Explainable Credit Underwriting on FICO HELOC: A Framework for Counterfactual Recourse under Profit, Fairness, and Stability Constraints

Abstract

Credit underwriting models are increasingly accurate yet often opaque, creating tension between portfolio performance, regulatory expectations, and actionable customer communication. This paper presents a reproducible experimental study on the FICO HELOC dataset (10,459 applicants; 23 credit-bureau features) that integrates (i) predictive modeling, (ii) executable bad-rate-constrained threshold policies, (iii) SHAP-based explanations, and (iv) counterfactual recourse with feasibility constraints. We train logistic regression, random forests, and gradient boosting using a consistent preprocessing pipeline that treats the dataset’s special missing codes (-7, -8, -9) as missing values and applies median imputation. We then derive underwriting policies by selecting the minimum approval threshold that satisfies a target bad-debt rate among approved accounts, and evaluate “approval rate and profit at equal bad rate.” To assess responsible deployment aspects, we compute group fairness gaps using demographic parity (DP) and equal opportunity (EO) across operational score segments, and quantify explanation stability using bootstrap resampling. Empirically, at a 20% bad-rate constraint, random forests achieve the highest approval rate (30.9%) and the highest mean profit under a simple unit-profit model (+1 for Good, -3 for Bad). Gradient boosting yields coherent global explanations dominated by ExternalRiskEstimate, credit file age, and revolving utilization, while counterfactual recourse achieves 100% success for logistic regression and 89.5% for gradient boosting under actionable feature constraints. Bootstrap analysis shows strong SHAP stability (mean Spearman ρ=0.871 for global importance; mean cosine similarity=0.856 for local vectors). The results demonstrate how policy design, explainability, and stability can be evaluated jointly on a realistic credit dataset, yielding decision rules that are directly implementable as underwriting policy.