Bottom line
Four things, in the order I found them. All of it comes from the 48-trust panel of public auto securitizations, the same loan-month records behind the last five issues, read at the loan level through July 2026.
Charge-offs rose faster than delinquency at every subprime lender I hold, and the gap is widening. Take loans written in 2020 and loans written in 2023, and look at each at 24 months on book. At Santander the share reported 60 or more days past due went from 5.32% to 8.79%, up 65%; the share of original dollars charged off went from 6.21% to 12.68%, up 104%. At AmeriCredit, delinquency up 88% and charge-offs up 162%. At Exeter, delinquency went down, 9.91% to 9.60%, while charge-offs went from 11.16% to 17.93%. A point of reported delinquency now carries 1.44 points of loss at Santander where it carried 1.17, 1.87 at Exeter where it carried 1.13, 2.20 at AmeriCredit where it carried 1.58. The number everyone reads is measuring less of the thing it is supposed to measure, and less each year.
The loss is going around the delinquency bucket, and extensions are the road. In Issue 13 I counted how much of the subprime book was current only because of a recent extension. This time I followed those loans forward. Take every subprime loan that was active and under 60 days past due at the end of July 2025, 365,239 of them, so every one looks clean on the delinquency line. Split them by whether they had been extended in the prior six months. Over the next twelve months the extended ones charged off at 19.65% and the rest at 8.31%, and after controlling for score, loan-to-value, term, months on book, loan size and lender, the odds ratio is 2.60. Throw out every loan that was even 30 days late at the start and it is still 2.43. An extension does not cure the loan. It moves the default past the point where the delinquency rate would have caught it.
It is not survivorship and it is not the term. I decomposed the seasoning curve inside fifteen closed trust cohorts, where no new loans ever enter, into the part that comes from the mix of borrowers changing and the part that comes from the same borrowers getting worse. The mix explains 0.6%. The weighted original term on subprime paper moved from 71.0 months on 2020 originations to 72.7 on 2024, which is the Issue 17 finding again from the other side. The book is aging, and that is real, but it is aging because nothing leaves: loans prepaid by month 24 fell from 13.2% of the 2020 vintage to 5.3% of 2023, and an extended loan still carries two thirds of its original balance at month 24 against well under half for one that was never extended.
I · The rate and the loss came apart
Here is the table that started this. Every row is one lender and one origination year, read at 24 months on book so that seasoning is held equal. Delinquency is the share of outstanding balance 60 or more days past due that month. Charge-off is the share of the cohort's original dollars written off by that month. Both are gross. I show Exeter's 2022 row because it is the worst in the record; Santander's 2022 row is left out because the panel holds only 2,479 of that year's loans, which is too few to stand beside the others.
| Lender · vintage | Loans | Reported 60+ | Charged off, $ | Loss per point | Current but extended |
|---|---|---|---|---|---|
| Santander 2020 | 93,706 | 5.32% | 6.21% | 1.17 | 8.9% |
| Santander 2021 | 84,969 | 6.75% | 8.51% | 1.26 | 10.4% |
| Santander 2023 | 93,152 | 8.79% | 12.68% | 1.44 | 10.9% |
| Exeter 2020 | 63,800 | 9.91% | 11.16% | 1.13 | 25.7% |
| Exeter 2021 | 42,217 | 10.19% | 14.83% | 1.46 | 28.5% |
| Exeter 2022 | 32,909 | 9.88% | 18.33% | 1.86 | 28.4% |
| Exeter 2023 | 37,999 | 9.60% | 17.93% | 1.87 | 28.7% |
| AmeriCredit 2020 | 61,977 | 1.56% | 2.46% | 1.58 | 15.4% |
| AmeriCredit 2021 | 42,723 | 1.78% | 3.65% | 2.05 | 18.3% |
| AmeriCredit 2022 | 48,325 | 2.46% | 5.48% | 2.23 | 20.8% |
| AmeriCredit 2023 | 36,236 | 2.93% | 6.46% | 2.20 | 23.2% |
Read Exeter down. Reported delinquency: 9.91, 10.19, 9.88, 9.60. A flat line with a slight downward slope, the kind of series that gets a lender described as stable. Charge-offs over the same four rows: 11.16, 14.83, 18.33, 17.93. The 2023 loan at two years old is reported as slightly less delinquent than the 2020 loan was and has lost 61% more money. Nothing in the first column predicts the second.
