Abstract
In the financial sector, providing clear and well-supported reasons for credit denials is a regulatory mandate under the Equal Credit Opportunity Act and the Fair Credit Reporting Act. This paper proposes a robust framework using Baseline Shapley (B-Shap) to generate these "Adverse Action" (AA) explanations. By decomposing the model output difference between a declined applicant and a successful reference point, the method provides mathematically sound feature attributions for complex, non-linear models. The authors demonstrate that enforcing monotonicity and continuity in models like Neural Networks enhances both interpretability and predictive performance.
Introduction
When a financial institution denies a credit application, it must provide the applicant with specific reasons for the decision. While simple linear models allow for straightforward explanations via coefficients, modern machine learning models—such as Neural Networks or Gradient Boosting—involve complex feature interactions that obscure direct attribution. Existing local interpretability tools like LIME often fail in this context because they rely on local linear approximations that may not hold for declined applicants who are "far" from the acceptance threshold. This paper addresses the challenge of attributing the gap between a denial and an acceptance to specific features in a way that is both mathematically rigorous and computationally feasible.
Methodology
Reference Point Comparison
Rather than analyzing a declined applicant ($x^D$) in isolation, the authors compare them to a reference point ($x^A$) located within the "Accept" region of the model's feature space. The objective is to decompose the difference in the predicted log-odds of default, $f(x^D) - f(x^A)$, into contributions from individual features.
Baseline Shapley (B-Shap)
The framework adopts Baseline Shapley (B-Shap) as the primary attribution tool. Unlike standard SHAP, which integrates over a data distribution and is computationally expensive, B-Shap compares the function value at a specific point against a fixed baseline. This makes the process more tractable by requiring only simple function evaluations. For models with multi-way interactions, the authors provide explicit formulas to distribute interaction effects fairly among participating variables.
Model Constraints and Feature Engineering
The authors argue that for credit models to be truly explainable, they must adhere to two fundamental properties:
- Monotonicity: A change in a feature should have a consistent directional effect on risk (e.g., increasing debt should never decrease the probability of default).
- Continuity: The model should avoid "jumps" in explanations, which are common in piecewise constant models like basic decision trees.
Additionally, to handle highly correlated predictors (e.g., multiple measures of delinquency), the authors suggest "collapsing" them into single groups. This prevents redundant or confusing reason codes from being presented to the applicant.
Experimental Results
The methodology was tested on a simulated dataset of 50,000 customers with 10 credit-related predictors. Key findings include:
- Performance Gains: A Monotone Neural Network achieved a Test AUC of 0.797, outperforming a standard unconstrained Feedforward NN (0.787). This suggests that domain-knowledge constraints can serve as effective regularizers.
- Explanation Accuracy: In a case study of a declined applicant, B-Shap identified "Credit Age" (59.5%) and "Number of Inquiries" (31.2%) as the primary drivers of the denial.
- Clarity through Collapsing: Grouping correlated delinquency variables allowed a single "Number of Accounts Past Due" category to account for 75.7% of an explanation, providing a much clearer reason than if the attribution had been split across multiple similar variables.
Conclusion and Limitations
The B-Shap framework provides a computationally efficient, model-agnostic way to generate regulatory-compliant explanations. However, some challenges remain. The choice of the reference point ($x^A$) is somewhat subjective and can significantly influence the resulting explanation. Furthermore, while B-Shap is more efficient than standard SHAP, its complexity still grows exponentially with the number of features, potentially requiring further approximations for models with more than 25 variables.
Summary updated