A Bank Trained Its Own Foundation Model to Catch Fraud — Freddy Lecue, Wells Fargo
Wells Fargo built a transformer-based foundation model on tabular transaction data that outperforms XGBoost for fraud detection.
“the model I'm going to show you outperforms XGBoost even in terms of latency, and in fraud detection we get a much better false positive reduction and a higher detection rate”
Wells Fargo trained a proprietary transformer-based foundation model on billions of customer transactions — debit, credit, checks — to serve as a single backbone for fraud detection, anti-money laundering, credit scoring, and marketing. The model outperforms XGBoost (their previous production standard) on both latency and fraud detection metrics. This signals a broader enterprise trend of building domain-specific foundation models on structured/tabular data rather than relying on general-purpose LLMs or classical ML.