FinanceRateCalc · New finding · 2025 record

A lender is not a number. It is a conditional surface.

Where these numbers come from, and how to check them
Method is published, not proprietary. Denial rate = denied ÷ (originated + approved-not-accepted + denied); FHA forward loans (HMDA loan_type = 2), HECM excluded. Every figure carries a machine-readable claim passport with the source hash.
Peer-visible method paper (DOI) Eight-year panel (DOI) Measurement boundaries (SSRN) Full methodology Claim passports with hashes How our tables reconcile Public corrections log What we tested and found nothing
Independent status, stated plainly: this is an independent re-aggregation of the public CFPB HMDA record, not an official CFPB or FFIEC output, and no figure here has yet been independently reproduced. The underlying data is free under CC BY 4.0 so that anyone can check it. No lender, vendor or AI company funds this work.

Published denial rates treat a lender as one thing: softer or stricter than the rest. Inside the federal record that turns out to be wrong for some of the largest institutions. Compare a lender to its peers inside identical cells — same state, same loan amount, same income, same DTI, same leverage — and the advantage of some lenders reverses depending on the loan.

Largest conditional swing, one lender, same year
50.8
percentage points between its softest and strictest leverage band

What the comparison does

Every published cell has its own denial rate. For each lender with at least 30 decisions inside a cell, we take the difference: cell rate minus lender rate. Positive means the lender denied less than the cell it was competing in; negative means more. Then we average by leverage band. Because the cell rate includes the lender itself, the comparison understates rather than exaggerates the gap.

Every lender that qualifies — not a selected example

Lender<80% CLTV80–90%90–95%95–96.5%>96.5%Swingperm pFlip
United Wholesale Mortgage+14 / +14
129c · 7,083d
-1 / -2
39c · 2,257d
-4 / -4
42c · 2,591d
+2 / +2
334c · 32,960d
-37 / -36
19c · 1,058d
50.80.000u80 -> 80-90
ROCKET MORTGAGE+2 / +2
371c · 18,058d
-12 / -14
5c · 214d
-5 / -5
7c · 319d
-4 / -4
129c · 8,185d
14.10.000u80 -> 80-90
Lennar Mortgage, LLC-5 / -6
110c · 15,658d
-15 / -14
9c · 405d
10.30.003
DHI Mortgage Company, LTD.-1 / -2
7c · 299d
+3 / +3
8c · 306d
-4 / -3
182c · 23,582d
-6 / -7
9c · 445d
8.60.03080-90 -> 90-95
Ruoff Mortgage Company+3 / +3
5c · 182d
+7 / +7
8c · 329d
4.10.023
PENNYMAC LOAN SERVICES LLC+3 / +3
15c · 596d
-1 / -0
8c · 401d
4.00.083u80 -> 95-96.5
Kind Lending, LLC-10 / -11
15c · 1,027d
-7 / -7
5c · 215d
3.30.347
NEW AMERICAN FUNDING, LLC.+4 / +4
6c · 260d
+6 / +6
12c · 543d
1.80.685
CROSSCOUNTRY MORTGAGE, LLC+6 / +7
67c · 3,226d
+8 / +8
19c · 778d
1.50.161
MOVEMENT MORTGAGE, LLC+6 / +6
10c · 402d
+7 / +7
10c · 434d
1.40.335
CMG MORTGAGE INC+6 / +6
29c · 1,622d
+7 / +7
17c · 842d
1.30.421
loanDepot.com LLC-8 / -8
39c · 1,478d
-9 / -10
89c · 6,562d
1.30.480
GUILD MORTGAGE COMPANY+6 / +6
30c · 1,363d
+7 / +7
35c · 1,427d
0.70.538
FAIRWAY INDEPENDENT MORT CORP+5 / +5
5c · 182d
+5 / +6
15c · 606d
0.50.867

Each cell shows cell-equal mean / decision-weighted mean in points, then the number of cells and the lender decisions behind them. Positive = softer than the cell. perm p is a within-lender permutation test: a lender's own cell differences are reshuffled across its bands 1,000 times, and we ask how often a swing this large appears by chance. Flip marks where the sign reverses.

Four things that decide whether this is real

1. Weighting. Both are shown, because they can disagree: the cell-equal mean treats each cell once, the decision-weighted mean lets big cells count more. Where they agree, as they do in the extreme bands below, the pattern is not a weighting artefact.
2. Who counts as a peer. The other lenders inside the same published cell: same state, loan-amount band, income band, DTI band and leverage band, same program, same decisioned universe. The cell rate includes the lender itself, so every gap here is understated rather than inflated.
3. Multiple testing. There are 1,086 cells in this file, and scanning them for a dramatic example is exactly how a finding becomes an artefact. So: every lender meeting the thresholds is in the table above, not a chosen one. The leverage bands come from the published cell definition, fixed long before this analysis. And each lender carries its own permutation test. 5 of 14 lenders show a swing that shuffling cannot reproduce at the 5% level; the rest do not, and are printed anyway.
4. What a gap is not. A cell is not a matched borrower. HMDA has no credit scores, and cells differ in channel, product and documentation practice. An observed gap is consistent with policy, with composition, or with both.

