COVID fraud trends: Has your machine learning caught up?
August 11, 2020
Although machine learning (ML) relies on complex statistical methods and high-power computing, it’s essentially a very simple concept. By identifying and examining statistically high-and low-risk combinations of transaction data, machine learning software can learn to accurately predict which orders are fraudulent.
Machine learning begins with human-guided computing software that sifts through large volumes of historical data points and identifies the historically relevant relationships. Then, all this information is fed into a variety of algorithms to arrive at predictions and assumptions.
Once in place, these programs continue to consume large amounts of data and gradually get better at identifying the high- and low-risk relationships. This is where the actual learning takes place.
When a major event like COVID-19 shifts purchasing trends, machine learning programs must adapt to new transaction patterns to accurately detect fraud. We’ll break down how machine learning works and how you can help your fraud platform adapt to the current payment landscape.
Fraud detection: what machine learning can do to help
Facilitate real-time decision-making
At the core of many fraud-mitigation platforms is a rule-based system. Such systems rely on people creating rules that will determine which orders to accept, reject, or send to manual review. While these systems are often quite effective, they do require a great deal of time-consuming manual interaction. Machine learning can help reduce this time requirement by evaluating large amounts of transactional data in real time.
Criminals are increasingly creating more subtle and non-intuitive patterns to avoid detection. As these subtle patterns become more difficult for humans to identify, machine learning techniques can be incorporated to spot them.
Rapidly respond to change
Because fraudsters are always changing their tactics, it’s a constant cat-and-mouse game. Machine learning is continuously analysing and processing new data, then autonomously updating models to reflect the latest trends
Significant advances in technology have reduced the costs associated with machine learning solutions and the computing systems capable of running them. As machine learning helps improve accuracy, it can also reduce costly false positives, and the time and expense of manual reviews
COVID’s impact on machine learning fraud detection
The impact of COVID has simultaneously reinforced the importance of machine learning and exposed the risks of relying on this approach too much.
As we mentioned machine-learning models are built around historical or predicted patterns. However, the true implications of COVID is evolving daily. Machine learning only sees transactional data, it doesn’t know about external drivers affecting consumer and fraudster behaviours. COVID is a great example of this.
Therefore, the human factor is key. At Cybersource, our Managed Risk Analysts were critical in helping identify these trends and adjusting business rules to eliminate false positives.
Our merchants are seeing a surge in online transactions from first-time buyers, rises in mobile device usage, spikes in app downloads, evolving purchasing hours, more returning customers, higher ATV’s, and shorter durations on the checkout page.
To counteract false positives associated with changing consumer trends, you can combine an automated machine learning system with a rules-based approach. This is where our Managed Risk team comes into play. During this period, the Managed Risk team is helping our clients navigate these consumer changes and align business strategies to the ‘new normal’.
How Cybersource tackled COVID with machine learning
‘White box’ approach to ML was decisive
Cybersource’s fraud management platform, Decision Manager, uses 260+ checks to calculate the risk of an order. Our Managed Risk Analysts have nearly full visibility of all these checks, enabling them to responds decisively to COVID-19. When certain information codes had spikes or drops, they knew the why and how to resolve the discrepancy.
Augmenting ML with a custom score
A bespoke custom score card (profile score) was used to augment the current Cybersource machine learning models, giving it time to catch up to the changing patterns of this pandemic. Profile score has essentially become the ‘COVID-19 Score’ during this period.
90-Day vendor ML only deployment
Most companies offering ML require an average 90-day lead-time to implement a new model. However, we were able to adjust to COVID-19 almost immediately. For many companies, that 90-day lead time could be the difference between going under or surviving during the pandemic.
ML ready for a second wave
Cybersource’s existing ML models have built in functionality to help merchants respond to a second wave. Decision trees used in Decision Manager Replay can use sampling data from the first lockdown to accurately predict the fraud configuration for the second wave.
Pioneering data scientist team
Machine learning has served as a core innovation of Cybersource’s Decision Manager since 1999. Our data scientist team has a combined 100+ years working in modelling and AI. They have expertise building models against the backdrop of recessions, dramatic migrations to CNP, and unprecedented fraud attacks.
Expertise ready on ‘day one’
Machine learning takes time to recalibrate, and this can be time clients cannot afford to lose. Human expertise can instantly adapt from ‘day one’ when dramatic shifts in consumer purchasing patterns occur. The expertise of our Managed Risk team, combined with awareness of developing regulatory changes, allowed us to update strategies proactively.
Verification of evolving patterns
Decision quality depends on available data—the more data—the better the decisions are. If ML uses data from a single merchant, it will have limited power. Our Managed Risk team can confirm changing buying patterns through their extensive portfolio of clients across multiple industries. The team has access to a multitude of Visa and CyberSource data reporting sources that can be used to verify the latest trends. Once a trend is verified, we share it with our clients enabling businesses to realign their business strategies.
No single point of failure
If one point is impacted (i.e. machine learning), we have the other points ready to minimize the negative impact such as Decision Manager rules, our Managed Risk team and our Screen Management team. We allow merchants to “play by their own rules” so if machine learning needs time to recalibrate to new trends, they can adjust settings in Decision Manager to minimize any negative impact.
In summary, vendors who only offer machine learning may lack a contingency plan for COVID-19. It takes a multi-faucet approach to effectively utilize machine learning in your fraud and risk strategy.