Fraud Detection with Machine Learning
It seems at times that fraudsters and the ecommerce industry are caught in a never-ending war of escalation, each side always trying to get a technological advantage over the other. It’s certainly true that fraud has gotten ever more sophisticated, and that advanced anti-fraud tools have become a necessity for many merchants.
New methodologies, after all, are required to deal with new threats. In particular, machine learning techniques are increasingly touted as a way to screen out fraudulent transactions more effectively. How does machine learning work, and what makes it such an important part of current fraud prevention solutions?
In theory, every fraudulent transaction should leave some clues as to its true nature—some inconsistency in the data that reveals the person behind it to be a fraudster, not a real customer. Fraudsters have gotten better about masking the big, obvious clues.
For example, they might use proxy servers to conceal the fact that their geolocation data shows them accessing the internet from a country known to have high fraud rates, nowhere near the shipping and billing address of the cardholder they’re victimizing.
Still, the thinking goes, there should be subtle clues in the behavior and transaction data that can give the fraudster away despite their best efforts. Machine learning is a method for going after these clues by analyzing sets of data for the trends and patterns that can be used as a basis for making future predictions. Its demonstrated effectiveness in this capacity has made it an integral feature of many of the latest anti-fraud solutions to hit the market.
What is Machine Learning?
Machine learning is an offshoot of artificial intelligence that focuses on using large amounts of data to train a software program to identify patterns within the data and make predictions based on them. Once the program learns to recognize certain patterns within existing data sets, it can review incoming data to look for similar emergent patterns and predict the values of as-yet-unknown data within that pattern.
In other words, if you show a machine learning program enough examples of fraudulent transactions, it should be able to analyze your future transactions and predict the likelihood that they are fraudulent, even in the absence of obvious indicators that previous generations of anti-fraud tools would have relied upon.
In the past, anti-fraud tools were stuck using a rules-based approach. You could set rules for what was and was not a red flag for fraud, and the tool would screen transactions accordingly.
For instance, you could block orders from certain countries, or from any device or IP address that had previously submitted an order through a different customer account. The tool would dutifully obey, and it was up to the merchant to figure out how to handle the inevitable exception cases.
The preferred approach for both catching more fraud and reducing false positives is risk scoring, where fraud indicators are assigned a point value and the transaction ends up with a final score that tells the merchant whether to accept, reject, or hold the order for manual review.
Machine learning saves merchants from having to come up with arbitrary point values—and yet more exception cases—for a risk scoring system.
Instead, the machine learning process functions as a way to draw statistical correlations between data points to create an accurate and reliable model for scoring the risk of fraud.
How Does Machine Learning Improve Fraud Detection?
Machine learning is an especially promising solution for credit card fraud because the rules-based approach is increasingly confounded by consumer demands for fast, seamless mobile shopping experiences. Rule-based systems can get confused by shoppers who switch devices or networks frequently, and asking customers to provide extra forms of identity verification can lead to checkout friction and cart abandonment.
Instead, machine learning solutions can use behavior analytics to identify fraud signifiers, which can be especially useful in cases of account takeover fraud, which can fool rules-based systems into thinking the transactions are coming from known customers.
By looking at past transactions, order history, and other recorded aspects of a customer’s behavior, machine learning tools can tell when a new and unfamiliar individual may be at the helm of a trusted account.
What’s the Most Effective Way to Detect Fraud with Machine Learning?
When using a fraud solution that employs machine learning, two things are of vital importance. First, the data that is being used to train the machine learning system needs to be of high quality and relevance, and second, you have to have a way to monitor the effectiveness of the system and make adjustments as necessary.
In some cases, merchants might not have enough of their own historical transaction data to use for machine learning. For these merchants, the solutions providers might use pre-existing data sets or generic models as a starting point.
To monitor a machine learning anti-fraud solution and gauge its effectiveness, you’ll need to engage in manual review of flagged orders.
This will allow you to determine what is being correctly identified as fraud, what is being falsely flagged, and what adjustments might need to be made in your model.
This will help you generate more accurate scores and set risk scoring thresholds that minimize your false positive rate, your reliance on manual review, and most importantly, fraud itself.
Conclusion
Having the right anti-fraud tools is an essential component of any decent chargeback prevention strategy. True fraud chargebacks can’t be fought—they’re completely legitimate, and it falls upon merchants to take responsibility for doing the best they can to catch and reject fraudulent transaction attempts before they can be processed.
When fraud slips through a merchant’s defenses, the result is a chargeback, the added cost of chargeback fees, an uptick to their chargeback rate, and most likely a loss of reputation in the eyes of the victimized cardholder.
To combat the many-headed hydra that online fraud resembles today, merchants need a multi-layered approach that utilizes the most effective technologies. For merchants dealing with credit card fraud and true fraud chargebacks, tools that employ machine learning can be among the most powerful and effective when it comes to preemptively detecting fraud.
Thanks for following the Chargeback Gurus blog. Feel free to submit topic suggestions, questions or requests for advice to: win@chargebackgurus.com.
