Vintage analysis

Vintage analysis is a credit risk technique that groups loans by the period in which they were originated and tracks each group's performance by age, rather than by calendar date. Comparing cohorts at the same number of months on book removes the distortion caused by portfolio growth, and exposes changes in credit quality months before headline arrears ratios move.

Key takeaways

• A vintage is a cohort of loans disbursed in the same period — usually a month or a quarter.
• Performance is measured by months on book (MOB), so a loan three months old is only ever compared with other loans three months old.
• A growing portfolio automatically dilutes its own arrears ratio, because young loans have not had time to default. Vintage analysis is the standard correction for this.
• The output is a vintage curve: cumulative bad rate plotted against MOB, one line per cohort.
• Its main uses are early detection of credit deterioration, evaluation of policy or scorecard changes, comparison of channels and products, and lifetime loss estimation for expected credit loss models.

What is vintage analysis?

Every loan has two clocks. One is calendar time — the date on the statement. The other is age, or how long the loan has been outstanding. Credit risk follows the second clock, not the first. A loan disbursed last week cannot be ninety days past due; a loan disbursed two years ago has had every opportunity to fail.

Conventional portfolio reporting mixes the two. A ratio such as portfolio at risk over 30 days is calculated across every loan on the book on a given date, regardless of age. That number answers "how is the portfolio doing right now," which is useful, but it cannot answer "is what we are originating today better or worse than what we originated last year." Those are different questions, and only the second one tells you whether your credit policy is working.

Vintage analysis answers the second question by holding age constant. Group the loans by disbursement month, then measure each group at MOB 3, MOB 6, MOB 12 and so on. Every comparison is like for like.

Why growth hides bad lending

This is the specific failure vintage analysis exists to prevent.

A lender doubling its book each year has a portfolio in which most loans are young. Young loans are almost never delinquent — not because they are good, but because they have not had time to go bad. The denominator of the arrears ratio grows immediately while the numerator only responds with a lag of several months.

The result is a portfolio at risk ratio that stays flat, or even improves, while underwriting quality is collapsing. When growth slows, the accumulated bad loans surface all at once and the ratio deteriorates sharply, often blamed on the economic environment rather than on origination decisions made a year earlier.

Vintage analysis makes this visible while it is still happening.

The vintage table

The standard output is a triangle: origination cohorts down the rows, months on book across the columns. The lower-right of the table is empty because recent vintages have not yet aged.

Cumulative 90+ days past due rate, by origination month and months on book

• January: MOB 3: 0.8% | MOB 6: 2.1% | MOB 9: 3.4% | MOB 12: 4.0%
• February: MOB 3: 0.9% | MOB 6: 2.3% | MOB 9: 3.6% | MOB 12: 4.2%
• March: MOB 3: 1.1% | MOB 6: 2.8% | MOB 9: 4.3%
• April: MOB 3: 1.6% | MOB 6: 3.9%
• May: MOB 3: 2.1% | MOB 6: 4.8%
• June: MOB 3: 2.4%

Read down any single column. At MOB 3, the bad rate has tripled between January and June. At MOB 6, it has more than doubled between January and May. Something changed in the business around March or April — a policy loosening, a new channel, a scorecard cutoff change, a fraud ring, or a new branch onboarding weak borrowers.

Now consider what the portfolio-level arrears ratio would have shown over the same period. If the book was growing quickly, very little. The May cohort's 4.8% at MOB 6 does not fully surface in headline numbers until late in the year.

Reading the table in three directions:

• Down a column — compares cohorts at the same age. This is the comparison that matters for credit quality.
• Across a row — traces one cohort's maturation, showing how losses accumulate with age.
• Along a diagonal — represents a single calendar period, since each step right and down moves forward one month in real time. Diagonal patterns indicate a shock affecting all cohorts at once, such as a macroeconomic event, a collections disruption, or a payment channel outage.

That last distinction is important diagnostically. A problem visible down the columns originates in underwriting. A problem visible along a diagonal originates in the environment or in servicing, and affects loans that were fine when they were written.

Vintage curves

Plotting each row as a line — cumulative bad rate on the vertical axis, MOB on the horizontal — produces the vintage curve chart that gives the technique its visual signature.

Well-behaved curves share a shape: steep early rise, then gradual flattening as the cohort seasons. The point at which the curve flattens is the maturation point, and it defines how long a cohort must be observed before its final loss rate can be estimated reliably. For short-term consumer and microloans this may be six to twelve months; for longer-term secured lending it can be several years.

Two properties of the curves carry meaning:

• Height — where the curve sits. A higher curve means a worse cohort at every age.
• Shape — how quickly it rises. Curves that rise very steeply in the first few months point to origination problems: fraud, misrepresented income, or borrowers who never intended to repay. Curves that stay low then deteriorate later point to affordability problems that emerge as circumstances change.

First payment default — a borrower missing the very first instalment — deserves separate tracking. It is almost never a credit event in the ordinary sense. It usually indicates fraud, identity misuse, a data error, or a loan officer originating loans the borrower never actually received.

