Trustworthy Synthetic Financial Data Generation using CTGAN with Explainable AI: A Novel Trust Score Framework for Credit Ris5k Assessment

This paper presents a comprehensive framework for generating trustworthy synthetic financial data by combining Conditional Tabular GAN (CTGAN) with Explainable AI (XAI) techniques. We propose a novel multi-dimensional trust score integrating five key metrics: downstream utility (WU = 0.9776), SHAP feature importance alignment (WI = 0.8810), permutation importance (WP = 0.9241), LIME local explanation fidelity (WL = 0.8646), and correlation-structure fidelity (WC = 0.7572), achieving a composite score of 0.8954. The framework validates synthetic data quality across statistical fidelity, predictive utility, and explainability dimensions. Comparative analysis demonstrates CTGAN superiority over VAE with 35.4% higher feature interaction preservation. The framework enables practical deployment in regulated financial domains while maintaining privacy, interpretability, and regulatory compliance. Results show 97.76% downstream utility preservation and 1.46% AUC improvement when combining real and synthetic data for class imbalance mitigation.