AI-DRIVEN FRAUD DETECTION AND RISK FORECASTING FRAMEWORK FOR REAL-TIME FINANCIAL TRANSACTIONS

This paper aims to critically analyze AI-based fraud detection and risk prediction models in real-time financial operations to fill in the gaps in scalable, interpretable, and defensible framework in the context of increasing cyber threats. This study applies the methodology of secondary research by searching Scopus and Google Scholar with the keywords such as; AI fraud detection real-time, gathering 8 peer-reviewed sources in the year 2018-2025, and applied thematic analysis to the results through familiarization, coding (e.g., LSTM precision), theme clustering (ML models, analytics, forecasting, hybrids), and synthesis. Findings show dominance of machine learning: Random Forests and SVMs at 95 percent accuracy, Kafka-Flink streams at 89 percent card frauds in less than 50ms, Transformers at 5.2 percent MAPE error and neuro-symbolic hybrids at 98.5 percent F1. Such weaknesses as 20% accuracy drop due to PGD attacks and concept drift are revealed in critical discussion. The main advantages of triangulation were provided by secondary benefits: cost-efficiency, speed in synthesis. The results drive financial resilience, calling on adversarial training and XAI developments.