Disparity detection with defendable segmentation

Design meaningful analytical segments and document their limitations before interpreting differences in lending outcomes.

Choose segments for a reason

A useful segment corresponds to a product, decision process, channel, period, or other business distinction that the institution can explain. Define the question before inspecting the results so segmentation does not become a search for a preferred answer. Record the inclusion rules, exclusions, and treatment of incomplete observations. Keep a stable reference population so a reviewer can understand how a subset relates to the whole portfolio.

Reconcile coverage and missingness

Compare source counts with the records available for analysis. Identify fields with missing, inconsistent, stale, or ambiguous values and examine whether those gaps are concentrated in particular channels or periods. Preserve the exclusion counts alongside the reported outcomes. When a segment is too sparse or unreliable to support interpretation, flag that limitation and consider whether a different scope or additional evidence is needed.

Separate signals from explanations

An observed difference is a prompt for investigation. Review the decision context, policy changes, product mix, operating practices, and analytical assumptions before reaching a conclusion about its cause. Use appropriate statistical and domain expertise for the chosen method. Document sensitivity checks and unresolved questions rather than converting a complex result into a single green or red indicator with no supporting context.

Make the method reviewable

Maintain a versioned segment dictionary, the transformation and calculation logic, the selected observation period, and the outputs used in review. Assign an owner for methodological changes and document how changes affect comparisons with prior periods. Cicrim helps lending, data, compliance, and risk teams connect this analytical discipline to their review process and to a practical record of follow-up decisions.

Working-session checklist

• Question and segment rationale
• Population reconciliation
• Missingness and coverage review
• Method assumptions and sensitivity checks
• Documented follow-up decisions