Why Location-Level Inputs Matter
A catastrophe model can only use the risk information supplied to it. A policy total, a street address, and a construction code are not interchangeable descriptions of exposure: each represents a different piece of information at a different level of detail. If the model input combines buildings, omits a location, or places a risk at the wrong coordinates, the resulting analysis may not represent the insured portfolio the actuary intended to model.
Location-level preparation is therefore a data-control task as well as a modeling task. The actuary should be able to trace each modeled exposure to a distinct insured location, confirm the meaning and units of its attributes, reconcile its coverage values, and identify attributes that are missing or uncertain. This tutorial focuses on those inputs and diagnostics, not on event simulation or the calculation of catastrophe losses.
Fields to Assemble and Validate
The exact required fields depend on the model and the product, but a useful input record generally separates four groups of information. Field names and permitted values should be mapped to the model’s specifications rather than assumed to have the same meaning across systems.
| Field group | Examples | Key control |
|---|---|---|
| Location identity and geography | Policy and location identifiers; street address; city, state or province, postal code, country; latitude and longitude | Confirm that each record identifies the insured site, not merely a mailing address or a policyholder’s headquarters |
| Construction and physical attributes | Construction class; year built; number of stories; roof type; building area | Check definitions, allowed codes, units, and whether the value describes the insured structure |
| Occupancy and use | Occupancy class; business use; residential or commercial indicator | Use a documented mapping; do not infer an exact class from a broad description without recording the assumption |
| Coverage and exposure | Building, contents, and business-interruption values; currency; valuation basis; coverage period | Confirm amounts are assigned to the correct location, component, currency, and effective period |
The model input should preserve the distinction between a value of zero and an unknown value. Zero business-interruption coverage means the coverage amount is known to be zero; a blank field means it is not known or was not supplied. Treating a blank as zero can understate the recorded exposure, while treating a true zero as missing can lead to an unjustified imputation.
Location identifiers also require attention. A policy can contain multiple buildings at one address, and one building can be represented by multiple coverage records. Before aggregating or joining data, define what constitutes one modeled location and how the records map to it. Otherwise a duplicate may count exposure twice, or an overly broad grouping may combine risks that should remain separate.
Worked Example: Reconcile Location-Level Coverage Values
A hypothetical commercial property schedule has four distinct locations. All amounts are in dollars. The recorded building, contents, and business-interruption values are: Location A, $1,200,000, $300,000, and $500,000; Location B, $800,000, $200,000, and $0; Location C, $1,500,000, $500,000, and $400,000; Location D, $600,000, $100,000, and $200,000. Determine each location’s total recorded exposure and the portfolio total. Also identify two attribute exceptions: Location C has no year-built value, and Location D has no roof-type value.
Method. For this data-reconciliation check, define recorded exposure at a location as the sum of its three supplied coverage values. This is a schedule total, not a modeled loss. Assume all amounts are in the same currency and apply to the same coverage period.
Location A totals $1,200,000 + $300,000 + $500,000 = $2,000,000. Location B totals $800,000 + $200,000 + $0 = $1,000,000. The recorded zero at Location B is treated as a known value, not as missing data.
Location C totals $1,500,000 + $500,000 + $400,000 = $2,400,000. Location D totals $600,000 + $100,000 + $200,000 = $900,000.
Add the four location totals: $2,000,000 + $1,000,000 + $2,400,000 + $900,000 = $6,300,000. The same total is obtained by coverage component: building values total $4,100,000; contents values total $1,100,000; and business-interruption values total $1,100,000. Thus $4,100,000 + $1,100,000 + $1,100,000 = $6,300,000.
Check and trap. The component totals reconcile to the location totals. The two missing physical attributes should be flagged separately from the dollar reconciliation: a complete exposure total does not establish that the records have complete model inputs. A common mistake is to replace Location B’s known $0 business-interruption value with a missing-value flag, or to regard the $6,300,000 total as evidence that the construction data are complete.
Geocoding: Address, Coordinate, and Precision
Geocoding assigns geographic coordinates to a location, typically by matching supplied address information to a map or reference database. A rooftop or parcel match may be more spatially specific than a street-segment match; a postal-area centroid can be substantially less specific. These are not equivalent placements. If a geocoder cannot locate a complete address, the returned point may represent a broader area rather than the insured building.
