FinanceRateCalc · New finding · 2025 record
The builders' own lenders are a different animal
Sort the fifty largest FHA lenders by behaviour rather than by name, and one group separates itself without being asked to. Six of them are the finance arms of homebuilders — D.R. Horton, Lennar, NVR, PulteGroup, KB Home, Century Communities — and they do two things at once that the rest of the market does not.
| Builder-owned (6) | Everyone else (37) | Difference | Permutation p |
|---|
| Denials vs own applicant-mix expectation | 1.26× | 0.74× | +0.53 | 0.0063 |
| Median rate spread over APOR | -0.77 | +0.55 | -1.33 | 0.0001 |
| Raw denial rate | 22.48% | 14.48% | +8.0 | 0.113 |
| Applications in the model | 13,544 | 13,454 | +90 | 0.9928 |
Read the rows in this order
Price. The builders' lenders sit 1.33 points below the rest of the market relative to the benchmark rate, and shuffling the group labels twenty thousand times almost never reproduces a gap that large. This is the cleanest result on the page. It is also the least surprising once said aloud: a rate buydown is a sales incentive, and the incentive is cheaper than cutting the price of the house.
Denials. They deny at 1.26× what their own applicant mix predicts, against 0.74× for everyone else. Note which measure separates them: the raw denial-rate difference is not distinguishable from chance (0.113), while the peer-adjusted one is (0.0063). A headline denial rate would have missed this entirely.
Not size. The two groups are the same size, to within ninety applications on an average of thirteen thousand (p = 0.9928). Whatever this is, it is not a large-lender effect wearing a different hat.
The obvious explanation, and why we are not asserting it
A captive lender does not receive applications the way a retail lender does. Its pipeline arrives through a builder's sales office rather than through rate shopping, which means the applicant pool can differ in ways no cell of state, loan amount, income, DTI and leverage will capture. That would produce exactly this pattern — below-market pricing offered as an incentive, and a higher denial rate against expectation because the expectation is built from a different kind of applicant.
It is a good explanation. It is not a finding. The public record shows the two behaviours; it does not show the funnel that produces them, and we are not going to write a mechanism into a page that only measures an outcome.
Limits, plainly. Six lenders is a small group and every comparison here is permutation-tested rather than assumed. The builder classification is hand-coded from GLEIF legal names — a judgement about names, not a corporate-ownership dataset — and it is published in full so it can be disputed or extended. Rate spread is observed and unadjusted: buydowns, points and product mix all sit inside it. Nothing here is evidence of misconduct, nothing predicts an individual application, and a higher denial ratio is not a statement about anyone's underwriting quality.
Data, classification and all five tests.
Method family: the peer-adjusted screen · conditional surfaces · why a headline rate hides this · what we tested and found nothing
FinanceRateCalc · Measured, not assumed.