Deploying Next-Level AI to Optimize Payment Fraud Detection
Written by: Steven D’Alfonso, Research Director, IDC Financial Insights
Fraud and cybercrime have moved on and adopted new technologies and methods; however, many fraud detection systems have not. In today’s environment, bad actors are using machine learning technology to execute attacks with greater speed and scale. In addition, the threat landscape is widening as merchants and card issuers must contend with identity-based frauds like account takeovers as well as friendly-fraud schemes, promotions abuse, and loyalty program protection. Rules-based fraud systems and basic supervised machine learning are no longer enough to keep pace with the increasing sophistication of fraud and cybercrime. The need for intelligent fraud detection using next-generation artificial intelligence (AI) has become necessary.
Rules-based systems are a set of conditions, which if met, can label a transaction as fraudulent. They are simple and straightforward and perfectly fine in identifying fraud that is also simple and straightforward. The nature of ecommerce, new payment schemes, and faster payments has given way to more complex, difficult to identify fraud patterns. The next generation of AI fraud detection can identify complex relationships between entities and transactions at a scale and in the context of desired business outcomes, augmenting the human intelligence of fraud analysts.
Machine learning is an element of AI – think of it as a tool inside the AI toolbox. Understanding how AI can transform your fraud detection requires understanding what it can do for you and how you interact with it. IDC views AI-based automation or intelligence as a five-level evolution:
- Human led
- Human led, machine supported
- Machine led, human supported
- Machine led, human governed
- Machine led, machine governed
Determining the level at which you are operating requires answering three questions:
- Who produces insights?
- Who decides and how?
- Who acts based on decisions?
The answer to each question, human or machine, will help determine what stage you are at and what it will take to progress to future stages. Rules-based systems operate at the lowest level with limited technology and humans making decisions based on experience and rules. In today’s landscape of ecommerce, with new payment schemes and faster payments, fraud detection systems must have operational capabilities at the machine-led, human-governed level.
Operations at the machine-led, human governed level are characterized by the machine:
- Producing insights by analyzing hundreds of data points around merchants, consumers, orders, devices, intelligence across networks of merchants, et al.
- Making decisions within a framework of human governance, synthesizing all the data collected to determine that a transaction is fraudulent.
- Executing a decision based on analysis showing a transaction is fraudulent, issuing the decline, and transparently showing the parameters that led to the final decision.
The use of machine learning in fraud models is not new. However, in today’s environment, business line managers and fraud teams must work together to align fraud modeling and business outcomes. Elevating your fraud analytics engine to an optimized machine-led, human-governed ecosystem is essential to managing operational expenses, fraud losses, chargebacks, and, most importantly, customer experience.
To get to an optimized level, your fraud detection models should use a combination of supervised and unsupervised learning, advanced anomaly detection, and draw insights from network-based intelligence. Supervised learning is used to mimic the thinking of a fraud analyst. These models must be trained on enormous amounts of data for them to simulate how an analyst would logically think about elements of a fraud scenario. Unsupervised learning models aim to simulate the instincts of a seasoned fraud analyst and are good at picking up nuanced patterns.
Combining both the logic of supervised learning and the intuitiveness of unsupervised learning provides a basis on which a fraud management program can achieve optimized levels. Machine learning only captures the first part of being machine led and human governed. AI without the ability to control and govern how the machine behaves is ineffective. Machine algorithms can be programmed to catch fraud all day, every day, but that is not always ideal. An optimal solution will have the ability for business line managers, marketing, and fraud to work together to determine appropriate risk levels across segmented populations of customers. An optimized next-level AI fraud solution will connect to business outcomes by providing control over fraud risk thresholds, decline rates, and fraud operations costs.
Achieving payment fraud detection optimization requires several key elements:
- A blend of supervised and unsupervised machine learning models
- Advanced anomaly detection capabilities
- Network intelligence curated from scores of merchants and card issuers
Also required is a flexible policy engine that puts control in the hands of risk managers and marketers to effectively govern the machine-learning models and enable strategic-level thinking, connecting fraud management to business outcomes.
A next-generation AI fraud solution can bring all these required elements together to simulate the work of a fraud analyst at speed and scale that can match the speed and scale of today’s fraud and cybercrime.
Read the full IDC Technology Spotlight to learn how AI-based fraud management can help companies achieve desired business outcomes while maintaining risk mitigation.