Santander Consumer USA subprime vehicle loan scoring

A borrower with damaged credit walks into a dealership and leaves in a car. The dealership picks the vehicle and its price, the term, the down payment and the products financed into the amount owed, and enters what the borrower says they earn and what they say they pay for housing. The application goes to Santander Consumer USA Inc., an indirect lender: it does not lend across its own counter, it buys the retail installment contract from the dealership. The company's own FY2020 Form 10-K describes what happens next in two sentences that the pleadings do not really dispute. "The Company's robust historical data on both organically originated and acquired loans is used by Company to perform advanced loss forecasting. Each applicant is automatically assigned a risk score using information from Credit Bureau and credit application, placing the applicant in one of multiple pricing tiers." The company then "continuously maintains and adjusts the pricing in each tier to reflect market and risk trends", with interest rate, down payment and loan-to-value named as the material components of risk-based pricing. A manual underwriting team is retained "for manual review, consideration of exceptions, and review of deal structures with dealers". Nothing in the record describes a human underwriter reading an ordinary application.

The scale is the last full public year before the company was taken private. 5,576 employees. 1,938,764 retail installment receivables outstanding at the end of 2020, against 1,810,973 a year earlier. $32.9 billion of retail installment contracts held for investment and $26.6 billion of originations in the year. Average origination credit-bureau score 626, average annual percentage rate 14.1 per cent, against 598 and 16.3 per cent in 2019. The correction leg runs at the same order of magnitude: 285,661 repossessions in 2019, 15.7 per cent of average receivables outstanding, and 177,639 in 2020 at 9.3 per cent — a figure suppressed by a nationwide suspension of involuntary repossession at the pandemic's onset, restarted in the third quarter, with over a million extensions granted from March 2020 and roughly a third of customers on a deferral at year end. 2019 is the representative pre-remedy year.

Now read the grievance, and note that it is not the one you expect. California's complaint, filed on 19 May 2020 in San Mateo Superior Court, does not say the model was wrong. It says: "Although Santander has sophisticated models that forecast consumer default, Santander's policies with respect to stated income and expenses allow it to underestimate default risk in important ways." Its paragraphs 8 to 11 open the architecture — the only pleading in the multistate group that does. One model takes the consumer's borrowing history together with the applied-for deal's loan-to-value, debt-to-income, payment-to-income, mileage and term, and emits a probability that the consumer becomes severely delinquent inside a defined window; that probability is converted into a scaled score on a proprietary, FICO-like scale. A second, separate life-of-the-loan model maps a given proprietary score to a probability of default before the term ends. Paragraph 11 carries the figure everyone quotes: "for at least part of the time period examined by the People, Santander projected that these consumers with the lowest proprietary scores had a greater than 70% likelihood of default over the life of the loan." Carry all three qualifiers with it. It is an allegation in a pleading, hedged inside the pleading itself; it is scoped to the lowest scores rather than to subprime borrowers generally; and it is a modelled life-of-loan probability from a second model downstream of the score, not an observed default rate. The coalition's own shared sentence, repeated verbatim from state to state, is softer: Santander, "through its use of sophisticated credit scoring models to forecast default risk, knew that certain segments of its population were predicted to have a high likelihood of default."

The forecast set the price. It did not set the loan. That is the whole geometry of this case, and it is unusual: the model informs the tier, the tier is continuously re-priced, and the decision to fund is a separate act with a separate owner and a separate objective. Nowhere in the pre-2020 record does the forecast bind the funding decision.

What was it fed? California's paragraphs 17 to 21 give the answer. The operator generally let applicants state mortgage and rent expenses without proof and had "no apparent measure against falsified housing figures"; where housing cost was not stated it assumed an internal default amount "that would not be reasonably sufficient to pay for mortgage or rent in the vast majority of localities"; and from early 2013 it made "an aggressive push ... to waive proof of income on most applications." The complaint's conclusion is the mechanism sentence: "Santander employs modeling that makes use of housing costs that are based on faulty information and therefore likely incorrect." Vehicle equipment came from the dealership too, which is the surface exploited by "power booking" — misrepresenting what the car has. An accurate estimator fed a figure nobody obtained returns a confident number, and the confidence is about the arithmetic.

