Explainable Financial Anomaly Detection Through Multi-Source Risk Feature Fusion

Abstract

Financial anomaly detection is a critical task for identifying fraud, earnings manipulation, abnormal credit behavior, and suspicious transaction patterns. This article proposes a multi-source risk feature fusion framework for explainable financial anomaly detection. The framework combines accounting ratios, transaction-level signals, credit exposure variables, and external market indicators into a unified feature space. Machine learning classifiers are trained to detect abnormal entities or events, while post-hoc explanation methods are used to translate model predictions into understandable risk evidence. The article argues that explainability is not only a technical requirement but also a governance necessity in financial supervision and auditing. By identifying the most influential variables behind anomaly scores, the framework assists analysts in distinguishing between operational irregularities, temporary volatility, and deliberate fraud. The proposed approach also supports model validation by comparing algorithmic findings with traditional financial risk indicators. Overall, the study contributes to the development of transparent and auditable AI systems for financial risk monitoring.