Why Event Catalogs Matter
A catastrophe model needs a structured account of the events that could affect an insured portfolio. An event catalog supplies that account: it organizes events and their attributes so that each event can be considered against a defined set of exposures. The central distinction is whether the catalog describes events that were observed or events generated as plausible possibilities.
The distinction affects interpretation. Historical catalogs preserve evidence about past events, subject to the completeness and conventions of the underlying records. Stochastic catalogs represent modeled events that may not have occurred in the historical record. Neither catalog, by itself, is a complete estimate of future loss. In particular, a list of event rows does not automatically state how likely each event is.
An event is not the same thing as a source record. Several reports may refer to the same physical event, while two events of the same peril may occur close together in time. A catalog therefore depends on an event-definition and record-linkage convention. A footprint describes the locations or geographic area associated with an event; it is not itself a measure of damage or insured loss.
Historical and Stochastic Catalogs
| Feature | Historical catalog | Stochastic catalog |
|---|---|---|
| What it represents | Events reported or otherwise documented as having occurred | Events generated by a model as plausible possibilities |
| Event identity | Resolved from observed records, dates, sources, and event conventions | Assigned within the simulation; not necessarily tied to a named historical event |
| Geography | Observed or reconstructed event location and footprint | Modeled event footprint or geographic description |
| Main interpretive limitation | Past observation may be incomplete and may reflect a different exposure base | Results depend on model assumptions and generated possibilities, not direct observation |
A stochastic event need not be a forecast of a specific future date or a prediction that a named event will occur. It is a modeled possibility under the model’s assumptions. Likewise, a historical catalog is not necessarily a complete inventory of all events that could occur. The events, geographic detail, and record quality depend on available data and the rules used to assemble the catalog.
Catalog contents can be used in different ways. One analysis may compare footprints with a portfolio’s locations; another may use event-level loss estimates already produced for those events. Those uses require care about the exposure and policy basis of the estimates. The event catalog itself should not be mistaken for a frequency model: interpreting event rates or catalog weights is a separate step.
Worked Example: Resolving Historical Records Into Events
A hypothetical archive contains six reports: two reports describe the same windstorm on September 14 and have matching tracks; one report describes a flood in a separate basin on September 14; two reports describe the same hailstorm on September 20; and a final report describes a distinct hailstorm on September 21. Under the stated event-linkage rule, matching tracks and dates identify duplicate reports of one event. How many distinct events are represented, and how many are in each peril group?
Method. Group reports by physical event under the specified linkage rule before counting events. The two windstorm reports on September 14 form one event, not two. The two September 20 hail reports also form one event.
The six reports reduce as follows:
- Windstorm: 2 reports become 1 event.
- Flood: 1 report remains 1 event.
- Hail: 2 September 20 reports plus 1 distinct September 21 report become 2 events.
The distinct-event count is \(1+1+2=4\). The peril counts are 1 windstorm, 1 flood, and 2 hail events; these sum to 4.
Check. The original six reports contain two duplicate reports, so \(6-2=4\) distinct events, consistent with the peril-by-peril count. The common mistake is to treat every source report as a separate event, which would incorrectly produce six events.
Footprints and Exposure Context
A catalog’s footprint is important because an event can affect some portfolio locations and not others. A footprint may be represented in a catalog by a geographic boundary, a set of affected locations, or another spatial description. The precise representation depends on the data and model. At this stage, the key interpretation is that a footprint associates an event with geography; it does not say how much damage each location sustains.
For a simple portfolio summary, the values insured at locations identified as within a footprint can be added. This produces a measure of exposure associated with that footprint, not an estimate of event loss. It also requires a consistent location inventory and a clear rule for deciding which locations are included. A location inside two footprints can appear in both event-specific exposure totals.
Worked Example: Summarizing Overlapping Event Footprints
A hypothetical insurer has five locations with total insured values (TIVs) of $1.2 million, $0.9 million, $0.8 million, $2.0 million, and $1.5 million, respectively. A historical event footprint includes locations 1, 3, and 5. A stochastic event footprint includes locations 2, 3, and 4. Find the TIV associated with each footprint, the TIV at locations in both footprints, and the TIV across the union of the two footprints.
Method. Add the TIVs for the locations listed in each footprint, then identify the shared and combined location sets. All TIVs are given in millions of dollars.
For the historical event, the included locations are 1, 3, and 5:
Its associated TIV is $3.5 million. For the stochastic event, the included locations are 2, 3, and 4:
Its associated TIV is $3.7 million. The shared location is location 3, so the TIV in both footprints is $0.8 million. The union includes locations 1, 2, 3, 4, and 5 exactly once:
Thus the union TIV is $6.4 million. As a check, adding the two footprint totals gives $3.5+$3.7=$7.2 million, which exceeds the union total by $0.8 million, exactly the TIV at the shared location. The common mistake is to call the $3.5 million or $3.7 million a loss. These are exposure summaries only; no damage or policy payment has been calculated.
From Catalog Events to Loss Interpretation
An event catalog can be paired with a portfolio and a loss calculation process to obtain event-level loss estimates. The loss is not an intrinsic property of the event alone: it depends on the portfolio and the terms or valuation basis used. A useful way to express this dependence is
where \(e\) is an event, \(P\) is the portfolio being evaluated, and \(T\) denotes the applicable coverage and financial terms. This notation is a reminder that the same event can produce different loss estimates for different portfolios or terms. It does not prescribe how damage or payments are calculated.
