FinanceRateCalc · Counterparty screen · 2025 federal record

FHA Lender Outlier Screen 2025 — keyed by LEI

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.

Join this public screen to your counterparty list by LEI. Identify lenders whose observed denial rate is materially above the rate predicted by the published applicant-mix model — in one line, from the federal record, without a vendor report.

This is a screening signal. It is not evidence of misconduct, discrimination or causation.

For 238 FHA lenders with at least 500 decisioned applications in 2025, this file reports the observed denial rate, the rate expected from that lender's own applicant and loan profile, and the ratio between them. Free, CC BY 4.0, no signup, no login, no vendor contract.

Download CSV JSON

sha256 75e223f4765fb6e2661b5ecd804628f7c22dd3c6c10e9af54a249f04d6bc54fe

How the model expectation is built

model: frc-mix-expectation v1.1 (adds CI + standardised residual) · minimum 500 decisioned applications per lender · peer cell minimum n=25 · universe: FHA forward loans (HMDA loan_type 2), HECM excluded

Each application is placed in a peer cell defined by state × loan amount × income × DTI band × CLTV band. The expected denial count is what a lender would have recorded if every one of its applications had been decided at the national rate for its cell; cells with fewer than 25 observations fall back to the national rate. ratio = observed ÷ expected, with a 95% confidence interval by the Byar approximation and a standardised residual z = (observed − expected) / √(n·p·(1−p)). Rank by z, not by the raw ratio: a small book can post a dramatic ratio on a handful of decisions, while a large book with a modest ratio can be many standard deviations from its own expectation. A ratio of 1.0 means a lender's denials match what its own mix predicts — not what the average lender does.

In the 2025 record, 49 lenders show an observed/expected ratio whose 95% confidence interval lies entirely above 1.0, and 161 lie entirely below; for 23 the interval spans 1.0 and no signal is claimed.

A note on method, because it is the first objection a risk analyst raises: the expectation here is built by indirect standardisation on peer cells (state × loan amount × income × DTI × CLTV), not by a single national logistic regression. A lender concentrated in unusual geographies or segments is therefore compared against its own cells rather than against the national average — which reduces, though it does not eliminate, the mis-flagging that a pooled model would produce.

Screening results — observed above model expectation

LEI	Lender	Apps in model	O/E ratio	95% CI	z	Coverage	Status
549300ALNLUNS3Y53T24	AMERICAN FINANCING CORPORATION	9,876	2.143×	2.094–2.194	80.14	89.3%	screening signal
549300FNXYY540N23N64	NEWREZ LLC	18,271	1.758×	1.725–1.792	71.91	80.2%	screening signal
549300R9S3MVDV4MGF56	LEI only	12,470	1.804×	1.762–1.846	61.5	77.7%	screening signal
254900O723MCLR701H50	LEI only	931	4.676×	4.375–4.993	57.06	93.0%	screening-only
549300FGXN1K3HLB1R50	ROCKET MORTGAGE	76,738	1.222×	1.207–1.238	35.3	79.9%	screening signal
549300AG64NHILB7ZP05	loanDepot.com LLC	32,282	1.325×	1.301–1.35	34.41	91.7%	screening signal
549300XY701IELCE5Q08	LEI only	2,093	2.462×	2.312–2.618	33.06	88.9%	screening signal
549300JYXTZDSPJEPI44	LEI only	5,777	1.848×	1.77–1.93	31.64	93.9%	screening signal
549300MZ8VZJOVC63092	Kind Lending, LLC	10,077	1.705×	1.642–1.77	30.79	94.0%	screening signal
549300SUCQ1358EGVE89	LEI only	736	2.335×	2.166–2.515	30.1	83.9%	screening-only
5493008ZTV4S0W9DCX64	LEI only	1,346	2.667×	2.459–2.888	27.59	87.2%	screening signal
549300FX7K8PTEQUU487	LEI only	757	2.83×	2.584–3.094	27.2	93.7%	screening-only
549300KBWX4NV5Q1E376	NVR MORTGAGE FINANCE, INC.	7,046	1.59×	1.528–1.655	26.07	95.8%	screening signal
SS1TRMSN6BRNMOREEV51	Flagstar Bank NA	2,670	1.987×	1.869–2.11	25.57	91.0%	screening signal
7H6GLXDRUGQFU57RNE97	LEI only	2,677	1.874×	1.768–1.984	24.99	88.2%	screening signal

