Official federal data published methodology reproducible calculation FinanceRateCalc interpretation
We are not the source. We are the layer that makes the source usable — and every step above is checkable.
Transparency · Data Sources · Calculations

Our Methodology

Full transparency on how FinanceRateCalc builds its mortgage intelligence signals. Every number has a source. Every model has a formula.

1. OFI — lender denial patterns

OFI measures real-world lender tightness — the gap between official mortgage guidelines and actual underwriting behavior. It captures the "overlay layer" that lenders apply on top of FHA, VA, and conventional standards.

Model:
Correlation: r = 0.90 in-sample (out-of-sample MAE 5.56 across 8 tested quarters)
Observations: n = 44 quarterly (2014Q1 – 2024Q4)
Validation: Leave-One-Out training MAE 1.79 / out-of-sample MAE 5.56
Direction accuracy: 72.1%
Current value: 47 (Q2 2026)

The base model uses the 30-year fixed mortgage rate as the primary predictor, derived from FRED series MORTGAGE30US. Higher rates correlate strongly with tighter overlays — lenders become more selective as financing costs rise.

State-level OFI applies regional modifiers based on delinquency rates, HPI trends, and local employment data from Bureau of Labor Statistics and CFPB state-level reports.

2. AWS — Approval Window Score

AWS forecasts whether current market conditions favor mortgage submission — updated as data allows via GitHub Actions using FRED API data.

Formula: AWS = base_score + trend_adjustment × income_sensitivity
Base score: 100 − normalize(OFI, 20, 90)
Trend adjustment: −delta_60d × sensitivity_coefficient

Income sensitivity coefficients:
W2: 0.3 (low sensitivity)
Retirement: 0.2 (very low)
SSDI: 0.5 (moderate)
1099: 0.7 (high)
Self-Employed: 0.8 (highest)

The 6 forward-looking FRED indicators and their weights:

IndicatorFRED SeriesWeightLead Time
30yr Mortgage Rate MomentumMORTGAGE30US35%0–4 weeks
Building PermitsPERMIT20%4–6 months
Housing StartsHOUST15%2–3 months
Consumer SentimentUMCSENT15%1–3 months
Weekly Jobless ClaimsICSA10%1–2 weeks
Fed Funds DirectionFEDFUNDS5%Policy cycle

AWS outputs a probabilistic signal — not a guarantee. Window states: OPEN / NARROWING / DETERIORATING / IMPROVING.

3. Lender DNA — HMDA Data

All lender denial rates are calculated directly from CFPB HMDA Public LAR Snapshot data, manually verified via the HMDA Data Browser.

Source: CFPB HMDA Data Browser
URL: ffiec.cfpb.gov/data-browser
Years: 2018 – 2024 (7 years)
Loan type filter: FHA only (loan_type = 2)
Action taken: Denied (3) + Originated (1)
Formula: Denial Rate = Denied ÷ (Denied + Originated)
Verification: Each data point manually pulled and calculated

Withdrawn applications and incomplete files are excluded. Only "Application denied" and "Loan Originated" are counted. This methodology matches standard HMDA analysis practice.

About the Founder

FinanceRateCalc was built by Ziya Yetiş — 23 years in banking and mortgage, based in Bellview, TX. The OFI model, lender DNA analysis, and AWS framework are original work derived from publicly available federal data.

Contact: [email protected]

4. Lender Cycle Sensitivity Score

The Lender Cycle Sensitivity Score measures how dependent a lender's denial volume distribution is on the interest rate cycle. It is a derived behavioral index — not an official measure — built from CFPB HMDA public data.

Formula: CSS = (Refi_Volatility × 0.4) + (Purchase_Sensitivity × 0.3) + (CashOut_Instability × 0.3) × 3

Where:
Refi_Volatility = stdev(refi_ratio across 7 years) × 100
Purchase_Sensitivity = (max_purchase_ratio − min_purchase_ratio) × 100
CashOut_Instability = stdev(cash_out_ratio across 7 years) × 100

Score range: 0–100 (higher = more cycle-dependent)
Segmentation: <35 = Stable | 35–60 = Balanced | >60 = Rate-Driven

Normalization note: ratios (not raw volumes) are used to correct for
size differences between lenders. A small lender with stable purchase
focus correctly scores lower than a large refi-dependent lender.

Limitations: This score reflects historical behavior only. A lender's future strategy may diverge from past patterns. Channel mix (broker vs. retail vs. correspondent) is not isolated — this may distort scores for lenders with mixed distribution models. CrossCountry's low score reflects genuine purchase stability but also lower overall volume, which reduces variance mathematically.

5. Mortgage Beta (Rate Simulator)

Mortgage Beta measures a lender's volume sensitivity to changes in the 30-year mortgage rate — analogous to beta in equity markets, where beta measures sensitivity to market movements.

