Mortgage Application Outcomes: Theory of AI Assistance and HMDA Measurement Sensitivity

A new academic study argues that AI-assisted mortgage applications can distort the approval statistics regulators publish, with seven of 33 institutional purchase comparisons flipping sign when a single classification convention changes.

The finding strikes at how analysts, investors and the Consumer Financial Protection Bureau read Home Mortgage Disclosure Act data — the public record that underpins most fair-lending scrutiny.

The working paper, “Mortgage Application Outcomes: Theory of AI Assistance and HMDA Measurement Sensitivity,” combines a formal lender-choice model with a descriptive audit of HMDA records covering 2018 through 2025. Its institutional focus is Better Mortgage, the digital lender whose AI assistant expanded during the study window.

Here is why that matters beyond the journal page. Machine-learning underwriting has been promoted — and attacked — on the promise of measurable approval-rate effects. This paper argues the scoreboard itself is unstable.

Mortgage Approval Rates Shift with Denial Reclassification

  • Recorded mortgage approval rates respond to how documentation-related denials are classified — seven of 33 purchase comparisons reverse direction under one common reclassification fraction.
  • Under fixed underwriting rules, AI assistance can raise or lower conditional approval depending on which borrower types return to complete their files.
  • Funded-loan conversion stays intact under the reassignment; the measurement problem is confined to the decision stage, not outcomes on the books.

Two Numbers That Should Not Move Together

The starkest result is a sign reversal. Across 33 selected HMDA reporting entities, seven recent purchase comparisons take both directions as a single denominator-classification fraction varies, according to the paper. Documentation-only comparisons fare worse: eleven become direction-ambiguous at a five-percentage-point classification-spread benchmark, and thirteen at ten points. Those spreads are chosen for sensitivity analysis, not calibrated to observed industry practice.

The crossing point for Better’s approval comparison is quantified. Under a common classification fraction, the approval comparison crosses at 44.5% of the ambiguous-denial pool; after a state-and-channel adjustment, it moves to 60.1%.

Measure Value
Sign-reversing purchase comparisons (33 entities) 7
Ambiguous comparisons at 5-point spread 11
Ambiguous comparisons at 10-point spread 13
Approval crossing, common fraction 44.5% of ambiguous-denial pool
Approval crossing, state-and-route adjusted 60.1%

But the balance sheet tells a different story — or rather, the same one. The paper finds that observed funded conversion remains invariant to this reassignment. The loans that actually close do not move when the paperwork categories do.

Why AI Assistance Muddies the Denominator

The theory half of the paper explains the mechanism. Borrowers choose a lender before learning their private documentation costs. An AI assistant scales that burden — it can explain requirements, collect information and walk applicants through a file. Lower friction means marginal applicants who would have abandoned the process now reach a credit decision.

The math admits two outcomes. If lower-approval borrower types expand fastest into the applicant pool, the lender’s conditional approval rate falls even though nothing changed in the underwriting rule. Under the reverse conditions, it rises. A “worse” approval rate can therefore accompany more completed applications and more funded loans.

The authors model this with general differentiable documentation and perceived-cost functions. Privacy concerns register as a negative — anticipated data exposure — while trust and convenience improvements register as a positive. The literature supports both signs: experimental work documents algorithm aversion alongside algorithm appreciation, and the paper cites both without claiming to know which dominates in mortgage markets.

What HMDA Records Can and Cannot Show

The measurement problem turns on administrative resolution. A documentation-related denial enters the recorded decision denominator; an applicant who simply never closes out the file does not. The paper shows how the relative prevalence of those two groups can change recorded resolution and approval rates while leaving funded-loan counts untouched.

HMDA does not observe assistance intensity, usage, or individual documentation histories. The Better setting also layers technology rollout atop broader distribution changes. The authors are explicit that their institution-period comparisons are descriptive illustrations of a measurement problem, not estimates of the correct latent classification or of AI effects.

Study after study in this space — papers by Fuster and coauthors on fintech processing speeds and predictive underwriting, work on insurance acquisition bundling information production with customer selection — treats recorded approval rates as data. This paper treats them as both data and convention.

What Analysts Watching Better Should Carry Forward

For anyone benchmarking Better’s AI-driven turnaround against peers, the practical implication is sharp: comparisons exceeding one percentage point can swing in either direction under defensible classification choices, at either sensitivity benchmark the authors tested. Ratios inside that band are effectively noise. Funded-volume figures, by contrast, sit outside the blast radius of the denominator problem.

The 32-peer comparison ranges across classification endpoints by multiples of a percentage point, and the narrower documentation-only cohort shows the same fragility across six entities shared between alternative definitions. The count of sign reversals holds at seven even when ambiguity is restricted to denials reporting documentation reasons alone.

Where the Debate Goes From Here

The framework distinguishes the anticipated option at lender choice from the realized continuation cutoff — a distinction the authors offer as their contribution to the application-frictions literature. Whether the industry’s growing stack of AI assistants really moves credit access, the paper says, cannot be settled from public HMDA records alone, because borrower-level exposure is latent.

Until an independent measure of assistance usage exists, published approval-rate rankings between AI-forward lenders and traditional banks — in either direction — rest partly on a classification decision the regulator’s data leaves undefined. The funded-loan count, the one number in the comparison the reclassification cannot touch, is the number the cross-checking starts with.

Disclaimer: The information provided in this article is for educational and informational purposes only and does not constitute financial advice.

Photo of author

Newsroom

Newsroom at Archyde.

Duke Professor Emeritus Phil Cook Discusses Homicide Clearance Rate Decline