Poisson process simply because its count can be divided by years. Dependence among occurrences, a changing rate over time, or a model with a fixed annual event count can make the Poisson probability inappropriate. In such cases, the rate remains a rate estimate, but an annual probability should be calculated from the specified occurrence model.
Common Interpretation Errors
- Calling a rate a probability. A rate can exceed 1 event per year, whereas a probability cannot exceed 1. Convert a rate to a probability only under a stated occurrence model.
- Using the wrong time base. A count of 12 over 6 years gives 2 events per year, not 12 events per year. Confirm the covered exposure period.
- Assuming every weight is an annual rate. A normalized scenario probability, a sampling weight, and an annual-rate weight have different meanings. Read the weight definition and check its units and sum.
- Adding overlapping categories. Summing category rates is valid for disjoint categories. If an event can appear in more than one category, count the union without double counting.
- Confusing an event exceedance with an annual aggregate exceedance. The Poisson conversion using \(\lambda_x\) gives the probability of at least one event whose individual loss exceeds \(x\). It does not account for several smaller losses accumulating past \(x\).
- Ignoring the threshold condition. “Greater than \(x\)” excludes losses equal to \(x\); “at least \(x\)” includes them. Apply the stated condition to each catalog record before summing weights.
Check Your Understanding
For each problem, identify what the catalog quantity represents before calculating. Unless a problem says otherwise, use a Poisson model for the requested annual probability.
- A hypothetical historical catalog contains 9 qualifying events over 15 complete years. Calculate the observed annual rate and the probability of at least one event in a year under a Poisson model. (Answer: Rate \(=0.600\) events per year; probability \(=0.4512\).)
- A weighted catalog has three records, each weight an annual rate: 0.04, 0.09, and 0.07 events per year. Their event losses are $0.4 million, $1.1 million, and $2.4 million. Find the rate and annual probability for an event loss greater than $1 million. (Answer: Rate \(=0.16\) per year; probability \(=0.1479\).)
- A model has a total event rate of 0.20 per year. Conditional on an event, its three scenarios have probabilities 0.60, 0.25, and 0.15, with losses of $0.5 million, $1.0 million, and $3.0 million. Find the rate and annual probability of an event loss at least $1 million. (Answer: Rate \(=0.08\) per year; probability \(=0.0769\).)
- A catalog records 45 events over 12.5 complete years. Under a Poisson model with a constant rate, calculate the annual rate and the probability of at least one event in a year. (Answer: Rate \(=3.60\) per year; probability \(=0.9727\).)
- A weighted catalog has 20 qualifying records with annual rates of 0.03 each and 10 other qualifying records with annual rates of 0.01 each. Find the combined rate and the probability of at least one qualifying event in a year. (Answer: Rate \(=0.70\) per year; probability \(=0.5034\).)