Santander and AmeriCredit tell the same story at a different pitch. Both reported delinquency rising, both reported it rising by less than loss rose, and at both the ratio of loss to reported delinquency climbed through every vintage. At 36 months on book the pattern holds for Exeter and AmeriCredit and the ratio keeps rising, 1.57 to 2.04 at Exeter and 2.06 to 2.98 at AmeriCredit. At Santander the 36-month ratio is flat at about 1.4, because by three years Santander's delinquency has caught up with its losses. I note that because it is the one place in the record where the two lines come back together, and it says something about how long the detour takes.
II · Where the loss went: around the bucket
A loan can only leave the 60-day delinquency count three ways. It can cure, it can charge off, or it can be given a new due date. The third is an extension, and it is the one that costs nothing and changes nothing except the number. In Issue 13 I measured the stock of loans sitting in that state. The question this time is what happens to them next.
The test is built so that every loan in it looks identical on the delinquency line. I took the four subprime shelves with tapes running to July 2026, Santander, Exeter, AmeriCredit and Bridgecrest, and anchored on July 31, 2025. Every loan still active and under 60 days past due that month qualifies: 365,239 loans. Then I split them by one thing the delinquency rate cannot see, whether the loan had received an extension in the six months up to the anchor. 54,791 had, 15.0% of the clean book. Then I watched twelve months forward for a charge-off.
| Group at July 2025 | Loans | Charged off by July 2026 |
|---|---|---|
| Current, not extended in prior 6 months | 310,448 | 8.31% |
| Current, extended in prior 6 months | 54,791 | 19.65% |
That is 2.36 times, raw. The objection writes itself: lenders extend the loans that are already in trouble, so of course those loans default more. Two answers. First, control for everything the tape knows about the loan. In a logistic model with credit score, loan-to-value, original term, months on book, loan size and a fixed effect for each lender, the odds ratio on a recent extension is 2.60, with a 95% interval of 2.53 to 2.66, on 342,806 loans with complete fields. Second, and more to the point, remove every loan that was showing any lateness at all. Of the extended loans, 19.3% were between 30 and 59 days past due at the anchor against 8.7% of the others, so I cut both groups to loans at 0 to 29 days. Those are loans a servicer would call performing. The extended ones charged off at 16.34% over the next year and the rest at 6.72%: 2.43 times, and the controlled odds ratio is 2.56.
It holds in every lender and every credit band. By lender, raw: AmeriCredit 6.45% against 14.32%, Bridgecrest 13.38% against 28.50%, Exeter 9.57% against 24.29%, Santander 7.59% against 16.46%. By score band the ratio runs from 1.97 under 540 to 3.53 at 660 and above, which is to say the extension tells you most about the borrower whose score tells you least.
Put the two sections together and the path is visible. A borrower misses payments. Before the second missed payment turns into a 60-day flag, the servicer grants an extension and the loan reads current again. Some share of those loans then fail, at two and a half times the rate of the loans around them, and arrive in the charge-off line having never passed through the delinquency line at all. The share of each book sitting in that intermediate state is in the last column of the first table: 11% at Santander, 23% at AmeriCredit, 29% at Exeter. The lender with the largest reservoir is the one whose delinquency rate went down while its losses went up.
III · It is not who is left
My first theory for the record delinquency numbers was survivorship, and I want to show the test because it failed cleanly. The idea is that the better borrowers refinance out of a pool as it ages, so the pool that remains is worse, and the rising rate is composition rather than deterioration. It is intuitive, it is partly true, and it does not matter.
A securitization trust is the right place to test it, because it is a closed cohort. The loans are fixed at the cutoff date, nothing new enters, and anything that changes the mix is by definition loans leaving. So inside each of fifteen subprime trusts I wrote the pool's 60-day rate at each age as a weighted average across five credit score bands, and split the change from month 12 to the last observed month into the part explained by the weights moving and the part explained by the rate within each band moving. Summed across all fifteen, the change was 52.4 points of delinquency. The mix explains 0.3 of them. The same borrowers, within the same score band, getting worse as they age explains 51.7. That is 0.6% against 98.7%, with the small remainder an interaction term.
