Wells Fargo says its 1B-parameter model beats XGBoost on fraud detection

Wells Fargo's head of AI, Freddy Lecue, said in a talk at AI Engineer that the bank pre-trained a decoder-only transformer of about 1 billion parameters on billions of monthly transactions to detect card fraud, according to a StartupHub.ai report published Oct. 23.1

Lecue said the bank built the foundation model from scratch to replace one-off classifiers, with no weights borrowed from open models.1 He said the model reads a customer's stream of debit, credit and digital check transactions and predicts whether the next transaction is fraudulent given the history, which XGBoost struggles to encode without brittle manual features.1

Lecue said the model beats the production XGBoost baseline by raising the fraud detection rate and cutting false positives, and runs well under the 250 millisecond decline window, roughly 10 times faster than the budget requires; the report gave no figures for the fraud or false-positive improvement.1

The team cut thousands of hand-engineered features to about 100 raw features across five dimensions (who, what, how, when and where), Lecue said.1

Lecue said fewer false declines mean fewer call-center contacts and fewer lost sales, which he quantified as millions in savings versus XGBoost, per StartupHub.ai, without publishing precision-recall curves or a time-split evaluation window.1

Lecue said the same backbone will be tried for credit line increase, anti-money laundering and marketing, and an 11 billion parameter version is in work; StartupHub.ai said the talk showed no holdouts by geography, merchant category or time period, and no drift monitoring or adversarial robustness testing.1