An underwriter who kept asking how the figures were produced

My mother grew up in northern Ontario speaking French; my father grew up in Ottawa speaking English. I was raised in Orléans, at the eastern edge of the city, in both languages and with a surname that tells nobody which one I think in. In 2024, when our bank began examining how it recorded customers' language, I asked the team to run my own record through the rule we had used since 2009. It placed me in English. The rule inferred language of service from a customer's surname and home branch, it was wrong for about 61,000 of our customers, and I had been chief data officer for three years without looking at it.

I begin there because that is the kind of error I have spent twenty-seven years learning to find. It is rarely a broken system. It is a reasonable shortcut, taken by someone competent, that nobody revisited and that has been quietly deciding something about customers ever since.

I finished a commerce degree at Collège universitaire Sainte-Héloïse in Ottawa in May 1999 and joined Caisse populaire Marécourt that July as a mortgage underwriter. The caisse was a francophone credit union with about C$1.1 billion of assets and eleven branches in eastern Ottawa and the counties beyond it. Within a year I was also preparing its quarterly returns to the provincial regulator, because the manager who had done them retired and I was the person who kept asking how the figures were produced. I learnt that a return summarises thousands of decisions made at desks like mine, and that the codes an underwriter types on a busy Friday become the numbers a regulator reads three months later.

In April 2003 I moved to Toronto to join the regulatory returns team at Lachford Bank, one of the country's larger banks. I prepared the mortgage and retail lending schedules of the quarterly returns and, from 2005, tested retail exposure data for the bank's move to the Basel II capital framework. It was technical work in a large team, and it taught me to read a mapping table as an auditor reads a ledger. I documented 340 rules that translated product codes into reporting categories, and found twelve that nobody could explain.

Thornvale Trust Company, a federally regulated trust company that funded mortgages with term deposits, hired me in June 2007 to build its credit-risk data. I arrived months before the financial crisis. For two years my small team assembled monthly loan-level arrears data that had existed only as branch summaries, and reconstructed fourteen years of loss history from paper files so that the company could estimate what a default might cost. We also rebuilt depositor records so that Thornvale could show, account by account, which balances deposit insurance covered; duplicate records fell from 11% to under 1%. I studied applied statistics in the evenings at Avonleigh University and completed a master's degree in 2010.

In September 2012 I joined Kestrelholm Mortgage Corporation, a listed monoline lender that originated prime mortgages through brokers and funded much of its book by securitisation. As director of mortgage portfolio analytics, reporting to the chief risk officer, I built renewal, prepayment and arrears models for a book that grew to about C$19 billion under administration. Before building anything I found that servicing records and securitisation records disagreed on about 2,300 loans a month. We corrected the records first. That delayed the renewal model by six months and made it worth having.

I came home to Ottawa in March 2017 to lead enterprise data and regulatory reporting at Aubreuil, reporting to the chief financial officer. Internal audit had found that our returns depended on more than a thousand manual adjustments each quarter. We built a loan-level mortgage data mart in which each of 212 fields has a named business owner, and brought the adjustments down to 190. When the bank created the role of chief data officer in February 2021, I was appointed to it.

Since then my work has followed three questions. The first is the mortgage renewal wave. Customers who fixed their rates for five years in 2020 and 2021 are renewing at much higher rates, and our arrears figures, which move last, were telling the board very little. We built a register of payment strain and used it to call customers early and to provision for what it showed, which our chief financial officer resisted and which raised our allowance. The second is models. When OSFI revised its guideline on model risk management, bringing artificial intelligence clearly within its scope, we counted every candidate tool in the bank rather than updating the old list. The third is language. In a bank where four customers in ten prefer French, data that works well in English and badly in French is not a translation problem. It reaches credit decisions, renewal letters and complaints.

I work best through written disagreement. Our head of model validation reports to the chief risk officer and challenges my team's models for a living, and I would rather read her objections in a memo than have a director discover them later. I take decisions to the risk committee with the dissent attached, and I report my office's errors as prominently as its improvements. Colleagues would tell you that my first question about any figure is which customers it leaves out, and that I reply in whichever language the question arrived in.

Open to Board appointments · Advisory roles · Speaking invitations

Languages / English · French