Extended subprime charge-offs lose more. The data won't say why.

Subprime auto loans that charge off after a payment extension lose 50.1% of the balance at the low end of the loan-to-value range and 60.0% at the high end. The ones that charge off having never been extended lose 40.6% and 52.1%.

Same pool, same collateral bands, about nine points apart. The extended group loses more everywhere.

That part is clean. The rest of this post covers what the number does and does not tell you.

The two curves

Loss severity here means loss given default: net loss divided by the balance at charge-off, dollar weighted across the pool. A $20,000 balance that charges off and books $10,000 of loss after recoveries is 50%. It sets what a default costs, which is a different question from how often defaults happen.

Split the charged-off subprime population by origination loan-to-value (LTV), then by whether the loan had ever had a payment moved:

Both curves slope up, which is the expected part. A higher advance rate at origination means less collateral behind the balance, so a charge-off recovers less of it. Every loss model already knows that.

What I care about is the vertical distance. It runs from just under eight points to about nine and a half, and it never closes. Subprime loans at 130% LTV that were never extended lose 52.1%, barely worse than subprime loans under 100% LTV that were extended, at 50.1%. On this cut, an extension in the history moves a low-LTV charge-off almost as far as thirty points of advance rate does.

A scope note, because this is the kind of number that gets quoted without its qualifier. Everything in this post is subprime, FICO under 660 at origination. The eight to nine points is a subprime figure. The prime version of the same cut does not behave the same way and needs its own post. None of this carries across to prime paper.

It holds at every band

Consistency is what makes this worth writing up. A gap that showed up in one bucket and disappeared in the others would be a mix artifact, and mix artifacts are easy to manufacture: thin cells, one large issuer carrying a band. Neither is what is happening here. The extended curve sits above the never-extended curve at every step of the subprime LTV range, same direction, similar distance.

Monotonic in LTV, parallel across LTV. Whatever is behind the gap, it is not a property of one collateral band.

The clock is the other half

The severity numbers come with a timing number, and the timing number is the one I keep going back to.

Previously extended loans take roughly 15.5 months from their first modification to reach charge-off. The never-extended comparison group gets there in about five.

Ten extra months is a long time in auto collateral. The vehicle depreciates through all of it. The balance gets re-aged instead of amortized, because the whole mechanic of an extension is moving a payment to the back of the schedule rather than applying it now. By the time the repossession happens, the car is older and less principal has come off the balance than the original schedule would have delivered. Both of those push severity up on their own, with nothing happening to the borrower at all.

So one reading of the subprime gap is close to mechanical. These loans die slower, and a slow death in depreciating collateral is expensive.

The other explanation is just as good

Extensions are not handed out at random. A servicer decides which loans to extend, and servicers extend loans that are already in trouble. A borrower who is 45 days past due and calls in is a far more likely extension candidate than one who has never missed.

So the two groups were different populations before any payment was moved. Origination LTV is the only collateral control in the table above, and it is a snapshot from the day the contract was signed. It says nothing about what the vehicle was worth two years later, or how many times the account had already been worked by the point someone granted an extension. If servicers route the weaker accounts into extensions, the extended bucket shows worse severity whether or not the extension itself did anything.

Both explanations predict the table exactly as it sits. They also both account for the longer clock. The clock was the measure I expected to separate them, and it does not.

Why the record can't pick between them

Here is the part I am not going to resolve, because it cannot be resolved from this data.

ABS-EE files record what happened. They do not record what a servicer was looking at when it granted the extension, and they do not contain a second copy of each extended loan where nobody granted one. The two candidate stories, extra time in depreciating collateral and selection of weaker loans into the extended group, arrive bundled inside one observed number. There is no reason code in the file. There is no internal risk grade as of the decision date, and no loan-level view of what the car was worth on the day of the decision.

I could pick a share anyway. Six points to the clock, three to selection, or the reverse. It would be a guess wearing a decimal point. The record supports the gap and the shape of the gap. It does not support a split, and any split you see quoted on this without a designed comparison is someone's prior, not a measurement.

I have spent more time on this one than I want to admit, poking at it from a few angles to see whether it would break in a direction. It does not break. That is an unsatisfying place to stop, and it is where the data stops.

What to do with it anyway

An unresolved decomposition is not the same as an unusable number.

The severity gap prices regardless of its source. If you are marking a pool of previously extended subprime paper, those loans have historically lost roughly eight to nine points more at charge-off than never-extended subprime paper at the same origination LTV. Timing or selection, the dollar is the same. Which one it is only matters if you are trying to forecast a change in servicer behavior.

The extension flag also belongs in the severity model, not just the default model. Most surveillance treats a modification as a probability-of-default signal, which it is, then stops there. It carries loss-given-default information too, and the two are not redundant. A subprime pool can run an ordinary default rate and a worse loss curve than you penciled, because its defaults are arriving through the extension channel.

Then there is the horizon. Ten months between first modification and write-off means a pool's extension activity leads its loss curve by about that much. If you watched an extension wave build in a trust in early 2025, those charge-offs are landing now. That is a scheduling fact and does not depend on which explanation you believe.

Where this sits

This is the loss side of a set of findings that has been building. Subprime extension usage rose fastest at high LTV, the Extension Default Rate is worst at high LTV, and extension activity overall has stayed above its pre-pandemic norm. Now the charge-offs coming out the far end of that channel cost more per dollar than the ones that never entered it.

The thread through all of it is that extensions concentrate where the collateral is thinnest, which is also where a default costs the most. True of usage, true of the Extension Default Rate, true of severity. The record supports the extension channel marking worse-outcome loans. It does not support the claim that the channel produces them, and this data cannot close that gap.

Methodology

• Source: SEC ABS-EE loan-level filings parsed into LoanTape's cache, deduped to one record per loan per reporting month by latest filing, with data through April 2026.
• Population: subprime charged-off loans, where subprime is FICO under 660 at origination. Every number in this post is subprime. Prime is a separate population with a different pattern and is not covered here.
• Loss severity is loss given default: net loss divided by charged-off balance, dollar weighted and pooled rather than averaged per loan, so large balances carry proportional weight.
• Groups: "previously extended" is any loan with at least one payment extension or type-4 modification recorded before charge-off. "Never extended" is a charge-off with no such record at any point in its observed life.
• LTV is origination loan-to-value, not current. There is no loan-level current collateral value in the filings.
• Timing is measured from the first recorded modification to the charge-off month for the extended group, against the corresponding interval for the never-extended comparison group.
• Extensions are not randomly assigned. Servicers select which loans to extend, so the two groups differ in ways the filings do not report. The severity gap reflects both the additional time to write-off and that selection, and this data cannot separate the two.

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 vintage cuts, is in the Modification Effectiveness Deep Dive. For the loan-level files behind it, see LoanTape pricing.