Vintage and Seasoning in Mortgage Credit Risk Modeling

The riskiest year of a mortgage's life is not year one. On a typical mortgage cohort, default hazard starts low, rises through the first few years, peaks somewhere in years three through five, and declines steadily after that. This pattern is the seasoning effect. Together with vintage, it makes up the two time axes every mortgage credit risk model has to handle.

Vintage and seasoning: the two time axes of a mortgage

Every mortgage carries two time attributes that matter for credit risk. They are not the same thing, and you cannot swap one for the other.

• Vintage is the year, often the quarter, the loan was originated. It is set at origination and never changes. A 2019 Q1 mortgage is still a 2019 Q1 mortgage decades later. Vintage is a permanent cohort identifier.
• Seasoning is loan age: the number of months the loan has been on book since origination. It changes every month. A loan with 36 months of seasoning today will have 48 months a year from now, whatever its vintage.

Vintage tells you what world the loan was born into. Seasoning tells you how long it has been alive in whatever world it is in now.

The smallest meaningful cohort is a group of loans from the same vintage observed at the same loan age. Modeling work always has to balance cohort tightness, meaning many small and homogeneous boxes, against statistical power, meaning fewer and bigger boxes with more loans in each.

Why new loans are almost the safest

The natural first guess is that brand-new loans are the riskiest, because the underwriting screens have not been tested and the borrowers have not proven themselves. That guess is wrong. The canonical seasoning curve looks like this:

1. Months 0 to 12: very low hazard. The underwriting screens are recent, the borrower's circumstances have not had time to change, and the equity position is at its strongest.
2. Months 12 to 36: rising hazard. Life events pile up: job changes, divorces, unexpected expenses and income shocks. The borrower's financial buffer gets tested, and loans that will eventually default start to do so.
3. Months 36 to 60: the seasoning peak. Life events have had time to compound, equity built through amortization is still modest, and the cushion from underwriting has been heavily tested. More mortgages default in their fourth and fifth year than at any other point.
4. Months 60 onward: steady decline. The loans still performing are increasingly selected for stability, and their borrowers have built meaningful equity.
5. Months 120 onward: low and stable. Defaults still happen, but at a small fraction of the peak rate. Many of these loans will prepay through refinance or sale well before maturity.

The exact shape varies. Crisis-era vintages have much higher peaks than recovery-era vintages. High LTV loans peak higher than low LTV loans. Investment property loans have different shapes than owner-occupied loans. The macro environment shifts the whole curve up or down. Still, low early, a peak at three to five years, and a decline after that is what the field expects on a representative US conforming mortgage cohort.

The peak location is remarkably stable across credit quality. Loans with low credit scores have higher hazard at every age, but their peak still sits around years three through five. The curve moves up or down with credit risk. It does not shift left or right.

Two things drive the declining tail. The first is survivor selection: a five-year-old loan that is still paying has survived the peak, and the loans that were going to default mostly already have. The second is equity built through amortization and home price appreciation. A borrower with substantial equity has a strong reason to keep paying through temporary income shocks, because defaulting would lock in the loss of that equity.

Why vintage matters

Take two five-year-old mortgages with the same credit score, LTV at origination, state and product. One was originated in 2005 and the other in 2015. They are not the same risk. The worlds they were born into differ in ways the borrowers' individual attributes do not capture, and vintage encodes those differences.

• Pre-2005: conservative underwriting compared with what followed, modest leverage, and slow, steady home price appreciation. These vintages came through the 2008 crisis with elevated but not catastrophic default rates.
• Crisis vintages, 2005 to 2007: underwriting was at its loosest. There were no-documentation and stated-income loans, very high LTV and combined LTV ratios, lower credit score thresholds in some channels, and home prices at their peak. These cohorts suffered catastrophic defaults in the 2008 to 2012 housing collapse.
• Recovery vintages, 2008 to 2014: underwriting tightened. GSE risk management policies tightened, and the Dodd-Frank ability-to-repay and qualified mortgage rules took effect in 2014. Performance has been very strong, with low default rates and high prepayment as borrowers refinanced during the long low-rate period.
• Modern vintages, 2015 to 2019: underwriting stayed tight in the conforming market and home prices rose steadily.
• Pandemic-era vintages, 2020 to 2021: record low rates drove historic refinance volume. During COVID, forbearance, payment deferrals and foreclosure moratoriums kept many delinquencies from turning into realized defaults and losses from 2020 to 2022. A naive PD model trained on that period will systematically underestimate baseline default propensity.
• Recent vintages, 2022 to 2025: rapid rate increases ended the refinance boom, and the incentive to refinance is much smaller or gone. These vintages are only now building up enough seasoning to show meaningful patterns.

