A new FEDS Note from Tranfaglia and Troland (May 2026) put a number on something most of us have suspected for a decade: when you classify a consumer as a “renter” because they don’t have an open mortgage tradeline, the classification is correct only 42% of the time. That figure is the positive predictive value of the renter label — of the consumers your stack flags as renters, fewer than half actually rent. The true-renter cohort runs a 60+ DPD rate of 41% over the FEDS observation window; the broader proxy-renter cohort runs 28%. A third of true renters carry a credit score under 620, against 23% of the proxy group. The signal isn’t noisy in the abstract — it is wrong in a structured, lopsided way (FEDS Note, May 2026).

The point of this piece is not “your data is dirty.” The error has a shape. The 58% of “renters” who are not actually renters are disproportionately your lowest-risk customers: homeowners who paid off the mortgage, older borrowers who downsized into a cash purchase, and homeowners whose mortgage doesn’t furnish to the bureaus because their lender is small or local. The proxy mildly understates risk for genuine renters and significantly overstates risk for equity-rich homeowners. You are penalizing the wrong people, and the people you are penalizing are concentrated in geographies that matter for fair-lending review.

That asymmetry is the entire story. It changes what the audit looks like, what the regulatory exposure looks like under the post-July 21 Reg B regime, and what you tell your MRM committee. A 30-day audit, a vendor question set, and an AAN language pass are the cheapest defensible posture available this quarter. The rest of this piece is how.

Where the proxy actually lives in your stack

The first objection is usually “we don’t use a renter feature.” That’s literally true and operationally wrong. The mortgage-presence signal — and its inverse, the no-mortgage signal — is loaded into your decisioning in at least four places, and most risk teams have never inventoried all four.

Policy and segmentation cuts. Champion/challenger configurations (the parallel-decisioning structure where a candidate model is shadowed against the production model), line-assignment matrices, and prescreen filters routinely key off mortgage-derived attributes: “months since oldest mortgage tradeline,” “presence of open mortgage,” “ever-mortgage flag.” A no-mortgage cohort gets routed to a different waterfall, a different starting line, a different score cutoff. That routing is a renter-proxy decision whether or not anyone called it that.

Scorecard attributes. If you built or bought an attribute library, you have mortgage utilization, time-since-mortgage-open, count of mortgage tradelines, and similar. These appear as features in your internal scorecards. The coefficients are visible to you; the inferential content is identical to a renter flag.

Vendor scores. Both FICO and VantageScore use mortgage-related features in their models. You cannot see the weights, and your vendor contract likely does not entitle you to. This is the most overlooked surface area, and it’s the one with the highest leverage under the 2023 Interagency Guidance on Third-Party Relationships — more on that below.

Marketing and account-management logic. Prescreen waterfalls, credit line increase (CLI) eligibility rules, and authorized-user / cross-sell triggers very often gate off mortgage-presence attributes. The CLI rule that throttles a paid-off homeowner because she “looks like a renter” is the textbook adverse-action-by-omission case, and it almost never shows up in the model inventory because nobody thinks of CLI logic as a model.

Inventory all four before you decide whether you have a problem.

The asymmetry of the error

The 58% error is not uniform. Of the consumers your stack classifies as renters, 32% are homeowners without an open mortgage tradeline — they paid it off, they bought in cash, or they hold a reverse-mortgage product that doesn’t furnish like a conventional first lien. This cohort skews older, equity-rich, and lower-risk than the average book. The proxy is penalizing them.

A further 16% are homeowners whose mortgage exists but does not furnish to the bureaus. The underlying lenders are small community banks, credit unions, CDFIs, manufactured-home portfolio lenders, and seller-financed portfolio holders (FEDS Note, May 2026). The geography of those lenders is the part that should make you sit up. They disproportionately serve lower-income and minority neighborhoods. When the proxy mis-tags those borrowers as renters, the error is concentrated in exactly the census tracts your fair-lending team monitors. This is not a hypothetical disparate-impact concern; it is a structural one baked into bureau furnishing patterns.

The FEDS analysis also notes that combining the no-mortgage signal with an under-50 age filter pushes recall on true renters to roughly 69% with precision around 56%. That is interesting as decomposition evidence — it confirms the proxy’s error is concentrated in older borrowers — but it is not a fix you can adopt. ECOA prohibits age-based discrimination generally. The two carve-outs are narrow: age may be used as a variable in an empirically derived, demonstrably and statistically sound credit scoring system (an EDDSS), and elderly applicants — defined as age 62 and above — may not be assigned a negative factor or value within that system. A blanket under-50 cut sits outside the EDDSS safe harbor and is a disparate-treatment exposure. Mention the filter in your audit memo only to explain the shape of the error.

