Why Catastrophe Estimates Depend on Exposure Data
A catastrophe model can only assess the locations and characteristics it receives. A missing building, an incorrectly placed coordinate, or a policy-level limit repeated as though it were a location-level value can change the modeled loss and the apparent concentration of risk. These defects may affect both the estimated gross loss and the reinsurance protection that appears necessary.
Exposure data review is therefore not just a preliminary formatting task. It is a control over the model’s inputs and over the interpretation of its outputs. A model result may be internally consistent and still be misleading if the underlying exposure schedule is incomplete, duplicated, or mapped to the wrong locations.
What to Reconcile Before Modeling
A useful data review follows the exposure from source records to the model input. The first question is whether the records are complete: do the submitted locations and values reconcile to the relevant policy schedules and portfolio totals? The second is whether they are unique at the intended level: does each physical location appear once for each applicable coverage, or are values repeated through joins, schedules, or policy summaries? The third is whether the fields mean what the model user thinks they mean.
In particular, distinguish a building’s insured value from a policy’s total limit, a location’s value from a policy aggregate, and a coordinate from a postal centroid. A field name alone does not establish the value’s meaning. If a policy limit is repeated on every location row, summing that field as if it were location-level insured value can produce a false concentration. Conversely, if a location value is omitted, the portfolio can be understated.
Geographic checks should examine more than whether coordinates are present. Confirm that coordinates fall in plausible locations, that latitude and longitude have not been reversed, and that address, postal code, and coordinates are not materially inconsistent. A geocode can be technically valid but too coarse for the analysis: a postal centroid may place many properties at one point, obscuring their actual separation or creating an artificial cluster.
Identify the valuation date, covered policies, locations, coverages, and units included in the file. Establish whether the file is intended to represent locations, buildings, policies, or another level.
Compare row counts, unique location counts, and insured-value totals with source schedules. Investigate differences rather than treating a matching grand total as proof that individual records are correct.
Check address-to-coordinate consistency, coordinate plausibility, duplicate keys, currency, coverage definitions, and whether each amount is location-level or policy-level.
Compare model results and accumulation summaries before and after corrections, or under reasonable alternative treatments where uncertainty remains. Record assumptions and unresolved exceptions.
How Exposure Errors Change Loss Estimates
For a simple illustration, suppose a specified event applies a known damage proportion \(q_j\) to the insured value \(V_j\) at location \(j\). The event loss across the listed locations is the sum of location losses:
This is a deliberately simplified event calculation: it assumes the stated damage proportions already reflect the scenario and ignores policy terms unless they are explicitly applied. It shows why an omitted or duplicated value matters. If the damage proportion is positive, adding an omitted exposure increases this calculated loss; duplicating a record counts its value more than once.
For random losses, expected aggregate loss is also additive: if \(L=\sum_j L_j\), then \(E[L]=\sum_j E[L_j]\), provided the expectations exist. This follows from linearity of expectation and does not require the location losses to be independent. However, dependence matters when assessing the distribution of total event loss, the chance of exceeding a threshold, or the size of a reinsurance recovery. A geographic accumulation error can therefore distort tail and event-based measures even when a simple expected-loss total appears plausible.
Worked Example: Finding an Omitted and a Duplicated Location
A hypothetical portfolio is modeled for one wind scenario. The correct location schedule has three properties, with insured values shown in thousands of dollars. The model’s scenario damage proportions are supplied and are treated as fixed for this calculation.
| Location | Insured value ($000) | Scenario damage proportion |
|---|---|---|
| A | 2,000 | 0.10 |
| B | 3,000 | 0.20 |
| C | 1,000 | 0.30 |
A faulty submission omits location C and includes location B twice. Calculate the correct and faulty event losses, then measure the error.
