MARKETS, CREDIT & POLICY DAILY INTELLIGENCE
c.The Credit CurrentDAILY INTELLIGENCEWhat matters in Credit
Deep-dive library
Law & regulation

ECOA: making adverse-action reasons explain the actual decision

A decision-trace framework for adverse-action explanations, now including an evaluation checklist for hybrid underwriting systems.

Current version · 2 versions · Publication details

Added a current hybrid-underwriting example, explanation-fidelity tests and the September 17 Affirm sources; the earlier article remains available.

Version history

In this article

The central obligation

Regulation B, which implements the Equal Credit Opportunity Act, makes the explanation of a credit decision part of the decision process. Section 1002.9 generally requires notice within 30 days after receiving a completed application. Its written adverse-action framework calls for specific principal reasons, or notice of the applicant’s right to request those reasons. This article focuses on completed consumer applications; incomplete applications, counteroffers and business credit have additional provisions.

The official interpretation says reasons must reflect factors actually considered. A notice that merely says an applicant did not meet internal standards, or did not achieve a qualifying score, does not supply the required specificity. The operational implication is straightforward: a lender needs to reconstruct why this application failed under the policy and model version used at the time.

A score explanation is only one part of the chain

A decision can involve eligibility rules, an affordability calculation, a score cutoff, fraud review and an underwriter’s judgment. The interpretation addresses combined systems and automatic denial factors. It also distinguishes the principal reasons for an adverse credit decision from the key factors affecting a credit score. Those are related records, but one cannot automatically stand in for the other.

Analysis: create a decision trace before drafting the notice. It should connect the source data, transformations, rule outcomes, model output, human override and final action. If an override supplies the decisive reason, a score explanation alone may describe a step that did not determine the outcome. A polished explanation cannot repair an incomplete decision record.

Illustrative application

Assume a fictional applicant passes an identity check and the credit-score threshold but fails a documented debt-to-income limit. The relevant explanation is the actual income-and-obligations issue, expressed accurately for the consumer. Selecting “insufficient credit history” simply because that is a readily available model reason would misdescribe this example.

Now change the facts: the debt ratio passes, but a model cutoff causes the decline. The lender must identify the principal factors behind that model-based outcome. A method that ranks globally important variables across the portfolio may not explain this applicant. This is an analytical example, not prescribed notice language or a conclusion about any lender.

An operating control map

The following is a proposed control design. Its purpose is to make errors visible before they propagate to thousands of notices.

Scroll horizontally to see all columns.

StageEvidence to retainFailure to catch
ApplicationInput values, sources and completeness dateWrong data or an incorrectly started notice clock
DecisionRule and model versions; override rationaleReasons taken from the wrong decision component
ExplanationPrincipal-factor mapping and approved wordingGeneric wording or a factor not actually considered
DeliveryNotice contents, timing and delivery recordAccurate explanation sent too late
MonitoringSample reconstructions and complaint feedbackA mapping defect repeated across a product

Testing beyond a readable letter

Analysis: sample both approvals near the cutoff and declines, across products and decision paths. Reproduce each decision from the preserved inputs, then compare the proposed reasons with the decisive factors. Include missing data, jointly decisive factors, overrides and changed model versions. Test whether small, irrelevant input changes produce large or implausible reason changes.

A second review should consider whether consumers can understand the wording. That review complements technical fidelity. Simpler words are useful only when they still convey the actual reason. The official commentary notes that giving more than four reasons is generally unlikely to help; that observation is not permission to omit a principal reason or mechanically fill four slots.

New evidence: explaining a hybrid underwriting decision

Affirm’s September 17 announcement describes a proprietary explanation method for its hybrid underwriting system. It is a useful current example of the explanation challenge, rather than independent proof that a particular notice satisfies Regulation B.

Analysis: validate the full decision path, including eligibility rules, transformations, model output, cutoff, overrides and the reason selected for the notice. A faithful explanation of one component may not explain the final decline. Use preserved inputs to reproduce the decision and test the notice against the factors actually considered.

Test missing or corrected bureau information, conflicting policy and model outcomes, joint reasons, and a switch between model versions. Record disagreements and correction ownership. Improvements in conversion or risk ranking do not answer these explanation questions.

What would change the assessment

An improved model that increases approvals can still create an explanation problem if its reasons cannot be validated. Conversely, a complex model is not automatically unusable because it is complex. The relevant evidence is repeatable decision reconstruction, specific reason fidelity, fair-lending review and effective correction when errors surface.

This article uses the current CFPB-hosted regulation and official interpretation accessed for this revision. It does not treat older AI circulars as an independent statement of current enforcement policy. Revisit the article when the regulation, authoritative interpretation or a material judicial decision changes the notice requirements; preserve the prior research version for comparison.

Sources

  1. CFPB — Regulation B §1002.9 and official interpretation
  2. CFPB — Appendix C model notification forms
  3. Affirm — underwriting model announcement, September 17
  4. Affirm Technology — hybrid model design and evaluation