Vintage Analysis of a Loan Portfolio: A Lender's Guide
Vintage Analysis of a Loan Portfolio: Spotting Bad Cohorts Before They Hit Your P&L
Last reviewed: November 2026
Vintage analysis of a loan portfolio groups loans by the month or quarter they were disbursed, then tracks each group's delinquency or loss at the same months on book. Because every cohort is measured at the same age, it shows whether the loans you made recently are better or worse than the ones you made before.
This guide is for CROs, CFOs and Heads of Analytics at NBFCs, banks and fintech lenders. It explains how to build vintage curves, how to read them, how they connect to expected credit loss, and the data discipline the analysis depends on.
Key Takeaways
- Portfolio ratios suffer from the denominator effect: fast growth can keep NPAs flat while new cohorts deteriorate.
- The standard vintage measure is ever-30-plus or ever-90-plus DPD as a share of amount disbursed, by months on book.
- Frozen monthly loan snapshots are the foundation, because rebuilt history hides real account movement.
- Vintage curves feed Ind AS 109 ECL estimates, and commercial banks move to ECL provisioning from 1 April 2027.
Why Do Portfolio Averages Hide Bad Cohorts?
Portfolio averages blend old loans with new ones that have not had time to default. A growing book keeps adding clean-looking young loans to the denominator.
Consider an illustrative book that grows from ₹400 crore to ₹520 crore in a year, while its 30-plus DPD balance rises from ₹10 crore to ₹13 crore. The ratio is 2.5% at both dates, because ₹10 / ₹400 and ₹13 / ₹520 both equal 2.5%.
The board sees stability. Vintage analysis may show that the newest cohorts are going bad far faster than the old ones did.
How Do You Build a Vintage Curve?
You build it by fixing cohorts, a metric and a time axis, then filling a triangle month by month. Each design choice changes what the curve can tell you.
- Define cohorts. Group loans by disbursement month, or by quarter if monthly volumes are small.
- Choose the event. Common choices are ever 30-plus DPD, ever 90-plus DPD, or cumulative net loss.
- Fix the denominator. Use the original amount disbursed, not today's outstanding balance.
- Align by months on book. Month 1 is the first month after disbursement, regardless of calendar date.
- Fill the triangle. Older cohorts have more observed months, so the table has a stepped shape.
- Segment. Repeat by product, channel, geography and underwriting policy version.
| Design Choice | Options | Practical Recommendation |
|---|---|---|
| Cohort period | Month or quarter | Monthly for retail, quarterly for small books |
| Event | Ever 30+, ever 90+, current DPD, net loss | Ever 30+ for early warning, ever 90+ and loss for provisioning |
| Measure | Count or amount | Amount for P&L, count for underwriting diagnostics |
| Denominator | Disbursed or outstanding | Disbursed, so prepayment does not distort it |
| Restructured loans | Keep or exclude | Keep and flag them, never silently drop them |
| Write-offs | Exclude after write-off or keep | Keep as bad in cumulative curves |
"Ever" metrics are stable because cures do not erase them. Current-DPD curves are still useful for collections, but they swing with collection intensity.
What Does a Vintage Triangle Show? An Illustrative Example
This illustrative example uses assumed figures. An NBFC tracks ever-30-plus DPD as a share of amount disbursed for three quarterly cohorts of personal loans.
| Cohort (Disbursed) | Month 3 | Month 6 | Month 9 | Month 12 |
|---|---|---|---|---|
| January 2026 (₹100 crore) | 0.8% | 1.9% | 2.6% | 3.1% |
| April 2026 (₹110 crore) | 1.0% | 2.4% | 3.4% | not yet observed |
| July 2026 (₹120 crore) | 1.5% | 3.3% | not yet observed | not yet observed |
Take the July cohort at month 6. If ₹3.96 crore of its ₹120 crore disbursed has ever been 30-plus DPD, the rate is ₹3.96 / ₹120 = 3.3%.
Compare cohorts at the same age. At month 6, July's 3.3% is about 74% higher than January's 1.9%, because 3.3 / 1.9 = 1.74.
If the July cohort follows January's shape, its month-12 rate would be about 3.1% x 1.74 = 5.4%. That is a projection, not a forecast, but it is enough to question what changed in July.
