Where a bank's data decides something about a customer

Canadian mortgages renew every few years while amortising over decades, so a bank's exposure to rate changes arrives in waves, borrower by borrower, at each maturity. I build the analytics that let a lender see that wave before its arrears report does. At Aubreuil this means a register of every mortgage renewing within eighteen months, the projected payment at current rates, payment as a share of observable income and, where customers bank with us, behavioural signals from deposit and card accounts. I work with credit risk on how the register enters expected credit loss estimates under IFRS 9, with model validation on whether its signals hold up, and with our privacy office on the uses customers would recognise. I am explicit about what it cannot see: borrowers whose deposits sit at another institution, and households whose strain shows nowhere in our data until they tell us.

OSFI's revised guideline on model risk management defines models broadly, brings methods built with artificial intelligence and machine learning clearly into scope and expects institutions to manage risk across the whole life of a model, including its data. Model risk policy and independent validation belong to our chief risk officer. My part is the inventory's completeness and the data each model depends on. In 2024 we ran a census of 274 candidate tools across the bank and recorded a reason for every item judged not to be a model. My office maintains the inventory platform, traces each model's lineage to source systems and names a data owner for each high-materiality model. When a model performs differently for one group of customers, I want to know first whether the data feeding it is equally complete for that group.

I require fairness testing before any credit model my team builds is released, and I ask for the same test on the process it replaces. We compare approval rates and pricing for applicants of similar observed risk across province, sector, gender where it is known and language of service. Language is never a model input; we record it so that we can test with it. Our 2023 small-business model taught me that in a bilingual market much of the unfairness sits upstream of the score, in record matching that handles accented letters, compound surnames and French addresses worse than English ones. Canadian human rights law protects grounds such as national or ethnic origin, and language can stand in for them invisibly. So we now test whether each record join succeeds equally often for each language community before we test the scores.

I have worked on regulatory returns since 2000, at a credit union, a large bank, a trust company and Aubreuil. My approach is to report from loan-level data with owners rather than from summaries corrected by hand. Between 2017 and 2021 we reduced the manual adjustments in our quarterly returns from 1,140 to 190 by building a mortgage data mart in which every one of 212 fields has a defined meaning, a business owner and a reconciliation to the general ledger. The finance team owns the returns and their sign-off; my office owns the data, its lineage and the controls that show a figure can be traced to the accounts behind it. I report data-quality exceptions affecting returns to the audit committee twice a year, including those we found too late to correct before filing.

Under the federal consumer-driven banking framework, customers will be able to direct their bank to share their data with accredited participants, and the quality of that data will matter to decisions we never see. I am preparing Aubreuil's account data as a product with a stated standard rather than an interface built to pass accreditation. That means identifying the fields most likely to be shared, measuring how complete and accurate each one is, testing them separately for French and English descriptions, and giving customers a route to correct a description or category that misrepresents them. Since mid-2025 we have reduced unidentified payroll descriptions from about 6% to 2.4% and retrained our transaction categoriser on 180,000 French-language examples labelled by our own staff. The connection itself belongs to our technology team; the content is mine to answer for.