Analysis

Analysis that changed how the business runs.

Each study starts with a decision someone had to make, uses the company's full history, and ends in a change of practice. Effects are shown relative to a baseline group; the company's own rates aren't published here.¹

Early-warning study

Can the first 90 days of a loan tell us how it ends?

Yes. By day 90, a loan has shown how it will end, and roughly when.

15M+collector notes classified by channel, contact and promise, from 2021 on.

303,044payment promises judged kept or broken.

1.7×the signal of the best lever at signing, carried by behavior in the first 90 days.

Data: Every loan's payment history since 2016, every collector note since 2021, and every payment promise, all extracted and rebuilt nightly by the platform.
Method: Four flags read from the payment ledger at day 90: a late fee, paying under 70% of what's due, a gap of 2.5 times the payment interval, and a missed first payment. Outcomes are tracked by year funded, with open loans counted as still performing, so every effect is a floor. Risk bands come from how similar past loans ended, not from a fitted model.
Finding: A late fee is the strongest single flag. A late fee with slow pace defaults as often as three flags, on 2.3 times as many loans. The pattern holds in every year funded since 2016 and in both states. Promise-keeping stayed inside a 5-point band every year from 2021 to 2026, through the move from people to automation, so it measures the borrower, not the collections process.
Decision: Every open loan now carries a risk band, rebuilt nightly. A 90-day flag plan that flags these loans and monitors them through their first year launched in October 2026, as a pilot in one state with the other as the control.

Defaults by flags at day 90

Relative to loans with no flags, loans funded 2022 to 2024¹

3 or more flags against no flags, every year funded

How many times as often flagged loans defaulted. Recent years are still maturing.

Decision brief

Should we keep writing deferred down payments?

The risk followed the cash paid at delivery, not the down payment written on the contract.

Data: Every deferred-down deal delivered since February 2026, measured against full-down deals sold in the same months to the same credit tiers. Down payments read from the payment ledger, not the contract.
Method: Expected losses at full-down rates for each month and credit tier, compared with actual losses, then split by cash paid at delivery and by whether the rest of the down came in within 30 days. Texas books sales tax inside the contracted down payment; separating it kept about 80% of Texas deals from looking deferred.
Finding: Deals with nothing paid at delivery were 14% of deferred deals and about 70% of the extra losses. Deals that paid the rest within 30 days were lost no more often than full-down deals.
Decision: With early repossessions and first-payment defaults rising, leadership ended deferred down payments in October 2026.

14%of deferred deals had nothing paid at delivery.

~70%of the extra losses came from those deals.

30 daysDeals that paid the rest of the down within 30 days were lost no more often than full-down deals.

Deals still on the books after delivery

Illustration of the pattern, not company rates²

Full down payment Deferred, rest paid within 30 days Deferred, nothing paid at delivery

Models

Models leadership can act on.

A promise-to-pay model built from 227,789 past promises and shipped 5 days after the request. A repossession-value model. Survival and static-pool analysis for policy and lender questions.

Promise-to-pay calibration

Predicted against actual, by decile. Illustration.³

Repossession value, median error

Dollars off the realized value

Before the model$3,914

With the model$1,084

Promise-to-pay: within about 6 points in backtesting, with 4x discrimination.
Answered the question the role was created for: what makes a loan succeed.

Effects compare each group with loans that had no flags at day 90 (or with full-down deals), on the same cohort and the same outcome date. The company's own default and loss rates are not published. Counts as of October 2026.

The chart shows the shape of the finding with made-up curves. The figures in the text come from the review.

The calibration chart illustrates the method with made-up points. The production model was built from 227,789 past promises and landed within about 6 points in backtesting.