A lender's historical portfolio offers direct evidence about which borrower and collateral characteristics actually predicted good or poor outcomes—evidence that can refine approval criteria more reliably than industry convention alone. Federally regulated lenders now report to OSFI through a monthly stock dataset that tracks every outstanding mortgage and home-equity line of credit, and the Bank of Canada uses this data to study how borrower demographics, amortization paths and current loan-to-value distributions shift between origination and later observation [1]. Private lenders can build analogous feedback loops from their servicing records, though smaller volumes require more caution when drawing statistical conclusions.
Why historical portfolio data matters
Underwriting guidelines typically originate from regulatory requirements, insurer criteria or industry convention. Over time, however, a lender's own book of business becomes a rich source of evidence about which factors actually matter. A loan file captures a snapshot at origination—debt-service ratios, LTV, credit score, income verification, property type—but the servicing record reveals whether that snapshot translated into steady payments, early payoff, refinancing with another lender or default.
The Bank of Canada found that roughly 70 percent of outstanding mortgages as of September 2024 had a current LTV of 65 percent or lower—significantly better than the distribution at origination—because borrowers paid down principal and, in many cases, property values rose [1]. That pattern has direct underwriting relevance: loans originated at 80 percent LTV do not remain at 80 percent LTV; understanding the typical equity-build curve informs how conservatively to set origination thresholds.
Feedback loops between originations and outcomes
Federally regulated mortgage insurers are required under OSFI's B-21 guideline to maintain effective quality-assurance and risk-management processes, including ongoing review of their insured portfolio's performance [2]. Those insurers must monitor delinquencies, claims experience and patterns or systemic issues that may impair insurance quality [2]. The same principle benefits any prudent lender: comparing origination cohorts by year, product type or geography against their subsequent loss or prepayment rates identifies variables that improve—or fail to improve—prediction.
The Bank of Canada's 2024 analysis of mortgage stress-test policies illustrates this approach [3]. Researchers examined originations before and after the 2016 and 2018 stress-test rules, then tracked the same borrowers' subsequent delinquencies across multiple credit products. They found that areas more exposed to the 2016 policy—meaning a higher share of originations that would have failed the new test under prior underwriting—experienced lower increases in credit-card and auto-loan delinquencies when rates rose sharply in 2022–23 [3]. Tighter origination criteria, when measured against later outcomes, correlated with better borrower resilience.
Practical steps for portfolio-based learning
1. Tag every file at origination. Record the key underwriting metrics—LTV, GDS, TDS, credit score, income-verification method, property type, purpose—in a format that can be queried later. If fields migrate across systems during renewals or assumptions, preserve the origination values.
2. Define meaningful outcome events. Default is the obvious metric, but other outcomes carry information: voluntary prepayment may signal borrowers who would have renewed elsewhere at a lower rate; extended amortization suggests payment stress short of default; early refinancing with a larger balance may indicate extraction of appreciation. Choose outcomes relevant to your lending model.
3. Segment by vintage, geography and product. A loan originated in 2019 in the Greater Toronto Area under one product line exists in a different risk context than a 2022 origination in a secondary market under another product. Mixing all files together obscures which factors matter under which conditions.
4. Test assumptions against data. If current policy disqualifies self-employed borrowers with less than two years of filed tax returns, check whether the historical loss rate for that group actually differs from salaried borrowers with similar ratios. Anecdote is not evidence; portfolio data can confirm or contradict the assumption.
5. Update criteria incrementally. Large policy swings create origination cliffs that make later analysis harder. Small, documented adjustments—tightening a credit-score floor by ten points, adding a property-type restriction—allow measurement of marginal effects while retaining comparable cohorts.
What the data cannot tell you
Historical portfolios describe the loans you approved, not the loans you declined. A low default rate among borrowers with sub-700 credit scores may reflect conservative underwriting elsewhere in the file—low LTV, strong co-borrower—rather than the score being unimportant. Survivorship bias also applies: loans that prepaid quickly exit the dataset before they have time to default.
Past performance under stable or rising property values does not predict performance under a prolonged price decline. The Bank of Canada notes that about 60 percent of outstanding mortgages as of September 2024 were scheduled to renew before the end of 2026, and about 40 percent could face a higher interest rate at renewal over that period [1]. How those renewals perform will depend on future economic conditions, not just origination quality. Stress-testing exercises—separate from historical observation—remain necessary to estimate tail-risk exposure.
Scope differences: regulated versus private lenders
OSFI's data-collection and risk-management requirements bind federally regulated lenders and insurers, not provincially regulated credit unions, mortgage investment corporations or private lenders [1]. Private lenders may still benefit from the analytical approach, but they are not obligated to follow B-21's prescribed governance structure or to submit loan-level data to a regulator. The trade-off is autonomy against scale: a private lender with a few hundred loans lacks the statistical power that a national bank derives from millions of records. Pooling anonymized performance data through an industry consortium or purchasing third-party analytics—while verifying that any benchmarks reflect comparable loan types and jurisdictions—are possible workarounds, but the decision depends on each lender's volume and risk appetite.
What changes the answer
The usefulness of historical analysis depends on data quality, portfolio size and the stability of the economic environment relative to the observation period. A lender originating primarily short-term bridge loans in a single city will learn different lessons than a national lender writing 25-year amortizations across property types. No dataset automatically generates better underwriting; the value lies in disciplined comparison of origination characteristics to eventual outcomes, honest acknowledgment of data gaps, and willingness to revise criteria when evidence warrants.