In particular, a historical reported loss and a current-portfolio estimate for the same historical event need not match. The historical report reflects the exposures and reporting basis at the time of the event. Revaluing that event against a current portfolio asks a different question: what loss estimate would result if the event footprint were applied to the current portfolio under the current evaluation basis?
Determine whether each row is an observed event, a source report, or a model-generated event. Resolve duplicate records under a stated event convention.
Identify the footprint representation and the locations it associates with the event. Do not treat geographic overlap as a damage estimate.
Establish whether an event loss is a historical reported amount or an estimate on a specified current portfolio and coverage basis.
A list of losses can be ranked or summarized, but event rates, weights, and probabilities require additional modeling information.
Why the Two Catalog Types Complement Each Other
Historical catalogs provide observed evidence and can support examination of what happened, where it happened, and what was recorded. Their limitations include gaps in observation, changes in event definitions, and differences between past and current exposures. A short or incomplete historical record may omit plausible event patterns, and a well-documented past event still does not establish how often comparable events will occur.
Stochastic catalogs can represent a broader set of plausible event configurations than the observed record alone. They are useful for exploring modeled possibilities, including events absent from the historical archive. Their outputs are conditional on model construction and assumptions. A stochastic catalog should not be described as observed experience, and the mere presence of an event in it does not mean that the event is expected to occur in a particular year.
The two sources can therefore be used together, but they answer different questions. Historical evidence can help assess whether modeled event characteristics are credible, while stochastic events can broaden the set of scenarios considered. Neither comparison eliminates uncertainty: observed records may be incomplete, and simulated catalogs may omit or misrepresent relevant possibilities.
Common Mistakes
- Counting reports instead of events. Multiple sources can report the same physical event. Apply an explicit linkage rule before counting distinct events.
- Calling a stochastic event a forecast. A generated event is a modeled possibility, not a prediction of a specific future occurrence.
- Treating footprint overlap as loss. A location associated with an event may contribute exposure within the footprint, but damage and insured payment have not thereby been determined.
- Comparing historical loss directly with a current estimate. The exposure base, coverage, valuation, and reporting basis may differ. State which basis each loss represents.
- Reading probabilities from catalog rows. Counts or rankings in a catalog do not, without additional assumptions and information, establish event rates or probabilities.
Worked Example: Comparing Event Loss Outputs on a Common Basis
For a hypothetical current portfolio and a fixed evaluation basis, a model produces current-basis loss estimates of $0.6 million, $1.1 million, and $1.4 million for three historical events. It also produces estimates of $0.4 million, $0.9 million, $1.3 million, $1.8 million, and $2.2 million for five stochastic events. Compare the largest listed loss in each group, count the listed losses above $1.5 million, and state what can and cannot be concluded.
Method. Compare event-level outputs only after confirming they use the same portfolio and evaluation basis. Rank each group and apply the stated $1.5 million threshold. Do not infer probabilities from the number of rows.
The historical estimates in increasing order are $0.6 million, $1.1 million, and $1.4 million. Their maximum is therefore $1.4 million. None exceeds $1.5 million, so the historical count above the threshold is 0.
The stochastic estimates in increasing order are $0.4 million, $0.9 million, $1.3 million, $1.8 million, and $2.2 million. Their maximum is $2.2 million. The values $1.8 million and $2.2 million exceed $1.5 million, so the stochastic count above the threshold is 2.
The stochastic catalog’s largest listed loss exceeds the historical catalog’s largest listed loss by
The ratio of the two maxima is \(2.2/1.4=1.5714\) to four decimal places. Thus the largest listed stochastic loss is about 1.5714 times the largest listed historical loss in this comparison. The check is that the comparison uses the same dollar units and common portfolio basis. The common mistake is to conclude that two of five stochastic events exceeding the threshold means a 40% annual chance of exceeding it. That conclusion is unsupported: the catalog counts alone do not provide event rates, weights, or annual probabilities.
Check Your Understanding
Work each problem by identifying what the catalog data establish, and avoid treating an event count as a probability.
- A historical archive has seven reports. Two pairs of reports are confirmed duplicates, each pair referring to one physical event; the other three reports each describe a different event. How many distinct events are represented? (Answer: 5 distinct events.)
- A hypothetical stochastic event footprint contains locations with TIVs of $0.7 million, $1.1 million, and $1.6 million. What is the TIV associated with the footprint, and is that amount an event loss? (Answer: $3.4 million; no, it is an exposure summary, not a loss.)
- A historical event produced a reported loss of $0.9 million on the portfolio insured at the time. A current-basis revaluation of the same event is $1.3 million. By how much does the current-basis estimate exceed the historical reported amount, and are the amounts necessarily inconsistent? (Answer: $0.4 million; no, they can differ because the exposure or evaluation basis differs.)
- A stochastic catalog lists losses of $0.5 million, $1.7 million, $2.1 million, and $1.2 million on a common portfolio basis. How many exceed $1.5 million, and what is the largest listed loss? (Answer: 2 losses; $2.1 million.)
- An analyst observes that 3 of 12 event rows in a stochastic catalog exceed a specified loss threshold and reports a 25% annual exceedance probability. What is the interpretive error? (Answer: The row fraction is not an annual probability unless the catalog’s simulation structure and appropriate event rates or weights support that calculation.)