Finite-Difference Approximation for Explaining Neural Network Predictions in Financial Credit Evaluation

Financial credit evaluation is an essential process that enables lenders to assess the creditworthiness of applicants. While deep learning (DL) models have shown high predictive accuracy in financial risk assessment, their lack of interpretability limits trust and regulatory compliance. Explainable AI (XAI) methods such as SHAP and LIME offer feature attribution, but they suffer from instability and reliance on external data. To address these challenges, this paper introduces the One-sided Linear Gradient Approximation (OLGA), a deterministic and computationally efficient explainability method based on finite-difference approximations and direct perturbation. OLGA provides direct attributions without requiring background data or extensive sampling, making it scalable for high-dimensional financial datasets. We compare OLGA against SHAP and LIME using key explanation metrics, including infidelity, sparsity, and sensitivity, across four deep learning architectures: Multi-Layer Perceptron, Convolutional Neural Network, Transformer, and Autoencoder. Experimental results indicate that OLGA achieves competitive fidelity and provides computational efficiency and stability compared to LIME and SHAP.