Fintech's new lending, advisory and crypto platforms create risks traditional banking models can't measure, so AI and statistical learning must be built into financial risk management.
Machine learning can spot fraudulent financial transactions, but detection rates vary widely by method and by which data features are used.
Resampling credit card fraud data before the train/test split inflates accuracy above 99%, a leak that collapses to 1.81% precision when done correctly; deep reinforcement learning fared worse at 34.8% accuracy.
Reviewing only fraud-detection studies that keep real-world class imbalance, researchers find overused benchmark datasets, tree-based models dominating, misleading metrics, and almost no usable explainability.
Machine learning ensembles and neural networks beat traditional rule-based credit scoring and fraud detection on precision, recall and AUC, while staying explainable.
Machine-learning ensembles and neural networks beat traditional rule-based credit and fraud models on precision, recall and AUC, while staying explainable enough for real use.
A learning-as-it-goes reinforcement learning agent starts out worse than statistical anomaly detection at catching card fraud but keeps improving as fraud tactics change.
Storing bank transactions in a graph database and feeding those relationship features to machine learning models improves detection of fraudulent transactions.