Twenty-seven years from hail maps to the capital return

Mandate

Create Haverlock's first enterprise data function: one governed platform, accountable owners for the data behind pricing, claims and regulatory returns, and standards for how models may use customer data.

Outcomes

• Re-located 520,000 homes in wildfire-exposed areas to the building between 2023 and 2025; corrected data lowered wildfire loadings for about 61,000 homes and raised them for about 14,300.
• Gave 184 critical data elements behind the capital and annual regulatory returns named owners and tested lineage; quarter-end reconciliation fell from 17 working days to 7.
• Introduced a rating variable register in 2022; 31 variables reviewed and 4 removed, including a purchased lifestyle score.
• Extended Quebec-style automated-decision notices to all five provinces in March 2025; about 900 of 6,400 review requests in the first year changed the outcome.
• Moved most analytical workloads from five separate data stores onto one governed platform by mid-2026.

Decisions in context

Haverlock created the chief data officer role in 2021 after an internal audit found that pricing, claims and finance each kept their own version of the policy record and reconciled them by hand at quarter-end. I was appointed from claims analytics and given the platform programme, data governance, enterprise analytics and a new geospatial exposure team. The first year was dominated by lineage work for IFRS 17, which exposed the 2019 mapping error in our regulatory returns and led to the resubmission described on my impact page. From 2023 the work widened to the wildfire exposure rebuild, the rating variable register and the automated-decision notices we extended from Quebec to every province in 2025. I do not own pricing or reserving, which sit with our chief actuary, or the capital model, which sits with finance and risk. I own the data those functions use and the standards for how models may use customer data, and I present both to the risk and audit committees.

Reason for the move

Current appointment; completing the model-risk programme for machine-learning models and the extension of exposure data standards to water and storm perils.

Mandate

Improve fraud detection in Ontario auto accident benefits and property claims without delaying honest claimants, and give claims leadership reliable data on cycle times, severity and leakage.

Outcomes

• Replaced the claimant-level fraud score with network analysis; annual referrals to special investigations fell from 11,800 to 4,300 and the finding rate rose from 19% to 47%.
• Identified 14 organised networks in two years, associated with about C$27m of avoided or recovered payments.
• Reduced the median time to first accident benefits payment across all claims from 15 days to 10.
• Built the claims data mart used for reserving and leakage reviews during the 2020 move to remote claims handling.

Decisions in context

Promotion to vice-president gave me the fraud programme as well as claims analytics, and my first task was to examine the claimant-level fraud score I had built two years earlier. Its results were poor: most referrals closed with no finding, and referred claimants waited nearly three times as long for their first benefit payment. I proposed network analysis instead and spent much of 2019 persuading the special investigations unit, whose measures rewarded referral volume, to accept fewer and better files. The network model went live in early 2020, weeks before the pandemic moved claims handling to home working. I also built the claims data mart that reserving and leakage reviews still use. In 2020 I noticed an odd rise in home-based business claim counts in personal property after a system migration and mentioned it informally rather than in writing, a lapse I describe in my impact cases.

Reason for the move

Appointed Haverlock's first chief data officer when the board approved an enterprise data function.

Mandate

Build Haverlock's first claims analytics team and give adjusters and claims leaders measures they could act on.

Outcomes

• Built the first claimant-level fraud score for Ontario auto accident benefits claims in 2017, later retired in favour of network analysis.
• Introduced severity triage for property claims that sent complex losses to senior adjusters on the first day; the reopened-claim rate fell from 7.4% to 5.1%.
• Built subrogation analytics that raised annual recoveries by about C$8m.
• Wrote the first claims data dictionary, defining 340 fields used across claims, actuarial and finance.

Decisions in context

I joined Haverlock to move from pricing into claims, where I thought the data was poorer and the consequences of errors more immediate. The claims function had reports but no analytics team, and adjusters distrusted figures they had never been asked to help define. I started with a data dictionary, because two departments meant different things by a closed claim, and then built property severity triage and subrogation analytics, both of which paid for the team within two years. My least successful work was the 2017 fraud score. It met the measures we set for it, referring more claims and finding more fraud in absolute terms, while imposing delays on honest claimants that we did not measure at the time. That lesson shaped how I have evaluated every model since: by its effect on the people it does not catch as well as on those it does.

