What I examine before an insurer's model is allowed to act

My first professional dataset was a map of hail storms, and I have worked on catastrophe data from three sides: a farmer-owned mutual setting hail rates, a reinsurance broker modelling clients' earthquake and storm exposure, and now an insurer pricing wildfire. I start with location and building data, because the finest hazard model cannot correct a home placed at a postal-code centroid several kilometres from its roof. At Haverlock I own the exposure data standard: rooftop location, construction, roof material, year built and measured distance to wildland fuel for every home in a designated hazard area. I also own the limits note that accompanies every use of our wildfire model, stating where the model is least validated for Canadian conditions and which return periods we treat as directional. I do not own the reinsurance purchase or the capital model. I make sure the people who do can see how good the underlying data is, and where it falls short.

Pricing at Haverlock belongs to our chief actuary. What I own is the data pricing may use, governed through a rating variable register introduced in 2022. Before a pricing team may model with a new variable, it must pass four tests: a plain-language reason it should relate to claims, a way for the customer to know and correct it, consent that covers the use, and a proxy test against neighbourhood-level census measures to see whether it stands in for a protected characteristic we do not collect. Auto rating is regulated by the provinces where we write it, and each regulator takes its own view of acceptable variables; the register sits beneath those rules and does not replace them. Since 2025 the register has also fed our automated-decision notices, which require every model that can trigger one to state its main factors in words a customer would accept. Four variables have been removed under the register, one of which performed well statistically.

Fraud in Canadian auto and property claims is increasingly organised: staged collisions, rehabilitation providers billing for treatment that did not happen, tow and storage operators steering damaged vehicles, and waves of vehicle theft with export schemes behind them. Scoring individual claimants catches some opportunists and delays many honest people. I built such a score in 2017 and retired it in 2019 in favour of network analysis that links providers, vehicles, addresses, telephone numbers and representatives across thousands of claims. Investigators now receive networks with evidence rather than names with a score. I work with our special investigations unit on what a referral must contain, with claims leaders on how quickly a claim that is not referred must be paid, and with privacy counsel on what we may share with industry investigative services and the police. I measure the programme by the share of referrals ending in a finding and by the delay it imposes on everyone else.

A federally regulated insurer reports to OSFI on its capital, claims and business through returns that directors rarely see in detail. Since 2022 I have treated the chain that produces those returns as something to be governed, not merely operated. We identified 184 critical data elements feeding the capital and annual regulatory returns, gave each an owner in finance, actuarial or claims, documented its lineage and added automated reconciliations at every step. Mapping tables sit under change control, and manual adjustments carried for more than two quarters are reported with an explanation. The same discipline supported our IFRS 17 data, where grouping and cash-flow inputs depend on policy and claims records being classified consistently. I report twice a year to the audit committee on lineage coverage, mapping changes, manual adjustments and reconciliation breaks. I do not sign the returns or the actuarial opinion; the chief financial officer and the appointed actuary carry those responsibilities. My job is to make sure they can trust the data beneath what they sign.

Haverlock uses machine learning in three underwriting settings: a document-reading tool that extracts fourteen fields from brokers' small commercial submissions, a triage model suggesting which home applications need an inspection, and gradient-boosted components inside some pricing models. None of them declines a risk on its own. Underwriters confirm the extracted fields, the triage model recommends rather than decides, and pricing components are filed and governed like any other part of a rating model. I am bringing all three into the inventory, validation and monitoring cycle that OSFI's revised model risk guideline expects, which treats machine-learning models as models rather than as software. The practical questions are plainer than the technology: how often the extraction tool misreads handwritten schedules, whether the triage model drifts as our book changes, and whether an underwriter under time pressure still checks the fields. We measure all three, and the extraction tool's accuracy, 96% on average, is reported field by field rather than as one figure.