The K-shaped divergence
The headline number is calm and is not, and both are correct readings of the same data. The New York Fed's Q2 2026 report puts total household debt at $18.771tn — $13.117tn of mortgages, $1.713tn of auto loans, $1.651tn of student loans, $1.263tn of cards and $0.568tn of everything else — with overall delinquency improving to 4.7% from 4.8%. That aggregate is stable because most household debt is a mortgage, and most mortgages were written at low fixed rates to people who still have jobs and equity.
Now the tail. Sixty-day delinquency on subprime auto loans hit 6.9% on Fitch and Equifax data, against the 5.0% peak of 2008. Treat the precision carefully: a separate read of Fitch's June 2026 securitisation data puts the comparable series at 5.67%, and the two are not obviously measuring the same population. The less disputable figures are 's own. Auto balances 90 or more days past due were 5.49% in Q2 2026, the worst since 2010, with the flow into serious delinquency rising to 3.00% annualised from 2.93% a year earlier. Card balances 90+ days delinquent were 12.92%. KBRA's July 2026 servicing data had net losses on non-prime auto securitisations at 9.46% against 0.66% for prime, a fourteen-fold gap inside one market.
And the lenders' shares sit near their highs. Around 20 August 2026, Capital One was 2.5% off its high and up 9.9% in a month, Synchrony 1.7% off, Oportun up 35% in twenty days; even Credit Acceptance, the deep-subprime specialist and the most exposed of the group, was only 8.7% below its recent high. We explained why: the equity of a subprime lender is not a claim on the loans. Most of the credit risk was sold into securitisations. What the shareholder owns is origination and servicing fees plus a retained residual that only takes losses once the ABS buyers' is gone. Rising delinquencies eat before they touch the stock, and high delinquency rates come with high interest rates, which is income.
Phantom debt and stretched terms
The official statistics also understate the load, in two directions.
The first is buy-now-pay-later, which mostly does not appear in credit-bureau data and therefore mostly does not appear in the numbers above. A Federal Reserve staff note reported that major providers originated nearly $160bn of US BNPL credit in 2025; eMarketer's estimate of annual BNPL spending for the same year is $116.7bn. The Fed's 2025 SHED survey found 16% of US adults had used it in the prior twelve months, 26% of users had paid late at least once and 17% had been charged a late fee. The CFPB's term for the invisible portion is phantom debt, which is unusually vivid for a regulator and unusually accurate.
The second is the shape of the auto loan itself. The average monthly payment on a financed new car was $773 in Q1 2026 and $777 in Q2. Edmunds found that 23.9% of new-vehicle buyers in Q2 2026 took terms of 84 months or longer, a record, up from 22.9% in Q1. A Dealertrack/Cox dataset summarised in July 2026 had 56.7% of loans starting with negative equity rolled in from the trade-in. Seven-year loans on depreciating assets, with the last car's shortfall already stapled to the front. Originations still set a record at $211bn in Q2 2026.
The securitisation machine's new collateral
The reason risky consumer lending lives in cars and cards rather than houses is that the post-2008 rulebook was written about houses. Dodd-Frank required creditors to make a documented, good-faith ability-to-repay determination on mortgages, and the CFPB later set a servicing baseline. Auto lending got much less. The statute explicitly bars the CFPB from exercising authority over motor vehicle dealers predominantly engaged in selling and servicing vehicles, and its supervision otherwise reaches banks above $10bn in assets and certain larger non-banks. Risk did what risk does and moved to where the rules were not.
The funding followed. Subprime auto lenders grew by bundling loans into ABS and selling them on, which turns a balance-sheet constraint into a distribution problem. When there is more money hunting yield than there is ordinary collateral to buy, the machine goes looking for new collateral.
Which is how you arrive at securitised personal-injury lawsuits. Consumer legal funders advance cash to plaintiffs waiting for cases to resolve, at fees and interest averaging 35 to 45 per cent a year. The advances are typically $3,000 to $5,000. Because they are non-recourse — nothing owed if the case fails — they are arguably not loans, so consumer rate caps largely do not apply. The New York Times documented a New York woman who said turned her ankle on a sidewalk, took $76,500 in advances, and owed at least $1.4mn by the time her case settled. More than two dozen securitisation deals since 2020 have raised $2.8bn from investors, with six funders accounting for over 90% of advances nationwide.
We wrote this up on 19 August and again on 20 August, because the pitch is the tell. The selling point is that litigation outcomes are uncorrelated with unemployment, house prices and rates. What the pool is actually exposed to is court calendars and insurers' willingness to settle, both of which every deal at once, and neither of which has a cycle of loss history to model against. The same year produced securitisations of songwriting royalties, fast-food franchises and .
