Machine learning for fraud detection: using technology to find more fraud

Over the last few years, there has been an increase in the number of fraud services available in the market and with this it has become equally as important to provide a consolidated view of the overall fraud risk of activities such as credit applications, transactions, or log ins.

And whilst protecting customers through detecting fraud is high on organisations’ agendas, the need to uncover fraudulent activity must be balanced against the importance of not adversely affecting the experience of genuine customers by subjecting them to unnecessary delays and checks.

The emergence of more data sources, at a more granular level, with greater accuracy and control opens us up to new opportunities in detecting fraudulent attempts. Across the entire data-landscape, we are seeing how data is changing, and how using better, more accurate and validated data can help organisations make more robust decisions.

Whilst many organisations will have adopted traditional rules-based strategies, these require a lot of effort to manage and with the increase in services and data points that feed into them, the number of permutations is becoming too large for humans to deal with. With the increase in digital traffic and need to identify and accept good customers whilst preventing bad, and more data points to add into traditional risk models, optimal fraud prevention techniques have naturally evolved into ones powered by machine learning (ML).

Why adopt machine learning in fraud detection?

With the increased number of complementary fraud services available today, it has become more important to provide a consolidated view of the overall fraud risk.

As more and more consumers are requiring instant decisioning and fulfilment of services, there has been an increased need for more accurate fraud checks to be built into the customer journey, rather than delaying decisions or transactions by introducing offline authentications. When making online decisions where the customer journey will be affected, it is even more important that only those activities with a real risk of being fraudulent are prevented.

Conversely, it is also important to not adversely impact the experience of genuine customers by subjecting them to unnecessary delays whilst fraud checks are completed. Consumers may well walk away, resulting in the loss of a potentially profitable customer, if subjected to additional time consuming verifications.

Additionally, there is a limit in the effectiveness of a rule-based referral strategy as it requires a lot of effort to manage and the number of permutations, which can become too large for a human to deal with.

Adopting a machine-learnt approach can help organisations embed a smooth customer journey, whilst flagging potential fraudulent attempts which could have impacted bottom-line.

“Losses through fraud could be crippling if not kept under control. When considering where to invest, organisations don’t need to throw money at site-fraud teams, but rather consider if ML would be a suitable alternative.”

 

James Torselli, Product Manager, Experian UK&I

Common machine learning misconceptions

1. Machine learning is new technology

The techniques we have been using for over 15 years in our scorecards in both credit and identity and fraud are branches of machine learning. Whilst many believe machine learning to be a new solution, we have been refining techniques to give more accurate outputs for well over a decade.

2. Machine learning is self-learning

Not necessarily – whilst models can continually evolve based on recent experience, this doesn’t have to be the case. Within identity & fraud, we are not deploying auto-learning models on the basis that this makes the governance aspects much simpler. Model performance is continually monitored and where it has degraded, will be refreshed. Replacement models will only be deployed with full visibility of the differences between that and the previous model, allowing referral cut-off to be set based on the desired volumes of referrals/fraud detection.

3. “Supervised models” means people are involved and “unsupervised models” does not

“Supervised” and “unsupervised” models are terms relating to the availability of outcome data, not human involvement. All our models involve an element of human validation.

“Machine learning is really giving organisations the opportunity to streamline their decisioning process; increase the onboarding of good customers, find more fraud, whilst ultimately keeping operational costs down.”

 

Sarah McCallum, Product Manager, Fraud Solutions UK&I

Who can use machine learning to detect fraud?

Whilst many organisations already use machine learning within certain processes, implementing a greater depth of automation within fraud detection is still seen by some as expensive and inaccessible.

However, with the vast number of data points now needed to accurately identify fraud, the use of ML is becoming essential. Machine learning has the ability to look for and understand complex relationships between all the data points within an application, giving organisations highly predictive fraud models that enable them to minimise false positives, generate fewer referrals and produce better outcomes for good customers.

Although typically, access to machine learning for fraud has been limited to organisations with memberships to consortia and ability to invest in the top-line capabilities, solutions are becoming available to smaller start-ups, challengers or mid-market organisations wanting the benefit of machine learning, without the barriers. Access to richer, more granular data, enabling more effective fraud scoring, is no longer just for those organisations with the deepest pockets. But rather one available to all, to offer consumers and lenders, a more robust level of fraud-prevention.