Four tests it had to survive

TestWhat it asksResult
Bootstrap flip locationIs the sign change stable, or an accident of which cells we happened to have?The flip lands between the under-80% band and the 80–90% band in 88% of 2,000 resamples for the largest case, and 99% for the second.
Negative controlIf this were an artefact of how cells are built, the same lender should swing on other dimensions too.Across income bands the same lender swings 4.6 points, against 50.7 across leverage — eleven times smaller.
ShrinkageDo thin bands collapse when pulled toward the lender's own average?The extreme high-leverage band moves from −36.6 to −34.0. The 129-cell low-leverage band does not move at all.
Leave-one-state-outIs one state carrying the whole thing?Across 29 states the gap ranges 48.0 to 58.9 points against a full-sample 50.8. Removing the largest state leaves 48.

One thing we cannot do with this file: a continuous curve of the gap against leverage. The published cells band CLTV, so the transition can only be located between bands, never at a point. Anyone reporting this as “the 80% threshold” is reading more precision than exists — including us, which is why the wording on this page says between bands.

How to say this safely

The published 2025 cell comparison shows a positive peer-adjusted gap in lower-leverage cells and a negative gap in high-leverage cells for two large lenders. The transition is an observational pattern. It is not an approval prediction, not a causal explanation of underwriting, and not a statement that any institution has a policy threshold at any particular level.

The flip, not the average

The more useful statistic is not how soft a lender is on average but where its sign changes. Two of the largest lenders in this table cross from the soft side to the strict side at the same place: between loans under 80% leverage and loans at 80–90%. One of them then goes far further at maximum leverage. A borrower at 70% leverage and a borrower at 97% are not dealing with the same institution, even when the name on the door is identical.

The case that stands out

One large wholesale lender runs 14.2 points softer than its peer cells below 80% leverage — across 129 cells and 7,083 of its own decisions — and 36.6 points stricter above 96.5% leverage, across 19 cells and 1,058 decisions. Not one of those 19 high-leverage cells shows it on the soft side. The same institution, the same program, the same year.

Most lenders are not like this. CrossCountry, Guild, CMG and Movement sit a few points softer than their cells almost everywhere, with swings of 1 to 2 points: their advantage really is a constant. The interesting cases are the ones where it is not.

The question we cannot answer yet

The natural next question is whether a lender's conditional surface moves over time — whether the same institution, on the same kind of file, sits differently against its peers in different periods. It is a good question and we cannot test it here.

The public HMDA record carries the activity year and nothing finer. Application date and decision date are withheld from the public file for privacy reasons, so there is no month, no quarter and no decision date to interact with. Any calendar claim built on year-level data would be invented, so we record the boundary instead: temporal_boundary: not_identified.

If a dated source becomes available, the design is already written down in the data file — one lender, two bands, four pre-specified periods, with period permutation and leave-one-period-out required before anything is published. Writing the design before the data exists is the only way to keep it from being fitted to whatever the data turns out to show.

Why this matters for anyone shopping a file

A headline denial rate is an average over whatever mix of loans a lender happened to receive. If a lender's behaviour is conditional, that average describes no actual borrower. The practical question is not “is this lender strict?” but “is this lender strict on files like this one?” — and for at least some large institutions, the answer changes sharply with leverage.

This is also a caution about our own published tables. The lender ranking we publish is a level comparison. This page is the reminder that a level can hide a reversal.

What this does not show. These are observed, unadjusted comparisons inside published cells. Cell composition, origination channel (wholesale versus retail), product mix and documentation practice differ between lenders and can produce these gaps with no difference in underwriting policy at all. HMDA contains no credit scores, so a cell is not a matched borrower. Nothing here is evidence of misconduct or discrimination, nothing predicts an individual application, and we do not recommend lenders. Cells require 40 decisions; lender rows inside a cell require 30; bands require at least 5 cells. Full data, including confidence intervals and cell counts.

Method family: The Door Effect (SSRN 7309319) · What Denial Rates Cannot See (SSRN 7423798, doi:10.2139/ssrn.7423798), whose first boundary — that an aggregate is not a property of its members — is what this page measures directly · LEI-keyed screen · what we tested and found nothing

FinanceRateCalc · Measured, not assumed.