Bottom line
Forty-four public auto securitizations file a loan-by-loan tape every month under Form ABS-EE. That is 907,334 loans across 43 trusts and fourteen lenders, prime captives and deep subprime shelves side by side, each loan carrying its score at origination, its days past due, its extensions and its charge-off. Nothing in this paper is a published ratio. Every figure is counted from those individual records.
Counting them that way says the risk in subprime auto is concentrated in one place, and it is not where the index points. Borrowers with identical credit scores reach serious delinquency at wildly different rates depending on who lent to them. Once they get there, almost every lender's outcome converges. Entry is where the dispersion lives; exit is close to a constant.
The read
A credit score is supposed to be the thing that tells you how a borrower behaves. In these tapes it is a weak predictor of anything once you know the lender's name. A borrower scored between 661 and 780, which every bureau and every rating agency calls prime, goes 60 days past due within two years about 1.1% of the time at Ford and 29.3% of the time at Santander. Same band, same country, same two-year window, twenty-eight times the failure rate.
Then the convergence. Among loans that did reach 60 days down, the share charged off within the next twelve months runs 68.1% at Exeter, 68.9% at Santander, and 67.8% at World Omni, which is a Toyota-affiliated prime captive. Three lenders at opposite ends of the credit spectrum land within 1.1 points of each other. Whatever separates a prime book from a subprime one, it stops operating at the moment a borrower is two payments behind.
If exit odds are near constant, then loss severity is mostly fixed and the whole business reduces to controlling how many loans arrive at that door. That is an underwriting and servicing problem, not a borrower-quality problem, and it is why two lenders can hold the same credit band and write books that behave nothing alike.
I · Entry and exit
Here are the two distributions on one scale. The left column is how often a prime-band borrower reaches 60 days past due by month 24, one dot per lender. The right column is what share of loans that got there were charged off within a year.
The left column spans twenty-eight to one. The right column spans one and a half to one, and the ordering inside it does not sort by tier: CarMax, a prime shelf, sits at the bottom with 56.1%, and World Omni, also prime, sits at 67.8%, above AmeriCredit's subprime book at 57.5%. Bridgecrest is the one genuine outlier at 86.1%.
A borrower holding a 700 score who ends up at Exeter is a borrower the captives declined, for reasons the tape cannot see: payment-to-income, down payment, time in file, a recent derogatory. Some unknown part of that 27.9x is selection rather than lender behaviour. The tape carries no income and no debt-to-income field, so I cannot size it. What the tape does establish is that the score alone, which is what most of the public commentary leans on, carries far less information than its prominence suggests.
II · Same score, different lender
The full grid, read across a row for one lender over the credit spectrum and down a column for one score band across lenders.
| Lender | 300-540 | 541-600 | 601-660 | 661-780 | 781-900 |
|---|---|---|---|---|---|
| Ford | · | · | · | 1.1% | 0.1% |
| Hyundai | · | · | · | 1.6% | 0.1% |
| Ally | · | · | · | 2.8% | 0.5% |
| GM Financial | · | · | · | 2.9% | 0.4% |
| Carvana | · | 10.6% | 7.0% | 3.0% | 0.7% |
| CarMax | · | 23.3% | 12.4% | 3.1% | 0.4% |
| World Omni | · | · | 14.7% | 4.2% | 0.7% |
| AmeriCredit | 26.4% | 19.8% | 13.3% | 19.4% | · |
| Exeter | 42.3% | 36.4% | 29.8% | 24.1% | · |
| Santander | 41.7% | 35.7% | 35.0% | 29.3% | 8.7% |
Read Santander's row from left to right. Its deep subprime band runs 41.7% and its prime band runs 29.3%, a difference of twelve points across four hundred points of credit score. Now read the 661-780 column from top to bottom: 1.1% to 29.3%, twenty-eight points, across lenders who all call that paper prime. The column moves more than the row.
III · One lender's improvement, read as the market's
The standard account of the cycle says the 2022 and 2023 books were written at peak used-car prices to borrowers flattered by stimulus-era credit models, that those cohorts are the problem, and that they are amortising away. Consumer Portfolio Services gave the market the cleanest version of it on its Q2 2026 call, disclosing recovery rates by vintage of 22% for 2022, 25% for 2023, 37.5% for 2024 and 47.1% for 2025.
On recoveries, our own tapes agreed; that was Issue 15. On delinquency the tapes disagree, and the disagreement only shows up once you stop pooling lenders.
At month 18 on book, Exeter's cohorts run 32.5% for 2021, 32.3% for 2022, 28.2% for 2023 and 25.4% for 2024. That is the published story exactly: a peak in the pandemic-era books and steady improvement since. AmeriCredit's run 7.6% for 2020, 13.5% for 2021, 13.9% for 2022 and 17.9% for 2024. Its worst cohort is its newest, and the line is still rising. Santander sits between them and essentially flat, 29.4% for 2023 against 28.3% for 2024.
Pooling those three lenders into one curve, which is what an index does, produces a shape that belongs to none of them. In these trusts the 2022 cohort is 69% Exeter with no Santander paper in it at all, and the 2023 cohort is 65% Santander with no AmeriCredit. A year can look better or worse purely because a different lender wrote more of it.
I ran the pooled version first and it reproduced the published story convincingly. Holding the lender fixed is what broke it. Any vintage comparison drawn across lenders whose composition shifts year to year is measuring the composition.
