Behavioral and Temporal Pattern Detection Within Financial Data With Hidden Information

Abstract

This paper describes a technique for behavioral and temporal pattern detection within financial data, such as credit card and bank account data, where the required information is only partially visible. Typically, transaction amount, transaction date, merchant name and type, and location of transaction are all visible data items, i.e., they are readily available in the financial institutions database. In contrast, the transaction status as a business transaction (using a personal card), a personal transaction, an investment related transaction, or perhaps a suspicious transaction, is information not explicitly available in the database. Our behavioral pattern detection technique combines well-known Hidden Markov Model (HMM) techniques for learning and subsequent identification of hidden artifacts, with run-time pattern detection of probabilistic UML-based formal specifications. The proposed approach entails a process in which the end-user first develops his or her deterministic patterns, s/he then identifies hidden artifacts in those patterns. Those artifacts induce the state set of the identifying HMM, whose remaining parameters are learned using standard frequency analysis techniques. In the run-time pattern detection phase, the system emits visible information, used by the HMM to deduce invisible information, and sequences thereof; both types of information are then used by a probabilistic pattern detector to monitor the pattern.