NEW YORK — JPMorgan Chase has developed an artificial intelligence model that can detect payment fraud, predict card-network responses and assess payment retries, potentially reducing the need to build separate systems for tasks that rely on much of the same transaction data, according to a research paper published Nov. 11.
The bank’s researchers said the Large Payments Model, or LPM, is already processing millions of transactions daily. It completes 99% of requests within 60 milliseconds and costs less than one-hundredth of a cent per transaction to score, according to the paper.

Unlike AI models trained primarily on written language, LPM learns from structured payment records containing such information as dates, postal codes, transaction categories and descriptions. Its training data includes card transactions for which the world’s largest bank serves as the merchant acquirer, along with automated clearing house payments and wire transfers. The largest version was trained on as many as 400 billion tokens, or units of training data.
What Tests Revealed
In tests on card payments, a version that evaluated individual transactions improved fraud detection performance by 62.7% compared with JPMorgan’s model designed specifically for that task, the researchers reported. A second version that incorporated up to six months of transaction history associated with the same party improved the measure by 85.5%.
Those gains were measured using average precision, which evaluates how effectively a system ranks relatively rare fraudulent transactions above legitimate ones. They do not represent an 85.5% reduction in fraud losses.
The researchers also reported an 80% improvement in predicting which of roughly 100 responses a card network would return and a 22.8% improvement in predicting a network’s response to a retried payment.
Results on Task on Which it Was New
For a task on which it had not been trained — predicting whether a retried payment would ultimately be approved — LPM, with a small additional component, scored 0.5% better than a model developed specifically for that purpose, the paper said.
The researchers also evaluated the model on ACH and wire-transfer fraud, holding constant the number of alerts that fraud investigators could review. Under that constraint, LPM identified 27.4% more fraudulent ACH payments and 63.5% more fraud measured by dollar value.
More Modest Results
Its performance on wires was more modest: It detected 5.6% more fraudulent transfers but no additional fraud by dollar value. The researchers attributed the weaker results in part to wire transfers accounting for a smaller share of the training data.
The comparisons were against JPMorgan’s own task-specific models using the bank’s payment data. The paper did not identify those comparison models, disclose their absolute performance scores or evaluate LPM against a public benchmark. It also did not quantify changes in fraud losses, legitimate payments incorrectly declined or overall cost savings.
The researchers acknowledged limitations, including the potential need for additional training when new data fields are introduced and the continued need to summarize lengthy transaction histories.
The findings were described in a JPMorgan Chase research paper published Nov. 11 and available at arxiv.org/abs/2611.15814.