What to measure

The technique is indifferent to which metric you track, and the choice should follow the question being asked.

• Cumulative % ever 30 / 60 / 90+ days past due: Early risk signal, available soonest.
• Cumulative default rate: Formal credit outcome defined by the lender's default definition.
• Cumulative net loss rate: Write-offs less recoveries — the true economic outcome.
• Point-in-time delinquency rate: Snapshot rather than cumulative; noisier but simpler.
• First payment default rate: Evaluates fraud and origination integrity.
• Prepayment / early settlement rate: Measures revenue impact and behavioural loan life for pricing models.

Two methodological choices matter more than they appear to:

Count-weighted or balance-weighted? Measuring by number of loans treats every borrower equally; measuring by disbursed amount reflects economic exposure. They can tell different stories when loan sizes vary widely — a cohort can look fine by count and terrible by value if the large loans are the ones failing. Report both where possible.

Which denominator? For cumulative loss rates, the denominator should be the original disbursed amount of the cohort, held constant. Using the current outstanding balance reintroduces exactly the distortion the technique is designed to remove, because the balance amortises down as good loans repay.

Where vintage analysis is used

Credit policy evaluation. The only reliable way to know whether a policy change worked. Tighten a cutoff in March, and the March-onward vintages should sit below the earlier ones at equal MOB. Nothing else isolates the effect this cleanly.

Scorecard monitoring. Vintage curves split by score band should be ordered and separated. If bands converge or cross, the model has lost discriminatory power and needs recalibration.

Channel, product and branch comparison. Cohorts segmented by acquisition channel, branch, loan officer or product frequently reveal that aggregate performance is being carried by one strong segment while another is losing money.

Campaign and promotion assessment. Loans written under a promotional push form their own vintage and can be assessed on their own terms rather than blended into the book.

Expected credit loss modelling. Vintage curves are a direct empirical input to lifetime loss estimation under IFRS 9, since they describe how losses emerge over the life of a cohort.

Pricing and profitability. Expected lifetime loss by segment, taken from mature vintage curves, is the loss component of a risk-based price.

Forecasting. Immature cohorts can be projected forward by applying the shape of comparable mature curves, giving an early estimate of where a recent vintage will end up.

How vintage analysis compares to other techniques

• Vintage analysis: Tracks origination cohorts by age. Best for comparing underwriting quality over time. Limitation: Requires cohorts to season.
• Roll rate analysis: Tracks movement between delinquency buckets month to month. Best for short-term collections forecasting. Limitation: Does not reflect origination quality.
• Flow rate / transition matrix: Evaluates the probability of moving between risk states. Best for modelling near-term migration. Limitation: Assumes transition relationships remain stable.
• Portfolio at risk snapshot: Measures the whole book at a single point in time. Best for assessing current portfolio condition. Limitation: Distorted by portfolio growth and mix shifts.
• Static pool analysis: Tracks a fixed pool of accounts over time. Best for securitisation and pool performance. Limitation: Conceptually identical to vintage analysis under a different name.

These are complements. Roll rates tell you what will happen next month; vintage curves tell you whether the loans you are writing today are worth writing.

Data and practical requirements

• A disbursement date on every loan, immutable, that defines cohort membership permanently. A loan never changes vintage, including after restructuring.
• Historical performance snapshots, or a transaction history complete enough to reconstruct each loan's delinquency status at every past month-end.
• A stable definition of "bad" applied identically across the whole series. Changing the default definition mid-series invalidates every comparison.
• Sufficient cohort size. Small cohorts produce noisy curves. Where monthly volumes are low, group into quarters.
• Consistent treatment of restructures and write-offs. A restructured loan that stops being reported as delinquent will flatter its vintage. Decide the rule, document it, apply it throughout.
• Segmentation keys captured at origination — product, channel, branch, officer, score band. Segments cannot be analysed retrospectively if the attributes were never recorded.

Common mistakes

Comparing immature cohorts to mature ones at their final values. A cohort at MOB 4 will always look better than one at MOB 18. Only equal-age comparisons are valid — this is the entire point of the method, and it is still the most common error.

Using outstanding balance as the denominator. Reintroduces the growth distortion. Hold the original disbursed amount constant.

Ignoring mix shift. A cohort can deteriorate simply because its composition changed — more of a riskier product, more of a weaker channel — without any single segment worsening. Segment before concluding that underwriting has slipped.

Ignoring seasonality. In agricultural and school-fee-driven lending, cohorts originated in different months face genuinely different repayment environments. Compare like months across years as well as consecutive months.

Over-reading small cohorts. A few defaults in a small cohort produce dramatic percentage swings that mean nothing.

Stopping at the portfolio level. The aggregate curve is the average of segments that may be moving in opposite directions.

Treating a diagonal effect as an origination problem. If every cohort deteriorates in the same calendar month, the cause is external or operational, not underwriting.