Insight
Explainability that regulators accept: SHAP in fair-lending
Move beyond “black box” explanations. Cicrim helps regional and community banks use SHAP to produce examiner-ready narratives, evidence-based reason codes, and durable fair-lending documentation — without slowing down credit decisions.
Jan. 30, 2026 · Authored by Terrence A. Thomas
Regulators don’t ask for “AI vibes.” They ask for evidence: what drove a credit decision, whether the drivers are consistent across populations, and how you detect and correct drift before it becomes a fair-lending event. SHAP (Shapley Additive exPlanations) is one of the most practical ways to translate model behavior into defensible, auditable explanations — when it’s implemented with the right controls.
What examiners expect from explainability
“Explainable” is not the same as “interpretable.” For fair-lending and model risk, the bar is repeatability, stability, and traceability. Your program should be able to demonstrate:
- Reason-code integrity: customer-facing adverse action and internal reason codes align with the underlying model drivers and are consistently generated.
- Population-level evidence: feature influence is examined across protected classes and relevant segments (e.g., product, channel, geography).
- Governed inputs: clear lineage for every feature (source system, transformations, missing data handling, and leakage controls).
- Change control: versioned models, thresholds, and reason-code mappings with approvals and rollback paths.
How SHAP helps — and where banks get it wrong
SHAP produces consistent attributions that can be compared across borrowers, segments, and time. But SHAP is only regulator-ready when the surrounding program is strong. Common failure modes we see:
- Using SHAP charts in slide decks without translating them into decision artifacts (reason codes, narratives, evidence tables, and monitoring thresholds).
- Treating a single global importance chart as “the explanation,” instead of combining local (individual) + global (portfolio) views.
- Ignoring correlated features and data proxies that can cause unstable attributions or mask disparate impact risk.
- Failing to connect explainability outputs to fair-lending testing (e.g., matched pairs, outcomes testing, marginal effects, second-look reviews).
Cicrim’s regulator-ready SHAP workflow
Cicrim implements SHAP as part of an end-to-end, exam-ready explainability and fair-lending framework — designed for community and regional bank realities (limited teams, multiple vendors, and heavy audit pressure).
1) Build reason codes that map cleanly to policy
We convert SHAP drivers into a governed reason-code library aligned to underwriting policy, ECOA/Reg B adverse action needs, and your existing credit playbooks — including tie-breaking rules and stability checks.
2) Segment-level fairness and consistency testing
We evaluate SHAP distributions across protected classes and operational segments, flag “driver flips,” and quantify variance in feature influence that can indicate proxy risk, drift, or brittle models.
3) Monitoring that produces audit artifacts
We define thresholds and alerts for attribution drift, feature stability, and policy breakpoints — producing monthly evidence packs your second line and auditors can reuse.
4) Examiner-ready narrative templates
We generate standardized language for model documentation and exam discussions: what the model does, why the top drivers make sense, how you validated them, and how you monitor them over time.
What you can deliver in 30–60 days
Most banks don’t need a multi-quarter science project to get to “credible explainability.” With the right scoping, you can stand up a practical SHAP layer and fair-lending evidence trail quickly:
- A governed reason-code mapping tied to model versions and underwriting policy.
- Segmented SHAP drift monitoring with thresholds and monthly evidence outputs.
- A fair-lending test plan that integrates explainability artifacts into reviews.
- A packaged “exam conversation” narrative for model use, limits, and controls.
If your bank is already using vendor credit models or internal scorecards, you can still make explainability defensible. The key is to operationalize a repeatable workflow: reason codes, monitoring, and evidence packs — not one-off plots.
Cicrim can deliver an “Explainability & Fair-Lending Evidence Pack” tailored to your model(s), policies, and regulator expectations — including SHAP configuration guidance, documentation templates, and monitoring controls you can sustain.