AI Applications Using Financial Data: Models, Performance, and Systemic Implications

Abstract

This paper examines the application of artificial intelligence (AI) to financial data, integrating a structured literature review with an original experimental study of transaction-fraud detection. The review covers fraud detection and anti-money laundering, credit scoring and risk management, financial time-series forecasting, algorithmic trading, and financial sentiment analysis using domain-specific large language models (LLMs). Industry evidence reported in the supplied source indicates rapid production adoption of AI in financial services, including 88% AI/ML production use and 89% generative-AI adoption among the 56 institutions surveyed by IIF-EY. The experimental component uses a synthetic dataset designed to replicate the structure and extreme class-imbalance characteristics of the Kaggle ULB credit-card fraud benchmark, with 8,040 transactions, 40 fraud cases, 28 anonymized principal-component features (V1–V28), and Amount. Three classifiers—Logistic Regression, Random Forest, and Gradient Boosting—are evaluated using ROC-AUC, PR-AUC, precision, recall, and F1-score under a stratified 70:30 train-test split. Gradient Boosting records ROC-AUC 0.9999, PR-AUC 0.9860, precision 0.9091, recall 0.8333, and F1-score 0.8696. The source experiment also reports V14, V10, and V4 as dominant predictors, consistent with cited explainable-XGBoost literature. Because the experimental dataset is synthetic and intentionally structured to reproduce strong signal, the results are interpreted as a methodological demonstration rather than a universal benchmark of real-world fraud performance. The paper further discusses explainability, bias, data quality, hallucination, confidentiality, third-party concentration, model risk, and systemic implications under emerging financial-AI governance regimes.