A practical review compares the source address with the geocoded address and records the match type or confidence indicator when available. Useful exception checks include missing coordinates, coordinates outside the stated country or region, several unrelated addresses assigned identical coordinates, and an unexpectedly large distance between the supplied and returned locations. Distance checks should use an appropriate coordinate reference system; raw differences in latitude and longitude are not distances in metres.
There is no universal distance threshold that identifies every error. A company may set a review rule suited to its data and use case, then investigate records that exceed it. A point several hundred metres from an address might be acceptable in some circumstances and unacceptable in others. The important control is to distinguish verified locations from approximate or unresolved placements and to preserve that distinction for model-input review.
Worked Example: Measure Exposure With Geocoding Exceptions
A hypothetical four-location schedule has recorded exposures of $2,000,000, $1,000,000, $2,400,000, and $900,000. An address review compares source-address positions with model-input coordinates using a local projected coordinate system. The respective distances are 40 m, 310 m, 120 m, and 680 m. Apply the company’s stated review rule: flag a location when the distance is greater than 250 m. Calculate the number of flagged locations and the share of total exposure at flagged locations.
Method. Apply the threshold to each distance, then sum the exposure of flagged records. Assume the four exposure values use the same currency and period, and that the supplied distances are already measured in metres in a suitable projected system.
The distances 40 m and 120 m are not greater than 250 m, so Locations A and C are not flagged. The distances 310 m and 680 m exceed 250 m, so Locations B and D are flagged. That is 2 of 4 records, or 50.00% of the location count.
The flagged exposure is $1,000,000 + $900,000 = $1,900,000. Total exposure is $2,000,000 + $1,000,000 + $2,400,000 + $900,000 = $6,300,000. Therefore the flagged exposure share is $1,900,000 / $6,300,000 = 0.301587…, or 30.16% when rounded to two decimal places.
Check and trap. The location count share is 50.00%, but the exposure share is 30.16%; they answer different questions. A common mistake is to report the count percentage as the percentage of exposure needing review. These calculations identify records for investigation; they do not establish the amount of catastrophe loss associated with them.
Missing Attributes and Exposure-Weighted Diagnostics
Missingness should be summarized by both record count and exposure amount. A missing value at a small location and a missing value at a large location each count as one incomplete record, but they do not represent the same share of portfolio exposure. For a field such as construction class, the exposure-weighted missing share is calculated as the missing-field exposure divided by total exposure, with both quantities measured on the same basis.
Here, \(V_j\) is the chosen exposure measure for location \(j\), and \(M\) is the set of locations with a missing value for the field being reviewed. The measure might be building value when auditing building construction; it should be stated explicitly. The formula is a direct ratio: the numerator adds the exposure for incomplete records, and the denominator adds exposure for all records. It does not estimate modeled loss or prove that missing records have the same risk characteristics as complete records.
An imputation can be useful for sensitivity analysis or an operational workflow, but it is an assumption, not a recovered fact. Keep the original missing indicator, document the rule, and compare reasonable alternatives. In particular, imputing the most common class can conceal whether the missing records are concentrated in a different type of location. The impact of missingness should be assessed against the intended model use and available source documentation.
Worked Example: Quantify Missing Construction and Test an Allocation
A hypothetical five-location portfolio has total exposure of $9,000,000. The construction fields are complete for three locations: wood-frame exposure is $2,000,000, masonry exposure is $3,000,000, and steel-frame exposure is $2,200,000. Construction is missing for the remaining two locations, whose exposures are $1,000,000 and $800,000. Calculate the missing-record share, the missing-exposure share, and an exposure allocation that assigns the missing $1,800,000 in proportion to the observed construction mix. Then compare that allocation with assigning all missing exposure to masonry.
Method. Assume the listed exposure amounts are additive, measured in the same currency and coverage period, and that the three named construction categories cover all known records. Proportional allocation is only a stated sensitivity assumption; it does not identify the true construction of either incomplete location.
Two of five records are missing construction, so the record share missing is 2 / 5 = 40.00%. The missing exposure is $1,000,000 + $800,000 = $1,800,000. Its share of total exposure is $1,800,000 / $9,000,000 = 20.00%.