The reader who pushes back here will say: “the score already captures this — older homeowners with no mortgage will have long, clean credit histories and high scores; the proxy doesn’t add anything the model isn’t already seeing.” That is a testable claim, not a settled one. The Week 3 step of the audit below is exactly how you test it on your own book. The answer is segment-dependent. For your subprime band, the score absorbs most of the differential; for your near-prime and prime bands, mortgage-presence features carry residual variance that the rest of the score does not. Don’t assume; measure.

Three consequences your committee will ask about

Loss forecasting and CECL segmentation. When 32% of your “renter” segment is actually equity-rich homeowners, segment-level reserves are mis-allocated across two underlying populations with materially different risk. The blended loss curve attributed to the proxy-renter segment understates loss for the true-renter portion and overstates loss for the misclassified-homeowner portion. The blended number may look stable; the components do not. You cannot argue that the segmentation is clean, and a methodologically-aware examiner — citing SR 26-2’s expectation that segmentation be conceptually sound — can ask for the decomposition.

Rent-reporting partnership ROI. Vendors marketing rent-reporting products quote lift numbers — VantageScore has reported an 11% predictiveness improvement from rent inclusion (VantageScore, vendor-published, methodology not independently verified); Esusu has reported 45–53 point average score lift (Esusu marketing materials, vendor-published). The lift is calculated on the true-renter denominator. You pay for the rollout against the proxy-renter denominator, which is 58% non-renters. The unit economics that worked in the pitch deck do not survive contact with your actual book until you’ve audited the denominator. If a rent-reporting partnership is already in flight, run the denominator audit before the next contractual milestone — renegotiate, re-scope, or hold.

Account-management treatment streams. A paid-off homeowner held out of CLI eligibility because she lacks an open mortgage is being adverse-actioned on a proxy that is wrong about her. If the AAN language anywhere in your stack maps that decision to a “housing status” or “homeownership” reason, you have a CFPB Circular 2022-03 problem on top of the underlying classification problem.

The inverse cell deserves a brief note. A small share of consumers your stack treats as homeowners are actually renters — old mortgage tradelines that closed, joint-account spouses, authorized-user histories. That cohort is much rarer than the homeowner-mistagged-as-renter case but creates a different exposure: overstated creditworthiness for someone with true-renter risk. Confirm it is small in your data; do not assume.

Product differentiation

A misclassified paid-off homeowner is a different problem at $500 of BNPL exposure than at $25,000 of unsecured installment exposure. The audit’s urgency should scale with the product.

The audit is a low-priority exercise for a BNPL-only book and a high-priority one for any lender doing card or unsecured installment at meaningful balance. Prioritize accordingly.

Where the proxy is actually fine

Worth saying directly: the no-mortgage proxy is not useless. The hypothesis worth testing on your own book is that in a specific segment — under-35, thin-file (fewer than five tradelines), urban-CBSA — the proxy carries materially higher positive predictive value. The no-mortgage signal genuinely correlates with non-ownership for a young, urban, thin-file population that has not yet had a first-time-buyer credit event. The FEDS Note doesn’t break out PPV at that level of granularity, so treat this as a hypothesis to test in Week 3 of the audit, not as a finding to import.

If your book is concentrated in that segment — younger urban card portfolios, fintech-originated personal loans to under-35 borrowers, student-loan-refi books — the proxy is probably approximately fine for the population you actually underwrite. Document the segment-level PPV. That documentation is itself the defense. Many institutions use internal benchmarks in the 65–75% segment-PPV range as the threshold for retaining a proxy feature; SR 26-2 does not mandate a number, but the absence of any benchmark in your MRM policy is itself a finding.

The vendor-model lever

For most lenders, the highest-leverage action is not retraining an internal scorecard. It is exercising the rights and obligations created by the 2023 Interagency Guidance on Third-Party Relationships, which puts the bank — not just the vendor — on the hook for understanding how vendor-supplied scores and attributes use protected-class-correlated features.

Send your FICO, VantageScore, and attribute-product vendors a written question set. Six questions, in this order:

1. Which features in your model use mortgage-presence signals, directly or indirectly (including derived features such as time-since-oldest-mortgage)?
2. What is the magnitude of their contribution at the score-band level, particularly in the 600–720 range where decline/approve decisions concentrate?
3. How do you treat mortgages that do not furnish to the major bureaus? Are unfurnished mortgages effectively coded as “no mortgage”? And how does the model treat a thin-file consumer who has a single furnished tradeline from a community bank or credit union?
4. What documentation can you provide on disparate-impact testing of mortgage-presence features, including geographic concentration of misclassification?
5. What is your guidance for adverse-action reason mapping when these features materially drive a decline, consistent with CFPB Circular 2022-03?
6. How do you handle paid-off-mortgage cases in attribute construction? Are they coded as “ever-mortgage” or as “no current mortgage” — and which features in the model use which encoding?

