What Vintage does not do
Vintage measures how your loans have actually behaved and prices new loans from that history. Every number is meant to be re-derivable by hand from your own data. Keeping that promise means drawing some lines on purpose. Each is listed below with its reason and, where there is one, what to do instead.
These are boundaries of the method, not limits of your data.
It measures history; it does not forecast the economy
Section titled “It measures history; it does not forecast the economy”No machine learning, no fitted distributions, no hidden parameters. The method is historical averaging by loan age over a cohort (a group of loans that behave alike): a rate at each age, measured from your loans and accumulated as they season. Vintage does not fit a statistical distribution to your losses or train a model on them. The few fixed rules that carry a curve past the edge of your data are published with the method; see Prepayment speeds and Vintage analysis.
No macroeconomic or qualitative overlay. There is no forecast adjustment and no qualitative (Q-factor) layer on top of the historical experience. The curves show what your book did, as measured. If your process calls for a forward-looking adjustment, apply it outside Vintage, on top of them.
No rate-incentive prepayment model. The prepayment speed is the cohort’s historical speed. There is no refinance S-curve that speeds prepayment up when market rates fall below a loan’s note rate, or slows it when they rise. The speed reflects the rate environment in which your cohort’s history was earned. See Prepayment speeds.
It models two behaviors, for one segment at a time
Section titled “It models two behaviors, for one segment at a time”Two behaviors, not three. Vintage models credit loss and prepayment. Full payoffs and partial prepayments are folded into one prepayment measure, because both are principal returned ahead of schedule.
A defined segment, not the whole book. The Modeling screen shows results once you have narrowed the portfolio to a cohort of loans that behave alike: at least a single product type, and usually narrower still. Averaging auto loans against mortgages produces a curve that describes neither. Define the segment on the Segment Builder; see Segments.
A maximum segment size. A segment is modeled up to a maximum number of loans that belongs to your organization and depends on how many months of history your data spans; the screen names your number when a segment is too large. A cohort large enough to exceed it is also large enough that its averaged curves would describe no real group of loans. See The Modeling screen.
Balance-weighted, not count-weighted. All rates and curves weight each loan by its balance, the convention used in current expected credit loss (CECL) work and in mortgage-market (SIFMA) analytics. A $2 million loan moves a curve more than a $20,000 one.
It never measures against a schedule it made up
Section titled “It never measures against a schedule it made up”No reconstructed schedule in any measurement. Nothing Vintage measures from your history is computed against an amortization schedule it built. Prepayment is measured only from figures you reported: a prepayment amount, actual against scheduled principal, or actual against scheduled payment totals. The speed’s denominator is the balance at the start of the month less the scheduled principal you reported for it, never a scheduled figure Vintage computed. Where a reported figure is missing, the loan or the month is excluded from that measure and counted; it is never replaced by a modeled stand-in. What to do instead: report Scheduled Principal Due or Paid, or both payment totals. See Missing and unreported data.
Projections assume level payments. Anything Vintage carries forward past the edge of your data amortizes on a level-payment schedule: the projected part of the loss curve, the weighted-average remaining maturity (WARM) estimate used for a cohort too thin for a curve, the required-rate build-up, and the month-by-month projection. Vintage has no loan-type field, so an interest-only or balloon loan is projected as if it amortized, and its balance would in reality stay outstanding longer than shown. Every one of these figures states the assumption beside it and in the downloaded file, so a measured figure can always be told from an assumed one. Read any projection for an interest-only or balloon book as an understatement of how long its balance stays outstanding, and of the interest, funding cost and credit loss that come with it.
One term for the projected tail. The observed part of every curve respects each loan’s own term. The projected part uses one representative term for the cohort, its balance-weighted term. Vintage does not model separate term groups inside one slice. What to do instead: slice by term, so the loans in a segment were written for similar lengths.
It shows measured speeds, not a projected survival curve
Section titled “It shows measured speeds, not a projected survival curve”No remaining-balance (survival) curve on the Modeling screen, and no cohort expected life. A survival curve is a projection, one representative loan amortized on an assumed schedule, and shown beside measured speeds the two would be hard to tell apart. The per-age speeds carry the run-off story on the measurement side. The loan’s expected life lives in Pricing, as the adjusted term of the specific loan you price, where it includes defaults as well as prepayments. See Loan pricing.
Its pricing makes two standard simplifications
Section titled “Its pricing makes two standard simplifications”Funding from your par rates, not a bootstrapped curve. Vintage funds the loan as a blended strip of the rates you quote on your funding (FTP) curve (the funds-transfer pricing curve: your cost of funds at each term), read at each month’s term. It does not bootstrap a zero-coupon curve from them (derive the implied rate for a single payment at each future date). At the terms community banks and credit unions price, the difference is well under a basis point (a hundredth of a percentage point), and quoted par rates are what a treasury desk actually publishes.
Equity credited at the funding rate. On the return-on-equity (RAROC) path, the share of the loan funded by allocated capital is credited back at the same blended funding rate the loan is charged, not at a separate reinvestment rate. A second rate would be one more number to source and defend, for a second-order effect on a single loan.
It does not guess what your data does not say
Section titled “It does not guess what your data does not say”Vintage normalizes and reconciles the facts you supply, and never manufactures a fact that would change what a curve or a price means. In practice:
- It never guesses what a status or flag value means.
CO,Yor1marks an event only when you classified it as one. See Classifying values. - It does not age a loan it cannot place. A loan in your earliest month with no Origination Date is excluded, not estimated. See Origination and term.
- It does not treat a blank as zero. See Missing and unreported data.
- It does not show a zero it did not measure. When no charge-off amount is supplied anywhere in your accepted uploads, lifetime loss reads as a dash, not 0%, and credit loss is left out of a price rather than priced at zero.
Rows without a usable Loan ID are not guessed into a loan. Loan ID is the primary key that ties rows to a loan; a designated Tax ID is scrambled the same way and serves as a fallback join key. A row with neither usable identifier is kept as source data but is not linked to any loan, and Vintage never infers which loan it belonged to from its other values. See Removing personal information.