FRC Research · June 2026

6 Correlations Nobody
Found in Mortgage Data.
Until Now.

FRC Intelligence June 28, 2026 HMDA · Correlation Analysis · 8-year data

We ran correlation analyses across CFPB HMDA data (2018–2025), FHFA house price index, Fed unemployment, MBA delinquency data, and SLOOS credit standards. What we found contradicts conventional wisdom on nearly every front.

Finding 01 ⭐ r = +0.618
FHA Denial Rates Predict Delinquency 2 Years Out

When FHA denial rates rise, FHA delinquency rates follow — exactly two years later. The lag correlation is r = +0.618, strong enough to be actionable.

YearFHA Denial RateFHA Delinquency (2 yrs later)
202023.1%5.1% (2022)
202123.4%8.7% (2023)
202231.1%10.2% (2024)
202330.0%11.5% (2025)

Denial rising in 2022–2023 predicted the 2024–2025 delinquency surge — which reached 11.5%, the highest since 2021.

The 2025 denial rate is declining (27.1%). If the pattern holds, delinquency should ease by 2027. But FHA delinquency is still elevated — the divergence is worth watching.

Why it makes sense: Rejected borrowers don't disappear. They find alternative lenders with looser standards. Those loans perform worse. The denial rate is a leading indicator of who gets credit — and at what quality.
Finding 02 ⭐ r = −0.696
The Recession Paradox: Unemployment Up, Denials Down

Higher unemployment correlates with lower mortgage denial rates. r = −0.696. This is counterintuitive and consistent across the full dataset.

YearUnemploymentAvg. Denial Rate
20183.9%35.0%
20208.1%23.1%
20215.4%23.4%
20233.6%30.0%

When unemployment is high, the Fed cuts rates to near zero. Low rates reduce the DTI burden on every borrower — applications that would have failed at 6.5% sail through at 3%. The lender doesn't get more lenient. The math just gets easier.

"Getting a mortgage is easiest when the economy is worst" — the Fed's rate response creates a window that benefits borrowers even as the labor market deteriorates.
Finding 03 r = −0.583
Home Prices Rising = Denial Rates Falling

When home prices appreciate faster, denial rates fall. The correlation is r = −0.583 — moderate but consistent.

YearHPI ChangeAvg. Denial Rate
2020+11.0%23.1%
2021+17.1%23.4%
2022+7.2%31.1%
2025+2.2%27.1%

Rising home prices increase existing homeowners' equity, reducing effective LTV on refinances. They also tend to occur during low-rate periods. The 2025 HPI deceleration to +2.2% is the weakest in the dataset — if the pattern holds, denial pressure may build.

Update — June 2026: NAR's May 2026 Existing-Home Sales report confirms this deceleration. Median sales price rose just +1.3% YoY to $429,300 — consistent with the weak appreciation environment our model flagged. Months of supply held at 4.5, down only slightly from 4.6 a year earlier. The slowdown we identified in the HPI data is showing up in real-time NAR figures.

Further update: NAR's Pending Home Sales Index rose 3.8% month-over-month in May 2026 (+4.8% YoY) — the strongest signal of recovering demand in nearly four years, even as the 30-year fixed rate held in the mid-6% range. NAR Chief Economist Lawrence Yun called this "consumers' acceptance of above-6% mortgage rates as the new normal." This complicates Finding 02 (the recession paradox) — buyers are now moving despite elevated rates rather than waiting for them to fall, which may decouple future denial rates from rate-driven dynamics in ways our 2018-2025 model doesn't yet capture. Worth re-testing once 2026 HMDA data is available.

HPI is a denial leading indicator — not because of sentiment, but because of LTV math and the macro conditions that drive appreciation.
Finding 04 r = −0.735
The Composition Fallacy: Cash-Out Ratio Falls When Prices Rise

Conventional wisdom: when home values rise, homeowners cash out. The data says otherwise — cash-out ratio falls as HPI rises. r = −0.735.

