Why it works the way it does.
Vintage is built for numbers that have to survive a question. Not just producing a figure, but holding up when a board member, an auditor, or an examiner asks how you arrived at it. Everything below follows from that one requirement.
For the people who answer for the numbers.
Vintage is built for community banks and credit unions holding years of loan-level history in their core systems, roughly $250 million to $5 billion in assets. More exactly, it is built for the people who have to answer for what that history should produce:
- The CFOwho signs the reserve and defends it to the board and the examiner.
- The chief credit officerwho owns loss experience across the book.
- The ALM and finance leadswho set the assumptions and price the next loan.
The product is meant to stay legible to them, and to the examiner reading over their shoulder.
A figure comes from your data, or it is not there.
Software that quietly fills a gap with a plausible number is more dangerous than software that leaves the gap showing, because a gap you can see is one you can go and close. Vintage never invents a value that would change what a curve or a price means. Three commitments follow from that, and together they explain why the product looks the way it does.
Methods you can re-derive by hand.
Vintage uses historical cohort averaging: a life-table of per-age rates, weighted by exposure. There is no machine learning, no fitted distribution, and no hidden parameter.
A banker or an examiner can reproduce any figure on the screen from the same files, without taking the software's word for it.
Projections shown as projections.
Where a curve rests on real history it is marked observed. Where it runs past your data it is marked projected, and the confidence band widens as the record thins.
When a cohort is too young for a vintage curve, Vintage says so and falls back to a WARM estimate rather than pretending the curve is solid.
Missing data named, not filled in.
When your data cannot support a number, Vintage shows a dash and names the field that would supply it, instead of guessing.
The inferences it does make, such as reading an origination date from a loan's first appearance, are counted and labeled where you can see them.
Every value traces back to the upload, file, row, and column that supplied it, so any figure on the screen can be followed all the way to its source.
What Vintage does not do.
A short list, published on purpose. You will find these limits eventually, and it is better for both of us that you find them now rather than after a decision.
- No economic overlayVintage measures what your book has done. It does not layer a forecast or a qualitative factor on top of that experience. If your process calls for a Q-factor, that judgment stays yours.
- No rate-incentive modelPrepayment is your cohort's measured historical speed. Vintage does not model how that speed would shift if rates moved.
- Loans it cannot age are set asideA loan already present in your earliest snapshot month, with no origination date, cannot be aged. It is counted and left out of the age-indexed curves rather than guessed at. A two-column file of loan IDs and origination dates brings those loans in.
- One representative term per sliceThe projected tail of a curve assumes a single balance-weighted term. Vintage flags a slice whose terms are too dispersed for that to hold, but it does not yet model separate term cohorts inside one slice.
These are limits of the current method, not of your data. The same history can be re-analyzed under richer methods as they are added.
The argument, made concrete.
The product page follows the same path in detail, from the files you upload to one reconciled history to the modeling screen. If you would rather talk it through first, that door is open too.