This project aims to model the spending behavior of bank customers, based on the type, amount, and inter-arrival of their credit/debit card transactions.  Specifically, we plan to build probabilistic generative models that take as input the timeline of transactions of every customer within each specified duration (week, month, etc.).  Each transaction is specified by its timestamp, type (e.g., gas, grocery, theater, etc.), and dollar amount.  The goal is then to (i) identify different “spend-ing regimes” of the customer-base, and (ii) automatically segment each timeline into its regimes. Potential contributions of this project to various business analytics tasks include the following:

• Knowledge discovery: This project will help answer questions such as “what kind of human spending regimes are there?”, “which regimes exhibit self-exciting characteristics; in activity (i.e., a spending triggers other spendings soon after) and/or in amount (i.e., a high-amount spending triggers other high-amount spendings)?”.
• Sense-making of customer population: The composition of a transaction-timeline into various regimes, and the transitions between the regimes will help characterize different classes of weekly/monthly/etc.  spending behaviors.  Based on this characterization, one can further identify different groups of customers.
• Anomaly/Change-point detection: As we will construct and learn probabilistic models of spendings data, we can compute the likelihood of any sequence of transactions over time, and flag anomalies when probability drops significantly.  We can also identify change-points in  spending  behavior  of  a  customer  based  on  anomalous  transitions  between  regimes,  e.g., by evaluating a discrepancy between predicted probability distribution over regimes before and after observing the data.  Such change-points may be indicative of life-changing events or point-of-compromise for credit/debit cards.
• Spendings predictions: As our probabilistic models are generative by design, we can use them for future predictions of likely spendings.  In particular, our models can help answer questions such as “what is the most likely timeline of spendings next week?”, “what is the expected number of spendings of a certain type (e.g., gas) next month?”, etc.  Moreover, we not only can predict what type of transactions are likely but also when they would occur into the future as we explicity model inter-arrival times.

Project Lead