YearHPI ChangeCash-Out Ratio
2019+5.5%0.416
2020+11.0%0.208
2021+17.1%0.301
2023+5.7%0.534

In 2020–2021, rate-and-term refinancing exploded as rates hit historic lows. Total refi volume surged — but cash-out's share of that volume fell, even as absolute cash-out volume increased. The ratio is deceiving. This is a composition effect, not a behavioral one.

Anyone reading cash-out ratios without controlling for total refi volume is measuring the wrong thing. The headline number moved down while the actual dollars moved up.
Finding 05 ⭐ r = +0.61 to +0.66
The Fed Sensitivity Gene: Some Lenders Track the Fed. Others Don't.

When the Fed raises rates, three lenders' denial rates move up sharply. Two major lenders are almost completely immune.

loanDepot
r=+0.66
Freedom
r=+0.61
PennyMac
r=+0.54
Rocket
r=−0.09
CrossCountry
r=+0.02

loanDepot, Freedom, and PennyMac are rate-driven underwriters. When the Fed tightens, they tighten. Rocket and CrossCountry operate on a different model — their denial behavior is nearly decoupled from Fed policy.

Practical implication: in a tightening cycle, route away from rate-sensitive lenders. In an easing cycle, they'll loosen faster too. The "Fed gene" cuts both ways.
Finding 06 ⭐ r = −0.150
SLOOS vs HMDA: Lender Intent ≠ Lender Behavior

The Fed's Senior Loan Officer Opinion Survey (SLOOS) measures what lenders say about their standards. HMDA measures what they do. The correlation between the two is r = −0.150 — essentially zero, and in the wrong direction.

YearSLOOS (net tightening %)Avg. Denial Rate
2020+37% (major tightening)23.1% (low!)
2021−8% (easing)23.4%
2022+20% (tightening)31.1%
2019−5% (easing)26.9%

In 2020, lenders reported major tightening in SLOOS — yet denial rates hit a multi-year low. The refi boom brought in so many easy-to-approve refinancing profiles that overall denial rates fell, even as purchase standards tightened.

SLOOS measures the lender's intent. HMDA measures the applicant pool's outcome. They're different things. Regulators and analysts who use SLOOS as a proxy for what's actually happening in approval markets are measuring the wrong signal.

This has regulatory implications: if policymakers rely on SLOOS to assess credit availability, they may be systematically misreading the market. HMDA tells the truth. SLOOS tells the story lenders want to tell.
Finding 07 · NEW NAR + HMDA cross-reference
Student Debt May Explain Why Younger Borrowers Hit the DTI Cliff Hardest

NAR's 2026 Generational Trends Report shows a striking pattern in homebuyer debt: 39% of Younger Millennials carry student loan debt (median $30,000), compared to just 7% of Older Boomers (median $24,000). The share with student debt declines steadily with age.

GenerationHas Student DebtMedian BalanceFinanced Purchase
Younger Millennial39%$30,00080%
Older Millennial27%$40,00078%
Gen X18%$32,00076%
Younger Boomer11%$28,00068%
Older Boomer7%$24,00058%

We don't have age data in HMDA — lenders aren't required to report borrower age, so we cannot directly correlate student debt with our DTI cliff findings. But the mechanism is plausible: a $30,000-$40,000 student loan balance adds meaningfully to monthly DTI. Our Finding from the Genome Project shows several major lenders carry a DTI-44 cliff — a hard threshold where denial rates spike. Borrowers carrying significant student debt are mechanically more likely to land in that 44%+ band than debt-free borrowers with identical income.

This is a hypothesis, not a proven causal link — we're combining two separate datasets (NAR survey data and CFPB HMDA data) that don't share a borrower-level join key. But the directional logic holds: younger buyers carry more student debt, student debt raises DTI, and several lenders apply sharp denial penalties exactly at the DTI thresholds where added student debt would push a borrower.

