Skip to main content
Big Data in Fraud: A Need to Stream

Today’s buzz around Big Data is bringing renewed attention to the need for streaming analytics—a focus that, if you ask me, is long overdue. 

In the payment card space, we at FICO are celebrating the 20th anniversary of using streaming analytics for fraud detection.  Our FICO® Falcon® Fraud Manager models utilize cardholder transaction profiles “in the stream” to update transaction patterns and features indicative of fraud.  The models leverage this in real time to generate a score that issuing banks use to stop payment card fraud.  Over the years, FICO has continuously improved Falcon model performance with new streaming capabilities, as I describe in a prior post.

One critical component of streaming analytics, particularly in fraud detection, is the ability to periodically update models in production with new data.  To do this in Falcon Fraud Manager, we use merchant data, a highly predictive Big Data source.  By capturing this merchant view of transactions, our fraud analytics can detect suspicious activity not visible from cardholder behavior patterns alone. 

Through weekly updates of merchant fraud risk profiles, the models—which we call FICO® Fraud Predictor with Merchant Profiles —detect fraud more accurately, faster and with less impact on good customers. 

How does it work?  Well, we receive data on millions merchants worldwide on a daily basis.  We use this to develop merchant risk assessments depending on transaction type (card-present vs. card-not-present) and where the transactions happened (domestic vs. cross-border).  FICO then sends updated merchant profiles electronically to all Fraud Predictor customers on a weekly basis. 

We leverage Big Data assets like merchant data to improve the performance of our models.  The graphic above compares three UK credit models; the Fraud Predictor models leverage merchant data, and the Falcon model doesn't.  The older Fraud Predictor v12 model detects 10% more fraud than a newly trained Falcon v13 model because of the added predictive value that the merchant data provides in the stream.

related posts