Integrating Model Explainability and Uncertainty Quantification for Trustworthy Fraud Detection

Fraud and money laundering continue to threaten financial stability around the world, thereby creating a need for detection systems that are both trustworthy, accurate, and reliable. Regulators increasingly require fraud detection systems to be transparent, accountable, and reliable. To address these challenges, this paper presents an Integrated Transparency and Confidence Framework (ITCF) for automated fraud detection. The framework combines local explainability using Local Interpretable Model-Agnostic Explanations (LIME) with calibrated uncertainty quantification using split Conformal Prediction (CP), aiming to improve both the understanding and the confidence in artificial intelligence-driven decisions. The framework was evaluated using the PaySim mobile-money dataset, which contains 6, 362, 620 transactions. Random Forest (RF) and XGBoost (XGB) models were benchmarked for fraud detection. LIME provided clear, case-specific explanations and aggregated feature importance, helping financial crime analysts understand why individual transactions were flagged. Conformal prediction measures model uncertainty and offers distribution-free coverage guarantees under independent and identically distributed (IID) assumptions. When targeting 90% coverage (α = 0.1), both models achieved empirical coverage close to 0.90. However, XGBoost outperformed Random Forest by achieving a higher F1-score and recall while also producing smaller prediction sets, making it more effective for risk-sensitive decision-making. Overall, the ITCF supports risk-based workflows by triaging high-uncertainty transactions for human review, reducing alert fatigue, and improving auditability. By integrating explainability and uncertainty estimation, the framework enhances transparency, strengthens operational trust, and supports the responsible, regulation-aligned use of AI in financial fraud and money laundering detection.