IV · A fifth of the pool is current on an extension
An extension moves a past-due loan back to current on the tape without the borrower catching up. Add back every loan returned to current that way in the trailing six months and subprime 60+ delinquency goes from 8.10% to 26.55%, a gap of 1,844 basis points on the July 2026 tapes. Prime runs 0.73% against 3.91%, a gap of 318. The window is a stated convention, not a filed figure, and the magnitude scales with it: three months gives roughly 1,057 basis points on the subprime tier, twelve months roughly 3,105. The ordering across lenders holds at every window.
Read it July to July, which holds the seasonal position fixed, and the band has widened for three years: 1,098 basis points in 2024, 1,344 in 2025, 1,844 in 2026, with the share of the pool touched by an extension going 11.85% to 14.84% to 21.00%. Those three readings rest on 10, 12 and 10 deals, so the comparison is not resting on a changing pool count, but the month-to-month shape of the line is.
The timing is the part worth keeping. The adjusted measure peaked in January 2026 at 27.86% and the extension stock peaked a month later at 23.09%, while reported delinquency kept climbing until June. Reported delinquency rising through the first half of 2026 is partly the extension stock unwinding and handing loans back.
Whether an extension is a cure or a deferral is visible six months later. Across the tapes, loans back at 60 days down within six months of an extension run 40.4% at Exeter and 39.8% at Bridgecrest, against 18.0% at AmeriCredit, 9.1% at World Omni and 4.0% at Honda. Exeter extends about four loans in ten and four in ten of those come straight back.
V · What the bondholder sees
None of the above has reached a senior noteholder.
| Trust | Pool loss to date | Class A notes | Delinquency | Trigger |
|---|---|---|---|---|
| Exeter 2022-1 | 24.00% | all retired | 14.40% | 40.0% |
| Exeter 2023-1 | 22.81% | all retired | 12.04% | 40.0% |
| Exeter 2024-1 | 17.80% | all retired | 10.48% | 40.0% |
| Santander 2023-6 | 9.54% | all retired | 9.96% | 24.0% |
| Santander 2024-1 | 9.60% | all retired | 9.70% | 24.0% |
Exeter's 2022 trust has written off twenty-four cents of every dollar of original pool principal and every class A and class B holder was repaid at par. Its delinquency trigger sits at 40% and the deal has never been within twenty-five points of it. That is the answer to whether the asset is dangerous, and the answer is that it was priced, structured and tranched so that it would not be, for the people at the top of the stack.
VI · Where it does break
The failures of the last three years did not come through the collateral. American Car Center told staff it was closing in February 2023, the day after pulling a $222 million bond sale. Tricolor filed Chapter 7 in September 2025; federal prosecutors in the Southern District of New York allege it had pledged about $2.2 billion of collateral against roughly $1.4 billion of real loans. PrimaLend, which lent to buy-here-pay-here dealers rather than to borrowers, filed Chapter 11 in October 2025 after vehicle values and delinquencies moved against the collateral behind its dealer lines. Each arrived through the liability side.
Our own perimeter makes the same point by omission. Tricolor never filed a single ABS-EE loan tape, because every one of its deals was privately placed. Neither did Westlake, GLS, Flagship, First Investors or Consumer Portfolio Services, whose vintage recovery disclosure I quoted above and cannot check against collateral. The lender whose book could not be examined is the one that turned out not to have the book.
The listed operator that got closest to trouble on credit alone is America's Car-Mart, the only public pure-play deep subprime buy-here-pay-here lender. Its FY2026 10-K reports net charge-offs of 27.6% of average finance receivables, an allowance at 25.15%, and a provision running 40.8% of sales. It has filed for forty-six quarters and is still operating. In July 2025 it filed a non-reliance 8-K because it had omitted the ASC 310-10-50-42 disclosures covering loan modifications to borrowers in financial difficulty, declared a material weakness, and restated. What it now discloses is that 30.3% of its portfolio was modified at least once during FY2026 and that roughly half of its contracts are modified at some point over their life.
The one listed subprime lender forced to restate was restating about how much of its book it was extending. That is the same measurement the tapes make directly in section IV, arrived at from the opposite direction.
What I would watch
Not the delinquency headline. It mixes a growing share of smaller and newer issuers into the same number, it moves with tax-refund season, and it tells you nothing about severity. Watch entry rates inside a single lender and a single band, which is the only cut where a change means a change in behaviour. Watch the extension stock, because distress parked there is distress that has not been priced yet. And watch whether any long-tenured issuer loses warehouse or term ABS access, because that is the mechanism that has ended companies, and it leaves no trace in a delinquency index until after the fact.
Caveats
Score bands are the score at origination as the issuer filed it. Issuers do not all file the same kind of score and some file internal tier codes in the same field, so unscored loans are excluded from any weighted average rather than counted as zero. A loan that paid off or charged off before a pool's cutoff never appears on any tape, and a tape ends at the clean-up call, which over-represents survivors at long seasoning; the curves here stop where that begins to bite. Extension-adjusted delinquency is a modelling choice with a stated window and is not a filed number. The CPS vintage recoveries come from a call transcript, not a filed statement, and one transcript service renders the 22% figure as a 2020-to-2022 vintage rather than 2022 alone. The Tricolor criminal allegations are allegations until adjudicated. Figures for American Car Center, Tricolor and PrimaLend are drawn from the companies' own filings and from contemporaneous reporting, not from loan tapes, because none of those three filed any.