What's Happening?
A new experiment has demonstrated that foundation models, similar to those powering large language models like ChatGPT, can significantly improve fraud detection for card issuers. By training these models on sequences of actual card transactions, researchers
were able to enhance existing fraud detection systems. The model learns patterns of individual spending behavior by analyzing transaction history, including merchant, amount, channel, and time since the last transaction. This approach allows the model to identify deviations from normal spending without being explicitly told which transactions are fraudulent. When integrated into an existing fraud model, these foundation models caught between 24% and 35% more fraud at the same risk flagging rate. This improvement was observed even for client portfolios whose data was not part of the initial training set, indicating the model's ability to generalize spending patterns. The success is attributed to the nature of issuing data, which typically involves long, ordered transaction histories with limited detail per transaction, making it an ideal scenario for sequential analysis.
Why It's Important?
This development is crucial for the financial industry, particularly for card issuers, as it offers a substantial leap in combating payment fraud. With global e-commerce fraud losses projected to reach $43.6 billion by 2027, and 98% of merchants reporting fraud in the past year, more effective detection methods are strategically necessary. The ability of foundation models to identify fraud without explicit fraud labels means that new clients can benefit from enhanced detection immediately, rather than waiting for sufficient confirmed fraud data to accumulate. This reduces financial losses, protects customer trust, and supports regulatory compliance obligations. The approach also addresses the limitations of traditional feature engineering, which often flattens transaction data and loses the valuable sequential context. By leveraging the order and rhythm of transactions, these models can detect subtle anomalies that human-engineered features might miss, thereby strengthening the overall security posture of financial institutions and safeguarding consumers.
What's Next?
The research team is moving forward with productionalizing this proof of concept for live deployment. However, two key challenges need to be addressed: explainability and latency. Explainability is critical in regulated environments, as current foundation models struggle to provide clear, human-understandable reasons for their decisions, unlike traditional models where a specific feature can be identified as the cause. Latency is another concern, as running a customer's full transaction history through a foundation model in real-time for each transaction is computationally intensive. The proposed solution involves pre-computing and storing embeddings to trade some data freshness for increased speed. Further research is also being conducted to assess the effectiveness of this approach in other fraud use cases beyond card issuing, where data characteristics might differ, such as merchant or acquirer fraud detection, which typically involve richer event details but shorter transaction histories per account.
Beyond the Headlines
The successful application of foundation models in fraud detection highlights a broader shift in how artificial intelligence can be leveraged to solve complex problems in the financial sector. This innovation moves beyond traditional rule-based or simple machine learning models by tapping into the power of sequential data analysis, which mimics how large language models understand context. The ethical implications of AI-driven decision-making, particularly regarding false positives that can freeze accounts or decline payments, remain a significant consideration. While the current focus is on detection accuracy, the impact on individual customers and the need for human oversight in low-confidence cases are paramount. The challenge of explainability also underscores a fundamental tension between the power of advanced AI and the need for transparency and accountability in regulated industries. Overcoming these hurdles will not only refine fraud detection but also set precedents for the responsible deployment of sophisticated AI across various critical applications.