How Should You Read the Shape of a Curve?
The shape often points to the cause. These patterns are diagnostic prompts, not proof, so confirm each one with loan-level review.
- A steep jump in months 1 to 3: often first-payment default, which can signal fraud, identity issues or weak onboarding checks.
- A parallel upward shift: the whole cohort is riskier, which usually traces back to underwriting cut-offs or sourcing mix.
- Late divergence after month 9: cohorts start alike and separate later, which can reflect seasoning, economic stress or a regional event.
- Curves that flatten early: most losses arrive soon after disbursement, typical of short-tenor unsecured loans.
Overlaying the policy-change dates on the chart makes these patterns far easier to explain to a risk committee.
The lesson for leadership: the warning appeared three months after disbursement, long before the loans reached 90 days past due. A portfolio NPA ratio would not have shown it.
What Do Vintage Curves Tell the CRO and CFO?
They tell the CRO which decisions created risk and the CFO what that risk will cost. Both depend on tagging every cohort with what changed.
- Underwriting changes: a new cut-off or a relaxed policy should show up as a step in later cohorts.
- Channels and partners: a DSA or fintech partner with steeper curves needs tighter criteria or exit.
- Pricing: cohorts that lose more than their risk premium covers are unprofitable even if they grow volume.
- Provisioning: curves feed probability of default estimates for expected credit loss.
Under Ind AS 109, there is a rebuttable presumption that credit risk has increased significantly once payments are more than 30 days past due, and that default has occurred by 90 days. Cohort default curves help estimate 12-month and lifetime loss for each stage.
NBFCs that report under Ind AS already apply these rules alongside RBI's prudential norms on asset classification. The RBI's final ECL directions of 27 April 2026 bring in-scope commercial banks onto expected credit loss provisioning from 1 April 2027.
Digiqt's expected credit loss intelligence AI agent supports provisioning analysis. A credit portfolio stress testing AI agent can test how cohorts behave under adverse scenarios.
Can your risk team show, cohort by cohort, which decisions are driving next year's credit cost?
Talk to Our Lending Technology Specialists
Visit Digiqt to see how we build data pipelines and AI agents for lenders.
What Data and Systems Does Vintage Analysis Need?
It needs loan-level history that nobody can quietly rewrite. Most failures come from data, not from the method.
- Frozen month-end snapshots of every loan from the loan management system.
- Origination attributes: product, channel, partner, branch, score band and policy version.
- Status flags: restructuring, moratorium, write-off, settlement, foreclosure and sale.
- A data warehouse that keeps snapshots immutable and versioned.
- Dashboards that let users switch cohort, metric and segment without new code.
Snapshots matter most. If you rebuild history from today's data, accounts that cured or were restructured disappear from earlier months, and old cohorts look better than they were.
Early-warning tools add value on top of cohort views. An early delinquency warning AI agent flags individual accounts, while vintages show whether the problem sits in a whole cohort.
Which Vintage Analysis Mistakes Mislead Lenders?
Each of these makes weak cohorts look stronger than they are.
- Using outstanding balance as the denominator. Prepayments shrink it and inflate delinquency rates unpredictably.
- Judging cohorts too early. Month-2 differences can be noise, especially for small cohorts.
- Mixing products in one curve. A shift in product mix looks like a change in credit quality.
- Dropping restructured accounts. Restructuring resets DPD, so excluding these loans hides real stress.
What Is Changing Next?
The move to ECL for commercial banks from 1 April 2027 will raise expectations for cohort-level default data across the system. Co-lending and partnership arrangements may see more demand for cohort-level evidence from bank partners.
The RBI also issued draft guidance on model risk management in June 2026, covering all models, including third-party ones. Projection tools built on vintage curves could fall within its definition of a model once the guidance is final.
Want vintage curves that update themselves every month from clean, frozen data?
Talk to Our Lending Technology Specialists
Visit Digiqt to see how we build lending platforms and portfolio dashboards.
Sources
- RBI: Press release on final ECL-based provisioning directions for commercial banks, 27 April 2026
- RBI: NBFC Income Recognition, Asset Classification and Provisioning Directions, 2025
- IFRS Foundation: IFRS 9 Financial Instruments, the basis for Ind AS 109
- RBI: Draft guidance on model risk management, 24 June 2026
