Issuer Charge-Off Amounts Are Not Comparable in Auto ABS
Of the $39.58 billion of principal that had entered a first charge-off month across the twenty auto ABS issuers in LoanTape's coverage as of June 30, 2026, $2.49 billion never appears in the field the filings label as the charged-off amount. That is 6.3% of the pool's gross charge-off exposure, absent from any calculation that sums chargedoff_principal_amount as filed. The gap is not spread evenly. The large subprime shelves file 0.97 to 1.00 of the balance entering the charge-off month. Several prime captives file 0.29 to 0.62 of it.
This post is the account of how that surfaced, what the filing conventions are, and what I changed in the pipeline as a result. It is also the explanation behind the correction notice that has been on four posts since August 22.
A recovery rate of 101%
The audit started with a number that could not be right. While building the motorcycle ABS post, the first pass at recovery rates divided cash recoveries by the filed charged-off amount. On Harley-Davidson's completed 2019-2021 charge-off cohorts it returned 101%.
I ruled out the obvious explanation first. If recoveries were filed as running totals and re-filed each month, summing them would double count. They are not: the typical charged-off loan shows one or two recovery months, and the ninetieth percentile is three. That is a collateral sale, not a cumulative field.
The mechanism appeared when I split the 936 charged-off loans in those cohorts by the size of their recoveries relative to the filed amount, then reconciled each group against the balance the loan carried entering its charge-off month.
| Recoveries vs filed charge-off amount | Share of loans | Filed amount / entering balance (median) | Filed amount + recoveries / entering balance (median) |
|---|---|---|---|
| At or below 1x | 54.6% | 1.00 | 1.45 |
| 1x to 2x | 19.3% | 0.69 | 1.49 |
| Above 2x | 26.1% | 0.24 | 1.05 |
The top group files the full balance, and recoveries arrive on top of it. The bottom group files roughly a quarter of the balance, and adding recoveries back reconstructs the balance within about 5%. That is a charge-off filed net of expected sale proceeds, with the proceeds then filed a second time as recoveries. Dividing recoveries by the netted amount counts the sale twice. Net losses computed the same way printed near zero or negative for those loans.
Within one issuer this is a cleaning step. It became a pipeline correction when I ran the same signature test on the other nineteen issuers.
What the twenty issuers file
The ABS-EE field is named as if it were one quantity. Reconciling every charged-off loan's row against its own balance history shows it is at least four, and the pattern is issuer-specific rather than random.
| Filing convention | What the filed amount is | Filed amount / entering balance | Where recoveries go | What raw arithmetic produces |
|---|---|---|---|---|
| Gross | Full balance at charge-off | About 1.0 | Filed separately, on top | Correct gross and net loss |
| Net of sale proceeds | Balance less expected collateral proceeds | Well below 1.0; about 0.25 in the Harley sample | Filed again as recoveries | Understated gross loss; recovery rate near or above 100% |
| Near-zero placeholder | A token amount on a terminal charge-off | Close to 0 | Filed separately | Loss almost absent; the loan reads like a payoff |
| Open partial write-down | A partial amount on a loan that keeps reporting a balance | Below 1.0, and the loan is not terminal | Later, if at all | Correct only once the terminal event is observed |
The governed audit as of June 30, 2026 counts 2,285,012 charged-off retail loans across the twenty issuers. Of those, 23,039 carry an open partial write-down on a loan still reporting a balance, 1,708 filed a partial amount and then a later terminal charge-off, and 11,725 enter the data with the charge-off already observed, so there is no prior month's ending balance to read. Some loans also carry a charge-off zero-balance code with no filed amount at all; in the Harley book that was 137 loans, 1.5% of charge-offs.
The distribution across issuers is what matters. Measured as filed gross charge-off over the balance entering the first charge-off month, the large subprime shelves reconcile at 0.97 to 1.00. Several prime captives reconcile at 0.29 to 0.62, which means the filed field is missing between a third and two thirds of their charge-off exposure, and at those issuers the netted, partial, or near-zero convention covers most of their charged-off loans, not a tail. The filed field works for tracking one gross filer over time. It does not work for ranking issuers, for a pool-level loss curve, or for any segment cut that mixes filers, which is every cut by FICO band, vintage, or collateral type.
I am not naming the under-filing issuers, for the same reason the motorcycle post did not: each convention reconciles once it is identified, so this is a comparability finding, not a finding about any filer's accuracy. Harley-Davidson appears because that post already reported its numbers.
Why the pool-level reports cannot show it
Nothing in a monthly Form 10-D or a servicer remittance report surfaces this. Those documents report the pool's charged-off amount and recoveries as totals, on whatever basis the servicer uses, and the totals are internally consistent. A pool that files net of proceeds shows a low charge-off line and a high recovery line, and the two add up. There is no field on a 10-D that states which convention was applied, and no way to test one from outside the pool.
The check exists only at the loan level. Each charged-off loan in the ABS-EE data has a balance history: the ending balance the month before the event, the filed amount in the event month, and the recovery rows after it. Reconciling those three against each other, loan by loan, is what separates a gross filer from a net filer, and it is the only place the distinction is visible.
The unified basis
The fix is one definition applied to every issuer:
- Find the first charge-off event: a positive filed charge-off amount, or zero-balance code 4.
