Anatomy of a toxic book.

The book below is a blinded composite, a subprime auto lender on a senior warehouse line, anonymized and rounded. But every benchmark around it is real and cited, and the pattern it walks through is the same one that surfaced publicly at Tricolor, at American Car Center, and across the 2022-2023 ABS vintages. The method is exactly what we run on a live tape; the industry numbers are exactly what the regulators and rating agencies published.

The setup: a book that looks fine.

A subprime auto lender, roughly 2,000 active loans across 18 originating dealers, financed on a $40M senior warehouse facility. The line carries the standard covenant package: a cumulative net loss cap, a delinquency trigger, and a weighted-average FICO floor. Miss any one and the bank can sweep cash, freeze advances, or reprice the line.

At the portfolio level, nothing flashed. Blended 60+ DPD was 6.8 percent, within a rounding error of the industry benchmark. For context, the subprime 60+ rate hit 6.65 percent in late 2025, its highest reading since the 1990s, and Fitch’s subprime annualized net loss index reached 9.81 percent in January 2026, a post-pandemic high.¹ So a 6.8 percent book genuinely looked like the middle of the market. Cumulative net loss was tracking under the covenant cap. On the monthly compliance certificate the lender sent its bank, every box was green.

A blended average is a weighted blur. It tells you where the middle of the book sits. It tells you nothing about which loans are pulling the middle in which direction. The entire risk lived in the variance the average erased.

Security: finding the toxic paper.

The first cut is attribution by dealer. Rank all 18 originators not by volume but by the metrics that actually predict loss, early payment default, severe delinquency rate, and net loss per dollar originated. The distribution was not even close to uniform.

Three dealers originated 22 percent of the book but accounted for 51 percent of the severe lates. Their early-payment-default rate, borrowers who miss inside the first three payments, the cleanest signal of a bad deal at inception, ran roughly four times the network median. EPD that high is not a servicing problem or a macro problem. It is an underwriting and origination-quality problem at the source. These three dealers were manufacturing paper that was impaired the day it was written.

That concentration is not a modeling convenience. It is how subprime loss actually distributes. LendingTree’s analysis of the default population found 83.7 percent of defaulted auto loans sit with deep-subprime borrowers scoring under 580, and defaults cluster years two through four of the loan, not at random.² Loss is concentrated by origination source and credit tier; an average sprays it evenly across a book and hides exactly where it lives. At Tricolor the concentration was its own kind of tell: after the bankruptcy, roughly 40 percent of 70,000 active loans shared a VIN with at least one other loan, the same collateral counted twice.³

Predictability: seeing the loss before it lands.

Attribution tells you who. The vintage view tells you when. Group the loans those three dealers wrote by origination quarter and plot cumulative net loss against months on book. The 2023 second-half cohort was rolling from 30 to 60 to 90 days past due at roughly 1.4 times the pace of the lender's seasoned vintages at the same seasoning point. This is not a hypothetical failure mode: Fitch flagged the 2022-2023 vintages as the worst post-pandemic performers, originated when used-car prices peaked and stimulus-inflated FICO masked the real credit underneath.⁴ The cohorts that look fine on a blended line are precisely the ones a vintage cut exposes.

Roll rates are leading. Net loss is lagging. The 90-day bucket today is the charge-off in two to four months and the realized loss after that, once you net the depressed auction recovery against the deficiency. Project that cohort forward on its own roll behavior and it lands near 19 percent cumulative net loss, against an 11 percent book average. That single pocket, left alone, drags the whole book through the covenant.

Feed the trajectory into the covenant math, current cumulative loss, monthly slope, the cap in the facility agreement, and it returns a date. At the observed pace, the net loss covenant breaks in month 6. Not "someday." Month 6, triggering the cash sweep that starves originations exactly when the lender would need liquidity to reprice and recover.

Attribution into action.

Because the problem was named precisely, the fix could be precise too, no blunt, book-wide retrenchment that strangles the healthy 78 percent of the portfolio. Three moves, all targeted:

One, stop the bleeding at the source. Pause new originations from the three flagged dealers pending a network review. The healthy fifteen keep funding.

Two, fix the structural defect. The toxic cohort skewed toward older, high-mileage vehicles written on 72-month terms, loans engineered to outlive their collateral. Tighten maximum term against vehicle reliability so the loan can no longer mature years after the car is worth less than the payoff.

Three, reprice the risk that remains. Recut advance-rate assumptions on the impaired cohort so the warehouse exposure reflects the real recovery, not the originated balance.

Re-run the covenant projection with those three changes in place and the breach pushes out beyond the projection horizon. The line is preserved. On a representative book of this size, the targeted intervention protects on the order of 140 basis points of portfolio yield and restores roughly $1.8M of covenant headroom versus letting the cohort run, the difference between a managed quarter and an emergency call from your bank.

The whole job, in one line.

The aggregate said the book was fine. Attribution said it had six months. That gap, between the number you report and the slope underneath it, is where lenders get swept and where the work lives. Security is finding the toxic paper before it metastasizes. Predictability is seeing the loss before it prints. Attribution is naming the exact dealer and vintage so you can act with a scalpel instead of a sledgehammer. Deliver those three and the yield protects itself.