Method. Apply the stated damage proportion to each submitted location value, then sum. The correct values are:
The faulty submission contains A once, B twice, and C zero times:
The dollar overstatement is \(1{,}400-1{,}100=300\) thousand dollars. Relative to the correct loss, the overstatement is \(300/1{,}100=0.2727\), or 27.27% when rounded to two decimals. The net error is not simply the omitted location’s loss: duplication of B adds 600 thousand, while omission of C subtracts 300 thousand, for a net increase of 300 thousand.
Check. Correct total insured value is \(2{,}000+3{,}000+1{,}000=6{,}000\) thousand. Faulty submitted value is \(2{,}000+3{,}000+3{,}000=8{,}000\) thousand. The submission both fails the value reconciliation and produces the calculated loss difference. Common mistake avoided: treating the net loss difference as evidence that only one defect exists; the omission and duplicate have separate, offsetting effects.
Geocoding and Geographic Accumulation
Catastrophe estimates often depend on whether locations lie within a hazard footprint or on how hazard varies across an area. A wrong coordinate can assign a property to the wrong hazard zone. Even if total insured value reconciles perfectly, the geographic distribution may not. That can alter modeled event losses, the apparent concentration by territory, and conclusions about how much protection is needed.
The following example uses a hypothetical event affecting only one of two zones. The event loss is calculated from the supplied zone-specific damage proportions; it is not a general catastrophe model formula. The assumed reinsurance cover is an occurrence excess-of-loss layer with a 4 million dollar limit excess of an 8 million dollar attachment. For event loss \(L\), its recovery is the amount above the attachment, limited to the layer limit:
Worked Example: Correcting a Mislocated Property Before Applying a Layer
A hypothetical portfolio has three locations. Zone North is affected by the event, with a 50% damage proportion; Zone South is outside the event footprint, with a 0% proportion. The correct schedule places two locations in North and one in South. A geocoding error places the South location in North. Values are in millions of dollars.
| Location | Insured value ($m) | Correct zone | Zone in faulty file |
|---|---|---|---|
| 1 | 10 | North | North |
| 2 | 8 | North | North |
| 3 | 6 | South | North |
Correct event loss. Only the 18 million dollars in North is affected:
Faulty-file event loss. The misplaced 6 million dollar location is incorrectly assigned the North damage proportion:
Using the stated occurrence layer, the correct recovery is \(\min\{\max(9-8,0),4\}=\min\{1,4\}=1\) million dollars. The faulty estimate is \(\min\{\max(12-8,0),4\}=\min\{4,4\}=4\) million dollars. Thus the geocoding error overstates modeled event loss by 3 million and indicated layer recovery by 3 million.
Check. The portfolio’s total insured value remains \(10+8+6=24\) million in both files, so a grand-total check alone will not detect this defect. A zone-level reconciliation shows the issue: correct North value is 18 million, whereas the faulty file reports 24 million. Common mistake avoided: assuming that a correct portfolio total guarantees a correct catastrophe result.
Aggregation Levels and Reinsurance Decisions
Exposure aggregation must respect the meaning of each field and the terms being analyzed. A location-level building value can be summed across locations for an exposure total. A policy limit repeated on several location rows cannot be summed as though each row were a separate limit. Similarly, a policy aggregate, a coverage limit, and a location sublimit are not interchangeable. The appropriate aggregation depends on the question and the contract structure.
The next example isolates a common field-mapping error. It assumes that the supplied event damage proportions are applied to each location’s building value and that the resulting event loss is compared with a hypothetical occurrence layer. It does not apply a policy limit or deductible.
Worked Example: Preventing a Repeated Policy Limit from Inflating Exposure
A hypothetical policy covers three buildings with location-level insured values of 3 million, 2 million, and 1 million dollars. Its total policy limit is 6 million dollars. Because the policy limit is repeated on each location row, a faulty model mapping reads 6 million as the insured value at every building. The event damage proportions are 50%, 25%, and 10%, respectively. Determine the correct and faulty event losses and compare recoveries under a 3 million dollar limit excess of a 2 million dollar attachment.
