The topic of machine learning in the fraud prevention space is one which is constantly on the lips of both financial institutions and merchants looking to exploit advances in IT infrastructure and intelligent computing to protect their businesses from possible danger. However, we may find ourselves asking what really is machine learning? Is it actually that effective in not only detecting fraud, but preventing it too?
To explain the first question: machine learning relies on algorithms which employ pattern recognition techniques to explore and learn the underlying structures in the data. By using past transaction data from fraudulent activity, alongside information from genuine customer transactions, these algorithms can be used to build predictive models which can forecast the probability of a transaction being fraudulent.
Predictive models deliver very tangible results in fraud detection. Their ability to extract meaning from complicated data means that they can be used to identify patterns and highlight trends which are too complex to be noticed either by humans or through other automated techniques. By running specific, effective algorithms and using them to make automated decisions, or generate alerts for suspicious activity, these techniques can save manual review time, reduce the number of false positives and quickly stop attempted fraud.
However, this approach is by no means new. In fact, predictive models first became popular almost two decades ago, particularly with financial institutions which successfully used models to detect significant volumes of card-present fraudulent transactions and save millions.
Back then, however, fraud problems were simpler and patterns were easier to identify. Fraudsters have since become savvier and more innovative, driving demand for further change in fraud detection techniques to ensure that defensive capabilities can match fraudsters’ offensive capabilities.
Technology advances over the last decade in particular have aided the evolution of machine learning and ensured it has remained an effective fraud prevention measure. For instance, the increased availability and scale of raw computing power means that we can now process, segment and analyze data on a much larger and more complex scale. This allows fraud analysts to understand both localized and widespread occurrences of fraud. It also enables these complex processes to be accomplished faster, frequently in real-time.
Additionally, other information, such as data resulting from web-behavior analysis, can be fed into the predictive models, creating a new and valuable dimension to the model’s accuracy.
The development of new algorithms, machine learning techniques and programming expertise have also all kept pace with changes in the payments and ecommerce landscape, with these latest techniques giving businesses the power to explore a much larger search area in the model optimization space and increase detection rates.
While it is clear that machine learning has a lot to offer to financial institutions and merchants in an effort to detect and prevent fraud, the approach does have its limitations.
As they learn from experience, predictive models cannot learn or spot monolithic events such as data breaches. For these you need to be running a rules-based model which uses negative lists and, preferably, consortium data.
Predictive models are also less adaptive at learning one-off events or transient phenomena. Our experience with customers around the world has taught us that combining predictive models with a customized rules engine delivers the optimal fraud prevention solution. The ability and flexibility of a comprehensive rules engine to deal with seasonal changes, emerging trends and one-time events complements the sophisticated pattern recognition techniques deployed by predictive models.
In the future it seems that machine learning and predictive models will be an integral and vital part of a winning fraud strategy. Backed by these predictive models, rules-based systems are constantly updated to augment performance and provide multifaceted coverage and protection. It is this holistic approach to fraud prevention that provides effective protection against the risk of fraud without compromising customer service, driving costs further upwards, or increasing the demand on scarce, in-house resources.
Perhaps one of the main challenges will lie with the needs of the industry which are constantly changing. With this in mind, developments and enhancements will need to continue to meet the needs of the industry as both consumers and fraudsters adapt their behavior. As fraud develops, it is imperative that predictive models develop too.
One such way of achieving this would be through exploring the use of smaller, more focused and tactical models, trained specifically on a closely targeted set of data – for example, a specific merchant sector or geography. If this is carried out, merchants will benefit from more sophisticated solutions which are faster to deploy and designed to address their specific trading landscapes, enabling them to stay that vital one step ahead.