Mastercard has announced the development of a large tabular model (LTM) that is trained on transaction data rather than text or images. The model is intended to improve the company’s ability to detect fraud and verify the authenticity of digital payments. The announcement was made in a recent blog post by Mastercard’s data science team.
Training and Data Sources
The foundation model has been trained on billions of card transactions, with plans to expand the dataset to hundreds of billions in the future. The data set includes payment events, merchant location, authorization flows, fraud incidents, chargebacks, and loyalty activity. Mastercard stated that personal identifiers were removed before training began, and that the model focuses on behavioural patterns rather than individual identities. By excluding personal data, the technology reduces privacy risks that may affect other forms of AI in the financial services sector. The scale and richness of the data allow the model to infer commercially valuable patterns, even without per‑user information. Mastercard acknowledges that anonymisation removes some signals that could be useful for risk assessment, but argues that the large volume of behavioural data compensates for any loss of detail.
What Is a Large Tabular Model?
An LTM differs from large language models (LLMs), which are trained on unstructured inputs and predict the next token in a sequence. Instead, the LTM examines relationships between fields in multi‑dimensional data tables, making it closer to pure machine learning than to artificial intelligence. The model learns from raw inputs which relationships are predictable, enabling it to identify anomalous patterns that predefined rules may miss. Mastercard describes the LTM as an “insights engine” that can be integrated into existing products and workflows. The operational risk of a model that interacts with customers, such as an LLM, differs from that of a model used for internal decision‑making. The technical infrastructure for the LTM comes from Nvidia, which provides the computing platform, and Databricks, which handles data engineering and model development.
Deployment in Cybersecurity
Cybersecurity is the first area where the LTM will be deployed. Mastercard operates several fraud detection systems that analyze transaction data and require human input to define suspicious behaviour. Examples of suspicious patterns include sudden increases in transaction frequency or purchases made in different parts of the world within a short time span. Early results indicate that the new model improves performance over conventional techniques in specific cases. For instance, high‑value, low‑frequency purchases that are often flagged as anomalies by traditional models can be distinguished more accurately by the LTM. Mastercard plans to deploy hybrid systems that combine established procedures with the new model, reflecting the regulatory environment in which the company operates. The company acknowledges that no single model is likely to perform well in all scenarios, so the LTM will complement existing tools.
Broader Applications and Cost Efficiency
The model is also claimed to be useful for scanning activity on loyalty programmes, portfolio management, and internal analytics, where large volumes of structured data are available. Currently, many organisations deploy multiple models tailored to each task, which increases training costs and the effort required for validation and monitoring. A single foundation model that can be fine‑tuned for different tasks may simplify processes and reduce costs.
Risks and Future Plans
A multi‑function LTM carries the risk that a failure in a widely deployed model could have system‑wide consequences. This risk partly explains Mastercard’s strategy of applying the technology alongside existing detection systems for the time being. The company intends to increase the scale of the data used and the overall sophistication of the model. Plans also include providing API access and software development kits to allow internal teams to build new applications. Mastercard emphasises data responsibilities such as privacy, transparency, model explainability, and auditability. Regulatory scrutiny of any system that influences credit decisions or fraud outcomes is expected, in addition to scrutiny of the data practices involved in the LTM’s operation.
Implications for the Payments Industry
Large tabular models rely on highly structured data rather than text or images, and may represent a new generation of AI systems in core banking and payments infrastructure. Evidence to date is limited to vendor reports, so performance claims should not be regarded as conclusive. Robustness under adversarial conditions, long‑term post‑training costs, and regulatory acceptance will determine the pace and extent of adoption. Mastercard’s focus on the table‑based approach signals a strategic bet on this technology for the near future.
Looking Ahead
Mastercard is expected to continue expanding the dataset and refining the model’s capabilities. The company plans to integrate the LTM into hybrid fraud detection systems and to offer internal teams tools for building new applications. Regulatory reviews and real‑world performance data will shape the model’s eventual deployment across Mastercard’s global payment network.