Correct calculation. Use location-level values:
The layer recovery is \(\min\{\max(2.10-2,0),3\}=\min\{0.10,3\}=0.10\) million dollars.
Faulty calculation. The repeated policy limit is incorrectly used as each building’s value:
The faulty layer recovery is \(\min\{\max(5.10-2,0),3\}=\min\{3.10,3\}=3.00\) million dollars. The error overstates event loss by \(5.10-2.10=3.00\) million dollars and modeled recovery by \(3.00-0.10=2.90\) million dollars.
Check. Correct location-level insured values sum to \(3+2+1=6\) million, matching the policy limit in this example. The faulty mapping treats the same 6 million dollar limit as three separate location values, creating 18 million dollars of apparent exposure. Common mistake avoided: using a plausible-looking policy total without checking whether the field is a policy limit or a location value.
Accumulation Controls and Decision Use
An accumulation report is only meaningful when its grouping is explicit. A portfolio total by postal code, for example, answers a different question from an event-footprint loss estimate or a total by treaty occurrence. A control should state what is grouped, which peril and event assumptions apply, whether values are gross or net of policy terms, and whether the output is an exposure total or a modeled loss.
Useful controls include:
- Completeness and reconciliation: compare policy and location counts and insured-value totals to authoritative schedules; investigate missing, unmatched, or unexpected records.
- Duplicate detection: test unique keys at the appropriate grain, such as policy-location-coverage, and review repeated values that may arise from joins or repeated policy fields.
- Geographic validation: flag missing or implausible coordinates, reversed latitude and longitude, address-coordinate conflicts, unexpected concentrations at identical points, and geocodes that are too coarse for the intended analysis.
- Field and unit validation: confirm currency, valuation date, coverage definitions, units, and the distinction between insured value, limit, and aggregate.
- Impact assessment: compare totals by relevant zones and rerun or adjust the analysis after material corrections. Where location is uncertain, evaluate reasonable alternative placements rather than silently treating one geocode as certain.
- Audit trail: retain the source, correction, rationale, responsible reviewer, and before-and-after impact for material changes.
These controls support better decisions, but they do not establish the “correct” amount of catastrophe reinsurance by themselves. Reinsurance analysis also depends on contract terms, modeled event losses, uncertainty, and the insurer’s risk appetite. The data review tells the analyst whether those calculations are being made on a sufficiently reliable representation of the exposure.
A useful reasonableness check is to compare multiple views: total insured value by policy, by location, and by geographic grouping; counts of unique locations; and modeled losses under the same event assumptions. A discrepancy does not automatically prove an error, because the views may use different definitions. It does identify a question that should be resolved before relying on an accumulation or recovery estimate.
Check Your Understanding
For each hypothetical situation, identify the relevant data issue and calculate or state its consequence.
- A wind scenario assigns a 20% damage proportion to a location with 5 million dollars of insured value. The location is omitted from the model file. What event loss is omitted? (Answer: 1 million dollars.)
- Two records each show 4 million dollars of insured value and a 25% event damage proportion. The second record is a duplicate of the first. By how much is event loss overstated? (Answer: 1 million dollars.)
- A 7 million dollar location is correctly outside an event footprint but is geocoded into an affected zone with a 40% damage proportion. What is the loss overstatement under these assumptions? (Answer: 2.8 million dollars.)
- A policy’s 9 million dollar total limit is repeated across four location rows. The analyst sums it as location-level exposure. What apparent exposure total results, and how much does it exceed the policy limit? (Answer: 36 million dollars apparent exposure, exceeding the 9 million dollar limit by 27 million dollars.)
- A hypothetical event produces a corrected gross loss of 11 million dollars. The occurrence cover is 5 million dollars excess of 8 million dollars, with no other terms in the calculation. What is the recovery, and what data checks should precede relying on it? (Answer: 3 million dollars; reconcile location records and values, validate geocoding and aggregation level, and confirm the exposure and event assumptions.)