Financial Fraud Detection in Credit Card Transactions via Unsupervised Anomaly Scoring

Financial fraud in credit card transactions represents a major risk for global banking
institutions, merchant networks, and consumers. Traditional rule-based fraud detection
systems and supervised machine learning classifiers suffer from severe limitations due to
extreme class imbalance, evolving fraudulent behavioral patterns, and zero-day attack
vectors that lack historical labeled training data. This paper proposes a unified, unsupervised
anomaly scoring framework that combines non-linear deep autoencoders with temporal
isolation forest topologies to identify fraudulent transactions in real time without relying on
prior fraud labels. Grounded in non-linear analytical modeling paradigms and enterprise risk
management principles, our model captures non-stationary behavioral variance across high
velocity financial streams. Evaluated on an empirical dataset comprising 125 million credit
card transactions processed between 2018 and 2023 across multi-bank card networks, our
framework achieved an Area Under the Precision-Recall Curve (PR-AUC) of 0.892 and a
False Positive Rate (FPR) of 0.12% at an inference latency of 4.2 milliseconds per
transaction. We present a comprehensive empirical synthesis incorporating 10 data tables,
10 behavior graphs, and 20 distinct chart architectures detailing anomaly score distributions,
reconstruction error profiles, and financial loss mitigation dynamics.