Addressing Fraud with Advanced Analytics
Credit card fraud is on the upswing, impacting consumers and every company that handles credit card transactions. What can an enterprise do to prevent it and protect itself and its legitimate consumers?
- By Troy Hiltbrand
- April 14, 2016
Year over year, there is a significant increase in credit card fraud. This is scary enough for the end consumer, whose credit cards are compromised, but what about the businesses that are caught in the middle, left without product and without the funds to cover their expenses? As credit card fraud is identified by banks and credit card companies, they issue a chargeback to the company where the initial transaction took place. Oftentimes, at this point the company is responsible for any product that left the facility to a fraudulent destination, the cost of the product and shipping, and sometimes steep charges from the credit lending agencies.
What is a small to midsized business supposed to do when fraudulent chargebacks quickly increase and the cost of doing business becomes a significant burden? This is where advanced analytics can be implemented as part of the transaction to ensure that the fraudulent transactions never happen.
To accomplish this, the first step is to analyze the data and identify the characteristics of a fraudulent transaction. Some of these data elements will be held only by third parties, such as credit providers. Other data elements will be in the transactional system of the company doing the selling. The idea is to work backwards, identifying known cases of fraud in the past and identifying similar patterns within those transactions.
Examples of this could include who is making the purchase, where they want the item shipped, the IP address of the transaction, and whether the IP location matches the shipping location. Each of these attributes starts to develop a profile of the transaction. By comparing fraudulent and non-fraudulent transactions, patterns start to emerge. At times, these patterns are very subtle and can only be effectively extracted through statistical analysis.
Once these characteristics are identified, there is a maturity cycle for implementing not only fraud detection but a fraud prevention program.
The first step is to codify rules around these characteristics and implement them as part of the transactional process. Examples of these include prevention of shipping to certain countries, limiting transactions to a certain threshold (or adding an approval process only for transactions that fall outside those limits), and blacklisting certain IP addresses or IP address ranges from creating transactions. The implementation of manual rules can stop the first segment of easily identifiable fraud. The challenge comes when criminals identify these filters and evolve to avoid these rules. If their target payback on the fraud is not sufficient, many will move to other vendors that are less stringent or without any fraud controls in place.
The second step is to utilize machine learning to identify these characteristics and provide a score for each transaction. Machine learning uses historical data to establish a mathematical model representing the statistical probability of an instance being either fraudulent or clean. The machine-learning algorithm is fed examples of both fraudulent transactions and non-fraudulent transactions.
From this, it extrapolates what attributes are most statistically relevant to the categorization of the fraudulent transactions. With machine learning, the answer is not black and white, like the codified rules. Each transaction is scored on how likely the transaction is fraudulent. At this point, the company has to set a risk threshold for which transactions to allow through and which ones to stop.
With machine learning, the models can be regenerated over time. By providing new examples of fraud, the model will evolve and be able to address changing behaviors by the criminals.
This scoring model can be added as an up-front filter on the transaction to stop blatantly fraudulent transactions, allow others through with a warning flag associated with them, and cleanly pass others through for fulfillment.
Finally, as criminals become more complex, the machine learning needs to evolve. The most aggressive step goes beyond looking only at the attributes associated with the transaction and takes into consideration connections to other transactions. These connections include the implied social relationships that exist between the company's clients. Determining attributes such as the similarity of names or locations between fraudulent activities provides an enhanced set of attributes for machine learning to use in the model development process. This starts to create an implied social network of the customers and allows for a model that can flag complementary fraudulent activity in the system even when the transaction itself appears valid.
Credit card fraud is harmful to consumers and to companies trying to legitimately transact business in today's environment. As your company gains visibility, it also becomes a bigger target for fraudulent activity. Implementing rules and machine learning can help you ameliorate the risk your company carries by developing a profile of fraud and stopping it before it has a chance to be committed.
Troy Hiltbrand is the chief information officer at Amare Global where he is responsible for its enterprise systems, data architecture, and IT operations. You can reach the author via email.