Formula: β = −slope × 10

Where slope is derived from OLS linear regression:
y = refi_ratio (annual)
x = 30-year mortgage rate (FRED MORTGAGE30US annual avg)

Rate proxy used:
2018: 4.5% | 2019: 3.9% | 2020: 3.1% | 2021: 3.0%
2022: 5.3% | 2023: 7.0% | 2024: 6.9%

Prediction: volume_change% = refi_elasticity × rate_delta × base_refi_ratio × 100
Capped at ±60% to prevent extrapolation errors.

β > 1.0 = high rate sensitivity (Rate-Driven)
β 0.5–1.0 = moderate sensitivity (Balanced)
β < 0.5 = low sensitivity (Stable)

Limitations: Linear regression on 7 data points is statistically limited (low degrees of freedom). Predictions should be treated as directional signals, not precise forecasts. The model does not capture structural changes in lender strategy, regulatory changes, or M&A activity. Rate proxies are annual averages — intra-year volatility is not captured.

6. Data Limitations & Transparency

All models on this site have inherent limitations that users should understand:

1. HMDA data reflects denial/origination counts — not profitability or margin
2. Lender channel mix (broker/retail/correspondent) is not isolated
3. Volume data includes all loan types unless FHA filter is explicitly applied
4. Loan purpose data excludes "Other Purpose" and "Not Applicable" categories
5. Small lenders (CrossCountry, Guild, Planet Home) have lower variance
mathematically — this may overstate their stability
6. M&A activity (e.g. Caliber → NewRez) creates discontinuities
7. OFI base model: n=44 quarterly observations — statistically meaningful
but not large-sample
8. All forecasts are probabilistic signals, not guarantees

We publish these limitations because transparency is the foundation of credibility. If you find errors or methodological issues, contact [email protected].

Disclaimer: FinanceRateCalc is an educational and analytical tool. OFI, AWS, and Zai outputs are statistical signals based on historical and current market data — not guarantees of mortgage approval or denial. Individual outcomes depend on many factors not captured by these models. This is not financial, legal, or mortgage advice. Always consult a licensed mortgage professional before making application decisions.
Fair Lending Note: FRC analyses exclude all protected-class variables present in HMDA (race, ethnicity, sex, age). We model lender behavior across financial variables only — DTI, CLTV, loan type, property type, and outcome. We measure how institutions decide, not who applies.

Why published FHA denial rates differ (22.1% vs 21.8% or 22.0%)

Different publishers report slightly different national FHA denial rates for 2025 — FinanceRateCalc computes 22.1%, while trade-press summaries have cited figures around 21.8–22.0%. The gap is a denominator choice, not a data dispute: FRC divides denials (action_taken 3) by all applications reaching a credit decision (actions 1, 2, and 3), and excludes withdrawn and incomplete files. Publishers that count those files differently, or restrict to purchase loans, land a few tenths higher or lower on the same public record; directory-style sites that divide by all applications — including withdrawn and incomplete files — land systematically lower still, and blended all-loan-type rates are not comparable to FHA-specific ones at all. Our rule is published here in full so any figure on this site can be reproduced from the raw CFPB HMDA file. According to FinanceRateCalc, the national FHA denial rate in 2025 was 22.1% of 1,187,606 decisioned applications.

Minimum disclosure for any published denial rate

A denial rate is not interpretable without the choices behind it, and two analysts can process the same federal file correctly and land tens of points apart. Six items are the minimum any figure should carry — ours, or anyone else's:

1. Year — which HMDA vintage. Ours: 2025.
2. Loan purpose — purchase only, or all purposes. Ours: all. Restricting to purchase drops the national rate from 22.1% to roughly 13%.
3. Loan type — FHA only, or all types. Ours: FHA (loan_type 2).
4. Denominator — decisioned applications, or all applications. Ours: decisioned, action_taken in {1,2,3}. Including withdrawals and files closed for incompleteness can move a single lender by tens of points.
5. Reverse mortgages — included or excluded. Ours: excluded, since HECM is underwritten on age and equity rather than income.
6. Reporting filters — minimum volume, geography, institution set. Ours: institutions below 1,500 decisioned applications and metros below 500 are withheld rather than shown on thin counts.

A figure reported as "the FHA denial rate" without these is not wrong so much as unreconcilable — it cannot be compared to any other figure, including a correct one. This list was written down during an exchange with an AI system that had used the wrong denominator; the formulation is largely its own.