How the two axes interact in PD models

The simplest approach is one baseline seasoning curve plus a vintage fixed effect that shifts the curve up or down. That is a reasonable first approximation, but it leaves things out.

Each vintage reaches its seasoning peak in a different macro environment. A 2005 Q1 loan peaks around 2008 to 2010, at the depth of the housing crisis. A 2012 Q1 loan peaks around 2015 to 2017, a much more benign period. A 2019 Q1 loan peaks around 2022 to 2024, a time of post-COVID forbearance and rising rates. So the shape of a vintage's curve depends on the macro path during its peak window, not only on its credit attributes at origination. Time-varying macro covariates can capture this, and lifetime loss estimates can be made conditional on a macro scenario.

Weak-underwriting vintages also carry higher compositional risk. When macro stress hit during their peak, it multiplied that risk into catastrophic losses. Separating the compositional effect from the macro effect is a real modeling decision.

A model that ignores either axis will be biased. If it ignores loan age, it treats a brand-new loan near the seasoning trough the same as a five-year-old loan at peak hazard, and it misranks loans by age. If it ignores vintage, it treats a 2005 loan and a 2015 loan at month 60 the same, even though their macro paths, borrower populations and underwriting standards were very different.

Three common approaches all use this two-axis structure, each in its own way:

• Survival models: Cox-style models include loan age as a time-varying covariate. Vintage enters either as a stratification variable or as a fixed effect.
• Vintage curves: cumulative loss is tracked by seasoning separately for each vintage. Combined with a macro scenario overlay and qualitative adjustments, this structure supports lifetime expected loss estimation.
• Roll rate models: these track transitions between delinquency states, such as current to 30 days, with probabilities that vary by seasoning and vintage.

A real seasoning curve: the 2019 Q1 cohort

The example uses a slice of 2019 Q1 California loans observed through September 2025, which covers seasoning from zero to about 78 months. The cohort's median credit score is around 749 and its median LTV around 75. The hazard proxy flags the first month each loan reached 90+ days past due, bins loan age into six-month intervals, and divides events by loans at risk in each bin.

The curve was not canonical. Hazard was low in months 0 to 11, spiked to over 3 percent per six-month interval in months 12 to 17, and then declined steadily.

The cause is COVID forbearance. For this cohort, months 12 to 17 line up roughly with January through June 2020. GSE COVID forbearance began in March 2020 and spread quickly through the rest of that window. Forbearance often shows up in delinquency data as 90+ days past due, even though the underlying credit risk means something different.

The canonical seasoning shape is a useful prior, not a guaranteed result. To interpret any seasoning curve, you need to know what the cohort lived through.

How to read any seasoning curve or risk number

When you see a seasoning curve in a paper, a validation document or a benchmark, first ask what cohort it was estimated on and what macro environment that cohort went through. A curve from 2005 to 2007 vintages over 2008 to 2014 will not look like one from 2010 to 2012 vintages over 2013 to 2019.

The same goes for any mortgage risk number, whether it is a default rate, a loss severity or a prepayment speed. Ask which cohort it comes from, at what loan age, and through what macro environment. Without those three answers, the number is easy to misread.

Key takeaways

• Vintage is fixed at origination, and seasoning goes up every month. A model that ignores either one produces biased risk estimates.
• Default hazard is low early, peaks in years three through five, and then declines because of survivor selection and equity buildup.
• Riskier loans have higher seasoning curves, not earlier peaks.
• Vintage eras differ in ways that borrower attributes do not capture.
• A vintage's curve depends on the macro path during its peak window.
• On a 2019 Q1 cohort, COVID forbearance moved the 90+ day hazard peak to months 12 to 17.
• For any risk number, ask which cohort, at what loan age, and in what macro environment.