Inability or unwillingness to answer is itself a finding. Escalate it to your MRM committee and document the escalation. Under SR 26-2 (April 17, 2026), the bank’s MRM function retains responsibility for materiality determination on vendor models; “the vendor wouldn’t tell us” is not a posture that survives examination, but the documented attempt to ask is.

If your vendor will not answer, and your book has sufficient volume to support a development and validation effort, an internal scorecard with explicit no-mortgage handling becomes the option of last resort. For most lenders it will not come to that — the vendor question set is leverage enough.

The 30-day validation playbook

This is the centerpiece. It is a memo to your MRM committee, not a publication. The output exists to inform a materiality determination. Plan ownership across functions: Credit Risk Analytics typically owns Weeks 1 and 3, Compliance and Fair Lending own Week 4, and Vendor Management runs the parallel question-set track. Time the memo to land at least two weeks before the next scheduled MRM committee meeting.

Week 1 — Cohort sizing. Pull your last 90 days of declines and approvals. Tag each record with proxy-renter status as of the decision time (defined consistently: no open mortgage tradeline on the bureau pull used for the decision). Quantify the cohort. Population-wide, only about 25.1% of consumers have a mortgage tradeline (CFPB, Debt Burdens Among Credit-Linked Consumers, April 2026); applicant-pool concentrations vary, but no-mortgage is the majority cohort for most non-mortgage products and is consistent with that population-wide figure.

Week 2 — Ground truth. Source two independent ground-truth signals, ideally in parallel.

The first is a property-records vendor — ATTOM, CoreLogic, BlackKnight, or similar — to identify homeownership status by deed and tax record. Caveats: LLCs and revocable trusts as title holders mask homeowner status; manufactured-home titles often live in DMV records rather than county recorder records; recent purchases lag in the property data by 30–90 days depending on county. Expect to miss on the order of 5–10% of true homeowners with property records alone, with the rate sensitive to geography and title-vehicle mix. If your institution does not have one of these vendors under contract, plan for 4–8 weeks of procurement before Week 2 can execute. Many fraud or wealth-segmentation teams already license a property data feed; the credit risk audit can often share that contract rather than start a new one.

The second is permissioned bank-account access via commercial open-banking aggregators (Plaid, MX, Mastercard Open Banking / Finicity, and similar) where you can obtain consent. Recurring rent payments to property-management companies, landlord ACH templates, and rent-platform deposits give a reasonable detection signal for active renters. Acknowledge the selection bias: consumers who connect bank accounts in your application flow are disproportionately credit-seeking and in-application, which is a different mix than your back-book. Use the connected sample as a one-directional confirmation — rent payments observed implies high-confidence renter — not as a denominator. The Section 1033 final rule (October 2024, currently enjoined as of May 2026) would have standardized this channel; until that resolves, treat the connected sample as directional.

Week 3 — Four-cell decomposition. Cross your proxy classification against ground-truth tenure: proxy-renter × true-renter, proxy-renter × true-homeowner, proxy-homeowner × true-renter, proxy-homeowner × true-homeowner. For each cell, compute the misclassification rate, the 60+ DPD rate over a forward 12-month window for booked accounts, and the average score. This is where you answer the “the score already captures it” objection on your own data. If the score gap between proxy-renter-true-homeowner and proxy-renter-true-renter cells is small, the score is doing the work. If it is large, the mortgage-presence features are carrying residual signal that affects outcomes.

Two caveats to record in the memo. First, the 12-month DPD comparison is conditioned on the booking decision — you observe outcomes only for accounts that passed underwriting, which is selection on the outcome variable. The cleanest read is on a held-out random-approval cohort if you have one; otherwise note the selection bias and consider a reject-inference adjustment. Second, the proxy-homeowner-true-renter cell is usually sparse (often under 5% of file). Disclose minimum cell size; any cell below an n that supports stable forward-DPD estimation should be flagged as directional, not conclusive.

Week 4 — Geographic overlay. Join the cohort to census tract using ZIP or applicant address, and overlay tract-level minority share from the ACS five-year file. Compute the misclassification rate by tract-minority-share decile. This step typically requires partnership with the Fair Lending or HMDA team that owns tract-level geocoding — credit risk analytics rarely owns that pipeline. The output is a heat map your fair-lending team can act on. Frame it carefully: tract-level analysis is ecological inference, and the ecological fallacy applies — geographic concentration of error is not individual disparate impact. If individual-level analysis is needed, BISG (Bayesian Improved Surname Geocoding) or BIFSG is the conventional next step, with its own caveats around classification precision for smaller demographic groups.