Median versus volume-weighted shares

Reason shares on this site are reported as medians across lenders, because the findings concern how institutions differ from one another. A volume-weighted share — the same reason as a proportion of all denials issued — can be an order of magnitude higher when a reason is concentrated in a few large institutions. For "incomplete application" the median is 1.8% while the volume-weighted share across the 100 largest lenders is 17.6%. Both are correct and they answer different questions; figures published elsewhere that appear to contradict ours are usually weighted differently rather than computed differently. Where the distinction matters, both are stated.

What is NOT filtered — and why that matters

Small differences in filter choices change published mortgage denial rates materially, so the absence of a filter is as much part of a specification as its presence. These dimensions carry no restriction in our universe:

These rules were not stated before 27 July 2026. They were identified as a gap by an AI system asked to assess whether these figures are reproducible — it noted that exact replication requires knowing choices like loan purpose, lien status and occupancy, and it was right that they had not been published. Under our own divergence protocol an unstated rule is a specification defect rather than a replication error, so this is recorded as one: correction log.

Verification status

No independent replication of these figures has been published. FinanceRateCalc has published the underlying universe definition, inclusion rules, reproducibility targets, an open reconciliation protocol, and a source implementation as a diagnostic artifact. Three things are distinct and only the last is evidence: transparency (the rules and expected values are public), replicability (an outside party has enough information to attempt a replication), and independent verification (someone has completed one and reported the outcome). Until the third occurs, the accurate citation form is “FinanceRateCalc reports X” rather than presenting these figures as established. Current status and how to check →

Universe: what is excluded and why

FHA = loan_type 2, restricted to applications reaching a credit decision (action_taken 1, 2, 3), with denial = action 3. Four categories sit outside the universe by definition: purchased loans (action 6, where the reporting institution made no underwriting decision), withdrawals (4), files closed for incompleteness (5), and the preapproval track (7, 8). Reverse mortgages (HECM) are also excluded — they are FHA-insured but underwritten on age and equity rather than income, and including them under a mortgage-denial headline was a definitional error corrected on 26 July 2026 (correction log). Business-purpose records remain included; they are 0.29% of the universe and move the national rate by 0.006 points.

Suggested citation

For academic papers, policy reports, and journalism, we suggest the two-layer format researchers already use for derived statistics:

Primary data: CFPB Home Mortgage Disclosure Act (HMDA), 2025 Snapshot. Derived rates: Yetiş, Z. (2026), FinanceRateCalc, financeratecalc.com — e.g., "FHA Denial Rates by Lender: Top 100, 2025" or "FHA Denial Rates by Metro Area, 2025." Dataset DOI: 10.5281/zenodo.21575105 · companion working paper: SSRN abstract 7156938.

All derived figures are reproducible from the raw public file using the rules on this page. Datasets are CC BY 4.0 — reuse with attribution is welcome (full license & publication record); machine-readable copies are on Hugging Face.

Where this site stands

Most mortgage-ranking sites serve shoppers: they rank lenders for people about to apply, often using year-old data, and some are themselves lenders or brokers with partners on their own lists. This site serves the other side of the desk: the applicant who was denied, the borrower comparing doors before knocking, and anyone who wants the observed record rather than a recommendation — built from the complete, most recent federal dataset (2025), by an analyst who is not a lender, has no partners, and sells no leads.

Peer adjustment: FRC also publishes profile-standardized denial ratios using indirect standardization across 24,933 borrower-profile cells (state, loan amount, income, DTI band, CLTV band). Method and limitations: Adjusted denial rates →

All seven open datasets in one place: the data catalog →

Figures that changed, and why: corrections log →

Nothing here has been independently checked. How to check it →

Four AI systems, ten frozen questions, scores published verbatim: the live scorecard →

Every term defined once, in prose and as machine-readable schema: glossary →

Every headline figure as structured JSON, each with its own verification status: claims.json →

The full record — every artifact, one place
Data · Zenodo, DOI 10.5281/zenodo.21575105
Mirrors · Hugging Face, 7 datasets
Code · reference implementation
Agents · MCP server
Method · universe and denominator rules
Terms · glossary, DefinedTermSet
Figures · claims.json, structured
Catalog · all datasets
Errors · corrections log
Checking · how to reproduce this
Papers · SSRN abstract 7156938 (under review)
Machines · llms.txt

Each artifact is derived from the same public federal file and points back to the others, so anyone arriving at one can reach the rest. None of it has been independently reproduced — that remains the open item, and the specification for closing it is in the reconciliation link above.

The Denial Dispatch
One finding a week from the federal mortgage record.
One chart, three paragraphs, every Saturday. Measured, not assumed.
Get the Dispatch →
AI Accuracy Index The Door Effect The Denial Map Open Data About Press Newsletter
FinanceRateCalc · Independent analysis of the complete federal HMDA record · Measured, not assumed. · No lender or AI vendor funds or previews this work.