An independent measurement two months after the 2017 Massachusetts and Delaware settlements put a number on the same gap. Bloomberg News reported on 22 May 2017 from a Moody's Investors Service report of 17 May, using newly available asset-backed issuer data: income was verified on 8 per cent of borrowers in a Santander deal against 64 per cent for a contemporaneous AmeriCredit deal, and loans combining low or no credit score, no co-signer and no income verification were about 9 per cent of Santander's pool balance against under 1 per cent of AmeriCredit's. A separate Moody's report put roughly 42 per cent of Santander's 2009 to 2014 loans written through the dealers identified as high-risk in those settlements as defaulted or expected to default. The company's treasurer answered on the record: the practice had been consistent over time though lower than competitors', the higher losses in the loans backing the bonds had been visible to investors, and bondholders were protected by loss cushioning in the bonds.

Meanwhile the checks the organisation already owned were running. A problematic-dealer tracking process had operated since as early as 2010. Early-payment-default monitoring identified dealerships subject to more extensive documentation requirements or exclusion — that is the company's own description in its own annual report. Massachusetts alleged that the operator's own internal audit had concluded its dealer oversight was inadequate. California alleges what happened to all of it, and the reason it gives is organisational rather than technical: "internal tension at Santander between punishing problematic dealers and retaining Santander's market share", reluctance to act against flagged dealers so long as enough of their paper was profitable, a preferred-lender arrangement with Chrysler under which flagged dealers were allowed to participate, and a stated-income policy rolled out without barring dealers with a history of misstating income, which "led to a significant spike in the number of early payment defaults." An error signal was observed and the loop did not close.

On 19 May 2020 the coalition announced parallel consent judgments in 34 jurisdictions — 33 states and the District of Columbia — carrying approximately $550 million and effective from 1 May. The headline is a package rather than a payment, and it has to be broken out. Approximately $65 million is cash restitution to a settlement-administrator trust. $5 million goes to the states for fees and costs, and up to $2 million funds administration, reverting if unspent. Up to $45 million takes the form of letting defaulted consumers scoring 401 or below who had not yet been repossessed keep the vehicle and the title. And approximately $433 million is immediate forgiveness of deficiency balances on defaulted loans the operator still owned, with further waivers on loans it had to attempt to buy back. Roughly one part cash to eight parts forgone collection. Per-jurisdiction shares give the shape: California approximately $99 million; South Carolina up to $20 million, of which more than $3 million is restitution and almost $17 million deficiency waivers; the District of Columbia $2 million, of which $267,112 is restitution and $1,735,000 debt relief.

What the attorneys general asked for going forward is the most interesting document in this file, because it is a governance design rather than a fine. They did not ask for a better model. They froze the one that existed: "Santander shall not substantially change its loss forecasting score formula", and a substantial change requires sixty days' advance notice to a Monitoring Committee describing the change and its potential impact on the back-test. They banded the entire remedy on that frozen score at 401 or below, 501 or below, 502 to 600, and 601 or above. They added a hard gate the model does not compute: no purchase of a loan where the sole obligor's residual income at origination — gross monthly income less monthly debt obligations, less a reasonable estimate of basic living expenses, less a reasonable estimate of payroll taxes — is zero or negative, with a reasonable debt-to-income threshold re-evaluated at least annually and a quarterly statistically relevant sample tested for calculation accuracy and threshold compliance. They ordered a second model built by 31 December 2020, an "income reasonability model" over historical consumer, third-party and geographic data whose only job is to score confidence in stated income and to route low-confidence applications to additional manual review. They made dealer "Treatments" — screens, documentation requirements, stipulations for any dealer known or reasonably suspected of income inflation, expense deflation or power booking — non-waivable until the dealer has demonstrably fixed the problem, and barred the operator from requiring dealers to sell ancillary back-end products. And they made the model police itself retrospectively: quarterly for four years, every newly defaulted loan is re-checked against the residual-income rule, and where the rule fails the deficiency is waived and deletion of the credit-bureau tradeline requested — with the qualifying window widening as the original forecast worsened, eighteen months for a score of 501 or below, twelve for 502 to 600, six for 601 or above.

