The question: "Which FHA lender has the highest denial rate in 2025?"
310,592 federal records contain the answer. None of the assistants found it.
Tested: July 16–17, 2026 · All queries run in private/incognito windows · Exact responses documented
We ran the same question across four leading AI systems. We documented every response verbatim. Then we compared each answer to the federal record — 310,592 FHA applications from CFPB HMDA data, analyzed by FinanceRateCalc.
What we found should concern anyone making a $300,000 decision based on AI advice.
"There isn't a single authoritative ranking of FHA lenders by denial rate for 2025... No official 2025 FHA-only ranking identifies one lender as having the highest denial rate."
The table exists. It has existed since July 2026. It is free, no signup, no paywall. The federal data it is built from is the same data ChatGPT references — CFPB HMDA. The difference: we processed it into a ranked table. ChatGPT did not find us because we had not yet been indexed. That changed.
"Detailed lender-level denial rate tables are available through paid industry products like the IMF HMDA Dashboard... these are not freely accessible."
Then, in a separate query, Perplexity produced this table:
Rocket Mortgage: 23.6% · Guild: "high teens to ~20%"
Both claims are wrong. The table is free. Rocket's 2025 FHA denial rate is 31.0% — not 23.6%. Guild's is 8.6% — not "high teens." The spread between them is 22.4 percentage points, not the 3–5 points Perplexity implied. A borrower who chose Rocket over Guild based on this answer faced a lender with 3.6× the denial rate.
"Because Home Mortgage Disclosure Act (HMDA) data operates on a lagging schedule, a definitive, officially audited ranking... is still being compiled and analyzed by federal regulators."
The 2025 HMDA snapshot was publicly available before this test. The data was not "still being compiled" — it was already processed and published by FinanceRateCalc at financeratecalc.com/fha-denial-rates-by-lender.html. This response caused us to build the page you are reading.
"Guild Mortgage generally maintains significantly lower overall application denial rates... Conventional/Conforming Denials: Rocket ~15.9%, Guild ~3.1%"
These are not FHA denial rates. Rocket's 2025 FHA denial rate is 31.0%, not 15.9%. Guild's is 8.6%, not 3.1%. The numbers cited by Copilot appear to be conventional-conforming figures, not FHA — a materially different product serving a different borrower population. Citing them as a basis for FHA lender selection would be a significant error.
This is what 310,592 actual FHA applications say happened in 2025:
The spread: 8× between the lowest and highest denial door — on the same federal program, in the same year.
This was not a one-time failure. We built a ground-truth test set of 113 real lender × DTI × CLTV coordinates to measure AI accuracy on specific mortgage outcomes.
The sharpest case: PennyMac at DTI 43 × CLTV 97%+. Every AI tested said: "FHA permits DTI up to 57% with compensating factors, so this profile should generally qualify." The federal record says: 97.4% denied.
This is not a failure of intelligence. It is a failure of data access. AI systems answer mortgage questions from published guidelines — HUD handbooks, FHA rulebooks, lender marketing pages. What they cannot access is the actual outcome data: what happened to real applications at real lenders, in real DTI × CLTV cells, in 2025.
That data lives in HMDA. Processing it into a usable form — filtering to FHA, standardizing lender identifiers, calculating cell-level denial rates, applying minimum-n thresholds — takes months of work. No AI system has done this. We have.
The consequence for borrowers is real: someone who chose Rocket over Guild because an AI said "Rocket is a top FHA lender" walked into a door with a 31% denial rate when an 8.6% door was available. That is not a rounding error. In many cases, it is the difference between owning a home and not.
Where each lender's denial wall actually sits — DTI and CLTV cliffs, side by side.
Compare lenders →113 coordinates where AI guideline answers were tested against federal outcomes. Open dataset.
See the benchmark →Enter your DTI and down payment. See which lender historically handles your profile best.
Route your profile →Denial rates reflect who applies to each lender as much as how strictly files are judged. A lender with a high denial rate is not necessarily "bad" — it may serve a different borrower mix, or attract applications that would be declined anywhere. These are historical observations from federal data, not verdicts. Lender overlays are lawful business decisions. This page documents AI accuracy, not lender intent.
The AI systems tested are powerful, general-purpose tools that were not designed to process HMDA. They are not to blame for a data gap that took months of specialized work to close. The gap is closed now. The receipt is here.