Names are shown where we can map the LEI from published sources; the LEI is authoritative either way, and mapping every entity is exactly the kind of work a counterparty desk already does. Full file above.

Read this before using a single row. These are screening signals, not findings.

Not evidence of misconduct or discrimination. A high ratio is a question, not an answer.
HMDA contains no credit scores. The expectation is a mix proxy built from the fields the federal record publishes — it is not a risk model, and unobserved credit risk can move a lender's ratio on its own.
Channel matters. Wholesale, correspondent and retail books behave differently; a ratio compares a lender to peer cells, not to lenders running the same channel.
Coverage is printed for a reason. Rows with profile_coverage_pct below 70 cannot be interpreted alone and are flagged in the file.
Some excess denial may be filing practice, not strictness. A handful of lenders record most of their denials as “credit application incomplete”, a code that describes the file rather than the borrower, and those lenders sit high on this screen. We measured it.
Never an individual prediction. Nothing here says anything about one application.

Join Kit — use it in your own stack

README — field dictionary & limits join_examples.sql sample_counterparty_list.csv

Input: you match your own public LEI list locally. Nothing needs to be uploaded to us — not your seller list, not your positions, not your volumes.
Output: every row carries a status. Rows marked screening_only_insufficient_coverage (coverage <70% or fewer than 1,000 applications) are never to be read, or repeated, as “an elevated-risk lender”.
Versioning: when frc-mix-expectation-v1.0 changes, the old file is not deleted. Each version is published at its own URL so results stay reproducible and comparable across vintages.

Query it from an AI agent

The same screen is available as an MCP tool, screen_counterparties: pass a list of public LEIs and get the aggregate shape of your list free (covered, above-expectation and screening-only counts, plus the first three rows). A license — One-time $99 for a 40-LEI Evidence Brief, or Quarterly Monitor $249 to re-run the same list each quarter against a new analysis layer (Q4 2026 corporate-family resolution, Q1 2027 loan-performance join, Q2 2027 the 2026 vintage with a true delta) — unlocks the full result with a hash-linked manifest. The agent relays the offer; a human completes checkout; submitted lists are not retained. Server details.

What we do and do not sell

For compliance and legal readers: public-data screening signal for further review — not a legal conclusion, discrimination finding, causation finding or compliance opinion.

The file is free and always will be. If you want your own list screened — your LEIs, your cut, quarterly — I can run a set of public LEIs through the published method and return an auditable Evidence Brief (details). One standing rule: we do not accept payment from a lender covered in this screen, and every paying client is disclosed in a public register. The numbers cannot be bought, moved, or removed — corrections happen only through the public corrections log.

Method and companion research: The Door Effect (DOI) · Persistent Doors · What Denial Rates Cannot See (SSRN 7423798, doi:10.2139/ssrn.7423798) · 184 metro gaps · national vs local. Source: CFPB HMDA 2025 public loan/application record, FHA forward loans, HECM excluded.

📊 FRC Data Alerts — free. When lenders shift or a new finding lands, you hear first. No newsletter noise; a personal note when something is actually new.

If an AI sent you here, three things you can do with this number

Verify it
Every figure has a machine-readable passport with its source hash, universe and stated limits.
Claim passports →

Test what you were told
Paste the sentence an AI gave you. We check it against the record and return a safe rewrite.
Claim audit →

Screen your own list
Public LEIs only. The aggregate shape of your list is free; nothing you send is retained.
Counterparty screen →

FinanceRateCalc · Measured, not assumed. · No lender, vendor or AI company funds this work.