Mastercard created a tabular foundation model to combat payment fraud

Mastercard has presented a new type of foundation model for payment infrastructure — a large tabular model trained not on text, but on arrays of transactional data. The company expects to use it to more accurately detect fraud and verify the authenticity of operations in digital payments.

How LTM Works

Unlike LLMs, which work with unstructured data and predict the next token, LTMs are trained on tables with a large number of features. Mastercard's dataset includes billions of card transactions, and the company plans to scale the volume up to hundreds of billions of records in the future. The training used not only the payments themselves, but also related fields: merchant geography, authorization chains, confirmed fraud cases, chargebacks, and loyalty program activity. The model's task is to find behavioral relationships between features and notice anomalies that are not covered by pre-written rules.

Mastercard specifically emphasizes that personal identifiers were removed from the data before training. According to the company's design, the model should analyze not the identity of the cardholder, but the nature of behavior within the transaction flow. This reduces some of the privacy risks that typically accompany AI systems in finance, although it does remove some signals that could potentially be useful for risk assessment. The company believes that the loss of accuracy can be compensated for by the scale of the sample and the richness of context.

Nvidia and Databricks provided the technical platform for the project.

Where the Model Will Be Deployed

The first deployment zone became cybersecurity and anti-fraud. Mastercard already has several systems that track suspicious transactions, but many of them depend on manual tuning: analysts set patterns like a sharp increase in purchase frequency or transactions from different countries in a short time. LTM should complement these mechanisms and better see complex combinations of features without a rigid set of rules.

According to the company, the effect is particularly noticeable in rare and expensive purchases, which traditional models often flag as suspicious even when the transaction is legitimate.

But the company sees other scenarios for the model within payment infrastructure:

• analysis of loyalty program activity
• internal analytics on portfolios and transactions
• support for solutions in cybersecurity systems
• creation of new internal applications via API and SDK

The company does not plan to immediately replace existing tools with a single model. Rather, the current plan is to build hybrid systems where LTM works together with already proven procedures and detectors. This cautious approach is understandable: in the payments industry, mistakes are costly and regulatory requirements are high. However, a unified foundation model that can be fine-tuned for different scenarios could potentially reduce the costs of training dozens of separate models, their validation, monitoring, and maintenance.

Risks and Limitations

The approach has its weak points. If a multifunctional model becomes widely deployed and begins to make systematic errors, the consequences could affect multiple products or processes at once. Therefore, Mastercard is not yet putting LTM in the role of the sole arbiter of disputed transactions.

An additional question is the explainability of decisions: in anti-fraud and credit processes, it's not enough for a business to simply receive a risk signal; they also need to show why the system worked the way it did. Without transparency and auditability, such models are harder to defend before regulators and internal compliance.

There are also more practical questions to which the market has not yet received independent answers. The claims about effectiveness come from Mastercard itself, so they cannot be considered final proof of the advantages of LTM over conventional ML approaches. It remains unclear how the model will behave under attack, how much it will cost to operate in the long term after training, and how willing regulators will be to accept this class of systems in critical financial infrastructure. These factors, not just test quality, will determine the pace of adoption.

What This Means

Mastercard shows that the next wave of AI in finance may be built not around chatbots, but around foundation models for tabular data. If LTM truly reduces false positives and simplifies work with multiple anti-fraud scenarios, banks and payment providers will begin investing more actively in such systems instead of a set of narrow models for each task.