Adverse action reasons

Adverse action reasons

How lenders can use per-decision explanations to draft the specific reasons an adverse-action notice needs, and what they still have to check themselves.

Not legal advice This page describes how Kaimei's output can support a compliance process. It is not legal advice; your counsel decides what your notices must say.

#What the rules ask

In the United States, the Equal Credit Opportunity Act and Regulation B require a creditor that takes adverse action on a credit application to give the applicant a statement of specific reasons, or to tell them they can ask for one. The Fair Credit Reporting Act adds notice duties when the decision relied on a consumer report.

A complex model does not change the requirement. Regulation B asks for specific reasons that indicate the principal reasons for the adverse action (12 CFR 1002.9(b)(2)), and its commentary says they must relate to and accurately describe the factors actually considered or scored. Appendix C adds that simply checking the closest reason on a sample form does not meet the requirement when it is not the factor actually used. The CFPB's circulars on this point (2022-03 and 2023-03) were withdrawn in May 2025; the regulation itself did not change. The commentary to Regulation B notes that more than four reasons is unlikely to help the applicant.

#How Kaimei helps

  • Specific reasons per decision. Each explanation ranks the inputs that pushed this applicant toward decline, with the size of each push.
  • Reproducible. KernelSHAP is exact for models with up to 12 inputs, and a fixed seed makes the other methods repeatable, so the same application always gives the same reasons.
  • On the record. Each explanation is stored with your application id, the model version, the method and the reference population, for audits and complaints.
  • What would have changed the outcome. A counterfactual search, with protected and unchangeable inputs held fixed.

#A worked example

The applicant from Reading an explanation was declined. Compliance keeps a reason table that maps each model input to approved wording, and lists the inputs that must never be stated. The pushes toward decline, strongest first, become draft reasons:

InputValuePush toward declineDraft reason
debt_to_income0.38+8.0 ptsExcessive obligations in relation to income
num_delinquencies1+7.5 ptsDelinquent past or present credit obligations with others
home_ownershiprent+4.6 ptsNot in the reason table. Held for review.
credit_score681+2.1 ptsCredit score
age34+1.3 ptsNever stated. Age is a prohibited basis under the Equal Credit Opportunity Act, with narrow exceptions. Escalate for model review.
employment_years4+0.7 ptsLength of employment

Income, utilisation and the loan amount pushed toward approval, so they are not reasons for this decline. The notice lists up to four approved reasons once review has cleared the two flags; the example wording follows the style of Regulation B's sample forms, and your own table should be written and approved by compliance.

#Doing it in code

For the engineers on your team: the reason table is a plain mapping kept under change control. The explanation's top_features are already ranked, and target_label confirms which way a positive push points.

Every model input should appear either in the reason table or in the never-state list, so an unmapped input is caught before a notice goes out.

REASONS = {  # owned by compliance, under change control
    "debt_to_income": "Excessive obligations in relation to income",
    "num_delinquencies": "Delinquent past or present credit obligations with others",
    "credit_score": "Credit score",
    "employment_years": "Length of employment",
}
NEVER_STATE = {"age"}

explanation = client.explanations.create(
    model="credit-risk-v3",
    prediction_id=application_id,          # your id for this application
    input=application,
    method="shap",
    options={"seed": 0, "top_k": 20, "counterfactual": True, "frozen_features": ["age"]},
)
assert explanation.output["target_label"] == "decline"   # positive pushes point toward decline

reasons, review = [], []
for row in explanation.top_features:                     # strongest first
    if row["attribution"] <= 0:
        continue                                          # pushed toward approval
    if row["feature"] in NEVER_STATE:
        review.append(f'{row["feature"]} pushed toward decline: escalate')
    elif row["feature"] in REASONS:
        reasons.append(REASONS[row["feature"]])
    else:
        review.append(f'{row["feature"]} has no approved reason')

print(reasons[:4])
print(review)
Outputfor the applicant above
['Excessive obligations in relation to income', 'Delinquent past or present credit obligations with others', 'Credit score', 'Length of employment']
['home_ownership has no approved reason', 'age pushed toward decline: escalate']

#Choices that matter

  • The reference population. Reasons are pushes relative to the average applicant in the reference data. Choose that population deliberately, such as recent applications, document it, and refresh it when your policy changes. A different population can reorder the reasons.
  • The method. Use KernelSHAP with a fixed seed, so a complaint months later can be answered with the same reasons.
  • Held-fixed inputs. In counterfactual searches, hold protected and unchangeable inputs fixed with frozen_features.
  • Records. Pass your application id as prediction_id and keep the explanation as long as your retention policy requires. Kaimei keeps records until you delete them; retention schedules that delete them for you are coming soon.

#What stays with you

  • The actual decision. Reasons must reflect the whole decision, including policy rules applied outside the model, such as cut-offs, overlays and manual reviews. Recording those rules beside each explanation is coming soon; until then, they live in your decision system.
  • The reason table. Its wording, and which inputs may be used at all. A place to keep it in Kaimei, under change control, is coming soon; the wording stays yours.
  • Fair lending testing. Disparate-impact testing is coming soon to Kaimei. Until then, explanations can help investigate a finding but are not a test, and the testing program stays yours either way.
  • The notice. Its content, timing and delivery.