The class the settlement calls "Mandatory Relief Consumers" is the single richest object in the record, because it defines the harm by a four-way conjunction rather than by any one failure: a loss forecasting score of 501 or below, purchase of the vehicle from a dealer while that dealer sat on Santander's high-risk or dealer performance management list, Santander buying the loan while the dealer was on that list, no proof of income obtained, and default. Four controls the organisation already possessed, none of them holding on the same file.

The remedy's own limit is where the loan had been funded. The judgment defines "Owns" as on the company's balance sheet and not part of a securitization. Loans that had left the balance sheet were reachable only by repurchase at or below the price they were sold for, on best efforts, within 150 days of the effective date, with a per-consumer accounting of each attempt and each refusal handed to the states; the prospective back-test reaches securitised loans only "to the extent permitted by the relevant securitization documents." Whether a given borrower got relief depended in part on where their loan had been funded. That is a fact about ownership and standing to forgive. It is not a claim that risk was transferred to investors, and this file does not make one: the operator's securitisations were largely on-balance-sheet secured financings with approximately $26 billion outstanding, its treasurer said the higher losses were visible to investors and that bondholders were protected by loss cushioning, and Moody's made no claim that noteholders were at risk. The federal securitisation inquiries produced no securitisation finding. A Department of Justice civil subpoena under the Financial Institutions Reform, Recovery, and Enforcement Act sought documents on the underwriting and securitization of nonprime auto loans since 2007 and was last disclosed as open in the FY2019 Form 10-K, absent from FY2020, with no public resolution located. The Securities and Exchange Commission's investigation, opened as a securitization-practices inquiry in October 2014, resolved on 17 December 2018 as an accounting and internal-controls case about the credit loss allowance for certain impaired loans, with a $1.5 million penalty and no admission or denial.

Three further matters bound the case rather than extending it, and each names a different surface. On 22 December 2020 the Consumer Financial Protection Bureau found that between January 2016 and August 2019 the operator furnished credit-bureau information it knew or reasonably should have known was inaccurate, failed to promptly correct it, omitted dates of first delinquency and lacked reasonable written accuracy policies, imposing $4.75 million under the Fair Credit Reporting Act and Regulation V — a furnishing failure, on the very channel the multistate remedy then uses to deliver its tradeline deletions. The Department of Justice's two Servicemembers Civil Relief Act consent orders — 25 February 2015 in the Northern District of Texas over 1,112 vehicles repossessed without court orders between January 2008 and February 2013, at least $9.35 million plus a $55,000 fine, and 1 October 2021 over ten denied early lease terminations — concern repossession process and lease administration and contain no model at all. And Mississippi, never a coalition member, litigated separately from January 2017 and settled on 21 July 2021 for $3.7 million including $1.8 million of consumer restitution.

The record has not gone quiet since. On 18 February 2022 Massachusetts took a $5.56 million assurance of discontinuance for more than 1,000 borrowers over insufficient disclosure of how post-repossession deficiency balances were calculated. On 3 June 2026 the New York State Department of Financial Services settled for a $400,000 penalty plus more than $275,000 of restitution over undisclosed recurring monthly extension fees where the disclosure documents showed only a single $25 fee — approximately $237,000 collected and about $86,000 assessed but uncollected, on conduct predating 2018 that the company had voluntarily ceased. That is the same servicing joint the 2020 judgment's subparagraphs 18(o) and 18(p) had addressed, examined by a different regulator six years later.

Two things have to be said at the end because they govern everything above. Every characterisation of the operator's knowledge, intent and practice in this file is an allegation. The May 2020 judgment was entered "without the taking of proof and without trial or adjudication of any fact or law", is expressly "not ... an admission by Santander regarding any issue of law or fact alleged in the Complaint", is inadmissible in other cases, and states in terms that it "does not constitute an approval by the Signatory Attorney General of Santander's business practices". The Massachusetts, Delaware, Mississippi and New York matters resolved the same way. And the record contains no channel by which a borrower ever saw the loss forecasting score, learned what it predicted, or contested it. In the parallel Mississippi settlement, restitution eligibility turned on an internal score that, per the state's own communication, neither the consumers nor the Attorney General's office nor the settlement administrator were to know.