It's no secret that machine learning is essential for any company that wants to stay ahead in the fast-changing world of fraud. Forrester published a report in 2015 entitled "Stop Billions in Fraud Losses with Machine Learning" which found that "legacy fraud management mechanisms fail in today's economy" and that machine learning was not only beneficial but necessary. In the time since that report was published the truth of these conclusions has only become more clear.
Machine learning makes fraud detection more accurate, more scalable and much faster. Businesses can keep up with the enormous amount of data flowing in, prevent loss more effectively and provide a better customer experience since fewer good orders are declined or delayed.
Artificial Intelligence and Fraud Protection
It's easy to understand why machine learning has caused such a stir in the fraud detection community. In some ways, fraud prevention seems like the ideal use case for this form of what some are calling Artificial Intelligence. There's a huge amount of data attached to each transaction, there are strong trends that can be found in buying patterns and habits (both those of good customers and those of fraudulent ones), and there's an answer at the end of the day: a transaction will be charged back, or not. This means that you can tag the data and give the machine feedback, which is how it learns to be more accurate over time.
Many companies have welcomed unsupervised machine learning with open arms, and been intrigued by the unexpected connections the machine was able to make. Since the machine approaches the data objectively, and by nature is equipped to deal with large amounts of data, it is able to spot trends and links that humans, whose view is more limited, are not able to see.
However, unsupervised machine learning in fraud prevention will only take you so far. We may talk of Artificial Intelligence, but so far, although machines can learn, they cannot think. They can only process and analyze the information they're given. That means the results you get are dependent on the data you've given the machine. Put bad, incomplete or misleading data in, and you'll get bad, incomplete or misleading data out. A machine won't try to figure out what's wrong with the data, or work out a way around that problem, or be visited by a moment of insight that shows the whole thing in a new light.
The Specific Challenge of Subjects that Fight Back
Machine learning in fraud prevention has proved a tremendously valuable tool but not a silver bullet. While the speed and scale that machines are capable of have transformed what is possible in fraud detection, they aren't enough by themselves to solve the complex challenges involved. A machine can be better at predicting the weather, or at identifying a photo, than a person. But neither the weather nor the photos fight back. They're not trying to conceal themselves, or trick the machine into thinking they're something else, or developing new techniques.
Fraudsters, of course, do all these things. Their whole approach is to confuse the data, to make their profiles look more convincing than they are, to adopt known characteristics of good buyers, to fly under the radar. They are continually coming up with new ways to hide themselves, and to enable full identity shifts at a speed Clark Kent would envy. As well, they constantly develop new fraud attacks, with new techniques or new technology. Sometimes they'll work in rings to confuse the picture further.
All of this is confusing for a machine. When things change quickly, it can be hard for a machine to keep up. Machines excel at "big data" - at finding the patterns in truly huge datasets. But fraud, specifically, is often about "small data". You want to catch the fraudsters before they've successfully stolen tens of thousands of dollars with a new technique that they were able to use over and over again before your system caught on. Essentially, you need to out-think them.
Adding in the Human Element
What many fraud teams and companies do is use machine learning up to a point, and then hand the problem over to manual reviewers. The review team will be responsible for going over transactions which appear suspicious in some way. According to the Annual Fraud Benchmark Report of 2016, 83% of US merchants rely on manual reviews. They apparently have a lot of suspicious-looking orders, since an average of 29% are reviewed - meaning more than a quarter of customers are delayed, waiting in confusion for their order to be confirmed, and frustrated by longer fulfillment times. Of course, the manual reviews do stop some orders, but since 82% of those reviewed orders are eventually approved there's clearly a flaw in the system, particularly from the customer experience perspective.
Moreover false positives (good orders mistakenly rejected) are a growing problem. The Javelin Financial Impact of Fraud 2016 report found that 30% of orders declined as fraud are believed to be legitimate. In fact the report found that the amount lost to false positives outweighs the amount lost to chargebacks by more than 5:1. And that's at a time when chargebacks are a growing fear as well, as more fraudsters move online post-EMV adoption in the US. The Global Fraud Attack Index found that fraud attacks have increased notably every quarter since the adoption date. Online merchants from all industries have felt the pinch.
Machine learning isn't enough by itself, but manual reviews bring back all the problems that machines were helping to solve. They're slow, they're not scalable, and they're not that accurate. And a reviewer whose job it is to look at one transaction after another and provide a decision as fast as possible loses the powerful 'bird's eye view' impression of the data that helps the machine to find patterns and note trends.
Fraud departments know that this is a problem, but struggle to keep their reviewers up to date with developing trends and to provide time for research over and above that needed to decision individual transactions. The Financial Impact of Fraud study found that 56% of merchants say fraud mitigation training time is difficult to set aside. Even those who manage generally can't prioritize it - and fraudsters move fast, so constant innovation is essential.
Forter and Machine Learning
Forter's approach is different. We add the vital human touch to machine learning fraud prevention through ongoing research and analysis which goes to inform the machine. We do not perform manual reviews; the system is fully automated and provides instant approve/decline decisions. Looking at transactions on both an individual basis and on a macro level is how we deepen our understanding of the current buying trends, transaction patterns, behavioral analytics and so forth.
Machine learning is an integral part of Forter's fraud prevention solution. Instant, highly accurate decisions would not be possible without it. Forter’s machine learning technology combines advanced cyber intelligence with behavioral and identity analysis to create a multi-layered fraud detection mechanism that learns from every transaction, tailoring itself swiftly to a particular merchant’s risk profile.
It is, however, equally important to understand the role that human intelligence plays alongside and together with the artificial intelligence. Forter’s team of experts engage in constant research into transaction data, consumer buying patterns, technological possibilities and the fraudster ecosystem itself. They continually improve the system and adapt it to meet the latest fraudster techniques.
Working in combination with the machine, that means that the system learns automatically, all the time, and is also continually refined by the research and analysis of Forter’s expert team. Forter’s team, and the machine itself, ensure that the system is continually updating, reacting to new data to become more and more tailored to each merchant's risk profile.
Since the essence of the system is this unique combination of human expertise, experience and research with machine learning - the combination of human intelligence with artificial intelligence - we call this driving force behind our accuracy "Integrated Intelligence."
The Best of Both Worlds
With Integrated Intelligence, you can leverage all the advantages of machine learning without compromising on accuracy. Forter's fully automated system is fast, smooth and designed to scale. It provides sub-second decisions no matter how much traffic your site is getting. But as well, it's highly accurate - so accurate that our customers see both an increase in sales and a reduction in chargebacks.
To find out more about the way Forter combines machine learning with human expertise, and what it can do for your business, contact us.