Fix the address before you buy the wildfire model
Canadian insurers are licensing wildfire models faster than they are correcting where their policyholders' homes actually stand. A model run on postal-code locations produces confident numbers about the wrong houses. Exposure data should come first, and the model's limits should travel with every price it sets.
Most of the wildfire conversations I have had with Canadian underwriters over the last three years began with a vendor's hazard map and ended with a question about price. I think the order is wrong. An insurer that cannot place each home it covers within a few metres of its roof, and say what the home is built from and how far it stands from continuous forest fuel, should not yet be using a wildfire model to price that home. It should be fixing the address. Once it has a model, it should write the model's known limits into the underwriting file in plain sentences, so that nobody downstream mistakes a modelled number for a measured one.
I am not arguing against catastrophe models. I spent seven years running them for a reinsurance broker, and I rely on one at Haverlock Assurance today. Wildfire models for Canada have improved quickly, and the insured losses of recent seasons in Alberta, British Columbia and Nova Scotia have made it irresponsible to price the peril on history alone. My argument is about sequence and candour. A sophisticated hazard model applied to a crude location produces a precise-looking error, and precise-looking errors are the kind that survive committee meetings.
What a postal-code centroid hides
After the 2023 season, my team reviewed how Haverlock located the 520,000 homes we insured in postal areas exposed to wildfire. Forty-one per cent were placed at the centre of their postal code or interpolated along a street segment, rather than on the building. In a city that error is usually a few hundred metres and rarely matters for wildfire. In a rural postal code, which in parts of Alberta covers hundreds of square kilometres, the centroid may sit in a field, a slough or a stand of spruce several kilometres from the house it represents.
Wildfire risk changes over short distances. A home thirty metres from continuous forest and a home two kilometres away across cultivated fields face different ember exposure, different fuel and different odds of a fire front arriving at all. The hazard layers in current models are resolved finely enough to tell those homes apart. Our location data was not. When we compared early model scores on our original file with scores on corrected locations, about a quarter of the homes changed hazard band. Some moved up. More moved down.
Fixing it was slow and unglamorous. We matched policies to municipal parcel and building footprint data where provinces and municipalities publish it, and asked brokers to confirm about 38,000 rural addresses that could not be matched automatically. We added roof material, year of construction and a measured distance to continuous wildland fuel, and we commissioned 12,000 inspections of homes in the highest-hazard areas to check what our application forms had collected. The programme took fourteen months and cost about C$5.9 million. Our head of personal property underwriting argued, fairly, that we were delaying rate action on homes we already knew to be exposed. We did delay it, by one renewal cycle, while she applied interim restrictions in the worst areas.
Write the model's limits into the underwriting file
Every catastrophe model is a set of assumptions about hazard, vulnerability and the way a portfolio responds to an event. Vendors document those assumptions at length, in technical papers few underwriters read. What rarely reaches the underwriting file is a short statement of what the model has not been tested against. For Canadian wildfire, that list is not trivial. Ember transport through boreal forest has been studied less than fire behaviour in the shrubland and mixed conifer of the western United States, where much model development began. The effect of suppression, of municipal fuel management and of local building practice has thinner validation data here. And a model calibrated to the fire seasons of recent decades describes a climate that is changing beneath it.
At Haverlock, any price or underwriting rule that uses the wildfire model now carries a numbered limits note, written by my geospatial team and agreed with the chief actuary. It runs to seven items and fits on one page. It says, among other things, that modelled losses for homes within 100 metres of fuel are more sensitive to vegetation assumptions than to any other input, that the model has not been validated against a Canadian fire with heavy ember-driven losses inside a town, and that we treat modelled return periods beyond 1-in-200 as directional. When the underwriting committee approves a change, the note is in the pack. When the board risk committee sees our modelled wildfire loss, the note is attached.
This sounds like bureaucracy. In practice it changes conversations. An underwriter who has read that the model is least certain at the forest edge asks for an inspection there instead of trusting a score. A reinsurer who can see that we know our limits asks fewer questions about our data, because we have answered them first. And a director reading a catastrophe paper can tell which figures deserve confidence and which are informed guesses.
Better data lowers some prices, and that is the point
The result most colleagues did not expect was that the corrected file lowered more prices than it raised. When we licensed the model in 2025 and applied it to the rebuilt exposure data, wildfire loadings rose for about 14,300 homes and fell for about 61,000. Most reductions went to rural homes far from fuel that had been carrying a loading set by the worst address in their postal code. Some increases fell on homes we had treated as safe because their postal centroid sat in farmland.
Both groups had been mispriced, and an insurer that takes pricing fairness seriously should care about the second group as much as the first. It is easy to describe exposure data work as defensive, a way to stop under-charging the riskiest homes. It is equally a way to stop over-charging the people who were subsidising them, and it is the strongest answer I know to the public suspicion that catastrophe pricing is a way of raising premiums under cover of climate change.
There were costs we underestimated. About 1,900 customers complained after renewal, most of them people whose wildfire loading rose. Brokers in two Alberta regions told us, correctly, that we had warned them too late. The corrected data also produced a larger modelled loss for the portfolio as a whole, which took two difficult meetings to explain to colleagues in finance.
My suggestion to any board approving a wildfire model is to ask three questions first. What share of the homes it will price are located on the building rather than at a postal centroid? What does the model's limits note say, and who wrote it? And when the corrected data is run, how many prices fall? If management cannot answer the first, the model can wait. If there is no answer to the second, nobody has read the model closely enough. If the answer to the third is none, I would look at the data again.