- Take the balance entering that month, from the prior month's ending balance. Where no prior month exists, use current balance plus the first filed amount.
- Gross loss is the greater of that entering balance and the cumulative filed amount. Where the filed amount is the larger of the two, it is kept; that is the case on 65,590 loans, 2.9% of charge-offs. A partial write-down on a still-open loan stays at its filed value until the loan terminates.
- Net loss is gross loss minus every recovery row, cash basis.
For a gross filer the new basis equals the old one, which is why the subprime shelves barely move. For a net-of-proceeds filer, gross loss rises to the balance, and the recoveries that were being double counted now offset it once.
The basis is also prefix-stable: a terminal row or a later filed amount that arrives after a release cutoff contributes from that row forward and cannot restate a month already published. That is what lets a monthly publish run without rewriting history each time an issuer re-files.
At the pool level the correction moves every loss figure in the same direction:
| Metric, twenty issuers, as of June 30, 2026 | Filed basis | Unified basis |
|---|---|---|
| Gross charge-off | $37.09B | $39.58B |
| Net charge-off after recoveries | $20.38B | $22.87B |
| Recovery rate | 45.1% | 42.2% |
| Harley-Davidson completed 2019-2021 cohorts, recovery rate | 101% | 62-66% |
Pool gross loss rises 6.7% on the filed figure and the pool recovery rate falls by about three points. The pool numbers understate what happens inside prime, because prime charge-offs are where the netted conventions concentrate: an issuer filing 0.29 of its entering balance more than triples its gross loss on the unified basis, while an issuer filing 0.99 is unchanged. Every comparison between prime issuers, every prime loss-to-liquidation curve, and every recovery rate that had a netted filer in its denominator is affected.
What LoanTape did
On August 22 I put a correction notice on the four public posts whose figures were built from the filed field: the subprime loss-to-liquidation and prime loss-severity posts from July 26, the recovery-by-make post from August 4, and the loss-to-liquidation-by-FICO post from August 7. The notice reads the same on each: the loss values used issuer-filed charge-off principal, which is not economically comparable across filers, and every figure and chart is being recomputed on the first-charge-off exposure basis. The posts stay readable. Their loss charts are replaced with a placeholder until the corrected versions are rebuilt.
The premium decks were withdrawn rather than annotated, since a deck cannot carry a per-chart caveat. That covers the five State of Auto monthly decks from January through May 2026, the vehicle collateral loan-to-value (LTV) deep dive, and the payment extensions deep dive. Their URLs return 410 Gone. Five gallery entries with monetary loss content, the public performance snapshot CSV, one product-post screenshot, and two screenshots in the June newsletter were also withdrawn. Two charge-off research drafts that were queued for publication are held.
Delinquency and extension figures on the public trackers are unaffected and remain published; only the monetary loss observation is withheld. Delinquency, measured as days past due (DPD) against active balance, does not use the charge-off field.
On the pipeline side, the governed basis now lives in one helper that every loss surface calls: the dashboard, the premium builds, the recurring reports, the validation queries, and the research lane. Every published artifact carries a stamp, chargeoff_basis_version=first-chargeoff-entering-exposure-prefix-stable-v2, and the website refuses to serve an artifact without the exact stamp. A pre-correction bundle cannot be relabeled and served. A nightly check across the full issuer roster proves the exposure floor holds, gross loss is non-negative, and the published cohort tables reconcile to a direct recomputation from the loan-level rows.
Every loss surface is being recomputed on this basis. The auto dashboard reflects it once the stamped publish is deployed. The withdrawn decks and the four posts are restored one at a time, each after its rebuilt asset passes a before-and-after diff review against the frozen expected movement by issuer.
Reading issuer loss numbers in the meantime
Three things follow from the audit for anyone using filed charge-off data directly.
First, a recovery rate near or above 100% is a filing convention, not performance. The two-step check is short: the filed amount as a share of the entering balance, then filed amount plus recoveries as a share of the same balance. A net filer shows the first well below one and the second at roughly one.
Second, a single issuer can mix conventions. Harley-Davidson filed the full balance on 55% of its charged-off loans and a netted amount on 26%. A within-issuer time series on the filed basis is only as stable as that mix.
Third, the correction is concentrated in prime. On the subprime shelves the two bases agree within three percent. The 6.3% pool gap comes almost entirely from prime captive exposure, which is how an individual prime issuer's gross loss can rise by a multiple while the pool total moves by single digits.
The reconciliation that produced this, every charged-off loan's balance history against its filed amount and recoveries, is the same loan-level extraction we build for enterprise clients on any issuer or segment.
Methodology note
Source is loan-level ABS-EE data filed with the SEC, parsed monthly, restricted to retail loan trusts; lease charge-offs stay on a separate basis. A charge-off event is the first month with a positive filed charge-off amount or zero-balance code 4. Entering balance is the prior month's ending balance, falling back to current balance plus the first filed amount. Recoveries are the filed recovered-amount field, cash basis, the sole recovery input. The Harley-Davidson figures are from the August 22 post and cover the 936 charged-off loans in the completed 2019-2021 charge-off cohorts. The pool totals, loan counts, and per-issuer ratios are from the governed cross-issuer audit run as of June 30, 2026, over deduplicated loan-month rows for the twenty enabled issuers. Issuers are not named in the convention table by design; the audit output by issuer is part of the release evidence for the corrected publish.