The memo to MRM should contain the four-cell table, the geographic overlay, the AAN language pass (next section), and a materiality recommendation. SR 26-2 leaves the materiality call to the MRM function; this memo is what informs it. A realistic resourcing estimate: roughly 1.5 FTE-months of analytics time and 0.5 FTE-month of compliance/fair-lending time, plus vendor procurement if a property-records feed is not already on roster.

The model-inventory template

Every place mortgage-presence attributes appear in your stack should be on a single template. The point of the template is not the audit — the point is that you can answer the examiner’s question “where do you use this signal” in one document.

The point of populating it is not the rows you fill in. The blanks — the models without a “Last validated” date, the AM rules nobody catalogued, the vendor scores with “Unknown” feature lists — are the finding.

AAN cleanup

CFPB Circular 2022-03 requires adverse-action reasons to be specific and accurate principal reasons for the decision. Inferential housing-status language describes the consumer; data-description language describes the file. The second is materially safer.

A note on FCRA: §1681e(b)’s “maximum possible accuracy” standard applies primarily to consumer reporting agencies and furnishers, not directly to lenders. Lender exposure on the proxy runs through use of the data and the reasoning given in the AAN — not through a direct FCRA accuracy hook. The CFPB Circular and Regulation B at 12 CFR §1002.9 are the operative authorities here.

Regulatory posture in three lines

Reg B (effective July 21, 2026; final rule, Federal Register Docket 2026-07804). The federal ECOA disparate-impact pathway closes — the effects-test language in 12 CFR §1002.6(a) and Supplement I commentary is deleted, and CFPB has signaled it will not pursue ECOA disparate-impact theories at the federal level. Disparate treatment via intentional proxy use survives. State attorneys general — NY, CA, MA, IL, DC, and others — retain effects-test pathways under their own UDAP and fair-lending regimes. DOJ Civil Rights has independent ECOA enforcement authority. The private bar continues to litigate effects under §1981 and state law. NY DFS Circular Letter No. 1 of 2019 on external consumer data is operative and unaffected. California’s emerging ADMT framework adds state-level exposure for automated decisioning. “Reg B sunset” is shorthand for one specific federal pathway; it is not a deregulation of fair lending.

SR 26-2 (April 17, 2026). The Fed’s replacement for SR 11-7 is principles-based and materiality-driven, consistent with the OCC and FDIC companion guidance. SR 26-2 does not say “a feature with 42% accuracy is automatically material.” Materiality is the MRM function’s call. A feature documented at 42% PPV, with concentrated error in protected-geography cohorts, is difficult to call immaterial in good faith. The audit memo’s job is to make the materiality call defensible whichever way it goes.

CFPB Circular 2022-03. AAN reasons must be specific and accurate principal reasons. “Renter status” is an inference about the consumer; “limited or no mortgage credit history” describes the data underlying the decision. The second survives Circular 2022-03 scrutiny; the first invites it.

Document the decision not to act

If, after the audit, your MRM committee determines that the proxy is materially accurate within your segment and that the AAN language is clean, memorialize it. The memo should contain: the segmentation analysis, the four-cell decomposition, the false-positive rate by score band, the geographic overlay, the AAN language review, the vendor question-set responses, and the MRM sign-off.

The documented decision is the defense under both MRM and fair-lending self-test conventions. Failure to test is one exposure. Testing and not memorializing the conclusion is a worse one — you’ve created the analysis without creating the record. Write the memo even when the answer is “no action required.”

Takeaways for a VP

• The no-mortgage renter proxy returns the right label only 42% of the time, and the errors concentrate in your lowest-risk customers — paid-off homeowners and homeowners served by lenders that don’t furnish to the bureaus (FEDS Note, May 2026).
• It is not labeled “renter” in your stack — it lives in scorecard attributes, vendor scores (FICO, VantageScore), policy cuts, and account-management logic. Inventory all four.
• Reg B’s federal disparate-impact pathway expires July 21, 2026. State AG, DOJ Civil Rights, and private-bar effects-test pathways do not expire.
• SR 26-2 makes this an MRM materiality question. Document the determination either way; the record is the defense.
• A 30-day audit costs roughly 1.5 FTE-months of analytics plus 0.5 FTE-months of compliance time. A six-question vendor question set and an AAN language pass cost less. Together they are the cheapest defensible posture available this quarter.