Reason Codes

A game-theoretic framework that assigns each feature an importance value for a particular prediction. SHAP values are the mathematical backbone of many reason code generators, unifying several attribution methods to explain any model's output.

• Key property: Additivity ensures the sum of all feature SHAP values equals the difference between the prediction and the average prediction.
• Use case: Explaining why a specific transaction received a high fraud score by quantifying the contribution of features like transaction amount, device age, or IP geolocation.

A technique that explains any classifier's prediction by approximating it locally with an interpretable surrogate model, such as a sparse linear model, around the specific instance being explained.

• Mechanism: Perturbs the input instance, observes the black-box model's predictions, and fits a simple model weighted by proximity to the original instance.
• Output: The coefficients of the local linear model serve directly as reason codes, indicating which features pushed the prediction toward fraud or legitimate.

A method that identifies the minimal set of feature changes required to flip a model's decision to a desired outcome. For fraud detection, this answers: 'What would need to change for this transaction to be approved?'

• Example: 'If the transaction amount were $200 instead of $2,000 and the shipping address matched the billing address, this transaction would not be flagged.'
• Advantage: Provides actionable, human-readable guidance that aligns naturally with adverse action notice requirements.

Specific reason codes mandated by regulations like the Fair Credit Reporting Act (FCRA) and Equal Credit Opportunity Act (ECOA) that must be disclosed to consumers when a negative decision is made based on a model's output.

• Regulatory requirement: Lenders must provide the principal factors that drove the denial, not just a generic score.
• Mapping challenge: Translating technical feature attributions (e.g., 'avg_txn_7d_zscore') into compliant, consumer-facing language (e.g., 'Recent transaction activity exceeds typical patterns').

A property of models that are inherently understandable due to their structure, providing direct reason codes without post-hoc analysis. Glass-box models like the Explainable Boosting Machine (EBM) learn modular feature functions that are independently inspectable.

• EBM advantage: Each feature's contribution is a learned shape function that can be plotted and explained directly, eliminating the need for SHAP or LIME approximations.
• Trade-off: May sacrifice marginal predictive performance compared to deep neural networks, though often negligible in tabular fraud domains.

A comprehensive, chronological record of the data, model parameters, feature values, and decision logic used for a specific transaction. Reason codes are a critical component of this trail, providing the human-readable justification layer.

• Components: Input feature snapshot, model version, SHAP or LIME output, generated reason codes, and final decision timestamp.
• Purpose: Enables full traceability for regulatory review, internal model risk management, and customer dispute resolution.