Reason for the move

Promoted to vice-president with responsibility for the fraud programme as well as claims analytics.

Mandate

Replace one-way rating tables with multivariate pricing models for auto and home, and prepare model documentation for provincial rate filings.

Outcomes

• Moved auto and home rating from one-way tables to generalised linear models in both provinces by 2013.
• Prepared the modelling exhibits for eleven auto rate filings in Alberta and Ontario.
• Built territory definitions from claims experience instead of fixed postal-code groupings.
• Contributed to rating and underwriting changes that lowered the home book's loss ratio by 3.2 points over three years.

Decisions in context

Fenwold gave me my first experience inside an insurer, working for a chief pricing actuary who wanted statistical methods brought to a rating plan that had grown by accretion. We moved auto and home rating to multivariate models over two years, and I prepared the modelling exhibits for eleven auto rate filings. The filings taught me that a model is not finished when it validates; it is finished when someone outside the company can understand why each variable belongs in it. A regulator's analyst made that point to me about a territory variable I could not explain in a sentence, and I rebuilt our territories from claims experience as a result. When the British parent sold the Canadian business in 2015, the buyer planned to absorb pricing into its own team. I chose to leave once the integration was announced rather than wait for that outcome.

Reason for the move

Fenwold's parent sold its Canadian business to a larger insurer in 2015; I left after the integration was announced to lead claims analytics at Haverlock.

Mandate

Run vendor earthquake and severe-storm models for clients' reinsurance renewals and explain the results to clients and reinsurers.

Outcomes

• Introduced an exposure data quality grade for every client submission, which reinsurers were asking for by name by 2009.
• Rebuilt earthquake modelling for clients with exposure in British Columbia and Quebec, adding soil and construction-era adjustments.
• Grew the catastrophe analytics team from 3 to 9 analysts and trained them in exposure data review.

Decisions in context

I moved to Toronto to work on a wider range of perils than prairie hail, and spent seven years running earthquake, hail and windstorm models for clients' reinsurance renewals. The work taught me how sensitive a catastrophe model is to the exposure file it is given. A home placed at its postal centroid, or recorded with unknown construction, can change a modelled loss more than any choice between vendor models. I introduced a quality grade for each client submission, covering location precision, construction and occupancy completeness and duplicate records. Reinsurers started asking for the grade, and clients improved their files to earn a better one. I managed nine analysts from 2008. By 2011 I wanted to fix data at its source rather than grade it from outside, which meant joining an insurer.

Reason for the move

Wanted to work inside an insurer, where exposure data could be corrected where it was created rather than graded afterwards.

Mandate

Bring statistical methods to township-level hail rates that had been set largely by committee.

Outcomes

• Built a storm-path dataset from eleven seasons of adjusters' reports, replacing township averages that hid where storms had tracked.
• Revised hail rates for about 1,200 townships and published the increases and decreases to members with the reasons.
• Cut the share of claims inspected more than seven days after loss from 31% to 12% by routing adjusters along forecast storm paths.

Decisions in context

I joined Coteau as its first statistician, a few months after finishing my master's degree. The mutual's hail rates were set township by township and had drifted from experience, partly because the losses were averaged across areas that storms crossed unevenly. Over two winters I built a storm-path dataset from adjusters' reports, which let us see where hail had fallen season by season. The revised rates went up for some members and down for others, and Wilf Dunstable insisted that we explain both at meetings in community halls across Saskatchewan. Farmers corrected our maps more than once. The same storm-path work let us route adjusters more quickly after large storms, which mattered to me for personal reasons. I left in 2004 to learn catastrophe modelling across a wider range of perils.

Reason for the move

Moved to Toronto to model a wider range of catastrophe perils at a reinsurance broker.