How Mathematical Optimization Helps to Improve Customer Experience and Fraud Defense for Consumer Banking

Author’s note: For confidentiality, we intentionally modify and omit some details in this article so that it is general enough and hence cannot be associated with any specific business entity. However, it still contains sufficient information for illustration purposes.

Although mathematical optimization has been used in portfolio optimization for decades, we have not seen a great deal of new use cases in finance that leverage the power of optimization. Most practitioners today focus on the use of machine learning in banking and thus might overlook the power of the optimization modeling technique. In 2022, my team and I were tasked by one of the largest and leading technology-driven consumer banks in the U.S. to improve some of its critical business policies, which impact the banking experience of its millions of customers. By leveraging an integer programming-based solution, we were able to not only significantly improve its customer experience but also automate the previously manual and labor-consuming business procedures. This article aims to present this use case in detail to bring attention to the emerging use of mathematical optimization in the consumer banking space and stimulate research interest in this domain. In our opinion, this project could fit into current research fields of “predict-then-optimize” and/or “segment-then-optimize.”

With the recent broad use of mobile apps, bank customers can easily deposit a check at their fingertips. What is unknown to most customers is that once the deposit is made, banks need to decide how quickly deposits are available for further use such as external transfer or withdrawal, which is typically based on a funds availability (FA) policy. On one hand, if the FA policy allows for the deposit to be immediately available, banks would be subject to enormous fraudulent loss should the checks turn out to be fake or contain insufficient balances in the counterparty accounts (i.e., accounts from which the checks are issued). On the other hand, customers may experience too much of a delay if the FA policy holds deposits until banks are able to verify the authenticity of checks, which often takes up to four days. Therefore, the question is: How should banks set up the FA policy to balance the tradeoff between customer experience and fraud loss?

Although the bank we worked with (referred to as “Bank A” hereafter) already had a machine learning model in production to score the risk of every deposit transaction, its business team still needed to make decisions on daily availability amount for each check deposit over a period of four days. For instance, if a customer deposits $1,000 cash, should it make $500 available by Day 0, $250 by Day 1 and the rest available by Day 2, or should it make only $250 available by Day 0, $500 by Day 1 and the rest available by Day 2? Building on business experiences and professional judgment, Bank A set up a policy consisting of two major layers: 1) segmentation and 2) decisioning. The segmentation layer comprises variables and corresponding cutoff values to segment deposit transactions. Within each segment, the decisioning layer sets the maximal fund availability amount for each of the days after a check is deposited. Table 1 shows an example of a funds availability policy.

Table 1: Example of a Funds Availability (FA) Policy

To improve the performance of the existing policy, Bank A needed my team to optimize two sets of decisions within the policy. First, with a new machine learning model being put into production, the original score cutoffs are obsolete (see column “Machine learning risk score” in Table 1). Second, Bank A needed us to optimize the daily maximum availability amount for each segment (see column “Dollar limit by days”). Given that the model score cutoffs impact the segmentation, which then impacts the optimal solution of the daily maximum availability amount, we adopted a two-stage iterative approach. Within each iteration, we first used the simulated-annealing method to propose a new candidate cutoff value. Once the new cutoff value is proposed, we then formulated an integer programming model to obtain the optimal solutions for the daily availability amount across all segments. If the new cutoff value and optimized daily availability amount does not lead to a better solution in the current iteration, we reject the candidate cutoff value and move to the next iteration. In other words, the second-stage integer programming model obtains an optimal value, which is used to evaluate whether the first-stage simulated annealing method accepts or rejects the proposed cutoff value.

In production, our solution improved the FA policy performance substantially: 15% more deposit transactions become fully available by Day 0, and the fraud loss does not increase as a result of the faster funds availability. More importantly, the optimization-based solution significantly reduces the time needed to update a policy – whereas a business analyst previously needed approximately one month to manually update a new policy, it now only takes hours for the optimization-based framework to generate a new one. This reduction enables Bank A to be more agile in response to any changes in customer expectation or fraudulent activity patterns.

Although we only worked on the FA policy optimization related to check deposit transactions, many similar challenges exist in the consumer banking or e-commerce sectors. For instance, if a credit card user makes a payment, how soon should banks make the corresponding credit line available? Interested readers could refer to [1] for a similar use case in e-commerce. With machine learning and generative AI seemingly dominating the buzzword in the media, we hope this article presents emerging opportunities in which operations research and mathematical optimization contribute substantial value to the business world.

Reference

1. Pulkkinen, N. Tiwari, A. Kumar and C. Jones, 2018, “A Multi-Objective Rule Optimizer with an Application to Risk Management,” 17th IEEE International Conference on Machine Learning and Applications.