Initial source-linked research article with operating analysis and illustrative examples.
The question a headline loss rate cannot answer
A portfolio loss rate mixes together loans originated at different times. Newly booked loans have had less opportunity to become delinquent or charge off. Rapid growth can therefore lower a reported ratio even when underwriting is weakening. The OCC’s Retail Lending handbook explicitly discusses this distortion and describes vintage and lagged analysis as tools for addressing it.
A vintage is a group of loans originated during a common period, such as a month or quarter. Comparing vintages at the same months on book helps isolate how performance evolves after origination. It does not automatically control for economic conditions, product changes or borrower mix; those require additional analysis.
Define the metric before drawing the curve
For an installment portfolio, one useful measure is cumulative net principal losses divided by original principal. It keeps the denominator fixed while the cohort seasons. Another measure, delinquent balances divided by current balances, answers a different question and can rise as good loans amortize or prepay. Neither should be relabeled as the other.
Credit-card cohorts add complications because balances can grow after opening, customers can draw repeatedly and limits can change. An account-opening vintage and a purchase vintage are different populations. Analysis should specify which exposure is being followed, how transfers and recoveries are handled, and whether accounts that close remain in the denominator.
A worked example: the denominator trap
All values in this example are hypothetical. In period one, a lender has $100 million of average receivables and $5 million of annual net losses, a 5% rate. In period two, average receivables grow to $150 million and annual losses reach $6 million. The reported rate falls to 4%, even though loss dollars increase by 20%. Growth explains why the ratio alone cannot establish better credit quality.
| Metric | Period one | Period two |
|---|---|---|
| Average receivables | $100 million | $150 million |
| Annual net losses | $5 million | $6 million |
| Reported loss rate | 5.0% | 4.0% |
| New cohort: cumulative loss at month 6 | 1.0% of original principal | 1.4% of original principal |
Read the second signal carefully
In the example, equal-age cohort loss worsens by 40 basis points, or 40% relative to the earlier 1.0% rate. That is a warning worth investigating, not proof that a policy change caused it. The newer cohort may have experienced different unemployment, a changed merchant mix or a different collection process.
Analysis: split cohorts by a small number of material drivers—product, acquisition channel, risk band, term and policy version—before producing dozens of tiny segments. Compare adequate sample sizes and show observation counts. Very recent vintages have incomplete outcomes; do not extrapolate an entire lifetime curve from a few early missed payments without uncertainty bounds.
Turn observation into a decision
A useful review links the chart to actions. If deterioration is concentrated in a channel with a changed verification process, investigate that process and consider a controlled pause or tighter review while evidence develops. If all segments deteriorate together, a macroeconomic or servicing explanation deserves more weight.
The OCC also describes lagged analysis, which compares current losses or delinquencies with an earlier balance denominator. This can help expose growth effects, but the selected lag matters. It should correspond to the product’s loss emergence rather than being chosen to make a preferred result appear. Vintage, lagged and contemporaneous measures work best together.
What would strengthen or weaken the conclusion
Confidence improves when several cohorts show the same equal-age pattern, source data reconcile, and the result survives controls for borrower mix and calendar conditions. Confidence weakens if definitions change midstream, loans disappear after sale, recoveries are inconsistently assigned or the apparent difference comes from a tiny population.
Vintage analysis is a diagnostic framework, not a causal experiment or a valuation model. Use it alongside cash-flow economics, loss forecasts and model validation. Preserve the definition and data cutoff with every version so a later “improvement” cannot be created simply by changing the chart’s denominator or observation window.