Risk-Aware AI Models for Financial Fraud Detection: Scalable Inference from Big Transactional Data

Purpose: The purpose of this study is to develop a scalable, risk-aware artificial intelligence (AI) framework capable of detecting financial fraud in high-throughput digital transaction environments. The research addresses the limitations of traditional rule-based and single-model systems, which often suffer from high false positives, poor adaptability, and limited interpretability. By combining supervised machine learning, unsupervised anomaly detection, and engineered domain-specific risk features, the proposed hybrid architecture aims to enhance precision, recall, and transparency. The framework is designed to meet stringent regulatory and operational requirements, making it deployable across banks, fintechs, digital wallets, and other institutions handling large volumes of transactions. Design/Methodology/Approach: The framework integrates a modular, five-layer architecture: 1) real-time multi-source data ingestion, 2) domain-driven risk feature engineering, 3) hybrid AI inference combining XGBoost and autoencoder models, 4) interpretability via SHAP and counterfactual reasoning, and 5) seamless decisioning and integration with downstream fraud operations. Distributed cloud-native infrastructure using Kafka, Spark, Kubernetes, and ONNX Runtime ensures scalability and low latency. Training leverages imbalanced financial datasets using advanced sampling techniques and time-series aware cross-validation. Continuous monitoring for data drift, fairness, and explainability supports compliance. Experimental validation was conducted using over 100 million anonymized transactions from multinational banking partners under production-like conditions. Findings: The hybrid framework achieved significant improvements over legacy systems and single-model baselines. Precision, recall, and F1-score reached 0.94, 0.93, and 0.935, respectively, outperforming rule-based approaches by over 35%. Median inference latency was reduced to 126 ms under 10,000 transactions per second, supporting real-time operations. SHAP-based interpretability and counterfactual reasoning provided transaction-level explanations, meeting regulatory transparency requirements. Top-1% alerting achieved a 92% fraud hit rate, reducing analyst workload. Data drift monitoring maintained model stability for over 90 days post-deployment. These findings demonstrate that hybrid AI designs can balance speed, accuracy, scalability, and auditability for modern fraud detection. Research