The reason is that the mix barely moves. In Exeter's 2022 trust the share of balance in the under-540 band is 20.9% at month zero and 20.5% at month 54. Santander's 2024 trust goes from 13.9% to 14.7% over 31 months, AmeriCredit's 2023 trust from 17.0% to 18.5% over 42. Better borrowers do prepay faster; across the subprime shelves 37% of loans scored 660 or above have prepaid against 23% under 540. But a difference of that size, acting over four years, shifts the score composition of a pool by a point or two, and a point or two of composition cannot produce ten points of delinquency.
IV · It is not the term either
The other mechanical explanation is that longer loans spend more months in the ages where delinquency peaks, so a book of longer loans reports a higher rate at constant borrower quality. The premise is testable and it is false for subprime. The balance-weighted original term across the subprime shelves was 71.0 months for 2020 originations and 72.7 months for 2024. Issue 17 found the same thing from the payment side: there is almost no 84-month subprime paper, and the longest Exeter loan in the panel is 80 months. A term that moved 1.7 months did not age the book.
V · Why the book is growing anyway
The book is aging, though, and this is where the reader's question gets its answer. Share of original balance still outstanding at month 24, by origination year: 43.7% for 2020, 38.8% for 2021, then 53.7% for 2022 and 51.6% for 2023. Ten to fifteen points more balance sitting on the books at the same age, on loans of the same term. The paydown slowed, and two things slowed it.
Prepayment collapsed. Share of loans prepaid in full by month 24: 13.2% of the 2020 vintage, 6.5% of 2021, 7.6% of 2022, 5.3% of 2023. Over the same years the median origination loan-to-value rose from 108.8% to 114.1%. A borrower who owes more than the car is worth cannot trade out of it, and a borrower who financed at 2021 rates has nowhere cheaper to refinance to. So the loans that used to leave a pool early, the ones with equity and options, stopped leaving.
And extensions hold balance on the books. At month 24 a loan that has been extended at least once still carries 62% to 69% of its original balance, depending on vintage, against 32% to 47% for a loan that never was. An extension defers principal as well as the due date. Within each lender the share of loans extended by month 24 has risen through the vintages: AmeriCredit from 25.4% on 2020 paper to 38.9% on 2023, Exeter from 34.5% to 42.0%, Santander from 16.8% to 24.3%. More loans extended, each holding more balance for longer.
So originations grow and the outstanding book grows faster, because the old book is not amortising away underneath. The loans that would have left are stuck, the loans that would have charged off are extended, and the aggregate delinquency rate sits on top of a book that is older, carrying more loan against less car, and more extended than the one it sat on in 2021, reporting a number that the extension reservoir keeps from rising as fast as the losses behind it. Every bad number is at a record. The one that matters most, loss at constant age, is further past its old record than the one that gets quoted.
What I would watch
The loss-per-point ratio in the first table, per lender, per vintage. It is one division on two numbers every servicer already reports, and when it rises the delinquency rate has stopped telling you what it used to. The current-but-extended share of balance, by lender, with the lookback window stated. And Santander's 36-month ratio, which is the one case where the two lines rejoined; if the 2023 vintage holds that pattern at three years, the detour at Santander is about a year long. If it does not, the detour is getting longer there too.
VI · What this does not show
Three lenders carry the vintage table, because they are the subprime shelves with enough loans in enough vintages. Bridgecrest is in the extension test and not the vintage table, since the panel holds one Bridgecrest trust. Prime captives are not tested here at all; whether their loss-per-point ratio is also rising is an open question.
Charge-off is gross. Recoveries on these shelves differ in both size and filing convention, and the Issue 15 work found two of the three large subprime issuers reporting repossession proceeds in incompatible ways, so I have not netted them. A net figure would lower every loss number and would not change the direction of any comparison in this brief.
"Current but extended" depends on the lookback window. On the pooled subprime book at July 2026 the share is 10.4 points with a three-month window, 18.2 with six, 30.8 with twelve. Direction and lender ordering do not depend on the window; the level does, and I have used six months throughout and said so.
The extension test is forward-looking from one anchor date. Loans extended in early 2025 are compared with loans not extended then, and the lenders decided which loans to extend. The 0-to-29-day restriction and the controls narrow the selection problem; the inference box above says what is left of it.
At 36 months the 2023 vintage is first-half 2023 originations only, and the 2024 vintage cannot yet be read past month 24. Santander's 2022 vintage is excluded from the table for thin coverage.
