Decisions where better data cost something first

Rebuilding wildfire exposure data before licensing a model

After the 2023 wildfire season, Haverlock paid about C$71 million on 1,140 home claims in Alberta and Nova Scotia. Our post-event review found that 41% of the 520,000 homes we insured in wildfire-exposed postal areas were located at a postal-code centroid or interpolated along a street, not on the building. Several destroyed homes had been coded as low hazard because their centroid sat in farmland. Bronwen Caskie, our head of personal property underwriting, wanted to license a vendor wildfire model at about C$2.1 million a year and use its scores at the 2024 renewals. Reinsurers had asked pointed questions about our exposure data at the January renewal.

I proposed fixing the data first and licensing the model afterwards. Bronwen disagreed in writing: she knew which areas were exposed and did not want to wait a year to act. Our chief financial officer, Ruth Lathwell, questioned a C$5.9 million data programme with no premium attached to it. The chief executive accepted my sequence on condition that Bronwen could impose interim underwriting restrictions in the forty highest-hazard postal areas. We gave up one renewal cycle of risk-based wildfire pricing, which our actuaries valued at about C$11 million of premium adequacy.

Over fourteen months we matched policies to parcel and building footprint data, asked brokers to confirm about 38,000 rural addresses, added roof material, year built and measured distance to continuous wildland fuel, and commissioned 12,000 inspections. We licensed the model in early 2025 and wrote a seven-item limits note to accompany every use of it.

About 23% of homes changed hazard band on the corrected data. At the 2025 renewals, wildfire loadings rose for about 14,300 homes and fell for about 61,000. The modelled 1-in-200 wildfire loss was 18% higher than a preliminary run on the old file had suggested, and the risk committee lowered our catastrophe retention as a result. No new home in a designated hazard area is now bound without a rooftop location.

What I would do differently is the broker communication. We told brokers in two Alberta regions about the re-banding six weeks before renewals, which was too late, and about 1,900 customers complained. I would also write the limits note before choosing a vendor rather than after, because it would have sharpened our questions.

Declining a purchased lifestyle score in auto pricing

In mid-2022, a few months into my role as chief data officer, our pricing team brought a refined auto rating model for Ontario and Alberta to the rating variable register I had just introduced. Its strongest new input was a household lifestyle segment bought from a consumer data broker and built from purchasing and survey data. Our vice-president of pricing, Anwar Bhamani, estimated that the model would improve auto results by about C$14 million a year. His team had worked on it for sixteen months, and the filing was planned for early 2023.

The register asked four questions, and the segment failed three. Nobody could say in a sentence why it should predict claims. Customers could neither see nor correct it. The data broker's consent language covered marketing, not pricing. Our proxy test also found that the segment tracked neighbourhood-level census measures of immigration and home language closely, characteristics we neither collect nor would rate on.

I declined to approve the variable. Anwar argued that the regulator would judge the filing on actuarial grounds and that we were giving up a measurable gain for a hypothetical concern. Ruth Lathwell asked for the cost in writing. Both papers went to the chief executive, who backed the register and asked the pricing team to seek the lost lift from data we could defend. We gave up the C$14 million estimate and delayed the filing by five months.

The team rebuilt the model with vehicle-specific repair-cost data, verified claims history and, for customers who chose it, a telematics programme with explicit consent. The revised model recovered about C$5 million a year of the estimated benefit. The filing drew no questions about its variables, and the telematics programme had 210,000 enrolled drivers by 2025. Since 2022 the register has reviewed 31 variables and removed four.

What I would change is the timing. The register should have existed before a team spent sixteen months building on a variable it would reject; that waste was a management failure, not theirs. I would also have involved our privacy officer when the data was bought, at the point where the consent problem was easiest to see.

Moving fraud analytics from claimants to networks

In 2017, as director of claims analytics, I built Haverlock's first fraud score for Ontario auto accident benefits claims. It scored each claimant on features such as treatment patterns, prior claims and time taken to report. By 2018 it was referring about 11,800 claims a year to our special investigations unit. When I reviewed its results in early 2019, after promotion to vice-president, 81% of referrals had closed with no finding, and referred claimants waited a median of 34 days for a first benefit payment against 12 days for everyone else. Meanwhile, organised networks of clinics, tow operators and representatives were costing us an estimated C$40 million a year, and the score rarely found them, because each claim in a ring looked ordinary on its own.

I proposed retiring the claimant score and building network analysis linking providers, vehicles, addresses, telephone numbers and representatives across claims, so that investigators received networks with evidence. An individual claim would be referred only with a network link or a specific concern an adjuster could state. Terrence Hallam, who led the special investigations unit, opposed the change. His unit was measured on referrals, and he expected us to miss opportunistic individual fraud. He was right that we would. We estimated the forgone individual-claim savings at about C$6 million a year and rewrote his unit's measures around findings and recoveries.

The network model went live in early 2020. Annual referrals fell to about 4,300, and the share ending in a finding rose from 19% to 47%. Over two years we identified 14 networks involving nine clinics, with associated avoided and recovered payments of about C$27 million. The median time to first payment across all accident benefits claims fell from 15 days to 10.

Network evidence is now the default basis for referral, and the investigations plan is built around networks rather than volumes. What I would do differently is admit sooner what the first score had learnt. It was trained on past investigators' decisions, so it reproduced their habits, including a tendency to refer claimants from particular neighbourhoods. I should have tested the labels before I tested the model. I would also have shared network indicators with the industry's investigative service a year earlier than we did.

Resubmitting capital returns over an immaterial error

In 2022 my team was tracing data lineage for our IFRS 17 implementation when we found that a 2019 claims-system migration had mapped claims under home-based business endorsements into personal property instead of commercial property. The error affected about C$46 million of unpaid claims across eight quarters in the data feeding our regulatory returns, including the minimum capital test (MCT). Because capital factors differ by line of business, our reported MCT ratio had been about two points too high, on a ratio above 200%. Solvency was never in question.

The stakes were credibility rather than capital. Ruth Lathwell preferred to correct the mapping prospectively and note the change in the next filing, a defensible position for an immaterial amount. I argued that our OSFI relationship manager should hear about it from us before that filing, that the two most recent quarterly returns should be resubmitted and that the audit committee should see the root cause, because the error had passed every control we had. Gareth Penhallow and the audit committee chair, Stuart Kittredge, agreed.

The cost was time and sequence. The resubmission and root-cause work took nine weeks of finance and data effort in the middle of our IFRS 17 parallel run, and I deferred our customer data platform by six months. We then identified 184 critical data elements feeding the capital and annual regulatory returns, assigned owners, documented lineage and added automated reconciliations at each step, at a cost of about C$3.8 million over two years. Mapping tables went under change control.

By 2025 quarter-end reconciliation took seven working days instead of seventeen, open internal audit findings on reporting data had fallen from nine to two, and our external auditor's management letter carried no data finding on the returns for the first time in four years. I now report to the audit committee twice a year on lineage coverage, mapping changes, manual adjustments and reconciliation breaks.

What I would do differently predates my appointment. In 2020, as vice-president of claims analytics, I noticed that home-based business claim counts in personal property looked high after the migration. I mentioned it to a finance colleague and assumed someone would follow it up. Nobody did. I now raise an anomaly like that in writing, with an owner and a date.