The complete records total $2,000,000 + $3,000,000 + $2,200,000 = $7,200,000. Their observed construction proportions are: wood frame, $2,000,000 / $7,200,000 = 27.7778%; masonry, $3,000,000 / $7,200,000 = 41.6667%; and steel frame, $2,200,000 / $7,200,000 = 30.5556%.
Apply these proportions to the missing $1,800,000. The allocation is $1,800,000 × 27.7778% = $500,000 to wood frame; $1,800,000 × 41.6667% = $750,000 to masonry; and $1,800,000 × 30.5556% = $550,000 to steel frame. The allocations sum to $500,000 + $750,000 + $550,000 = $1,800,000.
After proportional allocation, the class totals are $2,500,000 wood frame, $3,750,000 masonry, and $2,750,000 steel frame. They sum to $9,000,000. If instead all missing exposure is assigned to masonry, the totals are $2,000,000 wood frame, $4,800,000 masonry, and $2,200,000 steel frame. These also sum to $9,000,000, but the exposure mix is different.
Check and trap. The two allocations preserve total exposure but produce different construction distributions. The proportional result is not a fact about the two locations; it is one assumption to test. A common mistake is to infer that an imputed portfolio total remains equally reliable as the original construction mix simply because the dollar total reconciles.
A Repeatable Input-Review Sequence
The following sequence keeps identity, geography, attributes, and exposure checks distinct. It also makes the resulting exceptions actionable: a reviewer can tell whether a record needs address correction, source-document follow-up, or a documented data assumption.
Establish what one location means for the model and how policy, building, and location identifiers relate. Check for duplicate identifiers and unintended aggregation.
Check address components, coordinate presence, country or region consistency, match type, and exceptions under a documented distance rule.
Review construction and occupancy codes, units, allowed values, and missingness. Keep unknown, zero, and not applicable distinct where the data design permits.
Check component amounts, currency, effective period, and location assignment. Compare location sums with the source schedule and investigate differences.
Show record counts and exposure amounts affected, describe any imputation or correction, and retain a traceable record of the chosen treatment.
Why the Distinctions Matter
A record-level error can have a different practical importance from a portfolio-level data gap. A missing field on a large location may deserve priority even if few records are incomplete. Likewise, a geocoding exception may involve a small number of locations but a substantial amount of exposure. Reporting both counts and exposure shares helps reviewers prioritize work without confusing data quality with modeled loss.
No single completeness percentage establishes that the inputs are suitable. A portfolio can have nearly complete addresses but weak construction detail, or complete coverage totals with location assignments that have not been verified. The appropriate response depends on the field, the amount affected, the source evidence, and the intended use. Corrections should be distinguished from assumptions; unresolved uncertainty should remain visible rather than being converted silently into an apparently precise value.
Check Your Understanding
For each problem, state any assumptions needed and distinguish record counts from exposure amounts.
- A two-location schedule reports building values of $900,000 and $600,000, contents values of $100,000 and $150,000, and business-interruption values of $0 and $50,000. Calculate each location total and the portfolio total. (Answer: $1,000,000 and $800,000 by location; $1,800,000 total.)
- A location schedule has exposures of $400,000, $900,000, and $700,000. Under a review rule that flags coordinate discrepancies greater than 300 m, the respective distances are 90 m, 310 m, and 450 m. How many records and what percentage of exposure are flagged? (Answer: 2 of 3 records; $1,600,000 / $2,000,000 = 80.00% of exposure.)
- A portfolio has 12 location records and $12,000,000 total exposure. Two records, with exposures of $1,100,000 and $1,900,000, have no occupancy class. Calculate the missing-record share and missing-exposure share. (Answer: 2 / 12 = 16.67% of records; $3,000,000 / $12,000,000 = 25.00% of exposure.)
- Known construction exposure is $1,200,000 wood frame and $2,800,000 masonry. Another $1,000,000 has missing construction. Allocate the missing amount in proportion to the observed mix, then give the resulting class totals. (Answer: observed proportions are 30.00% wood frame and 70.00% masonry; allocate $300,000 and $700,000; resulting totals are $1,500,000 wood frame and $3,500,000 masonry.)
- A geocoding audit flags 3 of 8 locations. The flagged locations contain $2,400,000 of exposure, while the full schedule contains $10,000,000. Give both flagged percentages and explain which measures portfolio exposure affected. (Answer: 37.50% of records and 24.00% of exposure; the exposure-weighted figure is 24.00%.)