Everything here is from registered public securitizations. Loans that lenders keep on balance sheet, sell privately or fund in 144A deals are not in the tapes, and a lender's securitized paper need not look like the rest of its book.
Proof: every figure, traced
Computed means counted from the Exhibit 102 loan tapes on file with the SEC, construction stated. Derived means arithmetic on figures above it.
| Claim as stated | Source and construction | Status |
|---|---|---|
| Panel: 48 public auto trusts, loan-months January 2021 to July 2026; subprime shelves used here are Santander (SDART, 5 trusts), Exeter (EART, 5), AmeriCredit (AMCAR, 4) and, in the extension test, Bridgecrest (BLAST, 1) | SEC Form ABS-EE Exhibit 102; fields per 17 CFR 229.1125 Item 3: originationDate, originalLoanAmount, originalLoanTerm, obligorCreditScore, vehicleValueAmount, reportingPeriodActualEndBalanceAmount, currentDelinquencyStatus, paymentExtendedNumber, chargedoffPrincipalAmount, zeroBalanceCode. Duplicate filed periods deduplicated on loan and period, keeping the later filing | Computed |
| Vintage table at month 24: Santander 2020 / 2021 / 2023 reported 60+ 5.32 / 6.75 / 8.79%, dollar charge-off 6.21 / 8.51 / 12.68%, loans 93,706 / 84,969 / 93,152; Exeter 2020 to 2023 60+ 9.91 / 10.19 / 9.88 / 9.60%, charge-off 11.16 / 14.83 / 18.33 / 17.93%, loans 63,800 / 42,217 / 32,909 / 37,999; AmeriCredit 60+ 1.56 / 1.78 / 2.46 / 2.93%, charge-off 2.46 / 3.65 / 5.48 / 6.46%, loans 61,977 / 42,723 / 48,325 / 36,236 | Tape. Cohort is every loan on the shelf with that origination year observed at or before month 24. Numerators and denominators summed across all trusts on the shelf before division. Reported 60+ balance-weighted at the month-24 tape; charge-off is summed chargedoffPrincipalAmount through month 24 over summed originalLoanAmount. Cells under 5,000 loans suppressed; Santander 2022 had 2,479 | Computed |
| Count-basis charge-off at month 24, same cohorts: Santander 7.49 / 10.10 / 15.01%; Exeter 13.63 / 17.77 / 22.29 / 21.85%; AmeriCredit 4.33 / 5.47 / 7.75 / 9.39% | Tape, loans with any positive chargedoffPrincipalAmount through month 24 over cohort loans. Shown so the dollar figures above can be checked against a second construction; ordering identical | Computed |
| Loss per point 1.17 → 1.44 (Santander), 1.13 → 1.87 (Exeter), 1.58 → 2.20 (AmeriCredit); percentage changes 2020 to 2023 of +65 / −3 / +88% in reported 60+ and +104 / +61 / +162% in dollar charge-off | Arithmetic on the vintage table | Derived |
| Month 36: Exeter 60+ 10.84 → 12.71%, charge-off 16.99 → 25.98%, ratio 1.57 → 2.04; AmeriCredit 2.12 → 3.58%, 4.35 → 10.68%, 2.06 → 2.98; Santander 6.55 → 12.52%, 9.20 → 17.35%, ratio 1.40 → 1.39 | Tape, same construction at month 36; 2023 cohort at 36 is first-half 2023 originations | Computed |
| Current-but-extended share of month-24 balance: Santander 8.9 / 10.4 / 10.9%; Exeter 25.7 / 28.5 / 28.4 / 28.7%; AmeriCredit 15.4 / 18.3 / 20.8 / 23.2% | Tape, balance under 60 days past due with positive paymentExtendedNumber in any of the six reporting periods ending at month 24, over total month-24 balance. paymentExtendedNumber verified per-period at every lender in the panel: within-loan values rise and fall, 0.12% of entries negative (reversals) and floored at zero | Computed |
| Extension test: 365,239 loans active and under 60 days at July 31, 2025 across the four shelves with tapes to July 2026; 54,791 extended February to July 2025; twelve-month charge-off 19.65% against 8.31%; by lender AmeriCredit 14.32 / 6.45%, Bridgecrest 28.50 / 13.38%, Exeter 24.29 / 9.57%, Santander 16.46 / 7.59%; by score band ratio 1.97, 2.18, 2.42, 2.48, 3.53 for under 540 through 660-plus | Tape. Trusts whose last tape precedes July 2026 excluded (AMCAR 2021-1, EART 2021-1, SDART 2021-1, SDART 2022-1) | Computed |
