Artificial Intelligence & Machine Learning,
Finance & Banking,
Industry Specific
Sequential Foundation Model Gives Banks More Context Beyond Transactions
In the realm of banking, distinguishing between a customer navigating a brief financial hiccup and a consistently untrustworthy consumer has often posed a challenge. However, financial data company Plaid has taken innovative strides to address this issue by developing a foundation model designed specifically for this purpose. This new tool aims to aid banks and lenders in recognizing the nuanced differences in customer financial behavior, even when their transaction histories appear, at face value, identical.
Plaid dedicated approximately one year to the creation of this foundation model, which meticulously analyzes the flow of money. The model focuses on not only how funds move but critically, the sequence of those movements. This sequential foundation model is anticipated to be introduced by the company toward the end of the third quarter or in the fourth quarter of the year. According to Suddu Seshadri, head of data and artificial intelligence solutions at Plaid, traditional transaction figures provide an incomplete picture of consumer spending and management.
“Financial knowledge is very unique because it’s not externally available, and you can’t necessarily pattern match on text,” Seshadri remarked during an interview with ISMG. His assertion underscores the importance of designing a model tailored specifically to understand complex financial behaviors from the ground up, rather than relying on generalized data.
The essence of Plaid’s model lies in its ability to interpret each transaction as more than just an isolated data point. Rather, it analyzes transactions through a prism of three critical layers of information: the meaning of the transaction, the relationships in timing and order of events, and the account attributes. This multifaceted approach enables the model to grasp transactions within the wider context of individual customers’ financial patterns.
The complexity of raw transaction data can often render it ambiguous. For instance, a single deposit could represent various sources of income such as a salary, severance package, or reimbursement. Misinterpretation of such data can lead to skewed perspectives on a consumer’s financial behavior. Recognizing this, Plaid undertook the challenge to create an extensive dataset that encompasses not only transaction figures but also the context in which those transactions occur.
In illustrating the model’s capabilities, Plaid offered an example involving two consumers who seemingly reflect identical financial profiles. Both individuals share the same monthly income, average balance, rent payment, overdraft fees, and spending categories. However, upon closer examination, their financial behaviors reveal stark contrasts.
For Consumer A, funds from their paycheck are promptly allocated to essential expenses such as rent and utilities. Although an unexpected repair incurs a single overdraft fee, their account recovers swiftly prior to the next paycheck, returning to a standard financial pattern. Conversely, Consumer B adopts a different approach. Upon receiving their paycheck, they immediately direct funds toward repaying loans and credit cards, which depletes their account nearly within a day. This consumer scrambles to make small transfers before the next paycheck, perpetuating a cycle of financial stress.
At first glance, based on mere transaction movement, both consumers may seem indistinguishable in their financial habits. However, Seshadri emphasizes that the sequence and timing of events reveal critical insights into these consumers’ behaviors, clearly illustrating that Consumer A does not habitually overdraw their account.
Plaid’s unique access to banking data enables it to gather this crucial information. The approach taken to train their model involved contrastive learning, a method that enhances the model’s predictive capabilities regarding financial activity while also safeguarding the integrity of the information being analyzed.
Plaid has put significant emphasis on ensuring robust security measures throughout this modeling process. Seshadri confirmed that the same strict security protocols used throughout the company’s infrastructure were applied when handling sensitive financial information. This diligence extends to not exposing any model intervals or thresholds to mitigate risks associated with potential cyber threats. Furthermore, Plaid obtained necessary approvals from its privacy and legal teams prior to broadening access to the model.
The protection of individual financial information remains a priority for Plaid. Seshadri stated that the company always seeks explicit consent from customers before accessing their data and rigorously adheres to existing privacy regulations and agreements with banks and applications.
As Plaid progresses with refining its model, Seshadri highlights the intention to fine-tune its applications further, broadening its usability across multiple scenarios, including instances of first-party fraud, identity theft, cash advances, and payment processes. By continuing to enhance its model, Plaid aims to empower financial institutions with deeper, contextual insights that could ultimately transform how they interact with their consumers.