Causal Modeling for Fraud Detection: Enhancing Financial Security with Interpretable AI

Financial fraud poses a significant threat to the stability of modern economic systems. However, traditional machine learning approaches to fraud detection-primarily correlation-based-remain limited in precision, interpretability, and adaptability when confronting the constantly evolving strategies of fraudsters. This study introduces a causal inference framework for fraud detection, leveraging recent advancements in causal analysis to identify and quantify the underlying causal relationships among transaction attributes, user behaviors, and fraudulent outcomes. The framework incorporates three key components: causal discovery algorithms (PC and FCI), robust effect estimation techniques (e.g., PSM and DML), and an interpretable rule-extraction module that translates causal patterns into actionable insights. Experiments were conducted on two real-world datasets: a credit card transaction dataset (284,807 records, 32% fraud rate) and an insurance claims dataset (350,000 cases, 8% fraud rate). Results show that the proposed model consistently outperforms leading correlation-based methods-including AdaBoost, GBDT, XGBoost, and LightGBM-achieving notable performance improvements: an average 9-percentage-point gain in overall accuracy, a 2% increase in F1 score (up to 11%), a 5% boost in AUPRC, and a 13.3% improvement in MCC. A key finding highlights a 47% higher fraud risk associated with atypical location changes combined with large-value transactions, directly addressing the "black-box" limitations of conventional models. Robustness analyses further confirm the model's resilience against confounding influences such as seasonal fluctuations and demographic shifts, underscoring its adaptability to emerging fraud patterns. By integrating causal inference with interpretable artificial intelligence, this research advances fraud detection toward more precise, transparent, and regulatory-compliant financial risk management.