Hundreds of fraud systems in the world claim to do the same thing: fight fraud and limit fraud losses. But how can an enterprise know which systems truly work? Which ones may be too restrictive, leading to false positives? Which may not be restrictive enough, allowing fraudulent charges to pass through?
Historically, most fraud systems were developed somewhat manually by analysts who learned about new fraud trends and techniques and responded by adding rules to their fraud system. In theory, the new rules would limit or block transactions believed to be fraudulent. Unfortunately, these rules are based on humans reviewing fraudulent transactions and looking for commonality. What they usually fail to do is evaluate the impact the rule will have on legitimate cardholders and the number of good transactions that are declined erroneously.
Over time, as more fraud trends emerged, and more fraudulent charges occurred, more and more rules would be added to the system, eventually resulting in a difficult, often contradictory list of rules and restrictions.
For example, a bank in Dallas, Texas noticed several of their cardholders had large charges from a country overseas that occurred online at odd hours. And sure enough, some of those cardholders called in to report that they have not been overseas or charged anything from an international website. With a rule based fraud system, the fraud administrator could code in a new rule that denies transactions from the country concerned that occur during specific times.
Although this could be temporarily effective, one of two things are likely to occur: the fraudster will mask the IP address to make it appear that the charge is coming from a different country, or a good cardholder will actually go overseas and their card will be declined. Both situations are potentially bad for the cardholder, the merchant, and the bank.
In the last decade, machine learning has taken on a much larger role within fraud systems. With the massive amounts of data readily available and machine learning models getting more and more sophisticated, it’s no surprise that machine learning based fraud systems are the way of the future.
How does machine learning work?
Machine learning consists of several elements including but not limited to: data, algorithms, and models. Let’s start with the data. Terabytes of transaction data are available with payment processing. This creates a terrific environment for a machine learning based system.
Then we have algorithms. Algorithms have been studied for decades and with the popular rise of computer science, algorithms have become more sophisticated over time. Some of the more common algorithms used within machine learning include neural networks, decision trees, and Bayesian networks, among others.
Essentially, algorithms allow for a faster, more accurate analysis of data. Algorithms can quickly identify patterns within the data. When we think about card fraud, if we have a data set that includes fraudulent transactions, we can test different algorithms to see which ones more accurately identify fraudulent transactions, and which do not.
This leads us to models. When you combine the mountains of data available with algorithms that have been optimized by testing different data sets, you can begin to build models. This is the part of machine learning which is classified as “learning.” As the models are developed, the algorithms can “learn” and “adapt” and make informed “decisions” as to which data may or may not be fraudulent. They can detect good versus fraudulent transactions more effectively than humans can.
This is the key reason why machine learning is so much more effective over rule based systems. With so much data available, it is impossible for humans to recognize new, potentially fraudulent patterns and create rules that won’t limit the transactions in a negative way.
Why rule based fraud systems fail
Rule based systems are limited to humans analyzing data, deciding what rules to write, and then adding the rules to the system. But humans are simply too underpowered to make accurate assumptions about data because there is too much data out there to dig through.
This is why machine learning has come of age: with so much data available, and decades of algorithm research, computers can process data so much faster than humans.
The important point, though, is that machine learning is only as good as the data and the model. Even though machine learning is “intelligent,” it still requires iterative improvements. This is called “training the models” and must occur on a regular basis in order to keep the model aligned with reality and current events.
Machine learning gives complicated data sets a more accurate meaning, helping fraud professionals make sense of large amounts of data instead of trying to interpret data patterns on their own. They can rely on well-designed models to quickly recognize patterns in the data and make informed decisions.
And not all machine learning systems are created equally. In a proof of concept Worldpay conducted among the top vendors in this area, it found significant differences among them. To address the complexities of fraud, Worldpay has teamed up with the best machine learning companies and is in the process of developing what just might become the industry’s best hope for combatting fraud. By combining leading machine learning with our own team of fraud experts, we are about to embark on a new day for fighting fraud - FraudSight, coming soon.