FinanceRateCalc · Evidence Navigator · v0

What are you really asking? We'll show you exactly where the evidence ends.

This tool does not predict approvals and never recommends a lender. It maps your question onto the complete 2025 federal HMDA record (1,187,606 FHA decisions), shows what that public evidence supports — with receipts — and tells you honestly when your question needs a licensed professional instead. Nothing you click is stored; no personal information is asked, ever.

SUPPORTED BY THE RECORD

About 38% of the explainable variation in 2025 FHA denial outcomes attaches to lender identity rather than the applicant profile (859,090 decisions; McFadden pseudo-R² 0.171 → 0.276). Raw rates across the 100 largest lenders span 1.8% to 78.7% under the same federal program.

full analysis · receipt · DOI

Boundary: a share of explained variation, not of all denials; associational, not causal — HMDA contains no credit scores. It is not a probability for any application.
SUPPORTED BY THE RECORD

Often enormously. The widest measured 2025 gap is Cleveland, OH: 6.4% vs 80.1% among high-volume lenders — a 73.7-point spread inside one metro. We publish this gap for 184 metros, each with a machine-readable receipt.

all 184 metros · Cleveland receipt

Boundary: unadjusted observed rates among lenders with ≥100 decisioned applications in the metro; consistent with overlays, channel and applicant mix — not evidence of misconduct.
SUPPORTED BY THE RECORD

It barely moves. Across 2018–2025, 80.8% of panel variance sits between lenders; the year-to-year rank correlation is 0.918, and the softest door was the same institution in 7 of 8 years. We've even pre-registered ten predictions about the 2026 data.

Persistent Doors, DOI · dataset

Boundary: institution-level history; four lenders show dated regime shifts (documented in the paper). History is not a guarantee of any future year — that's why the predictions page exists.
OUTSIDE WHAT THIS EVIDENCE CAN ANSWER

No honest answer exists in this data. The federal record contains no credit scores, no full underwriting file, and aggregate rates are not individual probabilities — anyone who converts them into “your odds” is guessing with confidence.

What to do instead — questions that get real answers from a licensed lender: “What are your FHA overlays on credit score and DTI, beyond FHA minimums?” · “Do you manually underwrite refer/eligible files?” · “Can I see your overlay sheet in writing?”

Why we refuse: converting group rates into personal predictions is the exact failure we measure in AI systems. We won't do it ourselves.
WE NEVER RECOMMEND LENDERS

No lender pays us and none ever will — so we don't rank “best” doors and we don't route applicants. What the evidence does support: denial rates differ persistently and enormously, so shopping more than one door is the single most data-supported move after a denial.

Compare observed rates yourself: top-100 table · your metro

Boundary: observed rates reflect applicant mix and channel, not just standards — a “soft” door for the average mix may not be soft for your file.
THE RECORD CANNOT SEE THIS

HMDA reports income amounts, not income types — the public record cannot distinguish W-2 from 1099 from SSDI. Any lender-by-income-type approval claim you see online is not computable from this data.

Real-answer questions for a licensed lender: “How many months of 1099 history do you require for FHA?” · “Do you accept SSDI award letters as qualifying income, and gross-up how much?” · “What documentation triggers manual underwrite for self-employed files?”

This is a data-boundary refusal, not a policy opinion: we answer only what the record can support.

Method & boundaries for everything above: research hub · public corrections log · agent access: llms.txt

FinanceRateCalc · Measured, not assumed. · Nothing stored, nothing sold, no lender relationships.