Practical implication: young borrowers with student debt should pay particular attention to which lenders carry steep DTI-44 cliffs (PennyMac, Freedom) versus which don't (Guild, CrossCountry) — the DTI threshold may matter more for this group than for any other.
🪦 The Hypothesis Graveyard
Hypotheses we tested — and killed

Most research pages only show what worked. We think the graveyard is half the credibility. These hypotheses were reasonable, we tested them with real data, and they failed. We're publishing them so you don't have to trust our wins on faith.

"State bankruptcy rates predict FHA denial tightening" — REFUTED

Plausible mechanism: bankruptcies signal regional stress, lenders tighten locally. The level correlation exists (r=+0.36, n=50 states, 2023 US Courts filings per capita vs FHA denial) — but the predictive test failed: bankruptcy intensity showed no relationship with subsequent denial-rate change 2023→2025 (r=−0.12), and the lag structure added nothing (t and t+2 nearly identical, the signature of a shared static factor, not a leading signal). The level link is mostly a poverty artifact, further contaminated by Chapter 13 legal culture: Alabama and Tennessee file at high rates partly because of local bankruptcy practice, not purely economic distress. A county-level test with chapter controls remains open — but the easy version of this hypothesis is dead.

"When a lender's profits fall, it tightens the next year" — REFUTED

Plausible mechanism: shrinking profitability cuts risk appetite; underwriting tightens with a lag. We built a panel of the five publicly traded lenders in our dataset (Rocket, UWM, PennyMac, loanDepot, Guild) using company-reported GAAP net income 2020–2024 against our HMDA denial series. The naive pooled correlation looks alive (r=−0.30) — but it's a mirage. Demean each transition year across lenders (removing the industry-wide cycle) and the lender-specific signal collapses to r=−0.09. What the naive number was really capturing: in 2021 everyone's profit crashed and in 2022 everyone's denials rose — rates drove both. Which lender bled most did not predict which lender tightened most. The profitability→underwriting channel is a market-wide phenomenon, not a lender-idiosyncratic one. This is also a case study in why pooled panel correlations lie without fixed effects.

"The mortgage-Treasury spread leads denial rates" — REFUTED

Plausible mechanism: the spread prices secondary-market risk appetite; underwriting should follow. Result: contemporaneous r=+0.12, one-year lag r=+0.04, and the sign flips between sample halves. At annual resolution, there is no there there. If a spread signal exists, it lives at monthly frequency — untestable with public annual HMDA.

Think we killed one wrongly, or want to nominate a hypothesis? [email protected] — refutations get credited.

"The mortgage market has six hidden feedback loops. None of them work the way the textbooks say."

See how lender behavior patterns translate to your specific borrower profile.

Route with Zai — free →

Compare lender behavioral DNA across all 11 major FHA lenders.

Lender Similarity Engine →

New: these findings led to Shock Beta — the first metric predicting which lenders swing hardest in the next macro shock.

⚡ Shock Beta →
← HMDA Selection BiasFRC RESEARCH SERIESShock Beta →
Data sources: CFPB HMDA Public Loan-Level Dataset 2018–2025 (FHA) · FHFA House Price Index (Purchase-Only, National) · U.S. Bureau of Labor Statistics Annual Unemployment Rate · MBA National Delinquency Survey (FHA 90+ day, Q4 annual) · Federal Reserve SLOOS (Senior Loan Officer Opinion Survey, net tightening balance, mortgage category) · Federal Reserve Bank of St. Louis FRED. All correlations computed using Pearson r. Lag analysis: 1-3 year offset. N=6-8 annual observations per correlation. Sample size is limited; findings are hypothesis-generating, not definitive. All-cash purchases excluded from HMDA analysis. For educational purposes only — not financial advice.

← All Tools  ·  HMDA Selection Bias  ·  Mortgage Genome Project  ·  Lender Similarity Engine  ·  Shock Beta  ·  Terms  ·  Privacy