Real-time Graph Analysis of Financial Data Creates Potential for Millions in Fraud Detection Savings

The Challenge

The financial services company collects a huge amount of data provided by its customers, as well as enriched data from outside vendors that needs to be analyzed in real time before a transaction can be approved. While the majority of these requests are approved or denied instantly through an automated fraud detection system, potentially fraudulent requests are submitted to an analyst for manual review.

The analyst has a dedicated transaction review tool encompassing all relevant third-party data. However, the analyst had to query a Microsoft SQL Server database to review customer history for an association with fraud rings. These queries can require four or more levels of JOINs, impacting the SLA for reviewing the transaction within the manual review queue.

“It was taking five minutes or more to run a query,” said a product manager for fraud detection solutions at the company. “And since our analysts were having to review 10,000 daily transactions, this wasn’t sustainable. Also, a relational database wasn’t the right solution to perform link analysis queries so it placed a huge burden on our database.”

Not only that, but the queries would return complicated data that the analyst had to review in a matter of minutes.

The company needed to find a more efficient way to analyze the data to save time for both their waiting customers and analysts. They explored a number of data package solutions and other data visualization tools, but “the performance wasn’t high enough, wasn’t available in real-time, and wasn’t scalable,” said the product manager.

The Solution

While the company had initially sought a turn-key solution to avoid a drawn-out development process, the actual development time with Neo4j wasn’t too complex or time-consuming with a pilot project being easily completed during spare time.

Neo4j provided both real-time results with connected data and data visualization that allowed analysts to make faster, more accurate decisions. This opened the door for newer, more extensive searches, which the company hopes to expand from four to 10 degrees of separation.

Also, the company analysts began noticing clusters and relationships between their data which uncovered new, previously unnoticed potential fraud connections. Each customer can be represented by up to 30 nodes, each with up to 60 properties, totaling over 216 million nodes and 680 million relationships and 20% annual growth in database size.

Download Case Study