How can we harness payments data and manipulate it to determine if a transaction is legitimate (or more accurately, if it is not) and reduce customer churn? Here’s a guide to some of the advanced analytics that can underpin a successful approach to fraud detection. This should allow you to start separating the hype around artificial intelligence, from the machine learning, neural and regression models and the rule-based logic we have all come to embrace in the practice of making sausage in a fraud shop.
The role of behavioral profiling and adaptive machine learning
Enter behavioral profiling, the capacity for one’s own behavior to influence fraud detection strategy. The merchants we typically visit, the fuel pumps where we get gas, common travel destinations, these are all a part of the behavioral profiling technology that we’ve been using for years. Questions arise… How often do we buy from retailers that sell women’s ready-to-wear clothes? What’s the potential that someone will spend their holiday in Belize? Is this an unusual amount that someone is taking out at this ATM and is it the first time they are using this specific terminal on the other side of town? The technology to evaluate these scenarios has been around for a while, it’s mature and widely accepted as the “table stakes” feature functionality to reduce risk of fraud and maintain a good customer experience.
Behavioral profiling is also useful in models, whether it be a regression model (expanding on the example above at the ATM), neural scoring models, or going down the rule-based machine learning path. Adaptive machine learning will typically leverage multiple data sources, using up-to-date variables to provide the greatest possible timeliness in both legitimate and fraud transactions. This integration of many data points and risk indicators may include; recent fraud trends, legitimate spending patterns including non-monetary transaction elements like the addition of beneficiaries or changes to demographics, end-point device intelligence, resident malware indicators, authentication results and third-party or internal scoring models.
So, while all these elements are assembled together into a larger complex regression model, there is one additional element that can be integrated into the model that enhances it with statistical properties; the risk-weighting of the various signals and data elements inside of the regression model. This allows for the model to accurately assign an optimized predictive capability to these data elements, which will then accurately calculate the relative risk of the data element. This process produces reliable and repeatable decisioning logic, identifies insights into the legitimacy of the transaction and provides value relative to historical relationships and trends, aligning the model logic to inbound transactions, recognizing patterns and delivering the value of advanced analytics in fraud detection.
But is AI ready for primetime?
Pattern recognition in supervised machine learning (where the machine is provided example inputs, perhaps exemplified in the data elements described above) is not a new science; we’ve been delivering these models for years. What is new is the hype around this process and the introduction of Artificial Intelligence (AI) concepts in the fraud detection space. Here’s the deal with unsupervised AI… it’s not ready for primetime, performing under the legacy supervised machine learning analytics applications that are deployed presently in the smarter financial institutions and processors. It’s simply not fast enough, smart enough or cheap enough to be implemented at scale, so when you hear people say AI and fraud in the same sentence, you might be smelling actual fraud.
AI is suggested, expected and advertised to be the technology holy grail that will reduce human supervision and oversight, minimizing the manual analytical work load and constant strategy maintenance that is the backbone of any fraud analytics team. The end goal of AI is to get the computer to minimize the amount of work from humans and transition it to the machine, ultimately to realize a lift in efficiency and accuracy in fraud detection. The computer can indeed do things faster than humans, but the human will always know better about the reasons behind the signals… and for this reason, while the goal is admirable, its unlikely to be fully realized.
Utilizing the existing supervised machine learning strategies is presently delivering the best probability of aligning the stars of a high detection rate AND a positive customer experience. This is again, table stakes for financial institutions as the culture of fraud detection moves further toward the best customer experience metrics. Because I shouldn’t get a decline when I am going back to Arizona for Christmas again or having my favorite meal at that one Poutinerie in Montreal with the best smoked meats. But if you’re my bank and you fail to detect more than a transaction in a country I’ve never been to, I’ll be upset over it.
Every new payment type brings new fraud challenges – download our whitepaper ‘The Fraud Trap: Optimizing Digital Payment Controls from Day One‘ [PDF] to find out more.