A lender may produce an accurate credit score yet still be unable to answer a borrower’s question: “Why was my application rejected?” Interpretability is the ability to connect a prediction or decision to reasons people can understand. In finance, those reasons must serve borrowers, frontline staff, risk auditors, and compliance teams—not just model developers. section 6.3 section 1.1
An interpretable credit workflow keeps the path from data to decision inspectable:
- Handle missing values, sentinel codes, and outliers consistently and document the choices. Unexplained data quirks can undermine both score stability and auditability. section 4.4.4
- Use binning and weight-of-evidence (WoE) transformations when a transparent scorecard is needed. Link each bin’s value to its effect on the score and to a decision reason. section 4.4.4 section 6.6
- For less transparent models, incorporate explainability methods such as LIME or SHAP rather than assuming predictive performance makes a decision understandable. section 8.1.3 section 6.6
- Review explanations alongside performance, fairness, and data-drift checks. Risk, compliance, and business leadership must assess explainability before deployment; monitoring remains necessary afterward. section 5.3.9 section 6.6
From a credit feature to a reason
Suppose an applicant falls into a “Number of Delinquencies: 3–5” bin. Its negative WoE value can be traced to a lower credit score, giving an auditor a defensible link among the input, transformation, and outcome. That link is useful when a lender must provide specific reasons for an adverse credit decision. section 4.4.4
Interpretability has different operational demands across models. Fraud analysts need to understand why an alert was raised; a graph neural network may require graph-specific explanation approaches such as GNNExplainer or attention visualization. LLM-based interfaces can make deterministic explanations easier to read, but should translate model outputs rather than invent reasons. section 8.1.3 section 9.4.1 section 6.6