The verification hole
An asset-backed security is a promise that thousands of specific things exist, are worth something, and are pledged to you and to nobody else. Nobody walks the lot. The check is contractual: a prospectus, a servicer's representations, an auditor's sample, a trustee holding paperwork.
Tricolor is what happens when that promise is false. The SEC sued three former executives of the subprime auto lender on 18 August 2026 — founder and chief executive Daniel Chu, finance chief Jerome Kollar and senior finance director Ameryn Seibold — alleging they raised more than $1.9bn from ABS investors on collateral that was not what it appeared to be, including double-pledging hundreds of millions of dollars of loans. Kollar and Seibold have pleaded guilty and are cooperating. Chu's lawyer calls the case a "rehash" of prior allegations and says many are inaccurate. The size is contested: Reuters reports "hundreds of millions", while Cryptobriefing put the complaint's figure at more than $800mn of fabricated receivables, with the scheme running from 2018. Tricolor filed for bankruptcy in September 2025, so the description survived years of investor diligence.
The SEC enforcement director's line, which we flagged at the time: "These defendants defrauded investors based on bogus collateral and violated the integrity of our markets." A subprime auto ABS is now described by its own regulator as private credit. That is the correct description, and it tells you where the boundary sits.
Why it matters to this crash
Our view is that household credit is not the fuse. It is the pressure gauge.
The honest historical analogue is not 2008. It is 1997, when Mercury Finance disclosed accounting irregularities in January, collapsed and liquidated, and took most of an industry with it — 24 of 26 public subprime auto finance companies ended up in Chapter 11, liquidation, restructuring or exit. Losses to securitisation investors were contained, because auto loans are small and short-dated and the paper matures faster than the pool can rot. Conseco's 2002 bankruptcy sits in the same lineage, though what its securitisation holders actually recovered is not something we have documented. The securitisation model survived 1997. That is the base case here too.
The stress is nonetheless real, persistent and now large enough to bend other data. The share of card balances flowing into 90-day-plus delinquency was 6.97% in Q2 2026, the tenth consecutive quarter above the roughly 6.75% level that the onset of the Great Recession. Ten quarters of crisis-era readings without a crisis is the point: as we put it on 18 August, the losses are absorbed in ABS tranches and lender provisions quarter by quarter rather than arriving at once. July 2026 retail sales fell 0.6% against expectations of a small rise, the first drop in nine months, with sales ex-autos down 0.3%. Serious student-loan delinquency is reported at 10.6% in the same Q2 release, though a separate read of it gives 7.83% against 12.88% a year earlier; the series is being measured more than one way and we would not build an argument on either figure alone.
What makes this a crash point rather than a grind is the second thing Tricolor demonstrated. The consumer books are where the complex sources a very large volume of collateral, and that collateral is verified by whoever is selling it. If the plumbing fails, it does not fail gently. The tranche does not degrade; the assets simply were not there.
What would make this dangerous
The equilibrium runs on employment. A lender charging high rates on a performing book earns more from the spread than it loses to a slow drift in charge-offs, for as long as most borrowers keep paying. That is the whole trade, and it is defensible while people have jobs.
Specific things we would treat as the equilibrium breaking:
Prime crossing over. KBRA's July 2026 prime auto net loss rate of 0.66% is the number that says this is a subprime problem. If prime losses start moving toward non-prime's 9.46%, it is not a subprime problem any more.
The flow rates accelerating rather than plateauing. Card inflows into 90-day delinquency have sat above 6.75% for ten quarters. Serious auto delinquency inflow at 3.00% annualised is creeping, not spiking. A step change in either, rather than another decimal place, is the signal.
Residuals exhausted. The lender equities are priced on fee and spread income because subordination protects the retained residual. Watch for consumer lenders taking impairments on retained interests, or ABS deals breaching triggers that divert cash away from the residual holder.
A second Tricolor. One case of double-pledged collateral is a fraud. Two in the same funding channel is a diligence failure, and it would rationally cause warehouse lines to be re-papered and re-verified across the sector at the same time — which is what a funding freeze looks like from the inside.
A ratings response to collateral with no history. The litigation-advance pools, at 35 to 45 per cent a year, depend on insurers settling on a predictable schedule. If defendants start litigating instead, extends, and at those rates extension is not a delay, it is a loss. There is no observable secondary market, so the first honest mark will come from a downgrade rather than a trade.