Property risk / How the models work
Two models your risk function can actually inventory.
Under OSFI E-23 a third-party model needs an inventory entry, a validation and ongoing monitoring. That requires knowing what the model is made of. Every weight, resolution, band edge and artifact vintage below is published by the engine and queryable, so the validation has something to work with.
Wildfire
EMBER: five weighted components over measured ground.
| Component | Weight | What it measures |
|---|---|---|
| fuel_exposure | 0.35 | Combustible fuel around the building, from 30 m land cover and 250 m fuel typing. |
| terrain | 0.15 | Slope and how it carries fire toward or away from the structure. |
| ignition_history | 0.20 | Seventy-five years of recorded ignitions near the point. |
| potential_behaviour | 0.20 | CFFDRS fire behaviour under a declared severe reference scenario. |
| suppression_context | 0.10 | Built density and road access around the property. |
Fire behaviour runs under a declared severe reference scenario, fixed and published. It is not today’s weather and not a forecast, so the assessment does not move because a dry week arrived. That is what makes it a standing property reading rather than a daily one.
Fluvial flood
CREST: two independent estimates, plus escalators that only ever raise.
| Component | Weight | What it measures |
|---|---|---|
| nrcan_flood_susceptibility | 0.75 | Natural Resources Canada’s published national susceptibility mapping, at 120 m. |
| hand_susceptibility | 0.25 | Height above the nearest drainage, sampled at 30 m and mapped through a frozen province-wide percentile reference. |
Escalators, not terms
A regulatory floodplain designation and recorded flood events can raise a band and can never lower one, because their absence is silence rather than safety. Designated floodplain mapping covers roughly 0.19% of British Columbia, so treating its absence as good news would mis-read almost the entire province.
Fluvial only, and it says so
CREST assesses riverine flood. Coastal and storm-surge, pluvial and surface-water, flash, and sewer-backup flooding are named in the response as not assessed, in the same categories the federal climate return uses. An unassessed peril is recorded as absence of assessment, never as a low reading.
Versioning
Two fingerprints, because “did it change?” is two questions.
methodology_fingerprint
Covers only what we control: versions, formulas, thresholds and artifact vintages. It answers “did you change?”. If it moves, your own model controls should fire.
assessment_fingerprint
Covers that, plus the values measured at the location. It answers “did anything change?”. Same methodology, different assessment means the world moved and no revalidation is implied.
One hash could not separate those. Re-run an address a year later, compare both, and you know whether the property changed or we did, which is the difference between a note in the file and a model revalidation.
What neither model claims
The limits are part of the specification.
- Not a probability of fire or flood, at any return period.
- No inundation depth, extent or velocity. No flame length delivered as a forecast.
- No loss, damage, premium or actuarial quantity. We never receive your sums insured and the interface has no field for them.
- A band is a position on a screening scale, not a statement that an event will or will not occur.
- A missing component makes an index unavailable rather than reweighted. An unavailable index, never a quietly lower one.
| Component | Resolution | Contribution |
|---|---|---|
| nrcan_flood_susceptibility | 120 m | •••• |
| hand_susceptibility | 30 m | •••• |
Assessed peril: fluvial. Coastal, pluvial, flash and sewer-backup flooding are not assessed, and are reported as not assessed rather than as a low reading.
methodology_fingerprint •••• · assessment_fingerprint ••••
Put it in front of your validation team.
Tell us the book you would run and we will come back to you about it.