What extended subprime loans actually realized, by vintage

Everything I have published on payment extensions so far has been a rate. Which loans get one. What share is 30 or more days past due two months later. A rate is an early read, and you take an early read when the loans have not finished.

These loans have finished, or enough of them have, so this post has no rate in it and nothing estimated. Just completed cash: the net loss extended subprime loans realized over their lives, per dollar of balance outstanding in the month the extension was granted, set against non-extended loans from the same issuer and the same origination year.

The answer depends almost entirely on which vintage you ask about. On 2018 originations, extended loans realized 2.47 cents per dollar more net loss than the non-extended loans in their cell. On 2022 originations, 12.95 cents. The gap more than quintuples across four origination years.

Two caveats belong in the same breath as those numbers. The later vintages are not finished, so their figures are floors. And this is not a like-for-like comparison, because servicers choose which loans to extend and they tend to choose weaker ones.

What "realized" means here

There is nothing to estimate. The measure is arithmetic on loans that have already ended.

Take a subprime loan that received an extension. Note its balance in the month of the extension. Follow it to termination and add up the net loss the trust actually took on it. Divide. Then set that figure beside non-extended loans in the same issuer-and-vintage cell, so a shelf that wrote 2019 paper is compared against its own 2019 paper rather than against somebody else's 2022 book.

That is the whole thing. Nothing here tries to say what an extended loan would have done had it not been extended, and nothing corrects for the fact that extended and non-extended loans were different loans to begin with. It reports two realized numbers and subtracts.

I find this more useful than it sounds, for a straightforward reason. Every other measure in this series stops the clock early and has to be defended on its horizon. This one has no horizon problem. It has a completeness problem instead, which is a different thing and a more visible one, and I would rather have the visible one.

The gap widens with vintage, then flattens

Read down the column. The first four rows climb hard: 2.47, then 3.95, then roughly a doubling into 2020, then most of another one into 2021. Four origination years, five times the gap.

Then 2022 comes in at 12.95 and the climb stops. That last step is 0.44 cents. After three years of the gap widening by half again or more each time, 2021 and 2022 sit almost on top of each other, and I did not expect that.

The flattening is the part I would test before I believed it, because 2022 is also the least finished row in the table. A floor that has not stopped rising can look a lot like a ceiling. Ask me again in a year.

The two ends of the series read nothing alike. On 2018 paper, 2.47 cents is small enough to argue about. A slightly different cell definition or loss convention could move it around, and I would not build a thesis on it. Thirteen cents is not that. Thirteen cents per dollar of balance shows up in a bond, and no definitional quibble makes it disappear.

Extended loans realized more loss in every vintage here, including 2018, where the margin is thin. The slope is what changed. Whatever separates the extended population from the rest of the book, that separation kept widening through 2021, and it did so on cash the trusts actually took rather than on a projection.

The obvious thing to reach for is the collateral. The 2021 and later vintages were written at high advance rates against vehicle prices that later came down, and we have covered that collateral break and the high loan-to-value (LTV) tilt in extension usage before. Both fit the shape. Neither is demonstrated by this measure, which knows nothing about why.

Issuer dispersion is wide

Pooled across the seasoned 2018-22 subprime vintages, two issuers show the spread clearly.

Exeter's extended loans realized roughly 32 cents of net loss per dollar against roughly 15 cents on its non-extended book. Carvana's came in at about 24 cents against about 8 cents. Subtracting rounded figures, both spreads land near 16 to 17 cents.

Do not stack those spreads against the vintage table and expect them to line up. The issuer figures pool 2018 through 2022 instead of splitting by year, and they cover two books rather than the whole subprime pool. The vintage numbers are within-cell gaps across the seasoned book. Different aggregations, which is why the issuer spreads come out wider than the per-vintage gaps.

Across both cuts, the distance between the extended and non-extended arms is large at the issuer level, and it differs by shelf. If you hold auto ABS paper, the shelf matters.

The first caveat: these vintages are not finished

Loans from the 2020, 2021 and 2022 cohorts are still outstanding, and the more recent the vintage the more of them there are. Data here runs through April 2026.

That makes those figures floors. Loss accumulates as loans resolve, so the numbers on the later vintages can rise from here, and I expect some of them will. Nobody should read 12.95 cents as a settled result. It is what the completed portion of that vintage has produced so far.

The honest version of the caveat goes one step further. The non-extended arm is incomplete on the same vintages. So the gap itself is not pinned. It could widen as the remaining loans resolve, or it could compress. What I can say is that the 2018 end of the series rests on a much more complete book than the 2022 end, so the slope runs from a solid point to a provisional one. The flattening at the top sits entirely in provisional territory.

The second caveat: this is not an experiment

Servicers decide which loans get an extension. They are not drawing at random.

We have already documented what that selection looks like. Extension usage climbs with origination LTV and with original term, and in subprime the extension usually lands on a loan that is already past due rather than one that is current. Those findings are in the posts on who gets a payment extension and on how high-LTV subprime extensions perform. The extended group was a weaker group before anyone moved a payment.

So the realized gap holds two things that this measure cannot pull apart: whatever followed the extension, and the difference between the loans picked for one and the loans that were not.

I am not going to resolve that here, and I would be suspicious of anyone who resolved it with a table. My position is narrower. The slope is measured on completed cash, which makes it worth pricing. It says nothing about what extensions do, and does not support claims that it does.

What to watch

Whether the flattening between 2021 and 2022 survives. That is the first thing I want another year of data on. If both points keep rising and stay level with each other, the gap found a ceiling. If 2022 pulls ahead as it resolves, the flat step was an artifact of completeness and the trend never broke. The 2023 book will help here too, once it is seasoned enough to measure the same way.

Issuer dispersion with vintage held constant. The Exeter and Carvana figures above pool five origination years, which blends a shelf's book with its vintage exposure. Splitting them is the cleaner comparison and it is the cut I want next.

And the early read next to the late one. The two-month Extension Default Rate tells you something within a quarter; realized loss tells you something years later. Running both on the same vintages is how you find out whether the fast metric earns its keep.

Methodology

The measure is realized lifetime net loss per dollar of balance outstanding at the month of the extension, on subprime loans that received a payment extension, compared with non-extended loans in the same issuer-and-vintage cell. It is observed cash on completed loans, not a projection, and it carries no adjustment for differences between the extended and non-extended populations.

Extensions are loans flagged with a positive normalized payment_extended_number or an SEC modification_type_code of 4 or 04. Subprime is FICO below 660 at origination. Vintage is origination year. Cells are issuer by origination year.

Source is SEC ABS-EE loan-level filings parsed into LoanTape's loan-level cache, deduped to one record per loan per reporting month by latest filing, with data through April 2026. Coverage is the seasoned 2018-22 subprime vintages. The 2020, 2021 and 2022 cohorts still carry outstanding loans, so figures on those vintages are floors and can rise as the remaining loans resolve.

This runs on the same loan-level infrastructure behind LoanTape's ABS-EE dataset, Form 10-D data, and ABS remittance data. The subscriber version, with issuer, trust, and post-extension outcome cuts, is in the Modification Effectiveness Deep Dive. For the loan-level files behind it, see LoanTape pricing.