| Controlled odds ratio 2.60 (95% CI 2.53 to 2.66) on 342,806 loans; among loans at 0 to 29 days at the anchor, 16.34% against 6.72% and odds ratio 2.56 (2.48 to 2.64); 19.3% of extended loans were 30 to 59 days at the anchor against 8.7% of the rest | Logistic regression of twelve-month charge-off on the extension flag with credit score, loan-to-value at origination, original term, months on book, log original amount and lender fixed effects; loans missing any control dropped | Computed |
| Decomposition: fifteen closed trust cohorts, change from month 12 to last observed month summed 52.42 points; mix 0.30, within-band rate 51.72, interaction 0.40; 0.6% against 98.7% | Tape. Shift-share over five score bands (under 540, 540 to 579, 580 to 619, 620 to 659, 660 and above), balance weights, within closed trusts SDART, EART, AMCAR and BLAST with at least 18 months after the base. Month-0 base rejected as degenerate, stated in the text | Computed |
| Score-band shares: Exeter 2022-1 under-540 20.9% at month 0, 20.5% at month 54; Santander 2024-1 13.9% to 14.7% at 31; AmeriCredit 2023-1 17.0% to 18.5% at 42; prepaid share 37% at 660-plus against 23% under 540 | Tape, balance by band per age; prepayment is zeroBalanceCode 1 on the five subprime trusts with tapes to July 2026 | Computed |
| Balance-weighted original term 71.0 months (2020 originations) to 72.7 (2024), subprime shelves at month 12 | Tape, originalLoanTerm weighted by end balance | Computed |
| Share of original balance outstanding at month 24: 43.7 / 38.8 / 53.7 / 51.6% for 2020 to 2023; prepaid by month 24 13.2 / 6.5 / 7.6 / 5.3% of loans; median origination LTV 108.8 / 109.9 / 113.3 / 114.1%; extended loans at month 24 hold 62.6 to 68.7% of original balance against 38.7 to 45.1% for never-extended (2020 and 2023 vintages, pooled subprime) | Tape. Outstanding is summed end balance at month 24 over summed originalLoanAmount of the cohort. LTV is originalLoanAmount over vehicleValueAmount. Never-extended means no positive paymentExtendedNumber through month 24 | Computed |
| Share of loans extended by month 24, within lender: AmeriCredit 25.4 / 31.0 / 34.6 / 38.9%; Exeter 34.5 / 40.4 / 40.3 / 42.0%; Santander 16.8 / 24.4 / 18.8 / 24.3% for 2020 to 2023, loans observed to 24 only | Tape. A pooled version of this series showed a step from 20.8% to 31.7% in 2022; that step was lender mix, since the 2022 cell held 2,479 Santander loans against 69,124 in 2023, and it is not reported | Computed |
| Lookback sensitivity of current-but-extended on the pooled subprime book at July 2026: 10.43 points at three months, 18.20 at six, 30.75 at twelve | Tape, same construction with the window varied | Computed |
Loan-level data. SEC Form ABS-EE, Exhibit 102, for the 48 public auto trusts in the panel; this brief uses the Santander Drive, Exeter, AmeriCredit and Bridgecrest shelves. Field definitions per 17 CFR 229.1125, Item 3. Tapes read through the July 2026 reporting period.
Method. Months on book counted from originationDate to the tape's reporting period end. Charge-off treated as absorbing on the first positive chargedoffPrincipalAmount. Within-lender rates built by summing numerators and denominators across every trust on the shelf before dividing. paymentExtendedNumber is a per-period count on every shelf in the panel and is floored at zero where filed negative. Balance under 60 days past due with a positive extension in the prior six periods is called current-but-extended throughout.
Companion issues. Issue 13 on extension-adjusted delinquency, which measured the stock this brief follows forward. Issue 08 on the method. Issue 17 on term. Issue 16 on entry against exit from distress.
Where this brief reasons beyond what the filings show, it is labeled as an inference. Point-in-time reading of filings through the July 2026 reporting period and the public record through October 25, 2026. LendRisk Analytics is an independent research publication with no position in, and no affiliation with, any institution mentioned. This is not investment, legal or accounting advice.