Real-Time Anomaly Detection Within Credit Card Transactions

[MUSIC PLAYING] In this digital age, credit card fraud has become a common occurrence that can mean huge losses for businesses and put consumers' personal information at risk Students at the University of Chicago set out to develop a real-time anomaly detection process to detect abnormal behavior within credit card transaction data

So how did they do it? First they needed to find realistic customers to study So they developed a methodology utilizing US census attributes of age, sex, and population density, which allowed them to create a complete, accurate, and synthetically generated data set Then they established customized spending profiles for each consumer segment This allowed them to represent and study people with diverse spending habits From there, the students identified, through a literature review, a set of significant variables that are indicative of fraudulent transactions

These fraud indicators were varied beyond normal limits to generate fraudulent transactions Their approach was based on finding the anomalies in transactional behavior by defining normal behavior and declaring any data occurrences that lie outside of that region as an anomaly They developed a statistical model to predict these anomalies Specifically, they used Markov chains, consisting of state transition matrices for each user, as well as aggregate state transition matrices for each consumer segment These state transition matrices allow them to define the probability of a user to transition between successive states

Categorized credit card transactions These probabilities, when examined over a rolling window of a user's five most recent credit card transactions, allow them to identify fraudulent transactional behavior through state transitions that had a low probability of occurring Ultimately, they were able to accurately detect approximately 835% of fraudulent transactions, with a false positive rate of roughly 05%

A very promising result